Baseline corrections in experimental and quasi-experimental clinical trials

Abstract

Three methods of correcting for baseline differences are evaluated for use with randomized experimental designs and nonrandomized quasi-experimental designs. Cases where baseline means differ significantly in spite of random assignment and cases where observed baseline means do not differ significantly in quasi-experimental designs are given special attention. For comparison of treatment-induced change in randomly constituted treatment groups, analysis of covariance (ANCOVA) provides appropriate type I error protection and superior power regardless of the apparent significance of chance baseline differences. Correcting for baseline differences by expressing outcome scores as percentage of the baseline value for each individual produces results that closely approximate ANCOVA under specified conditions. Tests on simple pre-post difference scores are nonconservative when baseline means differ significantly in spite of randomization. In quasi-experimental research where treatment groups represent samples from different predefined populations, ANCOVA and percentage change provide inadequate baseline corrections even when the observed baseline means do not differ significantly. Although no method of baseline correction is entirely satisfactory in the absence of random assignment, tests on simple pre-post difference scores are generally superior for quasi-experimental designs in which treatment groups represent samples from different populations.