Estimating the indirect costs of teaching

Medicare's Prospective Payment System pays teaching hospitals using a regression coefficient in a log-log cost function. Previous literature showed that this coefficient is sensitive to the covariates included in the function, but specified teaching intensity as the logarithm of one plus the intern and resident-to-bed ratio. Provided the true relationship is log-log, adding one biases the coefficient substantially but not predicted cost. In a re-specified equation that makes this bias negligible, the coefficient is not nearly as sensitive to the inclusion of other covariates. Because further issues remain to be explored, it is premature to use our results for policy purposes.