> top > docs > CORD-19:2dfc0df5fefbe3bd657f73f9c78b3d3278d8b20b

CORD-19:2dfc0df5fefbe3bd657f73f9c78b3d3278d8b20b JSONTXT

Supplementary appendix Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study Abstract Very severe LCWI ----- *In 2013 Pocket Book of Hospital Care for Children, children with lower chest wall indrawing (LCWI) and no danger signs/signs of respiratory distress (and without HIV, severe malnutrition and other underlying conditions) now considered non-severe pneumonia along with children with elevated respiratory rate and recommended for home care. **'Very severe' pneumonia now called 'severe' pneumonia in 2013 pocketbook for hospital care. 1 In young infants. 2 Head nodding not specifically stated under definition for severe pneumonia in text of 2005 Pocket Book of Hospital Care for Children, but 'severe respiratory distress' is. And, elsewhere in manual, head nodding is included as a sign of 'severe respiratory distress'. Page 17 defines severe respiratory distress as very fast, labored breathing with use of auxiliary muscles for breathing (head nodding). Child appears to tire easily and is not able to feed because of respiratory distress. 3 Head nodding and nasal flaring not specifically stated under definition for severe pneumonia in 2013 Pocket Book of Hospital Care for Children, but 'severe respiratory distress' is. And, elsewhere in manual (page 4, e.g.), head nodding and nasal flaring are included as signs of 'severe respiratory distress' (when occurring along with labored, fast, gasping breathing). Page 4 defines severe respiratory distress as "The breathing is very laboured, fast or gasping, with chest indrawing, nasal flaring, grunting or the use of auxiliary muscles for breathing (head nodding). Child is unable to feed because of respiratory distress and tires easily." 4 Multiple or prolonged convulsions. Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study Abbreviations: CXR+, chest radiograph positive (consolidation and/or other infiltrate). The analyses on HIV exposure status were restricted to South Africa and Zambia, where detailed information on maternal HIV status during and after pregnancy was collected and maternal HIV tests were performed at enrollment; *Data from the Mali was site was excluded from this analysis due to lack of standardization in how the data were recorded for oxygen use. In The Gambia, based on evidence from a concurrent project of clinical care, all children in PERCH were coded as receiving oxygen if they were hypoxemic. Missing data: All HIV-uninfected/CXR+ cases (N=3); Gambia (N=0); Kenya (N=1); Zambia (N=0); South Africa (N=2); Bangladesh (N=0); Thailand (N=0). Abbreviations: CXR+, abnormal chest radiograph. Prior exposure to antibiotics was defined as positive serum bioassay for both cases and controls (solid bars) and additional cases (hashed bars) who were serum bioassay negative for antibiotics but identified as having received antibiotics prior to specimen collection based on documentation at study or referral facility. Prior exposure to antibiotics in Zambia is very high due to a health system structure whereby children are hospitalised only after being assessed at a primary care facility where antibiotics are administered prior to transport to hospital. Other non-fermentative gram-negative Other Streptococci and Enterococci e 4 (0.1) Detection based on NP/OP PCR positivity except for pathogens with specified density thresholds: P. jirovecii, 4.0 log10 copies/mL; H. influenzae, 5.9 log10 copies/mL; CMV, 4.9 log10 copies/mL; S. pneumoniae, 6.9 log10 copies/mL. Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study Supplementary a Odds ratios adjusted for age (months) and site. b Odds ratios adjusted for age (months), site, and all other pathogens, which quantifies the degree of association between pneumonia case status and pathogen detection for persons who are otherwise similar with respect to age, site and the other measured pathogens. c NP/OP PCR thresholds: P. jirovecii, 4.0 log10 copies/mL; H. influenzae, 5.9 log10 copies/mL; CMV, 4.9 log10 copies/mL; S. pneumoniae, 6.9 log10 copies/mL. d High density S. pneumoniae defined as > 6.9 log10 copies/mL (NP/OP PCR). Pneumococcal serotypes were determined from culture isolates by Quellung method or PCR; for NP/OP-culture negative but PCR-positive specimens, microarray was used to determine serotype. e Measles PCR testing was done using a single-plex PCR method on the NP/OP viral transport medium specimen from cases meeting one or more of the following criteria: history of measles, measles admission diagnosis, measles discharge diagnosis, or measles rash at enrollment. Organism abbreviations: CMV, cytomegalovirus; HBOV, human bocavirus; HMPV, human metapneumovirus; PV/EV, parechovirus/enterovirus; RSV, respiratory syncytial virus. Note: no adjustments were made to the odds ratio for prior antibiotic use, which is known to influence bacterial positivity. Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study Page 37 of 105 b. By number of organisms detected *NP/OP PCR thresholds: P. jirovecii, 4.0 log10 copies/mL; H. influenzae, 5.9 log10 copies/mL; CMV, 4.9 log10 copies/mL; S. pneumoniae, 6.9 log10 copies/mL. P-values are from logistic regression adjusted for site and age (in months). *Analysis includes the use of NP/OP PCR thresholds: P. jirovecii, 4.0 log10 copies/mL; H. influenzae, 5.9 log10 copies/mL; CMV, 4.9 log10 copies/mL; S. pneumoniae, 6.9 log10 copies/mL. Gray: Odds ratios adjusted for age (months) and site. Black: Odds ratios adjusted for age (months), site, and presence of other pathogens from NP/OP PCR, which quantifies the degree of association between pneumonia case status and pathogen detection for persons who are otherwise similar with respect to age, site and the other measured pathogens. Note: no adjustments were made to the odds ratio for prior antibiotic use, which is known to influence bacterial positivity. Abbreviations: CMV, cytomegalovirus; HBOV, human bocavirus; HMPV, human metapneumovirus; PV/EV, parechovirus/enterovirus; RSV, respiratory syncytial virus. a High density S. pneumoniae defined as > 6.9 log10 copies/mL (NP/OP PCR). Pneumococcal serotypes were determined from culture isolates by Quellung method or PCR; for NP/OP-culture negative but PCR-positive specimens, microarray was used to determine serotype. b Measles PCR testing was done using a single-plex PCR method on the NP/OP viral transport medium specimen from cases meeting one or more of the following criteria: history of measles, measles admission diagnosis, measles discharge diagnosis, or measles rash at enrollment. *NP/OP PCR thresholds: P. jirovecii, 4.0 log10 copies/mL; H. influenzae, 5.9 log10 copies/mL; CMV, 4.9 log10 copies/mL; S. pneumoniae, 6.9 log10 copies/mL. Legionella species 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0) AnyInfluenza B 11 (1.0) 10 (0.4) 7 (1.1) 19 (0.8) 13 (1.1) 5 (1.0) Influenza C 8 (0.7) 15 (0.5) 2 (0.3) 14 (0.6) 7 (0.6) 3 (0.6) Measles b (n=10) 0 (0.0) -- 0 (0.0) -- 0 (0.0) 0 (0.0) Parainfluenza Page 42 of 105 Supplementary Abbreviations: OR, odds ratio, CI, confidence interval, WB, whole blood; CXR+, chest X-ray positive (consolidation and/or other infiltrate). Odds ratios adjusted for site and age in months. For sites with lung aspirate positive cases (The Gambia and South Africa), cases positive on lung aspirate ('known cause') were accounted for in the final population aetiology estimate by using their results to update the prior distributions