*****  To join INSNA, visit  *****

Entering this conversation late, so I may have missed some things, but 
I wanted to put a plug in for using a combined empirical + simulation 
based approach to the problem, in case this hasn't been mentioned before.

If diffusion is thought of as a combination of network 
connectivity/structure, transmission probability, and duration of 
"infectiousness" (this is the classic formulation in epidemiology), there 
are quite a few parameters that can determine the dynamics of the 
diffusion.  Unless you're interested in a black-box type of projection, 
understanding how each of these components influences the process is an 
important goal.

Simulation of the process (before getting into the empirical trenches) 
allows for quite a bit of exploratory analysis and theoretical 
development.  The factors that influence the formation/dissolution of ties 
in the network may or may not be different than the factors that influence 
transmission across the ties.  But the impact of these two types of 
components on diffusion dynamics is quite different.  Presence/absence of 
ties influence the connectivity of the network, which has the usual 
non-linearities and threshold properties of a complex system (think 
emergence of the giant component), and this can become a key limiting 
factor in sparse networks (sexual networks are the paradigmatic example).

For those of us interested in the role of network structure, it is 
important that the model used to simulate the network is one that can also 
be used to estimate the model parameters from data.  That's one of the 
things that makes ERGMs attractive -- they do this double duty for 
estimation and simulation, allowing you to simulate networks that match 
the observable summary network statistics from your data.

You would also need empirical estimates of the probability of diffusion 
and the "duration of infectiousness"  to simulate the whole system.

In contrast to a single realization of the stochastic diffusion process -- 
which is what most empirical studies would give you (RCT or observational) 
-- simulating the network (with empirically estimated parameters for 
network dynamics and diffusion parameters) allows you to map out the 
distribution of outcomes from the system, and to see how sensitive it is 
to small changes in the underlying parameters.  Arguably, a lot more 
insight into the phenomenon of interest.


On Wed, 2 Mar 2011, Sinan Aral wrote:

> *****  To join INSNA, visit  *****
> Arun,
> I agree. Important and useful research begins with the right questions - that 
> goes without saying.
> But the problem of looking under the lamp post is not solely a problem of 
> RCTs. The same problem arises when we consider what observational data we 
> have access to. In fact, it seems likely that this problem would be more 
> pronounced for observational data that we happen to get access to than for 
> experimental studies that we have to design explicitly and in advance in 
> order to examine a particular question or set of questions.
> That we should use theory and scholarly intuition to seek out interesting 
> questions and phenomena to study is clear. However, in this case, the value 
> of the "question" of estimating social influence (broadly defined) is already 
> quite clear and its relevance is relatively well accepted. The reason why 
> causal estimation is important in this already well defined research area 
> (and people have been writing about this for decades) is that separating 
> correlation from causation in this specific case can inform us about what the 
> effects of various policy alternatives might be for programs aimed at peer to 
> peer HIV prevention, smoking cessation, obesity prevention, product 
> marketing, and so on.
> I could not agree more that the study that finds a valid instrument and then 
> searches for the question that the instrument helps address is misguided. 
> But, it doesn't seem to me that finding novel solutions to causal estimation 
> in networks leads us to the question of estimating the magnitude of peer 
> effects. In most contexts, that question itself is already well motivated. To 
> the contrary, it seems to me that our methods lag behind the theoretical 
> development of the questions (and the theories that explain social influence) 
> in this case.
> Another danger, besides looking under the lamp post, is remaining content 
> with showing correlation and assuming causation. Based on your own previous 
> work and our previous work together, I know you believe this as well. But, 
> its worth repeating in order to extend this discussion a bit I think.
> Cheers,
> Sinan
> Sinan Aral
> Assistant Professor, NYU Stern School of Business.
> Research Affiliate, MIT Sloan School of Management.
> Personal Webpage:
> SSRN Page:
> WIN Workshop:
> Twitter:
> On 2/28/2011 11:36 PM, Arun Sundararajan wrote:
>> *****  To join INSNA, visit   *****
>> The Onion clip is wonderful. Very much in snyc with the fake news
>> often being more informative than the real...should be required
>> viewing.
>> It seems to say a lot more than just "give observational data a
>> chance" to me in the context of this larger discussion. it isn't
>> merely that a large fraction of the phenomena we want to study occur
>> before we can design trials to measure them (and economists have been
>> dealing with this reality for decades). or that observational data are
>> more likely to lead to interesting discoveries of new things. Or that
>> whatever the methods, there are always alternative explanations,
>> especially when dealing with people in social settings. It's also that
>> if we start to believe in experiments and RCT's as the holy grail,
>> there's a danger of focusing too much on the kinds of questions that
>> lend themselves to that specific methodology, rather than going after
>> the ones that matter. (Even in the context of social influence in
>> networks.) analogously, there's a lot of time spent by research in
>> economics and marketing looking for "natural experiments" for
>> identification, and this gets to the point sometimes where it seems
>> like the research question was designed merely to exploit the cool
>> natural experiment...
>> i think that many aspects of this RCT vs. observational data (or for
>> observational data, matched-sample versus "structural" methods for
>> claiming causation) debates aren't unique to the context of social
>> influence in networks. For example:
>> cheers, Arun.
>> On Mon, Feb 28, 2011 at 12:40 AM, James Fowler<[log in to unmask]>  wrote:
>>> *****  To join INSNA, visit    *****
>>> We have also relied on methods like the ones Tom Valente mentioned in many
>>> of our observational studies, and we summarize the pluses and minuses of
>>> this approach in a new paper here:
>>> I also in principle like the actor-oriented model approach of Siena, but 
>>> in
>>> the past I could never get the model to converge for networks larger than
>>> 1000 nodes (this might be my own failing, though, as there is always a bit
>>> of art to getting models like that to work).
>>> We also have relied on experiments like this one in PNAS:
>>> and I am a big fan of David Nickerson's voter experiment and Sinan's new
>>> RCT.
>>> But I would resist abandoning evidence from observational studies.
>>> The resistance to observational studies reminds me of this Onion story:
>>> Multiple Stab Wounds May Be Harmful To Monkeys
>>> :)
>>> j
>>> James H. Fowler
>>> Professor of Medical Genetics and Political Science
>>> UC San Diego
>>> _____________________________________________________________________
>>> SOCNET is a service of INSNA, the professional association for social
>>> network researchers ( To unsubscribe, send
>>> an email message [log in to unmask]  containing the line
>>> UNSUBSCRIBE SOCNET in the body of the message.
> _____________________________________________________________________
> 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.

  Professor of Sociology and Statistics
  Director, UWCFAR Sociobehavioral and Prevention Research Core
  Box 354322
  University of Washington
  Seattle, WA 98195-4322

  Office:	(206) 685-3402
  Dept Office: 	(206) 543-5882, 543-7237
  Fax: 		(206) 685-7419

[log in to unmask]

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.