Efficient and Globally Robust Causal Excursion Effect Estimation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Efficient and Globally Robust Causal Excursion Effect Estimation".
Jane: Here's the gist of the paper. Causal excursion effect (CEE) characterizes the effect of an intervention under policies that deviate from the experimental policy.
Tom: This paper proposes a semiparametric efficient estimation of CEE and derives a semiparametric efficiency bound with identity or log link functions in the context of micro-randomized trials.
Jane: Digital behavior change interventions are being developed in various domains including mobile health, education, and social justice.
Tom: CEE captures whether an intervention is effective and how the intervention effect changes over time or interacts with the individual’s context.
Paper summary: Tom: We talked about how this paper introduces CEE estimation, focusing on getting an efficient estimate with robustness built in. Basically, the thesis is that existing semiparametric models for CEE often have to choose between being very precise or being very robust to model errors.
Jane: That’s right, and they propose a new way using two-stage estimators where the first stage finds the best nuisance function and then estimates the causal parameter second, which achieves that theoretical efficiency bound.
Lu: They build on a general theory for globally robust Z-estimators, which is important because it covers both cross-fitted and non-cross-fitted nuisance parameters, giving us more flexibility in how we handle those background models.
Meng: I’m still focused on the practical side—how complex these two stages are. The paper suggests a framework that allows for extremely flexible nuisance parameter estimators, even if they converge at a very slow rate, which sounds like it could be computationally demanding to run reliably at scale.
Lalam: They establish the general asymptotic theory for this class of two-stage globally robust Z-estimators, showing consistency and asymptotic normality for the estimators under working assumptions. This theory is crucial because it proves that these estimators behave predictably in large samples.
Tom: So, while they are building a strong theoretical foundation proving they work asymptotically, the core claim is that their specific two-stage setup actually achieves the efficiency bound we were looking for in this field.
Jane: And when you look at the simulation studies, they validate this by showing a relative efficiency of up to one point two seven against existing methods on real data from a Drink Less MRT. That's a concrete number showing the benefit over what’s currently out there.
Lu: The paper also presents an implementation section detailing how to construct estimating functions for both identity and log link functions, which provides the technical blueprint for building these efficient estimators.
Meng: So, before we get into the real data application, I need to know if this method handles the non-Donsker convergence of those nuisance parameters mentioned in their theory gracefully when we try to deploy it on actual traffic data.
Lalam: They also present a reduced form of the estimating function that simplifies the implementation process, which is good for practical use because it cleans up the math before you have to deal with complex matrix inversions.
Conclusion: Tom: Looking at this paper, "Efficient and Globally Robust Causal Excursion Effect Estimation," what stands out is how they combine theoretical rigor with practical gain. They provide a semiparametric efficiency bound for CEE under identity or log link functions, which gives us a solid benchmark.
Jane: The implication for us is that we can move away from the trade-off between precision and robustness in CEE estimation and aim for methods that achieve this theoretical efficiency, especially when dealing with time-varying interventions.
Lu: What it means for the broader research area is that they establish a general asymptotic theory for globally robust Z-estimators, which could be applied to other settings where you need to use nuisance parameters to improve efficiency without needing full parameter identification.
Meng: For someone trying to deploy this in a startup environment, the main implication is that if we can achieve even one point two seven times better efficiency with less data, it directly translates into faster iteration cycles and lower costs for running experiments.
Lalam: The authors are showing that by nonparametrically estimating the optimal nuisance control model and optimally weighting across decision points, you can see a substantial efficiency gain compared to prior work.
Tom: So, it boils down to this: this method offers a way to improve the power of causal inference for adaptive interventions without getting bogged down in overly complex or fragile modeling assumptions.
ZhaoXi Cheng, Lauren Bell, Tianchen Qian
Department of Biostatistics, Harvard University · Leeds Institute of Clinical Trials Research, University of Leeds · Department of Statistics, University of California, Irvine
stat.ME
Submitted: 2023-11-28
Updated: 2026-10-30
Comments: 76 pages, 7 figures. Final version, published in the Electronic Journal of Statistics
Journal ref: Electron. J. Statist. 20(2) 5126-5199 (2026)
DOI: 10.1214/26-EJS2606
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 97/100
The gist: Causal excursion effect (CEE) characterizes the effect of an intervention under policies that deviate from the experimental policy.
Key concepts
- Causal Excursion Effect (CEE)
- CEE characterizes the effect of an intervention when the actual policy used differs from the experimental policy. It helps determine if an intervention is effective and how that effectiveness varies depending on the individual's context or time.
