Regency EF
Comparing Methods for Regression Models in Which the Dependent Variable Is Based on Estimates (304090)
*Yi Mu, Centers for Disease Control and Prevention
Keywords: Estimated dependent variable regression; Bayesian; weighted regression;
Researchers often use sampled data to estimate unknown quantities of larger population, it is not uncommon to regress these estimated values against one or more independent variables to generate the ultimate coefficients of interest. In this study, we used simulated data to compare three methods used for estimated dependent variable regression (EDV) for count data: 1) to regress on estimated quantities which would completely ignore the estimation errors; 2) to regress on sampled data with estimated weights, which is calculated by dividing estimated population quantities by the sampled as if the quantities had been observed; 3) to regress on estimated quantities by treating them as Poisson distribution to account for estimation errors . The results showed that approach 1) performed the best, with parameter estimates and standard errors were almost the same as the gold standard; approach 2) provided slightly better parameter estimates than approach 3), but approach 3) produced slightly smaller standard errors than approach 2); both approach 2) and 3) came with larger standard errors than the gold standard. We concluded EDV could be performed directly on estimated quantities.