Analyzing phase III studies in hospice/palliative care. a solution that sits between intention-to-treat and per protocol analyses: the palliative-modified ITT analysis.

Intention-to-treat (ITT) analyses are the standard way to evaluate randomized controlled trials (RCTs) to minimize Type I errors related to differential rates of noncompletion from one study arm. People in palliative care often die sooner than predicted as a direct result of disease progression, some of whom will be participating in RCTs and who will, therefore, withdraw or die after randomization for reasons unrelated to the intervention. This proportion of withdrawals is statistically negligible in other clinical disciplines, but commonplace in hospice/palliative care, creating a systematic bias away from the true effect. ITT analyses in hospice/palliative care that deem all withdrawals to be treatment failures or that impute data from deteriorating participants systematically underestimate the benefits of interventions, reducing the power of these studies. Equally unacceptable would be a per protocol analysis that excludes all withdrawals after randomization as this will underestimate toxicity. A modified analytic approach is needed on a continuum between ITT and per protocol analyses. To address data after randomization where there is a high rate of withdrawals because of death or deterioration, criteria need to include being: 1) prespecified in the original protocol; 2) clinically absolutely the result of disease progression; 3) identified by the blinded Independent Data Monitoring Committee as being unrelated to the intervention(s); and 4) accounted for in the study's CONSORT diagram. Such data should not be included in the analysis of the primary outcome. This article aims to define a better way of balancing Type I and Type II errors in hospice/palliative care RCT analyses using the palliative-modified ITT analysis. Arguably, the palliative-modified ITT analysis should be the primary evaluation of hospice/palliative care Phase III studies but, as a minimum, should routinely be the key sensitivity analysis.

Duke Scholars

Author Gregory P. Samsa Biostatistics & Bioinformatics, Division of Biostatistics