- Semiparametric Efficiency Bound
- This establishes a theoretical limit on how precise any estimation method can be, given certain assumptions. The authors derive this bound for CEE using identity or log link functions, providing a benchmark for the best possible estimation performance.
- Two-Stage Estimators
- The proposed method uses a two-stage approach: first, an optimal nuisance function is identified and fitted. Second, the causal parameter is estimated using this information. This structure allows for robust estimation even if the initial model assumptions are slightly incorrect.
- Globally Robust Z-Estimators
- These are estimators designed to be reliable even when the assumed models (nuisance parameters) are misspecified. The framework allows for flexible nuisance parameter estimators that can converge slowly, offering greater robustness than standard doubly robust methods.
Terminology
Summary
Causal excursion effect (CEE) characterizes the effect of an intervention under policies that deviate from the experimental policy. p. 0 uk
3Department of Statistics, University of California, Irvine, c
t. edu
Causal excursion effect (CEE) characterizes the effect of an intervention
under policies that deviate from the experimental policy. It is widely used to
study the effect of time-varying interventions that have the potential to be frequently adaptive, such as those delivered through smartphones. We study the
semiparametric efficient estimation of CEE and we derive a semiparametric
efficiency bound for CEE with identity or log link functions under working
assumptions, in the context of micro-randomized trials.Where the paper says it
Where the paper says it
Introduction and Motivation
Digital behavior change interventions are being developed in various domains including mobile health, education, and social justice. CEE captures whether an intervention is effective and how the intervention effect changes over time or interacts with the individual’s context. p. 1 , 2021a; Dempsey et al. CEE captures whether an intervention is effective and how the intervention effect
changes over time or interacts with the individual’s context, which are central questions to
the optimization of the interventions. CEE captures whether an intervention is effective and how the intervention effect
changes over time or interacts with the individual’s context, which are central questions to
the optimization of the interventions. Existing semiparametric models for CEE typically involve multiple nuisance function such as outcome regression models and stablizing weights
(Boruvka et al. , 2018; Qian et al. It is shown that the choice of the nuisance model
can substantially impact the efficiency of the CEE estimation (Qian et al. , 2018; Bell et al. Therefore, there is a critical need to develop more principled and efficient methods for CEE
inference in order to improve estimation precision and reduce the sample size required in an
MRT.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Estimation Methodology
The authors propose a class of two-stage estimators that achieve the efficiency bound, where in the first stage an optimal nuisance function is identified and fitted and in the second stage the causal parameter is estimated. They establish a general theory for globally robust Z-estimators with either cross-fitted or non-cross-fitted nuisance parameters. p. 0 We propose a class
of two-stage estimators that achieve the efficiency bound and are robust to
misspecified nuisance models. In deriving the asymptotic property of the estimators, we establish a general theory for globally robust Z-estimators with
either cross-fitted or non-cross-fitted nuisance parameters. We propose a class of two-stage estimators that achieve the efficiency bound, where in the first
stage an optimal nuisance function is identified and fitted and in the second stage the causal
parameter is estimated. We establish the consistency and asymptotic normality of the proposed estimators, and we show that the proposed estimator is robust to misspecified nuisance
models.Where the paper says it
Where the paper says it
Asymptotic Theory
The authors derive a semiparametric efficiency bound under working assumptions. They establish the consistency and asymptotic normality of the proposed estimators. p. 1 We propose a class of two-stage estimators that achieve the efficiency bound, where in the first
stage an optimal nuisance function is identified and fitted and in the second stage the causal
parameter is estimated. We establish the consistency and asymptotic normality of the proposed estimators, and we show that the proposed estimator is robust to misspecified nuisance
models. We propose a class
of two-stage estimators that achieve the efficiency bound and are robust to
misspecified nuisance models. In deriving the asymptotic property of the estimators, we establish a general theory for globally robust Z-estimators with
either cross-fitted or non-cross-fitted nuisance parameters.Where the paper says it
Where the paper says it
Simulation and Application
The authors conduct simulation studies to validate the efficiency gain of the proposed estimators compared to the ones in the literature. In real data application to the Drink Less MRT, they apply the proposed estimator to assess the effect of messages on user engagement, and it shows a relative efficiency of up to 1. p. 1 In simulation studies, the proposed estimator shows substantial efficiency gain compared
to existing ones in the literature with a relative efficiency of up to 3. In real data application
to the Drink Less MRT, we apply the proposed estimator to assess the effect of messages on
user engagement, and it shows a relative efficiency of up to 1. In real data application
to the Drink Less MRT, we apply the proposed estimator to assess the effect of messages on
user engagement, and it shows a relative efficiency of up to 1. 27 against existing ones.Where the paper says it
Where the paper says it
Technical Details
The technical derivation involves constructing estimating functions for both identity and log link functions. They define an efficient score function by projecting the general score function onto the score functions for the treatment assignment probability.
