Abstract: Agents in a wide range of settings exhibit variation in decisions and outcomes. We note that this variation is jointly determined by skills and preferences, and we develop a simple framework that uses the joint distribution of provider-specific decisions and outcomes to decompose variation into skills and preferences. We apply this framework in the setting of pneumonia diagnosis by radiologists. Radiologists vary in both their diagnostic rates (decision) and their false omission rates (outcomes), and radiologists with higher diagnostic rates have higher false omission rates. We rationalize these patterns with a model of diagnosis, in which radiologists with heterogenous diagnostic skill endogenously choose thresholds to minimize some function of false negatives and false positives. Radiologists wish to avoid false negatives more than false positives, and this imbalance increases with lower diagnostic skill. Variation in skills can explain 60% of variation in diagnostic decisions, and policies to improve skills perform better than those to reduce variation.