Finding Respondents in the Forest: A Comparison of Logistic Regression and Random Forest Models for Response Propensity Weighting and Stratification

摘要

Survey response rates for modern surveys using many different modes are trending downward leaving the potential for nonresponse biasesin estimates derived from using only the respondents. The reasons for nonresponse may be complex functions of known auxiliary variables orunknown latent variables not measured by practitioners. The degree to which the propensity to respond is associated with survey outcomescasts light on the overall potential for nonresponse biases for estimates of means and totals. The most common method for nonresponseadjustments to compensate for the potential bias in estimates has been logistic and probit regression models. However, for more complexnonresponse mechanisms that may be nonlinear or involve many interaction effects, these methods may fail to converge and thus fail togenerate nonresponse adjustments for the sampling weights. In this paper we compare these traditional techniques to a relatively new datamining technique- random forests – under a simple and complex nonresponse propensity population model using both direct and propensitystratification nonresponse adjustments. Random forests appear to offer marginal improvements for the complex response model over logisticregression in direct propensity adjustment, but have some surprising results for propensity stratification across both response models.