|
article-title
|
Predicting Social Distancing Intention and Behavior During the COVID-19 Pandemic: An Integrated Social Cognition Model
|
|
abstract
|
Abstract
Background
Social distancing is a key behavior to minimize COVID-19 infections. Identification of potentially modifiable determinants of social distancing behavior may provide essential evidence to inform social distancing behavioral interventions.
Purpose
The current study applied an integrated social cognition model to identify the determinants of social distancing behavior, and the processes involved, in the context of the COVID-19 pandemic.
Methods
In a prospective correlational survey study, samples of Australian (N = 365) and U.S. (N = 440) residents completed online self-report measures of social cognition constructs (attitude, subjective norm, moral norm, anticipated regret, and perceived behavioral control [PBC]), intention, action planning, habit, and past behavior with respect to social distancing behavior at an initial occasion. Follow-up measures of habit and social distancing behavior were taken 1 week later.
Results
Structural equation models indicated that subjective norm, moral norm, and PBC were consistent predictors of intention in both samples. Intention, action planning, and habit at follow-up were consistent predictors of social distancing behavior in both samples. Action planning did not have consistent effects mediating or moderating the intention–behavior relationship. Inclusion of past behavior in the model attenuated effects among constructs, although the effects of the determinants of intention and behavior remained.
Conclusions
Current findings highlight the importance of subjective norm, moral obligation, and PBC as determinants of social distancing intention and intention and habit as behavioral determinants. Future research on long-range predictors of social distancing behavior and reciprocal effects in the integrated model is warranted.
|
|
title
|
Abstract
|
|
sec
|
Background
Social distancing is a key behavior to minimize COVID-19 infections. Identification of potentially modifiable determinants of social distancing behavior may provide essential evidence to inform social distancing behavioral interventions.
|
|
title
|
Background
|
|
p
|
Social distancing is a key behavior to minimize COVID-19 infections. Identification of potentially modifiable determinants of social distancing behavior may provide essential evidence to inform social distancing behavioral interventions.
|
|
sec
|
Purpose
The current study applied an integrated social cognition model to identify the determinants of social distancing behavior, and the processes involved, in the context of the COVID-19 pandemic.
|
|
title
|
Purpose
|
|
p
|
The current study applied an integrated social cognition model to identify the determinants of social distancing behavior, and the processes involved, in the context of the COVID-19 pandemic.
|
|
sec
|
Methods
In a prospective correlational survey study, samples of Australian (N = 365) and U.S. (N = 440) residents completed online self-report measures of social cognition constructs (attitude, subjective norm, moral norm, anticipated regret, and perceived behavioral control [PBC]), intention, action planning, habit, and past behavior with respect to social distancing behavior at an initial occasion. Follow-up measures of habit and social distancing behavior were taken 1 week later.
|
|
title
|
Methods
|
|
p
|
In a prospective correlational survey study, samples of Australian (N = 365) and U.S. (N = 440) residents completed online self-report measures of social cognition constructs (attitude, subjective norm, moral norm, anticipated regret, and perceived behavioral control [PBC]), intention, action planning, habit, and past behavior with respect to social distancing behavior at an initial occasion. Follow-up measures of habit and social distancing behavior were taken 1 week later.
|
|
sec
|
Results
Structural equation models indicated that subjective norm, moral norm, and PBC were consistent predictors of intention in both samples. Intention, action planning, and habit at follow-up were consistent predictors of social distancing behavior in both samples. Action planning did not have consistent effects mediating or moderating the intention–behavior relationship. Inclusion of past behavior in the model attenuated effects among constructs, although the effects of the determinants of intention and behavior remained.
|
|
title
|
Results
|
|
p
|
Structural equation models indicated that subjective norm, moral norm, and PBC were consistent predictors of intention in both samples. Intention, action planning, and habit at follow-up were consistent predictors of social distancing behavior in both samples. Action planning did not have consistent effects mediating or moderating the intention–behavior relationship. Inclusion of past behavior in the model attenuated effects among constructs, although the effects of the determinants of intention and behavior remained.
|
|
sec
|
Conclusions
Current findings highlight the importance of subjective norm, moral obligation, and PBC as determinants of social distancing intention and intention and habit as behavioral determinants. Future research on long-range predictors of social distancing behavior and reciprocal effects in the integrated model is warranted.
|
|
title
|
Conclusions
|
|
p
|
Current findings highlight the importance of subjective norm, moral obligation, and PBC as determinants of social distancing intention and intention and habit as behavioral determinants. Future research on long-range predictors of social distancing behavior and reciprocal effects in the integrated model is warranted.
|
|
abstract
|
Subjective and moral norms and perceived control were correlates of social distancing intentions, and intentions, action planning, and habit were correlates of social distancing behavior, in samples of Australian and US residents
|
|
p
|
Subjective and moral norms and perceived control were correlates of social distancing intentions, and intentions, action planning, and habit were correlates of social distancing behavior, in samples of Australian and US residents
|
|
body
|
Introduction
The novel coronavirus disease 2019 (COVID-19) pandemic has emerged as a truly global public health crisis [1]. While symptoms of COVID-19 are relatively mild without serious consequences in the majority of cases [2], modeling data suggest that approximately 4% of the global population is at risk of severe COVID-19 if infected and may require hospital admission for treatment [3]. Furthermore, SARS-CoV-2, the virus that causes COVID-19, is highly contagious, spreading mainly through person-to-person contact. Government-mandated measures to reduce transmission include advocacy of behaviors like wearing face masks and social distancing and issuing “shelter-in-place” orders and bans on public gatherings [4].
Social distancing, defined as maintaining a distance of at least 3–6 feet (1–2 m) from other people not from the same household is considered particularly effective in minimizing SARS-CoV-2 transmission [5, 6]. “Shelter-in-place” orders (also referred to as “stay-at-home” or “lockdown” orders) represent means to mandate social distancing by minimizing incidences of person-to-person contact outside individuals’ immediate household. Similarly, bans on public gatherings seek to limit the frequency and number of people with whom they come into close contact. However, such actions do not eliminate all potential contact because individuals under such orders still need to break from shelter to fetch provisions and, for members of essential professions, to go to work. It is, therefore, imperative that individuals comply with public health guidelines advocating the practice of social distancing when they come into contact with others. Compliance with guidelines is also highly important in regions that have not issued formal “shelter-in-place” orders but have instead provided “safer-at-home” guidelines and in areas that have begun to lift “shelter-in-place” orders.
Public health organizations have been tasked with developing behavioral interventions that are efficacious in promoting social distancing behaviors among the general population [6]. Given that social distancing is a relatively novel behavior in many countries, identification of the determinants of social distancing behavior has become critical. Moreover, identifying determinants that are potentially modifiable through intervention, that is, can be targeted in messages or campaigns of behavioral interventions aimed at promoting social distance, is a recognized priority [7]. There have, therefore, been calls for research informed by behavioral science that identifies key determinants of preventive behaviors in the context of the current pandemic, particularly social distancing [7, 8]. However, there is relatively little research on the determinants of social distancing, particularly in the context of communicable disease prevention (e.g., influenza) in a global pandemic [9]. Previous research, for example, has tended to focus on the social cognition determinants of other preventive behaviors, such as facemask wearing [10], or focused on hypothetical scenarios [11] in the context of influenza prevention. To date, there are few studies informed by behavioral science on the individual determinants of social distancing in the context of the COVID-19 pandemic.
To fill this evidence gap, the current study aimed to identify the determinants of social distancing behavior among individuals subject to social distancing regulations during the COVID-19 pandemic. The research adopted an integrated theoretical approach based on social cognition theories to identify constructs that predict social distancing behavior and the processes involved. The research is expected to provide evidence of potentially modifiable targets for behavior change interventions aimed at promoting social distancing. Such interventions may contribute to reduced infection rates during the current pandemic and may assist in preventing a “second wave” of infections as “shelter-in-place” orders are lifted [5].
Social Distancing Determinants: An Integrated Social Cognition Approach
Research examining health behavior determinants has a long tradition of applying social cognition theories [12], which assume health behavior enactment is a reasoned process determined by beliefs, such as risk perception, attitude, social norm, and perceptions of control or self-efficacy. A prototypical social cognition approach is offered by the theory of planned behavior [13]. In the theory, individuals’ intention to perform the target behavior is proposed as the most proximal determinant of the performance of a future target behavior. Intention is a function of three constructs that summarize sets of beliefs regarding the future behavior: attitude (beliefs that the behavior will have advantageous or disadvantageous consequences), subjective norm (beliefs that significant others express support for performing the behavior), and perceived behavioral control (PBC; beliefs in the capacity to perform the behavior and to overcome barriers to the behavior). Intention is proposed to mediate the effects of attitude, subjective norm, and PBC on behavior. PBC is also proposed to predict behavior directly when it approximates actual control. Theory predictions have been supported in correlational and prospective research across multiple behaviors, contexts, and populations [14].
While the elegant parsimony of the theory of planned behavior is appealing, it is not without limitations. Research applying the theory has indicated that substantive variance in health behavior remains unexplained [14]. In addition, the size of the effect of intention on health behavior is often modest, suggesting a “shortfall” in those who report an intention to perform the behavior and those who act on their intention [15]. Researchers have, therefore, proposed modifications to the theory to resolve these limitations, such as integrating additional constructs from other theories, in the theory to predict behavior more effectively and address the intention–behavior “gap” [16].
Introducing additional constructs to the theory is one approach to increasing explained variance in health behavior. For example, researchers have examined relations between moral norms, an additional form of normative influence, and health behavior. Moral norms are considered particularly relevant when there is a moral imperative for acting (e.g., vaccination and blood donation) [17]. In the context of COVID-19, messaging from public health authorities on COVID-19-preventive behaviors has focused on protecting the vulnerable (e.g., immunosuppressed individuals, those with underlying health conditions, and the elderly) [3]. On this basis, we reasoned that moral norm would constitute a highly relevant determinant of social distancing intention and behavior in the context of the pandemic. In addition, anticipated regret has been shown to predict behaviors perceived likely to have adverse consequences or result in significant losses if not performed [17]. Failure to perform social distancing behaviors may be perceived as having highly undesirable consequences, such as becoming infected or infecting vulnerable others. We, therefore, included moral norm and anticipated regret as additional predictors of intention to perform social distancing behavior in our integrated model.
Researchers have applied “dual-phase” models as a means to resolve the limitation of the intention–behavior “gap.” Models like the model of action phases [18] and the health action process approach (HAPA) [19] propose that individuals need to augment their intentions with action plans in order to enact them. Action plans reflect the extent to which individuals have specified when, where, and how they will perform the intended behavior. The model of action phases [18] suggests that individuals will more likely enact their intentions if they form an action plan, so action plans are proposed to moderate the intention–behavior relationship. By contrast, the HAPA suggests that planning is part of the process of intention enactment such that action plans mediate the intention–behavior relationship [19]. Meta-analyses of studies in health behavior have supported both processes [20, 21], and we aimed to test both in our proposed integrated model of social distancing behavior.
While social cognition theories like the theory of planned behavior assume participation in health behavior to be a reasoned process, research applying such theories has shown that past behavior remains a pervasive determinant of behavior alongside the theory constructs [22, 23]. The inclusion of past behavior as an independent behavioral predictor in a social cognition theory is important because it provides a test of its sufficiency in accounting for unique variance in behavior. However, residual effects of past behavior on behavior are also assumed to model the effects of other unmeasured constructs on behavior [23]. One candidate construct is habit, which reflects the “nonconscious” or “automatic” enactment of a behavior developed through its repeated performance in stable contexts [24, 25]. Research examining the effects of habit in the context of social cognition theories has examined how self-reports of experiencing the behavior as “automatic” and “unthinking” predict health behavior independent of intentions [26]. The introduction of habit in our augmented model, therefore, may provide important information on the extent to which social distancing behavior is determined by reasoned or nonconscious processes [27].
The Present Study
The present study aimed to identify the determinants of participation in social distancing behavior among individuals in the context of COVID-19 using an integrated social cognition model that incorporated constructs from the theory of planned behavior with moral norm, anticipated regret, action planning, and self-reported habit. We tested predictions of the proposed model in a prospective correlational study in two separate samples of adults from Australia and the USA, respectively. These countries provide an opportunity to examine the determinants of social distancing because they experienced rapid increases in COVID-19 cases relatively early in the pandemic and introduced public health advice and “lockdown” measures to minimize transmission via social distancing. In our proposed model (Model 1; Fig. 1), attitude, subjective norm, PBC, moral norm, and anticipated regret were specified as predictors of intention, and intention, PBC, and habit as predictors of social distancing behavior. Intention was proposed to mediate the effects of the social cognition constructs on behavior. The role of action planning as a mediator and moderator of the intention–behavior relationship was also specified. We also specified a second model (Model 2; Fig. 1) in which past social distancing behavior was included as a direct predictor of all constructs in the model, providing a test of its sufficiency. Although research demonstrating that social distancing behavior clusters with other health behaviors indicates that application of social cognition theories is viable for this behavior [28], research is needed to verify this contention and the current study contributes to this goal. The research may assist in identifying potentially modifiable constructs that relate to social distancing behavior. Such information may provide useful information to inform social distancing interventions focused on reducing the spread of COVID-19 and, more broadly, other communicable diseases.
Fig. 1. Standardized parameter estimates of the integrated model. Upper panel presents the model excluding past behavior (Model 1) and the lower panel presents the model including past behavior (Model 2). Coefficients printed on the upper line are for the Australian sample and coefficients printed on the lower line are for the U.S. sample. †Effect is significantly different across the Australian and U.S. samples in multiple-group analyses. *p < .05, **p < .01, ***p < .001.
Method
Participants and Recruitment
Samples of Australian (N = 495, 50.1% female) and U.S. (N = 701, 48.9% female) residents were recruited via an online research panel company. To be eligible for inclusion, participants needed to be aged 18 years or older and not subject to formal quarantine for COVID-19. Participants were also screened for age, gender, and geographical region and quotas imposed during recruitment to ensure that the final samples closely matched the national distributions for these characteristics in each country. Data were collected between April 1 and May 6, 2020. All participants in the Australian sample were subject to a national “shelter-in-place” order issued by the federal government. However, issuance of orders in the USA was devolved to state governments resulting in some variations. The vast majority of participants in the U.S. sample (n = 610, 87.0%) were subject to “shelter-in-place” orders for the duration of the study. However, some states did not impose “shelter-in-place” orders at all (Arkansas, Iowa, Nebraska, North Dakota, South Dakota, Utah, and Wyoming), so a minority of participants in the U.S. sample (n = 37, 5.3%) were never subject to an order. Furthermore, in some cases in the U.S. sample (n = 47, 6.7%), “shelter-in-place” orders had been lifted prior to follow-up data collection. However, among the states in the USA that did not have “shelter-in-place” orders, or lifted their orders during the study, all issued social distancing guidelines and encouraged the population to follow those guidelines. Baseline sample characteristics are presented in Table 1.
Table 1. Sample characteristics and descriptive statistics for study variables at baseline and at 1 week follow-up
Variable Australia sample U.S. sample
Baseline Follow-up Baseline Follow-up
Participants 495 365 701 440
Age, M years (SD) 47.09 (17.11) 49.78 (16.89) 45.55 (17.40) 51.77 (16.26)
Gender, n (%)a
Female 252 (51.1) 182 (50.1) 352 (48.9) 205 (46.6)
Male 241 (48.9) 181 (49.9) 341 (50.5) 231 (52.5)
Not specified/prefer not to answer 0 (0.0) 0 (0.0) 4 (0.6) 4 (0.9)
Employment status, n (%)b
Currently unemployed/full-time caregiver 231 (46.7) 180 (49.3) 330 (47.3) 216 (49.5)
Part-time/casual employed 97 (19.6) 65 (17.8) 106 (15.2) 60 (13.8)
Currently employed full time 140 (28.3) 104 (28.5) 233 (33.4) 147 (33.7)
Leave without pay/furloughed 27 (5.5) 16 (4.4) 28 (4.0) 13 (3.0)
Marital status, n (%)c
Married 184 (37.2) 146 (40.0) 300 (43.0) 224 (51.4)
Widowed 8 (1.6) 7 (1.9) 22 (3.2) 18 (4.1)
Separated/divorced 53 (10.7) 39 (10.7) 69 (9.9) 47 (10.8)
Never married 160 (32.3) 103 (28.2) 255 (36.6) 126(28.9)
Married de facto 90 (18.2) 70 (19.2) 51 (7.3) 21 (4.8)
Ethnicity, n (%)d
Black 3 (0.6) 1 (0.3) 52 (7.5) 26 (6.0)
Caucasian/White 392 (79.2) 304 (83.3) 566 (81.2) 376 (86.2)
Asian (South-East Asia/South Asia) 71 (14.3) 43 (11.8) 39 (5.6) 24 (5.5)
Middle-Eastern 6 (1.2) 3 (0.8) 1 (0.1) 0 (0.0)
Other 13 (2.6) 6 (1.6) 27 (3.9) 8 (1.8)
Prefer not to answer 10 (2.0) 8 (2.2) 12 (1.7) 2 (0.5)
Income, n (%)e
Zero income 8 (1.7) 4 (1.2) 31 (4.4) 19 (4.4)
$1–$199 ($1–$10,399) 9 (2.0) 6 (1.8) 40 (5.7) 24 (5.5)
$200–$299 ($10,400–$15,599) 12 (2.6) 8 (2.4) 34 (4.9) 23 (5.3)
$300–$399 ($15,600–$20,799) 19 (4.1) 12 (3.6) 38 (5.5) 23 (5.3)
$400–$599 ($20,800–$31,199) 42 (9.2) 33 (9.9) 62 (8.9) 33 (7.6)
$600–$799 ($31,200–$41,599) 57 (12.4) 42 (12.6) 61 (8.8) 39 (8.9)
$800–$999 ($41,600–$51,999) 45 (9.8) 31 (9.3) 68 (9.8) 46 (10.6)
$1,000–$1,249 ($52,000–$64,999) 39 (8.5) 32 (9.6) 48 (6.9) 38 (8.7)
$1,250–$1,499 ($65,000–$77,999) 28 (6.1) 22 (6.6) 59 (8.5) 41 (9.4)
$1,500–$1,999 ($78,000–$103,999) 72 (15.7) 50 (15.0) 72 (10.3) 48 (11.0)
$2,000 or more ($104,000 or more) 81 (17.6) 62 (18.6) 108 (15.5) 74 (17.0)
Prefer not to answer 47 (10.2) 32 (9.6) 76 (10.9) 28 (6.4)
Education level, n (%)
Completed junior/lower/primary school 18 (3.6) 17 (4.7) 6 (0.9) 0 (0.0)
Completed senior/high/secondary school 133 (26.9) 98 (26.8) 265 (37.8) 132 (30.0)
Postschool vocational qualification/diploma 147 (29.7) 111 (30.4) 138 (19.7) 94 (21.4)
Undergraduate university degree 131 (26.5) 93 (25.5) 214 (30.5) 159 (36.1)
Postgraduate university degree 66 (13.3) 46 (12.6) 78 (11.1) 55 (12.5)
SD standard deviation.
aTwo participants in the Australian sample did not report their gender and four participants in the U.S. sample did not report their gender.
bFour participants in the U.S. sample did not report their employment status.
cFour participants in the U.S. sample did not report their marital status.
dFour participants in the U.S. sample did not report their ethnicity.
eThirty-one participants in the Australian sample did not report their income and four participants in the U.S. sample did not report their income.
Design and Procedure
The study adopted a prospective correlational design with self-report measures of social cognition constructs from the proposed integrated model, intention, and past social distancing behavior administered at an initial data collection occasion in a survey administered using the Qualtrics online survey tool. Social cognition measures included the theory of planned behavior (attitude, subjective norm, and PBC), moral norm, anticipated regret, action planning, and habit constructs. Participants were informed that they were participating in a survey on their social distancing behavior and provided with information outlining study requirements. They were required to provide informed consent before proceeding with the survey. Participants were also provided with instructions on how to complete study measures and a definition of the target behavior: “The following survey will ask about your beliefs and attitudes about ‘social distancing’. What do we mean by social distancing? Social distancing (also known as ‘physical distancing’) is deliberately increasing the physical space between people to avoid spreading illness. The World Health Organization and other world leading health authorities suggest that you should maintain at least a 1–2 meter (3–6 feet) distance from other people to lessen the chances of getting infected with COVID-19. When answering the questions in this survey, think about your social distancing behavior (i.e., maintaining at least a 1–2 m (3–6 ft) distance from other people).” One week later, participants were re-contacted by the panel company and asked to self-report their habit and social distancing behavior over the previous week using the same measures used at the initial data collection occasion. Participants received a fixed sum of money for their participation based on expected completion time consistent with the panel company’s published rates. Approval for study procedures was granted prior to data collection from the Griffith University Human Research Ethics Committee.
Measures
Study measures were multi-item self-report measures of constructs based on published guidelines and measures used in previous studies [13, 29, 30]. Participants provided their responses on scales with seven-point response options. Complete study measures are provided in Supplementary Appendix A.
Social cognition constructs
Multi-item measures of attitude, subjective norm, PBC, moral norm, anticipated regret, and action planning were developed according to published guidelines [13, 29]. Each measure made explicit reference to the target behavior of social distancing, and participants were reminded of the definition of social distancing before completing the measures.
Intention
Participants’ intention to participate in social distancing behavior over the next week was measured using a scale developed according to published guidelines [31].
Habit
Habit was measured at both time points using the behavioral automaticity items of Verplanken and Orbell’s self-report habit index [25]. The measure measures individuals’ reflections on the extent to which the behavior is experienced as automatic and enacted without thought.
Past behavior and behavior
Participants self-reported their participation in social distancing behavior to minimize the transmission of the SARS-CoV-2 virus that causes COVID-19. The measure comprised two items prompting participants to report their frequency of social distancing behavior in the previous week. This is based on previously used self-report behavioral measures that have demonstrated concurrent validity with non-self-report measures in other behavioral contexts [32].
Demographic variables
Participants self-reported their age in years, gender, employment status (currently unemployed/full-time caregiver, currently full-time employed, part-time employed, or on leave without pay/furloughed), marital status (married, widowed, separated/divorced, never married, or in a de factor relationship), annual household income stratified by 11 income levels based on Australian and U.S. national averages, highest level of formal education (completed junior/lower/primary school, completed senior/high/secondary school, postschool vocational qualification/diploma, further education diploma, undergraduate university degree, or postgraduate university degree), and ethnicity (Black, Caucasian/White, Asian, or Middle-eastern). Binary income (low income vs. middle/high income), highest education level (completed school education only vs. completed postschool education), and ethnicity (Caucasian/White vs. non-White) variables were computed for use in subsequent analyses.
Data Analysis
Hypothesized relations among the integrated model constructs were tested in the Australian and U.S. samples separately using variance-based structural equation modeling implemented in the WARP 7.0 analysis package [33]. Model parameters and standard errors (SEs) were computed using the “Stable3” estimation method, which has been shown to provide the most precise parameter estimates in complex structural models in smaller samples and outperforms bootstrapping methods in simulation studies [33]. Simulation studies have also shown this method to provide more consistent and precise estimates in data containing outliers, which may inflate SEs and lead to abnormally high p-values [33]. Two models were estimated in each sample: a model testing predictions of the proposed integrated model with the binary demographic variables also included as covariates (Model 1; Fig. 1, upper panel) and a model that included effects of past social distancing behavior (Model 2; Fig. 1, lower panel). All constructs were latent variables indicated by single or multiple items. There were no missing data for the social cognition and self-reported behavioral variables. There were a few instances of missing data for the demographic variables ranging from 0.5% to 8.8% in the Australia sample, and 0.9% to 6.4% in the U.S. sample. Missing data are reported in Supplementary Appendix B. Missing data were imputed using stochastic hierarchical regression [33].
The analysis afforded a number of analyses to evaluate the adequacy of measures used to indicate the latent variables in the model. Construct validity of the latent factors for the social cognition, intention, and behavioral variables was established using the normalized factor pattern loadings after oblique rotation and Kaiser normalization [33] and the average variance extracted (AVE), which should approach or exceed .700 and .500, respectively. Internal consistency of the factors was estimated using omega reliability coefficients (ω) and composite reliability coefficients (ρ), which should exceed .700 and ideally approach .900. We also conducted tests of the discriminant validity of the constructs in the model. Discriminant validity was supported when the square root of the AVE for each latent variable exceeded its correlation with other latent variables.
Adequacy of the proposed model in describing the data was established using the goodness-of-fit (GoF) index with values of .100, .250, and .360 corresponding to small, medium, and large effect sizes. Further information on model quality was provided by the average path coefficient and average R2 coefficient. These indices summarize the average parameter estimates of relations in the model and the amount of variance explained in each dependent variable, respectively, and should be statistically significant for a good-quality model. In addition, an overall GoF index is provided by the average block variance inflation factor for model parameters and the average full collinearity variance inflation factor, which should be equal to or lower than 3.3 for well-fitting models. These indices indicate the extent to which latent variables in the model overlap and contribute to model multicollinearity. They, therefore, provide an indication as to the uniqueness of the existing latent variables in the model. Four further indices were also used to evaluate model quality: the Simpson’s paradox ratio (SPR), R2 contribution ratio (R2CR), the statistical suppression ratio (SSR), and the nonlinear bivariate causality direction ratio (NLBCDR). The SPR indicates whether the model is free from incidences of Simpson’s paradox (i.e., when the path coefficient and the correlation associated with a latent variable have opposite signs), indicating a causality problem. The SPR should exceed .700 and ideally approach 1.000. The R2CR and SSR provide indication of the extent to which models are free from instances of negative R2 contributions and statistical suppression. The R2CR and SSR should exceed .900 and .700, respectively. The NLBCDR provides an estimate of the extent to which the proposed “causal” associations in the proposed model are more tenable than those in the opposite direction and provide an initial indicator of support for the hypothesized directions of the causal links in the proposed model compared to if the proposed direction were reversed. The NLBCDR should exceed .700 for high-quality models. Kock [33] provides further technical details on model fit and quality indices.
Model effects were estimated using standardized path coefficients with confidence intervals (CIs) and test statistics. Effect sizes were estimated using a variant of Cohen’s f-square coefficient and represent the individual contribution of the predictor variable to the R2 coefficients of the criterion latent variable. Values of .02, .15, and .35 represent small, medium, and large effect sizes, respectively. Differences in the path coefficients in the models across the samples were tested using multiple-group analysis using the Satterthwaite method with two-tailed significance tests.
We also tested whether the inclusion of participants that were never under a “shelter-in-place” order, or had the “shelter-in-place” order lifted during the study, affected predicted relations in the models. The small numbers of participants that were, at some point, not subjected to “shelter-in-place” orders meant we could not conduct a formal moderator analysis, so we conducted a sensitivity analysis testing whether effects in the models differed if data from these participants were excluded. Models excluding and including past behavior were estimated in samples excluding participants who were never subject to a “shelter-in place” order, and in the sample that were never subject to an order, or who had the order lifted at some stage during the study. Formal comparisons of parameter estimates in these models with those from the full sample were made using the Satterthwaite method. Data files, analysis scripts, and output files for all analyses are available online: https://osf.io/x9tms/.
Results
Participants
Attrition across the two data collection occasions resulted in final sample sizes of 365 (M age = 49.78, standard deviation [SD] = 16.89; 50.1% female; retention rate 73.73%) and 440 (M age = 51.77, SD = 16.26; 46.6% female; retention rate = 62.77%) participants in the Australian and U.S. samples, respectively. Sample characteristics at follow-up are presented in Table 1. Attrition analyses in the Australian sample revealed that participants lost to attrition were younger and were more likely to be non-White. However, there were no differences in proportion of gender, income, and education level. A MANOVA with the social cognition constructs and past behavior as dependent variables and attrition status (lost to attrition vs. included at follow-up) revealed no differences (Wilks’ Lambda = 0.969, F(1,9) = 1.70, p = .077, partial η 2 = .031). Attrition analyses in the U.S. sample also indicated that participants lost to attrition were younger, and more likely to be male, non-White, and lower educated and have low income, than those remaining in the study at follow-up. The MANOVA testing for differences on social cognition constructs and past behavior among participants lost to attrition and those included at follow-up revealed statistically significant differences (Wilks’ Lambda = 0.969, F(1,9) = 2.40, p = .010, partial η 2 = .031). Follow-up tests revealed that mean values for past behavior, attitude, subjective norm, intention, moral norm, and habit with respect to social distancing were significantly lower among participants lost to attrition compared to those retained at follow-up. However, effect sizes for these differences were small (ds < .25). Details of attrition analyses are presented in Supplementary Appendix B.
Preliminary Analyses
Factor loadings and AVE values exceeded recommended .700 and .500 cutoff values in all cases. Omega reliability coefficients, interitem correlations (for two-item scales), and composite reliabilities indicated good internal consistency of scales used. Latent variable correlations among social cognition constructs were all statistically significant. Correlations among the majority of constructs in the Australia sample were small-to-medium in size (r range = .161 to .564), with some smaller correlations involving the subjective and moral norms constructs and habit (r range = .094 to .118). Correlations were small-to-medium in size in the U.S. sample (r range = .266 to .620). Square roots of the AVE for each latent variable exceeded the correlation of that variable with all other latent variables supporting discriminant validity. Skewness and kurtosis estimates indicated many of the variables were not normally distributed, justifying the use of the variance-based structural equation modeling, which is a “distribution-free” analytic method. Factor loadings, reliability coefficients, and distribution statistics are presented in Supplementary Appendix C, and latent variable correlations for model variables in both are presented in Supplementary Appendix D.
