📊 statistics

Mitigating the Winner's Curse While Controlling Multiplicity: e-Process Methods for Anytime-Valid Inference in Dose-Ranging Trials

This paper introduces an anytime-valid e-process method that mitigates the winner's curse and controls multiplicity in dose-ranging trials by subtracting a predictable selection charge from the best observed effect, thereby enabling rigorous global testing and confidence bounds for the optimal dose.

Original authors: Victor K. de la Pena, Fangyuan Lin, Demissie Alemayehu, Victor H. de la Pena

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Victor K. de la Pena, Fangyuan Lin, Demissie Alemayehu, Victor H. de la Pena

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a drug developer running a race to find the best dose of a new medicine. You have K different doses (let's say 6) and a placebo (a control group). You don't know which dose works best, so you start giving them out and watching the results as they come in, day by day.

The paper you provided solves two major problems that happen when you do this kind of "watch-and-see" experiment:

The "Winner's Curse" (The Optimism Trap): If you keep looking at the data, one dose will naturally look like the "winner" just by chance, even if it's not actually the best. It's like flipping 100 coins; eventually, one will land on heads 10 times in a row just by luck. If you pick that coin as your "lucky coin," you are being overly optimistic.
The "Peeking" Problem (The Multiplicity Trap): If you check the results every day, you have many chances to make a mistake. If you flip a coin once, you have a 5% chance of getting a "false alarm" (thinking it's biased when it's not). But if you flip it 1,000 times and check every time, the chance of seeing a false alarm at least once becomes almost 100%.

The Paper's Solution: The "Toll Booth" System

The authors propose a new statistical rule that acts like a smart toll booth for your drug trial. This system allows you to check the data as often as you want ("anytime-valid") without breaking the rules of statistics, but it forces you to pay a price for your optimism.

Here is how the system works, using a simple analogy:

1. The "Raw Score" vs. The "Real Score"

Every day, you calculate the Raw Score: "Which dose looks best right now compared to the placebo?"

The Problem: This score is inflated. It includes the "luck" of being the current leader.

2. The "Selection Charge" (The Toll)

To fix the optimism, the system calculates a Selection Charge.

The Analogy: Imagine the doses are runners in a race. If two runners are tied for first place, the race is exciting, and there is a lot of "option value" (you might switch your bet to the other runner if they pull ahead). This uncertainty costs money.
The Rule: The system charges you a fee based on how many doses are currently tied for the lead.
- If 4 doses are tied, the charge is high (lots of uncertainty).
- If 1 dose is clearly winning and pulling away, the charge drops to zero.
Why? Once a clear winner emerges, there is no more "luck" in switching leaders. The paper proves that once a true best dose separates itself from the pack, the "tax" you pay for uncertainty disappears.

3. The "Monitoring Margin" (The Safety Net)

Because you are allowed to peek at the data every day, you need a second safety net to prevent false alarms.

The Analogy: This is like a "speed limit" that gets stricter the longer you drive. It ensures that even if you check the results 1,000 times, you won't accidentally declare a winner just because you got lucky with the timing.

The Final Decision Rule

To declare "GO" (meaning the drug is effective and we should move to the next phase), the system requires you to pass a three-part test:

Raw Best Effect minus Selection Charge minus Monitoring Margin must be greater than The Clinical Goal.

Raw Best Effect: What the data says right now.
Minus Selection Charge: Paying for the fact that you picked the "luckiest" looking dose.
Minus Monitoring Margin: Paying for the fact that you kept looking at the data.
The Result: If the remaining number is still high enough to meet your medical goal, you can stop the trial and say, "Yes, this drug works!"

What This Paper Actually Claims (and What It Doesn't)

What it DOES:

It provides a mathematical guarantee that you will never falsely claim a drug works more than 5% of the time (or whatever error rate you set), even if you check the data every single day.
It gives you a "lower confidence floor." This is a number that says, "We are 95% sure the true best effect is at least this high."
It works for both continuous data (like blood pressure numbers) and binary data (like "cured" vs. "not cured").

What it DOES NOT do:

It does not tell you which specific dose is the "true" best. It only tells you that at least one of the doses is good enough. If you need to know exactly which one to manufacture, you need a different tool.
It does not replace safety checks. It only looks at whether the drug works (efficacy), not if it is safe.
It does not assume a specific shape for the dose curve. It doesn't assume the drug gets better in a straight line; it just looks for the best performer.

Summary

Think of this method as a honest accountant for your drug trial. It lets you look at the books whenever you want, but it insists that you pay a "tax" for the excitement of a close race (selection charge) and a "fee" for checking the books too often (monitoring margin). Only after you pay these fees does it tell you if the drug is truly a winner. This prevents you from celebrating a "winner" that was just a fluke of luck.

Technical Summary: Mitigating the Winner's Curse While Controlling Multiplicity in Dose-Ranging Trials

Problem Statement
Phase II dose-ranging trials typically involve randomizing $K$ active doses against a concurrent control and repeatedly analyzing accumulating data to identify a dose with a clinically meaningful effect. This standard practice introduces two distinct statistical challenges:

The Winner's Curse (Selection Optimism): Selecting the dose with the largest observed effect ( $\hat{\delta}_{\max}$ ) creates an upward bias. Even if individual estimators are unbiased, the maximum of $K$ estimators is optimistically biased because it was selected for being large.
Multiplicity: Repeatedly checking data over time (interim analyses) and across multiple doses inflates the Type I error rate. Standard multiplicity corrections (e.g., Dunnett's test) control error at fixed looks but do not account for the bias of the selected maximum, while bias corrections do not inherently license repeated "peeking" at the data.

The paper addresses the need for a procedure that provides anytime-valid inference: a method that controls the global Type I error rate (the probability of ever falsely declaring a "go" decision) at any arbitrary stopping time, while simultaneously correcting for the optimism inherent in selecting the current best dose.

