Complexity and Social Networks Blog

« November 19, 2005 | Main | November 25, 2005 »

21 November 2005

A dictum regarding social network analysis and causal inference

Figuring out the direction of the causal arrow is perhaps the major methodological issue of social science. This challenge is particularly acute in the study of social networks. Does the position in the social network affect success, or does success affect position in the social network? Do birds of a feather flock together, or do dogs and their owners starting looking alike? The strong structuralist tradition in social network analysis, which posits that networks are out of the reach of the agency of individuals, obfuscated this issue for a long time. With the increased attention to dynamic networks, and the development of tools to study how networks and individuals change at the same time (e.g., see the fine work on p* models, and related software, such as SIENA--http://stat.gamma.rug.nl/snijders/siena.html) there has been a dramatic improvement in the statistical toolkit available to deal with these issues. However:

(1) Longitudinal data do not guarantee correct conclusions regarding cause and effect. For example, one can imagine omitted variables dynamically affecting network and/or individual level variables, resulting in a spurious inference of causation in longitudinal data.

(2) Cross-sectional data can, under the correct circumstances, allow reasonable inferences of causation. Festinger’s classic study of social influence is, arguably, one such example.

(3) Despite the massive upsurge of social network related research, only a fraction of published social network research use longitudinal data, and only a fraction of a fraction of the studies that use cross-sectional data even hazard a sentence on where the network in question came from.

So, let me propose the following dictum for social network research:

Any research on the impact of social networks must at least wrestle with the factors underlying the network(s) under study, considering the possibility that (a) the network being studied was the result of the purported “impact� (i.e., reverse causation) and (b) some plausible third factor has affected both the network and potential outcome (i.e., spurious inference).

This is a pretty low bar, actually, but in many fields I would guess that close to 0% of the research exceeds it.

Posted by David Lazer at 11:44 PM