Structural Equation Models
Single-sample analyses
GoF and quality indices of the structural equation models are presented in Table 2. The models that excluded (Model 1) and included (Model 2) past behavior exhibited adequate fit and quality indices in both the Australian and U.S. samples. Standardized parameter estimates for the proposed direct effects for each model in the Australian and U.S. samples are presented in Fig. 1. Full parameter estimates for models in both samples are presented in Supplementary Appendix E. Parameter estimates, CIs, and effect sizes for the indirect effects of the models in both samples are summarized in Table 3.
Table 2. Model quality and GoF statistics for the structural equation models of the integrated model in the Australian and U.S. samples and multigroup model
Sample Model APC AR2 AVIF AFVIF GoF SPR R2CR SSR NLBCDR
Australia 1 .104** .177*** 1.177 1.561 .391 .841 .977 .889 .873
2 .116*** .338*** 1.222 1.904 .543 .819 .991 .931 .785
USA 1 .098** .192*** 1.187 1.823 .410 .889 .995 .825 .754
2 .116*** .338** 1.222 1.904 .543 .819 .991 .931 .785
MS 1 .100*** .182*** 1.159 1.704 .398 .905 .995 .794 .817
2 .113*** .300*** 1.186 1.760 .511 .931 .997 .917 .840
Model 1 = model excluding past behavior; Model 2 = model including past behavior.
AFVIF average full collinearity variance inflation factor; APC average path coefficient; AR2 average R2; AVIF average block variance inflation factor; GoF Tenenhaus’s goodness-of-fit index; MS multiple sample analysis; NLBCDR nonlinear bivariate causality direction ratio; R2CR R2 contribution ratio; SPR Sympson’s paradox ratio; SSR statistical suppression ratio.
*p < .05, **p < .01, ***p < .001
Table 3. Standardized parameter estimates for indirect effects for the structural equation model of the integrated model in the Australian and U.S. samples
Effect Model excluding past behavior Model including past behavior
β p 95% CI ES Β p 95% CI ES
LB UB LB UB
Australian sample
Indirect effects
Att.→Int.→Beh. .011 .359 −.052 .074 .003 .004 .444 −.059 .067 .001
SN→Int.→Beh. .042 .094 −.021 .105 .016 .016 .312 −.047 .079 .006
MN→Int.→Beh. .068 .016 .005 .131 .024 .028 .192 −.035 .091 .010
AR→Int.→Beh. .011 .356 −.052 .074 .003 .003 .457 −.060 .066 .001
PBC→Int.→Beh. .040 .101 −.023 .103 .011 .016 .307 −.047 .079 .005
Int.→AP→Beh. .040 .106 −.023 .103 .014 .011 .365 −.052 .074 .004
Hab. (T1).→Hab. (T2)→Beh. .102 <.001 .041 .163 .016 .078 .007 .017 .139 .013
PB→Hab.→Beh. – – – – – .021 .214 −.030 .072 .011
PB→Beh.a – – – – – .081 .034 −.007 .169 .042
Total effectsb
Int.→Beh. .220 <.001 .134 .306 .081 .090 .022 .004 .176 .033
PBC→Beh. .126 <.001 .040 .212 .036 .055 .110 −.033 .143 .016
Hab. (T1)→Beh. .096 .016 .010 .182 .015 .076 .044 −.012 .164 .012
PB→Beh. – – – – – .494 <.001 .412 .576 .258
U.S. sample
Indirect effects
Att.→Int.→Beh. <.001 .495 −.052 .054 <.001 .004 .443 −.049 .057 .001
SN→Int.→Beh. .072 .003 .019 .125 .029 .023 .190 −.030 .076 .009
MN→Int.→Beh. .102 <.001 .051 .153 .044 .040 .067 −.013 .093 .017
AR→Int.→Beh. .023 .192 −.030 .076 .011 .001 .478 −.052 .054 .001
PBC→Int.→Beh. .088 <.001 .037 .139 .025 .038 .079 −.015 .091 .011
Int.→AP→Beh. .061 .011 .008 .114 .029 .004 .441 −.049 .057 .002
Hab. (T1).→Hab. (T2)→Beh. .212 <.001 .161 .263 .075 .166 <.001 .115 .217 .059
PB→Hab.→Beh. – – – – – .068 <.001 .025 .111 .043
PB→Beh.a – – – – – .178 <.001 .105 .251 .112
Total effectsb
Int.→Beh. .377 <.001 .306 .448 .177 .142 <.001 .069 .215 .066
PBC→Beh. .146 <.001 .073 .219 .042 .074 .024 .001 .147 .021
Hab. (T1)→Beh. .242 <.001 .169 .315 .086 .171 <.001 .098 .244 .061
PB→Beh. – – −.052 .074 – .673 <.001 .604 .742 .423
aSum of indirect effects of past behavior on behavior through all model constructs.
bTotal effect comprising sums of all indirect effects through model constructs plus the direct effect.
β standardized parameter estimate; 95% CI 95% confidence interval of standardized parameter estimate; AP action planning; AR anticipated regret; Att. attitude; Beh. behavior; ES effect size of the standardized parameter estimate; Hab. (T1) self-reported habit measured at baseline (T1); Hab. (T2) self-reported habit measured at follow-up (T2); Int. intention; LB lower bound of 95% CI; MN moral norm; PB past behavior; PBC perceived behavioral control; SN subjective norm; UB upper bound of 95% CI. Focusing on the model excluding past behavior (Model 1), intention, action planning, and habit at follow-up were statistically significant direct predictors of social distancing behavior, with effect size for intention and habit generally larger in the U.S. sample. PBC directly predicted behavior in the Australian sample only, also with a small effect size. Intention predicted action planning in both samples with large effect sizes. Subjective norm, moral norm, and PBC predicted intention in both samples, with small-to-medium effect sizes, but effects of attitude were not significant. There was a small effect of anticipated regret on intention in the U.S. sample only. Habit at baseline predicted habit at follow-up in both samples, with large effect sizes. There was also a small-sized effect of habit at baseline on intention in the U.S. sample only. Overall, the model accounted for significant variance in social distancing behavior (Australian sample, R2 = .198; U.S. sample, R2 = .361), intentions (Australian sample, R2 = .571; U.S. sample, R2 = .623), and habit at follow-up (Australian sample, R2 = .416; U.S. sample, R2 = .486). Intentions (Australian sample, R2 = .066; U.S. sample, R2 = .148), action planning (Australian sample, R2 = .029; U.S. sample, R2 = .044), and habit at follow-up (Australian sample R2 = .041; U.S. sample, R2 = .129) each accounted for substantive variance in behavior. Action planning significantly moderated the intention–behavior relationship in the Australian sample only. While the effect was not in the predicted direction, probing the interaction revealed that the intention–behavior relationship increased as the level of planning increased, consistent with theory. However, the intention–behavior relationship is more likely to be smaller at lower levels of planning, and it seems that planning makes less of a difference when the intention–behavior relationship is large. A plot of the interaction effect is presented in Supplementary Appendix F.
Turning to the indirect effects, there were significant indirect effects of subjective norm, moral norm, and PBC on social distancing behavior mediated by intention in the U.S. sample. By contrast, only the indirect effect of moral norm on behavior through intention was significant in the Australian sample. The smaller indirect effects in the Australian sample is principally due to the significantly smaller effect size for the intention–behavior relationship in this sample compared to the U.S. sample. Habit at baseline predicted behavior through habit at follow-up in both samples. Effect sizes in all cases were small. There were significant total effects of intention, PBC, and habit at baseline on behavior, with effect sizes larger in the U.S. sample than in the Australian sample.
For the model including past behavior, significant effects of past behavior on all model constructs were observed in both samples with effect sizes ranging from small to large. The effects of past behavior on social distancing behavior were particularly large. Inclusion of past behavior led to an attenuation of model effects in both samples. Specifically, the effects of intention and habit at follow-up on behavior were reduced but remained statistically significant in both samples with small effect sizes. In addition, effects of subjective norm, moral norm, and PBC on intention, and the effect of intention on action planning, remained statistically significant in both samples, with small-to-medium effect sizes. The effect of habit at baseline on habit at follow-up was statistically significant in both samples, with large effect sizes. Variance explained in social distancing behavior increased substantially with the inclusion of past behavior, with only modest changes in explained variance in intentions (Australian sample R2 = .598; U.S. sample, R2 = .702) and habit at follow-up (Australian sample R2 = .416; U.S. sample, R2 = .486). Specifically, intentions (Australian sample, R2 = .029; U.S. sample, R2 = .065), past behavior (Australian sample, R2 = .216; U.S. sample, R2 = .311), and habit at follow-up (Australian sample R2 = .031; U.S. sample, R2 = .101) each accounted for substantive variance in behavior.
Turning to indirect effects, we found significant indirect effects of habit at baseline on behavior mediated by habit at follow-up in both samples with small effect sizes. There were also significant total effects of intention and habit at baseline on behavior in both samples, and of PBC on behavior for the U.S. sample, with small effect sizes. There were significant total indirect and total effects of past behavior on behavior in both samples, with large effect sizes. There was a small-sized indirect effect of past behavior on behavior mediated by habit at both time points in the U.S. sample, but the effect was not significant in the Australian sample.
Multisample analyses
Multisample analyses permitted for tests of difference in parameter estimates for each model across the Australian and U.S. samples. For the model excluding past behavior (Model 1), only effects of intention on habit at baseline, habit at follow-up on social distancing behavior, and intention on action planning differed across samples. These effects were significantly larger in the U.S. sample. Some effects with observed differences across samples, such as the effect of habit at baseline on intention or the moderator effect of planning on the intention–behavior relationship, did not differ significantly across samples. For the model including effects of past behavior (Model 2), multisample analysis revealed no differences in effect size across samples, indicating that the attenuating effect of past behavior on model effects also had the effect of eliminating the few differences in model effects across samples. Full details of the multiple-group analysis are presented in Supplementary Appendix G.
Sensitivity analyses
We re-estimated both models in samples excluding participants who were never subject to a “shelter-in place” order, and in the sample that were never subject to an order, or who had the order lifted at some stage during the study. Comparisons of parameter estimates in these models with those from the models estimated in the full sample revealed no significant differences in any of the model parameters. Results are reported in Supplementary Appendices H and I.
Discussion
The present study aimed to identify the determinants of social distancing behavior in the context of COVID-19 through the application of an integrated social cognition model. The integrated model was based on the theory of planned behavior [13] augmented to include additional predictors relating to normative (moral norm), anticipated affect (anticipated regret), volitional (action planning), and nonconscious (habit) determinants of health behavior. The model was tested in data from a correlational prospective survey study in two samples of Australian and U.S. residents subject to national or local “shelter-in-place” orders. Results indicated that intention and habit were significant predictors of social distancing behavior in both samples. Subjective norm, moral norm, and PBC were significant predictors of social distancing intention. In addition, intention-mediated effects of these social cognition constructs on social distancing behavior in the U.S. sample, but did so only for moral norm in the Australian sample. Action planning did not mediate effects of intention on behavior in either samples but moderated the intention–behavior relationship in the Australia sample. Inclusion of past behavior attenuated effects of social cognition constructs in the models in both samples, although habit and intention remained significant determinants of social distancing behavior in both samples. Excluding participants in the U.S. sample not subject to formal “shelter-in-place” orders, or had the orders lifted during the study, did not affect the pattern or size of the effects in the model, providing evidence that formal orders did not have a substantive bearing on the determinants of social distancing behavior in this sample.
Current findings provide qualified support for some, but not all, predictions of the integrated social cognition model for social distancing behavior. A key assumption of the model, derived from the social cognition theories on which it is based, is that social distancing behavior is reasoned action and, therefore, determined predominantly by intention and the belief-based constructs that underpin them. Effects of intention on social distancing behavior and its mediation of constructs reflecting social reasons for acting, particularly beliefs relating to significant others and moral obligations to perform the behavior, and PBC is consistent with this assumption. This is unsurprising in this context, considering the widely publicized details of the relatively mild effects of the virus in the majority of the population. It is likely that the majority of individuals do not view themselves as at serious risk from COVID-19 but have internalized the view that significant others want them to engage in social distancing and feel a moral obligation to perform the behavior to protect the health of those most at risk. Such a finding is consistent with research on similar health behaviors, such as blood donation, where behavioral performance is likely to promote the health of others rather than the self [34]. Similarly, the impact of PBC indicates the importance of perceived personal agency in maintaining social distancing behavior, consistent with previous research on health behaviors [14]. Individuals that see fewer barriers to maintaining social distancing and have the confidence to do so are more likely to intend to perform these behaviors.
The effects of subjective and moral norms and PBC suggests that these should be viable targets for behavioral interventions aimed at promoting social distancing behavior based on the model. For example, messages promoting moral obligation (e.g., highlighting social responsibility for preventing transmission of the virus to vulnerable others through social distancing) and perceived control (e.g., demonstrating how to easily and successfully maintain appropriate social distance) may facilitate greater intention to socially distance. However, the intention–behavior relationship in the present study was relatively modest in size, particularly in the Australian sample, indicative of a substantive intention–behavior “gap” [15]. This suggests that interventions targeting change in intention determinants, such as moral norms and PBC, may have only small effects on social distancing behavior. It may be of value to explore how properties of intention may affect intention–behavior relations in the context of social distancing behavior [35]. Such properties may signal potential intervention strategies that may strengthen intention–behavior relations in conjunction with messaging targeting moral norms and PBC.
Current findings also indicated consistent effects of self-reported habits on social distancing behavior. Importantly, the effects of habit were direct and independent of intentions, consistent with the theory that suggests that effects of habits reflect nonconscious, automatic processes developed through consistent experience with the behavior in stable contexts over time. Habits also partially mediated the effects of past behavior on social distancing behavior, suggesting that past behavior effects, at least in part, reflect habits [27]. An implication of these findings is that facilitating habit development in behavioral interventions may be effective in promoting social distancing. Research suggests that strategies, such as providing successful experiences of the desired behavior consistently over time and creating environment conditions that facilitate the behavior (e.g., consistent reminders and environmental restructuring) are effective in inducing habits [36], but the efficacy of such strategies in the context of social distancing behavior need to be verified empirically. Furthermore, legislation restricting or mandating behavior change facilitates habit formation over time. This suggests that the introduction of “shelter-in-place” and other government-mandated restrictions may facilitate social distancing habits.
Inclusion of past behavior as a predictor of social distancing behavior at follow-up reduced the effects of intention on behavior to a trivial size in both samples and also attenuated the effects of the social cognition constructs on intention. Such effects are consistent with previous research [22] and raise questions over the sufficiency of the model in identifying the determinants of social distancing behavior. However, such findings must be interpreted in light of the current study design and how the effects of past behavior can provide important information on the determinants of social distancing behavior. The 1 week time lag means that past behavior was always likely to have a large effect because individuals’ behavior tends to be relatively stable over short periods [22]. A more complete evaluation of model sufficiency would be afforded by testing its long-range prediction, which has often been cited as a goal of social cognition theories [14], and should be considered a future research priority for research on social distancing behavior. However, past behavior effects can be informative on the determinants of social distancing behavior as it may reflect the effects of other unmeasured behavioral determinants. In particular, past behavior will likely reflect determinants that bypass the reasoned, intention-mediated processes that lead to behaviors, such as implicit attitudes and motives, personality traits, and variables reflecting the social and physical environment. The effects of such constructs are speculative and future tests of the integrated model that incorporate such factors alongside those from the current model may assist in resolving these effects.
Consistent with dual-phase models [18, 19], we also tested the extent to which action planning was implicated in the process by which individuals act on their intention. Two patterns of effects were tested: mediation and moderation effects of action planning on the intention–behavior relationship. The mediation effect was significant in the U.S. sample but not the Australia sample, while the moderation effect was significant in the Australia sample only. However, in both cases, the effects were small in size. The small size of the mediation effects suggests that action planning is a relatively trivial component of the link between social distancing intention and behavior, particularly when past behavior was taken into account. The moderation of the intention–behavior relationship by action planning in the Australian sample was negative in sign, which is contrary to predictions [18]. However, probing this interaction indicated that individuals with stronger intention were more likely to follow through on their social distancing behavior at both high and low levels of action planning, but the rate of increase was much steeper for low planning, which supports the prediction. However, when the intention–behavior relationship was strongest, planning had little effect, so planning may only be effective for those with lower intentions. As with the mediation effect, the moderation effect was no longer present once past behavior was included in the model. Taken together, current results do not provide strong evidence for the role of action planning in mediating and moderating the intention–behavior relationship for social distancing.
Limitations and Avenues for Future Research
Current findings should be interpreted in light of some notable limitations. First, attrition rates in both samples were relatively high given the relatively brief time between the baseline survey and follow-up. High attrition could lead to selection bias with those who are more motivated or engaged overrepresented in the sample. While participants were reminded multiple times to complete follow-up measures, we acknowledge that more intensive recruitment and incentivization of nonresponders may have minimized drop out. Attrition also affected the demographic profile of the sample, particularly among underrepresented groups. Although the effect sizes of these differences were small, they were not trivial. This is particularly pertinent in the current context given emerging data indicating that COVID-19 infection and mortality rates are significantly higher in underrepresented minority and socioeconomic groups [37]. A potential solution would be to oversample in underrepresented groups likely to have low retention rates and is a recommendation for future research. It is also important to note that, although our sampling strategy ensured that the distribution of participants in our samples matched those of the national population according to gender and state, we did not stratify the sample by key demographic or socioeconomic factors. The samples, therefore, should not be considered representative of the national populations of Australia or the USA. Taken together, the bias linked to attrition rates and nonrepresentativeness of the samples places limits on the extent to which current findings can be generalized to the broader population.
Second, the intention–behavior “gap” in the current study resulted in small indirect effects of intention determinants, such as subjective and moral norms and PBC on social distancing behavior. This is a limitation of the current model and means that intervention strategies aimed at changing intention determinants may have relatively modest effects on behavior change. However, small effects may still translate to large numbers of people changing if interventions targeting change in these constructs are administered at the population level. Future intervention research is, nevertheless, needed to verify the effects of targeting change in model constructs on behavior. Research should also adopt behavioral measures that can be converted to meaningful metrics that demonstrate practically significant changes in social distancing behavior (e.g., numbers of people complying with social distancing guidelines when venturing outside the home).
Third, the current study observed social distancing over a relatively brief time frame. Short-range prediction has value as it helps identify potential determinants of social distancing behavior. However, consistency in performing social distancing over time is important for the effective prevention of virus transmission, so research on the determinants of social distancing in the long term is a priority. The relatively short time lag is also likely to be the reason why past behavior had such a pervasive effect in predicting behavior and other constructs in the model. The relevance of past behavior is likely to wane over time, so examining prediction over time may be more revealing as to the social cognition predictors of this behavior and the processes involved.
Fourth, the correlational design precludes the inference of causal effects among the constructs in the current model, so the proposed direction of effects are inferred from theory alone, not the data. Causal sequencing among variables would necessitate experimental or controlled intervention designs. Verification of such effects will highlight the value of the model in informing interventions to promote changes in social distancing behavior. In addition, the inclusion of past behavior in the current analysis modeled change in behavior over time. Past behavior also had the effect of modeling residual effects of unmeasured constructs on behavior, such as past measures of the model constructs. However, the adoption of a cross-lagged panel design would better facilitate the examination of how the change in specific model constructs over time affects social distancing behavior and permit tests of reciprocal effects. It is also important that the effects of past behavior do not provide definitive evidence that affecting change in model constructs, such as intentions or habit, through intervention will lead to a concomitant change in social distancing behavior. This highlights the imperative of intervention research that tests the efficacy of manipulating constructs from the current model in promoting social distancing behavior and illustrates the extent to which model constructs can be modified.
Finally, the current research relies exclusively on self-report measures. While self-reported behavior has exhibited concurrent validity when evaluated against non-self-report measures, such as behavior measured using devices or direct observation, the potential for recall bias or inaccurate reporting likely introduces additional measurement error in the behavioral measure, which would affect model relations. Further, self-reported data are also at risk of self-presentation bias and socially desirable responding. Health behaviors, particularly social distancing behavior in the context of a pandemic, are likely to be considered desirable, which may have compelled respondents to provide positive responses, without even being aware of such biases. Although we stressed anonymity to participants to make it clear that they had license to report their behavior without prejudice, this is unlikely to have fully eliminated such biases. Current data should, therefore, be interpreted in light of these potential biases and their potential to contribute to error variance in observed effects. Future research may consider the use of devices, such as GPS tracking of cellular phones, as alternative means to measure social distancing behavior that do not rely on self-report.
Conclusion
The current research aimed to identify the determinants of social distancing behavior to prevent transmission of the SARS-CoV-2 virus in samples of Australian and U.S. residents. The research applied an integrated theoretical model that included multiple social cognition determinants relevant to the behavioral context, and the processes involved, with the potential to be modifiable through intervention. Results provided qualified support for the proposed model, highlighting the importance of social and moral beliefs, and perceptions of control, in predicting intention, and habit and intention in predicting behavior, in both samples, although effects were relatively modest, particularly when past behavior was accounted for. Findings suggest that interventions aimed at promoting social distancing behavior should provide messages highlighting individuals’ obligations to significant others and the moral imperative of protecting the most vulnerable as reasons for social distancing, provide environments (e.g., workplaces and grocery stores) that are barrier free and easy to socially distance and provide consistent opportunities in regular, stable contexts to engage in social distancing to develop habits. Future research should seek to provide longer-range prediction of social distancing behavior by the integrated model constructs and test the stability and reciprocal relations among its constructs using a cross-lagged panel design.
Supplementary Material
kaaa073_suppl_Supplementary_Appendix Click here for additional data file.
|
|
sec
|
Introduction
The novel coronavirus disease 2019 (COVID-19) pandemic has emerged as a truly global public health crisis [1]. While symptoms of COVID-19 are relatively mild without serious consequences in the majority of cases [2], modeling data suggest that approximately 4% of the global population is at risk of severe COVID-19 if infected and may require hospital admission for treatment [3]. Furthermore, SARS-CoV-2, the virus that causes COVID-19, is highly contagious, spreading mainly through person-to-person contact. Government-mandated measures to reduce transmission include advocacy of behaviors like wearing face masks and social distancing and issuing “shelter-in-place” orders and bans on public gatherings [4].
Social distancing, defined as maintaining a distance of at least 3–6 feet (1–2 m) from other people not from the same household is considered particularly effective in minimizing SARS-CoV-2 transmission [5, 6]. “Shelter-in-place” orders (also referred to as “stay-at-home” or “lockdown” orders) represent means to mandate social distancing by minimizing incidences of person-to-person contact outside individuals’ immediate household. Similarly, bans on public gatherings seek to limit the frequency and number of people with whom they come into close contact. However, such actions do not eliminate all potential contact because individuals under such orders still need to break from shelter to fetch provisions and, for members of essential professions, to go to work. It is, therefore, imperative that individuals comply with public health guidelines advocating the practice of social distancing when they come into contact with others. Compliance with guidelines is also highly important in regions that have not issued formal “shelter-in-place” orders but have instead provided “safer-at-home” guidelines and in areas that have begun to lift “shelter-in-place” orders.
Public health organizations have been tasked with developing behavioral interventions that are efficacious in promoting social distancing behaviors among the general population [6]. Given that social distancing is a relatively novel behavior in many countries, identification of the determinants of social distancing behavior has become critical. Moreover, identifying determinants that are potentially modifiable through intervention, that is, can be targeted in messages or campaigns of behavioral interventions aimed at promoting social distance, is a recognized priority [7]. There have, therefore, been calls for research informed by behavioral science that identifies key determinants of preventive behaviors in the context of the current pandemic, particularly social distancing [7, 8]. However, there is relatively little research on the determinants of social distancing, particularly in the context of communicable disease prevention (e.g., influenza) in a global pandemic [9]. Previous research, for example, has tended to focus on the social cognition determinants of other preventive behaviors, such as facemask wearing [10], or focused on hypothetical scenarios [11] in the context of influenza prevention. To date, there are few studies informed by behavioral science on the individual determinants of social distancing in the context of the COVID-19 pandemic.
To fill this evidence gap, the current study aimed to identify the determinants of social distancing behavior among individuals subject to social distancing regulations during the COVID-19 pandemic. The research adopted an integrated theoretical approach based on social cognition theories to identify constructs that predict social distancing behavior and the processes involved. The research is expected to provide evidence of potentially modifiable targets for behavior change interventions aimed at promoting social distancing. Such interventions may contribute to reduced infection rates during the current pandemic and may assist in preventing a “second wave” of infections as “shelter-in-place” orders are lifted [5].
Social Distancing Determinants: An Integrated Social Cognition Approach
Research examining health behavior determinants has a long tradition of applying social cognition theories [12], which assume health behavior enactment is a reasoned process determined by beliefs, such as risk perception, attitude, social norm, and perceptions of control or self-efficacy. A prototypical social cognition approach is offered by the theory of planned behavior [13]. In the theory, individuals’ intention to perform the target behavior is proposed as the most proximal determinant of the performance of a future target behavior. Intention is a function of three constructs that summarize sets of beliefs regarding the future behavior: attitude (beliefs that the behavior will have advantageous or disadvantageous consequences), subjective norm (beliefs that significant others express support for performing the behavior), and perceived behavioral control (PBC; beliefs in the capacity to perform the behavior and to overcome barriers to the behavior). Intention is proposed to mediate the effects of attitude, subjective norm, and PBC on behavior. PBC is also proposed to predict behavior directly when it approximates actual control. Theory predictions have been supported in correlational and prospective research across multiple behaviors, contexts, and populations [14].
While the elegant parsimony of the theory of planned behavior is appealing, it is not without limitations. Research applying the theory has indicated that substantive variance in health behavior remains unexplained [14]. In addition, the size of the effect of intention on health behavior is often modest, suggesting a “shortfall” in those who report an intention to perform the behavior and those who act on their intention [15]. Researchers have, therefore, proposed modifications to the theory to resolve these limitations, such as integrating additional constructs from other theories, in the theory to predict behavior more effectively and address the intention–behavior “gap” [16].
Introducing additional constructs to the theory is one approach to increasing explained variance in health behavior. For example, researchers have examined relations between moral norms, an additional form of normative influence, and health behavior. Moral norms are considered particularly relevant when there is a moral imperative for acting (e.g., vaccination and blood donation) [17]. In the context of COVID-19, messaging from public health authorities on COVID-19-preventive behaviors has focused on protecting the vulnerable (e.g., immunosuppressed individuals, those with underlying health conditions, and the elderly) [3]. On this basis, we reasoned that moral norm would constitute a highly relevant determinant of social distancing intention and behavior in the context of the pandemic. In addition, anticipated regret has been shown to predict behaviors perceived likely to have adverse consequences or result in significant losses if not performed [17]. Failure to perform social distancing behaviors may be perceived as having highly undesirable consequences, such as becoming infected or infecting vulnerable others. We, therefore, included moral norm and anticipated regret as additional predictors of intention to perform social distancing behavior in our integrated model.
Researchers have applied “dual-phase” models as a means to resolve the limitation of the intention–behavior “gap.” Models like the model of action phases [18] and the health action process approach (HAPA) [19] propose that individuals need to augment their intentions with action plans in order to enact them. Action plans reflect the extent to which individuals have specified when, where, and how they will perform the intended behavior. The model of action phases [18] suggests that individuals will more likely enact their intentions if they form an action plan, so action plans are proposed to moderate the intention–behavior relationship. By contrast, the HAPA suggests that planning is part of the process of intention enactment such that action plans mediate the intention–behavior relationship [19]. Meta-analyses of studies in health behavior have supported both processes [20, 21], and we aimed to test both in our proposed integrated model of social distancing behavior.
While social cognition theories like the theory of planned behavior assume participation in health behavior to be a reasoned process, research applying such theories has shown that past behavior remains a pervasive determinant of behavior alongside the theory constructs [22, 23]. The inclusion of past behavior as an independent behavioral predictor in a social cognition theory is important because it provides a test of its sufficiency in accounting for unique variance in behavior. However, residual effects of past behavior on behavior are also assumed to model the effects of other unmeasured constructs on behavior [23]. One candidate construct is habit, which reflects the “nonconscious” or “automatic” enactment of a behavior developed through its repeated performance in stable contexts [24, 25]. Research examining the effects of habit in the context of social cognition theories has examined how self-reports of experiencing the behavior as “automatic” and “unthinking” predict health behavior independent of intentions [26]. The introduction of habit in our augmented model, therefore, may provide important information on the extent to which social distancing behavior is determined by reasoned or nonconscious processes [27].
The Present Study
The present study aimed to identify the determinants of participation in social distancing behavior among individuals in the context of COVID-19 using an integrated social cognition model that incorporated constructs from the theory of planned behavior with moral norm, anticipated regret, action planning, and self-reported habit. We tested predictions of the proposed model in a prospective correlational study in two separate samples of adults from Australia and the USA, respectively. These countries provide an opportunity to examine the determinants of social distancing because they experienced rapid increases in COVID-19 cases relatively early in the pandemic and introduced public health advice and “lockdown” measures to minimize transmission via social distancing. In our proposed model (Model 1; Fig. 1), attitude, subjective norm, PBC, moral norm, and anticipated regret were specified as predictors of intention, and intention, PBC, and habit as predictors of social distancing behavior. Intention was proposed to mediate the effects of the social cognition constructs on behavior. The role of action planning as a mediator and moderator of the intention–behavior relationship was also specified. We also specified a second model (Model 2; Fig. 1) in which past social distancing behavior was included as a direct predictor of all constructs in the model, providing a test of its sufficiency. Although research demonstrating that social distancing behavior clusters with other health behaviors indicates that application of social cognition theories is viable for this behavior [28], research is needed to verify this contention and the current study contributes to this goal. The research may assist in identifying potentially modifiable constructs that relate to social distancing behavior. Such information may provide useful information to inform social distancing interventions focused on reducing the spread of COVID-19 and, more broadly, other communicable diseases.
