Sequential Advantage Selection for Optimal Treatment Regimes

Ailin Fan, Wenbin Lu, Rui Song

Abstract

Variable selection for optimal treatment regimes in clinical trials and observational studies is receiving increased attention. Most existing techniques focus on selecting variables important for prediction; however, variables that are poor predictors of overall outcome may still be critical for decision-making. A qualitative interaction occurs when a treatment effect changes direction based on a variable's value, indicating its importance for prescriptive decisions. While the S-score (Gunter et al., 2011) characterizes the magnitude of these interactions individually, it struggles with correlated variables and high-dimensional settings. We propose Sequential Advantage Selection (SAS), a method that selects qualitatively interacting variables sequentially to exclude redundant features and improve model reliability. With robust stopping criteria, SAS handles large covariate sets even with small sample sizes. Simulation results and an application to the STAR*D depression trial demonstrate that SAS provides more comprehensive and effective treatment regimes than traditional methods.

Introduction

In the era of personalized medicine, the goal is to move away from "one-size-fits-all" treatments toward regimes tailored to individual patient characteristics. To achieve this, clinicians must identify "prescriptive" variables—covariates that dictate which treatment will work best for a specific patient.

The primary challenge is that standard variable selection techniques, such as LASSO, are designed for prediction. They prioritize variables with a strong correlation to the final outcome. However, a variable might be a poor predictor of the overall outcome but a critical factor in decision-making. For example, a genetic marker might not indicate the severity of a disease, but it could determine whether a patient responds to a specific drug or suffers a toxic reaction. This phenomenon is known as a qualitative interaction.

Methodology: Sequential Advantage Selection (SAS)

Existing tools like the S-score rank variables individually based on their potential for qualitative interaction. However, these rankings often fail when variables are highly correlated or when the number of potential features ($p$) exceeds the number of patients ($n$). To address these limitations, we developed Sequential Advantage Selection (SAS), a stepwise algorithm that builds a joint model for treatment regimes.

1. The Modified S-Score

The foundation of our method is the S-score, which measures the "value of information." It quantifies how much a patient’s expected outcome improves if a specific variable is used to guide treatment selection compared to choosing the best treatment for the average population.

2. Sequential Selection Process

Unlike individual ranking, SAS employs a sequential approach. The algorithm identifies the single best variable, incorporates it into the model, and then searches for the next variable that provides the highest additional (sequential) advantage. This process accounts for correlations between variables and prevents the selection of redundant information, leading to a more parsimonious and reliable model.

3. Stopping Criterion

To prevent overfitting in high-dimensional datasets, we implemented a threshold-based stopping criterion (e.g., 1%). If the inclusion of a new variable adds less than the specified percentage to the total advantage already explained by the model, the selection process terminates. This allows SAS to remain effective even when $p > n$.

Experimental Evaluation

Simulation Studies

We evaluated SAS using simulated datasets containing 1,000 independent variables and only 200 observations. We tested three distinct models—ranging from simple linear baselines to complex non-linear interactions—across both randomized trials and observational study scenarios. In high-correlation scenarios, SAS consistently achieved a higher True Discovery Rate (TDR) and identified more True Positives (TP) than both the standard S-score and LASSO. Furthermore, treatment regimes derived via SAS resulted in higher expected mean responses and lower error rates.

Real-World Application: The STAR*D Study

We applied SAS to data from the "Sequenced Treatment Alternatives to Relieve Depression" (STAR*D) clinical trial. We focused on a subset of 319 patients randomized to switch to either Bupropion (BUP) or Sertraline (SER) after failing an initial treatment. This dataset included 305 potential covariates, including demographics and intermediate medical measurements.

  • LASSO: Failed to select any variables for treatment decision-making.
  • S-score: Selected 219 variables, a number far too large for practical clinical implementation.
  • SAS: Selected 33 variables, including specific psychiatric screening scores and side-effect inventories. The resulting regime was statistically superior to a non-personalized approach.

Discussion and Future Work

SAS represents a significant advancement in identifying prescriptive variables for personalized medicine. By focusing on the "signal" of qualitative interactions rather than the "noise" of predictive-only variables, SAS provides clinicians with more actionable insights.

While powerful, the current iteration of SAS is focused on binary treatment options and continuous outcomes. Future research will focus on extending the method to:

  1. Multi-arm trials: Accommodating more than two treatment options.
  2. Dynamic Treatment Regimes (DTRs): Adapting the sequential selection for multi-stage decision-making over time.
  3. Diverse Outcome Types: Modifying the algorithm for categorical data or censored time-to-event data, such as those found in oncology trials.

Summary updated

Abstract

Variable selection for optimal treatment regime in a clinical trial or an observational study is getting more attention. Most existing variable selection techniques focused on selecting variables that are important for prediction, therefore some variables that are poor in prediction but are critical for decision-making may be ignored. A qualitative interaction of a variable with treatment arises when treatment effect changes direction as the value of this variable varies. The qualitative interaction indicates the importance of this variable for decision-making. Gunter et al. (2011) proposed S-score which characterizes the magnitude of qualitative interaction of each variable with treatment individually. In this article, we developed a sequential advantage selection method based on the modified S-score. Our method selects qualitatively interacted variables sequentially, and hence excludes marginally important but jointly unimportant variables {or vice versa}. The optimal treatment regime based on variables selected via joint model is more comprehensive and reliable. With the proposed stopping criteria, our method can handle a large amount of covariates even if sample size is small. Simulation results show our method performs well in practical settings. We further applied our method to data from a clinical trial for depression.
BibTeX
@article{fan2016sequential,
  author = {Ailin Fan and Wenbin Lu and Rui Song},
  title = {{Sequential Advantage Selection for Optimal Treatment Regimes}},
  journal = {The Annals of Applied Statistics},
  year = {2016},
  doi = {10.1214/15-aoas849},
}
Figure