Data synthesis
Studies were grouped according to the data they contained relevant to this systematic review’s three co-primary objectives. Group 1 studies reported the proportion of hospital admissions attributed to ADRs, group 2 studies reported the proportion of hospital admissions prolonged by ADRs, and group 3 studies reported the proportion of in-hospital deaths attributable to ADRs; a study could be included in more than one group. Since findings could be reported on the level of the patient or on the level of the admission (i.e., allowing for re-admissions), we decided to use admission-level data if reported, and patient-level data if no admission-level data were reported.
In the primary analysis, we pooled data from all studies in each group, regardless of study quality, study setting, methodological considerations, or study populations. We tested for heterogeneity using I2 statistics to decide whether to conduct meta-analysis: in the presence of heterogeneity, we would summarise the proportions mentioned above as medians and interquartile ranges; in the absence of heterogeneity, we would conduct random-effects meta-analysis, calculating the pooled estimate proportion after Freeman-Tukey double arcsine transformation to stabilise the variances.
We explored the proportions mentioned above by subgroups relating to study quality, study setting, methodological considerations, and study populations. Depending on heterogeneity within each subgroup (using I2 statistics) we would proceed to synthesise the data as above.
A similar approach was followed for the secondary objectives, where we summarised serious preventable ADRs. For the remainder of the secondary and explorative objectives, we narratively summarised clinical presentations commonly reported, drugs commonly implicated, and the contribution of HIV, ART, and methodological and quality issues.