for the remaining consolidated cases. These remaining cases were then used in the PIA model to update the aetiology prior a second time to complete the analysis of all cases. Cases positive on pleural fluid similarly had a 'known cause' but due to the small numbers of remaining cases with a confirmed pleural effusion on CXR without pleural fluid findings, these positive results were not used to update their aetiology prior distributions but were used to determine the aetiology of those positive cases after the model was run. For sites with lung aspirate positives, cases are identified as being in one of two groups; one group representative of those with lung aspirates collected (cases with consolidation on CXR), and a second group of cases with other infiltrate. The aetiology analysis has been run on all sites and case groups combined, allowing the aetiologic distribution and false positive rates to vary across sites, and the aetiologic priors to vary by case sub-group within a site (e.g., based on lung aspirate or pleural fluid data), while the data informing sensitivity are from all the sites. The analysis allows for the estimation at the all-site level as well as the site-specific level. Page 45 of 105 Abbreviations: NP/OP, nasopharyngeal/oropharyngeal swab; PCR, polymerase chain reaction; S. pneu, Streptococcus pneumoniae; Mtb = Mycobacterium tuberculosis. For induced sputum M. tuberculosis analyses, positivity was assigned by first induced sputum, or by first gastric aspirate if induced sputum unavailable. Lung aspirate results are restricted to specimens obtained within 3 days of enrollment and those pathogens determined by the clinical review team to be noncontaminants. Cases positive on lung aspirate ('known cause') were accounted for in the final population aetiology estimate by using their results to update the prior distributions for the remaining consolidated cases. Page 46 of 105 Supplementary Supplementary Table 11 . Description of symbols: The size of the symbol is scaled based on the ratio of the estimated aetiologic fraction to its standard error. Of two identical aetiologic fraction estimates, the estimate associated with a larger symbol is more informed by the data. Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study Supplementary Description of symbols: Line represents the 95% credible interval; darker region of line represents the interquartile range. The size of the symbol is scaled based on the ratio of the estimated aetiologic fraction to its standard error. Of two identical aetiologic fraction estimates, the estimate associated with a larger symbol is more informed by the data. Page 53 of 105 Supplementary For each pathogen, the figures display the distribution of the individual case probability for that pathogen, excluding cases with an aetiologic probability < 5% for the pathogens so as to scale the y-axis and display the children with higher probability. Red: cases testing positive for the pathogen by NP/OP PCR (or WB PCR for S. pneumoniae only); blue: cases testing negative; gray: cases missing data; black: cases with a positive silver standard specimen (i.e., positive on lung aspirate, pleural fluid, or blood culture). Cases with more than one pathogen detected in silver standard specimen(s) had Evidence for pneumonia etiology due to S. pneumoniae is based on three independent pieces of information: high load nasopharyngeal/oropharyngeal (NP/OP) PCR, high density whole blood (WB) PCR, and blood culture. This sensitivity analysis evaluated the impact of lowering the blood culture sensitivity priors for S. pneumoniae only, while also specifying increased certainty. The midpoint value of the sensitivity prior ranges decreased from approximately 11% to 2%; no changes were made to the other sensitivity or aetiology priors. *Site-specific observed results. Interpretation: Decreasing the blood culture sensitivity priors for S. pneumoniae had little impact on the aetiology results. Evidence for pneumonia aetiology due to S. pneumoniae is based on three independent pieces of information: high load nasopharyngeal/oropharyngeal (NP/OP) PCR, high density whole blood (WB) PCR, and blood culture. At sites with weak blood culture evidence (i.e., low number of pneumococcal positive isolates), the model relies more on the evidence from NP/OP and WB PCR data. Therefore, changes in the blood culture sensitivity prior will have minimal impact when there are few blood culture positives. In sites with more blood culture positive cases there is still limited influence of the blood culture sensitivity prior due to the additional contributing data for S. pneumoniae provided by NP/OP PCR and WB PCR. Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study This sensitivity analysis evaluated the impact of lowering the sensitivity priors for all pneumococcal measurements (which should result in higher etiological estimates for pneumococcus), while also specifying increased certainty. The midpoint value of the sensitivity prior ranges decreased as described below; no changes were made to the other sensitivity or aetiology priors. Blood Culture (sensitivity analysis presented in Supplementary Figure 13A) Base-Case sensitivity priors Lower and narrower sensitivity priors Ratio Base:Lower No prior antibiotic exposure and adequate blood volume Interpretation: Use of a higher aetiology prior modestly increased S. pneumoniae aetiology by 1.1-4.7% compared to the base model, resulting in posterior S. pneumoniae aetiology estimates ranging from 2.6-20.0% across the sites. The posterior aetiology estimate for S. pneumoniae was consistently lower than the aetiology prior that was specified for that site. Even when the combined aetiology prior for VT and NVT S. pneumoniae was 34.7% for sites which had not introduced PCV (Zambia, Thailand, and Bangladesh), the resulting posterior aetiology estimate ranged from 2.6-7.3%. *Site-specific observed results. Interpretation: Increasing the aetiology prior of NoS from 3% (the base case) to 25% increased the aetiology attributed to NoS by 6.8-13.5% at all sites except Thailand (where evidence for causality was weakest), resulting in a posterior NoS estimate of 7.8-15.5%. At the Thailand site which had the fewest number of cases and fewest number of blood culture positives, NoS aetiology increased from 5.4% (base) to 23.9%. This finding demonstrates that the amount of pneumonia due to pathogens that are not measured (NoS), which is determined by lack of evidence for the measured pathogens, is hard to quantify. However, this also demonstrates there is evidence in the data for measured pathogens to suggest that NoS is likely less than the prior of 25% because the NoS estimate was >10% lower than the prior for all sites except Thailand, where there was weak data for aetiologic evidence. While a precise estimate for NoS is not possible, the relative magnitude of the aetiology for the measured pathogens remains constant regardless of this assumption (i.e., the distribution of the pathogens tested for is similar). Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study Screening was performed 24/7 at four sites (Kilifi, Sa Kaeo, Nakhon Phanom and Matlab) during which all eligible consenting cases were enrolled. At the remaining sites, screening was performed during established hours; all eligible consenting cases presenting during pre-defined screening hours were enrolled, with the exception of Mali which used a systematic sampling process. See below table with site-specific screening hours. Hospitalisation was an inclusion criterion at all sites except Bangladesh where PERCH study enrollment included children recommended for hospitalisation but not admitted (n=33); 1 for hospitalised children, case assessment occurred within 24 hours of admission. The majority of these children were not admitted because the parent refused admission (32/33). Controls were randomly selected from residents of the study catchment areas and frequency matched to cases by age group (1 to <6 months, 6 to <12 months, 12 to <24 months, and 24-59 months of age). 1 All sites aimed to enroll approximately 1 control per case. To achieve this, all the African sites and Nakhon Phanom, Thailand aimed to recruit a minimum of 25 controls each month, Matlab and Dhaka Bangladesh each aimed to enroll 10-15 controls per month, and Sa Kaeo, Thailand aimed to enroll 13 controls per month. In months where the number of cases exceeded the target control enrollment, sites enrolled additional controls to achieve a 1:1 ratio case:control ratio for that month. Cases underwent a comprehensive standardised clinical examination at admission, 24 and 48 hours (if still hospitalised), and vital status was assessed at 30 days. 2 Respiratory signs, anthropometry, level of consciousness, and oxygen saturation (on room air whenever possible) were recorded, among other signs. Cases still hospitalised at 24 and 48 hours after admission underwent follow-up clinical assessments. Vital status of cases was assessed during a follow-up visit or phone interview conducted 30 days after admission (window 21-90 days). Controls were similarly assessed for clinical findings at enrolment but had no follow-up assessments. Chest radiographs (CXRs), obtained from cases at enrolment, were interpreted blinded to site and clinical factors by two (of 14) radiologists and paediatricians trained in the WHO standardised interpretation of paediatric CXRs. [3] [4] [5] Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study CXRs were classified as consolidation, other infiltrate, both, normal, or uninterpretable. CXRs obtained more than 3 days following enrolment were excluded from analyses because these films were not routinely collected, would reflect clinical progression (improvement/deterioration), and may represent nosocomial disease. Lung aspirate collection was performed at four sites (The Gambia, Bangladesh, Mali, and South Africa). Eligible cases were those with large, dense peripheral consolidation on CXR. Contraindications for lung aspirate collection included: 1. Presence of pneumatocoeles on CXR. 2. Post measles pneumonia. 3. If the patient was clinically unstable as determined by a clinician, the procedure would be deferred until stabilisation. 4. Cardiorespiratory resuscitation (CPR) performed within the last 24 hours. 5. Parental refusal to have their child subjected to lung aspiration. Pleural fluid was collected from a minority of cases as indicated by attending clinicians. The methodology for obtaining pleural fluid followed local clinical practice guidelines, including standard safety precautions. Contraindications for pleural fluid collection included: 1. Coagulopathy or thrombocytopenia. 2. Haemodynamic or respiratory instability (unless therapeutic thoracentesis was required for management of clinical instability). Further details on lung aspirate and pleural fluid collection and results will be published separately. Antibiotic pre-exposure was defined as having either a positive serum bioassay at enrolment (cases and controls), or clinician-documented antibiotic administration at the referral or study hospital prior to specimen collection (cases only). 6 Children were defined as being HIV-infected if they had direct viral detection at any age or were HIV seropositive at >12 months of age. At the South African site, where routine clinical practice is to confirm all HIV-seropositive cases under-18 months of age using molecular tests (including HIV PCR and HIV viral load testing), all children under-18 months of age with positive HIV serology had their status confirmed using molecular tests. Others were defined as being HIV-uninfected, including children with unknown HIV status (n=385 cases, n=595 controls, mostly from Kenya, The Gambia and Mali) because HIV testing was not performed. Tachypnoea was defined as ≥60 breaths per minute (bpm) for children <2 months of age, ≥50 bpm for children 2-11 months, and ≥40 bpm for children 12-59 months. Hypoxaemia was defined as either 1) room air pulse-oximetry oxygen saturation <90% at the two sites with elevation above 1,200 metres (Zambia and South Africa) or <92% at all other sites, or 2) a child being treated with oxygen in the absence of a room air oxygen saturation reading. Fever was defined as > 38°C. Controls were considered to have respiratory tract illness (RTI) if they had 1) cough (observed or reported) or runny nose (reported), or 2) one of the following: ear discharge (reported), wheeze (reported), or difficulty breathing (reported), in the presence of either sore throat (reported) or fever (observed temperature ≥38°C or reported fever in the past 48 hours). The specimens collected, microbiology tests conducted, and density thresholds used in analyses have been described elsewhere. [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] In brief, we used quantitative multiplex polymerase chain reaction (PCR) (FTD Resp-33 kit; Fast-track Diagnostics, Sliema, Malta) and routine culture to test nasopharyngeal/oropharyngeal (NP/OP) swabs (cases and controls; culture for S. pneumoniae only), induced sputa (cases only), gastric aspirates (cases only, for mycobacteria), lung aspirates (cases only; at The Gambia, Bangladesh, Mali, and South Africa sites), pleural fluid (cases only), and blood (cases and controls for Streptococcus pneumoniae PCR; cases only for blood culture). Cases and controls were also tested for antibiotic activity in serum and C-reactive protein (all cases and subset of controls). 6, 16 Refer to the below sections and Section 4 for more details on the laboratory results and how they were used in the aetiology analysis. The Fast Track Diagnostics Respiratory Pathogens 33 (FTD Resp-33) multiplex PCR kit which was used in the PERCH study includes the following 33 viral, bacterial and fungal targets: Data from all nasopharyngeal/oropharyngeal swabs (NP/OP) PCR targets were used in the analysis except Klebsiella pneumoniae and Moraxella catarrhalis. Some targets underwent additional confirmatory testing, as described below. 10 • Klebsiella pneumoniae: Positive results were considered invalid because of poor assay specificity. 17 • Moraxella catarrhalis: This target was summarized in the descriptive NP/OP PCR analyses but excluded from aetiology analyses due to poor specificity (odds ratio significantly less than 1.0) and inability to determine a density threshold that improved specificity such that is was positively associated with case status. 12 M. catarrhalis is a commensal bacterium that can cause pneumonia but was observed more commonly among controls than among cases. There is some evidence in the literature that presence in the NP/OP may be protective against developing pneumonia. The hypothesized biological rationale is that if the child's microbiome is slanted more toward colonization of M. catarrhalis, it may out compete other more pathogenic organisms. 