orgtheory.net

greetings from St Pete’s Beach or, waking up from the doldrums…

Omar

This is my last day at the 28th annual International Sunbelt Social Network Conference in Florida. After coming to this conference three times, I have to say that this is certainly my favorite academic conference (but of course that might not mean much, since I’ve never been to the Academy of Management Meetings). Most papers are usually interesting, and even though you do you get your occasional terrible presentation, those are still (even with the exponential growth of participants in the last few years) few and far between.

As you might have gathered from the title of the this post, this year’s edition of the conference is in St Pete’s Beach, Florida, which I found out is actually a little island about 35 minutes away from Tampa Bay. I played the dork and only went to beach once, although the weather was kind of cloudy most of the time. Not that I’m complaining, because this certainly beats shoveling snow at Indiana.

Ok, so here are a couple of general impressions:

Regression-type models for network data appear to be all the rage. I must have seen at least four presentations in a span of two hours touting one type or another of multi-level, latent class, random-effects, event-history, etc. model that incorporated some sort of “network effects” (usually a proximity matrix, or indicators for certain forms of local dependence such as reciprocity or transitivity bias), even John Skvoretz, who usually thumbs his nose at “statistical” models in favor of “mathematical” ones, wrote a really, really, long likelihood equation for his presentation (it didn’t fit into a single power point slide). All of this is without counting the many applied presentations that made use of the increasingly popular Exponential Random Graph Models, which are so popular, that I just found out that the cool kids now pronounce the acronym as if it was a (apparently Hebrew) word. So, just so you don’t sound like a moron, it’s ERGUMS, not E-R-G-Ms.

I think that this development is symbolic of how different “third wave” network “analysis” is from its 1970s Harvard cousin. The dissolution of the barrier between regression-type (and also stochastic) models and now “traditional” network models based on exploring and calculating graph properties, make a lot of the early rhetorical “chutzpah” associated with the network paradigm seem a tad outdated. If you remember those early position papers by Wellman, Berkowitz, etc., you realize that a lot of the initial identity of the network community and network analysis as a specific form of doing social science, was bought at the price of making claims about how different “relational” thinking was from “traditional social science” by which they meant Michigan-Wisconsin-North Carolina regression models.

The rise of regression type models for network data seems to confirm (in my view) not only that a lot of this early positioning of the network paradigm as a radical alternative in social science was misguided (because what it really boiled down was the juxtaposition of graph/discrete math-based models against probability/linear-algebra based ones), but also that the general linear model (thinking of it in the vaguest of ways as: some outcome [continuous, discrete, etc] is thought of as a parametric or non-parametric, linear or non-linear function of some set of indicators) is in fact a more “general” one than graph theory based ones. Of course, the more conciliatory position, is that different models are useful for different questions, and we certainly would never have gotten the radical rethinking of the notions of role and social position that came with the network revolution had it not been for the “thinking tool” provided by the graph representation of social structure. So calm down.

Another good thing about regression type models for network data? A lot of the presentations actually had pretty pictures and results (!).

A keynote address is a very ritualistic occasion. The keynote speaker this year was network theorist and analyst extraordinaire Steve Borgatti, whom I mentioned in a post a while back in reference to the 2006 Sunbelt conference in Vancouver. Borgatti’s talk was pretty much based on that earlier talk, but the tone was radically different. As Durkheim would have predicted, the ritualistic character of the occasion necessitated that he toned down the very critical and dour mood of the earlier version of the presentation. Instead, Borgatti seemed a little defensive, listing a variety of criticisms of the network approach (determinism, lack of consideration of agency, etc.) and dismissing them as “overly simplistic.” Thus, it appears that the ritual occasion produced a very real form of cognitive constraint, so that what came out was a ritualistic celebration of the strengths of the network position, rather than the previous soul-searching regarding its new found status as “normal science.”

Borgatti repeated his notion that there is a “core” of network concepts, quasi-models and image-schemas, and that this core is good and represents the best of the network contribution to social-scientific thinking (he also repeated his complaint of networks being perceived as merely “method” and not theory). While noting that most network theories can be classified into two kinds: theories of performance (most social capital theories and positional models of role-structure fall in here) and theories of homogeneity and “flow” (most diffusion theories fall here; for the most neglected uber-classic in this genre, see Fararo and Sunshine 1964) (of course, it was the genius of Granovetter’s signal contribution to derive a theory of performance from “flow” insights).

The key point that he made, is in my view, that this core is incredibly simple and unitary: it consists mainly of a balance model of triadic bias (the core of the flow/homogeneity model) coupled with a power-dependence model of the structural sources of advantage in exchange relations (the core of the performance model). Once you have that, you can pretty much “derive” about 90% of what we consider contemporary network theory (structural and regular equivalence, structural holes, strength of weak ties, etc.). The second point that I would add to that is, as you might have guessed, that that core was already established in 1982, so that nothing new has been added to it a quarter of a century (as Borgatti was quick to point out, clustering coefficients are just a variation on the balance theme). But Borgatti was in total ritual-celebration mood, so he didn’t mention the quarter-century thing.

