Abstract Objective: The standard method of generating disorder‐specific disability scores has lay raters make rankings between pairs of disorders based on brief disorder vignettes. This method introduces bias due to differential rater knowledge of disorders and inability to disentangle the disability due to disorders from the disability due to comorbidities. Methods: We propose an alternative, data‐driven, method of generating disorder‐ specific disability scores that assesses disorders in a sample of individuals either from population medical registry data or population survey self‐reports and uses Generalized Random Forests (GRF) to predict global (rather than disorder‐specific) disability assessed by clinician ratings or by survey respondent self‐reports. This method also provides a principled basis for studying patterns and predictors of heterogeneity in disorder‐specific disability. We illustrate this method by analyzing data for 16 disorders assessed in the World Mental Health Surveys (n = 53,645). Results: Adjustments for comorbidity decreased estimates of disorder‐specific disability substantially. Estimates were generally somewhat higher with GRF than conventional multivariable regression models. Heterogeneity was nonsignificant. Conclusions: The results show clearly that the proposed approach is practical, and that adjustment is needed for comorbidities to obtain accurate estimates of disorder‐specific disability. Expansion to a wider range of disorders would likely find more evidence for heterogeneity