Relevance of correlation between treatment responses in clinical trials.

Background: We recently demonstrated that in self-controlled clinical trials using t-statistic the correlation level between treatment responses is an important determinant of the sensitivity of testing.

Methods: The current study uses hypothetical examples and real data to study whether this phenomenon also affects other parametric statistical procedures.

Results: With negative correlations between treatment responses, not only paired t-test, but also repeated measures analysis of variance (ANOVA), Bonferroni adjustments for ANOVA, and therapeutic equivalence testing, lack sensitivity to demonstrate significant results.

Conclusions: Negative correlations are increasingly encountered in clinical pharmacological research, and standard parametric statistical methods are increasingly inefficient to compare between treatment responses with decreasing correlation. Proposals for handling this problem include: increase the sample sizes in the trial; replace intraindividual comparison by parallel-group assessment; find the reason for the negative correlation, and reanalyze the data, for example by subgroup analysis; if the distribution of paired differences is skewed, use a non-parametric test.