Examines whether substituting a health maintenance organization (HMO) for traditional fee-for-service (FFS) Medicaid insurance reduces the cost of children's health care. In nonrandomized settings, unobserved selection could bias estimates of HMO performance. To control for selection, researchers often rely on parametric assumptions or use instrumental-variables methods. We pursued a different approach based on semiparametric maximum-likelihood techniques, which Monte Carlo studies have shown to be quite robust in many contexts. We applied this model to data from a self-selected sample of children in either a Medicaid HMO or traditional FFS in Florida. After controlling for selection, we estimated that the HMO reduced expenditures by 9.1%. A simple model assuming no selection predicts no savings. We validated our estimates by comparing our results with those obtained from a randomized sample of HMO and FFS enrollees. These comparisons indicate that the HMO reduces expenditures by 13.6%. We conclude that selection can substantially bias estimates of HMO impact and that our approach .provides a potentially useful method for accounting for this bias.