Friday 25 May 2012

9/11 as an Instrument

It's hard to tell in cross-sectional data whether depression is cause, consequence, or spurious correlate of risky sexual activity. Susan Averett and Yang Wang try to figure it out in the latest AER. Their trick? AddHealth data collected over a period that spanned 9/11 that lets them use the terrorist attacks as an instrument for depression.

Some summary statistics:
Our full sample consists of 4,151 women aged 18 to 26 years. Most of our respondents are white (68 percent) or African American (25 percent). The average years of education are 13.5 and the average age is 21 years. Thirty-five percent smoked for at least 30 days in the past year, and during the year before interview 72 percent of these women drank alcohol and 31 percent used marijuana. Selfrated health averages 2.05, which is slightly worse than “very good.” The mean CESD score for the nine questions asked in the Wave III 95 percent of these women are sexually active in their relationships. Sixty-seven percent and 11 percent of them engage in oral and anal sex, respectively. Fewer than 2 percent of the women in our sample report using condoms.
An active sample. And, an active sample that showed a sharp jump in depression scores in the period right after 9/11. They then use 2SLS with 9/11 as instrument and find that OLS estimation underestimates the effect of depression on participation in risky sexual activity.
So a one point increase in the CESD (depression) score correlates with a 0.4% increased likelihood of vaginal sex by OLS, but a 2.1% increase by 2SLS - and a 1.9% reduction in the likelihood of using a condom. 

It's worth keeping this kind of result in mind when we hear findings that alcohol use correlates with risky sexual behaviour [and, let's not forget, with more positive consequences of sexual experiences]. If depression correlates with heavier drinking, then it's pretty easy to conflate the effects of alcohol with a covariate, like depression, that drives both drinking and risky sexual practices. 


  1. Hm. No access to the paper right now, but - am I supposed to believe that 9/11 had no effects on sexual behaviour except via depression? That seems rather hard.

    1. I don't have access from home either, but it's 2SLS, so they're getting in the first stage the predicted effect of 9/11 on depression scores, then using predicted depression in the second stage.

    2. Yes, but 2SLS results (and IV results more generally) are only valid if the endogenous regressor (depression in this case) is the *only* channel through which the instrument affects the dependent variable. If the model were overidentified, you could run a Sargan test, but if I understand correctly, it's just-identified, so we can only assess this aspect by using intuition. And my intuition says that's not credible.

    3. I'd thought they'd tested for instrument exogeneity; I'll have to check the paper again. But you're entirely right that there are other plausible mechanisms for 9/11 to affect the dependent variable. I'd wonder too whether geographic controls for folks living in NYC or DC/NOVA might be warranted.