18 Because the underlying assumption of the analytic model is that what is measured in the NP is a signal of what is in the lung, the statistically significantly higher rate in controls who do not have pneumonia is in direct conflict with this assumption. Therefore, these NP/OP data do not provide useful diagnostic evidence of pneumonia etiology and so are not used. • Legionella species: Samples positive for Legionella species underwent confirmatory PCR testing using a different extraction method due to contamination of the original extraction reagents. Only one positive remained following confirmatory testing. Although there was only one positive, since we systematically tested for Legionella species and it is thought to be a potential cause of paediatric pneumonia, it was included as a possible cause in the aetiology analysis; this was handled consistently across all sites, including those where no Legionella species positives were detected. Quantitative PCR density thresholds were applied to NP/OP for pathogens with poor specificity (i.e., odds ratio <1) where a threshold that improved distinction between cases and controls could be identified: pneumococcus, H. influenzae, cytomegalovirus, and Pneumocystis jirovecii by NP/OP. [11] [12] [13] Measles was not systematically tested for in the PERCH study. Of all PERCH study enrolled cases, 33 cases were identified who met the PERCH study clinical criteria for measles testing (history of measles in past month, measles rash, measles admission diagnosis, measles discharge diagnosis); the NP/OP swabs from these 33 cases were tested for measles. Of the 33 suspected measles cases, 4 (12.1%) tested positive (Kenya n=2; Mali n=2; all had either an admission or discharge diagnosis of measles), none of whom had a positive CXR. Since measles was not systematically tested for, we elected to remove this as a potential cause in the aetiology analysis and simply describe the number of positives identified in the PERCH study. These four cases were excluded from the aetiology analyses and population aetiology distributions. Quantitative PCR density thresholds were applied to whole blood results to improve distinction between prevalence in cases and controls. 14,15 Pneumococcal serotypes were determined from culture isolates by Quellung and/or PCR, or by microarray (NP/OP only) for specimens that were NP-culture negative and PCR-positive. 10 Pneumococcal serotypes were stratified into vaccine-type (VT) or non-vaccine type (NVT), based on their inclusion in the 13-valent pneumococcal conjugate vaccine (the VT serotypes being 1, 3, 4, 5, 6A, 6B, 7F, 9V, 14, 18C, 19A, 19F, 23F). Handling mixed or ambiguous serotypes: Pneumococcal serotype results that were mixed or ambiguous following Quellung and/or PCR and unable to be resolved at the study sites were serotyped by Quellung reaction at a reference laboratory (National Institute for Communicable Diseases, Johannesburg, South Africa or the Institute of Environmental Science and Research (ESR), Porirua, New Zealand). Serotyping for all pneumococcal isolates from Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study sterile sites as well as a sample of pneumococcal isolates from the NP swab culture (50-70 per site) were verified by Quellung at the ESR laboratory. 10 Serotyping by microarray: Pneumococcal serotypes for specimens that were positive by NP/OP PCR but negative by nasopharyngeal (NP) skim milk, tryptone, glucose and glycerin medium (STGG) culture were determined through microarray testing. 20 Samples with pneumococcal NP/OP PCR results above the density threshold (>6·9 log10 copies/mL) were prioritised for testing. Of the 87 samples meeting this criterion, 58 had serotyping results available by microarray. Multiple pneumococcal serotypes were detected in 13/58 (22%) of all samples tested by microarray. The output from the microarray provides the relative abundance of each serotype detected in the specimen (e.g., 15B [69%] + 10F [31%]). Microarray results were adjudicated by PERCH study investigators, including the PERCH study laboratory director, to determine the final serotype based on the relative abundance of each. For mixed serotype specimens, the dominant serotype was selected as the final serotype result, with pneumococcal serotype data from other specimens used to assist in the serotype confirmation process when available. In addition, a limited number of culture-positive/PCR-positive STGG samples (n=45) were tested by both methods (culture and microarray) as part of an initial validation step for the microarray; for these children, the serotype result from the culture test was used as the final serotype. Of these tested by both methods, 34 (76%) had identical results by the two methods. Blood culture was performed at each site's laboratory for each case enrolled. Positive specimens underwent confirmatory testing at the central laboratory. 10 Below is a summary of the review process: • Out of 4,176 cases with blood culture data, 526 (12·6%) had an organism isolated from blood o 328/526 (63·4%) cases were classified as having only a contaminant organism isolated (based on a priori list of contaminants; see below Table 2 ) • The remaining 198 blood culture positive cases underwent individual case reviews by a panel of PERCH study senior investigators to determine whether the organism detected was the likely cause of the child's hospitalisation or a contaminant. The review included clinical, laboratory and risk factor data. o This review resulted in an additional 51 cases being classified as having only a contaminant organism isolated. o A total of 147 cases had one or more pathogen identified that were considered to be plausible causes of the hospitalisation. A similar review was conducted for PCR and culture results from lung aspirate and pleural fluid specimens. We planned to test all blood cultures that were alarm positive, culture negative with the BinaxNOW ® assay for pneumococcal antigen, however this was not implemented consistently at all sites. Of the 13 cases who met the criteria for testing, only seven were tested, and all were negative. An additional 20 samples that did not meet the criteria were tested; of those, two were positive (both in South Africa) and are described below: Pneumococcal BinaxNOW ® testing of pleural fluid specimens Only some sites implemented pleural fluid specimen testing with the BinaxNOW ® assay for pneumococcal antigen. Of the 22 cases with pleural fluid samples, only eight were tested and three were positive. Of those eight cases tested, three were positive for S. pneumoniae by culture or PCR of the pleural fluid; two of those three were also positive by BinaxNow ® . Among those tested who were S. pneumoniae negative by both culture and PCR of pleural fluid (data available for both culture and PCR), none were BinaxNow ® positive. There is no specimen identified among those without BinaxNow ® performed where an additional test by BinaxNow ® would have provided diagnostic information that was not already provided by blood culture, pleural fluid culture or pleural fluid PCR, given that among the tested cases who were S. pneumoniae negative by both culture and PCR none were BinaxNOW ® positive. Furthermore, among those 14 children who were not tested by BinaxNOW ® , S. aureus and other pathogens were detected in the pleural fluid culture, providing diagnostic information for a pathogen other than S. pneumoniae. In the aetiology analysis all three cases positive for pneumococcal antigen by BinaxNow ® were considered 'confirmed pneumococcal cases'; for the case positive for two specimens on pleural fluid (Table 4 ) the aetiology was divided evenly between the pathogens (see Section 7). We attempted to obtain a single sputum specimen from each case; in South Africa multiple sputum specimens were obtained from cases for M. tuberculosis testing as part of routine clinical practice. 