📊 statistics

Causal inference for N-of-1 trials

This paper establishes a formal causal inference framework for N-of-1 trials by defining a unit-specific conditional average treatment effect (U-CATE) and demonstrating how to identify and estimate it using simple mean differences or the time-varying g-formula, depending on the complexity of factors like carryover effects and time-varying confounders.

Original authors: Marco Piccininni, Mats J. Stensrud, Zachary Shahn, Stefan Konigorski

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Marco Piccininni, Mats J. Stensrud, Zachary Shahn, Stefan Konigorski

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 trying to figure out which video game controller makes you play the best. You could ask a thousand other gamers what they like, but that's like asking a crowd what they think of a specific flavor of ice cream; it tells you about the average person, not about your specific taste buds. This is the heart of "personalized medicine": the idea that what works for the "average" patient might not work for you. To solve this, scientists use a special kind of experiment called an N-of-1 trial. Think of it as a scientific "A/B test" for a single person. Instead of comparing two groups of people, you compare two versions of yourself over time: one week you take the medicine, the next week you take a sugar pill, and you switch back and forth. The goal is to see which one actually makes you feel better, cutting through the noise of daily life.

However, just because you switch back and forth doesn't mean the math is easy. Your body isn't a blank slate; it has rhythms, habits, and memories. If you take a pill on Monday, does it still be working on Wednesday? If you feel sick on a rainy Tuesday, is it the weather or the pill? This is where causal inference comes in. It's the branch of science that tries to answer the question: "Did this cause that?" rather than just "Did they happen at the same time?" For years, people have debated how to analyze these single-person trials. Some thought randomizing the schedule was enough to fix everything, while others worried that the order of events or the passage of time could trick the results. The big question has been: How do we get a clear, honest answer about what works for one specific person, even when their body is messy and complicated?

This paper steps in to bring some serious order to this chaos. The authors, a team of data scientists and mathematicians, built a formal "rulebook" for N-of-1 trials using the language of causal inference. They didn't just guess; they created a mathematical framework that treats a single person's data like a tiny, self-contained universe. They define a specific target called the U-CATE (which sounds like a fancy robot, but it just means the "average effect for this specific person"). They show that if the rules of the game are simple—meaning the medicine wears off quickly and doesn't leave a lingering "ghost" effect—then the math is straightforward. You can just average the "good days" and the "bad days" to find the answer.

But the authors know real life isn't simple. They tackle the messy scenarios where the medicine might have a "carryover" effect (like a hangover from a pill that lasts longer than the dose), where your mood might trend up or down over time, or where your body's reaction to the pill depends on what happened yesterday. In these complex situations, they show that a simple average isn't enough. Instead, they propose using a sophisticated mathematical tool called the g-formula. Think of this as a time-travel simulation engine. It takes all the data you have—your temperature, the time of day, your symptoms, and the treatment history—and runs a virtual simulation to ask, "What would have happened if you had taken the pill every single time?" versus "What if you never took it?"

The paper proves that under certain conditions, this simulation engine can accurately identify the true effect of a treatment for a single individual, even when the data is noisy and the effects linger. They tested these ideas on real data from people tracking their acne symptoms. In one case, a simple average worked fine. In another, where the skin's reaction seemed to depend on the time of day and temperature, they had to use their complex simulation tool. The result? They found that the acne gel worked great for one person but didn't seem to help another, highlighting exactly why we need these personalized methods. The authors also clarify a common confusion: while randomizing the order of treatments is helpful, it's not a magic wand that fixes everything if the underlying biology is complex. You still need the right math to interpret the results.

Ultimately, this paper doesn't just say "N-of-1 trials are cool"; it gives scientists the precise instructions on how to read the results without getting fooled by the quirks of human biology. It shows that with the right assumptions and the right tools, we can move beyond guessing what works for the "average" person and finally know what works for you.

Technical Summary: Causal Inference for N-of-1 Trials

Problem Statement

The conventional approach to causal inference in randomized controlled trials (RCTs) focuses on estimating the average causal effect within a population. However, the rise of personalized medicine has highlighted the limitations of this approach, as individuals often exhibit heterogeneous responses to treatments, possess unique preferences, or may not meet the eligibility criteria of existing population-level RCTs. N-of-1 trials, defined as within-person crossover trials where a single individual receives treatment and control in sequential periods, offer a potential solution for identifying individual-specific causal effects.

Despite their intuitive appeal, the analysis and interpretation of N-of-1 trials remain methodologically complex. Confusion persists regarding the role of randomization, the handling of confounding (specifically autocorrelation and carryover effects), and the appropriate causal estimands. Previous literature has offered conflicting views: some claim randomization eliminates the need for causal inference methods, while others argue that autocorrelation and carryover must be accounted for even in randomized designs. This paper addresses the need for a formal causal inference framework to ground intuitive claims, define precise estimands, and establish identification and estimation strategies for single N-of-1 trials.