Fig. 1. Standardized parameter estimates of the integrated model. Upper panel presents the model excluding past behavior (Model 1) and the lower panel presents the model including past behavior (Model 2). Coefficients printed on the upper line are for the Australian sample and coefficients printed on the lower line are for the U.S. sample. †Effect is significantly different across the Australian and U.S. samples in multiple-group analyses. *p < .05, **p < .01, ***p < .001.
|
|
title
|
Introduction
|
|
p
|
The novel coronavirus disease 2019 (COVID-19) pandemic has emerged as a truly global public health crisis [1]. While symptoms of COVID-19 are relatively mild without serious consequences in the majority of cases [2], modeling data suggest that approximately 4% of the global population is at risk of severe COVID-19 if infected and may require hospital admission for treatment [3]. Furthermore, SARS-CoV-2, the virus that causes COVID-19, is highly contagious, spreading mainly through person-to-person contact. Government-mandated measures to reduce transmission include advocacy of behaviors like wearing face masks and social distancing and issuing “shelter-in-place” orders and bans on public gatherings [4].
|
|
p
|
Social distancing, defined as maintaining a distance of at least 3–6 feet (1–2 m) from other people not from the same household is considered particularly effective in minimizing SARS-CoV-2 transmission [5, 6]. “Shelter-in-place” orders (also referred to as “stay-at-home” or “lockdown” orders) represent means to mandate social distancing by minimizing incidences of person-to-person contact outside individuals’ immediate household. Similarly, bans on public gatherings seek to limit the frequency and number of people with whom they come into close contact. However, such actions do not eliminate all potential contact because individuals under such orders still need to break from shelter to fetch provisions and, for members of essential professions, to go to work. It is, therefore, imperative that individuals comply with public health guidelines advocating the practice of social distancing when they come into contact with others. Compliance with guidelines is also highly important in regions that have not issued formal “shelter-in-place” orders but have instead provided “safer-at-home” guidelines and in areas that have begun to lift “shelter-in-place” orders.
|
|
p
|
Public health organizations have been tasked with developing behavioral interventions that are efficacious in promoting social distancing behaviors among the general population [6]. Given that social distancing is a relatively novel behavior in many countries, identification of the determinants of social distancing behavior has become critical. Moreover, identifying determinants that are potentially modifiable through intervention, that is, can be targeted in messages or campaigns of behavioral interventions aimed at promoting social distance, is a recognized priority [7]. There have, therefore, been calls for research informed by behavioral science that identifies key determinants of preventive behaviors in the context of the current pandemic, particularly social distancing [7, 8]. However, there is relatively little research on the determinants of social distancing, particularly in the context of communicable disease prevention (e.g., influenza) in a global pandemic [9]. Previous research, for example, has tended to focus on the social cognition determinants of other preventive behaviors, such as facemask wearing [10], or focused on hypothetical scenarios [11] in the context of influenza prevention. To date, there are few studies informed by behavioral science on the individual determinants of social distancing in the context of the COVID-19 pandemic.
|
|
p
|
To fill this evidence gap, the current study aimed to identify the determinants of social distancing behavior among individuals subject to social distancing regulations during the COVID-19 pandemic. The research adopted an integrated theoretical approach based on social cognition theories to identify constructs that predict social distancing behavior and the processes involved. The research is expected to provide evidence of potentially modifiable targets for behavior change interventions aimed at promoting social distancing. Such interventions may contribute to reduced infection rates during the current pandemic and may assist in preventing a “second wave” of infections as “shelter-in-place” orders are lifted [5].
|
|
sec
|
Social Distancing Determinants: An Integrated Social Cognition Approach
Research examining health behavior determinants has a long tradition of applying social cognition theories [12], which assume health behavior enactment is a reasoned process determined by beliefs, such as risk perception, attitude, social norm, and perceptions of control or self-efficacy. A prototypical social cognition approach is offered by the theory of planned behavior [13]. In the theory, individuals’ intention to perform the target behavior is proposed as the most proximal determinant of the performance of a future target behavior. Intention is a function of three constructs that summarize sets of beliefs regarding the future behavior: attitude (beliefs that the behavior will have advantageous or disadvantageous consequences), subjective norm (beliefs that significant others express support for performing the behavior), and perceived behavioral control (PBC; beliefs in the capacity to perform the behavior and to overcome barriers to the behavior). Intention is proposed to mediate the effects of attitude, subjective norm, and PBC on behavior. PBC is also proposed to predict behavior directly when it approximates actual control. Theory predictions have been supported in correlational and prospective research across multiple behaviors, contexts, and populations [14].
While the elegant parsimony of the theory of planned behavior is appealing, it is not without limitations. Research applying the theory has indicated that substantive variance in health behavior remains unexplained [14]. In addition, the size of the effect of intention on health behavior is often modest, suggesting a “shortfall” in those who report an intention to perform the behavior and those who act on their intention [15]. Researchers have, therefore, proposed modifications to the theory to resolve these limitations, such as integrating additional constructs from other theories, in the theory to predict behavior more effectively and address the intention–behavior “gap” [16].
Introducing additional constructs to the theory is one approach to increasing explained variance in health behavior. For example, researchers have examined relations between moral norms, an additional form of normative influence, and health behavior. Moral norms are considered particularly relevant when there is a moral imperative for acting (e.g., vaccination and blood donation) [17]. In the context of COVID-19, messaging from public health authorities on COVID-19-preventive behaviors has focused on protecting the vulnerable (e.g., immunosuppressed individuals, those with underlying health conditions, and the elderly) [3]. On this basis, we reasoned that moral norm would constitute a highly relevant determinant of social distancing intention and behavior in the context of the pandemic. In addition, anticipated regret has been shown to predict behaviors perceived likely to have adverse consequences or result in significant losses if not performed [17]. Failure to perform social distancing behaviors may be perceived as having highly undesirable consequences, such as becoming infected or infecting vulnerable others. We, therefore, included moral norm and anticipated regret as additional predictors of intention to perform social distancing behavior in our integrated model.
Researchers have applied “dual-phase” models as a means to resolve the limitation of the intention–behavior “gap.” Models like the model of action phases [18] and the health action process approach (HAPA) [19] propose that individuals need to augment their intentions with action plans in order to enact them. Action plans reflect the extent to which individuals have specified when, where, and how they will perform the intended behavior. The model of action phases [18] suggests that individuals will more likely enact their intentions if they form an action plan, so action plans are proposed to moderate the intention–behavior relationship. By contrast, the HAPA suggests that planning is part of the process of intention enactment such that action plans mediate the intention–behavior relationship [19]. Meta-analyses of studies in health behavior have supported both processes [20, 21], and we aimed to test both in our proposed integrated model of social distancing behavior.
While social cognition theories like the theory of planned behavior assume participation in health behavior to be a reasoned process, research applying such theories has shown that past behavior remains a pervasive determinant of behavior alongside the theory constructs [22, 23]. The inclusion of past behavior as an independent behavioral predictor in a social cognition theory is important because it provides a test of its sufficiency in accounting for unique variance in behavior. However, residual effects of past behavior on behavior are also assumed to model the effects of other unmeasured constructs on behavior [23]. One candidate construct is habit, which reflects the “nonconscious” or “automatic” enactment of a behavior developed through its repeated performance in stable contexts [24, 25]. Research examining the effects of habit in the context of social cognition theories has examined how self-reports of experiencing the behavior as “automatic” and “unthinking” predict health behavior independent of intentions [26]. The introduction of habit in our augmented model, therefore, may provide important information on the extent to which social distancing behavior is determined by reasoned or nonconscious processes [27].
|
|
title
|
Social Distancing Determinants: An Integrated Social Cognition Approach
|
|
p
|
Research examining health behavior determinants has a long tradition of applying social cognition theories [12], which assume health behavior enactment is a reasoned process determined by beliefs, such as risk perception, attitude, social norm, and perceptions of control or self-efficacy. A prototypical social cognition approach is offered by the theory of planned behavior [13]. In the theory, individuals’ intention to perform the target behavior is proposed as the most proximal determinant of the performance of a future target behavior. Intention is a function of three constructs that summarize sets of beliefs regarding the future behavior: attitude (beliefs that the behavior will have advantageous or disadvantageous consequences), subjective norm (beliefs that significant others express support for performing the behavior), and perceived behavioral control (PBC; beliefs in the capacity to perform the behavior and to overcome barriers to the behavior). Intention is proposed to mediate the effects of attitude, subjective norm, and PBC on behavior. PBC is also proposed to predict behavior directly when it approximates actual control. Theory predictions have been supported in correlational and prospective research across multiple behaviors, contexts, and populations [14].
|
|
p
|
While the elegant parsimony of the theory of planned behavior is appealing, it is not without limitations. Research applying the theory has indicated that substantive variance in health behavior remains unexplained [14]. In addition, the size of the effect of intention on health behavior is often modest, suggesting a “shortfall” in those who report an intention to perform the behavior and those who act on their intention [15]. Researchers have, therefore, proposed modifications to the theory to resolve these limitations, such as integrating additional constructs from other theories, in the theory to predict behavior more effectively and address the intention–behavior “gap” [16].
|
|
p
|
Introducing additional constructs to the theory is one approach to increasing explained variance in health behavior. For example, researchers have examined relations between moral norms, an additional form of normative influence, and health behavior. Moral norms are considered particularly relevant when there is a moral imperative for acting (e.g., vaccination and blood donation) [17]. In the context of COVID-19, messaging from public health authorities on COVID-19-preventive behaviors has focused on protecting the vulnerable (e.g., immunosuppressed individuals, those with underlying health conditions, and the elderly) [3]. On this basis, we reasoned that moral norm would constitute a highly relevant determinant of social distancing intention and behavior in the context of the pandemic. In addition, anticipated regret has been shown to predict behaviors perceived likely to have adverse consequences or result in significant losses if not performed [17]. Failure to perform social distancing behaviors may be perceived as having highly undesirable consequences, such as becoming infected or infecting vulnerable others. We, therefore, included moral norm and anticipated regret as additional predictors of intention to perform social distancing behavior in our integrated model.
|
|
p
|
Researchers have applied “dual-phase” models as a means to resolve the limitation of the intention–behavior “gap.” Models like the model of action phases [18] and the health action process approach (HAPA) [19] propose that individuals need to augment their intentions with action plans in order to enact them. Action plans reflect the extent to which individuals have specified when, where, and how they will perform the intended behavior. The model of action phases [18] suggests that individuals will more likely enact their intentions if they form an action plan, so action plans are proposed to moderate the intention–behavior relationship. By contrast, the HAPA suggests that planning is part of the process of intention enactment such that action plans mediate the intention–behavior relationship [19]. Meta-analyses of studies in health behavior have supported both processes [20, 21], and we aimed to test both in our proposed integrated model of social distancing behavior.
|
|
p
|
While social cognition theories like the theory of planned behavior assume participation in health behavior to be a reasoned process, research applying such theories has shown that past behavior remains a pervasive determinant of behavior alongside the theory constructs [22, 23]. The inclusion of past behavior as an independent behavioral predictor in a social cognition theory is important because it provides a test of its sufficiency in accounting for unique variance in behavior. However, residual effects of past behavior on behavior are also assumed to model the effects of other unmeasured constructs on behavior [23]. One candidate construct is habit, which reflects the “nonconscious” or “automatic” enactment of a behavior developed through its repeated performance in stable contexts [24, 25]. Research examining the effects of habit in the context of social cognition theories has examined how self-reports of experiencing the behavior as “automatic” and “unthinking” predict health behavior independent of intentions [26]. The introduction of habit in our augmented model, therefore, may provide important information on the extent to which social distancing behavior is determined by reasoned or nonconscious processes [27].
|
|
sec
|
The Present Study
The present study aimed to identify the determinants of participation in social distancing behavior among individuals in the context of COVID-19 using an integrated social cognition model that incorporated constructs from the theory of planned behavior with moral norm, anticipated regret, action planning, and self-reported habit. We tested predictions of the proposed model in a prospective correlational study in two separate samples of adults from Australia and the USA, respectively. These countries provide an opportunity to examine the determinants of social distancing because they experienced rapid increases in COVID-19 cases relatively early in the pandemic and introduced public health advice and “lockdown” measures to minimize transmission via social distancing. In our proposed model (Model 1; Fig. 1), attitude, subjective norm, PBC, moral norm, and anticipated regret were specified as predictors of intention, and intention, PBC, and habit as predictors of social distancing behavior. Intention was proposed to mediate the effects of the social cognition constructs on behavior. The role of action planning as a mediator and moderator of the intention–behavior relationship was also specified. We also specified a second model (Model 2; Fig. 1) in which past social distancing behavior was included as a direct predictor of all constructs in the model, providing a test of its sufficiency. Although research demonstrating that social distancing behavior clusters with other health behaviors indicates that application of social cognition theories is viable for this behavior [28], research is needed to verify this contention and the current study contributes to this goal. The research may assist in identifying potentially modifiable constructs that relate to social distancing behavior. Such information may provide useful information to inform social distancing interventions focused on reducing the spread of COVID-19 and, more broadly, other communicable diseases.
Fig. 1. Standardized parameter estimates of the integrated model. Upper panel presents the model excluding past behavior (Model 1) and the lower panel presents the model including past behavior (Model 2). Coefficients printed on the upper line are for the Australian sample and coefficients printed on the lower line are for the U.S. sample. †Effect is significantly different across the Australian and U.S. samples in multiple-group analyses. *p < .05, **p < .01, ***p < .001.
|
|
title
|
The Present Study
|
|
p
|
The present study aimed to identify the determinants of participation in social distancing behavior among individuals in the context of COVID-19 using an integrated social cognition model that incorporated constructs from the theory of planned behavior with moral norm, anticipated regret, action planning, and self-reported habit. We tested predictions of the proposed model in a prospective correlational study in two separate samples of adults from Australia and the USA, respectively. These countries provide an opportunity to examine the determinants of social distancing because they experienced rapid increases in COVID-19 cases relatively early in the pandemic and introduced public health advice and “lockdown” measures to minimize transmission via social distancing. In our proposed model (Model 1; Fig. 1), attitude, subjective norm, PBC, moral norm, and anticipated regret were specified as predictors of intention, and intention, PBC, and habit as predictors of social distancing behavior. Intention was proposed to mediate the effects of the social cognition constructs on behavior. The role of action planning as a mediator and moderator of the intention–behavior relationship was also specified. We also specified a second model (Model 2; Fig. 1) in which past social distancing behavior was included as a direct predictor of all constructs in the model, providing a test of its sufficiency. Although research demonstrating that social distancing behavior clusters with other health behaviors indicates that application of social cognition theories is viable for this behavior [28], research is needed to verify this contention and the current study contributes to this goal. The research may assist in identifying potentially modifiable constructs that relate to social distancing behavior. Such information may provide useful information to inform social distancing interventions focused on reducing the spread of COVID-19 and, more broadly, other communicable diseases.
|
|
figure
|
Fig. 1. Standardized parameter estimates of the integrated model. Upper panel presents the model excluding past behavior (Model 1) and the lower panel presents the model including past behavior (Model 2). Coefficients printed on the upper line are for the Australian sample and coefficients printed on the lower line are for the U.S. sample. †Effect is significantly different across the Australian and U.S. samples in multiple-group analyses. *p < .05, **p < .01, ***p < .001.
|
|
label
|
Fig. 1.
|
|
caption
|
Standardized parameter estimates of the integrated model. Upper panel presents the model excluding past behavior (Model 1) and the lower panel presents the model including past behavior (Model 2). Coefficients printed on the upper line are for the Australian sample and coefficients printed on the lower line are for the U.S. sample. †Effect is significantly different across the Australian and U.S. samples in multiple-group analyses. *p < .05, **p < .01, ***p < .001.
|
|
p
|
Standardized parameter estimates of the integrated model. Upper panel presents the model excluding past behavior (Model 1) and the lower panel presents the model including past behavior (Model 2). Coefficients printed on the upper line are for the Australian sample and coefficients printed on the lower line are for the U.S. sample. †Effect is significantly different across the Australian and U.S. samples in multiple-group analyses. *p < .05, **p < .01, ***p < .001.
|
|
sec
|
Method
Participants and Recruitment
Samples of Australian (N = 495, 50.1% female) and U.S. (N = 701, 48.9% female) residents were recruited via an online research panel company. To be eligible for inclusion, participants needed to be aged 18 years or older and not subject to formal quarantine for COVID-19. Participants were also screened for age, gender, and geographical region and quotas imposed during recruitment to ensure that the final samples closely matched the national distributions for these characteristics in each country. Data were collected between April 1 and May 6, 2020. All participants in the Australian sample were subject to a national “shelter-in-place” order issued by the federal government. However, issuance of orders in the USA was devolved to state governments resulting in some variations. The vast majority of participants in the U.S. sample (n = 610, 87.0%) were subject to “shelter-in-place” orders for the duration of the study. However, some states did not impose “shelter-in-place” orders at all (Arkansas, Iowa, Nebraska, North Dakota, South Dakota, Utah, and Wyoming), so a minority of participants in the U.S. sample (n = 37, 5.3%) were never subject to an order. Furthermore, in some cases in the U.S. sample (n = 47, 6.7%), “shelter-in-place” orders had been lifted prior to follow-up data collection. However, among the states in the USA that did not have “shelter-in-place” orders, or lifted their orders during the study, all issued social distancing guidelines and encouraged the population to follow those guidelines. Baseline sample characteristics are presented in Table 1.
Table 1. Sample characteristics and descriptive statistics for study variables at baseline and at 1 week follow-up
Variable Australia sample U.S. sample
Baseline Follow-up Baseline Follow-up
Participants 495 365 701 440
Age, M years (SD) 47.09 (17.11) 49.78 (16.89) 45.55 (17.40) 51.77 (16.26)
Gender, n (%)a
Female 252 (51.1) 182 (50.1) 352 (48.9) 205 (46.6)
Male 241 (48.9) 181 (49.9) 341 (50.5) 231 (52.5)
Not specified/prefer not to answer 0 (0.0) 0 (0.0) 4 (0.6) 4 (0.9)
Employment status, n (%)b
Currently unemployed/full-time caregiver 231 (46.7) 180 (49.3) 330 (47.3) 216 (49.5)
Part-time/casual employed 97 (19.6) 65 (17.8) 106 (15.2) 60 (13.8)
Currently employed full time 140 (28.3) 104 (28.5) 233 (33.4) 147 (33.7)
Leave without pay/furloughed 27 (5.5) 16 (4.4) 28 (4.0) 13 (3.0)
Marital status, n (%)c
Married 184 (37.2) 146 (40.0) 300 (43.0) 224 (51.4)
Widowed 8 (1.6) 7 (1.9) 22 (3.2) 18 (4.1)
Separated/divorced 53 (10.7) 39 (10.7) 69 (9.9) 47 (10.8)
Never married 160 (32.3) 103 (28.2) 255 (36.6) 126(28.9)
Married de facto 90 (18.2) 70 (19.2) 51 (7.3) 21 (4.8)
Ethnicity, n (%)d
Black 3 (0.6) 1 (0.3) 52 (7.5) 26 (6.0)
Caucasian/White 392 (79.2) 304 (83.3) 566 (81.2) 376 (86.2)
Asian (South-East Asia/South Asia) 71 (14.3) 43 (11.8) 39 (5.6) 24 (5.5)
Middle-Eastern 6 (1.2) 3 (0.8) 1 (0.1) 0 (0.0)
Other 13 (2.6) 6 (1.6) 27 (3.9) 8 (1.8)
Prefer not to answer 10 (2.0) 8 (2.2) 12 (1.7) 2 (0.5)
Income, n (%)e
Zero income 8 (1.7) 4 (1.2) 31 (4.4) 19 (4.4)
$1–$199 ($1–$10,399) 9 (2.0) 6 (1.8) 40 (5.7) 24 (5.5)
$200–$299 ($10,400–$15,599) 12 (2.6) 8 (2.4) 34 (4.9) 23 (5.3)
$300–$399 ($15,600–$20,799) 19 (4.1) 12 (3.6) 38 (5.5) 23 (5.3)
$400–$599 ($20,800–$31,199) 42 (9.2) 33 (9.9) 62 (8.9) 33 (7.6)
$600–$799 ($31,200–$41,599) 57 (12.4) 42 (12.6) 61 (8.8) 39 (8.9)
$800–$999 ($41,600–$51,999) 45 (9.8) 31 (9.3) 68 (9.8) 46 (10.6)
$1,000–$1,249 ($52,000–$64,999) 39 (8.5) 32 (9.6) 48 (6.9) 38 (8.7)
$1,250–$1,499 ($65,000–$77,999) 28 (6.1) 22 (6.6) 59 (8.5) 41 (9.4)
$1,500–$1,999 ($78,000–$103,999) 72 (15.7) 50 (15.0) 72 (10.3) 48 (11.0)
$2,000 or more ($104,000 or more) 81 (17.6) 62 (18.6) 108 (15.5) 74 (17.0)
Prefer not to answer 47 (10.2) 32 (9.6) 76 (10.9) 28 (6.4)
Education level, n (%)
Completed junior/lower/primary school 18 (3.6) 17 (4.7) 6 (0.9) 0 (0.0)
Completed senior/high/secondary school 133 (26.9) 98 (26.8) 265 (37.8) 132 (30.0)
Postschool vocational qualification/diploma 147 (29.7) 111 (30.4) 138 (19.7) 94 (21.4)
Undergraduate university degree 131 (26.5) 93 (25.5) 214 (30.5) 159 (36.1)
Postgraduate university degree 66 (13.3) 46 (12.6) 78 (11.1) 55 (12.5)
SD standard deviation.
aTwo participants in the Australian sample did not report their gender and four participants in the U.S. sample did not report their gender.
bFour participants in the U.S. sample did not report their employment status.
cFour participants in the U.S. sample did not report their marital status.
dFour participants in the U.S. sample did not report their ethnicity.
eThirty-one participants in the Australian sample did not report their income and four participants in the U.S. sample did not report their income.
Design and Procedure
The study adopted a prospective correlational design with self-report measures of social cognition constructs from the proposed integrated model, intention, and past social distancing behavior administered at an initial data collection occasion in a survey administered using the Qualtrics online survey tool. Social cognition measures included the theory of planned behavior (attitude, subjective norm, and PBC), moral norm, anticipated regret, action planning, and habit constructs. Participants were informed that they were participating in a survey on their social distancing behavior and provided with information outlining study requirements. They were required to provide informed consent before proceeding with the survey. Participants were also provided with instructions on how to complete study measures and a definition of the target behavior: “The following survey will ask about your beliefs and attitudes about ‘social distancing’. What do we mean by social distancing? Social distancing (also known as ‘physical distancing’) is deliberately increasing the physical space between people to avoid spreading illness. The World Health Organization and other world leading health authorities suggest that you should maintain at least a 1–2 meter (3–6 feet) distance from other people to lessen the chances of getting infected with COVID-19. When answering the questions in this survey, think about your social distancing behavior (i.e., maintaining at least a 1–2 m (3–6 ft) distance from other people).” One week later, participants were re-contacted by the panel company and asked to self-report their habit and social distancing behavior over the previous week using the same measures used at the initial data collection occasion. Participants received a fixed sum of money for their participation based on expected completion time consistent with the panel company’s published rates. Approval for study procedures was granted prior to data collection from the Griffith University Human Research Ethics Committee.
Measures
Study measures were multi-item self-report measures of constructs based on published guidelines and measures used in previous studies [13, 29, 30]. Participants provided their responses on scales with seven-point response options. Complete study measures are provided in Supplementary Appendix A.
Social cognition constructs
Multi-item measures of attitude, subjective norm, PBC, moral norm, anticipated regret, and action planning were developed according to published guidelines [13, 29]. Each measure made explicit reference to the target behavior of social distancing, and participants were reminded of the definition of social distancing before completing the measures.
Intention
Participants’ intention to participate in social distancing behavior over the next week was measured using a scale developed according to published guidelines [31].
Habit
Habit was measured at both time points using the behavioral automaticity items of Verplanken and Orbell’s self-report habit index [25]. The measure measures individuals’ reflections on the extent to which the behavior is experienced as automatic and enacted without thought.
Past behavior and behavior
Participants self-reported their participation in social distancing behavior to minimize the transmission of the SARS-CoV-2 virus that causes COVID-19. The measure comprised two items prompting participants to report their frequency of social distancing behavior in the previous week. This is based on previously used self-report behavioral measures that have demonstrated concurrent validity with non-self-report measures in other behavioral contexts [32].
Demographic variables
Participants self-reported their age in years, gender, employment status (currently unemployed/full-time caregiver, currently full-time employed, part-time employed, or on leave without pay/furloughed), marital status (married, widowed, separated/divorced, never married, or in a de factor relationship), annual household income stratified by 11 income levels based on Australian and U.S. national averages, highest level of formal education (completed junior/lower/primary school, completed senior/high/secondary school, postschool vocational qualification/diploma, further education diploma, undergraduate university degree, or postgraduate university degree), and ethnicity (Black, Caucasian/White, Asian, or Middle-eastern). Binary income (low income vs. middle/high income), highest education level (completed school education only vs. completed postschool education), and ethnicity (Caucasian/White vs. non-White) variables were computed for use in subsequent analyses.
Data Analysis
Hypothesized relations among the integrated model constructs were tested in the Australian and U.S. samples separately using variance-based structural equation modeling implemented in the WARP 7.0 analysis package [33]. Model parameters and standard errors (SEs) were computed using the “Stable3” estimation method, which has been shown to provide the most precise parameter estimates in complex structural models in smaller samples and outperforms bootstrapping methods in simulation studies [33]. Simulation studies have also shown this method to provide more consistent and precise estimates in data containing outliers, which may inflate SEs and lead to abnormally high p-values [33]. Two models were estimated in each sample: a model testing predictions of the proposed integrated model with the binary demographic variables also included as covariates (Model 1; Fig. 1, upper panel) and a model that included effects of past social distancing behavior (Model 2; Fig. 1, lower panel). All constructs were latent variables indicated by single or multiple items. There were no missing data for the social cognition and self-reported behavioral variables. There were a few instances of missing data for the demographic variables ranging from 0.5% to 8.8% in the Australia sample, and 0.9% to 6.4% in the U.S. sample. Missing data are reported in Supplementary Appendix B. Missing data were imputed using stochastic hierarchical regression [33].
The analysis afforded a number of analyses to evaluate the adequacy of measures used to indicate the latent variables in the model. Construct validity of the latent factors for the social cognition, intention, and behavioral variables was established using the normalized factor pattern loadings after oblique rotation and Kaiser normalization [33] and the average variance extracted (AVE), which should approach or exceed .700 and .500, respectively. Internal consistency of the factors was estimated using omega reliability coefficients (ω) and composite reliability coefficients (ρ), which should exceed .700 and ideally approach .900. We also conducted tests of the discriminant validity of the constructs in the model. Discriminant validity was supported when the square root of the AVE for each latent variable exceeded its correlation with other latent variables.
Adequacy of the proposed model in describing the data was established using the goodness-of-fit (GoF) index with values of .100, .250, and .360 corresponding to small, medium, and large effect sizes. Further information on model quality was provided by the average path coefficient and average R2 coefficient. These indices summarize the average parameter estimates of relations in the model and the amount of variance explained in each dependent variable, respectively, and should be statistically significant for a good-quality model. In addition, an overall GoF index is provided by the average block variance inflation factor for model parameters and the average full collinearity variance inflation factor, which should be equal to or lower than 3.3 for well-fitting models. These indices indicate the extent to which latent variables in the model overlap and contribute to model multicollinearity. They, therefore, provide an indication as to the uniqueness of the existing latent variables in the model. Four further indices were also used to evaluate model quality: the Simpson’s paradox ratio (SPR), R2 contribution ratio (R2CR), the statistical suppression ratio (SSR), and the nonlinear bivariate causality direction ratio (NLBCDR). The SPR indicates whether the model is free from incidences of Simpson’s paradox (i.e., when the path coefficient and the correlation associated with a latent variable have opposite signs), indicating a causality problem. The SPR should exceed .700 and ideally approach 1.000. The R2CR and SSR provide indication of the extent to which models are free from instances of negative R2 contributions and statistical suppression. The R2CR and SSR should exceed .900 and .700, respectively. The NLBCDR provides an estimate of the extent to which the proposed “causal” associations in the proposed model are more tenable than those in the opposite direction and provide an initial indicator of support for the hypothesized directions of the causal links in the proposed model compared to if the proposed direction were reversed. The NLBCDR should exceed .700 for high-quality models. Kock [33] provides further technical details on model fit and quality indices.
Model effects were estimated using standardized path coefficients with confidence intervals (CIs) and test statistics. Effect sizes were estimated using a variant of Cohen’s f-square coefficient and represent the individual contribution of the predictor variable to the R2 coefficients of the criterion latent variable. Values of .02, .15, and .35 represent small, medium, and large effect sizes, respectively. Differences in the path coefficients in the models across the samples were tested using multiple-group analysis using the Satterthwaite method with two-tailed significance tests.
We also tested whether the inclusion of participants that were never under a “shelter-in-place” order, or had the “shelter-in-place” order lifted during the study, affected predicted relations in the models. The small numbers of participants that were, at some point, not subjected to “shelter-in-place” orders meant we could not conduct a formal moderator analysis, so we conducted a sensitivity analysis testing whether effects in the models differed if data from these participants were excluded. Models excluding and including past behavior were estimated in samples excluding participants who were never subject to a “shelter-in place” order, and in the sample that were never subject to an order, or who had the order lifted at some stage during the study. Formal comparisons of parameter estimates in these models with those from the full sample were made using the Satterthwaite method. Data files, analysis scripts, and output files for all analyses are available online: https://osf.io/x9tms/.
|
|
title
|
Method
|
|
sec
|
Participants and Recruitment
Samples of Australian (N = 495, 50.1% female) and U.S. (N = 701, 48.9% female) residents were recruited via an online research panel company. To be eligible for inclusion, participants needed to be aged 18 years or older and not subject to formal quarantine for COVID-19. Participants were also screened for age, gender, and geographical region and quotas imposed during recruitment to ensure that the final samples closely matched the national distributions for these characteristics in each country. Data were collected between April 1 and May 6, 2020. All participants in the Australian sample were subject to a national “shelter-in-place” order issued by the federal government. However, issuance of orders in the USA was devolved to state governments resulting in some variations. The vast majority of participants in the U.S. sample (n = 610, 87.0%) were subject to “shelter-in-place” orders for the duration of the study. However, some states did not impose “shelter-in-place” orders at all (Arkansas, Iowa, Nebraska, North Dakota, South Dakota, Utah, and Wyoming), so a minority of participants in the U.S. sample (n = 37, 5.3%) were never subject to an order. Furthermore, in some cases in the U.S. sample (n = 47, 6.7%), “shelter-in-place” orders had been lifted prior to follow-up data collection. However, among the states in the USA that did not have “shelter-in-place” orders, or lifted their orders during the study, all issued social distancing guidelines and encouraged the population to follow those guidelines. Baseline sample characteristics are presented in Table 1.
