Gender Coordination and the Sustainability of Housing Investments: Experimental Evidence from Rural Peru

Households are randomly assigned to one of three arms. The control group receives a placebo video unrelated to housing maintenance. The Sphere arm receives a video that assigns four maintenance tasks, sealing wall cracks, cleaning roof gutters, keeping upper windows unobstructed, and lubricating hinges and locks, to familiar household roles, promoting role clarity and individual internalization of responsibility. The Coordination arm receives a video with identical practical content but a different relational narrative: it frames under-maintenance as a coordination failure, corrects pessimistic beliefs about the partner's willingness to contribute, and closes with a joint implementation intention. The video is shown on a tablet during the household visit immediately after a pre-video survey instrument, and assignment is auto-loaded by the survey software. The experiment is implemented in two waves. The first wave (2024 cohort, approximately 870 households) includes Control and Sphere only. The second wave (2025 cohort, approximately 870 households) includes all Sphere and Coordination only. The primary contrast, Coordination versus Sphere, is estimated from the second wave.

2026-08-08

2026-11-30

The study has three sets of primary outcomes, all measured at the monitoring visit (midline) and endline.
1) Maintenance condition (direct observation).
2) Task completion and intra-household task allocation.
3) Decision-making and gender norms.

The study has three sets of primary outcomes, all measured at the monitoring visit (midline) and endline.

Maintenance condition (direct observation). A standardized index of housing module condition based on direct enumerator observation of five structural elements: floor, walls, roof, doors, and windows. Each element is scored on a condition scale and averaged with equal weights, standardized using the control-arm mean and standard deviation.

Task completion and intra-household task allocation. For each of five maintenance task types: cleaning and order (sweeping floors, cleaning walls and gutters), wall repair (sealing cracks), roof maintenance (checking and tightening screws and calaminas), door maintenance (varnish, hinges, frames), and window maintenance (varnish, hinges, glass), the survey records both whether the task was performed and who performed it: the respondent alone, the spouse alone, both together, another family member, a technician, or nobody. The share of tasks completed jointly (both partners) and the share completed at all are primary behavioral outcomes. Complementary summary measures ask who mainly performs structural maintenance tasks and who mainly performs cleaning and order tasks.

Decision-making and gender norms. Who decides when to do maintenance and how much to spend on it. Agreement with statements that men are more suited for repairs, that women are more responsible for cleaning, and that whoever knows should do it regardless of sex. Community-level perceptions of whether maintenance is a male or female task.

The primary confirmatory contrast, Coordination versus Sphere, is expected to increase joint task performance and shift decision-making toward shared responsibility, relative to role-based assignment. Sphere relative to Control tests whether any gender framing improves on the status quo.

1) Frequency of maintenance in the past 12 months
2) Frequency of couple conversation about the module's condition in the past 6 months
3) Whether there were disagreements in the household about whether to do maintenance
4) Partner expectation that the respondent participates in maintenance
5) Self-reported belief that preventive maintenance prolongs module life
6) Belief that a small crack not repaired quickly becomes a major problem
7) Descriptive norm: estimated share of neighbors who perform preventive maintenance
8) Thermal comfort inside the module during cold months
9) Life satisfaction
10) Information diffusion
11) Post-video comprehension
12) Immediate behavioral intention
13) Incentivized choice
Exploratory health outcomes (framed as exploratory given sample size and causal chain length):

Number of respiratory infections (IRAS) in the past 12 months
Number of diarrheal illness episodes (EDAS) in the past 12 months
Skin infections in the past 12 months
Healthcare expenditure on respiratory and gastrointestinal illness
School absenteeism due to respiratory or diarrheal illness (children aged 4–17)

