📊 statistics

Separating Biological and Choice Effects in Randomized Trials: Completing Causal Inference with β-Identification

This paper proposes a causal inference framework that decomposes randomized trial outcomes into biological effects ( $\alpha$ ) and preference-mediated choice effects ( $\beta$ ), arguing that formally identifying both components resolves ambiguity caused by nonadherence and crossover while providing a more complete understanding of treatment efficacy in preference-sensitive care.

Original authors: Ogan Gurel, James Weinstein

Published 2026-08-20

📖 9 min read🧠 Deep dive

CC BY 4.0

Original authors: Ogan Gurel, James Weinstein

Original paper licensed under CC BY 4.0 (https://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

In the world of medical research, the gold standard for proving a treatment works is the randomized controlled trial. In these studies, patients are randomly assigned to receive either a new treatment or a standard one, like flipping a coin to decide their path. The goal is to isolate the biological effect of the medicine or procedure itself, stripping away other variables to see if the treatment truly changes the outcome. For decades, this method has been the bedrock of evidence-based medicine, guiding doctors on which drugs to prescribe and which surgeries to perform. However, a persistent problem arises when the biological differences between treatments are small. In these cases, the results of a trial can become muddy or inconclusive. Patients often do not follow the assigned path; they might cross over to the other treatment, or they might refuse the assigned one entirely. Traditionally, researchers have viewed this non-adherence as a flaw, a source of noise that weakens the study's ability to find a clear answer. They have treated the patient's choice to switch treatments as a mistake to be corrected or ignored, believing that the true story lies only in the biology of the drug or procedure.

A new framework proposed by researchers Ogan Gurel and James Weinstein challenges this long-held view. They argue that in many modern medical situations, the patient's choice is not just a distraction but a fundamental part of the outcome. When the biological difference between two credible treatments is small, the decision a patient makes—driven by their personal values, fears, and preferences—becomes a powerful force that shapes their health just as much as the treatment itself. The researchers suggest that by treating patient choice as a measurable, causal factor rather than a statistical error, we can finally make sense of trials that have previously seemed to fail. This approach does not discard the old methods but expands them, allowing scientists to separate the biological effect of a treatment from the effect of the patient's decision to take it. By doing so, they can explain why some treatments work well in practice even when the biology suggests they shouldn't, and why some trials remain confusing until we acknowledge the power of human agency.

The core of this new thinking comes from re-examining a famous study known as the Spine Patient Outcomes Research Trial, or SPORT. This trial investigated treatments for a painful condition in the lower back called lumbar disc herniation. The study was designed with rigorous scientific standards, yet its results were famously inconclusive. The data showed that the difference in outcomes between patients who had surgery and those who received non-surgical care was very small. For years, this was interpreted as a sign that surgery offered no real biological advantage over non-surgical care. However, the researchers point out that the trial was also marked by high rates of "crossover." Many patients assigned to non-surgical care eventually chose to have surgery, and some assigned to surgery opted for non-surgical care instead. In the traditional view, this crossover was a problem that blurred the lines between the two groups, making it impossible to tell which treatment was better. The researchers argue that this view missed a crucial point: the crossover was not just noise; it was a signal.

By applying their new framework, the researchers were able to decompose the results of the SPORT trial into two distinct parts. The first part, which they call the biological effect, represents the pure physiological impact of the treatment itself, assuming a patient follows the plan perfectly. The second part, which they term the choice effect, captures the impact of the patient's decision to take or switch treatments. When they separated these two components, a different picture emerged. The biological difference between surgery and non-surgical care was small, and the choice effect was found to be clinically meaningful. Patients who chose their own treatment, regardless of what they were originally assigned, had outcomes that were jointly shaped by biology and choice. This suggests that the success of the treatment was not just about the surgery or the physical therapy, but about the alignment between the treatment and the patient's own preferences. The act of choosing a treatment that felt right to the patient was a causal factor in their recovery, just as real as the biological mechanism of the surgery.

This insight transforms how we should interpret clinical trials, especially in fields where multiple treatments exist and the biological differences between them are narrowing. The researchers propose a new way to classify trials based on the balance between these two forces. In some cases, the biological difference is so large that the patient's choice matters very little; in these scenarios, the old methods of analysis work perfectly fine. But in many other cases, particularly as patients age and face complex health issues, the biological differences shrink, and the patient's choice becomes the dominant factor. The researchers identify four distinct regimes for these trials. In one regime, biology is the clear winner, and doctors can recommend the superior treatment with confidence. In another, both biology and choice matter, requiring a more nuanced approach where doctors and patients decide together. In a third, neither factor is strong enough to give a clear answer, and the trial simply tells us we need more research. Finally, in the fourth regime, the biological difference is negligible, but the patient's choice drives the outcome. Here, the most important thing a doctor can do is not to prescribe a specific treatment, but to help the patient make the choice that fits their life and values.

