Selection Biases in Sociological Data

The potential for selection artifacts is enormous in sociological research. The selection problem and all of its solutions rest fundamentally on a researcher's ability to model properly both the substantive process and the selection process in the original population. Given a correct specification, the importance of explicit selection is determined relatively easily. Goldberger's work can be used to obtain a precise characterization of one's problems when multivariate normality among the regressors is a reasonable assumption. When multivariate normality cannot be asserted with confidence, Goldberger's results may still be sufficiently accurate for a first attack at the problem. It is difficult to know how widespread explicit selection problems actually are in the sociological literature; however, incidental selection clearly constitutes the most widespread threat to sociological research, but its actual significance is difficult to determine. Much rests on the size of the correlation between error processes in the substantive and selection equations. The more difficult problem of incidental selection is coupled with weaker diagnostic procedures and more sensitive statistical corrections. Finally, there is the issue of how research designs can be improved to minimize selection artifacts in advance. Perhaps most important, probability sampling whenever possible will surely improve matters as long as it is fully understood that because of the infinite regress problem, there will always be some original population from which a nonrandomly selected population has been culled. Forty-eight references are provided.