Learning in sender-receiver games

Stimulus-response (SR) and belief-based learning (BBL) models are estimated with experimental data from sender-receiver games and compared using the Davidson and MacKinnon P-test for non-nested hypotheses. Depending on a certain adjustment parameter, the P-test favors the SR model, the BBL model or neither of the models. Following Camerer and Ho, the models are also compared to a hybrid model that incorporates a mixture of both types of learning. The hybrid model is frequently not significantly better than either the SR or the BBL model. The sensitivity of the results to observations taken after learning has ceased is investigated.