Replicating "Male Backlash and Female Guilt" — Women's Employment and IPV in Urban India

For the final assignment in my Econometrics A course, my co-author Saurabh Roy and I replicated the core empirical results of Dhanaraj and Mahambare (2022), “Male Backlash and Female Guilt: Women’s Employment and Intimate Partner Violence in Urban India,” published in Feminist Economics. This post documents what we did, what we found, and the methodological choices that made this replication both instructive and occasionally tricky.

The Paper and Why It Matters

The relationship between women’s economic empowerment and intimate partner violence (IPV) has no clean theoretical prediction. Greater economic independence should, in principle, strengthen a woman’s bargaining power and reduce her tolerance for violence. But in contexts where traditional gender norms run deep, a woman entering paid work can also trigger a compensatory response from her husband — the male backlash channel — as he reasserts household authority.

What makes Dhanaraj and Mahambare (2022) distinctive is the introduction of a third channel: female guilt. Women who internalise patriarchal norms may, paradoxically, justify IPV against themselves upon entering paid work, raising their own measured exposure. Distinguishing among three theoretically plausible mechanisms — female autonomy, male backlash, and female guilt — using observational survey data is a genuinely hard problem. Their strategy for doing so is what drew us to this paper.

Data and Sample

We used unit-level microdata from the Domestic Violence module of the National Family Health Survey 2015–16 (NFHS-4), accessed through the DHS Programme repository. The DV module was administered to a 15 percent subsample of total households, and after restricting to currently married women (dropping divorced, widowed, and separated respondents for whom spousal controls are unavailable), we arrived at an analysis sample of 18,347 currently married urban women aged 15–49.

The outcome variable, any_violence, equals 1 if the woman reported experiencing any physical, sexual, or emotional IPV by her husband in the past twelve months. The treatment variable, paid_work, equals 1 if she engaged in paid work in the same period. In our sample, 23.8 percent of women reported being in paid work.

The Identification Problem

The central challenge is that paid_work is endogenous. Women who face higher IPV may exit the labour force; conversely, women with certain unobserved traits (say, higher agency) may both seek employment and resist violence more effectively. Either story contaminates a naive probit estimate.

Following the original paper, we instrument paid_work with district-level female labour force participation (FLFP) rates, constructed from the larger NFHS-4 state module sample (~37,415 urban women). The instrument captures local labour market conditions and prevailing cultural attitudes toward female employment — both of which plausibly shift a woman’s paid work decision — without directly influencing her individual IPV exposure once we condition on district-level controls (mean schooling, median wealth, past IPV rates) and state fixed effects.

The estimation framework is a bivariate probit model with endogenous treatment, estimated jointly:

\(\Pr(\text{IPV}_{i} = 1) = \Phi(\alpha \cdot \text{PaidWork}_{i} + X_i'\beta + \delta_s)\) \(\Pr(\text{PaidWork}_{i} = 1) = \Phi(\gamma \cdot \text{FLFP}_{d} + X_i'\pi + \delta_s)\)

where $\Phi$ is the standard normal CDF, $X_i$ is the full covariate vector, $\delta_s$ are state fixed effects, and $d$ indexes districts.

Replicating the Baseline Results

Our first task was to reproduce Table 4 of the original paper. The probit estimate of paid_work on any_violence is 0.025 (SE: 0.007). This looks small, but it understates the true causal effect because it absorbs the endogeneity. Once we switch to the bivariate IV probit, the average marginal effect rises to 0.083 (SE: 0.029) — more than three times larger.

* Probit baseline
probit any_violence paid_work $controls i.state, vce(cluster district)
margins, dydx(paid_work)

* Bivariate IV probit
biprobit (any_violence = paid_work $controls i.state) ///
         (paid_work = district_flfp $controls i.state), vce(cluster district)
margins, dydx(paid_work)

The Wald test of $\rho = 0$ rejects at the 5 percent level ($\chi^2(1) = 4.03$), confirming that the two equations’ error terms are correlated — i.e., the endogeneity concern is statistically warranted. The instrument is very strong: the first-stage coefficient on district_flfp is 3.328 (SE: 0.126).

Decomposing the Channels (Table 5)

The more interesting exercise is Table 5, which introduces interaction terms to isolate the three mechanisms. The original paper constructs three gender-attitude indices:

• Female autonomy: decision-making authority and mobility measures
• Male backlash: husband’s controlling behaviours
• Female guilt: wife-beating justification attitudes

Column (1) enters the main effects of all three indices. Column (2) adds their interactions with paid_work. All reported estimates are average marginal effects from the bivariate probit.

The headline result from Column (2) is:

The female autonomy interaction is essentially zero and insignificant. The male backlash interaction is positive and marginally significant, consistent with husbands responding to their wives’ employment by intensifying control. The female guilt interaction is positive but falls short of conventional significance thresholds. Dhanaraj and Mahambare (2022) interpret these results as partial evidence for the backlash channel and suggestive evidence for guilt, while the pure bargaining/autonomy channel finds no support.

Heterogeneity

We also replicated the three subsample analyses. A few patterns stood out to us:

By education: The paid work effect is largest and marginally significant among women with secondary education (AME = 0.089, SE: 0.047), but insignificant for both lower-educated and higher-educated women. This non-monotonicity is interesting — it may reflect that secondary-educated women are visible enough in the labour market to trigger a backlash response, without yet having the bargaining leverage that higher education confers.

By wealth: The paid work effect is concentrated entirely in the wealthiest quartile (AME = 0.194, SE: 0.075), and is essentially zero for the other three. This is perhaps the most striking heterogeneity result in the paper.

By region: The effect is statistically significant only in the North/West subsample (AME = 0.086, SE: 0.041). In South/East India, the point estimate is close to zero. The female guilt interaction, interestingly, is significant only in South/East (AME = 0.042, SE: 0.025), which the authors attribute to potentially stronger internalisation of patriarchal norms in that regional context.

What We Learned from the Replication

A few practical lessons from this exercise:

On the instrument construction. Building district-level FLFP rates requires merging two NFHS-4 files — the DV module (smaller, 15% subsample) and the women’s module (full sample). Getting the merge keys right and ensuring the district identifiers align between the two files took more time than we expected.

On the bivariate probit. biprobit in Stata does not natively report average marginal effects for the endogenous treatment variable in the way margins handles standard probit. We had to be deliberate about which equation we were calling margins on, and the syntax for recovering treatment-effect marginal effects differs from the non-IV case.

On replication fidelity. Our estimates match the original paper’s reported figures closely — within rounding — across all three main tables. The heterogeneity results replicate as well, including the wealth-quartile and regional breakdowns. This is a reassuring sign that the NFHS-4 data and the variable construction are well-documented enough to reproduce from scratch.

Final Thoughts

This replication confirmed the main findings of Dhanaraj and Mahambare (2022) cleanly. The positive and significant effect of paid work on IPV exposure in urban India, once endogeneity is corrected for, is a robust result. The channel decomposition is more tentative, but the partial evidence for male backlash is consistent with a broader literature on patriarchal norm reassertion in response to female economic gains.

For anyone working with NFHS-4 data or the bivariate probit estimator, I am happy to share the replication code — reach out via email.

References

• Dhanaraj, S., and V. Mahambare. 2022. Male backlash and female guilt: Women’s employment and intimate partner violence in urban India. Feminist Economics, 28(1): 170–198. https://doi.org/10.1080/13545701.2021.1986226
• International Institute for Population Sciences (IIPS) and ICF. 2017. National Family Health Survey (NFHS-4), 2015–16: India. Mumbai: IIPS.