Conclusion and Efficiency Gains
The proposed estimators show substantial efficiency gain compared to existing ones in the literature with a relative efficiency of up to 3. The efficiency improvement comes from nonparametrically estimating an optimal nuisance control model and optimally weighting across decision points. p. 5 6
In both methods, a linear nuisance model bt(Ht)
Tα with ad hoc selected features is used to
fit the exposure-free outcome to reduce variance. The estimators we propose achieve higher
efficiency partly through identifying and fitting optimal nuisance parameters. , 2018; Bell et al. Therefore, there is a critical need to develop more principled and efficient methods for CEE
inference in order to improve estimation precision and reduce the sample size required in an
MRT.Where the paper says it
Where the paper says it
Implementation Details
The implementation details include defining matrix inverses as Moore-Penrose generalized inverses and providing analytical forms for the estimating functions. They also present a reduced form of the estimating function that simplifies the implementation of the estimator. p. 6 However, we found the orthogonalization step not straightforward here due to the non-nested {St}t∈[T]
. Furthermore,
we conjecture that the form of the general efficiency bound will involve across-time correlation for many (t, u) combinations, and the need to model these correlations will make the
corresponding estimator implementation challenging with small to moderate sample sizes.Where the paper says it
Robustness and Flexibility
The estimators are robust to mispecification of nuisance parameter models. The framework allows for extremely flexible nuisance parameter estimators that can be non-Donsker and converge at an arbitrarily slow rate. p. 1 We establish the consistency and asymptotic normality of the proposed estimators, and we show that the proposed estimator is robust to misspecified nuisance
models. In addition, we show that due to its global robustness property, which we define in
Section 5, the proposed estimator allows extremely flexible nuisance parameter estimators
that can be non-Donsker and converge at an arbitrarily slow rate. Furthermore, there is
no requirement on the rate of convergence for the nuisance parameter estimators (as long as
they converge at oP (1)), which allows a wider range of nonparametric methods to be used
such as kernel or nearest neighbor estimators. This is less stringent than the literature on
doubly-robust average treatment effect estimation, where a faster than n
−1/4
rate is typically
required for nuisance parameter estimators (Van der Vaart, 2000; Kennedy, 2016).Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Robust Causal Excursion Effect Estimation: The AI system can estimate causal effects under policies that deviate from experimental policies in micro-randomized trials by proposing a class of two-stage estimators that achieve semiparametric efficiency bounds and are robust to misspecified nuisance models.
-
Flexible Nuisance Model Fitting: The system can fit optimal nuisance functions using flexible, nonparametric models such as generalized additive models, random forests, or stacking ensembles for estimating the effect of time varying interventions.
-
Cross-Fitting for Robust Inference: The AI system can incorporate cross-fitting techniques into its estimators to improve robustness and confidence interval coverage without requiring the Donsker condition on non-Donsker nuisance parameter estimators.
-
Efficiency Gain in Intervention Optimization: By using these methods, the AI system can achieve substantial efficiency gains compared to existing estimators in literature, potentially reducing the sample size required for analyzing time varying effect moderation with fewer samples.
-
Handling Complex Outcome Types: The framework is unified to handle continuous, binary (logistic), and count outcomes for Causal Excursion Effect estimation using either identity or log link functions.
Abstract
Causal excursion effect (CEE) characterizes the effect of an intervention under policies that deviate from the experimental policy. It is widely used to study the effect of time-varying interventions that have the potential to be frequently adaptive, such as those delivered through smartphones. We study the semiparametric efficient estimation of CEE and we derive a semiparametric efficiency bound for CEE under working assumptions, in the context of micro-randomized trials. We propose a class of two-stage estimators that achieve the efficiency bound and are robust to misspecified nuisance models. In deriving the asymptotic property of the estimators, we establish a collection of easy-to-use theorems for globally robust Z-estimators with either cross-fitted or non-cross-fitted nuisance parameters that are applicable beyond the proposed estimators for CEE. We demonstrate efficiency gain of the proposed estimator compared to existing ones through simulations and a real data application using the Drink Less micro-randomized trial.
Sources
- Incorporating nonparametric methods for estimating causal excursion effects in mobile health with zero-inflated count outcomes
- How estimating nuisance parameters can reduce the variance (with consistent variance estimation)
- A Meta-Learning Method for Estimation of Causal Excursion Effects to Assess Time-Varying Moderation
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes