Choosing models for health care cost analyses: issues of nonlinearity and endogeneity

Objective: To compare methods of analyzing endogenous treatment effect models for nonlinear outcomes and illustrate the impact of model specification on estimates of treatment effects such as health care costs.

Data sources: Secondary data on cost and utilization for inpatients hospitalized in five Veterans Affairs acute care facilities in 2005-2006.

Study design: We compare results from analyses with full information maximum simulated likelihood (FIMSL); control function (CF) approaches employing different types and functional forms for the residuals, including the special case of two-stage residual inclusion; and two-stage least squares (2SLS). As an example, we examine the effect of an inpatient palliative care (PC) consultation on direct costs of care per day.

Data collection/extraction methods: We analyzed data for 3,389 inpatients with one or more life-limiting diseases.

Principal findings: The distribution of average treatment effects on the treated and local average treatment effects of a PC consultation depended on model specification. CF and FIMSL estimates were more similar to each other than to 2SLS estimates. CF estimates were sensitive to choice and functional form of residual.

Conclusions: When modeling cost or other nonlinear data with endogeneity, one should be aware of the impact of model specification and treatment effect choice on results.