Information or Incentives? A Field Experiment on Local Disaster Prevention

The intervention gives municipal governments credible, district-specific information about their El Niño flood risk before the 2026-27 rainy season. We built that information ourselves, from public data. A district damage model combines the national emergency registry for 2012-2025, gridded rainfall and the 2017 Census. For each district it predicts the people affected and the houses destroyed if the coming season repeated the rains of 2017 or of 2023. Out of the 1,891 municipalities in force in 2026, the model selects the 1,248 where it predicts at least one house destroyed and ten people affected.

Each of these municipalities receives a letter from a researcher of the Pontificia Universidad Católica del Perú, on university letterhead. The letter goes to the mayor by e-mail and through the municipality's online document-reception system. It also goes, in copy, to the municipal manager and the civil-defence office. The letter frames its request under the constitutional right of petition, which obliges the municipality to reply in writing. Every letter reports the official El Niño outlook and the national damage of 2017 and 2023. It asks one verification question with tick boxes, whether people or infrastructure sit in the district's exposed areas, and an optional question on planned prevention works.

The letters differ across three groups of 416 municipalities, which we assigned at random within strata of department, natural region and official hazard rating. The control letter carries only the national information. The first treatment letter adds a block on the district with its predicted damage under the two scenarios. The block also reports the district's recorded damage in 2017, 2023 and 2025 and its official hazard rating. The second treatment letter carries the same block and states that the same information went, on the same date, to the national civil-defence agency. A separate letter to that agency, sent the same day, lists these municipalities with their figures and makes the statement true.

The letters go out in three waves within one season. Wave 1 reaches the outgoing mayor on 20 October 2026, and wave 2 the incoming mayor on 5 January 2027. Wave 3, on 10 March 2027, compares the predicted damage with the damage observed in January and February. The assignment stays the same in the three waves. After the season we share the study's results with every municipality and publish the predictions for all districts.

2026-10-20

2027-03-31

Both end points measure preventive spending from the resources under municipal control, which are own revenue and formula transfers. The spending is the municipality's accrued execution on the prevention and preparedness phases of budget program 0068, from October 2026 to June 2027. The first end point is an indicator equal to one if that execution is positive. The second is the same execution per inhabitant in soles, winsorized at the 99th percentile of the control group.

We read the outcome from the monthly budget execution of program 0068 by municipality, which the Ministry of Economy and Finance publishes as open data. Accrued execution records the municipality's obligation to pay once it has received the good, service or work. Each record carries its funding source, and we group the sources into four classes. Own revenue covers municipal taxes and directly collected revenue, and formula transfers cover the municipal compensation fund and the canon. Central transfers are the ordinary resources of the treasury, and the fourth class holds donations, transfers and credit. The money that arrives after a disaster, through emergency decrees and the disaster fund, comes as central transfers, so the primary outcome leaves them out.

Each record also carries a product of the program, and we fixed the rules that assign it a phase before the draw. Products 3000734 and 3000739 are preparedness, products 3000735 and 3000736 are prevention, and product 3000516 is response. Product 3000001, common actions, never counts as prevention or preparedness. We classify investment projects by name with patterns applied in a fixed order: recovery, response, preparedness, and prevention for every remaining project. A season whose accruals net to a reversal counts as zero execution. The denominator is the 2025 population of the district. The 150 provincial municipalities execute province-wide but enter with the population of their capital district, so we also report the estimates without them.

The indicator captures the extensive margin, where the model predicts that new information moves the municipalities whose belief crosses the cost of the first sol. We winsorize the amount because a few municipalities execute hundreds of soles per inhabitant in a season while the median executes about one. As robustness, we estimate a Poisson pseudo-maximum-likelihood regression on the unwinsorized level. We extract the data after each quarter closes and archive every extraction with its date and hash. The primary reading is the July 2027 extraction, and later readings of the same months are robustness. The outcome exists for all 1,248 municipalities, whether or not they reply to the letter.

H1 states that district information raises preventive spending, and we test it with T1 against control. H2 states that national observability changes that effect, and we test it two-sided with T2 against T1. Each hypothesis carries one randomization-inference p-value for its two forms, based on the larger absolute t-statistic. We adjust the two p-values by Holm's method. H1 holds if the estimate is positive in the form that attains the maximum and its adjusted p-value is below 0.05. H2 holds, in either direction, if its adjusted p-value is below 0.05.

