Enrichment designs using placebo nonresponders

Abstract

Enrichment designs that select placebo nonresponders have gained much attention during the last years in areas with high placebo response rates, eg, in depression. Proposals were made that re-randomize patients who did not respond to placebo during a first study phase as the sequential parallel design (SPD). This design uses in a second phase an enriched patient population where the treatment effect is expected to be more pronounced. This may be problematic if an effect in the overall population is claimed. Proposals were made to combine the treatment effects in the overall population from study phase 1 and the enriched population from study phase 2, alleviating but not solving the issue of a potential selection bias. This paper shows how this bias corresponding to the effect difference between the overall population and the enriched population depends on the variability of a potential subject-by-treatment interaction. Sample sizes are given, which lead to a significant result in the combining test with a given probability if actually the average effect in the overall population is zero. If, on the other hand, no subject-by-treatment interaction is given, the enrichment is shown to be inefficient. We conclude that enrichment designs using placebo nonresponders are not able to claim a positive average effect in the overall population if a subject-by-treatment interaction cannot be excluded. It cannot be used to demonstrate positive efficacy in the overall population in a pivotal phase III trial but may be used in early phases to demonstrate varying treatment effects between patients.

Usefulness of the placebo lead-in design for clinical trials with binary outcomes

“…It can be resolved if the correlation is the same for both the placebo–drug and placebo–placebo groups. However, as pointed out by Benda and Haenisch 21 in the context of the sequential parallel comparison design, in turn, enriching for placebo nonresponders is inefficient for a trial power. As shown in our numerical investigation results, it also holds good for the PLD.…”

A Test for Treatment Effects Based on the Exact Distribution of an Ordinary Least-Square Estimator in Sequential Parallel Comparison Design

“…Cui, Ogbagaber, and Hung (2019) improved the OLS estimator by considering multiple correlation structures between Stages 1 and 2. Similar to the suggestion by Chen et al (2011), Benda and Haenisch (2020) suggested a bias in the OLS estimator and proposed a new estimator constructed by introducing an interaction between the treatment effects of Stages 1 and 2. Rybin et al (2018) proposed a novel treatment effect estimation method that uses an expectation-maximization (EM) algorithm wherein the outcomes corresponding to placebo responders are regarded as missing variables.…”

“…In the SPCD framework, it is often assumed that δ 1 = δ 2 because the estimand attributes of δ 1 and δ 2 can be assumed as equivalent. However, Doros et al (2013), Chi et al (2016), and Benda and Haenisch (2020) suggested that δ 2 generally differs from δ 3 . Therefore, the hypothesis of the statistical test should be carefully defined considering the definition of the treatment effects.…”