About a month ago, another new paper on obesity (Christakis, N.A. and Fowler, J.H., 2007. The spread of obesity in a large social network over 32 years. New England Journal of Medicine 357:370-379) made headlines. The paper is based on an enormous data set that started being compiled in 1948, consisting of health information of thousands of residents of the town of Framingham, Massachussets. The original purpose of the study was to learn about causes of heart disease. Everyone in the study has received complete physicals every couple of years throughout their lives, and the data collection has continued on to the second and third generation of patients, which will presumably provide some information about the genetics of heart disease in addition to external causes. Christakis and Fowler use information about relatives and friends of the study subjects, included as part of the original data set, to assert that people whose friends get fat are more likely to get fat themselves.

The authors make the case that "[t]he spread of obesity in social networks appears to be a factor in the obesity epidemic." Although they never use the specific word "disease" in reference to obesity as some sensationalizing media outlets do, the thrust of the paper is that there is yet another cause of our obesity out there that is not our fault.

The advantage of using such a data set in this paper is that there is a lot of data about a lot of people over a long time period, and it is certainly understandable that scientists might conceive of other uses than the original purpose. The disadvantage is that the data set was not really designed to draw conclusions about obesity - and one of the major problems with the study is that the authors are promoting a cultural influence - based on interpersonal relationships - on weight gain using an extremely homogeneous sample, which does not represent a real cross-section of society. But, approached with an understanding of its assumptions and limits, Framingham-type data can indeed be useful for secondary studies.

The biggest limit of this analysis is its dependence on overlapping relationships among people. Although "social-network analysis" is not a technique with which I am highly familiar, it appears to consist mainly of high levels of pseudoreplication, which is a major problem for statistical analyses. Pseudoreplication is the use of data that are not independent, violating an important assumption upon which proper statistical analysis depends. There were 5124 focal subjects, and over 12,000 people total in the study, with an average of 7.5 social ties per person. The math on this clearly indicates that some people were analyzed as friends of more than one person. Thus, these data are not exactly independent. If the data were on people scattered about the country, so that each person's social network was independent of everybody else's, pseudoreplication would be avoided. With the data used as is, the statistical assumption of independence has been violated (although one can only conclude this in a roundabout manner; the author's use of jargon and limited statistical explanation makes their methods difficult to discover).

The most sensational assertion of the paper, that physical distance from one's friend does not affect the probability of becoming obese - and thus obesity of friends cannot be explained simply by them all having bad habits together - is undermined by the actual data, which are not nearly so conclusive. The authors broke physical distance into 6 rather absurd categories: 0 miles, 0.26, 1.5 miles, 3.4 miles, 9.3 miles, and 471 miles. Effectively, only the last group has true physical distance. Their conclusion is based on a nonstatistical difference, which may just mean that variation in the data is too large to detect a difference. In fact, the variation in their data is huge, with 95% confidence intervals (the statistical standard) often ranging over 50 or more percentage points. There is only confidence that having fat friends makes you fatter if the confidence interval does not overlap with a probability of zero. Looking at category 6 (471 miles) compared to the other groups in the figure below from the paper, four of the confidence intervals overlap with zero, as opposed to not more than 1-2 in the other distance categories, and upper confidence levels of probabilities of becoming obese are much lower than with shorter distances. The difference between category 6 and the others may not have been significant, but it is quite a stretch to conclude from this that distance from the friend does not make a difference in the probability of following him or her into obesity. (The six bars in each category represent data from six different health examinations over a person's life.)

For the authors, this paper was a no-brainer in two ways, though. If you can find a way to publish another reason why it isn't really someone's fault they are obese, by implying that the condition spreads from person to person like a disease, you've struck gold. In addition, because the conclusions are not that surprising at all (although they were spun as such by the NEJM media machine) it's easier not to pay attention to the statistical problems in the paper. But surely the authors' explanation that people thing it's more acceptable to be obese if their friends are is laughable to anyone. Do you know anyone who chose to be fat? "It really must be OK to let myself go if my friend has" just does not seem like a thought many people would have. But people who are friends will do things together, and if one of them can no longer do something physical, the other will end up hanging out watching TV with them, in a lowest-common-denominator effect. Because the assertion about physical distance not mattering is essentially bogus, a much simpler one is mutual lifestyle choice. It just wouldn't be very nice to dump a close friend because she got fat and couldn't ride bikes anymore.

Labels: health, humans, sociality