The paper argues that failing to recognize this fourth regime has led to many trials being labeled as failures when they were actually successful in a different way. When a trial shows no clear biological winner, but patients who choose their own path do well, the traditional conclusion is that the treatments are equal and the trial is inconclusive. The new framework suggests that the trial is actually telling us something profound: that the outcome depends on the patient's engagement and preference. By ignoring the choice effect, we lose the ability to understand why a treatment works in the real world. The researchers demonstrate that by measuring both the biological effect and the choice effect on the same scale, we can see the full picture. This allows us to distinguish between a treatment that is biologically weak and a treatment that is biologically neutral but highly effective when patients choose it willingly.

This shift has significant implications for how medical research is designed and how doctors make decisions. The researchers suggest that future trials should be designed to preserve the ability to measure patient choice, rather than trying to force patients to stick to a protocol. This might mean allowing patients to cross over or switch treatments if they wish, and then analyzing the data to see how those choices affected the outcome. It also means that the goal of a trial is not just to find the "best" treatment in a vacuum, but to understand how different treatments perform when patients are free to choose. In the realm of artificial intelligence and decision support tools, this distinction is critical. An AI system that only looks at biological data might recommend a treatment that is biologically sound but fails in practice because it does not align with patient preferences. By incorporating the choice effect, these tools can be designed to defer to human judgment when the biology is unclear, helping doctors facilitate shared decision-making rather than simply ranking treatments.

The researchers emphasize that this approach does not replace the rigorous standards of randomized trials but completes them. It acknowledges that in a world of complex, chronic conditions, the patient is not a passive recipient of care but an active participant whose decisions shape the result. The framework provides a way to quantify this participation, turning what was once considered a statistical nuisance into a vital piece of evidence. By doing so, it offers a more honest and complete account of what happens in a clinical trial. It explains why some treatments seem to work better in the real world than in the controlled setting of a study, and why some trials that look like failures on paper actually provide valuable guidance for doctors and patients. The ultimate goal is to move beyond a narrow focus on biology to a broader understanding of health that includes the human element of choice, ensuring that medical evidence reflects the reality of how people live with and manage their conditions.

The study relies on a specific mathematical decomposition to separate these effects, but the concept is straightforward. The researchers take the observed difference in outcomes between patients who received one treatment and those who received another, and they subtract the biological difference that would exist if everyone had followed their assignment perfectly. What remains is the effect of the choice itself. This remaining piece is not random noise; it is a measurable, causal force. The researchers show that in the SPORT trial, this choice effect was large enough to be clinically meaningful, explaining why the trial's results were so different from what a purely biological analysis would predict. They argue that this same pattern likely exists in many other areas of medicine, from heart disease to cancer, where patients have to weigh the benefits of a treatment against its burdens and risks.

By making the choice effect visible, the researchers hope to change the conversation in medical research. Instead of asking only "Does this treatment work?", we can also ask "Who does it work for, and under what conditions of choice?" This shift allows for a more personalized approach to medicine, where the focus is not just on the drug or the surgery, but on the partnership between the doctor and the patient. It recognizes that the best treatment is not always the one with the strongest biological effect, but the one that the patient is most likely to stick with and that fits their life. The researchers conclude that as medicine becomes more complex and the biological differences between treatments become smaller, the ability to measure and understand the power of choice will become increasingly important. It is a way to complete the picture of causal inference, ensuring that the evidence we gather reflects the full reality of human health and decision-making.

Technical Summary: Separating Biological and Choice Effects in Randomized Trials

Problem Statement
The paper addresses a central limitation in modern clinical trials, particularly within "preference-sensitive" care settings where biological differences between treatments narrow. In these scenarios, outcomes are jointly determined by biological efficacy ( $\alpha$ ) and patient preference-mediated choice ( $\beta$ ). Traditional randomized controlled trial (RCT) analysis often treats nonadherence, crossover, and heterogeneous treatment uptake as sources of noise or bias that attenuate the treatment signal toward the null. Consequently, trials may yield equivocal or inconclusive results because the analysis remains " $\alpha$ -centric," failing to quantify the causal contribution of patient choice. The authors argue that without explicitly separating these components, trials cannot distinguish between weak biological efficacy and strong preference-mediated influence, leading to a "paradox of rigor without resolution."

