Causal Inference under Treatment Noncompliance

21 concepts21 papers

In an ideal randomized controlled trial (RCT), the experimenter has complete control over the environment. You randomly assign a participant to a treatment group ($Z=1$) or a control group ($Z=0$), and they receive exactly what was assigned ($D=Z$).

But when experiments involve human agency, this control breaks down. A clinician can prescribe a daily vitamin, but the patient might forget to take it, as observed in long-term clinical studies like the MAVIDOS Maternal Vitamin D Osteoporosis Study. A doctor might recommend a traditional herbal protocol for fatty liver disease, but the patient might fail to adhere to the lifestyle adjustments, a common challenge highlighted by Qian Huang et al. in Effectiveness and safety analysis of Danggui Shaoyao Powder for the treatment of non-alcoholic fatty liver disease. Similarly, a website might prompt a user to subscribe to a premium tier, but the user can simply ignore the recommendation.

This discrepancy between the assigned treatment ($Z$) and the received treatment ($D$) is known as treatment noncompliance or imperfect compliance. It introduces a fundamental dilemma that invalidates standard statistical tools.

The Fundamental Dilemma of Noncompliance

When participants do not comply, researchers typically default to one of two naive analyses, both of which are flawed:

  1. Intention-to-Treat (ITT): This compares outcomes based purely on the original assignment ($Z$), ignoring what participants actually did. While ITT preserves the benefits of randomization and remains unbiased for the effect of the assignment, it severely dilutes the estimated effect of the treatment itself. If only 10% of people take a drug, the ITT effect will look tiny, even if the drug is a miracle cure for those who take it.
  2. As-Treated / Per-Protocol: This compares those who actually took the treatment ($D=1$) against those who did not ($D=0$). While this sounds intuitive, it completely destroys the benefit of randomization. The decision to comply is rarely random; it is driven by self-selection. Sicker patients might be more motivated to take a drug, or healthier individuals might experience fewer side effects. This self-selection introduces unobserved confounding, biasing the treatment effect estimate. An early attempt to model this complex clinical uncertainty without restrictive assumptions used Bayesian graphical models and Gibbs sampling in A Clinician’s Tool for Analyzing Non-Compliance.

To recover the true causal effect of the treatment without introducing self-selection bias, we must view the problem through the lens of Principal Stratification.

Principal Stratification and the Instrumental Variable Solution

To understand why we can isolate causal effects despite noncompliance, we categorize the population into four latent subgroups (strata) based on how they would react to their assignment:

  • Compliers: Take the treatment if assigned to it, and do not take it if assigned to control ($D(1)=1, D(0)=0$).
  • Always-Takers: Take the treatment regardless of their assignment ($D(1)=1, D(0)=1$).
  • Never-Takers: Refuse the treatment regardless of their assignment ($D(1)=0, D(0)=0$).
  • Defiers: Do the exact opposite of their assignment ($D(1)=0, D(0)=1$).

In most applications, researchers invoke the monotonicity assumption, which asserts that there are no defiers.

Under this framework, the randomized assignment $Z$ acts as an Instrumental Variable (IV). It affects the final outcome $Y$ only through its ability to change the received treatment $D$ (a condition known as the exclusion restriction).

Because Always-Takers and Never-Takers do not change their behavior regardless of whether $Z=0$ or $Z=1$, the assignment has zero effect on their treatment status. Under the exclusion restriction, the assignment also has zero effect on their final outcomes. Therefore, any observed difference in outcomes between the randomized groups ($Z=1$ vs $Z=0$) must be driven entirely by the Compliers.

This realization allows us to identify the Local Average Treatment Effect (LATE)—also known as the Complier Average Causal Effect (CACE). As Guido Imbens argues in Better LATE Than Nothing, prioritizing the internal validity of LATE provides a mathematically rigorous foundation for causal inference, isolating a clean treatment effect for a specific, well-defined subpopulation without relying on untestable assumptions about the entire population.

$$\text{LATE} = \frac{\text{ITT}_{\text{outcome}}}{\text{ITT}_{\text{compliance}}} = \frac{\mathbb{E}[Y \mid Z=1] - \mathbb{E}[Y \mid Z=0]}{\mathbb{E}[D \mid Z=1] - \mathbb{E}[D \mid Z=0]}$$

By dividing the total assignment effect ($\text{ITT}_{\text{outcome}}$) by the proportion of the population that actually complied ($\text{ITT}_{\text{compliance}}$), we scale the diluted effect back up to its true magnitude for the complier subgroup.

Identifying and Testing Latent Strata

The core challenge of principal stratification is that we cannot directly observe an individual's stratum. If a participant is assigned to control ($Z=0$) and does not take the treatment ($D=0$), they could be a Complier or a Never-Taker; their counterfactual behavior under $Z=1$ remains hidden.

To conduct rigorous frequentist inference in this setting, Wen Wei Loh et al. developed a finite-population generalized likelihood ratio test in A Finite Population Likelihood Ratio Test of the Sharp Null Hypothesis for Compliers. Their method computes exact $p$-values to test whether a treatment has a causal effect specifically for the unobserved Compliers, bypassing the need to identify individuals directly.

Real-world data often compounds noncompliance with missing outcomes. Sierra M. Li addresses this in randomized encouragement designs—where participants are merely encouraged to take a treatment, such as a flu shot—by introducing moment and maximum likelihood estimators that handle both noncompliance and missing data simultaneously (ITT analysis of randomized encouragement design studies with missing data).

