Event Study and DiD Show Opposite Effects for Women – Interpretation Help

[Jheelum Sarkar]

Hi everyone,
Using repeated cross sectional survey data, I am studying the gendered effects of a flood on paid labor using difference-in-differences (DiD) and event study analysis. I estimate the effects separately for men and women.

For men, both methods tell a consistent story: hours worked per week declined after the flood.

However, for women, I see conflicting results:

• Event study analysis: Shows that hours worked increased at lag 1 (p-value is 0.096 during one period after the flood), but the effect becomes insignificant afterward.
• DiD regression: Shows a negative and statistically significant effect, suggesting that women’s hours worked declined after the flood—which contradicts the event study.
• Adding controls:
  • In the event study, adding controls makes all estimates insignificant for women.
  • In DiD, the treatment effect remains negative and significant, even after adding controls.

Variable Descriptions:

• tot_hrs – Total hours worked per week in paid labor (dependent variable).
• treat_krl – Treatment indicator (1 = individual is in a flood-affected area, 0 = control group).
• post – Post-flood period indicator (1 = after the flood, 0 = before the flood).
• time_treat – Event time relative to the flood (negative values = pre-flood, 0 = event period, positive values = post-flood).
• age – Age of the individual.
• no_yrs_formal_edu – Number of years of formal education.
• marital_status – Categorical variable indicating marital status.
• religion_new – Categorical variable indicating religion.
• social_group – Categorical variable representing caste or social classification.
• sex_ratio – Sex ratio in the district (used as a control).
• district – District fixed effects (categorical variable).
• weight – Sampling weight for survey data.
• gender – Indicator for gender (1 = female, 0 = male).

Stata Code:

Event Study Model:

eventdd tot_hrs i.treat_krl i.district [pweight=weight] if gender==1 & age>18, timevar(time_treat) vce(robust)
eventdd tot_hrs i.treat_krl age no_yrs_formal_edu i.marital_status i.religion_new i.social_group sex_ratio i.district [pweight=weight] if gender==1 & age>18, timevar(time_treat) vce(robust)
DiD Model:

reg tot_hrs i.treat_krl##i.post i.district [pweight=weight] if gender==1 & age>18, vce(robust)
reg tot_hrs i.treat_krl##i.post age no_yrs_formal_edu i.marital_status i.religion_new i.social_group sex_ratio i.district [pweight=weight] if gender==1 & age>18, vce(robust)

I have added the event study graphs (without and with controls) and DiD result tables (without and with controls) for women.
Concerns:

1. Why does the event study show an initial increase in hours worked for women at lag 1, but DiD shows a decline?
   • Could this be due to a temporary labor supply response by women right after the flood, followed by a reversion?
   • Or is it a methodological issue (e.g., treatment timing, selection bias, or controls affecting estimates differently)?
2. Why do results become insignificant after adding controls in the event study but remain significant in DiD?
   • Are the controls absorbing most of the variation in the event study but not in DiD?

Figure 1. Event Study Results after adding controls.

Figure 2. Event Study Results without adding controls.

Figure 3. Diff-in-Diff Results without adding controls.

Figure 4. Diff-in-Diff Results after adding controls (Due to screen size, I did not show the coefficients corresponding to each district)

[Carlo Lazzaro]

Jheelum:
some comments about your (too long) post (please, see the FAQ on how to post more effectively. Thanks):
1) why not using -didregress- or xtdidtregress to address your research question?
2) your second OLS specification shows a higher (altough still poor) R-Sq. I would check if your model is correctly specified using -linktest-.