Methodology: The $\beta$ -Identification Framework
The authors propose a causal inference framework that decomposes observed outcomes into two co-estimands:

$\alpha$ (Biological Effect): The causal effect of treatment assignment, estimated via Intention-to-Treat (ITT) or, under nonadherence, as the Local Average Treatment Effect (LATE) or Complier Average Causal Effect (CACE) using randomized assignment as an instrumental variable (IV).
$\beta$ (Preference-Mediated Effect): The causal effect of patient choice, defined as the difference between the observed "as-treated" outcome difference ( $\Delta_{OBS}$ ) and the causally identified biological contrast ( $\Delta\alpha_{causal}$ ).

The framework relies on six axioms (A1–A6) and five operational rules (R1–R5) to guide design and interpretation. Key methodological requirements include:

Dual Causation: Acknowledging that outcomes arise from both biology and choice.
Common Scale: Estimating $\alpha$ and $\beta$ on the same clinical outcome scale, anchored to a Minimal Clinically Important Difference (MCID).
Non-Interference: Ensuring that the measurement of preference does not itself act as an intervention altering treatment uptake.
Identification Logic: Utilizing randomized assignment to identify $\alpha$ while preserving observable treatment uptake (crossover/nonadherence) to identify $\beta$ .

The framework introduces a Regime Map classifying trials into four quadrants based on the magnitude of $|\Delta\alpha|$ and $\beta$ relative to the MCID:

B1 ( $\alpha$ -Dominant): Biology is decisive; standard RCTs suffice.
B2 (Mixed): Both biology and choice are clinically material; requires joint inference.
B3 (Decision-Limited): Neither is decisive, or identification is weak; results in diagnostic restraint.
B4 ( $\beta$ -Dominant): Biology is equivocal, but choice is decisive; requires $\beta$ -first inference and shared decision-making.

Key Contributions

Reframing Nonadherence: The paper reclassifies crossover and nonadherence from "bias" to "identifying variation." In $\beta$ -dominant settings, these deviations provide the necessary variation to quantify the causal effect of choice.
Canonical Decomposition: It provides a formal algebraic definition for $\beta$ ( $\beta = \Delta_{OBS} - \Delta\alpha_{causal}$ ), allowing for the separation of biological and choice effects on a common scale.
Regime-Routed Design: The authors propose that trial design, power, and interpretation must be routed by the identified regime rather than a fixed design label. For instance, in $\beta$ -dominant regimes, suppressing crossover (to enforce adherence) would destroy the ability to identify $\beta$ .
Extension of Causal Inference: The framework extends standard instrumental variable and LATE methodologies to include preference-mediated effects as primary estimands, rather than residual noise.

Results and Empirical Placement
The paper applies this framework to re-evaluate existing trials, using the Spine Patient Outcomes Research Trial (SPORT) for lumbar disc herniation as a primary exemplar.

SPORT Reinterpretation: While originally judged inconclusive due to high crossover, the authors argue SPORT is a $\beta$ -dominant (B4) trial. The biological contrast ( $\Delta\alpha$ ) was small, but the preference-mediated effect ( $\beta$ ) was clinically significant. The "inconclusive" nature of the original analysis stemmed from an $\alpha$ -centric interpretation that could not account for the choice effect.
Other Exemplars: The framework places other trials into specific regimes:
- $\beta$ -Dominant (B4): ProtecT (prostate cancer), CODA (appendicitis), and COURAGE (coronary disease), where biological differences are minimal but treatment burdens and preferences drive outcomes.
- Mixed (B2): SPORT (spinal stenosis), where both biology and choice are material.
- Decision-Limited (B3): ISCHEMIA, where neither biology nor choice is sufficiently identified to drive a decisive recommendation.

Significance and Claims
The paper claims that $\beta$ -identification completes causal inference in preference-sensitive care rather than replacing standard RCT analysis. Its significance lies in:

Resolving Equivocal Trials: It transforms "null" or "inconclusive" findings into interpretable structures, distinguishing between weak biology and strong choice effects.
Strengthening Evidence-Based Medicine: By explicitly defining the domain where $\alpha$ -centric inference is valid (B1) versus where it fails (B2, B4), the framework prevents the misapplication of biological rankings in contexts governed by patient values.
Separating Generalizability: It separates invariant biological effects ( $\alpha$ ) from context-dependent choice effects ( $\beta$ ), clarifying external validity.
AI and Decision Support: The authors argue that clinical AI systems must defer to human judgment or explicit preference elicitation in $\beta$ -dominant regimes, as biological ranking alone is insufficient and potentially misleading.

The authors conclude that as clinical complexity increases (aging, multimorbidity), the field will shift from $\alpha$ -dominant to mixed or $\beta$ -dominant regimes. Therefore, treating patient choice as a causal structure to be measured, rather than noise to be suppressed, is a methodological necessity for modern clinical research.

Technical Summary: Separating Biological and Choice Effects in Randomized Trials

More like this