Frequency of maintenance in the past 12 months (ordered categorical: monthly, every 2–3 months, every 6 months, once a year, less than once a year)
Frequency of couple conversation about the module's condition in the past 6 months
Whether there were disagreements in the household about whether to do maintenance
Partner expectation that the respondent participates in maintenance (0–10 scale)
Self-reported belief that preventive maintenance prolongs module life (0–10 scale)
Belief that a small crack not repaired quickly becomes a major problem (0–10 scale)
Descriptive norm: estimated share of neighbors who perform preventive maintenance (0–100)
Thermal comfort inside the module during cold months
Life satisfaction (0–10 scale)
Information diffusion: whether the respondent discussed maintenance with neighbors or recommended maintenance to others outside the household
Post-video comprehension: correct recall of gutter-cleaning frequency and crack-sealing materials (treatment arms only)
Immediate behavioral intention: number of maintenance tasks planned in the next 7 days (treatment arms only)
Incentivized choice: selection of maintenance supply card over lottery (treatment arms only)
Exploratory Health Outcomes: Health outcomes are classified as exploratory because the study is not powered to detect health effects and the causal chain from video framing to health involves multiple intermediate steps: the video must change maintenance behavior, improved maintenance must translate into better thermal performance of the module, and better thermal conditions must reduce illness incidence. Each link attenuates the intent-to-treat effect on health, making detection unlikely at the available sample size. Results will be reported descriptively alongside the pre-specified behavioral and attitudinal outcomes.

Randomization is at the household level. The trial is not cluster-randomized. Assignment is stratified by village, with households allocated across arms as evenly as village size permits. Standard errors are clustered by village at the analysis stage.

The 2024 cohort assignment was selected from 1,000 candidate random draws evaluated on the joint balance of six pre-specified covariates: sex of beneficiary, age, household size, a maintenance index, module material type, and ecoregion, using an omnibus F-test with village fixed effects. The draw maximizing the F-test p-value was selected. The same procedure will be applied to the 2025 cohort. Randomization files, seeds, and checksums are archived before field implementation.

The experiment is implemented in two waves. The first wave (2024 cohort, 870 households) uses a two-arm design with Control and Sphere. The second wave (2025 cohort, approximately 870 households) assigns all households to Sphere or Coordination with no control group, and is the confirmatory wave for the primary Coordination-versus-Sphere contrast. Both waves use identical randomization code and survey instruments.

Assignment is loaded into the CAPI survey software before fieldwork begins and is revealed to the enumerator only after the pre-video instrument is complete. Enumerators cannot select or modify the assigned arm.

Not available

Computer-generated randomization using a documented seed and rerandomization algorithm, implemented in Stata before field implementation. Randomization is stratified by village (núcleo ejecutor): within each village, floor(N/2) households are randomly assigned to Control and the remainder to Sphere. The retained assignment is selected from 1,000 candidate draws to maximize the p-value of an omnibus F-test of covariate balance, estimated with department-by-maintenance-tertile fixed effects.

Household. Within each village, households are individually randomized to treatment arms.

No

NA

Approximately 1,700 households across two intervention cohorts (870 households in the 2024 cohort and approximately 870 households in the 2025 cohort).

2024 cohort: 435 households Control, 435 households Sphere.
2025 cohort (planned): 435 households Sphere, 435 households Coordination, 0 households Control.

The following MDEs assume 80 percent power, a two-sided 5 percent significance level, individual-level randomization stratified by village, and OLS estimation with village fixed effects. Standard errors are clustered by village. Primary contrast (Coordination vs. Sphere, Cohort 2 only, 435 vs. 435): 0.161 standard deviations on the standardized maintenance condition index. Mechanism outcomes (Coordination vs. Sphere, Cohort 2 only): 0.176 standard deviations on the shared maintenance responsibility index and the joint decision-making index. Sphere vs. Control (pooled Cohorts 1 and 2, 870 Sphere vs. 435 Control): 0.140 standard deviations on the maintenance index. Binary outcomes: approximately 7–10 percentage points, depending on baseline prevalence. Note on design: The Coordination arm exists only in Cohort 2. The Coordination-versus-Sphere comparison therefore uses Cohort 2 households only (435 per arm under the preferred allocation). The original design target of 0.127 SD is not achievable under the two-wave structure because it would require approximately 700 households per arm in Cohort 2 alone. The pre-specified allocation for Cohort 2 is 0 Control, 435 Sphere, and 435 Coordination, which maximizes power for the primary mechanism contrast. The treatment-versus-control comparison is identified from Cohort 1 (435 Control vs. 435 Sphere) and the pooled Sphere sample.