|
Abstract:
|
For a substantial class of sampling problems, the sample design not only provides the sampling distribution but also imposes a superpopulation model on the sample data. For many of these sampling problems this duality is an academic curiosity but there are some cases where inference based on probabilities of selection would be statistical mal-practice. In such situations, the model provides a vital backup and a Best Linear Unbiased Estimator (BLUE) that avoids Horvitz-Thompson (HT) based methodologies. Since the design imposes the model, model failure is a peripheral issue. An example is presented for mail sampling and other situations that encounter similar problems are mentioned. For the examples presented here, an adequate frame is not available for sample design but is only available after the sample has been selected and the data collected.
|
