A Systematic Review and Meta-Analysis of Practices Exposing Humans to Avian Influenza Viruses, Their Prevalence, and Rationale

Abstract. Almost all human infections by avian influenza viruses (AIVs) are transmitted from poultry. A systematic review was conducted to identify practices associated with human infections, their prevalence, and rationale. Observational studies were identified through database searches. Meta-analysis produced combined odds ratio estimates. The prevalence of practices and rationales for their adoptions were reported. Of the 48,217 records initially identified, 65 articles were included. Direct and indirect exposures to poultry were associated with infection for all investigated viral subtypes and settings. For the most frequently reported practices, association with infection seemed stronger in markets than households, for sick and dead than healthy poultry, and for H7N9 than H5N1. Practices were often described in general terms and their frequency and intensity of contact were not provided. The prevalence of practices was highly variable across studies, and no studies comprehensively explored reasons behind the adoption of practices. Combining epidemiological and targeted anthropological studies would increase the spectrum and detail of practices that could be investigated and should aim to provide insights into the rationale(s) for their existence. A better understanding of these rationales may help to design more realistic and acceptable preventive public health measures and messages.

Bias due to confounding: -If present, were confounding factors adjusted for? Bias in the selection of participants into the study: -(Cohort studies) Were exposed and unexposed participants drawn from the same population? -(Case-control studies) Were controls sampled from the same population as cases? -(cross-sectional studies) Were inclusion and exclusion criteria applied uniformly? Bias in measurement of exposures: -Are exposures well defined? -Was information on exposures unaffected by knowledge of the infection status or risk of infection? Bias due to missing data: -Were infection and exposure data reasonably complete? Bias in measurement of outcome: -(cohort and cross-sectional studies) Was the case definition objective? -(cohort studies) Can we be confident that the outcome of interest was not present at start of the study? -(case-control studies) Were the case and control definitions objective and consistent?
Based on these questions, a judgement was made on the extent to which the results of the studies were at risk of bias. The options for the assessment of each domain, and for the overall assessment were: low, moderate, serious and critical risk of bias. If assessment in any domain was serious (or critical) the overall assessment was deemed serious (or critical) otherwise it was deemed moderate unless all domains were assessed as low. The results of the assessment are provided in Supplementary table 1.
Bias due to confounding. All 10 case-control studies either matched controls and cases or stratified the analysis according to participants' age, sex and/or location. However, only 6 case-control studies 2-7 conducted multivariate analysis and presented odds ratios adjusted for other poultry exposures. They were classified as being at moderate risk of bias. Six [8][9][10][11][12][13] out of the 18 casecontrol, cross-sectional and cohort studies only provided, for each exposure variable, crude odds ratios which were not adjusted for other exposure variables hence considered as having scope for serious bias from confounding. Because of their ecological nature, the 5 ecological studies were likely to be impacted by biases due to confounding. Ecological studies exploring the impact of live bird market closure on the risk of infection 14,15 did not assess the influence of some seasonal factors, such as seasonal variations in poultry production and marketing, which were likely to impact on the extent of viral circulation in poultry.
Bias in the selection of participants into the study. All study participants were generally all sampled from the same populations. In two case-control studies patients with respiratory illness were more likely to be tested for H5N1 infection if they reported recent contact with any or sick poultry, which might have biased the results towards those exposures 3,11 . Convenience sampling was used to recruit participants in one cross-sectional study 16 , and was likely to generate a selection bias limiting the external validity of the study. Information about the recruitment of study participants was limited for 3 other studies [17][18][19] . In a cross-sectional study, a third of recruited participants were excluded from the sample as their questionnaires were judged invalid. If there was an association between exclusion of participants and some features associated with poultry exposures (e.g. level of education, time available to answer questionnaire), this may generate bias limiting the validity of the study 13 .
Bias in measurement of exposures. All cohort, case-control and cross-sectional studies assessed exposures using self-report data, therefore, some misclassification of exposures was likely. Moreover, all case-control studies using clinical infection as part of the case definition included a substantial proportion of cases that had died for which information about exposure was obtained through the interview of proxies 2, 4-8, 11, 20 . These proxies might have been less aware of specific activities, resulting in a bias toward the underestimation of the association between an exposure and infection. For studies conducted during new epidemics (e.g. 4,5,8,11 ), media reports and high levels of alarm among the public might have led to difference in recall between cases and controls. The substantial delay between measurements of outcome and exposures in most case-control studies was also a source of recall bias 2, 4-8, 10, 20 . In one study, controls were interviewed later than cases, when the public was more aware of the role of poultry exposure as a risk factor for infection. Controls might have been more likely to report such exposures, resulting in an underestimation of the association between exposure and infection 6 . For all studies using serology to define cases, the presence of antibodies may not imply recent exposure given that the virus of interest was endemic in most investigated settings. Therefore, there was uncertainty about whether the reported exposures preceded, or not, the infection. Exposures were assessed in ecological studies using aggregated data, such as density of poultry and density of live bird markets. If the quality of this data was spatially variable, it might have led to bias. The studies exploring the impact of live bird market closure on the risk of infection 14,15 assumed that the infection pressure was constant before, and after live bird market closure, respectively. Yet, this infection pressure depended on the prevalence of infection among incoming poultry, which was likely to vary over time 21 . It might have resulted in a miss-estimation of the reduction in the probability of infection.
Bias due to missing data. All studies were judged at low risk of bias due to missing data. Infection and exposure data were reasonably complete, and there was no indication that possible missing data were non-randomly distributed among study participants.
Bias in measurement of outcome. In the cohort study, some participants were seropositive at the start of the follow-up period, but were kept in the cohort 17 . They were therefore classified as having not sero-converted, which might have biased the results towards an underestimation of the association between some exposures and the outcome. In studies using clinical infection as part of the case definition, it was unclear whether other potential disease aetiology which could have affected the case in combination with AIVs (eg Dengue 11 ) were investigated, and could have led to a misclassification of these cases. Studies using serological results as a marker of infection could have misclassified cases due to cross-reactivity between viral subtypes. This is more likely to be the case using hemagglutinin inhibition (HI) titres compared to microneutralisation (MN) titres. A serological survey defined individuals as infected based on a low antibody titre threshold, which might have led to false positive 18 . Conversely HI titres are less sensitive than MN titres leading to false negatives. Therefore, the use of HI titres only was considered at moderate risk of bias. While individuals had to be seronegative to be recruited as controls in most case-control studies, the serological status of controls was not assessed in two studies 6,11 . Spatial and temporal variation in surveillance effectiveness would have impacted the validity of the ecological study results. SUPPLEMENTAL