So the question for Borgatti was: what’s outside of this core, and of this peripheral stuff what has the potential of making it into the core? Surprisingly, there was little outside of this core concepts that would be recognizable as “network theory.” Other candidates seemed variation on a theme (i.e. our notions of what “good position” is in a network are mostly based on models of “information flow” once we think about other kinds of things that flow on a network and about their inherent properties, such as duplicability, then notions of “centrality” and  power-dependence have to be modified accordingly; on the first score see Borgatti 2005, and on the second, see Schaefer 2007). In any case, it seemed to me that most of the stuff that Borgatti identified as possible candidates for entrance into the “core” had a lot of work to do before they get there (they are waiting for their Granovetter?). So the “innovation” problem in network theory (not network methods!) seems like it will be lasting a little bit longer.

Written by Omar

January 27, 2008 at 2:26 pm

Posted in networks, omar, sociology

11 Responses

Subscribe to comments with RSS.

  1. Hasn’t the core changed somewhat as people are increasingly interested in the network (i.e., positions in a network, network characteristics) the dependent variable? Dynamic network models, for instance, that try to explain how networks evolve over time seem completely different from the normal flow and performance models.

    Like

    brayden

    January 27, 2008 at 2:57 pm

  2. this certainly beats shoveling snow at Indiana.

    Amen to that brother. I could use a little beach, warm weather, and some general network dorkery right now, as I contemplate trying to scrape the ice that’s built up on my driveway.

    Like

    brayden

    January 27, 2008 at 4:50 pm

  3. Hi Omar,
    Well said. Was good to get to talk to you a bit while here.
    Wanted to say that i also had a reaction to some of what you said similar to Brayden’s – in that i think that the intent behind some of the Siena and “errghum” models is an attempt to retain this “analyzing social space in a fundamentally different way.” But as far as how those then get “applied” to “outcomes” beyond just tie-formation/dissolution – as you said, the normal-sciencyness of it definitely comes though in the nature of a lot of the talks going on here these days.

    Interesting that you see some of the bits that SB described being external to that core as still closely related. i will continue chewing through that some, but i think that i was focused on how much content he included in that core and how readily the story “worked” for them hanging together – more than most INSNA-ers probably previously considered. As such, i didn’t consider quite as much whether i find the things he framed as external to actually be so.

    Like

    shrinkingisaac

    January 27, 2008 at 6:33 pm

  4. btw – What’s going on with the time-stamps here? Did i miss something, or has it always been that way?

    Like

    shrinkingisaac

    January 27, 2008 at 6:34 pm

  5. shrinkingisaac – I think it’s always been that way. The time stamp is the time you posted the comment according to universal time (UTC), I think.

    Like

    brayden

    January 27, 2008 at 9:53 pm

  6. Brayden, you are absolutely right that the new stochastic models transform the ones and zeros of the usual proximity matrix into a long “response” vector and put them on “left-hand” side of the equation (as dependent variables). This had allowed network theorists to begin to wean themselves from the constipating aspects of thinking in graph theoretic terms (we all familiar with the “liberating” [aka, relational-manifestos type] aspects of graph theory), in particular thinking of the observed data as “the network” and thus as static and infra-structural. However, the novelty of the approach has not necessarily (yet!) translated into theoretical gains , for when we look at the “right hand” side of the equation our old friends are still lurking there: indicator variables for dyadic, reciprocity, triadic bias, etc. in other words: GOFNET (good old fashioned network theory) (BTW, you better have those in there, because they always seemed to have significant effects). In this sense ERGMs are a cutting edge tool, but they are still conceptually parasitic on flows and performance type of insights that form the core of traditional network theory (this was in fact shown in the title of Tom Snijders talk , which was something like “Beyond transitivity and reciprocity: What else is there? Let’s use SIENA to find out!”)

    Hi jimmy: great to get a chance to meet you!

    Like

    Omar

    January 27, 2008 at 11:19 pm

  7. What would Omar do?

    Go to Florida … to listen to network theory lectures!!!

    Like

    fabiorojas

    January 28, 2008 at 1:47 am

  8. Omar: I haven’t been to the Sunbelt in years, although it is a good conference. Was it all ERGMs, or did you see work using the Handcock/Raftery/Hoff latent space models too? The latter seems to me to be the future, but I have not read enough of the Snijders stuff to really do a fair comparison.

    Like

    Steve Morgan

    January 28, 2008 at 3:13 pm

  9. Hi Steve,

    It was mostly ERGMs, I think. I saw a couple of presentations that attempted to extend ERGMs by introducing some sort of discrete latent variable, essentially bringing in a sort of LCA logic into the analysis (i.e. assuming conditional independence within discrete classes). But none that I saw were conceiving of the latent space as a continuous, multidimensional space like HRH (I have only glanced at a couple of those papers, so I’ll follow your advice and read up on some on them!). Maybe there is a structural hole separating the Social Networks from the JASA crowd?

    Like

    Omar

    January 28, 2008 at 3:53 pm

  10. There are many structural holes in the networks crowd, and that is indeed quite surprising!

    I am reading a lot on causal effects and network analysis right now, and this is about the best thing I have found out there:

    Handcock, M.S., Raftery, A.E. and Tantrum, J. 2007. Model-based clustering for social networks. Journal of the Royal Statistical Society, Series A.

    It seems as if there is great potential in the approach. Perhaps it will be the fad at next year’s Sunbelt.

    Like

    Steve Morgan

    January 28, 2008 at 4:50 pm

  11. this is way later so noone will probably see it, but there were a rnumber of us taht were preplexed that Borgotti did not mention the very complete network theories of Harrison White. Perhaps the new version of Identity and Control coming out May 8th will change this roblem.

    Like

    Don Steiny

    March 9, 2008 at 9:41 pm


Comments are closed.

%d bloggers like this: