Instrumental variables and obesity: micro vs macro, FIGHT!

The restaurants==more calories assumption of the obesity paper by Gomis-Porqueras and Peralta-Alva I wrote about the other day is suspect given the findings of this paper:

But simple correlations between restaurants and overeating may conflate the impact of changes in supply and demand. People choose where and how much to eat. A key question is whether the growth in restaurant supply, in terms of both number of establishments and portion sizes, is contributing to the obesity epidemic, or whether it merely reflects changes in consumer preferences.

What’s driving the spurious correlation?

First, there is selection bias in who eats at restaurants; people who eat at restaurants also consume more calories when they eat at home. Second, when eating relatively large portions at restaurants, people tend to reduce other calorie consumption at other times during the day. After accounting for these factors, eating a meal at a restaurant is associated with only 24 additional calories.

Restaurants aren’t the cause of the obesity problem, but they’re targeted with expensive regulations:

Restricting a single source – restaurants – is therefore unlikely to affect obesity, as confirmed by our findings. This mechanism may also underlie the apparent failure of so many targeted obesity interventions (Kolata 2006). Despite their ineffectiveness, such policies have the potential to generate considerable deadweight loss. We measure the potential deadweight loss of policies targeted at restaurants and find it to be as high as $33 billion annually.

Because the Gomis-Porqueras/Peralta-Alva paper wasn’t about obesity per se, the Anderson/Matsa’s conclusions don’t really contradict it. People are going to restaurants more, but that’s not why they’re getting fatter.

Having been baptized into the DSGE faith, I’m compelled by solemn duty to uphold the tenets of the Prescottian Creed and defend macro from the heresies of the unbelievers in the Public and Labor offshoot of the Reduced Form sect. In other words, I have issues with the Anderson/Matsa paper.

The biggest issue I have is, like too many micro papers, their conclusions are much more broad than their findings. Their sample includes only a handful of States and because they use highway access as a way to measure access to restaurants (and there’s good reasons to do so), they limit their study to rural zip codes that lie within 10 miles of an interstate.

In this sub-sample, they found a positive relationship between distance to a highway and the distance to restaurants. They also found no correlation between BMI and distance from highways. Suggestive, ain’t it. They then claim to be shocked (SHOCKED!) to not have found any confounding variables that when considered might remove an underlying relationship between BMI and access to freeways. Thus, restaurants have nothing to do with obesity anywhere. QED.

I feel like there might be a missing step or two connecting their findings to their conclusion. Well, ok in their conclusion section of the paper they threw skeptics like me a bone: “Although our results apply specifically to rural consumers, the central conclusions are likely to generalize to urban consumers as well.” Given this is all they say about generalizability, its not much of a bone.

Also, they’re measuring distance from the home in this study, but where do people (or husbands or wives) work? In town, near the restaurant maybe? Maybe people that live far away from restaurants do have access to them after all.