Page 85 - IJPS-11-4

International Journal of
Population Studies Droughts and intimate partner violence

PV increased by 50%, with statistical significance at the 5% ranges from 0.5 to 1.0 and is used to measure the model’s
level. EV also saw an increase of 46% (p < 0.05). However, ability to distinguish between subjects who experience the
the estimates for SV in rural areas were not statistically outcome of interest and those who do not (Hosmer et al.,
significant. In contrast, the urban sample exhibits a 98% 2013). The area under the ROC curve for our PV model
increase in the odds of SV, which is significant at the 10% is 0.8207, for the SV model, it is 0.8461, and for the EV
level. model, it is 0.8281. According to Hosmer et al. (2013),
these values indicate that the models demonstrate excellent
3.4. Robustness, model diagnostics, and goodness discrimination and fit well.
of fit
2
For the PV model with full controls, the pseudo R value
3.4.1. Robustness check is 0.2402, indicating that the model explains approximately
Falsification test: To ensure the internal validity of our 24% of the variation in the data. Similarly, for the SV and
2
results, we conducted two falsification tests: EV models, the pseudo R values are 0.2274 and 0.2324,
a. Reversing the timeline: We reversed the “pre” and respectively, demonstrating good model fit.
“post” periods. The estimation results were the 4. Discussion
exact opposite of the actual results and statistically
significant (Table S7). Recent studies have explored the complex relationship
b. Switching the timeline: We also tested by limiting the between droughts and IPV, producing varying results
“pre” period to 2015 and grouping data from 2016, (Cools et al., 2020; Cooper et al., 2021; Epstein et al., 2020).
2019, and 2020 as the “post” period. Table S7 presents Rai et al. (2021) examined this relation within the Indian
these results, which show that exposure to drought is context but found no statistically significant link, likely
no longer associated with increased odds of IPV. This due to the overlap in the data collection period with the
further supports the internal validity of our original drought, which may have limited the ability to capture the
results, as the effect of drought exposure is no longer full impact of the drought. In contrast, our study aimed to
observed. resolve this ambiguity using a dataset in which the recall
period for domestic violence modules aligns with the
Both falsification tests, which involved changing the timeframe of the NEM drought, allowing us to assess the
drought exposure timeline, confirm that our findings are drought’s effect more comprehensively.
not a result of mechanical errors and that there is a strong
association between drought exposure and increased odds We began by examining whether the treatment and
of IPV. control groups had similar IPV incidences during 2015
– 2016. Using proportions tests (Table 2), we found that
3.4.2. Model diagnostics and goodness of fit the treatment group had a marginally higher proportion
To assess the goodness of fit of our estimated model, we of respondents experiencing IPV, except for those with SV,
applied the classification test and the receiver operating in 2015 – 2016. The differences increased further in 2019 –
characteristic (ROC) curve. The classification test 2021, where all three forms of IPV registered a statistically
compares the model’s predicted response (positive for IPV significant increase for states exposed to the NEM drought.
or negative) with the actual observations. A well-fitted These results are consistent with Epstein et al. (2020)
model should correctly identify both positive and negative findings and contradict those of Cools et al. (2020) and
outcomes. Here, we discuss the results of the PV model. Cooper et al. (2021). In the Indian context, Rai et al. (2021)
also reported increased PV post-exposure to drought;
Table S8 shows that the model predicted positive however, their result was not statistically significant. Our
responses for 5607 observations, of which 3807 were dataset captured the full effects of the drought, enabling us
correctly classified as positive (y = 1), whereas 1800 were to find statistically significant results.
incorrectly classified because the actual response was The NFHS asks a question regarding the number of
negative (y = 0). Of the 24,493 observations for which the “control issues” the respondent faces. We explored how
model predicted a negative response, 19,909 were correctly the number of reported control issues changed in 2015
classified, whereas 4584 were incorrectly classified. The – 2016 and 2019 – 2021 (Table S3). The negative sign of
overall classification accuracy of the PV model is 78.79%. the point-biserial correlation coefficient (−0.0142; n =
For the SV and EV models, the correct classification rates 29652; p < 0.05) indicates that the number of control issues
were 94.02% and 87.97%, respectively (Tables S9 and S10).
reported by respondents from states with no exposure
We further calculated the area under the ROC curve, to rainfall shock has a negative relationship with the
as shown in Figures S1-S3. The area under the ROC curve number of control issues. This suggests that the number

Volume 11 Issue 4 (2025) 79 https://doi.org/10.36922/ijps.3065