Is Height Contagious? Detecting Implausible Social Network Effects 

 Some of you may be familiar with the recent work on social network effects in public health, where several studies have found significant networks effects on outcomes such as obesity, smoking, and alcohol use. We have blogged about some of this work  here   and  here . One key question with these findings is whether the observed relationship between one's own health outcome and the health status of other individuals in one's reference group is indeed causal or driven primarily by selection effects. Many have argued that confounding seems like a serious concern given that one's friends are not chosen at random. But at the end of the day it remains an empirical question whether the study design is able to account for these selection effects or not. 

 In " Detecting implausible social network effects in acne, height, and headaches: longitudinal analysis " Ethan Cohen-Cole and Jason Fletcher add to this debate with a series of interesting placebo studies. They demonstrate that the same empirical specification used in previous studies (ie. logistic regression of own health ~ friends' health + X) also "detects" significant and fairly large network effects for implausible outcomes such as acne, height, and headaches. For example, having a friend with headache problems increases the respondent's chances of headache problems by about 47% on average. These implausible placebo findings suggest that previous findings may have been driven by confounding. Similar placebo tests have been used in a variety of papers such as DiNardo and Pishke (1997), Rosenbaum (2002), Abadie and Gardeazabal (2003), Angrist and Krueger (1999), and Auld and Grootendorst (2004) to name just a few, but this study is another great example that demonstrates the power of such tests for social science research. I will use this as a teaching example I think. 

 Interestingly, Cohen-Cole and Fletcher also show that their implausible effects go away once they augment the standard model by adjusting for environmental confounders that may affect both an individual and her friends simultaneously. They conclude that "There is a need for caution when attributing causality to correlations in health outcomes between friends using non-experimental data." 

 I wonder how this debate will evolve. The ultimate test to disentangle correlation and causation would be to find a good natural experiment or to run field a experiment where social ties are exogenously assigned. Does anybody know of ongoing research that does this? It seems difficult to get something like this approved by research review boards of course.