Replication Report: On the Robustness and Provenance of the Gambler's Fallacy by Xiang, Dorst, and Gershman (2025)

Xiang et al. (2025) investigate whether the gambler's fallacy-the false belief that a random event is less likely to occur if it has occurred recently-is robust to using probabilistic (versus point) predictions and independently and identically distributed (versus non-IID) sequences. In five experiments with 150 participants each observing 18 sequences of colored balls, the authors find strong evidence of the gambler's fallacy when eliciting point predictions or using non-IID sequences, but fail to observe a robust gambler's fallacy when eliciting probabilistic predictions over IID sequences. We successfully reproduce all main results from the processed data and analysis code. The absence of raw data precludes robustness checks incorporating response times, attention checks, or demographic controls. We conduct two main robustness analyses that both validate the main finding that point predictions are fundamentally different from probabilistic predictions, but also findstrong evidence of the gambler's fallacy for probabilistic predictions. First, we investigate the simplest form of the gambler's fallacy, finding that participants do expect negative autocorrelation based solely on the outcome of the final draw, which applies to both probabilistic and point predictions. Second,we find that this expectation of negative autocorrelation strengthens as the streak gets longer, and that this also applies to both probabilistic and point predictions.