Jean Opsomer
Professor
Department of Statistics
Colorado State University

Survey estimation of domain means that respect natural orderings

 In many surveys, estimates are often produced for large numbers of domains. For instance, the U.S. National Compensation Survey estimates mean wages for a large number of job types and levels, by U.S. region and large metropolitan area.  By dividing the overall sample into numerous domains, this can lead to small domain sample sizes, with resulting unreliable survey estimates.  At the same time, it is also common for variables in surveys to follow natural orderings that should be respected in estimates of domain means: in the case of the National Compensation Survey for instance, the mean wages for a job type in a particular region can reasonably be expected to be non-decreasing according to job level.  We propose to use isotonic regression ideas, which can produce domain estimators that respect this natural ordering, while remaining in a classical design-based inferential framework. We also extend the Cone Information Criterion to the survey setting, as a practical approach to detecting possible departures from monotonicity in the data.

Refreshments at 3:45 pm in Snedecor 2101.