We group the secondary outcomes into five families. Administrative response holds delivery of the letter and a reply within 30 business days and within six months. It also holds the content of the reply, meaning the box ticked and any planned works named. The last outcome of the family is the change in the modified budget of program 0068 per inhabitant from its opening value. National instruments hold the index of hypothesis H3 and its four components. The first component is the execution of central transfers in the response and recovery phases from October 2026 to June 2027. From October 2026 to December 2027, the other three count emergency declarations covering the district, disaster-fund allocations in it and national transfers to it. Damage holds people affected, houses destroyed and the number of records the district files in the national emergency registry during the season. Persistence holds the primary outcome from July to December 2027 and over fiscal year 2027. The last family repeats the primary outcome counting own revenue alone.

Replies arrive by e-mail and through each municipality's online document-reception system. We code each reply for its date, its sender and its attachments, the box it ticks and whether it names planned works or cites investment codes. A municipality that has not replied six months after dispatch counts as no reply, which is a value of the outcome and not attrition. We read damage from the national emergency registry, keep the excess-water hazards and remove duplicate records with our documented rule. Emergency declarations come from the official gazette, and national transfers to municipalities come from the budget data. The investment bank's registry locates disaster-fund allocations by district, whatever level of government executes them.

Secondary outcomes use the same two equations as the primary ones, and none of them is confirmatory. Within each of the five families they carry sharpened false-discovery-rate q-values. H3 combines the four national instruments into one standardized index, and we report its components. The dilemma pattern requires a positive index effect that survives the adjustment at 5 percent. It also requires an H1 estimate whose 90 percent confidence interval lies below half the minimum detectable effect of H1. The belief-gap pattern is the converse, and we report any other combination as mixed.

H4 interacts treatment with an indicator that the district received reconstruction or disaster-fund spending in 2017-2025. Rescue went to the districts with damage, and damage also sharpens the prior. The regression therefore also interacts treatment with people affected per inhabitant in 2017 and 2023. The robustness form replaces the indicator with rescue per unit of damage. H5 interacts treatment with two indicators, each with a predicted sign and with its lower-order terms. The first marks no people affected in 2017-2025 together with a prediction above the median of the eligible. The second marks an incoming mayor who has not held municipal office before, which the election of 4 October 2026 already fixes. H4 and H5 carry no multiplicity adjustment, and H4 can detect only a large attenuation.

We also report, without adjustment, heterogeneous effects of H1 and H2 by natural region, official hazard rating and municipality type. We add terciles of predicted people affected per 1,000 inhabitants, of past preventive execution and, after the season, of realized rainfall. Spillovers use the share of treated municipalities among the other eligible ones in the same river basin and in the same province. We assign each district to one of the 231 basins of the national water authority by its centroid. Triplets make that share depend on a municipality's own arm, so we recentre it on its expectation over the re-draws of the assignment. The outcome is prevention for the province route and damage for the river route, and the analysis is exploratory.

The trial is a randomized experiment with three arms, and the municipality is the unit of assignment and of analysis. The universe is the 1,891 municipalities in force in 2026, 1,695 district and 196 provincial. A district damage model, estimated on a 2012-2025 district-year panel, predicts the damage of a repeat of the 2017 and of the 2023 seasons. A municipality is eligible if the model predicts at least one house destroyed and ten people affected under either analog, which leaves 1,248. These municipalities account for 98 percent of the people affected nationally in 2017 and in 2023. Seventeen districts created after the 2017 Census inherit the prediction of their district of origin. The model defines the frame and supplies the figures of the letters, and it plays no part in the assignment.

We assign the 1,248 municipalities to control, T1 and T2 in equal numbers, 416 each, within 63 strata. The strata cross department, natural region and official hazard rating, and a rule fixed in advance merges cells with fewer than three municipalities. Within each stratum we sort municipalities by type, hazard rating and population. We then give each consecutive triplet an independent random permutation of the three arms. Leftover units go to the arms with the fewest leftovers across strata, and a random draw breaks ties. We drew the assignment once, on 6 October 2026, with a public seed, and we do not re-randomize. We deposit the frozen list, its hash and the script that draws the assignment with the plan.