Table 1. Sample characteristics and descriptive statistics for study variables at baseline and at 1 week follow-up
Variable Australia sample U.S. sample
Baseline Follow-up Baseline Follow-up
Participants 495 365 701 440
Age, M years (SD) 47.09 (17.11) 49.78 (16.89) 45.55 (17.40) 51.77 (16.26)
Gender, n (%)a
Female 252 (51.1) 182 (50.1) 352 (48.9) 205 (46.6)
Male 241 (48.9) 181 (49.9) 341 (50.5) 231 (52.5)
Not specified/prefer not to answer 0 (0.0) 0 (0.0) 4 (0.6) 4 (0.9)
Employment status, n (%)b
Currently unemployed/full-time caregiver 231 (46.7) 180 (49.3) 330 (47.3) 216 (49.5)
Part-time/casual employed 97 (19.6) 65 (17.8) 106 (15.2) 60 (13.8)
Currently employed full time 140 (28.3) 104 (28.5) 233 (33.4) 147 (33.7)
Leave without pay/furloughed 27 (5.5) 16 (4.4) 28 (4.0) 13 (3.0)
Marital status, n (%)c
Married 184 (37.2) 146 (40.0) 300 (43.0) 224 (51.4)
Widowed 8 (1.6) 7 (1.9) 22 (3.2) 18 (4.1)
Separated/divorced 53 (10.7) 39 (10.7) 69 (9.9) 47 (10.8)
Never married 160 (32.3) 103 (28.2) 255 (36.6) 126(28.9)
Married de facto 90 (18.2) 70 (19.2) 51 (7.3) 21 (4.8)
Ethnicity, n (%)d
Black 3 (0.6) 1 (0.3) 52 (7.5) 26 (6.0)
Caucasian/White 392 (79.2) 304 (83.3) 566 (81.2) 376 (86.2)
Asian (South-East Asia/South Asia) 71 (14.3) 43 (11.8) 39 (5.6) 24 (5.5)
Middle-Eastern 6 (1.2) 3 (0.8) 1 (0.1) 0 (0.0)
Other 13 (2.6) 6 (1.6) 27 (3.9) 8 (1.8)
Prefer not to answer 10 (2.0) 8 (2.2) 12 (1.7) 2 (0.5)
Income, n (%)e
Zero income 8 (1.7) 4 (1.2) 31 (4.4) 19 (4.4)
$1–$199 ($1–$10,399) 9 (2.0) 6 (1.8) 40 (5.7) 24 (5.5)
$200–$299 ($10,400–$15,599) 12 (2.6) 8 (2.4) 34 (4.9) 23 (5.3)
$300–$399 ($15,600–$20,799) 19 (4.1) 12 (3.6) 38 (5.5) 23 (5.3)
$400–$599 ($20,800–$31,199) 42 (9.2) 33 (9.9) 62 (8.9) 33 (7.6)
$600–$799 ($31,200–$41,599) 57 (12.4) 42 (12.6) 61 (8.8) 39 (8.9)
$800–$999 ($41,600–$51,999) 45 (9.8) 31 (9.3) 68 (9.8) 46 (10.6)
$1,000–$1,249 ($52,000–$64,999) 39 (8.5) 32 (9.6) 48 (6.9) 38 (8.7)
$1,250–$1,499 ($65,000–$77,999) 28 (6.1) 22 (6.6) 59 (8.5) 41 (9.4)
$1,500–$1,999 ($78,000–$103,999) 72 (15.7) 50 (15.0) 72 (10.3) 48 (11.0)
$2,000 or more ($104,000 or more) 81 (17.6) 62 (18.6) 108 (15.5) 74 (17.0)
Prefer not to answer 47 (10.2) 32 (9.6) 76 (10.9) 28 (6.4)
Education level, n (%)
Completed junior/lower/primary school 18 (3.6) 17 (4.7) 6 (0.9) 0 (0.0)
Completed senior/high/secondary school 133 (26.9) 98 (26.8) 265 (37.8) 132 (30.0)
Postschool vocational qualification/diploma 147 (29.7) 111 (30.4) 138 (19.7) 94 (21.4)
Undergraduate university degree 131 (26.5) 93 (25.5) 214 (30.5) 159 (36.1)
Postgraduate university degree 66 (13.3) 46 (12.6) 78 (11.1) 55 (12.5)
SD standard deviation.
aTwo participants in the Australian sample did not report their gender and four participants in the U.S. sample did not report their gender.
bFour participants in the U.S. sample did not report their employment status.
cFour participants in the U.S. sample did not report their marital status.
dFour participants in the U.S. sample did not report their ethnicity.
eThirty-one participants in the Australian sample did not report their income and four participants in the U.S. sample did not report their income.
|
|
title
|
Participants and Recruitment
|
|
p
|
Samples of Australian (N = 495, 50.1% female) and U.S. (N = 701, 48.9% female) residents were recruited via an online research panel company. To be eligible for inclusion, participants needed to be aged 18 years or older and not subject to formal quarantine for COVID-19. Participants were also screened for age, gender, and geographical region and quotas imposed during recruitment to ensure that the final samples closely matched the national distributions for these characteristics in each country. Data were collected between April 1 and May 6, 2020. All participants in the Australian sample were subject to a national “shelter-in-place” order issued by the federal government. However, issuance of orders in the USA was devolved to state governments resulting in some variations. The vast majority of participants in the U.S. sample (n = 610, 87.0%) were subject to “shelter-in-place” orders for the duration of the study. However, some states did not impose “shelter-in-place” orders at all (Arkansas, Iowa, Nebraska, North Dakota, South Dakota, Utah, and Wyoming), so a minority of participants in the U.S. sample (n = 37, 5.3%) were never subject to an order. Furthermore, in some cases in the U.S. sample (n = 47, 6.7%), “shelter-in-place” orders had been lifted prior to follow-up data collection. However, among the states in the USA that did not have “shelter-in-place” orders, or lifted their orders during the study, all issued social distancing guidelines and encouraged the population to follow those guidelines. Baseline sample characteristics are presented in Table 1.
|
|
table-wrap
|
Table 1. Sample characteristics and descriptive statistics for study variables at baseline and at 1 week follow-up
Variable Australia sample U.S. sample
Baseline Follow-up Baseline Follow-up
Participants 495 365 701 440
Age, M years (SD) 47.09 (17.11) 49.78 (16.89) 45.55 (17.40) 51.77 (16.26)
Gender, n (%)a
Female 252 (51.1) 182 (50.1) 352 (48.9) 205 (46.6)
Male 241 (48.9) 181 (49.9) 341 (50.5) 231 (52.5)
Not specified/prefer not to answer 0 (0.0) 0 (0.0) 4 (0.6) 4 (0.9)
Employment status, n (%)b
Currently unemployed/full-time caregiver 231 (46.7) 180 (49.3) 330 (47.3) 216 (49.5)
Part-time/casual employed 97 (19.6) 65 (17.8) 106 (15.2) 60 (13.8)
Currently employed full time 140 (28.3) 104 (28.5) 233 (33.4) 147 (33.7)
Leave without pay/furloughed 27 (5.5) 16 (4.4) 28 (4.0) 13 (3.0)
Marital status, n (%)c
Married 184 (37.2) 146 (40.0) 300 (43.0) 224 (51.4)
Widowed 8 (1.6) 7 (1.9) 22 (3.2) 18 (4.1)
Separated/divorced 53 (10.7) 39 (10.7) 69 (9.9) 47 (10.8)
Never married 160 (32.3) 103 (28.2) 255 (36.6) 126(28.9)
Married de facto 90 (18.2) 70 (19.2) 51 (7.3) 21 (4.8)
Ethnicity, n (%)d
Black 3 (0.6) 1 (0.3) 52 (7.5) 26 (6.0)
Caucasian/White 392 (79.2) 304 (83.3) 566 (81.2) 376 (86.2)
Asian (South-East Asia/South Asia) 71 (14.3) 43 (11.8) 39 (5.6) 24 (5.5)
Middle-Eastern 6 (1.2) 3 (0.8) 1 (0.1) 0 (0.0)
Other 13 (2.6) 6 (1.6) 27 (3.9) 8 (1.8)
Prefer not to answer 10 (2.0) 8 (2.2) 12 (1.7) 2 (0.5)
Income, n (%)e
Zero income 8 (1.7) 4 (1.2) 31 (4.4) 19 (4.4)
$1–$199 ($1–$10,399) 9 (2.0) 6 (1.8) 40 (5.7) 24 (5.5)
$200–$299 ($10,400–$15,599) 12 (2.6) 8 (2.4) 34 (4.9) 23 (5.3)
$300–$399 ($15,600–$20,799) 19 (4.1) 12 (3.6) 38 (5.5) 23 (5.3)
$400–$599 ($20,800–$31,199) 42 (9.2) 33 (9.9) 62 (8.9) 33 (7.6)
$600–$799 ($31,200–$41,599) 57 (12.4) 42 (12.6) 61 (8.8) 39 (8.9)
$800–$999 ($41,600–$51,999) 45 (9.8) 31 (9.3) 68 (9.8) 46 (10.6)
$1,000–$1,249 ($52,000–$64,999) 39 (8.5) 32 (9.6) 48 (6.9) 38 (8.7)
$1,250–$1,499 ($65,000–$77,999) 28 (6.1) 22 (6.6) 59 (8.5) 41 (9.4)
$1,500–$1,999 ($78,000–$103,999) 72 (15.7) 50 (15.0) 72 (10.3) 48 (11.0)
$2,000 or more ($104,000 or more) 81 (17.6) 62 (18.6) 108 (15.5) 74 (17.0)
Prefer not to answer 47 (10.2) 32 (9.6) 76 (10.9) 28 (6.4)
Education level, n (%)
Completed junior/lower/primary school 18 (3.6) 17 (4.7) 6 (0.9) 0 (0.0)
Completed senior/high/secondary school 133 (26.9) 98 (26.8) 265 (37.8) 132 (30.0)
Postschool vocational qualification/diploma 147 (29.7) 111 (30.4) 138 (19.7) 94 (21.4)
Undergraduate university degree 131 (26.5) 93 (25.5) 214 (30.5) 159 (36.1)
Postgraduate university degree 66 (13.3) 46 (12.6) 78 (11.1) 55 (12.5)
SD standard deviation.
aTwo participants in the Australian sample did not report their gender and four participants in the U.S. sample did not report their gender.
bFour participants in the U.S. sample did not report their employment status.
cFour participants in the U.S. sample did not report their marital status.
dFour participants in the U.S. sample did not report their ethnicity.
eThirty-one participants in the Australian sample did not report their income and four participants in the U.S. sample did not report their income.
|
|
label
|
Table 1.
|
|
caption
|
Sample characteristics and descriptive statistics for study variables at baseline and at 1 week follow-up
|
|
p
|
Sample characteristics and descriptive statistics for study variables at baseline and at 1 week follow-up
|
|
table
|
Variable Australia sample U.S. sample
Baseline Follow-up Baseline Follow-up
Participants 495 365 701 440
Age, M years (SD) 47.09 (17.11) 49.78 (16.89) 45.55 (17.40) 51.77 (16.26)
Gender, n (%)a
Female 252 (51.1) 182 (50.1) 352 (48.9) 205 (46.6)
Male 241 (48.9) 181 (49.9) 341 (50.5) 231 (52.5)
Not specified/prefer not to answer 0 (0.0) 0 (0.0) 4 (0.6) 4 (0.9)
Employment status, n (%)b
Currently unemployed/full-time caregiver 231 (46.7) 180 (49.3) 330 (47.3) 216 (49.5)
Part-time/casual employed 97 (19.6) 65 (17.8) 106 (15.2) 60 (13.8)
Currently employed full time 140 (28.3) 104 (28.5) 233 (33.4) 147 (33.7)
Leave without pay/furloughed 27 (5.5) 16 (4.4) 28 (4.0) 13 (3.0)
Marital status, n (%)c
Married 184 (37.2) 146 (40.0) 300 (43.0) 224 (51.4)
Widowed 8 (1.6) 7 (1.9) 22 (3.2) 18 (4.1)
Separated/divorced 53 (10.7) 39 (10.7) 69 (9.9) 47 (10.8)
Never married 160 (32.3) 103 (28.2) 255 (36.6) 126(28.9)
Married de facto 90 (18.2) 70 (19.2) 51 (7.3) 21 (4.8)
Ethnicity, n (%)d
Black 3 (0.6) 1 (0.3) 52 (7.5) 26 (6.0)
Caucasian/White 392 (79.2) 304 (83.3) 566 (81.2) 376 (86.2)
Asian (South-East Asia/South Asia) 71 (14.3) 43 (11.8) 39 (5.6) 24 (5.5)
Middle-Eastern 6 (1.2) 3 (0.8) 1 (0.1) 0 (0.0)
Other 13 (2.6) 6 (1.6) 27 (3.9) 8 (1.8)
Prefer not to answer 10 (2.0) 8 (2.2) 12 (1.7) 2 (0.5)
Income, n (%)e
Zero income 8 (1.7) 4 (1.2) 31 (4.4) 19 (4.4)
$1–$199 ($1–$10,399) 9 (2.0) 6 (1.8) 40 (5.7) 24 (5.5)
$200–$299 ($10,400–$15,599) 12 (2.6) 8 (2.4) 34 (4.9) 23 (5.3)
$300–$399 ($15,600–$20,799) 19 (4.1) 12 (3.6) 38 (5.5) 23 (5.3)
$400–$599 ($20,800–$31,199) 42 (9.2) 33 (9.9) 62 (8.9) 33 (7.6)
$600–$799 ($31,200–$41,599) 57 (12.4) 42 (12.6) 61 (8.8) 39 (8.9)
$800–$999 ($41,600–$51,999) 45 (9.8) 31 (9.3) 68 (9.8) 46 (10.6)
$1,000–$1,249 ($52,000–$64,999) 39 (8.5) 32 (9.6) 48 (6.9) 38 (8.7)
$1,250–$1,499 ($65,000–$77,999) 28 (6.1) 22 (6.6) 59 (8.5) 41 (9.4)
$1,500–$1,999 ($78,000–$103,999) 72 (15.7) 50 (15.0) 72 (10.3) 48 (11.0)
$2,000 or more ($104,000 or more) 81 (17.6) 62 (18.6) 108 (15.5) 74 (17.0)
Prefer not to answer 47 (10.2) 32 (9.6) 76 (10.9) 28 (6.4)
Education level, n (%)
Completed junior/lower/primary school 18 (3.6) 17 (4.7) 6 (0.9) 0 (0.0)
Completed senior/high/secondary school 133 (26.9) 98 (26.8) 265 (37.8) 132 (30.0)
Postschool vocational qualification/diploma 147 (29.7) 111 (30.4) 138 (19.7) 94 (21.4)
Undergraduate university degree 131 (26.5) 93 (25.5) 214 (30.5) 159 (36.1)
Postgraduate university degree 66 (13.3) 46 (12.6) 78 (11.1) 55 (12.5)
|
|
tr
|
Variable Australia sample U.S. sample
|
|
th
|
Variable
|
|
th
|
Australia sample
|
|
th
|
U.S. sample
|
|
tr
|
Baseline Follow-up Baseline Follow-up
|
|
td
|
Baseline
|
|
td
|
Follow-up
|
|
td
|
Baseline
|
|
td
|
Follow-up
|
|
tr
|
Participants 495 365 701 440
|
|
td
|
Participants
|
|
td
|
495
|
|
td
|
365
|
|
td
|
701
|
|
td
|
440
|
|
tr
|
Age, M years (SD) 47.09 (17.11) 49.78 (16.89) 45.55 (17.40) 51.77 (16.26)
|
|
td
|
Age, M years (SD)
|
|
td
|
47.09 (17.11)
|
|
td
|
49.78 (16.89)
|
|
td
|
45.55 (17.40)
|
|
td
|
51.77 (16.26)
|
|
tr
|
Gender, n (%)a
|
|
td
|
Gender, n (%)a
|
|
tr
|
Female 252 (51.1) 182 (50.1) 352 (48.9) 205 (46.6)
|
|
td
|
Female
|
|
td
|
252 (51.1)
|
|
td
|
182 (50.1)
|
|
td
|
352 (48.9)
|
|
td
|
205 (46.6)
|
|
tr
|
Male 241 (48.9) 181 (49.9) 341 (50.5) 231 (52.5)
|
|
td
|
Male
|
|
td
|
241 (48.9)
|
|
td
|
181 (49.9)
|
|
td
|
341 (50.5)
|
|
td
|
231 (52.5)
|
|
tr
|
Not specified/prefer not to answer 0 (0.0) 0 (0.0) 4 (0.6) 4 (0.9)
|
|
td
|
Not specified/prefer not to answer
|
|
td
|
0 (0.0)
|
|
td
|
0 (0.0)
|
|
td
|
4 (0.6)
|
|
td
|
4 (0.9)
|
|
tr
|
Employment status, n (%)b
|
|
td
|
Employment status, n (%)b
|
|
tr
|
Currently unemployed/full-time caregiver 231 (46.7) 180 (49.3) 330 (47.3) 216 (49.5)
|
|
td
|
Currently unemployed/full-time caregiver
|
|
td
|
231 (46.7)
|
|
td
|
180 (49.3)
|
|
td
|
330 (47.3)
|
|
td
|
216 (49.5)
|
|
tr
|
Part-time/casual employed 97 (19.6) 65 (17.8) 106 (15.2) 60 (13.8)
|
|
td
|
Part-time/casual employed
|
|
td
|
97 (19.6)
|
|
td
|
65 (17.8)
|
|
td
|
106 (15.2)
|
|
td
|
60 (13.8)
|
|
tr
|
Currently employed full time 140 (28.3) 104 (28.5) 233 (33.4) 147 (33.7)
|
|
td
|
Currently employed full time
|
|
td
|
140 (28.3)
|
|
td
|
104 (28.5)
|
|
td
|
233 (33.4)
|
|
td
|
147 (33.7)
|
|
tr
|
Leave without pay/furloughed 27 (5.5) 16 (4.4) 28 (4.0) 13 (3.0)
|
|
td
|
Leave without pay/furloughed
|
|
td
|
27 (5.5)
|
|
td
|
16 (4.4)
|
|
td
|
28 (4.0)
|
|
td
|
13 (3.0)
|
|
tr
|
Marital status, n (%)c
|
|
td
|
Marital status, n (%)c
|
|
tr
|
Married 184 (37.2) 146 (40.0) 300 (43.0) 224 (51.4)
|
|
td
|
Married
|
|
td
|
184 (37.2)
|
|
td
|
146 (40.0)
|
|
td
|
300 (43.0)
|
|
td
|
224 (51.4)
|
|
tr
|
Widowed 8 (1.6) 7 (1.9) 22 (3.2) 18 (4.1)
|
|
td
|
Widowed
|
|
td
|
8 (1.6)
|
|
td
|
7 (1.9)
|
|
td
|
22 (3.2)
|
|
td
|
18 (4.1)
|
|
tr
|
Separated/divorced 53 (10.7) 39 (10.7) 69 (9.9) 47 (10.8)
|
|
td
|
Separated/divorced
|
|
td
|
53 (10.7)
|
|
td
|
39 (10.7)
|
|
td
|
69 (9.9)
|
|
td
|
47 (10.8)
|
|
tr
|
Never married 160 (32.3) 103 (28.2) 255 (36.6) 126(28.9)
|
|
td
|
Never married
|
|
td
|
160 (32.3)
|
|
td
|
103 (28.2)
|
|
td
|
255 (36.6)
|
|
td
|
126(28.9)
|
|
tr
|
Married de facto 90 (18.2) 70 (19.2) 51 (7.3) 21 (4.8)
|
|
td
|
Married de facto
|
|
td
|
90 (18.2)
|
|
td
|
70 (19.2)
|
|
td
|
51 (7.3)
|
|
td
|
21 (4.8)
|
|
tr
|
Ethnicity, n (%)d
|
|
td
|
Ethnicity, n (%)d
|
|
tr
|
Black 3 (0.6) 1 (0.3) 52 (7.5) 26 (6.0)
|
|
td
|
Black
|
|
td
|
3 (0.6)
|
|
td
|
1 (0.3)
|
|
td
|
52 (7.5)
|
|
td
|
26 (6.0)
|
|
tr
|
Caucasian/White 392 (79.2) 304 (83.3) 566 (81.2) 376 (86.2)
|
|
td
|
Caucasian/White
|
|
td
|
392 (79.2)
|
|
td
|
304 (83.3)
|
|
td
|
566 (81.2)
|
|
td
|
376 (86.2)
|
|
tr
|
Asian (South-East Asia/South Asia) 71 (14.3) 43 (11.8) 39 (5.6) 24 (5.5)
|
|
td
|
Asian (South-East Asia/South Asia)
|
|
td
|
71 (14.3)
|
|
td
|
43 (11.8)
|
|
td
|
39 (5.6)
|
|
td
|
24 (5.5)
|
|
tr
|
Middle-Eastern 6 (1.2) 3 (0.8) 1 (0.1) 0 (0.0)
|
|
td
|
Middle-Eastern
|
|
td
|
6 (1.2)
|
|
td
|
3 (0.8)
|
|
td
|
1 (0.1)
|
|
td
|
0 (0.0)
|
|
tr
|
Other 13 (2.6) 6 (1.6) 27 (3.9) 8 (1.8)
|
|
td
|
Other
|
|
td
|
13 (2.6)
|
|
td
|
6 (1.6)
|
|
td
|
27 (3.9)
|
|
td
|
8 (1.8)
|
|
tr
|
Prefer not to answer 10 (2.0) 8 (2.2) 12 (1.7) 2 (0.5)
|
|
td
|
Prefer not to answer
|
|
td
|
10 (2.0)
|
|
td
|
8 (2.2)
|
|
td
|
12 (1.7)
|
|
td
|
2 (0.5)
|
|
tr
|
Income, n (%)e
|
|
td
|
Income, n (%)e
|
|
tr
|
Zero income 8 (1.7) 4 (1.2) 31 (4.4) 19 (4.4)
|
|
td
|
Zero income
|
|
td
|
8 (1.7)
|
|
td
|
4 (1.2)
|
|
td
|
31 (4.4)
|
|
td
|
19 (4.4)
|
|
tr
|
$1–$199 ($1–$10,399) 9 (2.0) 6 (1.8) 40 (5.7) 24 (5.5)
|
|
td
|
$1–$199 ($1–$10,399)
|
|
td
|
9 (2.0)
|
|
td
|
6 (1.8)
|
|
td
|
40 (5.7)
|
|
td
|
24 (5.5)
|
|
tr
|
$200–$299 ($10,400–$15,599) 12 (2.6) 8 (2.4) 34 (4.9) 23 (5.3)
|
|
td
|
$200–$299 ($10,400–$15,599)
|
|
td
|
12 (2.6)
|
|
td
|
8 (2.4)
|
|
td
|
34 (4.9)
|
|
td
|
23 (5.3)
|
|
tr
|
$300–$399 ($15,600–$20,799) 19 (4.1) 12 (3.6) 38 (5.5) 23 (5.3)
|
|
td
|
$300–$399 ($15,600–$20,799)
|
|
td
|
19 (4.1)
|
|
td
|
12 (3.6)
|
|
td
|
38 (5.5)
|
|
td
|
23 (5.3)
|
|
tr
|
$400–$599 ($20,800–$31,199) 42 (9.2) 33 (9.9) 62 (8.9) 33 (7.6)
|
|
td
|
$400–$599 ($20,800–$31,199)
|
|
td
|
42 (9.2)
|
|
td
|
33 (9.9)
|
|
td
|
62 (8.9)
|
|
td
|
33 (7.6)
|
|
tr
|
$600–$799 ($31,200–$41,599) 57 (12.4) 42 (12.6) 61 (8.8) 39 (8.9)
|
|
td
|
$600–$799 ($31,200–$41,599)
|
|
td
|
57 (12.4)
|
|
td
|
42 (12.6)
|
|
td
|
61 (8.8)
|
|
td
|
39 (8.9)
|
|
tr
|
$800–$999 ($41,600–$51,999) 45 (9.8) 31 (9.3) 68 (9.8) 46 (10.6)
|
|
td
|
$800–$999 ($41,600–$51,999)
|
|
td
|
45 (9.8)
|
|
td
|
31 (9.3)
|
|
td
|
68 (9.8)
|
|
td
|
46 (10.6)
|
|
tr
|
$1,000–$1,249 ($52,000–$64,999) 39 (8.5) 32 (9.6) 48 (6.9) 38 (8.7)
|
|
td
|
$1,000–$1,249 ($52,000–$64,999)
|
|
td
|
39 (8.5)
|
|
td
|
32 (9.6)
|
|
td
|
48 (6.9)
|
|
td
|
38 (8.7)
|
|
tr
|
$1,250–$1,499 ($65,000–$77,999) 28 (6.1) 22 (6.6) 59 (8.5) 41 (9.4)
|
|
td
|
$1,250–$1,499 ($65,000–$77,999)
|
|
td
|
28 (6.1)
|
|
td
|
22 (6.6)
|
|
td
|
59 (8.5)
|
|
td
|
41 (9.4)
|
|
tr
|
$1,500–$1,999 ($78,000–$103,999) 72 (15.7) 50 (15.0) 72 (10.3) 48 (11.0)
|
|
td
|
$1,500–$1,999 ($78,000–$103,999)
|
|
td
|
72 (15.7)
|
|
td
|
50 (15.0)
|
|
td
|
72 (10.3)
|
|
td
|
48 (11.0)
|
|
tr
|
$2,000 or more ($104,000 or more) 81 (17.6) 62 (18.6) 108 (15.5) 74 (17.0)
|
|
td
|
$2,000 or more ($104,000 or more)
|
|
td
|
81 (17.6)
|
|
td
|
62 (18.6)
|
|
td
|
108 (15.5)
|
|
td
|
74 (17.0)
|
|
tr
|
Prefer not to answer 47 (10.2) 32 (9.6) 76 (10.9) 28 (6.4)
|
|
td
|
Prefer not to answer
|
|
td
|
47 (10.2)
|
|
td
|
32 (9.6)
|
|
td
|
76 (10.9)
|
|
td
|
28 (6.4)
|
|
tr
|
Education level, n (%)
|
|
td
|
Education level, n (%)
|
|
tr
|
Completed junior/lower/primary school 18 (3.6) 17 (4.7) 6 (0.9) 0 (0.0)
|
|
td
|
Completed junior/lower/primary school
|
|
td
|
18 (3.6)
|
|
td
|
17 (4.7)
|
|
td
|
6 (0.9)
|
|
td
|
0 (0.0)
|
|
tr
|
Completed senior/high/secondary school 133 (26.9) 98 (26.8) 265 (37.8) 132 (30.0)
|
|
td
|
Completed senior/high/secondary school
|
|
td
|
133 (26.9)
|
|
td
|
98 (26.8)
|
|
td
|
265 (37.8)
|
|
td
|
132 (30.0)
|
|
tr
|
Postschool vocational qualification/diploma 147 (29.7) 111 (30.4) 138 (19.7) 94 (21.4)
|
|
td
|
Postschool vocational qualification/diploma
|
|
td
|
147 (29.7)
|
|
td
|
111 (30.4)
|
|
td
|
138 (19.7)
|
|
td
|
94 (21.4)
|
|
tr
|
Undergraduate university degree 131 (26.5) 93 (25.5) 214 (30.5) 159 (36.1)
|
|
td
|
Undergraduate university degree
|
|
td
|
131 (26.5)
|
|
td
|
93 (25.5)
|
|
td
|
214 (30.5)
|
|
td
|
159 (36.1)
|
|
tr
|
Postgraduate university degree 66 (13.3) 46 (12.6) 78 (11.1) 55 (12.5)
|
|
td
|
Postgraduate university degree
|
|
td
|
66 (13.3)
|
|
td
|
46 (12.6)
|
|
td
|
78 (11.1)
|
|
td
|
55 (12.5)
|
|
table-wrap-foot
|
SD standard deviation.
aTwo participants in the Australian sample did not report their gender and four participants in the U.S. sample did not report their gender.
bFour participants in the U.S. sample did not report their employment status.
cFour participants in the U.S. sample did not report their marital status.
dFour participants in the U.S. sample did not report their ethnicity.
eThirty-one participants in the Australian sample did not report their income and four participants in the U.S. sample did not report their income.
|
|
footnote
|
SD standard deviation.
|
|
p
|
SD standard deviation.
|
|
footnote
|
aTwo participants in the Australian sample did not report their gender and four participants in the U.S. sample did not report their gender.
|
|
p
|
aTwo participants in the Australian sample did not report their gender and four participants in the U.S. sample did not report their gender.
|
|
footnote
|
bFour participants in the U.S. sample did not report their employment status.
|
|
p
|
bFour participants in the U.S. sample did not report their employment status.
|
|
footnote
|
cFour participants in the U.S. sample did not report their marital status.
|
|
p
|
cFour participants in the U.S. sample did not report their marital status.
|
|
footnote
|
dFour participants in the U.S. sample did not report their ethnicity.
|
|
p
|
dFour participants in the U.S. sample did not report their ethnicity.
|
|
footnote
|
eThirty-one participants in the Australian sample did not report their income and four participants in the U.S. sample did not report their income.
|
|
p
|
eThirty-one participants in the Australian sample did not report their income and four participants in the U.S. sample did not report their income.
|
|
sec
|
Design and Procedure
The study adopted a prospective correlational design with self-report measures of social cognition constructs from the proposed integrated model, intention, and past social distancing behavior administered at an initial data collection occasion in a survey administered using the Qualtrics online survey tool. Social cognition measures included the theory of planned behavior (attitude, subjective norm, and PBC), moral norm, anticipated regret, action planning, and habit constructs. Participants were informed that they were participating in a survey on their social distancing behavior and provided with information outlining study requirements. They were required to provide informed consent before proceeding with the survey. Participants were also provided with instructions on how to complete study measures and a definition of the target behavior: “The following survey will ask about your beliefs and attitudes about ‘social distancing’. What do we mean by social distancing? Social distancing (also known as ‘physical distancing’) is deliberately increasing the physical space between people to avoid spreading illness. The World Health Organization and other world leading health authorities suggest that you should maintain at least a 1–2 meter (3–6 feet) distance from other people to lessen the chances of getting infected with COVID-19. When answering the questions in this survey, think about your social distancing behavior (i.e., maintaining at least a 1–2 m (3–6 ft) distance from other people).” One week later, participants were re-contacted by the panel company and asked to self-report their habit and social distancing behavior over the previous week using the same measures used at the initial data collection occasion. Participants received a fixed sum of money for their participation based on expected completion time consistent with the panel company’s published rates. Approval for study procedures was granted prior to data collection from the Griffith University Human Research Ethics Committee.
|
|
title
|
Design and Procedure
|
|
p
|
The study adopted a prospective correlational design with self-report measures of social cognition constructs from the proposed integrated model, intention, and past social distancing behavior administered at an initial data collection occasion in a survey administered using the Qualtrics online survey tool. Social cognition measures included the theory of planned behavior (attitude, subjective norm, and PBC), moral norm, anticipated regret, action planning, and habit constructs. Participants were informed that they were participating in a survey on their social distancing behavior and provided with information outlining study requirements. They were required to provide informed consent before proceeding with the survey. Participants were also provided with instructions on how to complete study measures and a definition of the target behavior: “The following survey will ask about your beliefs and attitudes about ‘social distancing’. What do we mean by social distancing? Social distancing (also known as ‘physical distancing’) is deliberately increasing the physical space between people to avoid spreading illness. The World Health Organization and other world leading health authorities suggest that you should maintain at least a 1–2 meter (3–6 feet) distance from other people to lessen the chances of getting infected with COVID-19. When answering the questions in this survey, think about your social distancing behavior (i.e., maintaining at least a 1–2 m (3–6 ft) distance from other people).” One week later, participants were re-contacted by the panel company and asked to self-report their habit and social distancing behavior over the previous week using the same measures used at the initial data collection occasion. Participants received a fixed sum of money for their participation based on expected completion time consistent with the panel company’s published rates. Approval for study procedures was granted prior to data collection from the Griffith University Human Research Ethics Committee.
|
|
sec
|
Measures
Study measures were multi-item self-report measures of constructs based on published guidelines and measures used in previous studies [13, 29, 30]. Participants provided their responses on scales with seven-point response options. Complete study measures are provided in Supplementary Appendix A.
