The Economics of Polycrisis Prevention: Quantifying the Human and Economic Returns of Resilience

Across 170 countries and 25 years, 72% of country-years have no recorded damage and 48% no recorded deaths. The mean damage across all country-years is 0.23% of GDP, but the worst 1% of country-years (about 42) account for 58% of the total and the worst 5% for 86%. Deaths are even more concentrated: 76% of the death-rate total sits in the worst 1%. Sixteen countries recorded a year with damage of 10% of GDP or more, 27 of 5% or more and 70 of 1% or more. The 1.55 million recorded deaths include 233,000 in Haiti (the 2010 earthquake), 188,000 in Indonesia (the 2004 tsunami), 141,000 in Myanmar (Cyclone Nargis in 2008), 114,000 in China, 91,000 in India and 89,000 in Pakistan: together 55%.

The income groups differ less than might be expected on the mean and more on what the median country experiences. Mean damage is 0.21% of GDP in high-income countries, 0.27% in upper-middle, 0.26% in lower-middle and 0.15% in low-income countries. But the median country's own mean is just 0.05%, 0.15%, 0.02% and 0.07% respectively, because a few countries with huge years pull the means up. For deaths the mean is 1.0 per 100,000 in high-income countries, 0.8 in upper-middle, 2.7 in lower-middle (inflated by Haiti 2010) and 0.6 in low-income, against medians of 0.5, 0.1, 0.2 and 0.2. A government deciding what to spend is therefore choosing against a record in which the typical year is uneventful and the average is carried by tails.

Two caveats matter for what follows. EM-DAT includes an event if it killed ten or more people, affected a hundred or more, or triggered a declaration or appeal, and leaves the damage field empty for many events; in low-income countries especially, a recorded zero often means unmeasured. And disasters are one channel of polycrisis. They are used here because their record is public, annual and comparable across 170 countries, and because the early-warning investment that addresses them has a published unit cost and effect size. Section 5 sets out the evidence for other channels.

Take the reference investment at face value: $0.50 per person a year, avoiding 30% of the damage and 30% of the deaths that the 2000–2024 record shows. Applied to all 8.0 billion people it would cost $4.0 billion a year and, on the 2000–2024 record, would have avoided about $62 billion of damage and 21,200 deaths a year (the 2000–2024 death rates imply about 70,700 deaths a year at today's population; recorded deaths averaged 62,000 a year at the populations of the time). The combined asset return is 15.5 to 1, but it is dominated by rich countries with valuable assets: 51 to 1 in high-income countries, 14.5 in upper-middle, 3.2 in lower-middle and 0.5 in low-income countries. At the country level the median asset return is 3.1 and 60% of countries clear break-even.

On lives, the picture is different. The expected cost per life saved is $155,000 in high-income countries, $217,000 in upper-middle, $166,000 in lower-middle and $333,000 in low-income countries, with $189,000 overall. The lives are concentrated where the big events were: lower-middle-income countries account for 41% of expected lives saved on 36% of the population. In the scatter, countries in the top right (Haiti, Myanmar, Sri Lanka, Indonesia) return both; those at the top left (Somalia, for example, with an asset return of 1.8) save lives for little in assets; those in the bottom right pay in assets only (Japan, New Zealand, the United States, Chile). Many poor countries sit at the left, where assets alone barely repay the investment on the record.

Two readings follow. First, for the poorest countries the recorded asset return is a floor, not a ceiling: damage is under-recorded, informal assets and livelihoods are not counted, and the valuation of a life is separate. A life valued at 100 times GDP per person is worth about $70,000 in a country with $700 GDP per person, so the value of lives saved would add little to the dollar return in a low-income country; the case for early warning there is humanitarian, and the calculator reports the value of a life at which the investment would break even. Second, the unit cost is an average: UN estimates apply the 50 cents to the people not yet covered, so the actual bill is $0.6 billion a year, not $4.0 billion, and the return on the right target is higher than on this uniform benchmark.

