ORGANIZATIONAL BEHAVIORAND HUMAN PERFORMANCE27, 411-422 (1981) General Conditions for the Success of Bootstrapping Models COLIN CAMERER Center for Decision Research, Graduate School of Business, University of Chicago Linear models which fit regression equations to clinical judgments, then use the fitted parts of judgments as "bootstrapped" judgments, have outperformed clinical judgments in many tasks. Empirically, the phenomenon has been per- vasive, but general conditions for the success of bootstrapping models have never been explicitly linked to cross-study data. This link, combined with psychologically plausible evidence about the relationships between judgmental variables, shows that bootstrapping will improve judgments slightly under al- most any realistic task conditions..This result allows one to apply bootstrap- ping blindly in cases where criterion information is missing or vague (precisely the cases where bootstrapping models are useful), and be confident that pre- dictions are being improved. A simple comparison of bootstrapping models with equal weighting models is also made, but general conditions for relative success of those two models are not specified. For many years psychologists fruitfully studied clinical judgment by regressing judgments against cues used in judgment, and studying the derived cue weights along with measures of correlation with outcome variables. Then it was discovered that the fitted parts of those regressions--i.e., the clinical judgments less any regression residual-- often correlated more highly with the outcome variable being predicted, than the judgments themselves did. The regression models, called bootstrapping models, pulled judges up by their proverbial bootstraps (see Dawes, 1971; Goldberg, 1970). Since the residuals from such a regres- sion equation generally represent random variance in clinical judgments, bootstrapping models work because they eliminate a source of judgmental variance that doesn't provide information about outcomes. When residuals from the judge's linear regression model are correlated with outcomes (which might be true if the judge senses environmental nonlinearities or uses valid cues not included in the linear model), things are more complex. In those cases, when will bootstrapping work? If envi- ronments are perfectly predictable, and all the relevant cues properly specified, then bootstrapping will always work. If judges are perfectly Thanks to Michael Doherty, Robyn Dawes, Lewis Goldberg, Coleman Kendall, Robin Hogarth, members of the Center for Decision Research workshop, and especially Hillel Einhorn for comments and ideas. Support from the Graduate School of Business is gratefully acknowledged. Requests for reprints should be sent to: Dr. Colin Camerer, Graduate School of Business, University of Chicago, Chicago, IL 60637. 411 0030-5073/81/030411 - 12502.00/0 Copyright(~) 1981by Academic Press,Inc. All rightsof reproductionin any form reserved,