Social cognition constructs
Multi-item measures of attitude, subjective norm, PBC, moral norm, anticipated regret, and action planning were developed according to published guidelines [13, 29]. Each measure made explicit reference to the target behavior of social distancing, and participants were reminded of the definition of social distancing before completing the measures.
Intention
Participants’ intention to participate in social distancing behavior over the next week was measured using a scale developed according to published guidelines [31].
Habit
Habit was measured at both time points using the behavioral automaticity items of Verplanken and Orbell’s self-report habit index [25]. The measure measures individuals’ reflections on the extent to which the behavior is experienced as automatic and enacted without thought.
Past behavior and behavior
Participants self-reported their participation in social distancing behavior to minimize the transmission of the SARS-CoV-2 virus that causes COVID-19. The measure comprised two items prompting participants to report their frequency of social distancing behavior in the previous week. This is based on previously used self-report behavioral measures that have demonstrated concurrent validity with non-self-report measures in other behavioral contexts [32].
Demographic variables
Participants self-reported their age in years, gender, employment status (currently unemployed/full-time caregiver, currently full-time employed, part-time employed, or on leave without pay/furloughed), marital status (married, widowed, separated/divorced, never married, or in a de factor relationship), annual household income stratified by 11 income levels based on Australian and U.S. national averages, highest level of formal education (completed junior/lower/primary school, completed senior/high/secondary school, postschool vocational qualification/diploma, further education diploma, undergraduate university degree, or postgraduate university degree), and ethnicity (Black, Caucasian/White, Asian, or Middle-eastern). Binary income (low income vs. middle/high income), highest education level (completed school education only vs. completed postschool education), and ethnicity (Caucasian/White vs. non-White) variables were computed for use in subsequent analyses.
|
|
title
|
Measures
|
|
p
|
Study measures were multi-item self-report measures of constructs based on published guidelines and measures used in previous studies [13, 29, 30]. Participants provided their responses on scales with seven-point response options. Complete study measures are provided in Supplementary Appendix A.
|
|
sec
|
Social cognition constructs
Multi-item measures of attitude, subjective norm, PBC, moral norm, anticipated regret, and action planning were developed according to published guidelines [13, 29]. Each measure made explicit reference to the target behavior of social distancing, and participants were reminded of the definition of social distancing before completing the measures.
|
|
title
|
Social cognition constructs
|
|
p
|
Multi-item measures of attitude, subjective norm, PBC, moral norm, anticipated regret, and action planning were developed according to published guidelines [13, 29]. Each measure made explicit reference to the target behavior of social distancing, and participants were reminded of the definition of social distancing before completing the measures.
|
|
sec
|
Intention
Participants’ intention to participate in social distancing behavior over the next week was measured using a scale developed according to published guidelines [31].
|
|
title
|
Intention
|
|
p
|
Participants’ intention to participate in social distancing behavior over the next week was measured using a scale developed according to published guidelines [31].
|
|
sec
|
Habit
Habit was measured at both time points using the behavioral automaticity items of Verplanken and Orbell’s self-report habit index [25]. The measure measures individuals’ reflections on the extent to which the behavior is experienced as automatic and enacted without thought.
|
|
title
|
Habit
|
|
p
|
Habit was measured at both time points using the behavioral automaticity items of Verplanken and Orbell’s self-report habit index [25]. The measure measures individuals’ reflections on the extent to which the behavior is experienced as automatic and enacted without thought.
|
|
sec
|
Past behavior and behavior
Participants self-reported their participation in social distancing behavior to minimize the transmission of the SARS-CoV-2 virus that causes COVID-19. The measure comprised two items prompting participants to report their frequency of social distancing behavior in the previous week. This is based on previously used self-report behavioral measures that have demonstrated concurrent validity with non-self-report measures in other behavioral contexts [32].
|
|
title
|
Past behavior and behavior
|
|
p
|
Participants self-reported their participation in social distancing behavior to minimize the transmission of the SARS-CoV-2 virus that causes COVID-19. The measure comprised two items prompting participants to report their frequency of social distancing behavior in the previous week. This is based on previously used self-report behavioral measures that have demonstrated concurrent validity with non-self-report measures in other behavioral contexts [32].
|
|
sec
|
Demographic variables
Participants self-reported their age in years, gender, employment status (currently unemployed/full-time caregiver, currently full-time employed, part-time employed, or on leave without pay/furloughed), marital status (married, widowed, separated/divorced, never married, or in a de factor relationship), annual household income stratified by 11 income levels based on Australian and U.S. national averages, highest level of formal education (completed junior/lower/primary school, completed senior/high/secondary school, postschool vocational qualification/diploma, further education diploma, undergraduate university degree, or postgraduate university degree), and ethnicity (Black, Caucasian/White, Asian, or Middle-eastern). Binary income (low income vs. middle/high income), highest education level (completed school education only vs. completed postschool education), and ethnicity (Caucasian/White vs. non-White) variables were computed for use in subsequent analyses.
|
|
title
|
Demographic variables
|
|
p
|
Participants self-reported their age in years, gender, employment status (currently unemployed/full-time caregiver, currently full-time employed, part-time employed, or on leave without pay/furloughed), marital status (married, widowed, separated/divorced, never married, or in a de factor relationship), annual household income stratified by 11 income levels based on Australian and U.S. national averages, highest level of formal education (completed junior/lower/primary school, completed senior/high/secondary school, postschool vocational qualification/diploma, further education diploma, undergraduate university degree, or postgraduate university degree), and ethnicity (Black, Caucasian/White, Asian, or Middle-eastern). Binary income (low income vs. middle/high income), highest education level (completed school education only vs. completed postschool education), and ethnicity (Caucasian/White vs. non-White) variables were computed for use in subsequent analyses.
|
|
sec
|
Data Analysis
Hypothesized relations among the integrated model constructs were tested in the Australian and U.S. samples separately using variance-based structural equation modeling implemented in the WARP 7.0 analysis package [33]. Model parameters and standard errors (SEs) were computed using the “Stable3” estimation method, which has been shown to provide the most precise parameter estimates in complex structural models in smaller samples and outperforms bootstrapping methods in simulation studies [33]. Simulation studies have also shown this method to provide more consistent and precise estimates in data containing outliers, which may inflate SEs and lead to abnormally high p-values [33]. Two models were estimated in each sample: a model testing predictions of the proposed integrated model with the binary demographic variables also included as covariates (Model 1; Fig. 1, upper panel) and a model that included effects of past social distancing behavior (Model 2; Fig. 1, lower panel). All constructs were latent variables indicated by single or multiple items. There were no missing data for the social cognition and self-reported behavioral variables. There were a few instances of missing data for the demographic variables ranging from 0.5% to 8.8% in the Australia sample, and 0.9% to 6.4% in the U.S. sample. Missing data are reported in Supplementary Appendix B. Missing data were imputed using stochastic hierarchical regression [33].
The analysis afforded a number of analyses to evaluate the adequacy of measures used to indicate the latent variables in the model. Construct validity of the latent factors for the social cognition, intention, and behavioral variables was established using the normalized factor pattern loadings after oblique rotation and Kaiser normalization [33] and the average variance extracted (AVE), which should approach or exceed .700 and .500, respectively. Internal consistency of the factors was estimated using omega reliability coefficients (ω) and composite reliability coefficients (ρ), which should exceed .700 and ideally approach .900. We also conducted tests of the discriminant validity of the constructs in the model. Discriminant validity was supported when the square root of the AVE for each latent variable exceeded its correlation with other latent variables.
Adequacy of the proposed model in describing the data was established using the goodness-of-fit (GoF) index with values of .100, .250, and .360 corresponding to small, medium, and large effect sizes. Further information on model quality was provided by the average path coefficient and average R2 coefficient. These indices summarize the average parameter estimates of relations in the model and the amount of variance explained in each dependent variable, respectively, and should be statistically significant for a good-quality model. In addition, an overall GoF index is provided by the average block variance inflation factor for model parameters and the average full collinearity variance inflation factor, which should be equal to or lower than 3.3 for well-fitting models. These indices indicate the extent to which latent variables in the model overlap and contribute to model multicollinearity. They, therefore, provide an indication as to the uniqueness of the existing latent variables in the model. Four further indices were also used to evaluate model quality: the Simpson’s paradox ratio (SPR), R2 contribution ratio (R2CR), the statistical suppression ratio (SSR), and the nonlinear bivariate causality direction ratio (NLBCDR). The SPR indicates whether the model is free from incidences of Simpson’s paradox (i.e., when the path coefficient and the correlation associated with a latent variable have opposite signs), indicating a causality problem. The SPR should exceed .700 and ideally approach 1.000. The R2CR and SSR provide indication of the extent to which models are free from instances of negative R2 contributions and statistical suppression. The R2CR and SSR should exceed .900 and .700, respectively. The NLBCDR provides an estimate of the extent to which the proposed “causal” associations in the proposed model are more tenable than those in the opposite direction and provide an initial indicator of support for the hypothesized directions of the causal links in the proposed model compared to if the proposed direction were reversed. The NLBCDR should exceed .700 for high-quality models. Kock [33] provides further technical details on model fit and quality indices.
Model effects were estimated using standardized path coefficients with confidence intervals (CIs) and test statistics. Effect sizes were estimated using a variant of Cohen’s f-square coefficient and represent the individual contribution of the predictor variable to the R2 coefficients of the criterion latent variable. Values of .02, .15, and .35 represent small, medium, and large effect sizes, respectively. Differences in the path coefficients in the models across the samples were tested using multiple-group analysis using the Satterthwaite method with two-tailed significance tests.
We also tested whether the inclusion of participants that were never under a “shelter-in-place” order, or had the “shelter-in-place” order lifted during the study, affected predicted relations in the models. The small numbers of participants that were, at some point, not subjected to “shelter-in-place” orders meant we could not conduct a formal moderator analysis, so we conducted a sensitivity analysis testing whether effects in the models differed if data from these participants were excluded. Models excluding and including past behavior were estimated in samples excluding participants who were never subject to a “shelter-in place” order, and in the sample that were never subject to an order, or who had the order lifted at some stage during the study. Formal comparisons of parameter estimates in these models with those from the full sample were made using the Satterthwaite method. Data files, analysis scripts, and output files for all analyses are available online: https://osf.io/x9tms/.
|
|
title
|
Data Analysis
|
|
p
|
Hypothesized relations among the integrated model constructs were tested in the Australian and U.S. samples separately using variance-based structural equation modeling implemented in the WARP 7.0 analysis package [33]. Model parameters and standard errors (SEs) were computed using the “Stable3” estimation method, which has been shown to provide the most precise parameter estimates in complex structural models in smaller samples and outperforms bootstrapping methods in simulation studies [33]. Simulation studies have also shown this method to provide more consistent and precise estimates in data containing outliers, which may inflate SEs and lead to abnormally high p-values [33]. Two models were estimated in each sample: a model testing predictions of the proposed integrated model with the binary demographic variables also included as covariates (Model 1; Fig. 1, upper panel) and a model that included effects of past social distancing behavior (Model 2; Fig. 1, lower panel). All constructs were latent variables indicated by single or multiple items. There were no missing data for the social cognition and self-reported behavioral variables. There were a few instances of missing data for the demographic variables ranging from 0.5% to 8.8% in the Australia sample, and 0.9% to 6.4% in the U.S. sample. Missing data are reported in Supplementary Appendix B. Missing data were imputed using stochastic hierarchical regression [33].
|
|
p
|
The analysis afforded a number of analyses to evaluate the adequacy of measures used to indicate the latent variables in the model. Construct validity of the latent factors for the social cognition, intention, and behavioral variables was established using the normalized factor pattern loadings after oblique rotation and Kaiser normalization [33] and the average variance extracted (AVE), which should approach or exceed .700 and .500, respectively. Internal consistency of the factors was estimated using omega reliability coefficients (ω) and composite reliability coefficients (ρ), which should exceed .700 and ideally approach .900. We also conducted tests of the discriminant validity of the constructs in the model. Discriminant validity was supported when the square root of the AVE for each latent variable exceeded its correlation with other latent variables.
|
|
p
|
Adequacy of the proposed model in describing the data was established using the goodness-of-fit (GoF) index with values of .100, .250, and .360 corresponding to small, medium, and large effect sizes. Further information on model quality was provided by the average path coefficient and average R2 coefficient. These indices summarize the average parameter estimates of relations in the model and the amount of variance explained in each dependent variable, respectively, and should be statistically significant for a good-quality model. In addition, an overall GoF index is provided by the average block variance inflation factor for model parameters and the average full collinearity variance inflation factor, which should be equal to or lower than 3.3 for well-fitting models. These indices indicate the extent to which latent variables in the model overlap and contribute to model multicollinearity. They, therefore, provide an indication as to the uniqueness of the existing latent variables in the model. Four further indices were also used to evaluate model quality: the Simpson’s paradox ratio (SPR), R2 contribution ratio (R2CR), the statistical suppression ratio (SSR), and the nonlinear bivariate causality direction ratio (NLBCDR). The SPR indicates whether the model is free from incidences of Simpson’s paradox (i.e., when the path coefficient and the correlation associated with a latent variable have opposite signs), indicating a causality problem. The SPR should exceed .700 and ideally approach 1.000. The R2CR and SSR provide indication of the extent to which models are free from instances of negative R2 contributions and statistical suppression. The R2CR and SSR should exceed .900 and .700, respectively. The NLBCDR provides an estimate of the extent to which the proposed “causal” associations in the proposed model are more tenable than those in the opposite direction and provide an initial indicator of support for the hypothesized directions of the causal links in the proposed model compared to if the proposed direction were reversed. The NLBCDR should exceed .700 for high-quality models. Kock [33] provides further technical details on model fit and quality indices.
|
|
p
|
Model effects were estimated using standardized path coefficients with confidence intervals (CIs) and test statistics. Effect sizes were estimated using a variant of Cohen’s f-square coefficient and represent the individual contribution of the predictor variable to the R2 coefficients of the criterion latent variable. Values of .02, .15, and .35 represent small, medium, and large effect sizes, respectively. Differences in the path coefficients in the models across the samples were tested using multiple-group analysis using the Satterthwaite method with two-tailed significance tests.
|
|
p
|
We also tested whether the inclusion of participants that were never under a “shelter-in-place” order, or had the “shelter-in-place” order lifted during the study, affected predicted relations in the models. The small numbers of participants that were, at some point, not subjected to “shelter-in-place” orders meant we could not conduct a formal moderator analysis, so we conducted a sensitivity analysis testing whether effects in the models differed if data from these participants were excluded. Models excluding and including past behavior were estimated in samples excluding participants who were never subject to a “shelter-in place” order, and in the sample that were never subject to an order, or who had the order lifted at some stage during the study. Formal comparisons of parameter estimates in these models with those from the full sample were made using the Satterthwaite method. Data files, analysis scripts, and output files for all analyses are available online: https://osf.io/x9tms/.
|
|
sec
|
Results
Participants
Attrition across the two data collection occasions resulted in final sample sizes of 365 (M age = 49.78, standard deviation [SD] = 16.89; 50.1% female; retention rate 73.73%) and 440 (M age = 51.77, SD = 16.26; 46.6% female; retention rate = 62.77%) participants in the Australian and U.S. samples, respectively. Sample characteristics at follow-up are presented in Table 1. Attrition analyses in the Australian sample revealed that participants lost to attrition were younger and were more likely to be non-White. However, there were no differences in proportion of gender, income, and education level. A MANOVA with the social cognition constructs and past behavior as dependent variables and attrition status (lost to attrition vs. included at follow-up) revealed no differences (Wilks’ Lambda = 0.969, F(1,9) = 1.70, p = .077, partial η 2 = .031). Attrition analyses in the U.S. sample also indicated that participants lost to attrition were younger, and more likely to be male, non-White, and lower educated and have low income, than those remaining in the study at follow-up. The MANOVA testing for differences on social cognition constructs and past behavior among participants lost to attrition and those included at follow-up revealed statistically significant differences (Wilks’ Lambda = 0.969, F(1,9) = 2.40, p = .010, partial η 2 = .031). Follow-up tests revealed that mean values for past behavior, attitude, subjective norm, intention, moral norm, and habit with respect to social distancing were significantly lower among participants lost to attrition compared to those retained at follow-up. However, effect sizes for these differences were small (ds < .25). Details of attrition analyses are presented in Supplementary Appendix B.
Preliminary Analyses
Factor loadings and AVE values exceeded recommended .700 and .500 cutoff values in all cases. Omega reliability coefficients, interitem correlations (for two-item scales), and composite reliabilities indicated good internal consistency of scales used. Latent variable correlations among social cognition constructs were all statistically significant. Correlations among the majority of constructs in the Australia sample were small-to-medium in size (r range = .161 to .564), with some smaller correlations involving the subjective and moral norms constructs and habit (r range = .094 to .118). Correlations were small-to-medium in size in the U.S. sample (r range = .266 to .620). Square roots of the AVE for each latent variable exceeded the correlation of that variable with all other latent variables supporting discriminant validity. Skewness and kurtosis estimates indicated many of the variables were not normally distributed, justifying the use of the variance-based structural equation modeling, which is a “distribution-free” analytic method. Factor loadings, reliability coefficients, and distribution statistics are presented in Supplementary Appendix C, and latent variable correlations for model variables in both are presented in Supplementary Appendix D.
Structural Equation Models
Single-sample analyses
GoF and quality indices of the structural equation models are presented in Table 2. The models that excluded (Model 1) and included (Model 2) past behavior exhibited adequate fit and quality indices in both the Australian and U.S. samples. Standardized parameter estimates for the proposed direct effects for each model in the Australian and U.S. samples are presented in Fig. 1. Full parameter estimates for models in both samples are presented in Supplementary Appendix E. Parameter estimates, CIs, and effect sizes for the indirect effects of the models in both samples are summarized in Table 3.
Table 2. Model quality and GoF statistics for the structural equation models of the integrated model in the Australian and U.S. samples and multigroup model
Sample Model APC AR2 AVIF AFVIF GoF SPR R2CR SSR NLBCDR
Australia 1 .104** .177*** 1.177 1.561 .391 .841 .977 .889 .873
2 .116*** .338*** 1.222 1.904 .543 .819 .991 .931 .785
USA 1 .098** .192*** 1.187 1.823 .410 .889 .995 .825 .754
2 .116*** .338** 1.222 1.904 .543 .819 .991 .931 .785
MS 1 .100*** .182*** 1.159 1.704 .398 .905 .995 .794 .817
2 .113*** .300*** 1.186 1.760 .511 .931 .997 .917 .840
Model 1 = model excluding past behavior; Model 2 = model including past behavior.
AFVIF average full collinearity variance inflation factor; APC average path coefficient; AR2 average R2; AVIF average block variance inflation factor; GoF Tenenhaus’s goodness-of-fit index; MS multiple sample analysis; NLBCDR nonlinear bivariate causality direction ratio; R2CR R2 contribution ratio; SPR Sympson’s paradox ratio; SSR statistical suppression ratio.
*p < .05, **p < .01, ***p < .001
Table 3. Standardized parameter estimates for indirect effects for the structural equation model of the integrated model in the Australian and U.S. samples
Effect Model excluding past behavior Model including past behavior
β p 95% CI ES Β p 95% CI ES
LB UB LB UB
Australian sample
Indirect effects
Att.→Int.→Beh. .011 .359 −.052 .074 .003 .004 .444 −.059 .067 .001
SN→Int.→Beh. .042 .094 −.021 .105 .016 .016 .312 −.047 .079 .006
MN→Int.→Beh. .068 .016 .005 .131 .024 .028 .192 −.035 .091 .010
AR→Int.→Beh. .011 .356 −.052 .074 .003 .003 .457 −.060 .066 .001
PBC→Int.→Beh. .040 .101 −.023 .103 .011 .016 .307 −.047 .079 .005
Int.→AP→Beh. .040 .106 −.023 .103 .014 .011 .365 −.052 .074 .004
Hab. (T1).→Hab. (T2)→Beh. .102 <.001 .041 .163 .016 .078 .007 .017 .139 .013
PB→Hab.→Beh. – – – – – .021 .214 −.030 .072 .011
PB→Beh.a – – – – – .081 .034 −.007 .169 .042
Total effectsb
Int.→Beh. .220 <.001 .134 .306 .081 .090 .022 .004 .176 .033
PBC→Beh. .126 <.001 .040 .212 .036 .055 .110 −.033 .143 .016
Hab. (T1)→Beh. .096 .016 .010 .182 .015 .076 .044 −.012 .164 .012
PB→Beh. – – – – – .494 <.001 .412 .576 .258
U.S. sample
Indirect effects
Att.→Int.→Beh. <.001 .495 −.052 .054 <.001 .004 .443 −.049 .057 .001
SN→Int.→Beh. .072 .003 .019 .125 .029 .023 .190 −.030 .076 .009
MN→Int.→Beh. .102 <.001 .051 .153 .044 .040 .067 −.013 .093 .017
AR→Int.→Beh. .023 .192 −.030 .076 .011 .001 .478 −.052 .054 .001
PBC→Int.→Beh. .088 <.001 .037 .139 .025 .038 .079 −.015 .091 .011
Int.→AP→Beh. .061 .011 .008 .114 .029 .004 .441 −.049 .057 .002
Hab. (T1).→Hab. (T2)→Beh. .212 <.001 .161 .263 .075 .166 <.001 .115 .217 .059
PB→Hab.→Beh. – – – – – .068 <.001 .025 .111 .043
PB→Beh.a – – – – – .178 <.001 .105 .251 .112
Total effectsb
Int.→Beh. .377 <.001 .306 .448 .177 .142 <.001 .069 .215 .066
PBC→Beh. .146 <.001 .073 .219 .042 .074 .024 .001 .147 .021
Hab. (T1)→Beh. .242 <.001 .169 .315 .086 .171 <.001 .098 .244 .061
PB→Beh. – – −.052 .074 – .673 <.001 .604 .742 .423
aSum of indirect effects of past behavior on behavior through all model constructs.
bTotal effect comprising sums of all indirect effects through model constructs plus the direct effect.
β standardized parameter estimate; 95% CI 95% confidence interval of standardized parameter estimate; AP action planning; AR anticipated regret; Att. attitude; Beh. behavior; ES effect size of the standardized parameter estimate; Hab. (T1) self-reported habit measured at baseline (T1); Hab. (T2) self-reported habit measured at follow-up (T2); Int. intention; LB lower bound of 95% CI; MN moral norm; PB past behavior; PBC perceived behavioral control; SN subjective norm; UB upper bound of 95% CI. Focusing on the model excluding past behavior (Model 1), intention, action planning, and habit at follow-up were statistically significant direct predictors of social distancing behavior, with effect size for intention and habit generally larger in the U.S. sample. PBC directly predicted behavior in the Australian sample only, also with a small effect size. Intention predicted action planning in both samples with large effect sizes. Subjective norm, moral norm, and PBC predicted intention in both samples, with small-to-medium effect sizes, but effects of attitude were not significant. There was a small effect of anticipated regret on intention in the U.S. sample only. Habit at baseline predicted habit at follow-up in both samples, with large effect sizes. There was also a small-sized effect of habit at baseline on intention in the U.S. sample only. Overall, the model accounted for significant variance in social distancing behavior (Australian sample, R2 = .198; U.S. sample, R2 = .361), intentions (Australian sample, R2 = .571; U.S. sample, R2 = .623), and habit at follow-up (Australian sample, R2 = .416; U.S. sample, R2 = .486). Intentions (Australian sample, R2 = .066; U.S. sample, R2 = .148), action planning (Australian sample, R2 = .029; U.S. sample, R2 = .044), and habit at follow-up (Australian sample R2 = .041; U.S. sample, R2 = .129) each accounted for substantive variance in behavior. Action planning significantly moderated the intention–behavior relationship in the Australian sample only. While the effect was not in the predicted direction, probing the interaction revealed that the intention–behavior relationship increased as the level of planning increased, consistent with theory. However, the intention–behavior relationship is more likely to be smaller at lower levels of planning, and it seems that planning makes less of a difference when the intention–behavior relationship is large. A plot of the interaction effect is presented in Supplementary Appendix F.
Turning to the indirect effects, there were significant indirect effects of subjective norm, moral norm, and PBC on social distancing behavior mediated by intention in the U.S. sample. By contrast, only the indirect effect of moral norm on behavior through intention was significant in the Australian sample. The smaller indirect effects in the Australian sample is principally due to the significantly smaller effect size for the intention–behavior relationship in this sample compared to the U.S. sample. Habit at baseline predicted behavior through habit at follow-up in both samples. Effect sizes in all cases were small. There were significant total effects of intention, PBC, and habit at baseline on behavior, with effect sizes larger in the U.S. sample than in the Australian sample.
For the model including past behavior, significant effects of past behavior on all model constructs were observed in both samples with effect sizes ranging from small to large. The effects of past behavior on social distancing behavior were particularly large. Inclusion of past behavior led to an attenuation of model effects in both samples. Specifically, the effects of intention and habit at follow-up on behavior were reduced but remained statistically significant in both samples with small effect sizes. In addition, effects of subjective norm, moral norm, and PBC on intention, and the effect of intention on action planning, remained statistically significant in both samples, with small-to-medium effect sizes. The effect of habit at baseline on habit at follow-up was statistically significant in both samples, with large effect sizes. Variance explained in social distancing behavior increased substantially with the inclusion of past behavior, with only modest changes in explained variance in intentions (Australian sample R2 = .598; U.S. sample, R2 = .702) and habit at follow-up (Australian sample R2 = .416; U.S. sample, R2 = .486). Specifically, intentions (Australian sample, R2 = .029; U.S. sample, R2 = .065), past behavior (Australian sample, R2 = .216; U.S. sample, R2 = .311), and habit at follow-up (Australian sample R2 = .031; U.S. sample, R2 = .101) each accounted for substantive variance in behavior.
Turning to indirect effects, we found significant indirect effects of habit at baseline on behavior mediated by habit at follow-up in both samples with small effect sizes. There were also significant total effects of intention and habit at baseline on behavior in both samples, and of PBC on behavior for the U.S. sample, with small effect sizes. There were significant total indirect and total effects of past behavior on behavior in both samples, with large effect sizes. There was a small-sized indirect effect of past behavior on behavior mediated by habit at both time points in the U.S. sample, but the effect was not significant in the Australian sample.
