Agent-Based Modeling is a powerful method for examining social processes. Social disciplines, including social psychology, are also concerned with social processes but use different methods and frameworks. A vast experimental psychological literature exists to describe human behavior, much of which is tested and communicated through linear regression models. The current work presents a method for leveraging published regression models to inform the construction of Agent-Based Models. When statistical models are translated from the literature into Agent-Based Models, the resulting simulations preserve many aspects of the original theoretical construction. We found that maintaining a common theoretical “language” facilitated communication among psychologists, even those with very little ABM experience.
What are your thoughts on fuzzy thresholds with the same approach? Assuming it would be of use with data where behavior is graded like the common 1-5 or 0-10 ranges
Thanks for your question!
I think fuzzy thresholds are functionally equivalent to the approach I presented. The random “fuzz” factor inherent in linear regression comes from the error term.
So, let’s assume some measure ranges from 1-5. Then, let’s presume that, for our linear model, values of 1 produce lower scores than values of 5. Then, let’s randomly sample from the gaussian distribution and add that to our score. When we add noise, this inherently fuzzes the scores.
On that basis, scores of 1 still have a chance of producing above-threshold outcomes - but this is proportional to the likelihood of sampling a sufficiently-large value from the error distribution. In contrast, a score of 5 is much more likely to be above threshold - but the error term could randomly drop it below threshold. It all comes down to the error distribution.
What are your thoughts about all that?
I have just uploaded a pre-print that discusses this method in greater detail:
This is chapter 4 of my dissertation, which I just defended last month.
It’s a work-in-progress and I am very interested to hear feedback!
In Europe there are quite a number of social psychologists using ABM (@wander.jager, Hans-Joachim Mosler, etc). Some of them, like Mosler, also base their models on observed regression models, while others, like @wander.jager, define process based models. Any reflections on this?
And yes you can make assessments of what would happen with an ABM, but how do you know your model is sufficiently capturing the phenomenon in order to make such extrapolations?
Thank you for your question! To be sure, ABM is has been used in social psychology for some time. For example: Jackson, et al. (2017) provide a tutorial-style introduction. Or, Macy & Willer (2002) for the sociologists.
My observation in social psychology (and behavioral economics) is that the type of questions that have been explored by ABM thus far are strongly influenced by economic decision-making games - and the types of research that make use of those games - because they are the most straight-forward to adapt to a computational model. However, I have not found that these games generalize particularly well to other kinds of behavior - and as such, the decision-games approach to modeling hasn’t worked well for my research.
The underlying principle is that whatever paradigm makes sense to the empirical research must also make sense in a modeling context. Decision games are one example of a strong match between empiricism and modeling. I suggest that regression models of behavior are another example - so long as a little care is taken when translating the regression to code.
I think the key here is to take the regression model as-is, straight from the literature. Based on your suggestion, I found Tobias & Mosler (2017), in which they do construct a linear model to drive their ABM. First up, thank you for the pointer! However, it looks to me that they’ve actually built their own theoretical model as an amalgamation of several other findings - and they have chosen to represent it as a linear model. This theoretical construction is subtly different from directly adapting a regression - whole, and in its entirety - from an article.
My claim is that if the regression is adapted, as-is, then the ABM preserves the properties of the original model of the phenomenon. Thus, if a regression model explains 40% of the variance, then an ABM directly based on that regression can reproduce at most 40% of the original phenomenon. So I am not able to claim that it captures the phenomenon - but I think we can make a strong claim about how much of the phenomenon it could capture.
Basically, the logic goes: the regression is a “True Enough” correspondence to the phenomenon (p < 0.05) - and insofar as it is True Enough, the ABM simulation of that model also has a fair shot at being True Enough. Unlike with the variance argument, I’m not going to claim that we could translate a p-value from a regression to an ABM - but I do think this is a future direction. I have some thoughts about this in the chapter I posted earlier in this thread.
Finally, the strongest way to argue that the model captures the extended phenomenon is to run a confirmatory study.
- Step 1 is the empirical observation and subsequent production of a linear model.
- Step 2 is the ABM implementation, which is used to create predictions (this is what my presentation is about).
- Step 3 would then be another empirical step in which the ABM predictions are either confirmed or falsified.
I haven’t used this study pattern yet - but again, I discuss it in the chapter I posted - because I think this would be a strong argument that the model has captured the extended phenomenon.
I haven’t really shared this argument widely before - so I’m genuinely curious to hear what you think about this line of thinking.
Jackson, J. C., Rand, D., Lewis, K., Norton, M. I., & Gray, K. (2017). Agent-Based Modeling: A Guide for Social Psychologists. Social Psychological and Personality Science , 8 (4), 387–395. https://doi.org/10.1177/1948550617691100
Macy, M. W., & Willer, R. (2002). From Factors to Actors: Computational Sociology and Agent-Based Modeling. Annual Review of Sociology , 28 (1), 143–166. https://doi.org/10.1146/annurev.soc.28.110601.141117
Tobias, R., & Mosler, H.-J. (2017). Optimizing Campaigns for Changing Routine Behaviors by Using an Empirically Calibrated Microsimulation Model. Social Science Computer Review , 35 (2), 184–202. https://doi.org/10.1177/0894439315620866