21 In situations when more than one sputum or gastric aspirate specimen was collected, only the first specimen result for Mycobacterium tuberculosis was used; hence, if for example a child had two induced sputum specimens and three gastric aspirates submitted and only the second induced sputum specimen was positive for M. tuberculosis culture, such a case was not deemed to be a case of tuberculosis. Induced sputum specimens were tested by culture and using the FTD Resp-33 multiplex PCR. For the majority of pathogens the IS PCR data was strongly correlated with the NP/OP PCR data. 22 Because IS specimens were not collected from controls, we could not estimate and account for specificity which we assumed was not 100% (i.e., not a 'silver standard' measure like blood culture). Given lack of information on specificity and the strong correlation with NP/OP PCR data, we concluded that IS data would not meaningfully contribute to and improve the analysis, so, except for induced sputum culture for M. tuberculosis, they were not used in the determination of aetiology. To estimate the aetiology of pneumonia, we needed an analytic method that could integrate the multiple specimens and tests results from the cases and the controls. 23 We developed a novel analytic method called the Bayesian Analysis Kit for Etiology Research (BAKER), which is a nested, partially latent class analysis. [24] [25] [26] This method was used to integrate the PERCH study data to estimate the aetiology distribution for each individual case and for the population of cases. The purpose of this appendix is to provide additional details related to the analytic inputs and methods specific to the PERCH Integrated Analysis (PIA) beyond that presented in the main paper. In this section we describe nuances related to the specimen and laboratory measurements and pathogens included in the integrated aetiology analysis. Gold None Blood culture (following clinical review to remove contaminants). Pleural fluid and lung aspirate (culture or PCR; following clinical review to remove contaminants) by 1) updating the aetiology priors for cases with consolidation (lung aspirate results only), and 2) manually updating individual probabilities for pleural fluid and lung aspirate positive cases. Induced sputum culture for Mtb (positive by first IS specimen, or first gastric aspirate if IS unavailable). NP/OP PCR, with density thresholds for the following targets: • S. pneumoniae: 6.9 log10 copies/mL • H. influenzae and H. influenzae type b: 5.9 log10 copies/mL • CMV: 4.9 log10 copies/mL • PCP: 4 log10 copies/mL Ignore NP/OP PCR data for Moraxella catarrhalis and Klebsiella pneumoniae (see Section 2.1 -Laboratory Procedures). WB PCR (S. pneumoniae only) with density threshold (2.2 log10 copies/mL). Coronavirus OC43, NL63, HKU1, and 229E were measured as separate targets but grouped as 'Coronaviruses' for the aetiology analysis. Additional bacteria detected in blood culture (no corresponding NP/OP PCR measurement used in the aetiology analysis): • As described in Section 2.1, the FTD Resp-33 panel includes some organisms grouped at the target level (e.g., HMPV A/B, RSV A/B), and some organisms split across multiple targets (e.g., Parainfluenza types 1, 2, 3 and 4 and, Influenza A, B and C). Analyses were performed using the results of each target wherever possible as opposed to grouping across targets for a given organism, except for Coronavirus which we grouped across targets (coronaviruses NL63, 229E, OC43 and HKU1) due to the low numbers of positives. Another exception is S. pneumoniae, for which we applied the serotyping data to subsequently split pneumococcus into relevant subgroupings (e.g., PCV13-type vs not) rather than analysing each pneumococcal serotype separately which would have spread those data too thinly. The aetiology results for pathogens estimated at the sub-species level were subsequently aggregated at the species level for presentation in the most figures ( Table 5 ). Note that because the analysis had an equal aetiology prior for each pathogen, pathogens that are post-hoc grouped (e.g., Influenza) have in aggregate a larger cumulative aetiology prior than other pathogens without sub-species or do not have a FastTrack target at the subspecies level (e.g., Influenza virus has 3 times the aetiology prior as Rhinovirus because each Influenza type [A/B/C] is given the same aetiologic prior as Rhinovirus when the output is combined as 'Influenza virus'). As the PIA is a Bayesian analysis, we needed to specify starting values for the distribution of the aetiology fractions (aetiology prior distributions). To reduce a priori bias, we used uninformed aetiology priors wherein each pathogen tested for has an equal starting probability to be the cause; in this study the aetiology prior for all pathogens was 1/34 (i.e., 33 pathogens plus a 'Not Otherwise Specified' (NoS) category), a 34-dimension symmetric Dirichlet distribution with hyperparameter = 0.1. The exchangeable nature of the symmetric Dirichlet prior treats all pathogens equally likely to be the most important cause prior to being updated by PERCH study data. The distribution and uncertainty around the aetiology priors were also specified (i.e., not a fixed point estimate) and assumed that the majority of pneumonia cases were caused by a subset of the pathogens, without specifically indicating which pathogens; this was implemented by using a small hyperparameter ( = 0.1) to make the prior distribution more flexible and the prior uncertainty larger in contrast to under larger values. Even with this assumption, all organisms are included in the analysis (Table 6 ) and have a non-zero aetiologic fraction estimated by the analysis, in contrast to other analytic approaches that exclude pathogens not associated with case status. Abbreviation: PCV13, 13-valent pneumococcal conjugate vaccine. The lung aspirate and pleural fluid specimens provide valuable information about a child's pneumonia cause, but few children had this information and extrapolating their results to other children would have been challenging because of small numbers and because they represented only a fraction of the cases (e.g., results from pleural fluid are only representative of children with pleural effusion, not all pneumonia cases). In this section we describe methods used to incorporate this valuable information into the integrated aetiology analysis, which was done by updating the aetiology priors for the subset of representative cases. There were three primary challenges with incorporating the lung aspirate and pleural fluid data into the analysis: 1. Very few cases had available specimens and only a subset were positive. 2. Of the children who were positive, many were positive for multiple organisms. This creates conflict between the PIA assumption of a single-pathogen cause and that test results from these specimens have 100% specificity. 