Multisample analyses
Multisample analyses permitted for tests of difference in parameter estimates for each model across the Australian and U.S. samples. For the model excluding past behavior (Model 1), only effects of intention on habit at baseline, habit at follow-up on social distancing behavior, and intention on action planning differed across samples. These effects were significantly larger in the U.S. sample. Some effects with observed differences across samples, such as the effect of habit at baseline on intention or the moderator effect of planning on the intention–behavior relationship, did not differ significantly across samples. For the model including effects of past behavior (Model 2), multisample analysis revealed no differences in effect size across samples, indicating that the attenuating effect of past behavior on model effects also had the effect of eliminating the few differences in model effects across samples. Full details of the multiple-group analysis are presented in Supplementary Appendix G.
Sensitivity analyses
We re-estimated both models in samples excluding participants who were never subject to a “shelter-in place” order, and in the sample that were never subject to an order, or who had the order lifted at some stage during the study. Comparisons of parameter estimates in these models with those from the models estimated in the full sample revealed no significant differences in any of the model parameters. Results are reported in Supplementary Appendices H and I.
|
|
title
|
Results
|
|
sec
|
Participants
Attrition across the two data collection occasions resulted in final sample sizes of 365 (M age = 49.78, standard deviation [SD] = 16.89; 50.1% female; retention rate 73.73%) and 440 (M age = 51.77, SD = 16.26; 46.6% female; retention rate = 62.77%) participants in the Australian and U.S. samples, respectively. Sample characteristics at follow-up are presented in Table 1. Attrition analyses in the Australian sample revealed that participants lost to attrition were younger and were more likely to be non-White. However, there were no differences in proportion of gender, income, and education level. A MANOVA with the social cognition constructs and past behavior as dependent variables and attrition status (lost to attrition vs. included at follow-up) revealed no differences (Wilks’ Lambda = 0.969, F(1,9) = 1.70, p = .077, partial η 2 = .031). Attrition analyses in the U.S. sample also indicated that participants lost to attrition were younger, and more likely to be male, non-White, and lower educated and have low income, than those remaining in the study at follow-up. The MANOVA testing for differences on social cognition constructs and past behavior among participants lost to attrition and those included at follow-up revealed statistically significant differences (Wilks’ Lambda = 0.969, F(1,9) = 2.40, p = .010, partial η 2 = .031). Follow-up tests revealed that mean values for past behavior, attitude, subjective norm, intention, moral norm, and habit with respect to social distancing were significantly lower among participants lost to attrition compared to those retained at follow-up. However, effect sizes for these differences were small (ds < .25). Details of attrition analyses are presented in Supplementary Appendix B.
|
|
title
|
Participants
|
|
p
|
Attrition across the two data collection occasions resulted in final sample sizes of 365 (M age = 49.78, standard deviation [SD] = 16.89; 50.1% female; retention rate 73.73%) and 440 (M age = 51.77, SD = 16.26; 46.6% female; retention rate = 62.77%) participants in the Australian and U.S. samples, respectively. Sample characteristics at follow-up are presented in Table 1. Attrition analyses in the Australian sample revealed that participants lost to attrition were younger and were more likely to be non-White. However, there were no differences in proportion of gender, income, and education level. A MANOVA with the social cognition constructs and past behavior as dependent variables and attrition status (lost to attrition vs. included at follow-up) revealed no differences (Wilks’ Lambda = 0.969, F(1,9) = 1.70, p = .077, partial η 2 = .031). Attrition analyses in the U.S. sample also indicated that participants lost to attrition were younger, and more likely to be male, non-White, and lower educated and have low income, than those remaining in the study at follow-up. The MANOVA testing for differences on social cognition constructs and past behavior among participants lost to attrition and those included at follow-up revealed statistically significant differences (Wilks’ Lambda = 0.969, F(1,9) = 2.40, p = .010, partial η 2 = .031). Follow-up tests revealed that mean values for past behavior, attitude, subjective norm, intention, moral norm, and habit with respect to social distancing were significantly lower among participants lost to attrition compared to those retained at follow-up. However, effect sizes for these differences were small (ds < .25). Details of attrition analyses are presented in Supplementary Appendix B.
|
|
sec
|
Preliminary Analyses
Factor loadings and AVE values exceeded recommended .700 and .500 cutoff values in all cases. Omega reliability coefficients, interitem correlations (for two-item scales), and composite reliabilities indicated good internal consistency of scales used. Latent variable correlations among social cognition constructs were all statistically significant. Correlations among the majority of constructs in the Australia sample were small-to-medium in size (r range = .161 to .564), with some smaller correlations involving the subjective and moral norms constructs and habit (r range = .094 to .118). Correlations were small-to-medium in size in the U.S. sample (r range = .266 to .620). Square roots of the AVE for each latent variable exceeded the correlation of that variable with all other latent variables supporting discriminant validity. Skewness and kurtosis estimates indicated many of the variables were not normally distributed, justifying the use of the variance-based structural equation modeling, which is a “distribution-free” analytic method. Factor loadings, reliability coefficients, and distribution statistics are presented in Supplementary Appendix C, and latent variable correlations for model variables in both are presented in Supplementary Appendix D.
|
|
title
|
Preliminary Analyses
|
|
p
|
Factor loadings and AVE values exceeded recommended .700 and .500 cutoff values in all cases. Omega reliability coefficients, interitem correlations (for two-item scales), and composite reliabilities indicated good internal consistency of scales used. Latent variable correlations among social cognition constructs were all statistically significant. Correlations among the majority of constructs in the Australia sample were small-to-medium in size (r range = .161 to .564), with some smaller correlations involving the subjective and moral norms constructs and habit (r range = .094 to .118). Correlations were small-to-medium in size in the U.S. sample (r range = .266 to .620). Square roots of the AVE for each latent variable exceeded the correlation of that variable with all other latent variables supporting discriminant validity. Skewness and kurtosis estimates indicated many of the variables were not normally distributed, justifying the use of the variance-based structural equation modeling, which is a “distribution-free” analytic method. Factor loadings, reliability coefficients, and distribution statistics are presented in Supplementary Appendix C, and latent variable correlations for model variables in both are presented in Supplementary Appendix D.
|
|
sec
|
Structural Equation Models
Single-sample analyses
GoF and quality indices of the structural equation models are presented in Table 2. The models that excluded (Model 1) and included (Model 2) past behavior exhibited adequate fit and quality indices in both the Australian and U.S. samples. Standardized parameter estimates for the proposed direct effects for each model in the Australian and U.S. samples are presented in Fig. 1. Full parameter estimates for models in both samples are presented in Supplementary Appendix E. Parameter estimates, CIs, and effect sizes for the indirect effects of the models in both samples are summarized in Table 3.
Table 2. Model quality and GoF statistics for the structural equation models of the integrated model in the Australian and U.S. samples and multigroup model
Sample Model APC AR2 AVIF AFVIF GoF SPR R2CR SSR NLBCDR
Australia 1 .104** .177*** 1.177 1.561 .391 .841 .977 .889 .873
2 .116*** .338*** 1.222 1.904 .543 .819 .991 .931 .785
USA 1 .098** .192*** 1.187 1.823 .410 .889 .995 .825 .754
2 .116*** .338** 1.222 1.904 .543 .819 .991 .931 .785
MS 1 .100*** .182*** 1.159 1.704 .398 .905 .995 .794 .817
2 .113*** .300*** 1.186 1.760 .511 .931 .997 .917 .840
Model 1 = model excluding past behavior; Model 2 = model including past behavior.
AFVIF average full collinearity variance inflation factor; APC average path coefficient; AR2 average R2; AVIF average block variance inflation factor; GoF Tenenhaus’s goodness-of-fit index; MS multiple sample analysis; NLBCDR nonlinear bivariate causality direction ratio; R2CR R2 contribution ratio; SPR Sympson’s paradox ratio; SSR statistical suppression ratio.
*p < .05, **p < .01, ***p < .001
Table 3. Standardized parameter estimates for indirect effects for the structural equation model of the integrated model in the Australian and U.S. samples
Effect Model excluding past behavior Model including past behavior
β p 95% CI ES Β p 95% CI ES
LB UB LB UB
Australian sample
Indirect effects
Att.→Int.→Beh. .011 .359 −.052 .074 .003 .004 .444 −.059 .067 .001
SN→Int.→Beh. .042 .094 −.021 .105 .016 .016 .312 −.047 .079 .006
MN→Int.→Beh. .068 .016 .005 .131 .024 .028 .192 −.035 .091 .010
AR→Int.→Beh. .011 .356 −.052 .074 .003 .003 .457 −.060 .066 .001
PBC→Int.→Beh. .040 .101 −.023 .103 .011 .016 .307 −.047 .079 .005
Int.→AP→Beh. .040 .106 −.023 .103 .014 .011 .365 −.052 .074 .004
Hab. (T1).→Hab. (T2)→Beh. .102 <.001 .041 .163 .016 .078 .007 .017 .139 .013
PB→Hab.→Beh. – – – – – .021 .214 −.030 .072 .011
PB→Beh.a – – – – – .081 .034 −.007 .169 .042
Total effectsb
Int.→Beh. .220 <.001 .134 .306 .081 .090 .022 .004 .176 .033
PBC→Beh. .126 <.001 .040 .212 .036 .055 .110 −.033 .143 .016
Hab. (T1)→Beh. .096 .016 .010 .182 .015 .076 .044 −.012 .164 .012
PB→Beh. – – – – – .494 <.001 .412 .576 .258
U.S. sample
Indirect effects
Att.→Int.→Beh. <.001 .495 −.052 .054 <.001 .004 .443 −.049 .057 .001
SN→Int.→Beh. .072 .003 .019 .125 .029 .023 .190 −.030 .076 .009
MN→Int.→Beh. .102 <.001 .051 .153 .044 .040 .067 −.013 .093 .017
AR→Int.→Beh. .023 .192 −.030 .076 .011 .001 .478 −.052 .054 .001
PBC→Int.→Beh. .088 <.001 .037 .139 .025 .038 .079 −.015 .091 .011
Int.→AP→Beh. .061 .011 .008 .114 .029 .004 .441 −.049 .057 .002
Hab. (T1).→Hab. (T2)→Beh. .212 <.001 .161 .263 .075 .166 <.001 .115 .217 .059
PB→Hab.→Beh. – – – – – .068 <.001 .025 .111 .043
PB→Beh.a – – – – – .178 <.001 .105 .251 .112
Total effectsb
Int.→Beh. .377 <.001 .306 .448 .177 .142 <.001 .069 .215 .066
PBC→Beh. .146 <.001 .073 .219 .042 .074 .024 .001 .147 .021
Hab. (T1)→Beh. .242 <.001 .169 .315 .086 .171 <.001 .098 .244 .061
PB→Beh. – – −.052 .074 – .673 <.001 .604 .742 .423
aSum of indirect effects of past behavior on behavior through all model constructs.
bTotal effect comprising sums of all indirect effects through model constructs plus the direct effect.
β standardized parameter estimate; 95% CI 95% confidence interval of standardized parameter estimate; AP action planning; AR anticipated regret; Att. attitude; Beh. behavior; ES effect size of the standardized parameter estimate; Hab. (T1) self-reported habit measured at baseline (T1); Hab. (T2) self-reported habit measured at follow-up (T2); Int. intention; LB lower bound of 95% CI; MN moral norm; PB past behavior; PBC perceived behavioral control; SN subjective norm; UB upper bound of 95% CI. Focusing on the model excluding past behavior (Model 1), intention, action planning, and habit at follow-up were statistically significant direct predictors of social distancing behavior, with effect size for intention and habit generally larger in the U.S. sample. PBC directly predicted behavior in the Australian sample only, also with a small effect size. Intention predicted action planning in both samples with large effect sizes. Subjective norm, moral norm, and PBC predicted intention in both samples, with small-to-medium effect sizes, but effects of attitude were not significant. There was a small effect of anticipated regret on intention in the U.S. sample only. Habit at baseline predicted habit at follow-up in both samples, with large effect sizes. There was also a small-sized effect of habit at baseline on intention in the U.S. sample only. Overall, the model accounted for significant variance in social distancing behavior (Australian sample, R2 = .198; U.S. sample, R2 = .361), intentions (Australian sample, R2 = .571; U.S. sample, R2 = .623), and habit at follow-up (Australian sample, R2 = .416; U.S. sample, R2 = .486). Intentions (Australian sample, R2 = .066; U.S. sample, R2 = .148), action planning (Australian sample, R2 = .029; U.S. sample, R2 = .044), and habit at follow-up (Australian sample R2 = .041; U.S. sample, R2 = .129) each accounted for substantive variance in behavior. Action planning significantly moderated the intention–behavior relationship in the Australian sample only. While the effect was not in the predicted direction, probing the interaction revealed that the intention–behavior relationship increased as the level of planning increased, consistent with theory. However, the intention–behavior relationship is more likely to be smaller at lower levels of planning, and it seems that planning makes less of a difference when the intention–behavior relationship is large. A plot of the interaction effect is presented in Supplementary Appendix F.
Turning to the indirect effects, there were significant indirect effects of subjective norm, moral norm, and PBC on social distancing behavior mediated by intention in the U.S. sample. By contrast, only the indirect effect of moral norm on behavior through intention was significant in the Australian sample. The smaller indirect effects in the Australian sample is principally due to the significantly smaller effect size for the intention–behavior relationship in this sample compared to the U.S. sample. Habit at baseline predicted behavior through habit at follow-up in both samples. Effect sizes in all cases were small. There were significant total effects of intention, PBC, and habit at baseline on behavior, with effect sizes larger in the U.S. sample than in the Australian sample.
For the model including past behavior, significant effects of past behavior on all model constructs were observed in both samples with effect sizes ranging from small to large. The effects of past behavior on social distancing behavior were particularly large. Inclusion of past behavior led to an attenuation of model effects in both samples. Specifically, the effects of intention and habit at follow-up on behavior were reduced but remained statistically significant in both samples with small effect sizes. In addition, effects of subjective norm, moral norm, and PBC on intention, and the effect of intention on action planning, remained statistically significant in both samples, with small-to-medium effect sizes. The effect of habit at baseline on habit at follow-up was statistically significant in both samples, with large effect sizes. Variance explained in social distancing behavior increased substantially with the inclusion of past behavior, with only modest changes in explained variance in intentions (Australian sample R2 = .598; U.S. sample, R2 = .702) and habit at follow-up (Australian sample R2 = .416; U.S. sample, R2 = .486). Specifically, intentions (Australian sample, R2 = .029; U.S. sample, R2 = .065), past behavior (Australian sample, R2 = .216; U.S. sample, R2 = .311), and habit at follow-up (Australian sample R2 = .031; U.S. sample, R2 = .101) each accounted for substantive variance in behavior.
Turning to indirect effects, we found significant indirect effects of habit at baseline on behavior mediated by habit at follow-up in both samples with small effect sizes. There were also significant total effects of intention and habit at baseline on behavior in both samples, and of PBC on behavior for the U.S. sample, with small effect sizes. There were significant total indirect and total effects of past behavior on behavior in both samples, with large effect sizes. There was a small-sized indirect effect of past behavior on behavior mediated by habit at both time points in the U.S. sample, but the effect was not significant in the Australian sample.
Multisample analyses
Multisample analyses permitted for tests of difference in parameter estimates for each model across the Australian and U.S. samples. For the model excluding past behavior (Model 1), only effects of intention on habit at baseline, habit at follow-up on social distancing behavior, and intention on action planning differed across samples. These effects were significantly larger in the U.S. sample. Some effects with observed differences across samples, such as the effect of habit at baseline on intention or the moderator effect of planning on the intention–behavior relationship, did not differ significantly across samples. For the model including effects of past behavior (Model 2), multisample analysis revealed no differences in effect size across samples, indicating that the attenuating effect of past behavior on model effects also had the effect of eliminating the few differences in model effects across samples. Full details of the multiple-group analysis are presented in Supplementary Appendix G.
Sensitivity analyses
We re-estimated both models in samples excluding participants who were never subject to a “shelter-in place” order, and in the sample that were never subject to an order, or who had the order lifted at some stage during the study. Comparisons of parameter estimates in these models with those from the models estimated in the full sample revealed no significant differences in any of the model parameters. Results are reported in Supplementary Appendices H and I.
|
|
title
|
Structural Equation Models
|
|
sec
|
Single-sample analyses
GoF and quality indices of the structural equation models are presented in Table 2. The models that excluded (Model 1) and included (Model 2) past behavior exhibited adequate fit and quality indices in both the Australian and U.S. samples. Standardized parameter estimates for the proposed direct effects for each model in the Australian and U.S. samples are presented in Fig. 1. Full parameter estimates for models in both samples are presented in Supplementary Appendix E. Parameter estimates, CIs, and effect sizes for the indirect effects of the models in both samples are summarized in Table 3.
Table 2. Model quality and GoF statistics for the structural equation models of the integrated model in the Australian and U.S. samples and multigroup model
Sample Model APC AR2 AVIF AFVIF GoF SPR R2CR SSR NLBCDR
Australia 1 .104** .177*** 1.177 1.561 .391 .841 .977 .889 .873
2 .116*** .338*** 1.222 1.904 .543 .819 .991 .931 .785
USA 1 .098** .192*** 1.187 1.823 .410 .889 .995 .825 .754
2 .116*** .338** 1.222 1.904 .543 .819 .991 .931 .785
MS 1 .100*** .182*** 1.159 1.704 .398 .905 .995 .794 .817
2 .113*** .300*** 1.186 1.760 .511 .931 .997 .917 .840
Model 1 = model excluding past behavior; Model 2 = model including past behavior.
AFVIF average full collinearity variance inflation factor; APC average path coefficient; AR2 average R2; AVIF average block variance inflation factor; GoF Tenenhaus’s goodness-of-fit index; MS multiple sample analysis; NLBCDR nonlinear bivariate causality direction ratio; R2CR R2 contribution ratio; SPR Sympson’s paradox ratio; SSR statistical suppression ratio.
*p < .05, **p < .01, ***p < .001
Table 3. Standardized parameter estimates for indirect effects for the structural equation model of the integrated model in the Australian and U.S. samples
Effect Model excluding past behavior Model including past behavior
β p 95% CI ES Β p 95% CI ES
LB UB LB UB
Australian sample
Indirect effects
Att.→Int.→Beh. .011 .359 −.052 .074 .003 .004 .444 −.059 .067 .001
SN→Int.→Beh. .042 .094 −.021 .105 .016 .016 .312 −.047 .079 .006
MN→Int.→Beh. .068 .016 .005 .131 .024 .028 .192 −.035 .091 .010
AR→Int.→Beh. .011 .356 −.052 .074 .003 .003 .457 −.060 .066 .001
PBC→Int.→Beh. .040 .101 −.023 .103 .011 .016 .307 −.047 .079 .005
Int.→AP→Beh. .040 .106 −.023 .103 .014 .011 .365 −.052 .074 .004
Hab. (T1).→Hab. (T2)→Beh. .102 <.001 .041 .163 .016 .078 .007 .017 .139 .013
PB→Hab.→Beh. – – – – – .021 .214 −.030 .072 .011
PB→Beh.a – – – – – .081 .034 −.007 .169 .042
Total effectsb
Int.→Beh. .220 <.001 .134 .306 .081 .090 .022 .004 .176 .033
PBC→Beh. .126 <.001 .040 .212 .036 .055 .110 −.033 .143 .016
Hab. (T1)→Beh. .096 .016 .010 .182 .015 .076 .044 −.012 .164 .012
PB→Beh. – – – – – .494 <.001 .412 .576 .258
U.S. sample
Indirect effects
Att.→Int.→Beh. <.001 .495 −.052 .054 <.001 .004 .443 −.049 .057 .001
SN→Int.→Beh. .072 .003 .019 .125 .029 .023 .190 −.030 .076 .009
MN→Int.→Beh. .102 <.001 .051 .153 .044 .040 .067 −.013 .093 .017
AR→Int.→Beh. .023 .192 −.030 .076 .011 .001 .478 −.052 .054 .001
PBC→Int.→Beh. .088 <.001 .037 .139 .025 .038 .079 −.015 .091 .011
Int.→AP→Beh. .061 .011 .008 .114 .029 .004 .441 −.049 .057 .002
Hab. (T1).→Hab. (T2)→Beh. .212 <.001 .161 .263 .075 .166 <.001 .115 .217 .059
PB→Hab.→Beh. – – – – – .068 <.001 .025 .111 .043
PB→Beh.a – – – – – .178 <.001 .105 .251 .112
Total effectsb
Int.→Beh. .377 <.001 .306 .448 .177 .142 <.001 .069 .215 .066
PBC→Beh. .146 <.001 .073 .219 .042 .074 .024 .001 .147 .021
Hab. (T1)→Beh. .242 <.001 .169 .315 .086 .171 <.001 .098 .244 .061
PB→Beh. – – −.052 .074 – .673 <.001 .604 .742 .423
aSum of indirect effects of past behavior on behavior through all model constructs.
bTotal effect comprising sums of all indirect effects through model constructs plus the direct effect.
β standardized parameter estimate; 95% CI 95% confidence interval of standardized parameter estimate; AP action planning; AR anticipated regret; Att. attitude; Beh. behavior; ES effect size of the standardized parameter estimate; Hab. (T1) self-reported habit measured at baseline (T1); Hab. (T2) self-reported habit measured at follow-up (T2); Int. intention; LB lower bound of 95% CI; MN moral norm; PB past behavior; PBC perceived behavioral control; SN subjective norm; UB upper bound of 95% CI. Focusing on the model excluding past behavior (Model 1), intention, action planning, and habit at follow-up were statistically significant direct predictors of social distancing behavior, with effect size for intention and habit generally larger in the U.S. sample. PBC directly predicted behavior in the Australian sample only, also with a small effect size. Intention predicted action planning in both samples with large effect sizes. Subjective norm, moral norm, and PBC predicted intention in both samples, with small-to-medium effect sizes, but effects of attitude were not significant. There was a small effect of anticipated regret on intention in the U.S. sample only. Habit at baseline predicted habit at follow-up in both samples, with large effect sizes. There was also a small-sized effect of habit at baseline on intention in the U.S. sample only. Overall, the model accounted for significant variance in social distancing behavior (Australian sample, R2 = .198; U.S. sample, R2 = .361), intentions (Australian sample, R2 = .571; U.S. sample, R2 = .623), and habit at follow-up (Australian sample, R2 = .416; U.S. sample, R2 = .486). Intentions (Australian sample, R2 = .066; U.S. sample, R2 = .148), action planning (Australian sample, R2 = .029; U.S. sample, R2 = .044), and habit at follow-up (Australian sample R2 = .041; U.S. sample, R2 = .129) each accounted for substantive variance in behavior. Action planning significantly moderated the intention–behavior relationship in the Australian sample only. While the effect was not in the predicted direction, probing the interaction revealed that the intention–behavior relationship increased as the level of planning increased, consistent with theory. However, the intention–behavior relationship is more likely to be smaller at lower levels of planning, and it seems that planning makes less of a difference when the intention–behavior relationship is large. A plot of the interaction effect is presented in Supplementary Appendix F.
Turning to the indirect effects, there were significant indirect effects of subjective norm, moral norm, and PBC on social distancing behavior mediated by intention in the U.S. sample. By contrast, only the indirect effect of moral norm on behavior through intention was significant in the Australian sample. The smaller indirect effects in the Australian sample is principally due to the significantly smaller effect size for the intention–behavior relationship in this sample compared to the U.S. sample. Habit at baseline predicted behavior through habit at follow-up in both samples. Effect sizes in all cases were small. There were significant total effects of intention, PBC, and habit at baseline on behavior, with effect sizes larger in the U.S. sample than in the Australian sample.
For the model including past behavior, significant effects of past behavior on all model constructs were observed in both samples with effect sizes ranging from small to large. The effects of past behavior on social distancing behavior were particularly large. Inclusion of past behavior led to an attenuation of model effects in both samples. Specifically, the effects of intention and habit at follow-up on behavior were reduced but remained statistically significant in both samples with small effect sizes. In addition, effects of subjective norm, moral norm, and PBC on intention, and the effect of intention on action planning, remained statistically significant in both samples, with small-to-medium effect sizes. The effect of habit at baseline on habit at follow-up was statistically significant in both samples, with large effect sizes. Variance explained in social distancing behavior increased substantially with the inclusion of past behavior, with only modest changes in explained variance in intentions (Australian sample R2 = .598; U.S. sample, R2 = .702) and habit at follow-up (Australian sample R2 = .416; U.S. sample, R2 = .486). Specifically, intentions (Australian sample, R2 = .029; U.S. sample, R2 = .065), past behavior (Australian sample, R2 = .216; U.S. sample, R2 = .311), and habit at follow-up (Australian sample R2 = .031; U.S. sample, R2 = .101) each accounted for substantive variance in behavior.
Turning to indirect effects, we found significant indirect effects of habit at baseline on behavior mediated by habit at follow-up in both samples with small effect sizes. There were also significant total effects of intention and habit at baseline on behavior in both samples, and of PBC on behavior for the U.S. sample, with small effect sizes. There were significant total indirect and total effects of past behavior on behavior in both samples, with large effect sizes. There was a small-sized indirect effect of past behavior on behavior mediated by habit at both time points in the U.S. sample, but the effect was not significant in the Australian sample.
|
|
title
|
Single-sample analyses
|
|
p
|
GoF and quality indices of the structural equation models are presented in Table 2. The models that excluded (Model 1) and included (Model 2) past behavior exhibited adequate fit and quality indices in both the Australian and U.S. samples. Standardized parameter estimates for the proposed direct effects for each model in the Australian and U.S. samples are presented in Fig. 1. Full parameter estimates for models in both samples are presented in Supplementary Appendix E. Parameter estimates, CIs, and effect sizes for the indirect effects of the models in both samples are summarized in Table 3.
|
|
table-wrap
|
Table 2. Model quality and GoF statistics for the structural equation models of the integrated model in the Australian and U.S. samples and multigroup model
Sample Model APC AR2 AVIF AFVIF GoF SPR R2CR SSR NLBCDR
Australia 1 .104** .177*** 1.177 1.561 .391 .841 .977 .889 .873
2 .116*** .338*** 1.222 1.904 .543 .819 .991 .931 .785
USA 1 .098** .192*** 1.187 1.823 .410 .889 .995 .825 .754
2 .116*** .338** 1.222 1.904 .543 .819 .991 .931 .785
MS 1 .100*** .182*** 1.159 1.704 .398 .905 .995 .794 .817
2 .113*** .300*** 1.186 1.760 .511 .931 .997 .917 .840
Model 1 = model excluding past behavior; Model 2 = model including past behavior.
AFVIF average full collinearity variance inflation factor; APC average path coefficient; AR2 average R2; AVIF average block variance inflation factor; GoF Tenenhaus’s goodness-of-fit index; MS multiple sample analysis; NLBCDR nonlinear bivariate causality direction ratio; R2CR R2 contribution ratio; SPR Sympson’s paradox ratio; SSR statistical suppression ratio.
*p < .05, **p < .01, ***p < .001
|
|
label
|
Table 2.
|
|
caption
|
Model quality and GoF statistics for the structural equation models of the integrated model in the Australian and U.S. samples and multigroup model
|
|
p
|
Model quality and GoF statistics for the structural equation models of the integrated model in the Australian and U.S. samples and multigroup model
|
|
table
|
Sample Model APC AR2 AVIF AFVIF GoF SPR R2CR SSR NLBCDR
Australia 1 .104** .177*** 1.177 1.561 .391 .841 .977 .889 .873
2 .116*** .338*** 1.222 1.904 .543 .819 .991 .931 .785
USA 1 .098** .192*** 1.187 1.823 .410 .889 .995 .825 .754
2 .116*** .338** 1.222 1.904 .543 .819 .991 .931 .785
MS 1 .100*** .182*** 1.159 1.704 .398 .905 .995 .794 .817
2 .113*** .300*** 1.186 1.760 .511 .931 .997 .917 .840
|
|
tr
|
Sample Model APC AR2 AVIF AFVIF GoF SPR R2CR SSR NLBCDR
|
|
th
|
Sample
|
|
th
|
Model
|
|
th
|
APC
|
|
th
|
AR2
|
|
th
|
AVIF
|
|
th
|
AFVIF
|
|
th
|
GoF
|
|
th
|
SPR
|
|
th
|
R2CR
|
|
th
|
SSR
|
|
th
|
NLBCDR
|
|
tr
|
Australia 1 .104** .177*** 1.177 1.561 .391 .841 .977 .889 .873
|
|
td
|
Australia
|
|
td
|
1
|
|
td
|
.104**
|
|
td
|
.177***
|
|
td
|
1.177
|
|
td
|
1.561
|
|
td
|
.391
|
|
td
|
.841
|
|
td
|
.977
|
|
td
|
.889
|
|
td
|
.873
|
|
tr
|
2 .116*** .338*** 1.222 1.904 .543 .819 .991 .931 .785
|
|
td
|
2
|
|
td
|
.116***
|
|
td
|
.338***
|
|
td
|
1.222
|
|
td
|
1.904
|
|
td
|
.543
|
|
td
|
.819
|
|
td
|
.991
|
|
td
|
.931
|
|
td
|
.785
|
|
tr
|
USA 1 .098** .192*** 1.187 1.823 .410 .889 .995 .825 .754
|
|
td
|
USA
|
|
td
|
1
|
|
td
|
.098**
|
|
td
|
.192***
|
|
td
|
1.187
|
|
td
|
1.823
|
|
td
|
.410
|
|
td
|
.889
|
|
td
|
.995
|
|
td
|
.825
|
|
td
|
.754
|
|
tr
|
2 .116*** .338** 1.222 1.904 .543 .819 .991 .931 .785
|
|
td
|
2
|
|
td
|
.116***
|
|
td
|
.338**
|
|
td
|
1.222
|
|
td
|
1.904
|
|
td
|
.543
|
|
td
|
.819
|
|
td
|
.991
|
|
td
|
.931
|
|
td
|
.785
|
|
tr
|
MS 1 .100*** .182*** 1.159 1.704 .398 .905 .995 .794 .817
|
|
td
|
MS
|
|
td
|
1
|
|
td
|
.100***
|
|
td
|
.182***
|
|
td
|
1.159
|
|
td
|
1.704
|
|
td
|
.398
|
|
td
|
.905
|
|
td
|
.995
|
|
td
|
.794
|
|
td
|
.817
|
|
tr
|
2 .113*** .300*** 1.186 1.760 .511 .931 .997 .917 .840
|
|
td
|
2
|
|
td
|
.113***
|
|
td
|
.300***
|
|
td
|
1.186
|
|
td
|
1.760
|
|
td
|
.511
|
|
td
|
.931
|
|
td
|
.997
|
|
td
|
.917
|
|
td
|
.840
|
|
table-wrap-foot
|
Model 1 = model excluding past behavior; Model 2 = model including past behavior.
