STATISTICAL HAZARDS IN THE DETERMINATION OF ADVERSE IMPACT WITH SMALL SAMPLES

In 1975, a small but prestigious fellowship program was charged with implementing a selection process that discriminated against women. One set of relevant data described the selection rates for the three most recent years of the program, while another set consisted of rankings based on written qualifications and on performance during a weekend interview, available only for 1975. The numbers of candidates involved were so small that adequate statistical power could not be assured; results were equivocal. To increase power, the probabilities for the three years considered separately can be combined. Disagreement between the expert witnesses for the plaintiff and the defendant were based on differing assumptions about appropriate statistical procedures.