SUPPLEMENTARY TEXT 2. SENSITIVITY ANALYSIS.
For practices which were informed by at least 3 studies, pooled ratio estimates were re-assessed with studies grouped according to their risk of bias, moderate or serious. For 5 out of 6 practices, pooled odds ratios estimated based only on studies at moderate risk of bias were slightly lower than combined estimates (Supplementary table 2). This new estimates did not affect the conclusions we made in the main text. Therefore, we presented combined estimates. In the main text, the estimations of pooled odds ratios were for practices leading to direct exposure to poultry (touching, cleaning, preparing poultry) irrespective of the location of exposure. We re-computed these estimates by stratifying studies according to the location of the exposures of interest, households or premises (i.e. commercial farms, live bird markets and abattoirs). Pooled odds ratio estimates based on study conducted in households were generally higher than the combined estimates (Supplementary table 3). However, their associated confidence intervals were much wider, as the I 2 index also increased, meaning that study-specific odds ratios for households were highly heterogeneous. The small number of studies and the high level of heterogeneity between studies prevent us from drawing general conclusions about a possible increase in the strength of the association between direct exposure and infection when considering direct exposures taking place in households only. Moreover, several studies were assumed to be "household" studies but in fact did not specify the location where participants were exposed to poultry, and might have included some participants which were directly exposed to poultry in farms or markets. For instance, in 5 , 6% of cases and 1% of controls were defined as poultry workers (defined as deriving at least half of their income from work involving poultry), and the location where some exposures took place, such as "Poultry contact during slaughtering or processing" and "Poultry consumption" was not specified. Also, some studies only included poultry workers as participants 9,13,[17][18][19]24 . For these studies, all non-exposed participants (i.e. no reported direct exposure with poultry) in the baseline group were at least indirectly exposed to poultry. In contrast, when participants were selected from the general population, some non-exposed participants in the baseline group may not have had any indirect exposure to poultry. This may explain that some pooled odds ratio estimates based on study conducted in premises were lower than the combined estimates (Supplementary table 3).
The range of prevalence of each reported practice was re-assessed based on the countries where the studies took place (Supplementary table 4). For some practices, it appeared that most of the heterogeneity across studies could be partially explained by their geographical location. For instance, the proportion of study participants reporting burying/incinerating dead poultry ranged from 2% to 95%. However, the prevalence estimated in the 3 studies conducted in Bangladesh ranged between 2%-12%, and between 16%-19% in the 2 studies conducted in Egypt. It was much higher in Lao and Vietnam, ranging between 78%-87% and 76%-79%, respectively. However, it has to be kept in mind that some research groups produced several estimates for a given practice within a given country. Any differences in prevalence may therefore be more likely to be real as while these estimates were produced at different times, the study population may have remained the same, and the study protocol may be unchanged. As discussed in the main manuscript, the way in which prevalence were assessed-generally based on self-report-was likely to affect the accuracy of any estimates. Finally, the ranges of prevalence did not differ between the 2 study quality groups for most practices (Supplementary table 5). The proportion of participants who reported "selling sick and dead birds" and "throwing dead and sick birds in open spaces" was lower for quality 1 than quality 2 studies. It has to be noted that this quality assessment was only based on whether sampling was random and if regionally or nationally representative or not. Based on the information provided in the articles, it was not possible to assess the reliability of the answers provided by respondents. SUPPLEMENTAL

SUPPLEMENTAL TABLE 12
Risky and protective practices of which their association with AIV infection was assessed in the reviewed risk factor studies Exposure H. L.

Non-adj OR (95% CI)
Adj OR (95% CI) Author (year) Poultry in the vicinity: neighbors raising poultry Live poultry in neighborhood