3. Obtaining a lung aspirate or pleural fluid specimen from a child was determined based on strict clinical characteristics, therefore these children were unlikely to be clinically or aetiologically representative of all PERCH study cases. We made four key decisions in determining how the lung aspirate and pleural fluid data were to be used in the analysis of determining the cause for the latent (unknown) cases: 1. Instead of using these silver standard measurements as direct inputs in the PIA model (where 97% of cases would have missing data), the positive results were used to 'update' the uniform aetiology prior values such that those pathogens detected on the silver standard measurement had a higher probability of being the cause. • This process was done on a site-specific basis, i.e., the distribution of pathogens found was only applied to other cases at the same site. 2. The updated aetiology priors were only applied to the subset of cases who were deemed representative of those with the corresponding specimen obtained (see next section). 3. Given the small number of cases with pleural effusion confirmed on chest x-ray (n=22, ranging from n=3 to n=7 across sites) who would be representative of those with pleural fluid specimens, we decided against using the pleural fluid data to update the aetiology priors, as done for the lung aspirate positive cases, because the analytical group to which these would be applied was too small for a robust analysis (i.e., requires stratified analysis). For these confirmed cases the pleural fluid results were used to inform their individual aetiology but not extrapolated to other cases. 4. For the cases with positive lung aspirate or pleural fluid results, those with more than one pathogen detected had aetiology apportioned equally across the pathogens. Decision 2 above required us to define the subgroups of cases representative of those with specimens obtained. For lung aspirates, we identified cases with consolidation on chest radiograph (CXR), an eligibility criterion for obtaining the specimen, and analyzed them separately, updating their aetiology prior using the lung aspirate results. The lung aspirate results used to update the aetiology priors were restricted by the following criteria: 1. The specimen was obtained within 3 days of enrollment to exclude any nosocomial infections. 2. The positive result was confirmed to be a cause of the child's pneumonia by the Silver Standard Working Group (see Section 2.4) based on their clinical presentation, course of treatment, and the laboratory data (e.g., weakly positive PCR results were excluded). 3. The child was not HIV+ (as the primary analysis excludes HIV+ children). To inform the aetiology priors the following rules were applied to the eligible specimens to determine the updated aetiology priors used in the PIA (Table 7) : Abbreviations: LA, lung aspirate; PCV13, 13-valent pneumococcal conjugate vaccine. Use of non-informative values of sensitivity (i.e. uniform likelihood from 0-100%) contradicts our belief that (a) NP/OP PCR is highly sensitive (i.e., likely greater than 50%) for detecting pathogens given they are in the lung and (b) blood culture (BCx) is poorly sensitive (i.e. likely less than 20%). The sensitivity priors for each pathogen/test were selected a priori to analysing the PERCH study data by an internal working group. Available information external to the PERCH study was evaluated to determine the sensitivity of the measurements. When developing the sensitivity priors, we are referring to diagnostic sensitivity for a given organism, which is the probability the test correctly identifies those children with pneumonia caused by the organism. In the below sections, we detail the final sensitivity priors for each pathogen/test combination as well as the processes for developing those priors. The PIA integrates these sensitivity prior distributions with the observed data, changing ('updating') the priors with contributions of the study data which results in posterior sensitivity estimates for each pathogen/test combination. Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multicountry case-control study 85 6.2. Blood cultures The final sensitivity priors for pathogens measured by blood culture are summarized below. The sections that follow describe the information used to inform the sensitivity priors. Diagnostic sensitivity of blood cultures The common pneumonia pathogens associated with bacteraemia are typically in concentrations of at least 10-100 CFU/mL. Therefore, 1 mL of blood inoculated into a blood culture bottle should have high sensitivity (estimate 90%) to detect the pathogen given that it entered the blood, as only one bacterium in the bottle is needed to eventually alarm the blood culture instrument. Antibiotic pretreatment and contamination will lower this estimate. If there are no skin contaminants, culture sensitivity is about 100% for detecting pathogens in blood assuming microbial identifications are accurate. Therefore, sensitivity of blood culture to detect pathogens in the lung is driven primarily by the prevalence of bacteraemia. There is direct evidence of the diagnostic sensitivity for Streptococcus pneumoniae and Haemophilus influenzae from vaccine probe studies (see section 6.2.2), which was estimated to be between 5-20%. For all other pathogens we set the base blood culture sensitivity prior to 5-15%, except for Salmonella species, Enterobacteriaceae and Neisseria meningitidis, for which we selected wider priors (10-50%) to reflect their greater uncertainty. The results of the pneumococcal conjugate vaccine (PCV) trials in South Africa and The Gambia were reanalysed, restricting to PERCH-like conditions (i.e., severe and very severe pneumonia cases only), to estimate the number of pneumococcal blood culture confirmed cases relative to all pneumococcal pneumonia cases (i.e., to estimate the percent of pneumococcal pneumonia cases that were bacteraemic). The number of pneumococcal blood culture confirmed cases (numerator) was estimated by the number BCx+ for pneumococcus in the control group who met the PERCH study case definition. The denominator was estimated by the number of PERCH-like pneumonias in the control group that had a BCx taken multiplied by the vaccine efficacy against PERCH-defined pneumonia, adjusted for the vaccine efficacy against vaccine-type BCx+ and the percent of BCx+ that was vaccine-type: results were 6.1% in South Africa and 17.7% in The Gambia (both with wide confidence intervals due to small number of vaccine-type BCx+ in controls (n=6 in SA and n=14 in The Gambia). We also conducted a sensitivity analysis assuming the vaccine efficacy against vaccine-type pneumonia was lower for non-bacteraemic cases (50%) than for BCx+ cases (4/6=67% in South Africa and 11/14 =79% in The Gambia): results were 4.6% in South Africa and 11.3% in The Gambia. Restricting analyses to CXR+ cases only reduced the sample size of vaccine-type BCx+ in controls to only 2 in South Africa (too small to analyse) and 10 in The Gambia, which produced similar results as for all cases (18.5% and 13.2% for the sensitivity analysis). Taking the lowest and highest values from these four results (4.6% to 18.5%) and rounding up (because impact of prior antibiotics was not accounted for in these analyses but is adjusted for the in the PIA -see section 6.2.3) produced the range 5-20%. A similar re-analysis exercise was intended for H. influenzae using results of The Gambian Hib probe trial; however, the data from that 20+ year old trial could not be located. Given the similarities to pneumococcus, we elected to apply the S. pneumoniae blood culture priors to H. influenzae. Both antibiotic exposure prior to specimen collection and low blood volume reduce pathogen detection in blood cultures for all bacteria by approximately 50%. 