The bias of the estimation of income distribution parameters in non-probability surveys: an experiment based on EU-SILC microdata for Poland
The bias of the estimation of income distribution parameters in non-probability surveys: an experiment based on EU-SILC microdata for Poland

Summary/Abstract: The selection bias is a major issue when using non-probability samples for inference about finite populations. Minimizing it requires sufficient auxiliary power from external sources and proper modelling. This article examines the potential bias of estimates of income distribution parameters using real-life data from a household survey (European Union Statistics on Income and Living Conditions, EU-SILC) and computer simulations. It investigates the change in the performance of the estimates depending on data defectiveness, the sample size, the pool of auxiliary variables, sources of auxiliary information, and the estimation method. The non-probability samples had non-ignorable bias intentionally introduced by designing selection mechanisms directly tied to income or the access to a computer and the Internet. The results indicated that the bias of the estimates behaved predictably for simple parameters, such as the median and mean. However, for more complex parameters, the bias patterns were less stable. The use of sociodemographic variables for weighting did not consistently reduce the bias; on the contrary, in certain cases, it led to the aggravation of the bias. The number of auxiliary variables and the sample size were as important for the bias reduction as the actual estimation methods, of which the doubly robust estimation performed best.

• Details
• Contents