Similarly, Trang Quynh Nguyen et al. provide a unifying framework to identify both complier and noncomplier average causal effects under "latent missing-at-random" (LMAR) conditions, offering researchers a robust mathematical toolkit when perfect compliance and complete outcomes are unavailable (Identification of complier and noncomplier average causal effects in the presence of latent missing-at-random (LMAR) outcomes).

In modern machine learning contexts, mapping individuals to these latent groups is crucial for estimating Individual Treatment Effects (ITE). Céline Beji et al. demonstrate that while standard ITE models assume perfect compliance, adapting these frameworks to account for noncompliance is a critical frontier for accurately discovering latent causal populations (Estimating Individual Treatment Effects through Causal Populations Identification).

Beyond Averages and One-Sided Compliance

While classical LATE focuses entirely on the average treatment effect for compliers, recent research expands this toolkit in three major directions: distributional effects, population extrapolation, and two-sided noncompliance.

Distributional Effects

Instead of looking only at how a treatment shifts the mean outcome, researchers often want to see how it alters the entire distribution (e.g., variance, quantiles). Undral Byambadalai et al. combine instrumental variables with machine learning-based regression adjustments to isolate Local Distributional Treatment Effects (LDTE) under imperfect compliance, preserving statistical efficiency even within complex randomization schemes (Beyond the Average: Distributional Causal Inference under Imperfect Compliance).

Population Generalization

Because LATE is local to compliers, it cannot automatically generalize to the entire population—a major hurdle for policymakers. To bridge this gap, Jason Poulos introduced a reweighting method called PATT-C, which leverages covariate information to extrapolate from compliant subgroups and estimate the average treatment effect for the entire population (Estimating population average treatment effects from experiments with noncompliance).

Two-Sided Noncompliance

Many digital and medical systems suffer from "two-sided" noncompliance, where control participants can actively seek out and receive the treatment anyway. In low-compliance prescription settings, Thibaud Rahier et al. developed a meta-estimator (C-IPE) that decomposes the overall prescription effect into two distinct, easier-to-estimate components: the raw efficacy of the treatment and the probability of compliance, drastically reducing variance (Individual Treatment Prescription Effect Estimation in a Low Compliance Setting).

Relaxing Monotonicity: Bounding and Proxies

What happens if the monotonicity assumption fails? If "defiers" exist, the standard IV estimator breaks down because we can no longer guarantee that the instrument's effect flows unidirectionally. When exact point identification is impossible, researchers turn to optimization and proxy variables.

Junzhe Zhang et al. address this by introducing a "Restricted Instrumental Variable" model. Instead of forcing a single point estimate, they construct tractable linear programs to calculate sharp upper and lower bounds on causal effects for continuous outcomes in high-dimensional contexts, using these bounds to restrict exploration spaces in online learning (Bounding Causal Effects on Continuous Outcome).

Alternatively, point identification can be rescued without monotonicity if auxiliary data is available. Yuta Kawakami demonstrated that by pairing an instrumental variable with a "proxy variable" that correlates with unobserved potential outcomes, researchers can mathematically recover exact causal effects without relying on the restrictive monotonicity assumption (Instrumental Variable-based Identification for Causal Effects using Covariate Information).

Noncompliance in Sequential and Adaptive Settings

The challenges of noncompliance are no longer confined to static, offline clinical trials. They are increasingly vital in online, sequential decision-making environments like multi-armed bandits and A/B testing platforms.

Standard multi-armed bandits assume that the chosen action is the one executed. If a recommendation system serves an ad, it assumes the user experiences it. When users exercise agency and ignore recommendations, standard bandit algorithms suffer from severe bias and fail to find optimal policies.

Standard Bandit:    [Algorithm Choice]  =======================> [Observed Outcome]
                                         (Assumes Full Adherence)

Instrument Bandit:  [Algorithm Choice]  ---> [User Compliance] ---> [Observed Outcome]
                                         (Unobserved Confounding)

To solve this, Nathan Kallus formalized the Instrument-Armed Bandit (IAB) framework in Instrument-Armed Bandits. By treating the bandit's choices as instruments rather than direct treatments, the proposed algorithms achieve sublinear regret by explicitly focusing exploration and optimization on the subpopulation of compliers.

Expanding this to broader experimental designs, Miruna Oprescu et al. developed frameworks for efficient adaptive experimentation with noncompliance (Efficient Adaptive Experimentation with Noncompliance). Their approach mitigates the unobserved confounding that arises when participants choose whether to adhere to sequential recommendations, optimizing data collection on the fly.

This integration of noncompliance into advanced experimental frameworks highlights a major gap in traditional statistical adjustments. For instance, high-dimensional covariate adjustment methods, such as the Lasso-based adjustments for randomized block designs studied by Hanzhong Liu et al., still fundamentally assume full compliance (Design-Based Theory for Lasso Adjustment in Randomized Block Experiments and Rerandomized Experiments). Extending these high-dimensional regression adjustments to handle participant non-adherence remains an active and vital frontier in causal inference.

Key Foundational Work

A Clinician’s Tool for Analyzing Non-Compliance (1996) — contributed 1 concept elements to this direction.
ITT analysis of randomized encouragement design studies with missing data (2005) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_treatment_noncompliance_273acfea,
  title = {treatment noncompliance},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/treatment-noncompliance/rdr_273acfead61c43349c1dd86795fc6f78},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_treatment_noncompliance_273acfea_v1,
  title = {treatment noncompliance},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/treatment-noncompliance/rdr_273acfead61c43349c1dd86795fc6f78/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake