In 2007, a study published in the New England Journal of Medicine generated widespread media coverage for its claims that obesity can be transmitted via social networks, such as friendship, familial relationship or marriage. Details about the study appear in Issue 225 of this Update. The authors wrote additional papers on other personal characteristics, including smoking cessation, happiness and loneliness, concluding in each that a process of contagion or infection within the social network transmits the characteristics and that the transmission occurs up to three steps in the network, thus providing evidence of a “’three degrees of influence’ rule of social network contagion.”
A new study published in a lesser known journal, contends that the authors’ statistical analyses do not support their conclusions. Russell Lyons, “The Spread of Evidence-Poor Medicine via Flawed Social-Network Analysis,” Statistics, Politics, & Policy, Vol. 2, Issue 1 (2011). According to Russell Lyons, an Indiana University mathematician, the 2007 paper was based on insufficient attention to assumptions and misinterpretation of results not only by the authors, but also by its reviewers. He points to a corollary illustrated by this inadvertent misuse of statistics, i.e., “that top journals do not serve as rigorous judges of quality, due to lack of statistical competence.” Lyons concludes that “we need to improve our statistics education” and, given the difficulty he had getting his paper reviewed and published due to an apparent distaste for critiques, recommends the establishment of a journal specifically devoted to critiques.
At the core of Lyons’s critique is that the 2007 study authors analyzed associations “calculated from statistical models whose parameters are estimated by using the observational data.” They argued that the associations were not just associations, but measured causal effects, by ruling out the equally plausible possibilities, according to Lyons, that the associations were a result of “homophily (or selection), which is the fact that people tend to associate with others like themselves, and a shared environment (also called ‘confounding’ or ‘contextual influences’ by other researchers).” Lyons also points to a lack of statistical significance to the directional estimates in the papers and questionable assumptions made in the use of statistical models.