Individual Strategy Choice in the Repeated Prisoner’s Dilemma

Reanalyzing 12 experiments on the repeated prisoner’s dilemma (PD), we find strong evidence for players’ use of behavior strategies. Starting with unrestricted memory-1 strategies, the most parsimonious non-rejected representation of behavior distinguishes three subject types: defectors, cautious cooperators and strong cooperators. The defectors defect with a high probability in every round. Both cooperating types play semi-grim behavior strategies with different cooperation rates in round 1. This simple three-type mixture fits significantly better than 1046 combinations of (generalized) pure strategies from the literature, which we fitted at the treatment level. Semi-grim behavior strategies fit better than all 1046 mixtures of (generalized) pure strategies even when we use a constant and pre-defined specification, without using free parameters or any kind of post-hoc econometric magic. Furthermore, the resulting type shares correlate with the treatment parameters in a predictable manner, and the strategies themselves are largely predictable thanks to their approximate invariance, but the strategies cannot be rationalized as responses to expected payoffs.