Methodology
The authors propose a new statistical framework based on e-processes and the selection premium identity. The core innovation is treating the "winner's curse" not just as a bias to be subtracted, but as a predictable drift that must be accounted for in a martingale-based testing framework.

The Selection Premium Identity:
The method utilizes a "selection premium" function, $sp_L(u)$ , which quantifies the expected one-step gain of retaining the option to switch the leader (the current best dose) after observing new noise.
- Let $S_{t,k}$ be the cumulative centered noise for dose $k$ at look $t$ .
- The selection premium is defined as $sp_L(u) = E[\max_k(u_k + \eta_k)] - \max_k u_k$ , where $\eta$ represents the distribution of the next noise increment.
- This premium represents the "option value" of waiting to see if a different dose becomes the leader.
The Equal-Margin Reference:
To calculate the premium without knowing the true underlying effects, the method uses an "equal-margin" reference model where all doses have the same effect $\delta$ . The authors prove that the selection premium under this equal-margin configuration serves as a conservative envelope for the expected optimism under any heterogeneous configuration of true dose effects. This allows the procedure to charge a "selection fee" that is valid regardless of the true dose-response shape.
Construction of the E-Process:
The procedure constructs a corrected residual process $R_t(\delta)$ :
$R_t(\delta) = G_t(\delta) - A_t(\delta)$
Where:
- $G_t(\delta)$ is the cumulative score (raw best effect minus the candidate margin $\delta$ ).
- $A_t(\delta) = \sum_{s=1}^t a_s(\delta)$ is the cumulative selection charge, where $a_s$ is the selection premium calculated at each step based on the current state (specifically, the number of tied leaders).
- By subtracting this predictable upward drift ( $A_t$ ), the residual $R_t$ becomes a supermartingale with non-positive drift.
Anytime-Valid Decision Rule:
The method converts the corrected residual into an e-process $E_t(\delta)$ using a half-normal mixture over a betting parameter $\lambda$ . The decision rule to declare a "global go" (that the best true effect $\delta^*$ exceeds a clinical margin $\delta_0$ ) is:
$\hat{\delta}_{\max, t} \geq \delta_0 + \frac{A_t(\delta_0)}{n_t} + \frac{q_\alpha(V_t; \rho)}{n_t}$
- $\hat{\delta}_{\max, t}$ : Raw observed best effect.
- $A_t(\delta_0)/n_t$ : The selection charge (correcting for winner's curse).
- $q_\alpha(V_t; \rho)/n_t$ : The monitoring margin (correcting for repeated looks, derived from the e-process boundary).
- $n_t$ : Cumulative sample size per arm.
The paper provides specific implementations for Gaussian outcomes (using numerical integration or Monte Carlo for the premium) and Binary outcomes (using a closed-form solution based on the number of tied leaders, $r_t$ ).

Key Contributions

Unified Correction: The paper separates the control of Type I error (via the e-process boundary) from the correction of selection bias (via the selection charge), yet unifies them in a single decision rule.
State-Dependent Premium: Unlike static corrections, the selection charge is dynamic. It is high when doses are tied (high uncertainty about the winner) and decays to zero once a unique leader emerges, reflecting the diminishing "option value" of switching.
Anytime-Valid Confidence Sequences: By inverting the test for all candidate margins $\delta$ , the method generates an anytime-valid lower confidence sequence for $\delta^*$ (the true best effect), ensuring that the lower bound converges to the true best effect as sample size increases.
Robustness: The method does not rely on parametric dose-response models (e.g., Emax, logistic) or trend assumptions, making it applicable to exploratory trials where the dose-response shape is unknown.

Results and Performance

Type I Error Control: Simulation studies confirm that the proposed rule strictly controls the Type I error rate (false "go" probability) at the nominal level $\alpha$ (e.g., 0.05) even under continuous monitoring and arbitrary stopping times. In contrast, applying fixed-look thresholds (like Dunnett's) to unplanned interim looks results in severe Type I error inflation (up to 24% in simulations).
Power and Efficiency: The method demonstrates reasonable power to detect effects exceeding the clinical margin. While the "selection charge" imposes a cost early in the trial (when ties are common), this cost vanishes asymptotically as a unique leader emerges.
Binary Outcomes: For binary endpoints, the method is computationally efficient, requiring only the count of tied leaders to calculate the selection charge. Simulations show that the procedure correctly identifies the "go" decision when the true effect is sufficiently large, with the lower confidence floor converging to the true best effect.
Comparison: The proposed method is compared against armwise union e-processes and fixed-look max-z tests. While armwise e-processes may offer slightly higher power in some scenarios, the proposed method provides a distinct advantage by explicitly modeling the selection process of the maximum rather than testing individual arms independently.

Significance and Claims
The paper claims to provide a mathematically explicit and operationally transparent solution for program-level proof-of-activity in dose-ranging trials.

Scope: The method is designed to test whether at least one dose in the range exceeds a clinical margin and to provide a lower bound for the best true effect.
Limitations: The authors explicitly state that the method does not identify the true best dose (localization), nor does it provide unbiased estimates for the selected dose's effect. It does not replace dose-response modeling, safety assessments, or benefit-risk evaluations.
Value: The primary contribution is the ability to make "winner-bias-aware" decisions in flexible, adaptive trial designs without inflating Type I error. The selection charge acts as a "ledger" that quantifies the cost of optimism, ensuring that a "go" decision is only made when the raw evidence remains robust after paying for both the selection of the winner and the repeated monitoring of the data.

The paper concludes that this approach offers a rigorous framework for anytime-valid inference in settings where the "winner" is selected dynamically, bridging the gap between selection bias correction and sequential hypothesis testing.