the declining role of simulations in sociology
Over the weekend, I got into an exchange with UMD management student Robert Vesco over the computer science/sociology syllabus I posted last week. The issue, I think, is that he was surprised that the course narrowly focused on topic modelling – extracting meaning from text. Robert thought that maybe there should be a different focus. He proposed an alternative – teaching computer science via simulations. Two reactions:
First, topic modelling may seem esoteric to computer scientists but it lies at the heart of sociology. We have interviews, field notes, media – all kinds of text. And we can move beyond the current methods of having humans slowly code the data, which is often not reliable. Also, text is “real data.” You can easily link what you extract from a topic modelling exercise to traditional statistical analysis.
Second, simulations seem to have a historically limited role in sociology. I find this sad because my first publication was a simulation. I think the reason is that most sociologists work with simple linear models. If you examine nearly all quantitative work, you see that most statistical analyses use OLS and its relatives (logits, event history, Tobit. Heckman, etc). There’s always a linear model in there. Also, in the rare cases where sociologists use mathematical models for theory, they tend to use fairly simple models to express themselves.
Simulation is a form of numerical analysis – an estimate of the solutions of a system of equations that is obtained by random draws from the phase space. You would only need to do this if the models were too complicated to solve analytically, or the solution is too complex to describe in a simple fashion. In other words, if you have a lot of moving parts, it makes sense to do a simulation. Since sociological models tend to be very simple, there is little demand for simulations.
Robert asked about micro-macro transitions. This proves my point. A lot of micro-macro models in sociology tend to be fairly simple and stated verbally. For example, many versions of institutionalism predict diffusion driven by elites. Thus, downward causation is described by a simple model. More complex models are possible, but people seem not to care. Overall, simulation is cool, but it just isn’t in demand. Better to teach computer science with real data.
50+ chapters of grad skool advice goodness: Grad Skool Rulz/From Black Power
orgtheory.net 
Fabio’s suggestion that most sociologists don’t care about complex models makes for an interesting contrast to people who do care. You do find public servants and public policy practitioners thinking about complexity, both in the context of policy making within any one country and for international development.
Most interesting policy questions, when addressed in a forward looking manner, present themselves as complex. For example, whether elite diffusion is valid or not, it is very hard to manage all the moving parts entailed in institutional reform in a developing country – and most such policy reform efforts fail. Further, using insights from simple models as a basis for policy in such circumstances is deeply dangerous in that it invites technocratic top-down decision making that lends itself to the worst kind of unintended consequences. Thorny policy questions are often sensitive to initial starting conditions, have a host of interacting factors (many of which are specific to the local conditions), and pose huge difficulty in isolating cause and effect. Perhaps simple models are good for posing post-hoc explanations, but less so for forward looking policy guidance.
Some types of problems are poorly suited to linear modeling, but then that’s not all that sociology has to offer. Unintended consequences and emergent phenomena are core concepts in sociology. I’m not convinced by simulations as a tool myself (side note: if this is the decline, what was the high point?), but simulations are trying to take complexity seriously and there’s a strong argument to be made that many interesting policy questions (especially when looking forwards rather than backwards) should be approached as complex and adaptive systems rather than linear ones.
LikeLike
ijlaln
September 30, 2014 at 7:49 am
While it looks like the focus of the course is on text analysis, only one week of the syllabus is about topic modeling. The rest of syllabus looks to be an introduction to techniques that are popular in the data analytics world, like clustering and classification, that can equally be applied to textual and non-textual data. Sadly most of the classification algorithms require that “humans slowly code the data.”
Additionally, I wouldn’t say that topic models “extracting meaning from text.” Topic models are a data reduction technique, similar to factor analysis. We don’t say that reducing 20 Census tract variables to 3 factors “extracts meaning”, and there’s nothing magical about topic models or text data.
That said, when I think about “Computational Social Science”, I usually think of it is as quantitative text analysis + networks + agent based modeling.
LikeLike
neal caren
September 30, 2014 at 1:18 pm
The high point was probably the publication of Schelling’s segregation model in J Math Soc. I pay attention to Michael Macy and his colleagues at Cornell, but I suppose Fabio is half right: there isn’t much interest in sociology in simulations. There should be, though.
LikeLike
Randy
September 30, 2014 at 1:39 pm
I agree that there’s not as much interest in simulations as there should be but I’m very skeptical that it’s declining. I published a simulation a few months ago in Sociological Science. Likewise McClelland had a simulation in the last Sociological Theory. And of course you have young faculty like Centola who are primarily known for simulation work. And then you’ve got people like Bruch or Garip who do simulations that are heavily constrained by empirical data.
LikeLike
gabrielrossman
September 30, 2014 at 4:29 pm
@gabriel and others: Fair point. It isn’t exactly zero, but in grad school there were multiple efforts to make simulation platforms for social science (http://repast.sourceforge.net/ and http://en.wikipedia.org/wiki/Swarm_Development_Group). These, for the most part, are ignored in sociology. But times change, perhaps Big Data will draw more attention to computation in general and people will want to learn about related ideas like simulation.
LikeLike
fabiorojas
September 30, 2014 at 6:55 pm
1) As an “overview” course in computational social science, as it was billed in the description, I think simulations and modelling are deserving of being included. Not necessarily as “the focus” but at least as a component.
2) The problem with simulations and their popularity is partly endogenous and partly due to prior technical constraints. The endogenous part is that they are not popular, so few people teach and understand them, so few people do them, so they remain unpopular.
The technical part is just as the power of computers is making many kinds of big data / machine learning methods feasible, it is also doing the same for simulations and network modelling. Though this application of computing power has been a bit slower to catch on. Simulations that used to take days, can now be done in hours if not minutes. Moreover, we have powerful graphical and statistical tools to make the output of simulations more accessible than ever before. The tools Fabio mentioned are also helpful in making them more accessible though I’m not a fan of black boxes.
In short, I’m excited and bullish about the future of simulations. I think modern computing power will allow us to not only engage in “big data” but also in “big theory”. While the former has been leading the way, I suspect the latter will soon follow. And hopefully we’ll be combining them together. But to get there we will need more people teaching these methods and making them well-known. Fabio’s rabble-rousing is a good start :)
LikeLike
Robert V
October 4, 2014 at 2:17 pm