I would like to thank Michael New for his research seeking to understand how public policy can most effectively reduce the number of abortions in America. His studies have made valuable contributions to this research agenda and my personal communications with him have prodded me to think a bit harder about our empirical tests. That said, I have a couple of initial responses to his most recent article.
– We were very clear in the Catholics in Alliance (CACG) study to report and highlight the finding that the abortion rate increases when states allow Medicaid funding to pay for abortions. We believe this finding is consistent with our overall findings which suggest that the abortion rate is sensitive to economic factors. We highlighted this finding when the study was introduced at the Democrats for Life Panel in Denver, and twice in the press conference that followed. We have also highlighted and discussed this finding in subsequent public blog discussions. It is inaccurate to suggest that we have tried to hide one of the findings of our study.
– Our initial data analysis found many economic factors that affect the abortion rate, in addition to Temporary Assistance for Needy Families (TANF) spending. For example, we found evidence that male employment reduces the abortion rate. In subsequent research, we have found that male employment and TANF spending are especially important in states with relatively low average incomes. We have also found that male employment is especially important in states with high marriage rates. In the conclusion to our study, we were clear about the implications of our research, suggesting that having sufficient socioeconomic resources to raise a child may be an important factor in whether pregnant women seek abortions.[ref]From the conclusion: “Socioeconomic policies are very blunt instruments for pursuing the goal of abortion reduction. Nonetheless, the results suggest that the economic status of pregnant women factors prominently into their abortion decision. That these factors matter at all, much less that the magnitude of their effect is so large, suggests that having sufficient economic and social resources to raise a child may be an important determinant of whether women carry a pregnancy to term, and that better targeted policies may be even more effective in reducing the number of abortions.”
[/ref] Further, the CACG study voiced support for laws such as parental consent and informed consent, even though the statistical analysis did not yield positive findings for these variables. To state that the study focused solely on trying to show that welfare spending reduces abortions is misleading.
– Weighting data is usually done to correct for the possibility that some states’ calculation (or estimate) of the abortion rate is more accurate than other states’ calculation of the abortion rate. Thus, to weight the data by population, as Michael New has done, implies that he assumes the data from high-population states are much more accurate than data from low-population states. For example, weighting by population assumes that the measure of the abortion rate in California is ten times as accurate as the measure of the abortion rate in Mississippi or Iowa, as the population of California is over ten times that of these other two states. We do not think this is an appropriate use of weights. If the concern is that states with low populations have systematically larger changes in the abortion rate during the 1990s than high-population states, then logged population may be appropriate as a weight without making unrealistic assumptions about the magnified accuracy of the data in states with large populations. Second, if the concern is that the variance in the errors for some observations is larger than for other observations (in the language of statisticians, heteroskedasticity), the standard method of addressing this issue is to weight observations by the inverse of the variance of the abortion rate at different levels of the explanatory variables (in this case population of women aged 15-44). To our knowledge, New’s results are based on simply weighting by population, and not the standard method of weighting used to address heteroskedasticity concerns. Our main findings remain if: (1) we use no weights; (2) we weight the data using standard methods to address heteroskedasticity; and (3) if we weight by the log of population.
– Regarding data that is potentially biased because it is reported by hospitals only, we agree that this may be a concern. These data would be biased if the change in the abortion rate reported in hospitals is systematically different from the change in the abortion rate in hospitals and clinics combined. To exclude this data out of concern for bias in the estimate of the abortion rate, the researcher should provide evidence that these abortion rates differ systematically. Further, even if bias is present, there is no evidence to suggest whether this helps or hurts the central findings. The best strategy, in our minds, is to report models that both include and exclude potentially biased data. In our working paper, we do this, and find that excluding the potentially biased states does not alter the main result. Finally, New’s studies commissioned by the Heritage Foundation have incorrectly excluded data from West Virginia when there is no indication from the original Centers for Disease Control and Prevention (CDC) sources that data from West Virginia were collected only from hospitals.[ref]For example, upon inspection of Table 3 in the following link, the reader can see that the data from 1990 for Alabama, Alaska, Iowa, New Hampshire, and Oklahoma all have two stars next to the number of abortions, indicating that the data are from hospitals. The data from West Virginia have no stars, indicating that these data were collected in a similar manner to the majority of states: http://www.cdc.gov/mmwr/preview/mmwrhtml/00031585.htm[/ref]
– The main difference between the statistical model that we employ and that which New uses is that our model allows for the possibility that factors such as male employment, TANF spending, or Medicaid funding can have both a short-term and a long-term effect on the abortion rate. In the model that New uses, the effects of the explanatory variable are assumed to occur only in the current year. This difference is important because some factors that affect the abortion rate may do so over the long term. For example, male employment may change the decision calculus of women considering an abortion both in the current year and over the longer term. This is also true of laws that restrict abortions. One argument in favor of laws that restrict abortion is that they provide a “teaching” effect, which, if true, may not be confined to the current year but would shape behavior in the long run. Thus, we believe that a statistical model that captures both short- and long-term effects is a better strategy than a model that assumes only short-term effects.