Across six Indian cities, the neighbourhoods most likely to flood receive 12.2% less capital per standard deviation of flood hazard (p = 0.0001). Bengaluru — the one city publishing at ward level — shows why: its flood-prone wards do tilt spending toward drainage, they are simply handed smaller budgets. The misallocation sits one level above where climate-budget audits look.
Hover any unit to preview it · click to keep it pinned · click it again to release · the table and the map are one selection.
Effect on stormwater spending per standard deviation of flood hazard, within city. Four of six are negative; the two positives have error bars so wide (Mumbai ±80pp, Surat ±100pp) that −12% sits comfortably inside them. Formally there is no detectable disagreement between the six — they are consistent with one common effect, which is why the combined estimate above is the headline rather than any single city.
Robust to estimator: OLS −9.6%, PPML −9.6%, median regression −12.2%.
Observed flood density rises monotonically across modelled-hazard quartiles — 1.81× from lowest to highest (Spearman ρ = +0.26). The hazard surface finds floods that actually happened.
The same ₹23,081 crore is 1.3% or 45.9% drainage depending on which of three defensible keyword definitions you apply — a 35.2-fold range.
Reading 300 orders shows why: the medium tier is 99.1% precise with 95.5% recall, but 81% of the money it flags is bundled “roads and drains” work, where the full amount is charged to drainage.
Those 300 were labelled by reading each description, independently, twice — the passes agreed on 298 of 300. The first version of this check used a regex and reported exactly 100% precision. That was the tell: the adjudicator shared its keywords with the classifier it was grading. Re-labelled properly the two agree only 90% of the time.
Against 660 km of mapped drainage line-work. 87% of the effect survives. The p-value weakens only because the control is collinear with hazard by construction — the standard error widens 1.21× while the coefficient barely moves. Reading the p-value alone would mistake multicollinearity for the effect vanishing.
Per SD of annual rainfall, within ward. Naively this says a wet year cuts drainage spending. It does not — the order count falls too, and lighting and buildings fall harder than drainage. Rain stops construction. These are payments, not budgets, so only the share is interpretable, and on the share nothing responds.
Alignment with the ruling party buys a ward nothing, and the hazard penalty is unmoved by party controls. Whatever produces the budget gap, it is not partisan targeting.
Hazard is nonetheless politically distributed: opposition INC wards are more flood-prone than BJP wards, and the seven independents are the most exposed of all.
The categories with no flood-protective function — buildings, lighting, water supply — sit flat, which is what makes the hazard measure credible: it is not simply correlated with spending in general.
Zero of five contrasts are significant. Drainage moves with roads and parks and cannot be told apart from either. High-hazard wards tilt toward outdoor civil works generally — what low-lying, less-built-up land needs — not toward flood protection.
This sharpens the headline rather than denting it: flood hazard predicts a smaller total budget, and there is no drainage-specific compensation inside it.
The first line matters most: trunk drains cross ward boundaries and are built where water collects, so the money that cannot be assigned sits disproportionately in high-hazard wards. The true gap is probably larger than reported.
Why Bengaluru stops at FY2022. The city redrew its wards from 198 to 243, then split into five separate corporations in 2025 — so the unit of analysis stops existing. Tested anyway: post-2022 data gives −5.8% on the old map and +3.5% on the new one, neither significant. The other five cities run to FY2024–26, so the pooled panel spans FY2013–2026.
Why floods and not heat. The standard climate grid takes six distinct values across Bengaluru's 198 wards — a three-day spread across an entire city. And no municipal budget has a heat line: cooling is scattered across parks, housing, transport and water. Floods are the one hazard with both a measurable geography and a nameable budget line.
Effect on the drainage share per standard deviation of hazard, each arm re-estimated with ward-clustered and Conley spatial standard errors. Under the narrow definition — dedicated stormwater assets only — the effect is significantly positive. Dedicated drainage does track hazard; bundled road-and-drain money does not.
The headline budget penalty ranges −8% to −12.8% depending on how much terrain is absorbed into the controls, and is significant in every specification.
Significant at every threshold and strongest with no threshold — the opposite of a threshold artefact. Dropping population entirely gives −8.0%; using satellite built-up area instead gives −7.5%; using no controls at all gives −11.1%. The controls shrink the effect rather than create it.
The headline survives any correction. The +1.61pp share result does not, and is reported as suggestive rather than established — the main claim rests on the total-budget channel, which is unaffected.