War’s Civilian Deaths Are Mostly Not Caused by Weapons. Of 29 Million Attributed to Wars Since 1990, 21 Million Were Infection, Childbirth and Hunger.
Key takeaways · 12 min read
- In the largest global analysis, wars between 1990 and 2017 were associated with 29.4 million excess civilian deaths after battle deaths had been removed from the data.
- 21.0 million of those were communicable, maternal, neonatal and nutritional; 6.0 million were non-communicable disease; 2.4 million were injuries other than combat, mostly self-harm and interpersonal violence.
- Only wars registered. Minor conflicts, and conflict treated as a simple yes-or-no exposure, showed no statistically significant association with civilian mortality.
- Absolute harm was greatest in adults over 69 (317 per 100,000) and children under five (264 per 100,000); in relative terms it fell steadily with age.
Between 1990 and 2017 there were 1,118 separate armed conflicts somewhere in the world. The battle deaths were counted, argued over, and published. Then a team at Imperial College London asked a different question: once you subtract every battle death, what happens to the ordinary death rate of the civilians who are still alive?
Their answer, published in BMC Medicine in 2020, is that wars were associated with an extra 29.4 million civilian deaths over those twenty-eight years — none of them battle deaths, because battle deaths had already been removed from the data. Twenty-one million of the twenty-nine million were from communicable, maternal, neonatal and nutritional causes. Pneumonia. Diarrhoea. Childbirth. Malaria. Measles.
This is the part of war that does not photograph. It is also the part that most of the arguing about casualty figures is really about, because indirect deaths are estimated rather than counted, and estimates move. This article is about what the better estimates say, how much they move, and why one of the most quoted numbers in the field lost more than half its value when somebody changed a single assumption.
What “indirect” actually means
A direct death is a person killed by a weapon: shot, blown up, crushed by a collapsing building. An indirect death is a person who dies of something ordinary, at a time when they would not otherwise have died of it, because the war removed whatever was keeping them alive. The clinic that has closed. The vaccine cold chain that has no power. The road the ambulance cannot use. The crops nobody planted. The overcrowded camp where measles moves through a population of children whose immunisation was interrupted three years ago.
The distinction matters administratively as well as morally. Direct deaths get counted by journalists, human rights monitors and armies, and they can in principle be verified one at a time. Indirect deaths can only be inferred: you have to know what the death rate would have been without the war, and nobody observes that. Every indirect death estimate is therefore a comparison against a counterfactual that has to be constructed, and the construction is where the disagreements live.
What the 29.4 million were
Excess civilian deaths associated with wars, 1990–2017, after battle deaths were removed from the mortality data.
Jawad, M. et al., “Estimating indirect mortality impacts of armed conflict in civilian populations: panel regression analyses of 193 countries, 1990–2017”, BMC Medicine, 2020. Mortality rates were corrected to exclude battle-related deaths before analysis.
The headline effect size is that war — defined as a conflict producing at least 1,000 battle deaths in a country-year — was associated with an increase in age-standardised all-cause civilian mortality of 81.5 per 100,000 population, with a confidence interval running from 14.3 to 148.8. That interval is wide, and the authors do not hide it.
Only wars register at all
The most useful thing in the Imperial paper is a null result. The authors ran four different ways of measuring conflict exposure. When conflict was treated as a simple yes-or-no variable, the association with civilian mortality was not statistically significant — 26.4 per 100,000, with an interval running from minus 12.9 to plus 65.8, comfortably including zero. Minor conflicts, those producing between 25 and 999 battle deaths in a year, showed no significant association either. Only the top quintile of exposure, more than five battle deaths per 100,000 population, and the category defined as war produced a detectable signal.
This cuts against the intuition that any armed conflict damages a population’s health measurably. At country level, over a year, most of them do not — or at least, the damage is smaller than the noise in the data. What the evidence supports is narrower and more brutal: it is intensity that does the killing, and the relationship is not linear.
Who the indirect deaths fall on
Absolute increase in age-standardised all-cause mortality associated with war, per 100,000 population.
In relative terms the ranking changes: the proportional increase was largest in children under five and fell steadily with age. The elderly figure is large in absolute terms because their baseline death rate is already high.
Jawad et al., 2020. Female all-cause estimate 66.6 per 100,000, 95% CI −0.6 to 133.8 — the interval crosses zero, so the sex difference should not be read as established.
The number that halved
The single most quoted figure in this entire field is 5.4 million: the International Rescue Committee’s estimate of excess deaths in the Democratic Republic of the Congo between 1998 and 2007, produced from a series of five household surveys and reported in The Lancet. It was cited in Security Council debates and by every organisation working in the region, and it is the reason people confidently say that modern wars kill nine people indirectly for every one they kill with weapons.
In 2013 and 2014, a group of demographers at Southampton and Portsmouth went back to it with a different method — census data from 1984, two Multiple Indicator Cluster Surveys, and the 2007 Demographic and Health Survey, combined into a cohort component projection. Their estimate of excess population loss for the same conflict was 2.4 million for a closed population, and 1.7 million once migration was accounted for. Their conclusion was blunt: the choice of mortality baseline determines the level of excess population loss, and the IRC approach may have overestimated the scale.
Nothing about that finding says the Congolese wars were not catastrophic. The same authors note that Congolese mortality was exceptionally high regardless of which baseline or assumptions were used. What it says is that the difference between 5.4 million and 1.7 million is not a difference in the war. It is a difference in what you assume people would have died of anyway — and in a country with a very high peacetime death rate, that assumption is doing most of the work.
Same conflict, same decade, two methods
| IRC surveys | Demographic reconstruction | |
|---|---|---|
| Estimate | 5.4 million excess deaths, 1998–2007 | 2.4 million excess population loss (closed population); 1.7 million with migration |
| Method | Five retrospective household mortality surveys, 2000–2007, with a regional pre-war baseline | 1984 census, 1995 and 2001 MICS, 2007 DHS; indirect demographic techniques and cohort component projection |
| What drives the difference | The assumed counterfactual death rate. Both teams agree Congolese mortality was among the highest in the world with or without the war. | |
Coghlan et al. and successive IRC mortality surveys; Kapend, R., Hinde, A. and Bijak, J., “The Democratic Republic of Congo conflict (1998–2004): assessing excess deaths based on war and non-war scenarios”, 2013, and Kapend, R., doctoral thesis, University of Southampton, 2014.
How close is close enough to be at risk
A 2021 Lancet series paper on women and children took a geospatial approach instead: rather than asking how many died, it asked how many people live near enough to a conflict to be affected by one. In 2000, an estimated 185 million women and 250 million children were living within 50 kilometres of an armed conflict. By 2017 that had risen to 265 million women and 368 million children. A further 36 million children and 16 million women were displaced.
People living within 50 kilometres of an armed conflict
Non-displaced women and children, estimated from geospatial analysis. Nobody in these counts has to be shot at.
A further 36 million children and 16 million women were displaced in 2017.
Bendavid, E. et al., “The effects of armed conflict on the health of women and children”, The Lancet, 2021.
The same paper estimates that more than 10 million deaths in children under five between 1995 and 2015 can be attributed to conflict, and that women of reproductive age living near high-intensity conflicts have roughly three times the mortality of women in peaceful settings. Note the framing: near, not in. Nobody in these estimates has to be shot at. The mechanism is that a hospital fifty kilometres away stops functioning, or the road to it becomes unusable, or the midwife leaves.
What breaks, in what order
The cause-specific pattern in the Imperial data is a fairly precise description of a collapsing health system. Respiratory infections and tuberculosis rise, which is what happens when people are crowded together and treatment courses are interrupted. Enteric infections rise, which is water and sanitation. Malaria and neglected tropical diseases rise, which is vector control and drug supply. Maternal and neonatal disorders rise, which is skilled birth attendance.
The non-communicable disease finding is the more modern one, and the one most likely to matter in the conflicts of the next decade. A 3.4 per cent relative increase in cardiovascular, kidney, cancer and digestive deaths sounds trivial until you apply it to a middle-income country where those diseases were already the leading killers. Six million of the twenty-nine million deaths in the estimate are in this category. Insulin has a supply chain. Dialysis needs electricity. Chemotherapy needs an appointment system. A war does not have to reach a person to end their treatment.
And the injuries category, once battle deaths are removed, is driven by self-harm and interpersonal violence — two things that rise together in populations under sustained stress and with weapons close to hand.
Three sentences that are all defensible
Holding all three at once is the honest position. Discarding any one of them produces a version that is easier to argue with and less true.
What the authors themselves flag
The Imperial team list their own limitations at length, and they point in both directions. The conflict data come from the Uppsala Conflict Data Program, which relies on journalism and will miss conflicts nobody reported. The mortality data come from the Global Burden of Disease study, whose modelling smooths sudden shifts — which would make their estimates too low. Their analysis is ecological and at country level, so it cannot see the sub-national concentration of harm. Refugees who died in other countries are not in it at all.
