Protocol and registration
The review was developed in line with the PRISMA guidelines (8). The original review protocol, as well as its amendments, were registered in PROSPERO (CRD42020203924).
We sought records published after 2010 which report on pneumonia in children between four weeks and five years of age in SSA. Studies reporting exclusively on children below four weeks of age were excluded because the etiological pattern of pneumonia is different in this age group (9). Studies were included if they reported on the frequency of bacterial causes of pneumonia in the relevant population, with or without comparison. Studies using culture or molecular methods on blood or any type of respiratory sample (nasopharyngeal swabs, induced sputum, lung aspirate) were eligible. We only included primary research and considered the following study designs: case series, surveillance, cross-sectional, case-control, cohort, and interventional studies. Modelling studies and reviews were not eligible.
We searched the following electronic databases without language restrictions: MEDLINE using the PubMed interface (last search 10 October 2020), Web of Science database (last search conducted 16 October 2020), and African Index Medicus (last search 2 October 2020). The MEDLINE search was restricted to articles published after 1 January 2010; no restrictions were applied to other searches. We manually searched the reference lists of included records for other potentially relevant records.
Our search strategy combined the key themes of the review question: (a) bacterial pneumonia (b) children and (c) sub-Saharan Africa. For each of the themes, we applied alternate terms and spelling combinations, including truncations and wildcards to improve sensitivity. This search strategy was applied to MEDLINE and Web of Science; in the search of African index Medicus, we omitted the theme of SSA. Full details of the search strategies and syntaxes are available as supplementary material (supplement 1).
Screening of titles and abstracts and full-text screening for eligibility was conducted by blinded double-voting, with a third vote to resolve disagreements. CO, BE and OI screened the titles and abstracts while VW resolved disagreements. Potentially eligible records from the title and abstract screening were considered for full-text assessment. The assessment of the full-texts was conducted by BE, OI and VW, with CO acting as a tiebreaker to resolve disagreements. MJ and CO subsequently searched the reference list of records included in the review for potentially relevant records. We used the Covidence platform (https://www.covidence.org/about-us-covidence/) to organize the screening and selection of records.
Data collection process
We developed a data extraction form, implemented it in Covidence, and refined it after a pilot phase using five included records. Next, one member of the review team extracted the relevant data items from all the included papers and a second member checked the extracted data. Disagreement between the primary extraction and data check was resolved by consensus between voting members in consultation with a third member of the team. No additional information was sought from investigators or authors.
The following categories of information were extracted: (a) study characteristics (study aim, design, and start and end date), (b) characteristics of the study population (description of cases, pneumonia case definition, method of recruitment, the severity of pneumonia and number of children screened for pneumonia if applicable); (c) type of outcome measure (sample type, method of sample collection, method of bacterial identification, total number of samples collected, number of samples with positive test results for bacteria, and number of specific bacterial isolates). For case-control studies, we extracted similar data items for the control subjects.
Risk of bias in individual studies
Two members of the review team assessed the risk of bias, with disagreement resolved by consensus. We used the Joanna Briggs Institute (JBI) quality assessment tools for assessing the quality of included studies (10). As this review focuses on the cases with pneumonia we used the JBI tool for case-series for both case series and case-control studies.
Summary measure and analyses
Since this review aimed to summarize the prevalence of specific bacterial agents among cases with pneumonia, our main summary measure was the proportion of pneumonia cases with specific isolates. We first conducted meta-analyses of these proportions per sample type and per pathogen, using a random-effects model and after a variance-stabilizing transformation (double arcsine transformation). Second, for case-control studies of nasopharyngeal isolates, we also conducted a meta-analysis of the crude odds ratios of bacterial isolation comparing children with and without pneumonia. We assessed heterogeneity by computing Cochrane's Q and I2 statistics which measure the proportion of the variation between studies that is due to heterogeneity and not by chance (11, 12). R (package metafor) and STATA version 16 were used to conduct the analyses and to produce forest and funnel plots (13).
Risk of bias across studies
Assessing the risk of publication bias in meta-analyses of prevalence studies is not straightforward, as the prevalence is expected to vary across studies and funnel plots may not be relevant (14). We, therefore, discussed the possible presence of bias across studies and the implications it may have had on our findings without making a quantitative evaluation. For the second type of meta-analysis in this review, i.e. the association between pneumonia and nasopharyngeal isolation of S. pneumoniae (expressed as odds ratio (OR)), we did construct a funnel plot. To assess the funnel plot symmetry, we relied mostly on visual inspection of the plot, with support from formal statistical tests (formal tests for asymmetry are underpowered when the funnel plot has fewer than 10 studies)(14).