A scientific method for sample selection is the probability sample. All individuals should have a non-zero probability of selection and all of them should be known. This facilitates precise and unbiased estimates of population characteristics while quantifying the precision as well by means of confidence intervals or margin of error. Though this is the ideal situation, there are many factors that lead to poor estimates, the most conspicuous one is non-response bias. One major impact of non-response is that the sample size gets reduced. This can be overcome by increasing the initial sample size. The more serious problem is the bias potential on the population estimates. This can happen because of over- or under-representation of certain groups due to non-response. Often these differences in a population are due to ethnic or socioeconomic factors. To avoid this, the amount of non-response should be kept to a minimum. Estimating response probabilities depend on the model that is used. The most common model used is a logit model and this study compares this with the simple linear model. Estimation of response probabilities need individual values of auxiliary variables for both respondents and non-respondents which may not be available. This article explores approaches that do not depend on such preconditions such as using weights that can be used to estimate the response probabilities. These estimated response probabilities can be used for analysis of non-response, correction for non-response and as a measure of representativity indicator. This study also compares the response probability estimation using a logit model that is the benchmark for response probabilities from the linear model, transforming weights into probabilities using regression estimation and transforming weights by raking ratio estimation into estimated response probabilities.
Jelke Bethlehem
展开 >
Leiden University, Institute of Political Science, Albert Verweystraat 21, 2394 TK Hazerswoude-Rijndijk, The Netherlands
NSTL
