Study Selection Procedures
Figure 1 summarizes the results of the study search and selection process. A total of 10,196 studies were identified in the databases after removal of duplicates. During title and abstract screening, 10,068 were excluded, leaving 129 studies for full-text review, of which, 24 studies were included in the systematic review and 6 in the meta-analysis.
The main reasons for exclusion in the review were (1) Reports outside the study scope, (2) Studies not related to review objectives, (3) estimation of excess mortality among patients with a specific disease instead of a population and/or cohort, and (4) the use of a comparator which was less than 1 year in the estimation of the expected number of deaths in the calculation of excess mortality. The main reasons for exclusion from the meta-analysis were that studies did not specify the population size, the number of expected deaths (all-cause mortality), the number of observed deaths, or the methods for estimating excess mortality.
Study Characteristics
The characteristics of the 24 included studies are summarized in Table 1. Studies were published between 2020 and 2023 but most were published in 2021 (13 studies). Five studies were conducted in low-income countries and 19 in lower-middle-income countries (Figure 2). Most of the studies were conducted in Asia, including Iran (7). India (4), Bangladesh (2), and Indonesia (2). There were 6 studies from Africa and none from Latin America or the Caribbean. Sanmarchi et al.30 reported estimates from 5 countries, making it a total of 29 countries in the review (Figure 3).
For the meta-analysis, 10 countries were included from 6 studies. In 7 countries, the observed deaths were higher than expected [India (Leffler et al.31 and Lewnard et al.22), Iran (Safavi-Naini et al.33), Kyrgyzstan (Sanmarchi et al.30), Uzbekistan (Sanmarchi et al. 30), Tunisia, (Sanmarchi et al.30), Bolivia (Sanmarchi et al.30)]. In three countries, negative excess mortality was recorded, thus the observed deaths were lower than the number expected in the absence of the pandemic [Indonesia (Wijaya et al. 34), Kenya (Otiende et al. 35) and Mongolia (Sanmarchi et al.30)].
Risk of bias (quality) assessment
Quality assessment was conducted for the studies according to the Newcastle-Ottawa Scale (NOS) score by two reviewers, and the disparity was solved by discussion and/or consulting a third reviewer. A typical instrument designed for meta-analysis of observational research is the Newcastle-Ottawa Scale (NOS). Because it is simple to use and has gained widespread adoption as a result of Cochrane Collaboration guidelines, the NOS is appropriate for systematic reviews.
Estimate of excess mortality in LLMICs
Table 2 provides an overview of population and mortality data reported by the studies included in the meta-analysis. During the COVID-19 pandemic, of the total 1,398,858,717 individuals/populations, 3,555,880 all-cause deaths were reported, while 2,152,474 deaths were expected from the eleven countries. The pooled excess mortality was 100.3 deaths per 100,000 population. The excess risk of death was 1.65 (95% CI: 1.649, 1.655 p<0.001). There was a high heterogeneity as indicated by the I2 of 100%.
In 7 countries, the observed deaths were higher than expected, whilst, in three countries, negative excess mortality was recorded, thus the observed deaths were lower than the number expected in the absence of the pandemic.
Methods and data used to estimate excess mortality in LLMICs.
The 24 articles used four distinct methods/study designs to determine excess mortality. The largest group of studies (15 articles) used retrospective data of already existing mortality datasets (Hanifi et al.40; Ahmadi et al.41; Jha et al.42; Tadbiri et al.43; Ghafari et al.44; Otiende et al.35; Watson et al.45; Safavi‑Naini et al.33; Lewnard et al., 22; Ghafari et al.46; Rasambainarivo et al.47; Acosta et al. 32; Elyazar et al.48; Hedstrom et al.39) to estimate excess mortality. Two studies used quantification of burial sites by observing the increase in the number of burial grounds to estimate excess mortality (Besson et al.,25; Warsame et al.,26). One study used a cross-sectional survey through a household census (Barnwal et al.,27) and another used grey literature (use of already published figures from journalists and organizations) (Leffler et al.31) to estimate excess mortality.
With regard to the source of data, four studies used more than one data source to estimate excess mortality. This included burials in public cemeteries + civil death registration + health authority death registration (Elyazar et al.48), daily mortality/incidence data from the Syrian Ministry of Health + Excess all-cause mortality data from a statement by the Damascus governorate + obituary notification data from Facebook page (Watson et al.45), National survey data + health facility deaths (Jha et al.42) and figures published by regional governments and Indian journalists + government hospital data + funeral counts + handwritten death registers (Leffler et al.31).
All other studies relied on only one data source. Five studies used National Civil Registration Data (Safavi‑Naini et al., 33; Tadbiri et al.43; Ghafari et al., 44; Ghafari et al., 46; Lewnard et al., 22). Two studies each used the Health and Demographic Surveillance System (Otiende et al.35; Hanifi et al., 40), death registers (Acosta et al.32; Rasambainarivo et al.47) and imaging of burial sites/grounds (Besson et al.25; Warsame et al.26). One study used only primary data (census/survey) data (Barnwal et al., 27) and another study used Bureau of Vital Statistics data (Ahmadi et al. 41; Wijaya et al. 34) to estimate excess mortality.
Studies used several different methods to determine the expected deaths that were used to calculate excess mortality. Twelve studies used modelling techniques to estimate excess mortality. Of these, five studies used linear regression (Lewnard et al.; 2021; Wijaya et al., 34; Leffler et al.31; Tadbiri et al. 43; Ghafari et al., 44), two studies used auto-regression modelling techniques (Rasambainarivo et al. 47; Safavi‑Naini et al., 33). Two other studies used geospatial analysis which involves identifying new grave plots and measuring changes in burial surface area over a period (Besson et al., 25; Warsame et al. 26) and two studies used estimation of death counts (Hedstrom et al., 39; Elyazar et al.48). Other modelling techniques used included Cox proportional hazard models, Auto-Regressive Integrated Moving Average, model fit, multilevel regression model (full bayesian model).
Factors influencing excess mortality in LLMICs.
In assessing the factors that might have influenced excess mortality, of the 24 studies, only one reported differences in mortality between rural and urban areas (Jha et al. 42). They found that excess deaths in the first wave of the pandemic were concentrated in urban areas, while deaths in the second wave affected both urban and rural areas. Other studies speculated what could have caused excess morality without empirical evidence in their data. No study reported disaggregated information by socio-economic status.