Column 1 assumes that the relationship between vaccination camps and gender imbalance is quadratic and estimates the coefficient on the interaction between girl and number of camps and the interaction between girl and the number of camps squared (as before, the number of camps is top-coded at ten). The coefficient estimates do point to a non-monotonic relationship. The linear interaction term is negative (increases in vaccination camps increase discrimination), but the squared term is positive. In this table, as in the following ones, for simplicity I show only the coefficients on the variables of interest. However, Appendix W.1 (available on the author’s web-site) shows the complete regressions in all cases.

Despite the evidence in the previous subsection that the placement of camps is unrelated to village-level socioeconomic status, there may still be concerns about these issues. Column 2 of Table 4 therefore replicates Column 1 but also includes controls for the interaction between gender and income, gender and income squared, and gender and education (linear and squared), as well as the existing controls for gender interactions with village population and distance. If the result on vaccination camps is being driven by some non-linear interaction of gender and another control, this specification should identify it. In fact, the coefficients in Column 2 are extremely similar to Column 1, suggesting that any correlations here make virtually no difference to the results.

Finally, in Column 3 of Table 4, I explore the relationship between camps and vaccination in a less parametric way. In this case, I include dummy variables for number of camps (1,2,3,4,5 and 6 or more – 0 is the excluded category) and number interacted with gender. This allows us to see exactly what is driving the results in Columns 1 and 2. The results are similar to what we would expect based on Figure 3. The only significant interactions are between gender and two or three camps; in areas with two to three camps the inequality is significantly worse than in areas with none. However, despite the lack of significance on the other coefficients, we see a pattern that mimics Figure 3: the coefficient on two camps and three camps is large and negative, and then we see less negative coefficients on interactions with more than three camps.

Table 5 replicates Table 4 but uses as the dependent variable a binary measure of whether the child has any vaccinations. The results mimic Table 4: we see evidence of a non-monotonic relationship, both when looking at the interaction with continuous number of camps (Columns 1 and 2) and in the interaction with number of camps measured discretely (Column 3).

Proposition 2 in the Section 2 suggests that the non-linear effect identified above should be stronger in areas with stronger son preference. To test this, I divide the sample based on each woman’s reported ideal sex ratio and replicate the regression in Column 1 of Table 4 for each half of

18