Set against that: they detect an effect only in the most intense conflicts, and panel regression can be biased by unmeasured confounders that drive both conflict and mortality — state failure being the obvious candidate, since it causes wars and kills people by itself. The authors are explicit that their model is an association across an “average country” and not a causal claim about any particular war.
Which way each limitation pushes the estimate
| Toward too low | Toward too high, or simply unknown |
|---|---|
| Global Burden of Disease modelling smooths sudden shifts in mortality | Panel regression can be biased by unmeasured confounders that cause both conflict and death — state failure being the obvious one |
| Conflicts too small to be reported by journalists are missing from the conflict data | Country-level analysis cannot see sub-national concentration, so the “average country” effect may not describe any real place |
| Refugees who died outside the country are not counted at all | The association is an association, and the authors do not claim it is causal for any particular war |
Limitations as stated by the authors, Jawad et al., 2020.
As a sanity check they applied their own coefficient to the 2003 invasion of Iraq and got 20,902 indirect deaths for 2004, against Burnham and colleagues’ survey-based average of 16,181 a year. For Yemen in 2016 their model gives 21,603 against a UNDP estimate averaging 32,750. Two comparisons, one high and one low. That is roughly the accuracy the field currently has.
Questions people ask
Is it true that nine people die indirectly for every one killed by weapons?
That ratio comes largely from the IRC’s Congo surveys, and it is the figure most affected by the baseline problem described above. There is no stable ratio across conflicts: it depends on the intensity of the fighting, the strength of the health system beforehand, and the peacetime death rate you compare against. A war in a country with good hospitals and low background mortality produces a very different ratio from one in a country without them.
Why not just count the indirect deaths properly?
Because there is nothing to count. An indirect death is an ordinary death that happens to be premature; the death certificate, if one exists, says pneumonia. The only way to identify it is statistically, by comparing observed mortality with what would have been expected. That comparison is a model, and models are arguable in a way that a body count is not.
Does this mean casualty figures reported during a war are wrong?
Not wrong, incomplete. Figures reported during a conflict are almost always direct deaths, because those are what can be observed in real time. The indirect toll accrues over years and is usually estimated afterwards, if at all. The two are not competing numbers; they are counting different things.
Who is most at risk?
In absolute terms, adults over 69 and children under five. In relative terms, children under five, decreasing steadily with age. The effect estimate for males was significant and for females it was not, but the female confidence interval only barely crosses zero, and the study cannot distinguish a genuinely smaller effect from an imprecisely measured one.
Does this apply to a short, contained conflict?
The evidence says probably not, or not measurably. Minor conflicts — under 1,000 battle deaths in a country-year — showed no significant association with civilian mortality in the global analysis. The indirect toll appears to require sustained intensity, and it appears to accumulate over years rather than weeks.
The short version
- In the largest global analysis, wars between 1990 and 2017 were associated with 29.4 million excess civilian deaths after battle deaths had been removed from the data.
- 21.0 million of those were communicable, maternal, neonatal and nutritional; 6.0 million were non-communicable disease; 2.4 million were injuries other than combat, mostly self-harm and interpersonal violence.
- Only wars registered. Minor conflicts, and conflict treated as a simple yes-or-no exposure, showed no statistically significant association with civilian mortality.
- Absolute harm was greatest in adults over 69 (317 per 100,000) and children under five (264 per 100,000); in relative terms it fell steadily with age.
- The most quoted figure in the field — 5.4 million excess deaths in the DR Congo — was re-estimated at 2.4 million, or 1.7 million allowing for migration, by demographers using a different baseline.
- By 2017, 265 million women and 368 million children were living within 50 kilometres of an armed conflict, up from 185 million and 250 million in 2000.
- More than 10 million deaths in children under five between 1995 and 2015 have been attributed to conflict; women near high-intensity conflicts have about three times the mortality of women in peaceful settings.
- Every one of these numbers is a model output. The authors of the largest of them say their estimates are conservative in some respects and vulnerable to unmeasured confounding in others.
This article summarises published epidemiological and demographic research on mortality associated with armed conflict. It is not a political analysis, it takes no position on any particular conflict or its causes, and it is not guidance for anyone currently in or near one. Where estimates are contested, the contest is described rather than resolved.