Methodology

Framework and Notation

The authors adopt a formal causal inference framework based on structural causal models (SCMs) and counterfactual reasoning. They define an N-of-1 trial as a study where a treatment schedule $Z$ is assigned at baseline and repeats cyclically over time. The treatment status $A_k$ at time $k$ is a deterministic function of $Z$ . The analysis focuses on a single individual, conceptualized as a draw from a near-infinite superpopulation, allowing for the definition of individual-specific causal effects.

Causal Models

The paper develops two primary causal models:

The Basic Scenario:
- Assumptions: The outcome $Y_k$ is generated by a structural equation $Y_k = f^Y(A_k, U, \epsilon_k)$ , where $U$ represents unmeasured individual-specific characteristics (baseline covariates) and $\epsilon_k$ is time-varying noise.
- Key Constraints: The model assumes no carryover effects (treatment at time $k$ does not affect outcomes at $k' > k$ ), no outcome-outcome effects, and no time-varying common causes of the outcome whose distribution changes over time.
- Stationarity: It assumes strong conditional stationarity: the distribution of counterfactual outcomes is constant over time given $U$ .
The Relaxed Causal Model:
- Motivation: To address threats to stationarity common in clinical settings, such as carryover effects, autocorrelation, and time trends.
- Structure: This model introduces a time-varying covariate $L_k$ (e.g., environmental factors, physiological states) that may be affected by previous treatment, previous outcomes, and previous covariate states. The outcome $Y_k$ depends on $A_k$ , $Y_{k-1}$ , $L_k$ , $L_{k-1}$ , and individual-specific variables $U$ .
- Assumptions: While strong stationarity is relaxed, the model imposes regularity constraints: the functional forms generating $L_k$ and $Y_k$ do not change over time, and noise distributions remain stationary. It allows for direct and indirect carryover effects and outcome-outcome dependencies.

Identification and Estimation

Target Estimand: The paper defines the U-CATE (Conditional Average Treatment Effect conditional on $U$ ) as the primary estimand. This represents the contrast between the expected counterfactual outcome under an "always treated" strategy versus a "never treated" strategy for a specific individual.
Basic Scenario Estimation: Under the basic model, the U-CATE is identified by the simple mean difference between observed outcomes during treatment periods and control periods. The authors prove that this simple difference is an unbiased, asymptotically normal estimator.
Relaxed Model Estimation: Under the relaxed model, simple mean differences are biased due to carryover and time trends. The authors show that the U-CATE can be identified using a longitudinal g-formula. This requires estimating conditional distributions of the outcome and covariates given their direct causes (e.g., $P(Y_k | A_k, Y_{k-1}, L_k, L_{k-1}, U)$ ).
Role of Randomization: The authors argue that randomization of the treatment schedule is not strictly necessary for identification in these settings because the inference is conditional on the individual's characteristics ( $U$ ) and the specific schedule ( $Z$ ). Randomization serves primarily to ensure exchangeability if one wishes to make marginal inferences over time points or different schedules, but the core individual-specific inference relies on the structural assumptions and stationarity conditions.

Key Results

Formalization of N-of-1 Inference: The paper successfully formalizes N-of-1 trials within a superpopulation framework, defining the U-CATE as a rigorous individual-specific causal effect.
Validity of Simple Estimators: In the "basic scenario" (no carryover, no time trends), a simple unpaired mean difference is a valid, unbiased estimator for the U-CATE. This provides theoretical justification for the use of simple t-tests in specific N-of-1 contexts.
Identification under Complexity: In more realistic settings involving carryover effects, autocorrelation, and time-varying confounders, the paper demonstrates that the U-CATE is still identifiable via the g-formula, provided that specific stationarity conditions hold and sufficient data (repeated cycles) are available.
Data Application: The authors applied their methods to public data from two N-of-1 trials regarding acne severity.
- Under the basic model, one participant showed a significant treatment effect, while the other did not.
- Under the relaxed model (accounting for time of day, temperature, and potential carryover), the g-computation algorithm confirmed a favorable treatment effect for the second participant but found no evidence of an effect for the first.
- This application illustrates how different assumptions about the data-generating process lead to different analytical strategies and results, emphasizing the importance of model specification.

Significance and Claims

The paper claims to provide a necessary theoretical foundation for the analysis of N-of-1 trials, moving beyond ad-hoc statistical approaches. By grounding these trials in a formal causal framework, the authors aim to:

Clarify misconceptions regarding the necessity of randomization and the handling of confounding in single-subject designs.
Justify the use of simple estimators in specific, well-defined scenarios.
Provide a rigorous pathway (via the g-formula) for analyzing complex N-of-1 trials where carryover and time trends are present.
Highlight that while results from a single N-of-1 trial are specific to the individual (limited generalizability), aggregating a series of N-of-1 trials can yield efficient estimates of population-level average treatment effects under specific assumptions (e.g., no effect modification by $U$ ).

The authors maintain a modest stance, acknowledging that the validity of their results depends heavily on the plausibility of the underlying causal assumptions (e.g., stationarity, correct model specification). They note that while the framework allows for the identification of individual effects, the transportability of these results to future time points or other individuals remains a challenge inherent to the nature of personalized medicine.