Cloud asks
Have you seen this study? http://people.virginia.edu/~am5by/fertilitytiming_sept2009.pdf
And, more importantly, what do you think of it?
The author of that study was hot hot hot on the market this year. However, rumor is she’s decided not to move, despite offers.
The Effects of Motherhood Timing on Career Path
Abstract
This paper estimates the causal effects of motherhood timing on female career path, using national panel data from the NLSY79 and biological fertility shocks to instrument for the age at which a woman bears her first child. Motherhood delay leads to a substantial increase in earnings of 9 percent per year of delay, a smaller increase in wage rates of 3 percent, and an increase in hours worked of 6 percent. Supporting a human capital story, the postponement premium is largest for college-educated women and those in professional and managerial occupations. Family leave laws do not significantly influence the premium. Panel estimation reveals evidence of both fixed wage penalties and lower returns to experience for mothers: a “mommy track” is the likely channel for the timing effect.
All righty. Before we get into the study itself, Ima gonna explain a little bit about empirical econometrics. Economists are very interested in the idea of causality. We’re not the only social science that focuses on causality, but we’re more likely to use something we call “natural experiments” rather than actual experiments (as psychologists do) to explore questions of causality (psychologists call these quasi-experiments). X and Y are correlated, but does X cause Y, does Y cause X or is there a third variable, Q, that causes both?
One of the types of natural experiments we use is something magical called “Instrumental Variables” or IV for short. IV is really neat because it basically takes a variable Z that we know causes X and does not cause Y (except through its effect on X). The canonical example of a good IV is the Vietnam draft lottery number to study the effect of Vietnam service on labor market outcomes. A worse draft number is correlated with actually serving in Vietnam, but because it is randomly assigned, is not related to labor market outcomes except through its effect on Vietnam service. A “good instrument” will have these two qualities: Z will be correlated with X, and Z will not cause Y except through the channel of X. Most instruments are not as good as the Vietnam draft lottery, but we believe that they will tell us something anyway… but generally we don’t think these IV papers give the final word, just additional evidence.
In this paper, Amalia Miller uses several imperfect instruments to look at the effect of motherhood timing on female labor market outcomes. Remember, in order to be a good instrument, the instrument will have to be strongly correlated with timing of motherhood and not correlated with female labor market outcomes except through the channel of motherhood timing. The former we can test using statistics: there are some heuristics we use for the t-stat, and there are some tests for low coefficients or “weak instruments” that are more or less accepted depending on your training. The latter is where thinking is needed.
Miller’s chosen instruments are:
1. whether first pregnancy ended in miscarriage
2. whether conception of the first child occurred while using contraception
3. elapsed time from first conception attempt to first birth.
Now think: Are these related to female labor market outcomes through any means other than pregnancy timing?
I would argue yes: All three of these are related to maternal health (including obesity), and health is directly related to labor market outcomes. Miscarriage not only affects timing of motherhood, but also can cause depression, which relates to labor market outcomes. Different types of people use contraception (because of religion, sense of responsibility etc.) and these personality characteristics may be directly related to labor market outcomes. And infertility itself can be very time-consuming (I read a paper the other year on just on how much time is spent at the doctors, and how it has to be spent during the working day for the most part, for example). I’m sure you can think of many more ways that Z, the instrument, relates to Y, the outcome variable through channels other than X.
Does that kill her results? Well, no. It just means we can’t be confident in the answer given in this paper (and it took a while to get this paper published and it’s not in one of the standard top econ field journals, probably for the above reasons). If we attack the same question in many other imperfect ways and get the same results, we’ll feel more confident with the results in this paper. We don’t really know the answer, but this should update our Bayesian priors to the results that she finds. If we do come up with a perfect instrument or run a randomized controlled experiment (which we won’t do because we’re not Nazis) or get a much cleaner natural experiment that we can use a cleaner technique on (say, differences-in-differences) and get different results, then those results will trump these results. In the absence of that, this paper is providing more information, and provides an improvement over the previous literature which is, at best, getting only at correlation.
So that’s a critique of the coefficients she finds– are they really accurate? Probably not, but they’re probably closer to the “truth” than what the previous literature has found.
Let’s take the results as given. Are her conclusions merited given the results?
Supporting a human capital story, the postponement premium is largest for college-educated women and those in professional and managerial occupations. Family leave laws do not significantly influence the premium. Panel estimation reveals evidence of both fixed wage penalties and lower returns to experience for mothers: a “mommy track” is the likely channel for the timing effect.
Well, let’s throw that open to the grumpy nation. What do you think? What are alternate explanations?