Complexity and Social Networks Blog: A dictum regarding social network analysis and causal inference

February 2008
Sun	Mon	Tue	Wed	Thu	Fri	Sat
					1	2
3	4	5	6	7	8	9
10	11	12	13	14	15	16
17	18	19	20	21	22	23
24	25	26	27	28	29

« Google bombs - Voice option and collective action | Main | Are silo's always bad? »

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 November 21, 2005 11:44 PM

Comments

I think the problem in the existing research is that it is very tempting to just choose one direction of causality in order to prove the arguments one wants to make! I totally agree with you that this discussion should at least be part of any research that deals with the structure-performance issue. However, I wonder whether there is a third variable one has to look at when studying structure-performance questions. Let's take the example of open source software development:

(1) One could argue that only skilled programmers are running the OSS community and the other less skilled programmers simply follow. In consequence, such "lead developers" (based on von Hippel's lead user idea), over time, will become central in the network, just because they are (a) considered as a really valuable knowledge base, (b) contacted very frequently, and (c) exploit the diverse contacts they have to turn novel information into new program modules. Backed by Burt's structural hole argument (1992) you would conclude: individual performance affects network structure.

(2) Now let's think of a very similar, but still slightly different argumentation. One is interested in the question: how can such lead users exploit their social capital to increase their performance/programming capacity? building on Moran's (2005) insights one might argue: a central lead user's development performance is optimal as long as there are enough "structural holes", but the more the lead user's ego network reaches closure, the less performant she will be in processing non-complex/non-innovative tasks. Now the intersting part: Moran (2005) has also been able to show that once ego networks are closed managerial performance is going up again!! This in turn means that managerial performance//lead developer performance depends on network structure ..... which is exactly the opposite to the conclusion under (1)

The issue I want to raise here is the following: It might be possible that different results might stem from different levels of analysis: do I look at the (capabilities of a) network actor and how they translate into network strucutre (as done in 1) or do I look at the ties that link one actor with all the other actors in order to find out how a variation of these ties affects the actor's performance (as done in 2)?

Posted by: Thomas Langenberg at November 23, 2005 12:36 AM

Complexity and Social Networks Blog

February 2008

4

5

6

8

9

19

24

25

26

28

Editor Login

Convener in chief:

Editors:

Categories

Archives

Recent Entries

Recent Comments

Notification

A dictum regarding social network analysis and causal inference

Comments

Notification

Post a comment