Conferences

Due to the lack of an appropriate frame, many surveys in the social sciences apply a cluster sample design. This has some merits (e.g. reduced fieldwork costs) but also considerable drawbacks. One of the most prominent ones is an increase in variance of estimators when, at the estimation stage, the design is ignored and survey data is treated as having arisen from a simple random sample. The factor by which the variance of an estimator is increased as compared to the variance of the same estimator under simple random sampling is the design effect.
Design effects basically arise from two sources: clustering (deff_c) and b) weighting (deff_p). As in most fields of survey statistics, estimators of the design effect have been proposed motivated both from the design-based and the model-based perspective. Estimators derived under either approach have their merits and drawbacks in terms of quality and extensibility. The design-based approach considers both sources of variation directly as it is based on a direct comparison of an estimator’s variance under the complex and under a (hypothetical) simple random sample design. The model-based approach decomposes deff_c and deff_p explicitly and estimators of either component of the design effect have been proposed. Under this approach, the estimator’s quality depends mainly on the quality of an appropriate estimator of rho, the intra class correlation coefficient.
In this presentation the results of a simulation study on the quality of a number of estimators of both approaches is presented. The talk will summarize important findings and gives recommendations that can guide the researcher’s choice of an estimator.