***** To join INSNA, visit ***** Weihua (Edward):
Let me make a few observations which will be limited given this forum but may be relevant (for more information I conduct a 1/2 day workshop at Sunbelt):
1) One finding from our studies was a very simple, obvious, but critical observation: You can not seperate the intervention from the mode of delivery. In all 3 papers we published from 2 randomized trials we found strong and significant interaction effects between network and intervention conditions.  One needs to design the intervention with network delivery as part of the process, not modifying an existing intervention to be "networked."  This makes the problem of testing network interventions more complex, but also more interesting.
2) We opted for randomization at the classroom level for 2 reasons: (1) It provides more units of randomization, thus increasing power and (2) within school randomization helps reduce bias inherent in school-level randomization (students from the same schools being similar).  So  I would urge classroom level randomization if possible.  On the other hand, middle school social networks span classroom boundaries (although we have limited data on this) and so this is a limitation.  Conversely, students named an average of 4+ friends within the classroom in 6th grade middle schools so clearly there are considerable within class friendship networks.
3) I think it is considerably more interesting to test variations in networks for an intervention rather than modes of identifying central people.  The centrality measures are pretty strongly correlated (see Valente et al., 2008) and depending on the network are not likely to identify different people.
In sum, I think it is not possible to provide a response to your question without knowing more about the intervention, but encourage you to pursue this work since this is a great application of social network methods

Valente, T.W., Coronges, K., Lakon, C., & Costenbader, E. (2008).  How correlated are centrality measures?  Connections, 28 (2), 16-26.

- Tom

Weihua An wrote:
[log in to unmask]" type="cite">***** To join INSNA, visit ***** Dear all,

I wonder if you could please shed light on how to conduct a project involving social network experiment. Basically, suppose I can conduct the experiment in roughly 60 middle schools and the outcome I am interested in is the prevalence of cigarette smoking among students. What experimental design with social network component will you view or suggest as innovative and important?

I can think of this on two lines. One is to test an important proposition in social network analysis that opinion leaders can accelerate the diffusion of (positive) attitudinal and behavioral changes. There will be three experimental conditions. The control condition is with no intervention. The first treatment condition is randomly picking changing agents. The second treatment condition is picking the opinion leaders as change agents. Then we compare the efficacy of these three conditions. Of course, we can add some twist to the procedure of how to pick opinion leaders, like contrasting opinion leaders chosen using in-degree with those chosen with eigenvector centrality or others, etc. I know there were a lot of studies using variations of this kind of design (Kelly et al. 1991; Latkin 1998; Valente and Davis 1999; Sikkema et al. 2000; Larkey et al. 2002; Stoty et al. 2002; Woff et al. 2004; Valente 2005), but I did not see any studies directly and purely using this design in public health literature. Does this design seem too obvious? I am also concerned that the individual-based interventions of this design may not be effective at all.

The other is to use group intervention. For example, Wing and Jeffery (1999) showed that group treatment was more effective than individual treatment in a weight loss program.

Some studies like Buller et al. (2000) and Valente (2003) combined the above two approaches, using opinion leaders to lead group interventions. I am concerned with interference between groups or cliques as called in Buller et al. (2000).  Also, the combination makes it hard to disentangle which network feature, between opinion leaders and group cohesion, matters in terms of preventing cigarette smoking.

I am also concerned with the unit of analysis. Because the number of schools are small (some may even drop off when they are asked to sign up), I thought about conducting the randomization on classroom level. But the interference between classrooms may become a big problem for inference of the intervention effects.

Whether you happen to work on this area or not, I would really appreciate your comments on my concerns and suggestions on how to make an "innovative" social network experiment given the above constrains, etc. Any perspective will be extremely welcomed. Thanks very much!



Weihua (Edward) An

Ph.D. Candidate in Sociology
Doctoral Fellow in Social Policy
Graduate Associate of IQSS
Harvard University
568 William James Hall
33 Kirkland Street
Cambridge, MA  02138
_____________________________________________________________________ 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.

Thomas W. Valente, PhD
Director, Master of Public Health Program
Department of  Preventive Medicine
Keck  School of Medicine
University of Southern California
1000 S. Fremont Ave., Unit #8
Building A Room 5110		
Alhambra CA 91803		
phone: (626) 457-4139
cell: (626) 429-4123
fax: (626) 457-6699
email: [log in to unmask]

My personal webpage:

The Empirical Networks Project

Evaluating Health Promotion Programs
(Oxford U. Press):
_____________________________________________________________________ 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.