The letter from Drs. Gozalo, Mor, and Miller appears to raise 2 issues. One concerns potential bias in our hospice effect estimation related to controlling for consistently low service use for 24 months before death. We did test for the effect of this variable and stated that the results persisted among all sensitivity analyses, including "models that do not correct for consistently low Medicare use."

The other issue turns on the effectiveness of propensity scores to control for selection bias, particularly in study samples, such as ours, in which nursing home residents are combined with community dwellers. Our analyses, including calculation of propensity score, were stratified by age group and condition cohort. Within each of these relatively homogeneous strata, models estimating hospice effects controlled not only for propensity score and for nursing home residence but also for an array of measures previously shown to correlate with hospice use: duration of illness, disease burden, gender, race, Medicaid enrollment, and urban or rural influences. Our discussion of limitations acknowledged that even if poststratification is used to match decedents by age and condition and propensity scores are used to control for other hospice selection effects, some selection bias will inevitably remain in an observational study design.

We sought methods that improved upon those applied in previous research concerning hospice effects on Medicare expenditures. We welcome future studies that offer additional methodologic improvements. Our study mainly shows that the quality and costs of various strategies to serve patients coming to the end of life deserve careful study.