An investment whose expected benefit-cost ratio is B recovers its cost over a window when the window's average loss is at least 1/B of the long-run average. Because losses are concentrated, most windows fall below the average. For a ten-year window, the median window delivers 80% of the long-run average; 41% deliver less than half of it, 16% record nothing at all and 18% deliver more than double. As a result, an investment with an expected return of 2 to 1 recovers its cost in 59% of ten-year windows, 4 to 1 in 69%, and 10 to 1 in 75%. Over five years the figures are 43%, 51% and 58%; over twenty years 87%, 90% and 94%.

Deaths behave similarly. A 4 to 1 investment in lives recovers its cost in 72% of ten-year windows, and the lumps differ by income: in low-income and lower-middle-income countries the death record is smoother over a decade (87% for 4 to 1) than the damage record (62% and 67%), because deaths are recorded where damage is not, and large events there are more frequent; in high-income countries the damage record is smoother (74%) than deaths (55%).

The practical point is about appraisal, not about prevention's worth. A budget office that reviews a programme after three or five years will see a realised return below cost in about half of cases, even where the expected return is high, and may cancel what it should keep; one that sees a big event avoided will credit the programme with more than it earned. Expected returns, appraised ex ante on the full record, with multi-year commitments, risk pooling and contingent finance to carry the lumps, are the right basis. The 25-year record is short and the largest events dominate it, so the figures describe this record, not a law.

A government cannot choose where the next catastrophe will occur, but it can choose where to invest first. Ranking the 170 countries by mean death rate in 2000–2011 and testing the ranking on 2012–2024 gives a rank correlation of 0.51 for deaths and 0.52 for damage as a share of GDP: the record is informative but far from decisive. Aiming the reference investment at the worst 10% (17 countries, 1.1 billion people) would cost $0.5 billion a year, capture 31% of later deaths and save one life per $178,000. The worst 20% (34 countries, 3.1 billion people) captures 49% at $319,000 per life; the worst half 82% at $380,000; everyone 100% at $404,000.

Aiming works in this test, but only partly, because the events that dominate each period are in different places. The largest deaths of 2012–24 were not all in the countries with the largest deaths of 2000–11, and countries with no record in the first period have large events in the second. The implication is that targeting on history should be paired with a hazard-and-exposure model, not replace it, and that universal basic coverage retains a case: the cost per life of the uniform investment in 2012–24 is only about 2.3 times that of the aimed one, while the share of deaths missed by aiming at 10% is 69%.

Aim by damage rather than deaths and the picture differs: the countries with the highest damage ratios in 2000–11 yield higher asset returns (6.3 to 1 for the worst 10%, 18 to 1 for the worst 30%) but save fewer lives per dollar ($570,000 and $827,000 per life). Aiming for assets and for lives therefore selects different countries, which is the two-currency result of section 2 seen through targeting. A programme that wants both has to decide which it is buying, or to buy both separately.

Health prevention reports the highest ratios in dollars and the cheapest lives. Immunisation in 94 low- and middle-income countries returns about $26 per dollar on treatment costs and productivity in 2011–20 and $20 in 2021–30, and about $51–52 per dollar when the value people place on health is counted; a separate analysis of 73 Gavi-supported countries finds $54. The cost per life saved by the best global health charities has been estimated at $3,000–5,500 historically and about $8,000 at the margin in 2025. These are one to two orders of magnitude below the $189,000 of the uniform early-warning benchmark, which reflects how rare disaster deaths are compared with the burdens of disease, not a failure of early warning.

The disaster-specific evidence is thinner and more observational: the Global Commission on Adaptation's 30% damage reduction from 24 hours' notice, an association between early-warning coverage and mortality six times lower, and a before-and-after estimate that Ahmedabad's heat action plan avoided about 1,190 deaths a year. Cash transfers are associated with lower mortality (20% for adult women, 8% for children under five). None of these figures is a measured return on a defined budget in a defined place, and several are modelled or observational, which is why this report keeps its own estimates conditional on stated assumptions.

Reading across the table and the dashboard, three things stand out. Returns in lives are cheapest where preventable disease is common, not where disasters are; returns in dollars are largest where assets are most valuable; and returns to adaptation overall are well above one in published studies but with wide ranges. A prevention portfolio should therefore be built channel by channel, with the best-evidenced investments first, and with a clear view of which currency each is buying.

