We propose to use maximum likelihood econometric methods in survival analysis in large, secondary data sources to measure the effect of various characteristics on the lengths of hospital stays in psychiatric hospitals and the lengths of the subsequent community stays. These methods significantly improve upon methods presently being used in the psychiatric literature in that they explicitly allow for and measure the effect of duration dependence and unobserved heterogeneity. We also will demonstrate how to treat missing variables as a special type of unobserved heterogeneity that can be controlled for using these econometric techniques. We propose to use and improve these methods on two large data sets: a) the Virginia DMHMRSAS administrative records and b) the 1991 and 1986 MEDPAR data set. The large number of observations in each data set allows for rich specifications of the survival functions for both hospital stays and community stays. We also propose to develop new, semiparametric survival function estimators that a) allow for duration dependence to interact with observed covariates and b) do not rely upon somewhat arbitrary functional form assumptions typically made in the survival analysis literature. We will encourage use of both maximum likelihood techniques and our new techniques by providing copies of user-friendly FORTAN programs at minimal cost. Finally, we will show how the extra information derived from using these methods can be fruitfully used to inform policy decisions. We will measure the proportion of variation in hospital stay lengths and community tenure lengths that can be explained by the data a) using standard methods used in the psychiatric literature, b) using maximum likelihood techniques, and c) using our new techniques. Also, we demonstrate how these estimators can identify hospital or community characteristics that affect hospital stay and community stay lengths and psychiatric services costs. We also develop a comprehensive model of medical usage and community stays whose implementation is delayed for a future project.
| Garren, S T; Smith, R L; Piegorsch, W W (2000) On a likelihood-based goodness-of-fit test of the beta-binomial model. Biometrics 56:947-50 |
