IT FOCUSES ON PROBLEMS OF SELECTION IN RESEARCH DESIGNS WHERE ASSIGNMENT TO EXPERIMENTAL CONDITIONS OCCURS ON A NONRANDOM BASIS. FIVE DIFFERENT SOLUTIONS TO THE PROBLEMS OF SELECTION ARE DISCUSSED - RANDOMIZATION, CONVARIANCE ADJUSTMENT, GAIN SCORES, MATCHING, AND EXPLICIT SELECTION - AND THE CONDITIONS UNDER WHICH THESE SOLUTIONS ARE USEFUL ARE EXAMINED. IT IS CONCLUDED THAT WHEREVER POSSIBLE, RESEARCHERS SHOULD EMPLOY RANDOMIZATION. WHEN THIS APPROACH IS UNAVAILABLE, AND THE RESEARCHER HAS CONTROL OVER THE ALLOCATION OF OBSERVATIONS TO EXPERIMENTAL CONDITIONS, THE ASSIGNMENT OF OBSERVATIONS TO TREATMENT AND CONTROL CONDITIONS AS AN EXACT FUNCTION OF THEIR OBSERVED SCORES ON A SELECTION VARIABLE IS SUGGESTED. WHEN THE ALLOCATION OF OBSERVATIONS TO EXPERIMENTAL CONDITIONS IS NOT UNDER THE CONTROL OF THE RESEARCHER, AND WHERE THE ASSUMPTION OF LINEARITY IS APPROPRIATE, LINEAR STATISTICAL ADJUSTMENT IS JUDGED THE BEST ANALYTIC APPROACH. REFERENCES ARE INCLUDED. (AUTHOR ABSTRACT MODIFIED)