This study used 1992 and 1993 data from private employers to compare t
he performance of various risk adjustment methods in predicting the me
ntal health and substance abuse expenditures of a nonelderly insured p
opulation. The methods considered included a basic demographic model,
Ambulatory Care Groups, modified Ambulatory Diagnostic Groups and Hier
archical Coexisting Conditions (a modification of Diagnostic Cost Grou
ps), as well as a model developed in this paper to tailor risk adjustm
ent to the unique characteristics of psychiatric disorders (the ''como
rbidity'' model). Our primary concern was the amount of unexplained sy
stematic risk and its relationship to the likelihood of a health plan
experiencing extraordinary profits or losses stemming from enrollee se
lection. We used a two-part model to estimate mental health and substa
nce abuse spending. We examined the R-2 and mean absolute prediction e
rr or associated with each risk adjustment system. We also examined th
e profits and losses that would be incurred by the health plans sewing
two of the employers in our database, based on the naturally occurrin
g selection of enrollees into these plans. The modified Ambulatory Dia
gnostic Groups and comorbidity model performed somewhat better than th
e others, but none of the models achieved R2 values above .10. Further
more, simulations based on actual plan choices suggested that none of
the risk adjustment methods reallocated payments across plans sufficie
ntly to compensate for systematic selection.