Design and setting
This study used longitudinal (panel) regression models to assess the impact of armed conflict on mortality rates in 193 countries between 1990 and 2017. Country was the unit of analysis. Longitudinal regression methods are frequently employed in health and economic studies to examine dynamic and transitional associations over time using routine data; examples include the association between democratisation and recession on health [26, 27], and between government health spending on the incidence of disease [28]. Fixed effects specifications were employed, which adjust for time-invariant country-level factors, and can be considered similar to differences-in-differences approaches with multiple time points.
Data
We combined several datasets for our analysis. We obtained armed conflict data from the UCDP Georeferenced Event Dataset (GED) Global v19.1 [29]. The UCDP has collected country-year armed conflict data since 1946, and event-specific data since 1989, which have been used extensively for research purposes [5, 10, 12, 17, 18, 21]. The UCDP GED provides the best estimate of the number of battle-related deaths at the village-day level for all armed conflicts that meet the UCDP definition of having at least 25 battle-related deaths per calendar year [29]. We took data from 1990 to 2017 and converted it into a country-year dataset. The UCDP GED assigns armed conflict events in Palestine to Israel, misclassifying the former as being conflict-free, so both countries were removed from the dataset. Further details on the UCDP methodology are shown in Additional file 1 (Paragraph 1.1).
We obtained country-level mortality data between 1990 and 2017 from the Global Burden of Disease (GBD) study [30]. The GBD study utilises data from various sources to produce internally consistent, statistically modelled annual estimates of mortality by causes of death for each country. Estimates are not restricted to country citizens or nationals, but include migrants and refugees, including those displaced by conflict. The GBD excludes direct battle-related deaths from initial all-cause mortality calculations, reintroducing them in later calculations. Thus, battle-related deaths are listed as a separate cause of death (Category C3.3: terrorism and armed conflict) [30]. The GBD is appropriate for studying excess deaths from armed conflict because no indirect consequences of armed conflict are built into the underlying mortality estimations. Although GBD data are modelled with smoothing functions, investigation of mortality trends for selected countries experiencing conflict shows GBD smoothing functions did not obscure abrupt changes in mortality and are unlikely to introduce bias to this analysis (Additional file 1: Fig. S1).
We developed a conceptual framework (Additional file 2) using existing conflict and health literature which informed the selection of covariates. The framework consisted of seven pathways through which armed conflict may affect civilian mortality, and seven drivers of conflict which were used as potential confounders. To capture changes in country wealth and income, we used the gross domestic product (GDP) per capita in current US dollars, available from the World Bank [31], and membership of the Organisation for Economic Co-operation and Development (OECD) [32]. For changes in the degree of democratisation, we used the Varieties of Democracy (V-Dem) dataset, specifically the Multiplicative Polyarchy Index (continuous variable, range 0 to 1, created by multiplying five core components of electoral democracy) [33]. We used V-Dem rather than the more commonly used POLITY IV [5, 7, 10, 12, 13] due to better data completeness. For changes in demographic factors, we used data on the proportion of people living with a density > 1000 people/km2, the proportion living in urban areas (both taken from the Institute for Health Metrics and Evaluation (IHME) [34]), and the age dependency ratio (ratio of under 15s and over 64s to the working population) taken from the World Bank [31]. For changes in the average levels of educational attainment in countries, we used the mean years of educational attainment per capita, separately for males and females, available from the IHME [34]. We captured changes in ethnic group composition using the Historical Index of Ethnic Fractionalisation Dataset, which corresponds to the probability that two randomly drawn individuals within a country are not from the same ethnic group [35]. Finally, to capture changes and shocks in climate-related factors, we took data from the Emergency Disasters Database to control for the presence of droughts and earthquakes [36], and the IHME to control for the population-weighted mean temperature and the proportion of people living in the 5th quintile of annual rainfall. We were unable to find a suitable dataset for income inequality, although it is plausible that its effects are captured through GDP per capita and educational attainment. Unlike previous research, we did not adjust for health expenditure [5, 7,8,9], the prevalence of HIV/AIDS [11, 13], and refugee movement [5, 15] as these can be considered mediators (i.e. factors on the causal pathway between conflict and health) rather than confounders.
Measures
Our main outcome measures were all-cause and cause-specific mortality rates, as reported by the GBD study [30], but with deaths due to terrorism and armed conflict (Category C3.3) removed. Removing battle-related deaths from all-cause mortality prevents the bias of including battle-related deaths in both the explanatory variable and outcome measure which is commonly found in previous studies [6,7,8,9,10,11,12,13, 16, 24, 25]. The GBD study categorises cause-specific mortality into first-order (communicable, maternal, neonatal, and nutritional diseases; non-communicable diseases [NCDs]; injuries) and second-order causes [30].
As explanatory variables for armed conflict, we explored four different specifications. Firstly, we used a binary variable indicating the presence of armed conflict per country-year observation (0 = no [< 25 battle-related deaths per country-year], 1 = yes [≥ 25 deaths]) as per the UCDP [1]. This approach is limited as it groups all conflicts together regardless of intensity, so a second specification was the rate of battle-related deaths per 100,000 population as a continuous measure. We used the rate rather than absolute battle-related deaths to avoid bias from emphasising small conflicts in populous countries which are unlikely to exert country-wide effects. Thirdly, as previous research has shown the relationship between battle deaths and civilian mortality to be non-linear [24], we explored quintiles of the rate of battle-related deaths per 100,000 population. Fourthly, we used intensity cutoffs as per the UCDP [1] to create a categorical conflict variable (0 = no conflict [< 25 battle-related deaths per country-conflict-year], 1 = minor conflict [25–999 deaths], 2 = war [≥ 1000 deaths]). We used county-conflict-year for this final specification to prevent misclassifying large countries with multiple small conflicts as war-affected (should the total of these conflicts be greater than 1000 battle deaths).
Statistical analysis
We described our sample using frequencies and means and reported the rates of battle-related deaths and civilian mortality by the different conflict explanatory variables. We presented graphical time trends in the number of conflicts, the number and rate of battle-related deaths, and the mean battle-related deaths per country.
We then used fixed effects linear panel regression methods to assess the relationship between armed conflict and mortality, which was estimated using the following equation:
$$ {\mathrm{Mortality}}_{it}={\beta}_0+{\beta}_1{\mathrm{Conflict}}_{i\left(t-1\right)}+{\beta}_2{\mathrm{Covariates}}_{it}+i+t+{u}_{it} $$
where i is the country, t is the year, and u is the error term representing unexplained variation. We chose fixed effects over random effects given the likelihood that our error term u was correlated with our covariates (a key assumption that would be violated in random effects) and as indicated by the Hausman test. Fixed effects specifications control for time-invariant observed and unobserved country-level factors (as denoted by i in the equation above) and therefore only assess within-country associations rather than between-country associations. Model fit was assessed using scatter plots and post-regression diagnostics (including the variance inflation factor (VIF), studentised residuals, stem and leaf plots, Cook’s D, and DFITS) to identify collinear variables and outliers. Due to collinearity with female education (VIF > 30) and superior goodness-of-fit, we only included male education. Four data points were considered outliers and dropped (from the model only using a continuous measure for armed conflict) based on having large residuals and high leverage: Rwanda 1994, Bosnia and Herzegovina 1995, Congo 1997, and Eritrea 1999. Four of twelve covariates had missing data: age dependency ratio (4.3%), GDP per capita (5.8%), Multiplicative Polyarchy Index (11.7%), and Ethnic Fractionalisation Index (31.6%); the latter two were omitted from main analyses due to high levels of missing data, but their effects tested in sensitivity analyses (see below).
Our first models separately tested the association between the four explanatory variable specifications for armed conflict and age-standardised all-cause mortality, adjusting for ten covariates (GDP per capita, OECD member, population density, urban living, age dependency ratio, male education, temperature, rainfall, earthquakes, droughts), and year-country fixed effects. Mortality rates were standardised using the GBD world population standard. We included country-clustered Huber-White robust standard errors to account for possible heteroscedasticity and serial correlation. In all models, we lagged the armed conflict variable by 1 year unless otherwise stated to capture indirect deaths in the year after.
Because higher intensity armed conflicts drove the associations, our second models employed the same specification as above but using only the categorical armed conflict exposure variable for war (we present “war” vs. no conflict). These models used age-standardised all-cause and cause-specific mortality rates (first- and second-order causes; the latter helping to explain underlying drivers) and age- and sex-stratified all-cause and cause-specific mortality rates (first-order causes only). We then assessed the lagged effects of war (from 2 to 10 years with a separate model for each year lag) on age-standardised all-cause and cause-specific (first-order causes only) mortality to account for the varying progression of disease pathologies. We used the models’ “war” beta coefficients to calculate the absolute and relative change in civilian mortality and used post-estimation commands to calculate the number of civilian deaths.
In additional analyses, we explored whether armed conflicts that involved particular actors (“armed conflict actor type”) had differential associations with civilian mortality. In accordance with UCDP definitions, we categorised each armed conflict as being state-based (at least one actor is the state of a country), non-state (no actors are the state of a country), or one-sided violence (one actor in armed conflict with civilians). We analysed all four explanatory variable specifications for conflict (i.e. binary, continuous, quintiles, and categorical) by armed conflict actor type, and for countries with only one actor type in each year, we interacted actor type and the rate of battle-related deaths, leaving all specifications identical to prior models.
Sensitivity analyses
We undertook multiple sensitivity analyses to check the robustness of findings. First, we tested alternative model specifications with sequential addition of covariates, addition of new covariates that contained high rates of missing data, and random effects. Second, to address the possibility of misclassification bias between deaths due to armed conflict and deaths due to homicide, we repeated our main model removing deaths from interpersonal violence. Third, we used an alternative measure of armed conflict exposure derived from the Major Episodes of Political Violence dataset compiled by the Centre for Systemic Peace [37]. This dataset captures major conflicts globally, and its binary conflict variable (0 = no conflict, 1 = conflict) is 97.6% specific and 97.7% sensitive to the UCDP “war” binary variable (0 = no conflict, 1 = war). More details about how this dataset compares to the UCDP are presented in Additional file 1: Paragraph 1.2. All analyses were conducted in Stata 15.