6 Therefore, we halved the sensitivity priors of blood culture for children with evidence of either factor. For the analysis, prior antibiotic exposure was defined as having either a positive serum antibiotic bioassay or clinician report of antibiotics administered prior to specimen collection, and low blood volume was defined as <1.5 mL. To adjust the sensitivity prior, the midpoint of the originally specified sensitivity prior range was calculated (e.g., 11 .4% is the midpoint of the sensitivity prior range of 5-20% for S. pneumoniae) then halved (5.7%). The beta distribution was applied to this halved midpoint to construct the adjusted range. We specify the two parameters of a beta probability density with mode at the midpoint sensitivity (Table 9 ). For children who were missing data for prior antibiotic exposure and/or blood volume, we developed rules for how to handle the missing data based on the results for children with data at the same site (Tables 10 and 11 ). When the proportion of cases with adequate blood volume (or no prior antibiotic use) was >70%, children with missing data were assumed to have adequate blood volume (or no prior antibiotic use) and no changes were made to sensitivity priors. If <30%, then children with missing data were assumed to have low blood volume (or prior antibiotic use) and their sensitivity priors were halved. When the proportion was between 30-70%, no assumptions could be made on the status for children with missing data (i.e., "unknown" status), and the sensitivity prior range was widened to the minimum of the adjusted range and maximum of the base range to reflect this uncertainty. Red = between 30-70%; set to "unknown" for cases missing data. Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study Unknown prior antibiotic use and unknown blood volume a South Africa Unknown prior antibiotic use and unknown blood volume a a Use wider sensitivity prior range that spans the minimum of the adjusted range and maximum of the base range to reflect this uncertainty. The final sensitivity priors for pathogens measured by NP/OP PCR are summarized below. , the sensitivities are also likely to be high. Based on this evidence, we set the lower bound of the sensitivity prior to 50% for these pathogens. For Salmonella species and Legionella species, there was no evidence regarding frequency of colonisation so a wide sensitivity prior range was selected (0.5-90%). For relevant pathogens, applying PCR density thresholds is likely to decrease the percentage of confirmed cases positive on NP/OP PCR. [11] [12] [13] Evidence in PERCH study data: We evaluated PERCH study data that were not directly being used in the PIA to provide additional evidence on which to base the NP/OP PCR sensitivity priors. There were eight cases positive on lung aspirate (culture and/or PCR) for S. pneumoniae. Of the seven with NP/OP PCR data, all were positive for S. pneumoniae, but only five were above the density threshold (6.9 log10 copies/mL). The two cases below the density threshold (6.2 log10 and 6.8 log10 copies/mL, respectively) did not have evidence of prior antibiotic use. 6, 13 The sensitivity prior range was changed from 50-90% (midpoint 72%) to 15-55% (midpoint 33%). For the analysis, prior antibiotic exposure was defined by having either positive serum bioassay (cases and controls) or clinician report of antibiotics administered prior to specimen collection (cases only). Two induced sputum specimens were routinely obtained from cases at the PERCH study South Africa site in order to estimate the sensitivity of Mtb culture using capture-recapture methods comparing the first and second induced sputum results. Using this method, the sensitivity of a single IS specimen was estimated to be between 15-50%. However, this is a setting with high HIV prevalence which may have higher sensitivity. The standard thinking is that the sensitivity of sputum or gastric aspirate has an upper bound of ~30% and a lower bound of about 10%. 27 Based on discussions with the Executive Committee it was decided that the South Africa PERCH study estimate in a high Mtb prevalence setting might be too high to apply to the other sites so the group decided to use the standard (10-30%) as the sensitivity prior for Mtb for all sites (i.e., those with high rates of TB as well as those with low rates). The lung aspirate and pleural fluid data were used to estimate aetiology for the cases with positive results. For extrapolation of the lung aspirate data to cases with unknown aetiology, these results were used to update the aetiology priors for other cases who were (1) at the same site and (2) clinically representative of those with a specimen obtained (see Section 5.1). As such, sensitivity priors were not a relevant parameter for these data in the analysis. Sensitivity of pneumococcal WB PCR was estimated by comparing the fraction of blood culture positive specimens that were also PCR positive. In the PERCH study, among cases with pneumococcus detected in blood by culture, 68% had pneumococcus detected by PCR in whole blood; this reduced slightly to 62.5% after applying the optimal quantitative PCR load threshold. 14, 15 We observed a similar proportion in pneumonia cases enrolled in a pilot study at the Kilifi site. 14 Because the integrated aetiology analysis estimates the sensitivity of WB given blood culture positivity, it cannot be also used to set the priors because these data then would be incorporated twice. Therefore, to determine the sensitivity prior of WB PCR we considered the evidence from the lung aspirate and pleural fluid specimens. Children who are positive on a pleural fluid specimen are likely different from those positive on lung aspirate since they have advanced to the state where the pathogen is able to travel outside the lung and into the pleural space; as such, they are likely to have a higher sensitivity for detection of pneumococcus on whole blood (i.e., more likely that the pathogen also moved into the blood) than for most pneumonia cases. Indeed, the evidence in the data supports this hypothesis, since high density WB PCR among cases positive for pneumococcus on pleural fluid was 80%, while for lung aspirate it was 33%. Starting from the non-informative sensitivity range of 2.8-97% (a distribution of beta(1,1)), we updated the range using the lung aspirate and pleural fluid data. For cases with consolidation (representative of those with lung aspirate specimens obtained), the 3/9 positive lung aspirate cases resulted in an updated sensitivity prior range of 12-65%. Given the small number of cases with pleural effusion to whom the pleural fluid results would apply, and in the absence of data to inform the sensitivity priors for cases with other infiltrate or normal CXR findings, we applied the lung aspirate range (12-65%) to all cases. Sensitivity prior Streptococcus pneumoniae 12-65% A partially latent class model (PERCH Integrated Analysis [PIA]), the BAKER method, was used to integrate the blood culture and induced sputum (TB only) results from the cases, and the NP/OP PCR and WB pneumococcal PCR results from both cases and controls to estimate the aetiology distribution for each individual case and for the population of cases, with the probability for each organism ranging from 0% to 100%. 