Causal ABMs: Learning Plausible Causal Models using Agent-based Modeling

We present Causal ABM, a methodology to derive causal structures describing complex underlying behavioral phenomena. Agent-based models (ABMs) have powerful advantages for causal modeling that have not been explored sufficiently. Unlike traditional causal estimation approaches which often result in "one best" causal structure that is learned, two properties of ABMs - equifinality (the ability of different sets of conditions or model representations to yield the same outcome) and mutlifinality (the same ABM might yield di.erent outcomes) - can be exploited to learn multiple diverse "plausible causal models" from data. Using an illustrative example of news sharing on social networks we show how this idea can be applied to learn such causal sets. We also show how genetic algorithms can be used as a estimation technique to learn multiple plausible causal models from data due to their parallel search structure. However, significant computational challenges remain before this can be generally applied, and we, therefore, highlight specific key issues that need to be addressed in future work.