***** To join INSNA, visit ***** Dear Carolyn

I'm not an addict of power analysis although it does have its uses in the right circumstances.  But to apply it to a social network may be to miss the point - with all the data dependencies in a social network, it is likely that standard power analytic procedures, based as they are on techniques for independent observations, are quite inappropriate. And we don't know which graph distributions to apply to the data until we collect it and fit some models.

Of course, there are lots of social network analytic techniques that are not based on statistical models (eg cohesive subset analysis, structural equivalence, etc) and for these the issue of power should never arise.

But, sometimes, some people just insist that you need to do a power analysis.  In these circumstances I usually tell them that I am going to fit an exponential random graph or p* model, and if I were to do that using pseudo-likelihood estimation (not the optimal way to proceed but usually sufficient for this argument) I will be using logistic regression with the number of possible arcs as the cases.  So with 200 nodes, and a directed graph, you have of the order of 40,000 cases for one logistic regression.  They usually find this convincing enough.  (Of course, whether you can fit your research question into this framework is another matter...)

Good luck.

Garry Robins

Date: Wed, 07 Jul 2004 20:00:05 -0500
From: "Carolyn.Birmingham-1" <[log in to unmask]>
Subject: figuring statisitcal power?
Sender: Social Networks Discussion Forum <[log in to unmask]>
To: [log in to unmask]
Reply-to: "Carolyn.Birmingham-1" <[log in to unmask]>
X-Mailer: iPlanet Messenger Express 5.2 HotFix 1.16 (built May 14 2003)
X-Accept-Language: en
Priority: normal
X-Spam-Status: hits=0.001, required=5, tests=BAYES_50
X-Scanned-By: CNS Open Systems Group
X-Spam-Score:  (0)

*****  To join INSNA, visit *****


I got asked a question today that I had no clue how to answer... When doing network stuff how does one figure out statistical power (meaning a formula to put in how many nodes, variables, kinds of ties, etc.)? Is it dependent upon the software used? I just sort of figured with 200 plus nodes and 100% response rate that this wouldn't be an issue...but a regular stats jock wants a more concrete answer...

U of OK, Norman

Dr Garry Robins
Department of Psychology
School of Behavioural Science
University of Melbourne
Victoria 3010

Tel: 61 3 8344 6372
Fax: 61 3 9347 6618
_____________________________________________________________________ SOCNET is a service of INSNA, the professional association for social network researchers ( To unsubscribe, send an email message to [log in to unmask] containing the line UNSUBSCRIBE SOCNET in the body of the message.