H1 compares T1 with control and measures the effect of district information. H2 compares T2 with T1 and measures the effect of telling the mayor that the national agency holds the same figures. The pooled contrast compares any district letter with control. The letters go out in three waves within the 2026-27 rainy season, in October, January and March, and the assignment stays the same across waves. Every outcome comes from administrative records that the state publishes, so the study needs no survey and loses no municipality to attrition. The primary outcome covers October 2026 to June 2027, and we read it in the July 2027 extraction of the budget data.

The primary estimates regress each outcome on the arm indicators with stratum fixed effects. We use heteroskedasticity-robust standard errors, which are conservative under matched triplets. Randomization inference re-draws the whole assignment 10,000 times for the pooled contrast. For each pairwise contrast it holds the third arm fixed and swaps the two labels within triplets. We also estimate three robustness specifications, each one separate from the others. The first adds the pre-treatment mean of the outcome, and the second adds the balance covariates interacted with treatment. The third stacks the seasons 2022-23 to 2026-27 with municipality and stratum-by-season fixed effects.

Not available

We drew the assignment once, by computer in R 4.5.3, on 6 October 2026, before sending any letter. The input was the frozen list of 1,248 eligible municipalities, the file prediction_damage_fen.csv. Anyone can check the file against its SHA-256 hash, 62e0fcb4ea704b92341de1d0ca1b3b3dad56e89f8a9b4536e139d2a524a7000c. We fixed the rule for the seed in advance, so nobody chose the number. The rule takes the banking-system selling exchange rate that the Central Reserve Bank of Peru publishes in its series PD04640PD. It uses the rate of the last business day before the draw and multiplies it by one thousand. The rate for 5 October 2026 was S/ 3.435 per dollar, so the seed was 3435. The script fetched the number from the Bank's statistics API and wrote the series, date and value into every row of the assignment file. With the rate from the Bank's site and the deposited script, any reader can repeat the draw.

The procedure stratifies by department, natural region and official hazard rating and assigns the arms within ordered triplets, with a national rule for leftover units. We drew once and did not re-randomize, and the script stops if an assignment file already exists. Two of nineteen covariates differ between treated and control municipalities at the 5 percent level: the vulnerability index and the indicator of a high risk scenario. The joint test of all covariates on the treatment indicator also rejects balance (p = 0.004). As the plan specifies, we keep the draw and control for both covariates in the robustness specification. The script, the log, the session information, the assignment file and the balance table go to this registration as private documents until the trial ends.

We randomize at the level of the municipality, which is also the unit of analysis. We identify each municipality by the six-digit code of its district. A provincial municipality enters as one unit, with the code of its capital district. The 17 districts created after the 2017 Census enter as units in their own right.

No

1,248 municipalities. The treatment is not clustered, so each municipality is its own cluster.

1,248 municipalities

416 municipalities in the control arm, which receives national information. 416 municipalities in T1, which adds district information. 416 municipalities in T2, which adds that the national civil-defence agency received the same district information.

We computed minimum detectable effects by simulation on the 2023-24 season, from October to June, with the primary definition of the outcome. We drew 300 placebo assignments with the procedure of the actual draw and estimated the primary equation on each, with stratum fixed effects and no baseline. With 80 percent power and a two-sided test at 5 percent, the minimum detectable effect equals 2.8 times the standard deviation of the placebo estimates. We convert soles at S/ 3.44 per dollar, the rate that gave the seed. For H1, T1 against control, the minimum detectable effect is 8.4 percentage points on the share of municipalities that execute anything, from a base of 59.5 percent. On the winsorized amount it is S/ 4.3 (about USD 1.25) per inhabitant, on a mean of S/ 8.2 (USD 2.40), and both figures equal 0.17 standard deviations. For H2, T2 against T1, the figures are 8.7 percentage points and S/ 4.4, and for the pooled contrast they are 7.1 points and S/ 3.6. The baseline-adjusted robustness specification lowers the minimum detectable effect of H1 to 7.2 percentage points and S/ 3.6. On own revenue alone, whose base is 21.9 percent, H1 detects 6.6 percentage points, or 0.16 standard deviations. The interaction of H4 with past rescue detects an attenuation of about 25 points, three times the main effect. The Poisson robustness form detects only an increase of 1.56 log points, which multiplies the mean by 4.7. A few municipalities with very high execution dominate its variance. All figures cover one season, since the three waves fall within the same rainy season.