Data extraction

The following data was extracted from each study: first author, year published, country, the sex/es studied, sample size, study design, ART follow-up years, initiating ART regimen (if reported), outcome/s analyzed, variable scale transformation methods, criteria for model variable/s selection (e.g. statistical methods and/or a priori clinical information), assessment of confounding and covariates adjusted for in the final model. For each of the final model variables, the unit and scale of measurement, effect sizes, 95% confidence intervals and, where available, standard deviations were noted. Effect sizes were rounded off to the nearest whole number and 95% confidence intervals and standard deviation to one decimal place. If ‘immunologic failure’ was mentioned, we checked if it was defined according to the WHO criteria [4].

Risk of bias was also assessed in each study as follows: Low risk—covariate adjusted for in model based on its clinical/biological plausibility; medium risk—covariates included based on both biological and statistical significance; and high risk—model employed only statistical significance (p-value). The provision by authors of biological reasoning, including references, for their covariate adjustments was noted.