Presenters: Ned Smith and William Rand
Part of the Session on Social Systems
Link to model seems to be wrong.
Removed the bad model link (there is no code associated with this presentation). Thanks Miguel.
Very interesting application of ABM based on individual observable behaviours!
I have a question, though: going from micro to macro may overlook macro-level mechanisms that are acting and were not measured at the micro-level, while going from macro to micro may lead to a model that is fitted to a global pattern but completely misses the underlying micro-level mechanisms that generate the patterns, so what should be the best approach to follow in a model building cycle? What is the approach that will send us back to the drawing board more often?
On a side note, if I could get the salary of the highest paid friend of a friend on demand, the motivation to quit my job would be fairly high
I agree that the combination of experimental work and agent-based modeling can be a fruitful combination. In fact, there is quite some research on social dilemmas and behavioral finance where ABM and lab experiments are combined. Perhaps that is the research you capture under top-down approaches (given your Duffy 2006) reference.
It is good to realize that the goal of typical lab experiments is to test a hypothesis that builds on some theoretical foundation. The observed behaviors are not meant to be realistic outside of the lab (like physicists doing experiments in the lab using constructs that are also not observable in everyday life to test a theory). I looked up the Smith et al (2012) and noticed that the experiments (survey questions) are quite different than the experiments we typically use to combine ABM and experiments. So maybe you may want to contextualize the kind of experiments you are talking about (maybe those experiments you want to use do not focus on hypothesis testing)…
Interesting ABM study of social networks. Looking forward to seeing the act NetLogo model.
In your to be published paper, you described four approaches of validating ABMs: (1) micro-face validation; (2) macro-face validation; (3) empirical input validation; and (4) empirical output validation. The first two seems like model calibration by stylized facts, the third suggests calibration by empirical (input) data, and the fourth is the model validation by empirical data/patterns/stylized facts that is normally practiced. Is my understanding right?
Another question is about one of the model assumptions that agents’ wealth never goes down. It is an unrealistic assumption if empirical observations are used for validation of the model structure. What if you let agents spend a constant “metabolism” money whether employed or unemployed? Further, what if change the model so that each agent has a networking cost in some function of its social network? Sustaining a social network certainly costs money or other resources, and people who is more socially active is likely to spend more. I am curious how the simulation will look like if the two modifications are made. Will the results still be validated by empirical observations of job seeking?
You mentioned that he result of running the model was a disparity within the social network, as well as in wealth. However, the network already started as a power law distribution… which already has a few network-rich (super-connected) nodes and many more network-poor (low degree) nodes. Did you see the disparity actually increase substantially? Did the power law exponent change, or did it become more bimodal, with fewer nodes of medium status?
Very interesting question, I am always a fan of multiple methods, and multiple models to get at the underlying phenomenon. I do think that both micro->macro and macro->micro has its usefulness. I’m not 100% sure that there is a “best practice” for the cycle. In a paper I wrote with Uri Wilensky, called Full Spectrum Modeling, we discussed the desire to actually not focus on a single model, but on a suite of models. There we were primarily focused on the level of the detail of the models, but I could see it also applying to the modeling approach. I think often the best ordering would depend on where you have the richest knowledge before the modeling effort. If you have more low-level knowledge then start at the micro-level and vice versa if you have more aggregate level knowledge.
By the end of our model you only have a very small chance of getting your friend’s salary if you quite your job So it would be a gamble…
Marco, I think that many lab experiments that are confirming a hypothesis, are still doing that in the context of some underlying mechanism, and that mechanism could be tested further in an ABM. However, I agree that not all lab experiments will be equally well suited to this approach and it would be worth thinking about which ones those are.
I will put together a model website in the near future (hosted on OpenABM of course). Sorry just haven’t gotten around to it yet.
Great question about validation. I don’t know if I would call #1 and #2 validation by stylized facts, though that can play a role. We were more thinking of would an expert in the field consider the results to be true “on face”, so for instance, do people under job stress seek employment via their social networks? I"m not sure that’s a stylized fact, so much as a statement that most experts in the field would agree with.
Sure, we could also incorporate an expense function, but we felt that overly complicated the model. If we assume that everyone’s wealth goes down by roughly the same amount per tick, or some percentage per tick, the results will be equivalent. However, you have a very interesting point, if there was some way that those expenses were differentiated, such as different costs of social network expenditure then that would be interesting to explore.
Thanks for your reply.
You are completely right that the network already has a built-in disparity, but yes we do see that disparity increase over time. We did not focus too much on that since that was not the main point of this paper, but it would be interesting to look into in more detail. I didn’t do any rigorous tests of it.