There are two main reasons for using weights in sample surveys. The first is the need to account for unequal inclusion probabilities which may result from various processes such as disproportional selection of units within strata. In telephone surveys, inclusion probabilities vary for example because of differential amounts of telephone numbers people have or differential numbers of persons in the household belonging to the target population. Therefore, design weights have to be calculated and applied during data analysis. In our presentation we first demonstrate how design weighting was done in the study CELLA2. This is a parallel survey of landline and mobile phones conducted in Germany in the summer of 2010 with about 1.500 respondents each.
In addition to, the efforts involved in finding suitable frames without over- or under-coverage and in drawing gross samples accurately, there is a second reason why the data might need to be weighted. If nonresponse to a survey is systematic, the estimates will be biased. We describe how (design weighted) distributions of socio-demographic variables in CELLA2 differ from the corresponding distributions in the Microcensus, a 1 % sample of about 800,000 individuals which provides official representative statistics for the population of Germany. We explain how design weighting and an additional adjustment weighting can be combined in the GREG-estimator. It will be demonstrated that weighted estimates of socio-demographic variables are frequently closer to the Microcensus distributions than unweighted estimates. Thus, most of the important biases in telephone samples can be corrected if adequate weighting is applied.