Dan Hedlin: Assessment of bias induced by cut-off sampling

It is common in surveys to abstain from trying to observe some units even if they would have been of interest. For example, in business surveys it is not usual to take a sample among the smallest businesses. In other surveys remote areas or hard-to-reach individuals may be deliberately excluded. This leads to bias. The term cut-off sampling refers to this type of sampling design, in which some units that do belong to the target population are not sampled. The reasons why cut-off sampling is employed vary; it may not be feasible to include small businesses in samples, or it may be believed that they are so small that the bias is negligible.

The cut-off point in business surveys is often drawn at five or ten employees. Below the cut-off point the data source is usually only administrative data, above it there are survey data on the variable of interest. It is an ‘out-of-sample’ issue; the units we would like to extrapolate to are outside the scope of the sample. So, is this an impossible problem?

It may not come as a surprise that there is little research on bias and estimation issues associated with cut-off sampling, given the nature of the problem.

In my talk, I would like discuss different approaches to the issue of assessing bias due to cut-off sampling and also seek input from the audience.