1.7 Sensitivity Analysis
1.7.2 Alternative Sample Restrictions
An additional concern is that the results may be sensitive to the source of information on remarriage. Recall that knowledge of a widow’s remarriage is contingent on her communicating in some way with the pension board. Specifically, I observe a widow’s remarriage if her children file a minors’ claim, or if she files a new claim under the act of March 3, 1901. If the source of information is distributed differently among women who remarry before and after obtaining a pension, and if the source of this information is correlated with marital outcomes, this might bias my results. As an example, recall from table 1.1 that minors’ pension applications are the source of evidence for remarriage in 64 percent of cases that occur before a pension is granted and 84 percent of cases that occur after a pension is granted. This means that my sample of women who remarry before receiving a pension may be disproportionately composed of childless women who lived to 1901. These women may be younger and healthier by construction, and thus better marriage prospects.
I use two alternative sample restrictions to address this concern. First, I restrict the sample to women who have children under the age of 16 when they are widowed, and I stop following these women once their youngest child turns 16. So, the sample is restricted to women whose marital status might be known through a minor’s pension application. Second, I discard any information that comes from a source other than a General Law pension claim, either widow or minor. Thus, any woman whose marital status is known only from a pension application under the law of March 3, 1901 becomes an observation with missing marital status.
Panels A and B of figure 1.3 plot the empirical hazard rate of remarriage by pension status, in similar fashion to figure 1.2, under these two sample restrictions. While the overall picture looks similar, as time passes the rate of remarriage for women without pensions starts to lie solidly below that of women with pensions. This could reflect the fact that the sample size is substantially reduced by these restrictions.
It may also indicate that the effect of the pension on women’s behavior is simply smaller for those with small children, so differences by pension status shrink when the sample is restricted to these women. However, we cannot rule out the possibility that differences in the source of information on remarriage are biasing the estimated effect of the pension away from zero.
The model described in section 6 is estimated under these sample restrictions, and the results appear in table 1.9. The baseline results, with and without a correction for correlated unobserved heterogeneity, are repeated in panel A. Panels B and C contain results from the sample restrictions outlined above. As seen in panel C, the results are not sensitive to the omission of information from pension applications under the act of March 3, 1901. When the sample period is restricted to years in which the widow has a minor child, the estimate remains negative; however, it decreases in magnitude relative to the baseline, and the standard errors increase. In panel B, we can only say with about 80 percent certainty that the coefficient is different from zero. Still, these results broadly support the finding of a negative effect of the pension, even if the estimate becomes noisier under one of the sample restrictions.
Panel C of figure 1.3 and panel D of of table 1.9 impose a different sample re-striction. These use only women who are successfully linked to the census of 1870 and/or 1880. These data provide independent verification of the information on mar-ital status in the pension files. Women have an incentive to lie to the pension board about marital status; however, there should be no such incentive to lie to census
enumerators. By including only women whose marital status can be verified in the census, I mitigate accuracy issues that stem from pension fraud. Another benefit of the linked data is that it allows me to observe potentially important demographic variables such as birthplace and literacy. As seen in figure 1.3 and in table 1.9, the results are not sensitive to restricting the sample to women linked to the census, or to including controls for immigration and literacy. Panel D of figure 1.3 and panel G of table 1.9 restrict the sample to women widowed during the war years. Dying during the war is arguably more random than failing to recover from a non-life-threatening injury or disease contracted during the war, so it is worth verifying that the results are robust to this sample restriction. The restriction has little effect on the estimate.
Finally, I estimate OLS and 2SLS models that are similar to those in the previous subsection, restricting the sample to women who are linked to the census of 1870 or 1880 through children from their first marriages. As explained earlier, it is desirable to use an alternative way of identifying remarried widows, as the source of marriage information in the pension data may generate artificial differences between widows who remarry and those who do not. Table 1.10 contains results from regressions of an indicator for being remarried in the census on an indicator for having received a pension within five years of applying.58 These are similar to the regressions presented in table 1.8. In column 1, I use links to the 1870 census; in column 2, I use links to the 1880 census; and in column 3, I pool both years and cluster standard errors by widow.
Columns 4, 5, and 6 repeat these specifications using two stage least squares, where the instrument is the name homogeneity index used earlier. This instrument explains a reasonable amount of variation in pension status for the sample linked to the 1880 census, but it performs very badly for the sample linked to the 1870 census. This suggests that much of the variation being explained by the instrument is coming from
58I also try this with different time frames, and the results are similar.
women widowed in the later part of my sample.59 Still, while the number of women linked in this fashion is small, and the estimates are often noisy, these results broadly support the basic findings. The coefficient on pension income is always negative, and the 2SLS estimate is significant at the ten percent level when the 1880 census is used.