I just got back from the 2008 ASHE conference (see previous post) at Duke University. I had an absolutely stellar time. In fact, I'd say I got a lot more out of this conference than from those in the past. I think this has a lot to do with the fact that:
1) Having been to all those previous conferences, I actually knew people at this one. Not only is it fun to catch up with old friends, but having old friends makes it easier to make new ones.
2) I have a better sense of what I am interested in, so I can self-select into sessions that will better satisfy my curiousity.
3) The average presentation at ASHE was high quality and the conference was well-attended by the field's superstars.
4) My presentation went really well and was well received.
5) The conference was at Duke, my alma mater. I ended up taking my first economics class, taught by Allen Kelley, on a lark, which, along with two other Allen Kelley classes, planted the seeds for my decision to take up a PhD a few years later.
---------------------
Some broad themes from the conference. The sessions started with a keynote address by Mark McClellan, an MD/PhD economist with strong academic and policy credentials (he was in charge of the Centers for Medicare and Medicaid Services and the FDA), along with a now (in)famous brother. The basic theme of the talk was how health economics can help with policymaking, especially with an impending new administration. While it was exciting to hear that policy-makers are begging for policy-relevant evidence from health economists, it was depressing to hear that, despite nearly forty years of health economics, officials at the CBO and OMB are still forced to write sentences like "Evidence on the effectiveness of this intervention is inconsistent and limited" when trying to score new public health and health care policies.
Indeed, I saw this with my own eyes when contemplating the sum evidence from the numerous sessions on obesity. The papers therein covered a wide range of topics: the effects of food and drink prices, fast foods, law changes, and education on obesity; the impact of obesity on academic achievement and labor market earnings; and the economic costs of obesity. Each of these topics requires some heavy statistical lifting to get around the obvious causal inference issues. All of the conference papers came armed with the latest techniques. However, the sum knowledge from the slew of studies, in terms of policy relevance, appears to be limited. For example, the papers collectively suggested that food prices, fast foods, taxes and law changes appear to have only modest effects on obesity prevalence. As such policies centered around these factors, such as the popular "fast food tax" may lack heft as useful public health instruments.
--------------------
Since, I mentioned methods, I should point out that it was heartening to see such a high level of scientific rigor at these meetings. There was a healthy obsession with causal inference, and nearly all of the papers used fixed effects or instrumental variables. For those who are not into the jargon, fixed effects is best explained by an example: imagine you have data following the same individuals over time. You are interested in learning the effect of some X on Y, but are worried about bias from some omitted variable U, a characteristic about the individual. Fixed effects allow you to essentially subtract of the individual level mean from X and Y. If U doesn't vary across individuals, fixed effects basically washes this confounder out of the estimation procedure. Now you can get at the causal effect X on Y.
Sibling or twin studies are a good example of fixed effects. Let's say you want to learn the causal effects of birthweight on later wages. The worry is that birthweight is correlated with genetic factors or socioeconomic status early in life, which is correlated with outcomes later in adulthood. Using twin fixed effects, you can control for (or difference out) for genes and shared early life socieconomic status.
Of course, if the omitted variables that you worry about change over time for an individual or vary across a set of siblings, fixed effects estimates will not recover causal effects. To get around this researchers often use instrumental variable (IV) methods (either alone or in conjunction with fixed effects), where a factor that affects the regressor of interest, but not the outcome directly or indirectly through some other factor, is used to identify causal impacts. A famous example is the use of the quarter of birth or school cutoff dates to predict attained schooling (individuals born later and near the cutoff date are younger when they finish schooling or a full grade behind when they drop out).
The validity and theoretical basis for instruments are hotly debated, and researchers now must meet a pretty heavy burden of proof to get people interested and convinced. Also, it is now well-known that instruments must predict the regressor of interest strongly enough so as to yield valid estimates: "weak" instruments lead to biases of their own. It was nice to see that researchers at the conference went into great detail when defending their assumptions underlying the use of fixed effects and instruments. It was also nice to see the discussants call these assumptions into question and pointing out the additional implicit assumptions that needed to be made to achieve validity. This sort of discussion forces people to write down an explicit model and all the required identifying assumptions.
At some point though, I couldn't help but feel like there was some overkill and overpolicing, especially with the weak instruments issue (one researcher with a good IV strategy commented that she did not utilize the approach because her instruments were slightly weaker than what is considered kosher). Some new research on IV points suggests the use of estimation techniques that are robust to weak instruments: I was surprised that more conference-goers had not adopted these (indeed, mine was the only paper I saw that used these techniques). Second, some recent research suggests ways to proceed with instruments that might be "slightly" correlated with unobserved factors that influence the outcome and to test the sensitivity of IV results to varying degrees of instrument "badness." I would have liked to have seen a higher rate of adoption of these new techniques.
---------------------------------
In the next installment, I'll speak more about my ASHE experiences by summarizing some research on the early life origins of disease, behavioral economics, and the many benefits of education.
No comments:
Post a Comment