It is true that people with a work disability are not randomly sampled.  The CPS in general, however, is based on a large sample size that is statistically stratified (over 50,000 households and 100,000 people).  The overall sample size is large enough that data from the subpopulation is reliable.  If this were a problem, we would expect to see greater variability in employment rates from year to year.  The stability of these rates over time lends credence to the value of the work disability sample.

The criticism contradicts the frequent use that many vocational and economic experts make of the CPS data.  The sampling and weighting procedures in the CPS are based primarily on geographic area, household type, unemployment ratios, age, gender, race, and Hispanic origin.  Education is not a part of this procedure.  Despite this, many forensic experts value the CPS data and use average earnings and employment by education level in order to estimate lifetime earning capacity.  Based on the Skoog and Toppino (2002) criticism, however, this is wrong.  This criticism, then, appears to be carrying the sampling need to extremes with the sole purpose of denigrating the CPS disability data.

Skoog and Toppino (2002) state that disability surveys are only useful if they are extremely specific in their sampling methods.  They note:

For their Tables to have any validity, their statistics must be based on sensible (read statistically consistent at a minimum, if not statistically unbiased or efficient) structural parameter estimates. They must answer the question if one has this education, sex, age, and impairment/condition, what is the conditional probability such a person will be employed? (Emphasis in original; p. 85)

This is an impossible solution to the issue.  First, the sample size would have to be enormous in order to accommodate this desire – probably a sample size much larger than even the 5% sample of the Decennial Census!  Even if this could be accomplished to any degree, it would be impossible to anticipate every combination of age, gender, education level, and impairment that might arise in a forensic setting.