Breastfeeding Research and Intention to Treat 

 The current issue of The Atlantic features an interesting story by Hanna Rosin called " The Case Against Breastfeeding ." Rosin argues that the health benefits of breastfeeding have been overstated by advocates and professional associations and that, given the costs (in mothers' time and independence), mothers should not be made to feel guilty if they decide not to breastfeed for the full recommended period. One of her key points is that observational studies overstate the benefits of breastfeeding by failing to adequately adjust for background differences between mothers who breastfeed and those who don't. Observable differences, reports Rosin, are considerable: breastfeeding is more common among women who are "white, older, and educated; a woman who attended college, for instance, is roughly twice as likely to nurse for six months." In the course of making her argument Rosin provides a very nice layman's treatment of the difficulties of learning from observational studies; I think the article could be useful in teaching statistical concepts to a non-technical audience, although the politics of the issue might overwhelm the statistical content. 

 I followed up a bit on one experimental study she mentions in which researchers implemented an encouragement design in Belarus: new mothers in randomly selected clinics who were already breastfeeding were exposed to an intervention strongly encouraging them to nurse exclusively for several months, and the health outcomes of those babies as well as babies in non-selected clinics were tracked for several years. Rosin reports that this study found an effect of breastfeeding on gastrointestinal infection and infant rashes (and possibly IQ), but no effect on a host of other outcomes (weight, blood pressure, ear infections, or allergies).  

 I read what appears to be the first paper from the study ( published in 2001 in JAMA ), which reported that the intervention reduced GI infection and rashes. One thing that surprised me was that all health effects was reported in terms of "intention-to-treat," ie a raw comparison of outcomes in the treatment group and the control group, irrespective of whether the mother actually breastfed. The intervention increased the proportion of mothers doing any breast-feeding at 6 months from .36 to .5, so we know that whatever effects are found via ITT understate the impact of breastfeeding itself (because they measure the impact of being assigned to treatment, which changes breastfeeding status only for some mothers). (The authors know this too, and they raise the point in  a rejoinder .)  

 The standard approach I learned is to estimate a "complier average treatment effect" by essentially dividing the ITT by the effect of treatment assignment on treatment status, but the study appears to not do this. (The CATE for GI infection, according to my back-of-the-envelope calculation and assuming "no defiers," is about -.3, ie about a 30% decrease in the probability of infection for mothers who were induced to breastfeed by the intervention.) I suppose focusing on ITT's could be common in epidemiology because it addresses the policymaker's question of whether it's worth it to implement a similar program, assuming compliance rates would be similar. But for a mother thinking about what to do, the CATE gives much better information about whether or not to breast-feed.