Cognitive disability following traumatic experiences of disaster has been documented; however, little is known about heterogeneity in the association across individuals. In this natural experiment study of approximately 3000 Japanese older adults in an area directly affected by the 2011 Great East Japan Earthquake, the baseline survey was established 7 months before the 2011 earthquake. To inductively identify heterogeneity in postdisaster cognitive disability by predisaster characteristics, we applied a machine learning–based causal inference approach—generalized random forest. We identified strong evidence for heterogeneity in the association between home loss and cognitive disability objectively assessed 2.5 and 5.5 years after the 2011 earthquake. The subgroups with the strongest disaster-dementia associations tended to be from low socioeconomic backgrounds and have predisaster health problems. The study demonstrated that some subpopulations are particularly prone to experience cognitive disability after disasters, which could be overlooked in studies assessing population average associations only.