ObjectiveTo examine the ability of various models to prospectively identify a small group of individuals with predictable high future costs that may be mitigated through disease management.Data SourcesDiagnoses and medical costs for over a million members of employer-sponsored benefit plans from the Medstat MarketScan®Research Database (1997–1999).Study DesignA prior cost model, a diagnosis-based (diagnostic cost group [DCG]) model and a diagnosis + prior cost (combo) model were each calibrated on 1997–1998 data and applied to 1998 data to identify 0.5%-sized ‘top groups’ of individuals most likely to be expensive in 1999 (validation). An individual with a year 2 cost over $US10 000 was considered to be a ‘good pick’. The percentage of good picks, other features of the cost distribution, and the prevalence of ‘manageable’ i.e. commonly managed disease (diabetes mellitus, congestive heart failure, asthma/chronic obstructive pulmonary disease and depression) were compared in the three top groups. The performance of nine additional top groups – one for each model type fitted to costs top coded at each of $US100 000, $US50 000 and $US25 000 – was also investigated.ResultsIndividual R2values for the (full-range) prior cost, DCG, and combo models were 11, 16, and 21%, respectively; R2values increased to 27% and 31% for DCG and combo models top coded at $US25 000. The full-range model top groups contained 54, 62, and 70% good picks, and 42, 53, and 48% of their cases, respectively, had at least one manageable disease. Top coding the prior cost model did not produce better top groups. However, the top groups from the DCG and combo models top coded at $US25 000 contained the most good picks (65 and 76%, respectively) and commonly managed chronic conditions (61 and 47%, respectively).ConclusionThe DCG and combo models were better than the prior cost model for identifying groups rich in individuals who will be expensive the following year. Surprisingly, top groups based on top-coded models dominated their full-range model analogs, identifying more good picks and more people with manageable disease.