Properties of treatment effect estimates in switching regime models: impacts of sample size and strength and dirtiness of instruments.
- Presenter:
Chair: Susan Ettner; Discussant: Michael French Mon June 5, 2006 10:45-12:15 Room 226
Instrumental variables methods are widely used to adjust for unobserved variables that are correlated with both treatment selection and patient outcomes. However, there is a growing literature showing that such models often perform more poorly than ordinary least squares regression. The bias and precision of instrumental variables estimators have been shown to be sensitive to comparatively weak correlations of the instrumental variable with the residuals of the outcome equation. The bias and precision of instrumental variables estimators have also been found to be sensitive to the strength of the instruments. Larger sample sizes cannot overcome problems of bias but are helpful in improving precision.
To date, the examination of the properties of IV estimates has been confined to the single equation case. Single equation estimates of treatment effects do not account for the interactions of treatment cohorts with observed covariates. Nor do they account for differences in the distributions of patient characteristics between treatment cohorts. Switching regime models offer the potential to control for both of these factors-potentially enabling a more precise estimate of treatment effects. However, we know little about the sensitivity of treatment effects estimates obtained from switching regime models to issues such as the “dirtiness” of the IV, strength of the IV, or sample size. In this paper, we conduct simulation analysis to examine these properties for IV estimation of switching regime models when these factors are varied.
We present the results of a simulation study to demonstrate estimation error caused from correlation between the residual and the instrumental variable when IV estimation of switching regime models are employed to estimate of regression parameters in presence of an endogenous regressor.