|
Field
Last Published
|
Before
May 05, 2026 08:05 AM
|
After
August 28, 2026 11:03 AM
|
|
Field
Experimental Design (Public)
|
Before
A multi-round panel survey experiment with approximately 1,500 households in flood-prone districts of Bangladesh, surveyed four times between April and October 2026. Two respondents per household are interviewed with identical instruments. The study embeds several randomized experiments within the survey to identify sources of measurement error. The main randomizations vary survey mode (phone vs. in-person), respondent incentive levels, enumerator assignment, and the ordering of modules within the questionnaire. A digital monitoring component operates between survey rounds.
|
After
A multi-round panel survey experiment with approximately 1,500 households in flood-prone districts of Bangladesh, surveyed four times between April and October 2026. Two respondents per household are interviewed with identical instruments. The study embeds several randomized experiments within the survey to identify sources of measurement error. The main randomizations vary survey mode (phone vs. in-person), respondent incentive levels, enumerator assignment, and the ordering of modules within the questionnaire. A digital monitoring component operates between the final two rounds.
|
|
Field
Randomization Unit
|
Before
Multiple levels: - EA level: Incentive amounts, Resonance group assignment. - Household level (within EA): Survey mode (stratified by EA), recording visibility (household level for in-person interviews). - Respondent level (within household): Module ordering.
|
After
Multiple levels: - EA level: Incentive amounts. - Household level (within EA): Survey mode (stratified by EA), recording visibility (household level for in-person interviews). - Respondent level (within household): Module ordering, high-frequency monitoring via Resonance. - Interview-level: Enumerators to interviews (conditional on EA), audit announcement (conditional on enumerator selected for any audit announcement), planted errors in dwelling description. - Enumerator level: any audit announcements.
|
|
Field
Sample size (or number of clusters) by treatment arms
|
Before
Survey mode: ~750 households phone, ~750 households in-person (15 per arm per EA × 50 EAs).
Incentives: ~17 EAs low, ~17 EAs medium, ~16 EAs high (~510/510/480 households).
Module order: ~500 respondents per version (6 versions, respondent-level).
Recording visibility: Round 0: ~2,700 respondents with visible external microphone, ~300 respondents tablet-only;
High frequency monitoring system: ~20 EAs information + monitoring (~600 HH), ~20 EAs monitoring only (~600 HH), ~10 EAs pure control (~300 HH).
|
After
Survey mode: ~750 households phone, ~750 households in-person (15 per arm per EA × 50 EAs).
Incentives: ~17 EAs low, ~17 EAs medium, ~16 EAs high (~510/510/480 households).
Module order: ~500 respondents per version (6 versions, respondent-level).
Recording visibility: Round 0: ~2,700 respondents with visible external microphone, ~300 respondents tablet-only;
High frequency monitoring system: ca. 1000 respondents to receive invite, 500 willing respondents to receive no invite.
|
|
Field
Intervention (Hidden)
|
Before
We study the magnitude, sources, and remedies of measurement errors in survey data. The study covers approximately 1,500 households across 50 enumeration areas in four districts (Dhaka, Satkhira, Sirajganj, Sunamganj), interviewed in four rounds from April to October 2026. Two adult respondents per household are interviewed separately with identical instruments.
The embedded experiments include:
1. Survey mode: Households randomly assigned (stratified by EA) to phone or in-person interviews from Round 1 onward (Round 0 is always in-person).
2. Respondent incentives: Three EA-level arms (low/medium/high monetary incentive, linearly spaced). In Round 3, all households receive the high incentive, creating random variation in incentive changes.
3. Module order: Six questionnaire versions randomly assigned at the respondent level, rotating three module blocks (Mental Health + FIES, Housing + WASH, Income) across early/middle/late positions.
4. Recording visibility: Among in-person interviews, household-level randomization of visible external microphone vs. tablet-only recording.
5. Digital monitoring (Resonance): EA-level randomization into three groups: monitoring check-ins plus preparedness information, monitoring only, and pure control.
Additional experiments planned for Rounds 2–3 include enumerator audit announcements, planted errors in dwelling descriptions for enumerator diligence testing, enumerator re-randomization for rapport effects, individualized enumerator feedback, household roster omissions for attention testing, and SMS attrition reminders. These are described in the attached pre-analysis plan.
Data quality is assessed through benchmark comparisons (enumerator-observed vs. respondent-reported dwelling characteristics), embedded consistency checks (planted items in FIES and PHQ-8 modules, locus of control triad, paired climate adaptation questions), enumerator feedback, paradata indicators, audio recordings, and the RISSK machine-learning-based risk scoring toolkit.
|
After
We study the magnitude, sources, and remedies of measurement errors in survey data. The study covers approximately 1,500 households across 50 enumeration areas in four districts (Dhaka, Satkhira, Sirajganj, Sunamganj), interviewed in four rounds from April to October 2026. Two adult respondents per household are interviewed separately with identical instruments.
The embedded experiments include:
1. Survey mode: Households randomly assigned (stratified by EA) to phone or in-person interviews from Round 1 onward (Round 0 is always in-person).
2. Respondent incentives: Three EA-level arms (low/medium/high monetary incentive, linearly spaced). In Round 3, all households receive the high incentive, creating random variation in incentive changes.
3. Module order: Six questionnaire versions randomly assigned at the respondent level, rotating three module blocks (Mental Health + FIES, Housing + WASH, Income) across early/middle/late positions.
4. Recording visibility: Among in-person interviews, household-level randomization of visible external microphone vs. tablet-only recording.
5. Digital monitoring (Resonance): Respondent-level randomization into two groups: weekly monitoring check-ins between R2 and R3 and pure control.
6. Audit announcements: Randomized first between enumerators (none vs. any) and then within treated enumerators (50% of interviews randomly selected to be flagged to enumerator as being audio audited).
7. Planted errors: We randomly plant errors in dwelling descriptions and test whether enumerators catch them.
8. Flipping a question's framing: We randomize whether a question with obvious and unambiguous truth is framed as (i) whether the respondent has been interviewed previously; (ii) whether the respondent is interviewed for the first time.
Additional experiments planned for Round 3 include individualized enumerator feedback and SMS attrition reminders. All experiments are described in detail in the attached pre-analysis plan.
Data quality is assessed through benchmark comparisons (enumerator-observed vs. respondent-reported dwelling characteristics), embedded consistency checks (planted items in FIES and PHQ-8 modules, locus of control triad, paired climate adaptation questions), enumerator feedback, paradata indicators, audio recordings, and the RISSK machine-learning-based risk scoring toolkit.
|