The calculator takes a country's own 25-year record of damage (% of GDP) and death rates, draws the chosen number of years at random with replacement 2,000 times, and for each draw computes the damage and deaths avoided by the chosen effect sizes, converted to dollars at the country's mean 2015–24 GDP and population. It reports the expected return, the probability that avoided damage covers cost (and with lives valued, that avoided damage and lives together do), the probability that nothing at all is avoided, and the distribution of realised returns.

Reading the output: a country with an expected return well above one but a probability of payback near half is a country where the investment is good and the next decade could still disappoint. A country with a very low asset return and a break-even life value above what its income can support is one where the case rests on humanitarian grounds, or on losses the record does not capture. The break-even value of a life is reported both in dollars and as a multiple of GDP per person, so that readers can judge it against their own valuation.

The calculator describes this record, not the future. It uses recorded disaster loss only, applies the same effect size to every year, ignores that effects may be larger for the biggest events or smaller where infrastructure is weak, and treats years as independent. It does not model the other channels of polycrisis, whose returns appear in section 5. It is a way to see how much would have to be true, and how lumpy the result would be.

Appraise on expected returns and fund multi-year. A programme reviewed after three to five years will show a return below its cost in roughly half of cases even if its expected return is four to one. Finance ministries should appraise prevention on the full loss record or a catastrophe model, commit funds for a decade or more, and review on process and coverage indicators between events, not on realised avoided damage.

Finance the lumps. Because the return arrives in a few large events, it can be carried by pooling and pre-arranged finance: regional risk pools, contingent credit, catastrophe bonds and anticipatory financing triggered by forecasts. The tiny size of the reference bill against the interest burden of the squeezed countries ($1.4 billion against $495 billion) shows that the issue is the shape of the budget, not its size.

Aim with a hazard model, and count both currencies. History is a moderate guide, so targeting should combine past loss with hazard and exposure models, keep a floor of universal coverage, and state whether the programme is buying assets or lives. In the poorest countries, where asset return is low and recorded damage understated, the case is in lives and in measured but unrecorded losses; the evidence base there is thinnest, and better loss accounting is an investment in itself. Beyond disasters, the cheapest lives and the highest ratios are in health prevention, which belongs in the same portfolio.

Losses: Our World in Data / EM-DAT series "Economic damages from natural disasters as a share of GDP" and "Death rates from natural disasters" (all disasters), country-years 2000–2024; country-years without a record are zero. Countries are those of the World Bank with at least 0.5 million people in 2023 and a World Bank income group (170: 54 high income, 49 upper-middle, 44 lower-middle, 23 low). Economy: WDI NY.GDP.MKTP.CD, NY.GDP.PCAP.CD (mean 2015–2024) and SP.POP.TOTL (2023). Recorded damage is applied as a share of today's GDP, so dollar figures are loss ratios applied to the current economy, and deaths are rates applied to today's population.

Reference investment: cost $0.50 per person per year (the UN Early Warnings for All plan: $3.1 billion over five years, about 50 cents per person per year covered), avoiding 30% of damage (Global Commission on Adaptation, for 24 hours' notice) and, as an assumption, 30% of deaths. Expected asset return = 0.3 × mean damage % ÷ 100 × GDP per person ÷ 0.50; group returns are totals; lives per $1 million = 10⁶ × 0.3 × mean death rate ÷ 10⁵ ÷ 0.50. Lumpiness: for each country with a recorded loss, rolling windows of 5, 10 and 20 years; the realised return relative to expected is the window's mean loss divided by the country's 25-year mean; an investment with expected ratio B pays back in a window if this is at least 1/B. Targeting: countries ranked by mean death rate 2000–11, outcomes summed over 2012–24, Spearman rank correlations. Calculator: 2,000 bootstrap draws of years with replacement (seeded) from the country's record.

Caveats: EM-DAT includes events by thresholds and leaves damages blank for many events, mostly in poor countries; the 25-year record is short and dominated by a few events; windows overlap, so the pooled probabilities are not independent; the 30% effect sizes are assumptions drawn from a published estimate and an association; the unit cost is an average; disasters are one channel of polycrisis; the calculator treats years as independent and ignores trends in hazard and exposure; and the literature table's figures use different methods and are not comparable.