Sampling and Estimation for Business Surveys
In this chapter we give guidelines on many of the practical details of designing a business survey: the register and sampling frame, sample design and various aspects of estimation. The full details of the theory are well covered in other texts (Cochran 1977; Särndal et al. 1992, Lohr 2010a), and are not repeated here. We will almost exclusively consider design-based inference (including the model-assisted approach), in other words, estimation where the sampling probabilities are used explicitly. This is because the size disparity of businesses necessitates differential sampling in most circumstances, and therefore violates one of the conditions for effective model-based inference (i.e., that sampling should be ignorable). This does not mean that model-based inference could not be used in some circumstances, and interested readers can find details of this approach in Valliant et al. (2000), and of Bayesian survey inference in Ghosh and Meeden (1997).
5.1 Basic Principles
A piece of advice to start—a professor I know regularly accosts people with the question “What is the target of inference?”; this could advantageously be paraphrased as “What are you trying to measure?” It is well worth noting the answer to this question clearly before starting on a survey design, so that the target is always in mind; otherwise it is very easy to produce a survey whose purpose is unclear.
In the beginnings of business statistics ...