24-26 The PIA model was run to determine aetiology of the latent cases, i.e., those without known aetiology. Those children with a known aetiology were handled differently, as described in Figure 1 and Table 14 . Used within the model to inform on the aetiology of the latent cases. Blood culture positive for more than one pathogen (green box, B) Single case (positive for two bacteria on blood culture, Salmonella species and Streptococcus pyogenes). Handled separate from the model because blood culture is assumed to be 100% specific in the model. Only used to inform the individual child's aetiology. Lung aspirate positive (green box, C) Excluded from the model run and appended to the output dataset to inform on the individual child's aetiology. Lung aspirate results were used to update the aetiology prior distributions for the remaining cases with x-ray consolidation using a multinomial likelihood to inform on the population aetiology. Yes (through updated aetiology priors) Pleural fluid positive (green box, D) Excluded from the model run and appended to the output dataset to inform on the individual child's aetiology. a Due to the small numbers of latent cases with a confirmed pleural effusion on CXR (i.e., those to whom the pleural fluid results would be applicable, see section 5.1), these positive results were not used to update their aetiology prior distribution. Measles (green box, A) Excluded from the analysis and described separately (i.e., not appended to final aetiology distribution), see Section 2.1. Yes No a Two cases whose pleural fluid was positive also had blood cultures positive for the same pathogen as detected on pleural fluid remained in the PIA model so that their blood culture data could inform the aetiology for other cases, but their pleural fluid results were not used as input measurements for the reasons described above. To estimate the all-site aetiology attribution (main paper Figure 4 ), a single model was run that included the cases and controls from all the sites, adjusting for site and age (age < 1 year and age > 1 year). The adjustment consisted of stratifying the cases and controls within the model by age and site to allow the aetiology to vary by strata. Including all sites and age strata in a single model (i.e., instead of running separate site-age stratified models) enabled estimation of an all-sites-combined aetiology estimate across the strata. The algorithm does this by using the stratum-specific blood culture data, prevalence and odds ratios for NP/OP PCR and pneumococcal WB PCR to inform on aetiology. At the same time the algorithm updates the sensitivity priors of each measurement by combining evidence across the site-age strata to obtain the posterior sensitivity values for the measurements. For example, the model updates the S. pneumoniae NP/OP PCR sensitivity for sites with no S. pneumoniae positive blood cultures based on the fraction of S. pneumoniae positive blood culture cases at other sites who were also S. pneumoniae positive on NP/OP PCR. At sites with positive lung aspirate data (The Gambia and South Africa), cases were categorized as representative of those with the lung aspirate specimens collected versus those not; the aetiology priors were updated using those results, but only for the subset of cases representative of those with lung aspirates, and only for the respective site (The Gambia or South Africa), informed by the distribution of pathogens detected on lung aspirates (purple boxes in Figure 1 ; see Section 5.1). Aetiology priors were not updated at the other sites without lung aspirate data (Table 15 ). Analyses were run for 25,000 iterations; the first 5,000 were excluded as a 'burn-in' stage and the remaining were sampled every 5 th iteration and were averaged to obtain the aetiology probability distribution of each case, ranging from 0% to 100% for each pathogen. At each iteration of the analysis, PIA assigns each case to a specific pathogen based on their measurements, the measurements of the 'population' of the other cases, the control measurements and the sensitivity and aetiology priors. In this Bayesian model, the unknown parameters to be estimated are the true cause of the infection for each child (aetiology) and the sensitivities for each measurement. The set of unknowns is assumed to follow a joint prior distribution representing prior uncertainty about their values. The Markov Chain Monte Carlo (MCMC) estimation method simulates a large number of values for all of the unknowns from their posterior distribution given the observed data. In this simulation, the estimate of the aetiology for each pathogen is approximately the average of the individual case probabilities and is provided with a 95% credible interval (CrI), the Bayesian analogue of the confidence interval. Convergence was checked by repeating the analysis multiple times using different starting seed values and comparing the posterior distributions between the runs. Cases with known aetiology who were excluded from the models (see Figure 1 and Table 14 ) were appended to the output dataset from the model at each iteration with aetiology probability assigned as follows: 1. Cases with 1 pathogen detected: assigned 100% to the aetiology detected for all iterations. 2. Cases with more than 1 pathogen detected: assigned 100% to the aetiology detected, where the 100% aetiology assignment for a given pathogen is distributed equally across the iterations relative to the number of pathogens detected. For example, a case with two pathogens detected (pathogens A and B) would have 50% of their iterations indicated as 100% aetiology for Pathogen A and the remaining 50% of the iterations indicated as 100% aetiology for Pathogen B. To estimate the aetiology results stratified by age or severity, a single analysis was run with all sites combined, including site, the case-level CXR strata (described above), and the stratifying variable of interest (i.e., age or severity) as regression components. This allows the aetiologies to vary by strata. Stratum-specific estimates (e.g., age < 1 year and age > 1 year, or individual sites, or severe and very severe) were calculated, in addition to calculating the estimate combined across the strata, as described above. To evaluate how the pathogen aetiologic distribution of pneumonia cases varies by site, we standardised the sitespecific aetiology distributions by age and pneumonia severity strata (according to World Health Organization definitions used pre-2013), using the average of the age and severity distribution across the sites as the standard. The age by severity breakdown amongst CXR+/HIV-cases (Table 17a ) and all HIV-Cases (Table 17b ) by site and averaged across sites are shown below. Given the similarity in the case mix between the CXR+/HIV-and the all HIV-cases, we elected to use the same case mix for standardisation in each of these analyses. For the standardised results, at each iteration of the MCMC posterior sampling, we assigned for each case a weight between 0 and 1 according to the standardising distribution that defines the weight based on the individual's covariate profile. The population level aetiology results were estimated for each site using the weighted datasets. Results are presented in the paper both as observed and standardised. Average distribution of age*severity strata across all sites. Appendix: Causes of severe pneumonia requiring hospital admission in children without HIV infection from Africa and Asia: the PERCH multi-country case-control study

projects that include this document

Unselected / annnotation Selected / annnotation