2.4. Statistical analyses

To investigate whether there were systematic differences between those that successfully attended an RHU compared to those who did not, Student's t-test and Pearson's chi-square test were used. All demographic characteristics, TB symptoms and dimensions of social trust were tested.

Participants represented a wide range of communities from across the Philippines. To account for unique geographical factors of each community and distinctive dynamics and interactions from different ICM field office staff, multilevel modelling was used. All ICM regional field office sites and the communities were assigned a unique identifying number. The models were fitted using Markov chain Monte Carlo (MCMC) techniques provided by the ‘MCMCglmm’ package (Hadfield, 2010). Although the ensuing analyses take a Bayesian approach, no priors were defined. To our knowledge, this is the first investigation into the multilevel effects of social trust on TB health seeking behaviours in the context of extreme poor settings in the Philippines. Although there are studies that have examined social trust as a predictor for various health outcomes in other low and middle income countries (LMICs) (Agampodi et al., 2015; Grover et al., 2016), social processes differ considerably by socio-cultural context. Given the dearth of evidence in the Philippine context, we did not feel confident constructing priors that would substantially influence the outcomes of the analyses.

In the multilevel models, individual participants formed level 1, and both unique community ID and field office site ID were fitted as random effects, forming level 2. Univariate odds ratios (ORs) were first produced before the models were subsequently fitted using stepwise regression, with an aim to minimize the deviance information criterion (DIC). To examine the relative influence of the different variables, Model 1 adjusted only for the social trust variables, and models 2 and 3 adjusted for only the demographic and health-related variables, respectively. Model 2 had the lowest DIC but it was only marginally lower than the full Model 4 which adjusted for all the aforementioned variable domains and thus was selected as the final model. Model diagnostics were assessed to ensure convergence of the model (Supplementary File 1).

All analyses were conducted using R version 3.3.3 (R Core Team, R Foundation for Statistical Computing, Vienna, Austria).