Biased estimation in policy research: an illustrative example of ridge regression in a health system model

The paper develops an argument for the necessity of examining individual coefficients in policy models. As a result of this need, it is posited that something other than OLS estimators should be used since they are inflated and have extremely large variances when multicollinearity is present. Further, it is argued that policy models are by definition theoretically nonorthogonal. Ridge regression as one of a class of biased estimators is offered as one possible approach to dealing with the nonorthogonality problem in policy research. The logic of the approach is articulated and an empirical model of a health system is estimated with ordinary least squares and ridge estimators. The models are compared and implications discussed.