Norms Behind Closed Doors:
Misperceptions and Maternal Employment in Couples

What's missing: Individual choices don't reveal intra-household dynamics. Does the gender gap reflect preferences, or the fact that women anticipate they will absorb the household's need for flexibility?

Critical finding: Only "support for mothers working outside the home" moderates the gender gap in WTP. Gender norms — not just preferences — shape how couples allocate labor market investments. But we can't change those norms with a DCE.

What is missing in the literature: All prior studies observe only one partner — cannot measure mutual misperceptions or study how information corrects within-couple belief gaps. Course choices and lab outcomes are often hypothetical. None use a probability-sampled population as both respondents and norm-source.

Pluralistic Ignorance Pattern: Both spouses privately support maternal employment, but dramatically underestimate community support for fathers (−27.5 to −32.8 pp). Misperception about mothers' support is much smaller (−10.5 pp). Within-couple, spouses are optimistic (slightly overestimate each other's support). The community-level friction dominates, not spousal. Placebo shows this is gender-norm-specific.

Step 2 — IPWRA within each sample S (doubly robust ATT):

ATT̂ = (1/N_T^w) Σ_i w_i^S [ D_i(Y_i − m̂₀) + (1−D_i)·(ê_i/(1−ê_i))·(m̂₁ − m̂₀) ]

e(·) = treatment model Pr(D=1│X^D, strata); m_d(·) = outcome model E[Y│D=d, X^D, strata]. Step 1 weights enter as pweights. Consistent if either e(·) or m_d(·) is correctly specified.

Mechanism: Respondents are never told their partner's actual baseline beliefs. Any updating of spousal perceptions must arise from generalized introspection — corrected community norms lead individuals to reassess expectations about close others.

Sample 2∩3 (N=1,102): treated at midline, surveyed at endline (1–2 months later). Outcomes: (1) job mobility — changed jobs or started a business; (2) aspires to improve LM situation; (3) wants to balance work and family.

Pattern: treated women are more likely to have started a job, report better work–family balance, and higher career aspirations — consistent with a relaxed social-permission constraint among those already attached to the labor market.

Theory: generalized introspection (Boltz 2025) — credible information about the community realigns perceptions of all close others, including the partner, because the partner is part of "the community." The spousal effect is a downstream consequence of correcting community beliefs, not an independently informed margin.

Why this is "mechanism evidence" (not just heterogeneity): the pattern discriminates between hypotheses. If treatment moved preferences, effects should be uniform. The fact that only labor-attached households respond is consistent with information lifting a social-permission constraint that binds only when structural barriers (childcare, demand) are already low. Information complements — not substitutes for — structural policy.

⚠ Caveat — endogeneity: we did not randomize which spouse engaged with the WhatsApp module. Direct / indirect / joint exposure configurations are endogenous. We therefore interpret these results as descriptive, not causal.

Policy implication: Norm-correcting information and structural interventions (childcare, workplace flexibility) are complements, not substitutes. Each targets a different margin of the same problem.

Why this matters: the poll transforms a single-outcome RCT into a mechanism-identification design. It is also immediately replicable by any practitioner on any platform — a standalone policy tool for mapping community-norm misperceptions in real time.

Across all 8 gender-norm items, the same pattern holds: progressive private attitudes coexist with sizable misperceptions about others' views, particularly men's support. The placebo shows near-universal agreement and no misperceptions — confirming misperceptions are norm-specific, not general pessimism.

What we do: Romano-Wolf adjusts each p-value by simulating the joint distribution of all test statistics under the null (1,000 bootstrap replications, clustered by household). It produces a family-wise error rate–controlled p-value for every outcome, accounting for the dependence structure between them.

Plain English: "Even if I throw lots of outcomes at this experiment, here is the p-value adjusted for the fact that I am fishing in many ponds. Results that survive RW are not lucky strikes."

Notes: Romano-Wolf computed on OLS (unweighted) — conservative vs. IPWRA. 1,000 replications, seed(12345), clustered by household. Women's job mobility survives stepdown correction (RW p=0.013). Course nomination for men is marginal (RW p=0.077).

What we do: Trim observations from the lower-attrition arm to make response rates equal across treatment and control. Then compute the worst-case and best-case ATT — the interval brackets all possible values consistent with monotonicity (treatment doesn't change who attrits).

Plain English: "Imagine the absolute worst possible scenario about who dropped out. Even then, my treatment effect lies somewhere in this range. If both ends of the range exclude zero, my result holds even under unobservable bias."

Interpretation: Lee bounds apply when treatment monotonically increases probability of being in sample. For the course (Sample 2), this is satisfied by design (engagers). Both bounds strictly positive for men's course nomination and women's job search — key results hold under worst-case attrition scenarios consistent with monotonicity.

What we do: re-estimate the IPWRA ATT under 4 alternative weight constructions to check that headline results don't depend on the most extreme weights:

• (i) Baseline: probit propensity score (preferred)
• (ii) Winsorised: cap weights at 95th percentile
• (iii) Trimmed: drop observations with PS < 0.10 (≈ 1% of obs)
• (iv) Logit PS: alternative functional form for the selection model

Plain English: "If a handful of unusual observations were driving my result, the estimate would change a lot when I cap or drop them. It doesn't change → my result is robust, not artifact of extreme weights."

After IPWRA weighting, maximum absolute standardized mean differences (SMDs) are below 0.10 in all samples and genders. Some covariates show marginal imbalance in Sample 2∩3 (joint F tests reject), but effect sizes are small, and post-weighting balance is tight. The key variable — second-order belief about men's community support — does not differ significantly across treatment and control arms in any sample.

Propensity scores range from 0.06 to 0.76; overlap is adequate in all samples. Effective sample sizes remain large after weighting; mass outside common support is small. P-score densities from 0.06–0.76 → no extreme regions of non-overlap that would invalidate IPWRA.

IPWRA is the preferred specification chosen a priori to address selection into midline take-up. OLS and weighted-OLS produce nearly identical point estimates across all headline results — the choice of estimator does not drive the findings.

Job mobility effects are concentrated among employed (+14.8 pp) and unemployed (+8.9 pp) women. Inactive women show near-zero effects. The course nomination effect is also largest when the wife is employed (+13.4 pp, p<0.10). Together, these results suggest information works at the margin where action is already feasible.

Engager characterization (selective take-up): only 36% engage with the WhatsApp module. We compare engagers vs. non-engagers on (a) demographics and (b) the targeted second-order belief.

Plain English: "Engagers are more inactive (selection on demographics — fix with IPW step 2). But they hold the same prior on community support as non-engagers — so the disclosed norm is accurate for them, the people who actually received it."

Reference-group accuracy: max deviation of any subgroup's mean SOB from city-wide average is 3.3 pp. The misperception we are correcting is 28 pp → reference-group mismatch is < 12% of the corrected signal. Disclosed Bogotá-average norm is a valid proxy for every demographic subgroup.

Reference group accuracy: Max deviation of any subgroup's mean SOB from city-wide average = 3.3 pp (high-SES). The corrected misperception is ~28 pp → reference group mismatch is <12% of the corrected signal. Disclosed norm is valid for all demographic subgroups in the sample.

The mediation exercise supplements rather than replaces the reduced-form evidence. We treat it as a consistency check on the sign pattern and direction of the channels.

Replication data and code available at doi.org/10.7910/DVN/QYWHLA.