What can we do about exploratory analyses in clinical trials?

The research community has alternatively embraced then repudiated exploratory analyses since the inception of clinical trials in the middle of the twentieth century. After a series of important but ultimately unreproducible findings, these non-prospectively declared evaluations were relegated to hypothesis generating. Since the majority of evaluations conducted in clinical trials with their rich data sets are exploratory, the absence of their persuasive power adds to the inefficiency of clinical trial analyses in an atmosphere of fiscal frugality. However, the principle argument against exploratory analyses is not based in statistical theory, but pragmatism and observation. The absence of any theoretical treatment of exploratory analyses postpones the day when their statistical weaknesses might be repaired. Here, we introduce examination of the characteristics of exploratory analyses from a probabilistic and statistical framework. Setting the obvious logistical concerns aside (i.e., the absence of planning produces poor precision), exploratory analyses do not appear to suffer from estimation theory weaknesses. The problem appears to be a difficulty in what is actually reported as the p-value. The use of Bayes Theorem provides p-values that are more in line with confirmatory analyses. This development may inaugurate a body of work that would lead to the readmission of exploratory analyses to a position of persuasive power in clinical trials.