Further reading: the Jawad et al. paper is open access at BMC Medicine and is worth reading for its limitations section alone, which is unusually candid. The Lancet 2021 series on women and children in armed conflict is the best single overview of the geospatial evidence.
- Why We Fight, Christopher Blattman (2022). An economist synthesizing decades of quantitative conflict research, including on casualty-data quality.
- Noise, Daniel Kahneman, Olivier Sibony & Cass Sunstein (2021). On unwanted variability in judgment — relevant to how wildly indirect-death estimates can vary.
- How to Talk to a Science Denier, Lee McIntyre (2021). On engaging contested claims without dismissing or overselling them.
Sources
- Jawad, M., Hone, T., Vamos, E.P., Roderick, P., Sullivan, R. and Millett, C., “Estimating indirect mortality impacts of armed conflict in civilian populations: panel regression analyses of 193 countries, 1990–2017”, BMC Medicine 18:266, 2020. (1,118 unique armed conflicts. War associated with +81.5 per 100,000 age-standardised all-cause civilian mortality, 95% CI 14.3–148.8; binary conflict exposure +26.4, 95% CI −12.9–65.8, non-significant; minor conflict non-significant; fifth quintile of exposure +99.2, 95% CI 28.2–170.3. Post-estimation totals: 29.4 million (22.1–36.6) civilian deaths attributable to wars, of which 21.0 million (16.3–25.6) communicable, maternal, neonatal and nutritional, 6.0 million (4.1–8.0) non-communicable, 2.4 million (1.7–3.1) injuries. Under-5s +263.7 per 100,000, over-69s +317.3; males +97.0, females +66.6 with a confidence interval crossing zero. Model applied to Iraq gives 20,902 indirect deaths in 2004 against Burnham et al.’s 16,181 per year; applied to Yemen gives 21,603 in 2016 against a UNDP average of 32,750.)
- Kapend, R., Hinde, A. and Bijak, J., “The Democratic Republic of Congo conflict (1998–2004): assessing excess deaths based on war and non-war scenarios”, 2013; and Kapend, R., “The demography of armed conflict and violence”, University of Southampton, 2014. (Against the IRC’s 5.4 million excess deaths for 1998–2007, this reconstruction from the 1984 census, the 1995 and 2001 MICS and the 2007 DHS gives 2.4 million for a closed population and 1.7 million allowing for migration. The authors state that the choice of mortality baseline determines the level of excess population loss and that the IRC approach may have overestimated the scale; they also state that Congolese mortality was exceptionally high regardless of baseline.)
- Bendavid, E., Boerma, T., Akseer, N., Langer, A., Malembaka, E.B. and Okiro, E.A., “The effects of armed conflict on the health of women and children”, The Lancet, 2021. (Approximately 36 million children and 16 million women displaced in 2017. Population living within 50 km of armed conflict rose from 185 million women and 250 million children in 2000 to 265 million women and 368 million children in 2017. More than 10 million deaths in children under five attributed to conflict between 1995 and 2015. Women of reproductive age near high-intensity conflict have roughly three times the mortality of women in peaceful settings.)
- Roberts, L., Zantop, M., Ngoy, P., Lubula, C. and Mweze, L., “Elevated mortality associated with armed conflict — Democratic Republic of Congo, 2002”, JAMA, 2003, and the International Rescue Committee mortality survey series, 2000–2007. (The original survey programme behind the 5.4 million figure. Crude mortality of 3.5 per 1,000 per month in the east and 2.0 in the west in 2002; the majority of deaths attributed to preventable infectious disease rather than violence.)
- Wise, P.H., “The Epidemiologic Challenge to the Conduct of Just War: Confronting Indirect Civilian Casualties of War”, Daedalus, 2017. (Argument that most civilian casualties result from the destruction of food, water, shelter and health care rather than direct exposure to weapons, and that the ability to measure and to mitigate these effects has improved enough that ignoring them is no longer defensible.)
- Murray, C.J.L. et al., “Armed conflict as a public health problem”, BMJ 324:346, 2002. (Early statement of the measurement problem: conflict causes health consequences through displacement, breakdown of health and social services and heightened disease transmission, and quantification methods were then inadequate to capture them.)
- Guha-Sapir, D. and Terán Gómez, V., “Angola: The Human Impact of War”, Centre for Research on the Epidemiology of Disasters, 2006. (Field-survey review setting out the standard direct/indirect framework and the main causes of indirect death: economic collapse, food shortage and malnutrition, disruption of health systems, mass movement into overcrowded settlements.)
