Comparing agent-based models on experimental data of irrigation games

Agent based models are very useful tools for exploring and building theories on human behavior; however, only recently have there been a few attempts to empirically ground them. We present different models relating to theories of human behavior and compare them to actual data collected during experiments on irrigation games with 80 individuals divided in 16 different groups. We run a total of 7 different models: from very simple ones involving 0 parameters (i.e., pure random, pure selfish and pure altruistic), to increasingly complex ones that include different type of agents, learning and other-regarding preferences. By comparing the different models we find that the most comprehensive model of human behavior behaves not far from an ad hoc model built on our dataset; remarkably we also find that a very simple model presenting a mix of random selfish and altruistic agents performs only slightly below the best performing models