Statistical analyses.
We assigned a value of one-half the detection limit of the analytical method when a compound was not detected in a sample. OC concentrations had log-normal distributions and were log-transformed in all analyses. Therefore, results for contaminants are presented as geometric means. The correlation between contaminant concentrations was evaluated using Pearson’s method on log-transformed values. To evaluate associations between OC exposure and infection incidence rates, we used Poisson regression with quartiles of OC concentration as the main independent variable, and individual incidence rates as the dependent variable (both for bivariate and multivariate analyses). We categorized the exposure using quartiles boundaries, with the first quartile as the group of reference (Table 1). Regression results are, therefore, an estimate of the incidence rate ratios (RRs) for infants in the three highest quartiles of exposure, when infants in each of these quartiles are compared to infants in first quartile. To test the hypothesis of a dose–response association between incidence rates and OC concentrations (p-value for trend), we included the contaminant concentration (log-transformed) directly in the model and treated it as a continuous variable.
We based the selection of potential confounding variables on clinical knowledge and a literature review. Every identified potential confounding variable was tested in the model, but only those influencing the incidence rate ratios by > 5% were included in the final model. The variables initially excluded from the model were retested one by one in the final model to ensure that their exclusion did not influence the results. The variables included in the final model were maternal age at delivery (continuous), season of birth, year of birth (category), breast-feeding duration (categories), sex of the infant, socioeconomic status of the caregiver (continuous), smoking during pregnancy (yes/no), number of cigarettes smoked per day during pregnancy (continuous), number of children < 6 years of age living with the infant (continuous), and village of residence. The following variables were excluded from the final model because they did not significantly affect the association of interest: day care frequentation (ever/never), mean hours per week in day care (continuous), maternal omega-3 fatty-acid concentration in blood (continuous), proportion of omega-3 highly unsaturated fatty acids (continuous), number of smokers in the house where the infant resided (continuous), birth weight, gestational age, and reviewer of the medical chart. When postnatal exposure was investigated, we included in the model the infant’s age when the blood sample was drawn. We considered vaccination coverage a potential confounding factor. Information on vaccination was gathered through the review of the medical chart, but information was missing for many children. Preliminary analyses showed that vaccination coverage was not related to contaminant burden. We thus excluded it from the final models.
All modeling results are presented for both the crude model (only exposure categories) and the adjusted model (exposure categories and all the confounding variables mentioned above). Statistical analyses and database management were conducted using the SAS system 8.02 (SAS Institute, Cary, NC, USA). By convention, a p-value < 0.05 was considered significant.