AFVIF average full collinearity variance inflation factor; APC average path coefficient; AR2 average R2; AVIF average block variance inflation factor; GoF Tenenhaus’s goodness-of-fit index; MS multiple sample analysis; NLBCDR nonlinear bivariate causality direction ratio; R2CR R2 contribution ratio; SPR Sympson’s paradox ratio; SSR statistical suppression ratio.
*p < .05, **p < .01, ***p < .001
|
|
footnote
|
Model 1 = model excluding past behavior; Model 2 = model including past behavior.
|
|
p
|
Model 1 = model excluding past behavior; Model 2 = model including past behavior.
|
|
footnote
|
AFVIF average full collinearity variance inflation factor; APC average path coefficient; AR2 average R2; AVIF average block variance inflation factor; GoF Tenenhaus’s goodness-of-fit index; MS multiple sample analysis; NLBCDR nonlinear bivariate causality direction ratio; R2CR R2 contribution ratio; SPR Sympson’s paradox ratio; SSR statistical suppression ratio.
|
|
p
|
AFVIF average full collinearity variance inflation factor; APC average path coefficient; AR2 average R2; AVIF average block variance inflation factor; GoF Tenenhaus’s goodness-of-fit index; MS multiple sample analysis; NLBCDR nonlinear bivariate causality direction ratio; R2CR R2 contribution ratio; SPR Sympson’s paradox ratio; SSR statistical suppression ratio.
|
|
footnote
|
*p < .05, **p < .01, ***p < .001
|
|
p
|
*p < .05, **p < .01, ***p < .001
|
|
table-wrap
|
Table 3. Standardized parameter estimates for indirect effects for the structural equation model of the integrated model in the Australian and U.S. samples
Effect Model excluding past behavior Model including past behavior
β p 95% CI ES Β p 95% CI ES
LB UB LB UB
Australian sample
Indirect effects
Att.→Int.→Beh. .011 .359 −.052 .074 .003 .004 .444 −.059 .067 .001
SN→Int.→Beh. .042 .094 −.021 .105 .016 .016 .312 −.047 .079 .006
MN→Int.→Beh. .068 .016 .005 .131 .024 .028 .192 −.035 .091 .010
AR→Int.→Beh. .011 .356 −.052 .074 .003 .003 .457 −.060 .066 .001
PBC→Int.→Beh. .040 .101 −.023 .103 .011 .016 .307 −.047 .079 .005
Int.→AP→Beh. .040 .106 −.023 .103 .014 .011 .365 −.052 .074 .004
Hab. (T1).→Hab. (T2)→Beh. .102 <.001 .041 .163 .016 .078 .007 .017 .139 .013
PB→Hab.→Beh. – – – – – .021 .214 −.030 .072 .011
PB→Beh.a – – – – – .081 .034 −.007 .169 .042
Total effectsb
Int.→Beh. .220 <.001 .134 .306 .081 .090 .022 .004 .176 .033
PBC→Beh. .126 <.001 .040 .212 .036 .055 .110 −.033 .143 .016
Hab. (T1)→Beh. .096 .016 .010 .182 .015 .076 .044 −.012 .164 .012
PB→Beh. – – – – – .494 <.001 .412 .576 .258
U.S. sample
Indirect effects
Att.→Int.→Beh. <.001 .495 −.052 .054 <.001 .004 .443 −.049 .057 .001
SN→Int.→Beh. .072 .003 .019 .125 .029 .023 .190 −.030 .076 .009
MN→Int.→Beh. .102 <.001 .051 .153 .044 .040 .067 −.013 .093 .017
AR→Int.→Beh. .023 .192 −.030 .076 .011 .001 .478 −.052 .054 .001
PBC→Int.→Beh. .088 <.001 .037 .139 .025 .038 .079 −.015 .091 .011
Int.→AP→Beh. .061 .011 .008 .114 .029 .004 .441 −.049 .057 .002
Hab. (T1).→Hab. (T2)→Beh. .212 <.001 .161 .263 .075 .166 <.001 .115 .217 .059
PB→Hab.→Beh. – – – – – .068 <.001 .025 .111 .043
PB→Beh.a – – – – – .178 <.001 .105 .251 .112
Total effectsb
Int.→Beh. .377 <.001 .306 .448 .177 .142 <.001 .069 .215 .066
PBC→Beh. .146 <.001 .073 .219 .042 .074 .024 .001 .147 .021
Hab. (T1)→Beh. .242 <.001 .169 .315 .086 .171 <.001 .098 .244 .061
PB→Beh. – – −.052 .074 – .673 <.001 .604 .742 .423
aSum of indirect effects of past behavior on behavior through all model constructs.
bTotal effect comprising sums of all indirect effects through model constructs plus the direct effect.
β standardized parameter estimate; 95% CI 95% confidence interval of standardized parameter estimate; AP action planning; AR anticipated regret; Att. attitude; Beh. behavior; ES effect size of the standardized parameter estimate; Hab. (T1) self-reported habit measured at baseline (T1); Hab. (T2) self-reported habit measured at follow-up (T2); Int. intention; LB lower bound of 95% CI; MN moral norm; PB past behavior; PBC perceived behavioral control; SN subjective norm; UB upper bound of 95% CI.
|
|
label
|
Table 3.
|
|
caption
|
Standardized parameter estimates for indirect effects for the structural equation model of the integrated model in the Australian and U.S. samples
|
|
p
|
Standardized parameter estimates for indirect effects for the structural equation model of the integrated model in the Australian and U.S. samples
|
|
table
|
Effect Model excluding past behavior Model including past behavior
β p 95% CI ES Β p 95% CI ES
LB UB LB UB
Australian sample
Indirect effects
Att.→Int.→Beh. .011 .359 −.052 .074 .003 .004 .444 −.059 .067 .001
SN→Int.→Beh. .042 .094 −.021 .105 .016 .016 .312 −.047 .079 .006
MN→Int.→Beh. .068 .016 .005 .131 .024 .028 .192 −.035 .091 .010
AR→Int.→Beh. .011 .356 −.052 .074 .003 .003 .457 −.060 .066 .001
PBC→Int.→Beh. .040 .101 −.023 .103 .011 .016 .307 −.047 .079 .005
Int.→AP→Beh. .040 .106 −.023 .103 .014 .011 .365 −.052 .074 .004
Hab. (T1).→Hab. (T2)→Beh. .102 <.001 .041 .163 .016 .078 .007 .017 .139 .013
PB→Hab.→Beh. – – – – – .021 .214 −.030 .072 .011
PB→Beh.a – – – – – .081 .034 −.007 .169 .042
Total effectsb
Int.→Beh. .220 <.001 .134 .306 .081 .090 .022 .004 .176 .033
PBC→Beh. .126 <.001 .040 .212 .036 .055 .110 −.033 .143 .016
Hab. (T1)→Beh. .096 .016 .010 .182 .015 .076 .044 −.012 .164 .012
PB→Beh. – – – – – .494 <.001 .412 .576 .258
U.S. sample
Indirect effects
Att.→Int.→Beh. <.001 .495 −.052 .054 <.001 .004 .443 −.049 .057 .001
SN→Int.→Beh. .072 .003 .019 .125 .029 .023 .190 −.030 .076 .009
MN→Int.→Beh. .102 <.001 .051 .153 .044 .040 .067 −.013 .093 .017
AR→Int.→Beh. .023 .192 −.030 .076 .011 .001 .478 −.052 .054 .001
PBC→Int.→Beh. .088 <.001 .037 .139 .025 .038 .079 −.015 .091 .011
Int.→AP→Beh. .061 .011 .008 .114 .029 .004 .441 −.049 .057 .002
Hab. (T1).→Hab. (T2)→Beh. .212 <.001 .161 .263 .075 .166 <.001 .115 .217 .059
PB→Hab.→Beh. – – – – – .068 <.001 .025 .111 .043
PB→Beh.a – – – – – .178 <.001 .105 .251 .112
Total effectsb
Int.→Beh. .377 <.001 .306 .448 .177 .142 <.001 .069 .215 .066
PBC→Beh. .146 <.001 .073 .219 .042 .074 .024 .001 .147 .021
Hab. (T1)→Beh. .242 <.001 .169 .315 .086 .171 <.001 .098 .244 .061
PB→Beh. – – −.052 .074 – .673 <.001 .604 .742 .423
|
|
tr
|
Effect Model excluding past behavior Model including past behavior
|
|
th
|
Effect
|
|
th
|
Model excluding past behavior
|
|
th
|
Model including past behavior
|
|
tr
|
β p 95% CI ES Β p 95% CI ES
|
|
td
|
β
|
|
td
|
p
|
|
td
|
95% CI
|
|
td
|
ES
|
|
td
|
Β
|
|
td
|
p
|
|
td
|
95% CI
|
|
td
|
ES
|
|
tr
|
LB UB LB UB
|
|
td
|
LB
|
|
td
|
UB
|
|
td
|
LB
|
|
td
|
UB
|
|
tr
|
Australian sample
|
|
td
|
Australian sample
|
|
tr
|
Indirect effects
|
|
td
|
Indirect effects
|
|
tr
|
Att.→Int.→Beh. .011 .359 −.052 .074 .003 .004 .444 −.059 .067 .001
|
|
td
|
Att.→Int.→Beh.
|
|
td
|
.011
|
|
td
|
.359
|
|
td
|
−.052
|
|
td
|
.074
|
|
td
|
.003
|
|
td
|
.004
|
|
td
|
.444
|
|
td
|
−.059
|
|
td
|
.067
|
|
td
|
.001
|
|
tr
|
SN→Int.→Beh. .042 .094 −.021 .105 .016 .016 .312 −.047 .079 .006
|
|
td
|
SN→Int.→Beh.
|
|
td
|
.042
|
|
td
|
.094
|
|
td
|
−.021
|
|
td
|
.105
|
|
td
|
.016
|
|
td
|
.016
|
|
td
|
.312
|
|
td
|
−.047
|
|
td
|
.079
|
|
td
|
.006
|
|
tr
|
MN→Int.→Beh. .068 .016 .005 .131 .024 .028 .192 −.035 .091 .010
|
|
td
|
MN→Int.→Beh.
|
|
td
|
.068
|
|
td
|
.016
|
|
td
|
.005
|
|
td
|
.131
|
|
td
|
.024
|
|
td
|
.028
|
|
td
|
.192
|
|
td
|
−.035
|
|
td
|
.091
|
|
td
|
.010
|
|
tr
|
AR→Int.→Beh. .011 .356 −.052 .074 .003 .003 .457 −.060 .066 .001
|
|
td
|
AR→Int.→Beh.
|
|
td
|
.011
|
|
td
|
.356
|
|
td
|
−.052
|
|
td
|
.074
|
|
td
|
.003
|
|
td
|
.003
|
|
td
|
.457
|
|
td
|
−.060
|
|
td
|
.066
|
|
td
|
.001
|
|
tr
|
PBC→Int.→Beh. .040 .101 −.023 .103 .011 .016 .307 −.047 .079 .005
|
|
td
|
PBC→Int.→Beh.
|
|
td
|
.040
|
|
td
|
.101
|
|
td
|
−.023
|
|
td
|
.103
|
|
td
|
.011
|
|
td
|
.016
|
|
td
|
.307
|
|
td
|
−.047
|
|
td
|
.079
|
|
td
|
.005
|
|
tr
|
Int.→AP→Beh. .040 .106 −.023 .103 .014 .011 .365 −.052 .074 .004
|
|
td
|
Int.→AP→Beh.
|
|
td
|
.040
|
|
td
|
.106
|
|
td
|
−.023
|
|
td
|
.103
|
|
td
|
.014
|
|
td
|
.011
|
|
td
|
.365
|
|
td
|
−.052
|
|
td
|
.074
|
|
td
|
.004
|
|
tr
|
Hab. (T1).→Hab. (T2)→Beh. .102 <.001 .041 .163 .016 .078 .007 .017 .139 .013
|
|
td
|
Hab. (T1).→Hab. (T2)→Beh.
|
|
td
|
.102
|
|
td
|
<.001
|
|
td
|
.041
|
|
td
|
.163
|
|
td
|
.016
|
|
td
|
.078
|
|
td
|
.007
|
|
td
|
.017
|
|
td
|
.139
|
|
td
|
.013
|
|
tr
|
PB→Hab.→Beh. – – – – – .021 .214 −.030 .072 .011
|
|
td
|
PB→Hab.→Beh.
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
.021
|
|
td
|
.214
|
|
td
|
−.030
|
|
td
|
.072
|
|
td
|
.011
|
|
tr
|
PB→Beh.a – – – – – .081 .034 −.007 .169 .042
|
|
td
|
PB→Beh.a
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
.081
|
|
td
|
.034
|
|
td
|
−.007
|
|
td
|
.169
|
|
td
|
.042
|
|
tr
|
Total effectsb
|
|
td
|
Total effectsb
|
|
tr
|
Int.→Beh. .220 <.001 .134 .306 .081 .090 .022 .004 .176 .033
|
|
td
|
Int.→Beh.
|
|
td
|
.220
|
|
td
|
<.001
|
|
td
|
.134
|
|
td
|
.306
|
|
td
|
.081
|
|
td
|
.090
|
|
td
|
.022
|
|
td
|
.004
|
|
td
|
.176
|
|
td
|
.033
|
|
tr
|
PBC→Beh. .126 <.001 .040 .212 .036 .055 .110 −.033 .143 .016
|
|
td
|
PBC→Beh.
|
|
td
|
.126
|
|
td
|
<.001
|
|
td
|
.040
|
|
td
|
.212
|
|
td
|
.036
|
|
td
|
.055
|
|
td
|
.110
|
|
td
|
−.033
|
|
td
|
.143
|
|
td
|
.016
|
|
tr
|
Hab. (T1)→Beh. .096 .016 .010 .182 .015 .076 .044 −.012 .164 .012
|
|
td
|
Hab. (T1)→Beh.
|
|
td
|
.096
|
|
td
|
.016
|
|
td
|
.010
|
|
td
|
.182
|
|
td
|
.015
|
|
td
|
.076
|
|
td
|
.044
|
|
td
|
−.012
|
|
td
|
.164
|
|
td
|
.012
|
|
tr
|
PB→Beh. – – – – – .494 <.001 .412 .576 .258
|
|
td
|
PB→Beh.
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
.494
|
|
td
|
<.001
|
|
td
|
.412
|
|
td
|
.576
|
|
td
|
.258
|
|
tr
|
U.S. sample
|
|
td
|
U.S. sample
|
|
tr
|
Indirect effects
|
|
td
|
Indirect effects
|
|
tr
|
Att.→Int.→Beh. <.001 .495 −.052 .054 <.001 .004 .443 −.049 .057 .001
|
|
td
|
Att.→Int.→Beh.
|
|
td
|
<.001
|
|
td
|
.495
|
|
td
|
−.052
|
|
td
|
.054
|
|
td
|
<.001
|
|
td
|
.004
|
|
td
|
.443
|
|
td
|
−.049
|
|
td
|
.057
|
|
td
|
.001
|
|
tr
|
SN→Int.→Beh. .072 .003 .019 .125 .029 .023 .190 −.030 .076 .009
|
|
td
|
SN→Int.→Beh.
|
|
td
|
.072
|
|
td
|
.003
|
|
td
|
.019
|
|
td
|
.125
|
|
td
|
.029
|
|
td
|
.023
|
|
td
|
.190
|
|
td
|
−.030
|
|
td
|
.076
|
|
td
|
.009
|
|
tr
|
MN→Int.→Beh. .102 <.001 .051 .153 .044 .040 .067 −.013 .093 .017
|
|
td
|
MN→Int.→Beh.
|
|
td
|
.102
|
|
td
|
<.001
|
|
td
|
.051
|
|
td
|
.153
|
|
td
|
.044
|
|
td
|
.040
|
|
td
|
.067
|
|
td
|
−.013
|
|
td
|
.093
|
|
td
|
.017
|
|
tr
|
AR→Int.→Beh. .023 .192 −.030 .076 .011 .001 .478 −.052 .054 .001
|
|
td
|
AR→Int.→Beh.
|
|
td
|
.023
|
|
td
|
.192
|
|
td
|
−.030
|
|
td
|
.076
|
|
td
|
.011
|
|
td
|
.001
|
|
td
|
.478
|
|
td
|
−.052
|
|
td
|
.054
|
|
td
|
.001
|
|
tr
|
PBC→Int.→Beh. .088 <.001 .037 .139 .025 .038 .079 −.015 .091 .011
|
|
td
|
PBC→Int.→Beh.
|
|
td
|
.088
|
|
td
|
<.001
|
|
td
|
.037
|
|
td
|
.139
|
|
td
|
.025
|
|
td
|
.038
|
|
td
|
.079
|
|
td
|
−.015
|
|
td
|
.091
|
|
td
|
.011
|
|
tr
|
Int.→AP→Beh. .061 .011 .008 .114 .029 .004 .441 −.049 .057 .002
|
|
td
|
Int.→AP→Beh.
|
|
td
|
.061
|
|
td
|
.011
|
|
td
|
.008
|
|
td
|
.114
|
|
td
|
.029
|
|
td
|
.004
|
|
td
|
.441
|
|
td
|
−.049
|
|
td
|
.057
|
|
td
|
.002
|
|
tr
|
Hab. (T1).→Hab. (T2)→Beh. .212 <.001 .161 .263 .075 .166 <.001 .115 .217 .059
|
|
td
|
Hab. (T1).→Hab. (T2)→Beh.
|
|
td
|
.212
|
|
td
|
<.001
|
|
td
|
.161
|
|
td
|
.263
|
|
td
|
.075
|
|
td
|
.166
|
|
td
|
<.001
|
|
td
|
.115
|
|
td
|
.217
|
|
td
|
.059
|
|
tr
|
PB→Hab.→Beh. – – – – – .068 <.001 .025 .111 .043
|
|
td
|
PB→Hab.→Beh.
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
.068
|
|
td
|
<.001
|
|
td
|
.025
|
|
td
|
.111
|
|
td
|
.043
|
|
tr
|
PB→Beh.a – – – – – .178 <.001 .105 .251 .112
|
|
td
|
PB→Beh.a
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
–
|
|
td
|
.178
|
|
td
|
<.001
|
|
td
|
.105
|
|
td
|
.251
|
|
td
|
.112
|
|
tr
|
Total effectsb
|
|
td
|
Total effectsb
|
|
tr
|
Int.→Beh. .377 <.001 .306 .448 .177 .142 <.001 .069 .215 .066
|
|
td
|
Int.→Beh.
|
|
td
|
.377
|
|
td
|
<.001
|
|
td
|
.306
|
|
td
|
.448
|
|
td
|
.177
|
|
td
|
.142
|
|
td
|
<.001
|
|
td
|
.069
|
|
td
|
.215
|
|
td
|
.066
|
|
tr
|
PBC→Beh. .146 <.001 .073 .219 .042 .074 .024 .001 .147 .021
|
|
td
|
PBC→Beh.
|
|
td
|
.146
|
|
td
|
<.001
|
|
td
|
.073
|
|
td
|
.219
|
|
td
|
.042
|
|
td
|
.074
|
|
td
|
.024
|
|
td
|
.001
|
|
td
|
.147
|
|
td
|
.021
|
|
tr
|
Hab. (T1)→Beh. .242 <.001 .169 .315 .086 .171 <.001 .098 .244 .061
|
|
td
|
Hab. (T1)→Beh.
|
|
td
|
.242
|
|
td
|
<.001
|
|
td
|
.169
|
|
td
|
.315
|
|
td
|
.086
|
|
td
|
.171
|
|
td
|
<.001
|
|
td
|
.098
|
|
td
|
.244
|
|
td
|
.061
|
|
tr
|
PB→Beh. – – −.052 .074 – .673 <.001 .604 .742 .423
|
|
td
|
PB→Beh.
|
|
td
|
–
|
|
td
|
–
|
|
td
|
−.052
|
|
td
|
.074
|
|
td
|
–
|
|
td
|
.673
|
|
td
|
<.001
|
|
td
|
.604
|
|
td
|
.742
|
|
td
|
.423
|
|
table-wrap-foot
|
aSum of indirect effects of past behavior on behavior through all model constructs.
bTotal effect comprising sums of all indirect effects through model constructs plus the direct effect.
β standardized parameter estimate; 95% CI 95% confidence interval of standardized parameter estimate; AP action planning; AR anticipated regret; Att. attitude; Beh. behavior; ES effect size of the standardized parameter estimate; Hab. (T1) self-reported habit measured at baseline (T1); Hab. (T2) self-reported habit measured at follow-up (T2); Int. intention; LB lower bound of 95% CI; MN moral norm; PB past behavior; PBC perceived behavioral control; SN subjective norm; UB upper bound of 95% CI.
|
|
footnote
|
aSum of indirect effects of past behavior on behavior through all model constructs.
|
|
p
|
aSum of indirect effects of past behavior on behavior through all model constructs.
|
|
footnote
|
bTotal effect comprising sums of all indirect effects through model constructs plus the direct effect.
|
|
p
|
bTotal effect comprising sums of all indirect effects through model constructs plus the direct effect.
|
|
footnote
|
β standardized parameter estimate; 95% CI 95% confidence interval of standardized parameter estimate; AP action planning; AR anticipated regret; Att. attitude; Beh. behavior; ES effect size of the standardized parameter estimate; Hab. (T1) self-reported habit measured at baseline (T1); Hab. (T2) self-reported habit measured at follow-up (T2); Int. intention; LB lower bound of 95% CI; MN moral norm; PB past behavior; PBC perceived behavioral control; SN subjective norm; UB upper bound of 95% CI.
|
|
p
|
β standardized parameter estimate; 95% CI 95% confidence interval of standardized parameter estimate; AP action planning; AR anticipated regret; Att. attitude; Beh. behavior; ES effect size of the standardized parameter estimate; Hab. (T1) self-reported habit measured at baseline (T1); Hab. (T2) self-reported habit measured at follow-up (T2); Int. intention; LB lower bound of 95% CI; MN moral norm; PB past behavior; PBC perceived behavioral control; SN subjective norm; UB upper bound of 95% CI.
|
|
p
|
Focusing on the model excluding past behavior (Model 1), intention, action planning, and habit at follow-up were statistically significant direct predictors of social distancing behavior, with effect size for intention and habit generally larger in the U.S. sample. PBC directly predicted behavior in the Australian sample only, also with a small effect size. Intention predicted action planning in both samples with large effect sizes. Subjective norm, moral norm, and PBC predicted intention in both samples, with small-to-medium effect sizes, but effects of attitude were not significant. There was a small effect of anticipated regret on intention in the U.S. sample only. Habit at baseline predicted habit at follow-up in both samples, with large effect sizes. There was also a small-sized effect of habit at baseline on intention in the U.S. sample only. Overall, the model accounted for significant variance in social distancing behavior (Australian sample, R2 = .198; U.S. sample, R2 = .361), intentions (Australian sample, R2 = .571; U.S. sample, R2 = .623), and habit at follow-up (Australian sample, R2 = .416; U.S. sample, R2 = .486). Intentions (Australian sample, R2 = .066; U.S. sample, R2 = .148), action planning (Australian sample, R2 = .029; U.S. sample, R2 = .044), and habit at follow-up (Australian sample R2 = .041; U.S. sample, R2 = .129) each accounted for substantive variance in behavior. Action planning significantly moderated the intention–behavior relationship in the Australian sample only. While the effect was not in the predicted direction, probing the interaction revealed that the intention–behavior relationship increased as the level of planning increased, consistent with theory. However, the intention–behavior relationship is more likely to be smaller at lower levels of planning, and it seems that planning makes less of a difference when the intention–behavior relationship is large. A plot of the interaction effect is presented in Supplementary Appendix F.
|
|
p
|
Turning to the indirect effects, there were significant indirect effects of subjective norm, moral norm, and PBC on social distancing behavior mediated by intention in the U.S. sample. By contrast, only the indirect effect of moral norm on behavior through intention was significant in the Australian sample. The smaller indirect effects in the Australian sample is principally due to the significantly smaller effect size for the intention–behavior relationship in this sample compared to the U.S. sample. Habit at baseline predicted behavior through habit at follow-up in both samples. Effect sizes in all cases were small. There were significant total effects of intention, PBC, and habit at baseline on behavior, with effect sizes larger in the U.S. sample than in the Australian sample.
|
|
p
|
For the model including past behavior, significant effects of past behavior on all model constructs were observed in both samples with effect sizes ranging from small to large. The effects of past behavior on social distancing behavior were particularly large. Inclusion of past behavior led to an attenuation of model effects in both samples. Specifically, the effects of intention and habit at follow-up on behavior were reduced but remained statistically significant in both samples with small effect sizes. In addition, effects of subjective norm, moral norm, and PBC on intention, and the effect of intention on action planning, remained statistically significant in both samples, with small-to-medium effect sizes. The effect of habit at baseline on habit at follow-up was statistically significant in both samples, with large effect sizes. Variance explained in social distancing behavior increased substantially with the inclusion of past behavior, with only modest changes in explained variance in intentions (Australian sample R2 = .598; U.S. sample, R2 = .702) and habit at follow-up (Australian sample R2 = .416; U.S. sample, R2 = .486). Specifically, intentions (Australian sample, R2 = .029; U.S. sample, R2 = .065), past behavior (Australian sample, R2 = .216; U.S. sample, R2 = .311), and habit at follow-up (Australian sample R2 = .031; U.S. sample, R2 = .101) each accounted for substantive variance in behavior.
|
|
p
|
Turning to indirect effects, we found significant indirect effects of habit at baseline on behavior mediated by habit at follow-up in both samples with small effect sizes. There were also significant total effects of intention and habit at baseline on behavior in both samples, and of PBC on behavior for the U.S. sample, with small effect sizes. There were significant total indirect and total effects of past behavior on behavior in both samples, with large effect sizes. There was a small-sized indirect effect of past behavior on behavior mediated by habit at both time points in the U.S. sample, but the effect was not significant in the Australian sample.
|
|
sec
|
Multisample analyses
Multisample analyses permitted for tests of difference in parameter estimates for each model across the Australian and U.S. samples. For the model excluding past behavior (Model 1), only effects of intention on habit at baseline, habit at follow-up on social distancing behavior, and intention on action planning differed across samples. These effects were significantly larger in the U.S. sample. Some effects with observed differences across samples, such as the effect of habit at baseline on intention or the moderator effect of planning on the intention–behavior relationship, did not differ significantly across samples. For the model including effects of past behavior (Model 2), multisample analysis revealed no differences in effect size across samples, indicating that the attenuating effect of past behavior on model effects also had the effect of eliminating the few differences in model effects across samples. Full details of the multiple-group analysis are presented in Supplementary Appendix G.
|
|
title
|
Multisample analyses
|
|
p
|
Multisample analyses permitted for tests of difference in parameter estimates for each model across the Australian and U.S. samples. For the model excluding past behavior (Model 1), only effects of intention on habit at baseline, habit at follow-up on social distancing behavior, and intention on action planning differed across samples. These effects were significantly larger in the U.S. sample. Some effects with observed differences across samples, such as the effect of habit at baseline on intention or the moderator effect of planning on the intention–behavior relationship, did not differ significantly across samples. For the model including effects of past behavior (Model 2), multisample analysis revealed no differences in effect size across samples, indicating that the attenuating effect of past behavior on model effects also had the effect of eliminating the few differences in model effects across samples. Full details of the multiple-group analysis are presented in Supplementary Appendix G.
|
|
sec
|
Sensitivity analyses
We re-estimated both models in samples excluding participants who were never subject to a “shelter-in place” order, and in the sample that were never subject to an order, or who had the order lifted at some stage during the study. Comparisons of parameter estimates in these models with those from the models estimated in the full sample revealed no significant differences in any of the model parameters. Results are reported in Supplementary Appendices H and I.
|
|
title
|
Sensitivity analyses
|
|
p
|
We re-estimated both models in samples excluding participants who were never subject to a “shelter-in place” order, and in the sample that were never subject to an order, or who had the order lifted at some stage during the study. Comparisons of parameter estimates in these models with those from the models estimated in the full sample revealed no significant differences in any of the model parameters. Results are reported in Supplementary Appendices H and I.
|
|
sec
|
Discussion
The present study aimed to identify the determinants of social distancing behavior in the context of COVID-19 through the application of an integrated social cognition model. The integrated model was based on the theory of planned behavior [13] augmented to include additional predictors relating to normative (moral norm), anticipated affect (anticipated regret), volitional (action planning), and nonconscious (habit) determinants of health behavior. The model was tested in data from a correlational prospective survey study in two samples of Australian and U.S. residents subject to national or local “shelter-in-place” orders. Results indicated that intention and habit were significant predictors of social distancing behavior in both samples. Subjective norm, moral norm, and PBC were significant predictors of social distancing intention. In addition, intention-mediated effects of these social cognition constructs on social distancing behavior in the U.S. sample, but did so only for moral norm in the Australian sample. Action planning did not mediate effects of intention on behavior in either samples but moderated the intention–behavior relationship in the Australia sample. Inclusion of past behavior attenuated effects of social cognition constructs in the models in both samples, although habit and intention remained significant determinants of social distancing behavior in both samples. Excluding participants in the U.S. sample not subject to formal “shelter-in-place” orders, or had the orders lifted during the study, did not affect the pattern or size of the effects in the model, providing evidence that formal orders did not have a substantive bearing on the determinants of social distancing behavior in this sample.
