Accounting for non-ignorable sampling and non-response in statistical matching

Data for statistical analysis is often available from different samples, with each sample containing measurements on only some of the variables of interest. Statistical matching attempts to generate a fused database containing matched measurements on all the target variables. In this article, we consider the use of statistical matching when the samples are drawn by informative sampling designs and are subject to not missing at random non-response. The problem with ignoring the sampling process and non-response is that the distribution of the data observed for the responding units can be very different from the distribution holding for the population data, which may distort the inference process and result in a matched database that misrepresents the joint distribution in the population. Our proposed methodology employs the empirical likelihood approach and is shown to perform well in a simulation experiment and when applied to real sample data.