Inference on Categorical Survey Response: A Predictive Approach

We consider the estimation of finite population proportions of categorical survey responses obtained by probability sampling. The customary design-based estimator does not make use of the auxiliary data available for all the population units at the estimation stage. We adopt a model-based predictive approach to incorporate this information and make the estimates more efficient. In the first part of our paper we consider a multinomial logit type model when logit function is a known parametric function of the covariates. We then use it for the prediction of non-sampled responses. This together with sampled responses is used to obtain the estimates of the proportions. The asymptotic biases and variances of these estimators are obtained. The main drawback of this approach is, being a parametric model it may suffer from model misspecification and thus, may lose it’s efficiencies over the usual design-based estimates. To overcome this drawback, in the next part of this paper we replace the multinomial logit type model by a nonparametric model using recently developed random coefficients splines models. Finally, we carry out a simulation study. It shows that the nonparametric approach may lead to an appreciable improvement over both parametric and design-based approaches when the regression function is quite different from multinomial logit.