Current findings provide qualified support for some, but not all, predictions of the integrated social cognition model for social distancing behavior. A key assumption of the model, derived from the social cognition theories on which it is based, is that social distancing behavior is reasoned action and, therefore, determined predominantly by intention and the belief-based constructs that underpin them. Effects of intention on social distancing behavior and its mediation of constructs reflecting social reasons for acting, particularly beliefs relating to significant others and moral obligations to perform the behavior, and PBC is consistent with this assumption. This is unsurprising in this context, considering the widely publicized details of the relatively mild effects of the virus in the majority of the population. It is likely that the majority of individuals do not view themselves as at serious risk from COVID-19 but have internalized the view that significant others want them to engage in social distancing and feel a moral obligation to perform the behavior to protect the health of those most at risk. Such a finding is consistent with research on similar health behaviors, such as blood donation, where behavioral performance is likely to promote the health of others rather than the self [34]. Similarly, the impact of PBC indicates the importance of perceived personal agency in maintaining social distancing behavior, consistent with previous research on health behaviors [14]. Individuals that see fewer barriers to maintaining social distancing and have the confidence to do so are more likely to intend to perform these behaviors.
The effects of subjective and moral norms and PBC suggests that these should be viable targets for behavioral interventions aimed at promoting social distancing behavior based on the model. For example, messages promoting moral obligation (e.g., highlighting social responsibility for preventing transmission of the virus to vulnerable others through social distancing) and perceived control (e.g., demonstrating how to easily and successfully maintain appropriate social distance) may facilitate greater intention to socially distance. However, the intention–behavior relationship in the present study was relatively modest in size, particularly in the Australian sample, indicative of a substantive intention–behavior “gap” [15]. This suggests that interventions targeting change in intention determinants, such as moral norms and PBC, may have only small effects on social distancing behavior. It may be of value to explore how properties of intention may affect intention–behavior relations in the context of social distancing behavior [35]. Such properties may signal potential intervention strategies that may strengthen intention–behavior relations in conjunction with messaging targeting moral norms and PBC.
Current findings also indicated consistent effects of self-reported habits on social distancing behavior. Importantly, the effects of habit were direct and independent of intentions, consistent with the theory that suggests that effects of habits reflect nonconscious, automatic processes developed through consistent experience with the behavior in stable contexts over time. Habits also partially mediated the effects of past behavior on social distancing behavior, suggesting that past behavior effects, at least in part, reflect habits [27]. An implication of these findings is that facilitating habit development in behavioral interventions may be effective in promoting social distancing. Research suggests that strategies, such as providing successful experiences of the desired behavior consistently over time and creating environment conditions that facilitate the behavior (e.g., consistent reminders and environmental restructuring) are effective in inducing habits [36], but the efficacy of such strategies in the context of social distancing behavior need to be verified empirically. Furthermore, legislation restricting or mandating behavior change facilitates habit formation over time. This suggests that the introduction of “shelter-in-place” and other government-mandated restrictions may facilitate social distancing habits.
Inclusion of past behavior as a predictor of social distancing behavior at follow-up reduced the effects of intention on behavior to a trivial size in both samples and also attenuated the effects of the social cognition constructs on intention. Such effects are consistent with previous research [22] and raise questions over the sufficiency of the model in identifying the determinants of social distancing behavior. However, such findings must be interpreted in light of the current study design and how the effects of past behavior can provide important information on the determinants of social distancing behavior. The 1 week time lag means that past behavior was always likely to have a large effect because individuals’ behavior tends to be relatively stable over short periods [22]. A more complete evaluation of model sufficiency would be afforded by testing its long-range prediction, which has often been cited as a goal of social cognition theories [14], and should be considered a future research priority for research on social distancing behavior. However, past behavior effects can be informative on the determinants of social distancing behavior as it may reflect the effects of other unmeasured behavioral determinants. In particular, past behavior will likely reflect determinants that bypass the reasoned, intention-mediated processes that lead to behaviors, such as implicit attitudes and motives, personality traits, and variables reflecting the social and physical environment. The effects of such constructs are speculative and future tests of the integrated model that incorporate such factors alongside those from the current model may assist in resolving these effects.
Consistent with dual-phase models [18, 19], we also tested the extent to which action planning was implicated in the process by which individuals act on their intention. Two patterns of effects were tested: mediation and moderation effects of action planning on the intention–behavior relationship. The mediation effect was significant in the U.S. sample but not the Australia sample, while the moderation effect was significant in the Australia sample only. However, in both cases, the effects were small in size. The small size of the mediation effects suggests that action planning is a relatively trivial component of the link between social distancing intention and behavior, particularly when past behavior was taken into account. The moderation of the intention–behavior relationship by action planning in the Australian sample was negative in sign, which is contrary to predictions [18]. However, probing this interaction indicated that individuals with stronger intention were more likely to follow through on their social distancing behavior at both high and low levels of action planning, but the rate of increase was much steeper for low planning, which supports the prediction. However, when the intention–behavior relationship was strongest, planning had little effect, so planning may only be effective for those with lower intentions. As with the mediation effect, the moderation effect was no longer present once past behavior was included in the model. Taken together, current results do not provide strong evidence for the role of action planning in mediating and moderating the intention–behavior relationship for social distancing.
Limitations and Avenues for Future Research
Current findings should be interpreted in light of some notable limitations. First, attrition rates in both samples were relatively high given the relatively brief time between the baseline survey and follow-up. High attrition could lead to selection bias with those who are more motivated or engaged overrepresented in the sample. While participants were reminded multiple times to complete follow-up measures, we acknowledge that more intensive recruitment and incentivization of nonresponders may have minimized drop out. Attrition also affected the demographic profile of the sample, particularly among underrepresented groups. Although the effect sizes of these differences were small, they were not trivial. This is particularly pertinent in the current context given emerging data indicating that COVID-19 infection and mortality rates are significantly higher in underrepresented minority and socioeconomic groups [37]. A potential solution would be to oversample in underrepresented groups likely to have low retention rates and is a recommendation for future research. It is also important to note that, although our sampling strategy ensured that the distribution of participants in our samples matched those of the national population according to gender and state, we did not stratify the sample by key demographic or socioeconomic factors. The samples, therefore, should not be considered representative of the national populations of Australia or the USA. Taken together, the bias linked to attrition rates and nonrepresentativeness of the samples places limits on the extent to which current findings can be generalized to the broader population.
Second, the intention–behavior “gap” in the current study resulted in small indirect effects of intention determinants, such as subjective and moral norms and PBC on social distancing behavior. This is a limitation of the current model and means that intervention strategies aimed at changing intention determinants may have relatively modest effects on behavior change. However, small effects may still translate to large numbers of people changing if interventions targeting change in these constructs are administered at the population level. Future intervention research is, nevertheless, needed to verify the effects of targeting change in model constructs on behavior. Research should also adopt behavioral measures that can be converted to meaningful metrics that demonstrate practically significant changes in social distancing behavior (e.g., numbers of people complying with social distancing guidelines when venturing outside the home).
Third, the current study observed social distancing over a relatively brief time frame. Short-range prediction has value as it helps identify potential determinants of social distancing behavior. However, consistency in performing social distancing over time is important for the effective prevention of virus transmission, so research on the determinants of social distancing in the long term is a priority. The relatively short time lag is also likely to be the reason why past behavior had such a pervasive effect in predicting behavior and other constructs in the model. The relevance of past behavior is likely to wane over time, so examining prediction over time may be more revealing as to the social cognition predictors of this behavior and the processes involved.
Fourth, the correlational design precludes the inference of causal effects among the constructs in the current model, so the proposed direction of effects are inferred from theory alone, not the data. Causal sequencing among variables would necessitate experimental or controlled intervention designs. Verification of such effects will highlight the value of the model in informing interventions to promote changes in social distancing behavior. In addition, the inclusion of past behavior in the current analysis modeled change in behavior over time. Past behavior also had the effect of modeling residual effects of unmeasured constructs on behavior, such as past measures of the model constructs. However, the adoption of a cross-lagged panel design would better facilitate the examination of how the change in specific model constructs over time affects social distancing behavior and permit tests of reciprocal effects. It is also important that the effects of past behavior do not provide definitive evidence that affecting change in model constructs, such as intentions or habit, through intervention will lead to a concomitant change in social distancing behavior. This highlights the imperative of intervention research that tests the efficacy of manipulating constructs from the current model in promoting social distancing behavior and illustrates the extent to which model constructs can be modified.
Finally, the current research relies exclusively on self-report measures. While self-reported behavior has exhibited concurrent validity when evaluated against non-self-report measures, such as behavior measured using devices or direct observation, the potential for recall bias or inaccurate reporting likely introduces additional measurement error in the behavioral measure, which would affect model relations. Further, self-reported data are also at risk of self-presentation bias and socially desirable responding. Health behaviors, particularly social distancing behavior in the context of a pandemic, are likely to be considered desirable, which may have compelled respondents to provide positive responses, without even being aware of such biases. Although we stressed anonymity to participants to make it clear that they had license to report their behavior without prejudice, this is unlikely to have fully eliminated such biases. Current data should, therefore, be interpreted in light of these potential biases and their potential to contribute to error variance in observed effects. Future research may consider the use of devices, such as GPS tracking of cellular phones, as alternative means to measure social distancing behavior that do not rely on self-report.
|
|
title
|
Discussion
|
|
p
|
The present study aimed to identify the determinants of social distancing behavior in the context of COVID-19 through the application of an integrated social cognition model. The integrated model was based on the theory of planned behavior [13] augmented to include additional predictors relating to normative (moral norm), anticipated affect (anticipated regret), volitional (action planning), and nonconscious (habit) determinants of health behavior. The model was tested in data from a correlational prospective survey study in two samples of Australian and U.S. residents subject to national or local “shelter-in-place” orders. Results indicated that intention and habit were significant predictors of social distancing behavior in both samples. Subjective norm, moral norm, and PBC were significant predictors of social distancing intention. In addition, intention-mediated effects of these social cognition constructs on social distancing behavior in the U.S. sample, but did so only for moral norm in the Australian sample. Action planning did not mediate effects of intention on behavior in either samples but moderated the intention–behavior relationship in the Australia sample. Inclusion of past behavior attenuated effects of social cognition constructs in the models in both samples, although habit and intention remained significant determinants of social distancing behavior in both samples. Excluding participants in the U.S. sample not subject to formal “shelter-in-place” orders, or had the orders lifted during the study, did not affect the pattern or size of the effects in the model, providing evidence that formal orders did not have a substantive bearing on the determinants of social distancing behavior in this sample.
|
|
p
|
Current findings provide qualified support for some, but not all, predictions of the integrated social cognition model for social distancing behavior. A key assumption of the model, derived from the social cognition theories on which it is based, is that social distancing behavior is reasoned action and, therefore, determined predominantly by intention and the belief-based constructs that underpin them. Effects of intention on social distancing behavior and its mediation of constructs reflecting social reasons for acting, particularly beliefs relating to significant others and moral obligations to perform the behavior, and PBC is consistent with this assumption. This is unsurprising in this context, considering the widely publicized details of the relatively mild effects of the virus in the majority of the population. It is likely that the majority of individuals do not view themselves as at serious risk from COVID-19 but have internalized the view that significant others want them to engage in social distancing and feel a moral obligation to perform the behavior to protect the health of those most at risk. Such a finding is consistent with research on similar health behaviors, such as blood donation, where behavioral performance is likely to promote the health of others rather than the self [34]. Similarly, the impact of PBC indicates the importance of perceived personal agency in maintaining social distancing behavior, consistent with previous research on health behaviors [14]. Individuals that see fewer barriers to maintaining social distancing and have the confidence to do so are more likely to intend to perform these behaviors.
|
|
p
|
The effects of subjective and moral norms and PBC suggests that these should be viable targets for behavioral interventions aimed at promoting social distancing behavior based on the model. For example, messages promoting moral obligation (e.g., highlighting social responsibility for preventing transmission of the virus to vulnerable others through social distancing) and perceived control (e.g., demonstrating how to easily and successfully maintain appropriate social distance) may facilitate greater intention to socially distance. However, the intention–behavior relationship in the present study was relatively modest in size, particularly in the Australian sample, indicative of a substantive intention–behavior “gap” [15]. This suggests that interventions targeting change in intention determinants, such as moral norms and PBC, may have only small effects on social distancing behavior. It may be of value to explore how properties of intention may affect intention–behavior relations in the context of social distancing behavior [35]. Such properties may signal potential intervention strategies that may strengthen intention–behavior relations in conjunction with messaging targeting moral norms and PBC.
|
|
p
|
Current findings also indicated consistent effects of self-reported habits on social distancing behavior. Importantly, the effects of habit were direct and independent of intentions, consistent with the theory that suggests that effects of habits reflect nonconscious, automatic processes developed through consistent experience with the behavior in stable contexts over time. Habits also partially mediated the effects of past behavior on social distancing behavior, suggesting that past behavior effects, at least in part, reflect habits [27]. An implication of these findings is that facilitating habit development in behavioral interventions may be effective in promoting social distancing. Research suggests that strategies, such as providing successful experiences of the desired behavior consistently over time and creating environment conditions that facilitate the behavior (e.g., consistent reminders and environmental restructuring) are effective in inducing habits [36], but the efficacy of such strategies in the context of social distancing behavior need to be verified empirically. Furthermore, legislation restricting or mandating behavior change facilitates habit formation over time. This suggests that the introduction of “shelter-in-place” and other government-mandated restrictions may facilitate social distancing habits.
|
|
p
|
Inclusion of past behavior as a predictor of social distancing behavior at follow-up reduced the effects of intention on behavior to a trivial size in both samples and also attenuated the effects of the social cognition constructs on intention. Such effects are consistent with previous research [22] and raise questions over the sufficiency of the model in identifying the determinants of social distancing behavior. However, such findings must be interpreted in light of the current study design and how the effects of past behavior can provide important information on the determinants of social distancing behavior. The 1 week time lag means that past behavior was always likely to have a large effect because individuals’ behavior tends to be relatively stable over short periods [22]. A more complete evaluation of model sufficiency would be afforded by testing its long-range prediction, which has often been cited as a goal of social cognition theories [14], and should be considered a future research priority for research on social distancing behavior. However, past behavior effects can be informative on the determinants of social distancing behavior as it may reflect the effects of other unmeasured behavioral determinants. In particular, past behavior will likely reflect determinants that bypass the reasoned, intention-mediated processes that lead to behaviors, such as implicit attitudes and motives, personality traits, and variables reflecting the social and physical environment. The effects of such constructs are speculative and future tests of the integrated model that incorporate such factors alongside those from the current model may assist in resolving these effects.
|
|
p
|
Consistent with dual-phase models [18, 19], we also tested the extent to which action planning was implicated in the process by which individuals act on their intention. Two patterns of effects were tested: mediation and moderation effects of action planning on the intention–behavior relationship. The mediation effect was significant in the U.S. sample but not the Australia sample, while the moderation effect was significant in the Australia sample only. However, in both cases, the effects were small in size. The small size of the mediation effects suggests that action planning is a relatively trivial component of the link between social distancing intention and behavior, particularly when past behavior was taken into account. The moderation of the intention–behavior relationship by action planning in the Australian sample was negative in sign, which is contrary to predictions [18]. However, probing this interaction indicated that individuals with stronger intention were more likely to follow through on their social distancing behavior at both high and low levels of action planning, but the rate of increase was much steeper for low planning, which supports the prediction. However, when the intention–behavior relationship was strongest, planning had little effect, so planning may only be effective for those with lower intentions. As with the mediation effect, the moderation effect was no longer present once past behavior was included in the model. Taken together, current results do not provide strong evidence for the role of action planning in mediating and moderating the intention–behavior relationship for social distancing.
|
|
sec
|
Limitations and Avenues for Future Research
Current findings should be interpreted in light of some notable limitations. First, attrition rates in both samples were relatively high given the relatively brief time between the baseline survey and follow-up. High attrition could lead to selection bias with those who are more motivated or engaged overrepresented in the sample. While participants were reminded multiple times to complete follow-up measures, we acknowledge that more intensive recruitment and incentivization of nonresponders may have minimized drop out. Attrition also affected the demographic profile of the sample, particularly among underrepresented groups. Although the effect sizes of these differences were small, they were not trivial. This is particularly pertinent in the current context given emerging data indicating that COVID-19 infection and mortality rates are significantly higher in underrepresented minority and socioeconomic groups [37]. A potential solution would be to oversample in underrepresented groups likely to have low retention rates and is a recommendation for future research. It is also important to note that, although our sampling strategy ensured that the distribution of participants in our samples matched those of the national population according to gender and state, we did not stratify the sample by key demographic or socioeconomic factors. The samples, therefore, should not be considered representative of the national populations of Australia or the USA. Taken together, the bias linked to attrition rates and nonrepresentativeness of the samples places limits on the extent to which current findings can be generalized to the broader population.
Second, the intention–behavior “gap” in the current study resulted in small indirect effects of intention determinants, such as subjective and moral norms and PBC on social distancing behavior. This is a limitation of the current model and means that intervention strategies aimed at changing intention determinants may have relatively modest effects on behavior change. However, small effects may still translate to large numbers of people changing if interventions targeting change in these constructs are administered at the population level. Future intervention research is, nevertheless, needed to verify the effects of targeting change in model constructs on behavior. Research should also adopt behavioral measures that can be converted to meaningful metrics that demonstrate practically significant changes in social distancing behavior (e.g., numbers of people complying with social distancing guidelines when venturing outside the home).
Third, the current study observed social distancing over a relatively brief time frame. Short-range prediction has value as it helps identify potential determinants of social distancing behavior. However, consistency in performing social distancing over time is important for the effective prevention of virus transmission, so research on the determinants of social distancing in the long term is a priority. The relatively short time lag is also likely to be the reason why past behavior had such a pervasive effect in predicting behavior and other constructs in the model. The relevance of past behavior is likely to wane over time, so examining prediction over time may be more revealing as to the social cognition predictors of this behavior and the processes involved.
Fourth, the correlational design precludes the inference of causal effects among the constructs in the current model, so the proposed direction of effects are inferred from theory alone, not the data. Causal sequencing among variables would necessitate experimental or controlled intervention designs. Verification of such effects will highlight the value of the model in informing interventions to promote changes in social distancing behavior. In addition, the inclusion of past behavior in the current analysis modeled change in behavior over time. Past behavior also had the effect of modeling residual effects of unmeasured constructs on behavior, such as past measures of the model constructs. However, the adoption of a cross-lagged panel design would better facilitate the examination of how the change in specific model constructs over time affects social distancing behavior and permit tests of reciprocal effects. It is also important that the effects of past behavior do not provide definitive evidence that affecting change in model constructs, such as intentions or habit, through intervention will lead to a concomitant change in social distancing behavior. This highlights the imperative of intervention research that tests the efficacy of manipulating constructs from the current model in promoting social distancing behavior and illustrates the extent to which model constructs can be modified.
Finally, the current research relies exclusively on self-report measures. While self-reported behavior has exhibited concurrent validity when evaluated against non-self-report measures, such as behavior measured using devices or direct observation, the potential for recall bias or inaccurate reporting likely introduces additional measurement error in the behavioral measure, which would affect model relations. Further, self-reported data are also at risk of self-presentation bias and socially desirable responding. Health behaviors, particularly social distancing behavior in the context of a pandemic, are likely to be considered desirable, which may have compelled respondents to provide positive responses, without even being aware of such biases. Although we stressed anonymity to participants to make it clear that they had license to report their behavior without prejudice, this is unlikely to have fully eliminated such biases. Current data should, therefore, be interpreted in light of these potential biases and their potential to contribute to error variance in observed effects. Future research may consider the use of devices, such as GPS tracking of cellular phones, as alternative means to measure social distancing behavior that do not rely on self-report.
|
|
title
|
Limitations and Avenues for Future Research
|
|
p
|
Current findings should be interpreted in light of some notable limitations. First, attrition rates in both samples were relatively high given the relatively brief time between the baseline survey and follow-up. High attrition could lead to selection bias with those who are more motivated or engaged overrepresented in the sample. While participants were reminded multiple times to complete follow-up measures, we acknowledge that more intensive recruitment and incentivization of nonresponders may have minimized drop out. Attrition also affected the demographic profile of the sample, particularly among underrepresented groups. Although the effect sizes of these differences were small, they were not trivial. This is particularly pertinent in the current context given emerging data indicating that COVID-19 infection and mortality rates are significantly higher in underrepresented minority and socioeconomic groups [37]. A potential solution would be to oversample in underrepresented groups likely to have low retention rates and is a recommendation for future research. It is also important to note that, although our sampling strategy ensured that the distribution of participants in our samples matched those of the national population according to gender and state, we did not stratify the sample by key demographic or socioeconomic factors. The samples, therefore, should not be considered representative of the national populations of Australia or the USA. Taken together, the bias linked to attrition rates and nonrepresentativeness of the samples places limits on the extent to which current findings can be generalized to the broader population.
|
|
p
|
Second, the intention–behavior “gap” in the current study resulted in small indirect effects of intention determinants, such as subjective and moral norms and PBC on social distancing behavior. This is a limitation of the current model and means that intervention strategies aimed at changing intention determinants may have relatively modest effects on behavior change. However, small effects may still translate to large numbers of people changing if interventions targeting change in these constructs are administered at the population level. Future intervention research is, nevertheless, needed to verify the effects of targeting change in model constructs on behavior. Research should also adopt behavioral measures that can be converted to meaningful metrics that demonstrate practically significant changes in social distancing behavior (e.g., numbers of people complying with social distancing guidelines when venturing outside the home).
|
|
p
|
Third, the current study observed social distancing over a relatively brief time frame. Short-range prediction has value as it helps identify potential determinants of social distancing behavior. However, consistency in performing social distancing over time is important for the effective prevention of virus transmission, so research on the determinants of social distancing in the long term is a priority. The relatively short time lag is also likely to be the reason why past behavior had such a pervasive effect in predicting behavior and other constructs in the model. The relevance of past behavior is likely to wane over time, so examining prediction over time may be more revealing as to the social cognition predictors of this behavior and the processes involved.
|
|
p
|
Fourth, the correlational design precludes the inference of causal effects among the constructs in the current model, so the proposed direction of effects are inferred from theory alone, not the data. Causal sequencing among variables would necessitate experimental or controlled intervention designs. Verification of such effects will highlight the value of the model in informing interventions to promote changes in social distancing behavior. In addition, the inclusion of past behavior in the current analysis modeled change in behavior over time. Past behavior also had the effect of modeling residual effects of unmeasured constructs on behavior, such as past measures of the model constructs. However, the adoption of a cross-lagged panel design would better facilitate the examination of how the change in specific model constructs over time affects social distancing behavior and permit tests of reciprocal effects. It is also important that the effects of past behavior do not provide definitive evidence that affecting change in model constructs, such as intentions or habit, through intervention will lead to a concomitant change in social distancing behavior. This highlights the imperative of intervention research that tests the efficacy of manipulating constructs from the current model in promoting social distancing behavior and illustrates the extent to which model constructs can be modified.
|
|
p
|
Finally, the current research relies exclusively on self-report measures. While self-reported behavior has exhibited concurrent validity when evaluated against non-self-report measures, such as behavior measured using devices or direct observation, the potential for recall bias or inaccurate reporting likely introduces additional measurement error in the behavioral measure, which would affect model relations. Further, self-reported data are also at risk of self-presentation bias and socially desirable responding. Health behaviors, particularly social distancing behavior in the context of a pandemic, are likely to be considered desirable, which may have compelled respondents to provide positive responses, without even being aware of such biases. Although we stressed anonymity to participants to make it clear that they had license to report their behavior without prejudice, this is unlikely to have fully eliminated such biases. Current data should, therefore, be interpreted in light of these potential biases and their potential to contribute to error variance in observed effects. Future research may consider the use of devices, such as GPS tracking of cellular phones, as alternative means to measure social distancing behavior that do not rely on self-report.
|
|
sec
|
Conclusion
The current research aimed to identify the determinants of social distancing behavior to prevent transmission of the SARS-CoV-2 virus in samples of Australian and U.S. residents. The research applied an integrated theoretical model that included multiple social cognition determinants relevant to the behavioral context, and the processes involved, with the potential to be modifiable through intervention. Results provided qualified support for the proposed model, highlighting the importance of social and moral beliefs, and perceptions of control, in predicting intention, and habit and intention in predicting behavior, in both samples, although effects were relatively modest, particularly when past behavior was accounted for. Findings suggest that interventions aimed at promoting social distancing behavior should provide messages highlighting individuals’ obligations to significant others and the moral imperative of protecting the most vulnerable as reasons for social distancing, provide environments (e.g., workplaces and grocery stores) that are barrier free and easy to socially distance and provide consistent opportunities in regular, stable contexts to engage in social distancing to develop habits. Future research should seek to provide longer-range prediction of social distancing behavior by the integrated model constructs and test the stability and reciprocal relations among its constructs using a cross-lagged panel design.
|
|
title
|
Conclusion
|
|
p
|
The current research aimed to identify the determinants of social distancing behavior to prevent transmission of the SARS-CoV-2 virus in samples of Australian and U.S. residents. The research applied an integrated theoretical model that included multiple social cognition determinants relevant to the behavioral context, and the processes involved, with the potential to be modifiable through intervention. Results provided qualified support for the proposed model, highlighting the importance of social and moral beliefs, and perceptions of control, in predicting intention, and habit and intention in predicting behavior, in both samples, although effects were relatively modest, particularly when past behavior was accounted for. Findings suggest that interventions aimed at promoting social distancing behavior should provide messages highlighting individuals’ obligations to significant others and the moral imperative of protecting the most vulnerable as reasons for social distancing, provide environments (e.g., workplaces and grocery stores) that are barrier free and easy to socially distance and provide consistent opportunities in regular, stable contexts to engage in social distancing to develop habits. Future research should seek to provide longer-range prediction of social distancing behavior by the integrated model constructs and test the stability and reciprocal relations among its constructs using a cross-lagged panel design.
|
|
sec
|
Supplementary Material
kaaa073_suppl_Supplementary_Appendix Click here for additional data file.
|
|
title
|
Supplementary Material
|
|
label
|
kaaa073_suppl_Supplementary_Appendix
|
|
caption
|
Click here for additional data file.
|
|
p
|
Click here for additional data file.
|
|
back
|
Acknowledgments
Funding: Martin Hagger’s contribution was supported by a Finnish Distinguished Professor (FiDiPro) award from Business Finland (1801/31/2015).
Compliance with Ethical Standards
Authors’ Statement of Conflict of Interest and Adherence to Ethical Standards The authors declare no conflicts of interest.
Authors’ Contributions M.S.H., K.H., J.J.K., and S.R.S. conceptualized the data; M.S.H., K.H., J.J.K., S.A.M., and S.R.S. made contributions to methods and measures; J.J.K., S.A.M., and S.R.S. were involved in data collection and curation; M.S.H. analyzed the data; M.S.H. and K.H. written the original draft; M.S.H., K.H., J.J.K., S.A.M., and S.R.S. reviewed and edited the article.
Ethical Approval Research was conducted in compliance with National Health and Medical Research Council’s ethical standards for research involving human participants. The research was approved by the Griffith University Human Research Ethics Committee.
Informed Consent All participants provided informed consent prior to data collection.
|
|
ack
|
Acknowledgments
Funding: Martin Hagger’s contribution was supported by a Finnish Distinguished Professor (FiDiPro) award from Business Finland (1801/31/2015).
|
|
title
|
Acknowledgments
|
|
p
|
Funding: Martin Hagger’s contribution was supported by a Finnish Distinguished Professor (FiDiPro) award from Business Finland (1801/31/2015).
|
|
sec
|
Compliance with Ethical Standards
Authors’ Statement of Conflict of Interest and Adherence to Ethical Standards The authors declare no conflicts of interest.
Authors’ Contributions M.S.H., K.H., J.J.K., and S.R.S. conceptualized the data; M.S.H., K.H., J.J.K., S.A.M., and S.R.S. made contributions to methods and measures; J.J.K., S.A.M., and S.R.S. were involved in data collection and curation; M.S.H. analyzed the data; M.S.H. and K.H. written the original draft; M.S.H., K.H., J.J.K., S.A.M., and S.R.S. reviewed and edited the article.
Ethical Approval Research was conducted in compliance with National Health and Medical Research Council’s ethical standards for research involving human participants. The research was approved by the Griffith University Human Research Ethics Committee.
Informed Consent All participants provided informed consent prior to data collection.
|
|
title
|
Compliance with Ethical Standards
|
|
p
|
Authors’ Statement of Conflict of Interest and Adherence to Ethical Standards The authors declare no conflicts of interest.
|
|
p
|
Authors’ Contributions M.S.H., K.H., J.J.K., and S.R.S. conceptualized the data; M.S.H., K.H., J.J.K., S.A.M., and S.R.S. made contributions to methods and measures; J.J.K., S.A.M., and S.R.S. were involved in data collection and curation; M.S.H. analyzed the data; M.S.H. and K.H. written the original draft; M.S.H., K.H., J.J.K., S.A.M., and S.R.S. reviewed and edited the article.
|
|
p
|
Ethical Approval Research was conducted in compliance with National Health and Medical Research Council’s ethical standards for research involving human participants. The research was approved by the Griffith University Human Research Ethics Committee.
|
|
p
|
Informed Consent All participants provided informed consent prior to data collection.
|
