|
Abstract:
|
Dealing with nonresponse is a necessity in survey data analysis. Weighting is a nearly perfect technique because, in particular, it allows a consistent estimation of the distribution function. Imputation is practically mandatory in the case of partial nonresponse. Deterministic imputations are invalid for estimating nonlinear transformations of the variable (the distribution function). One often add a random term to the predictor with , which adds a parasite variance to all estimations. Using the estimating equations of the prediction model and some techniques of balanced sampling, this extra-variance can be considerably reduced. In the case of donor imputation, coherency between imputation and weighting can be achieved. For discrete variables, balance is obtained with the cube method. For numerical variables, new ideas based on elementary properties of martingales are necessary.
|
