Inference on Selected Subgroups in Clinical Trials

This note is based on the X. Guo and X. He (2020), Inference on Selected Subgroups in Clinical Trials, Journal of American Statistical Association

Subgroup Analyses and Selection bias Overview

Subgroup analyses figure out promising treatments to some subgroups. In practice, subgroup analyses consists of two steps: subgroup identification and subgroup confirmation. In the identification step, researchers looks for the best selected subgroup which comes from domain considerations or data-driven. The existing data-driven identification methods are mostly machine-learning based or model based. The confirmation step provides an inference for possibly additional trials on the promising subgroup. Generally, The inference to the selected subgroup may lead to an overly optimistic evaluation, which is called subgroup selection bias. This kind of bias is originated from the selection mechanism of maxi=1,⋯,k(βi)\max\limits_{i=1,\cdots,k}(\beta_i). Some of the existing methods for subgroup confirmation might be conservative, without theoretical justification, or model-dependent. Thus, this paper proposes a (data-adaptive) de-biasing bootstrap estimate and a valid one-sided confidence bound.

Procedure

For predifined subgroups, the interested estimators are the βs\beta_s (the true effect size of the subgroup with the largest estimated effect) and βmax\beta_{\max} (the largest true subgroup effect size). The procedure is:

  1. Obtain the largest modified bootstrapped subgroup effect βmax⁡,modified*=maxi=1,⋯,k(βi*+di)\beta^*_{\max,\text{modified}}=\max\limits_{i=1,\cdots,k}(\beta^*_i+d_i) with modification di=(1−nr−0.5)(β̂max−β̂i)d_i=(1-n^{r-0.5})(\hat\beta_{\max}-\hat\beta_i)
  2. Construct the bootstrap confidence interval β̂max±cα/n\hat\beta_{\max}\pm c_\alpha/\sqrt{n} with cα=quantile(n(maxi=1,⋯,k(βi,b*+di)−β̂max),1−α)c_{\alpha}=quantile(\sqrt{n}(\max\limits_{i=1,\cdots,k}(\beta^*_{i,b}+d_i)-\hat\beta_{\max}),1-\alpha) (asymptotically sharp interval for βs\beta_s).
  3. Derive the bias-reduced estimator β̂max⁡,reduced=β̂max−E*[βmax⁡,modified*−β̂max]\hat\beta_{\max,\text{reduced}}=\hat\beta_{\max}-E^*[\beta^*_{\max,\text{modified}}-\hat\beta_{\max}]
  4. For data-adaptive rr, tune rr with cross validation and bias-reduced estimator.

While considering the post-hoc subgroup identification, the procedure is similar to the predefined subgroup identification but no data-adaptive selection for rr.

Possible Extention

This method is useful for homogeneous subgroups, because the subgroup selection bias in the homogeneous subgroups might be larger than one in the heterogeneous subgroups. When the subgroup effects are heterogeneous, the naive approach of subgroup identification provide an efficient enough result with less empirical bias. However, the tradeoff among the tuning parameter rr, heterogeneity, and number of subgroups has not been discussed.

With the simulation provided in this paper, the bias and coverage get worse when the number of subgroups or rr gets large. When rr gets large, it means that the effect modification dd goes to 0. We could also discover that the empirical bias with small rr performs better than large rr in homogeneous subgroups setting. However, this situation is reversed in heterogeneous setting. Thus, the relationship between modification, heterogeneity, and number of subgroups might be a potential issue to work on.

Another possible extension associates with the confounding adjustment, especially for trial data and observational studies. This method allows overlapped subgroup; that is, we cannot think the subgroup as the stratification. With confounding adjustment, the subgroup effect size should be reconsidered.

Comments