Doubly Robust Estimation with Stabilized Weights for Binary Proximal Outcomes in Micro-Randomized Trials

arXiv:2510.08359 · stat.ME, math.ST, stat.TH · Submitted 2025-10-09 · Read on arXiv

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: "Doubly Robust Estimation with Stabilized Weights for Binary Proximal Outcomes in Micro-Randomized Trials".

Jane: Here's the gist of the paper. Micro-randomized trials (MRTs) are increasingly used to evaluate mobile health interventions with binary proximal outcomes because standard inverse probability weighting (IPW) estimators are unbiased but unstable in small samples or under extreme randomization.

Tom: Micro-randomized trials (MRTs) have emerged as a central experimental design for evaluating just-in-time adaptive interventions (JITAIs) in mobile health (mHealth) studies.

Jane: At each decision time, participants are randomized to treatment or control, and proximal outcomes are observed over short windows.

Tom: This sequential structure enables the estimation of excursion effects, the causal effect of treatment delivery conditional on availability and context.

Paper summary: Tom: So we're talking about this paper today called "Doubly Robust Estimation with Stabilized Weights for Binary Proximal Outcomes in Micro-Randomized Trials". The core idea here is that micro-randomized trials are super useful for studying mobile health interventions because you can randomize participants at every decision time and see outcomes over short windows.

Jane: That's right, and the paper points out that standard methods, like inverse probability weighting or IPW, are unbiased but they get really unstable when you have small sample sizes or when the randomization is extreme.

Lu: The paper addresses this by proposing a new approach called DR-EMEE which combines inverse probability weighting with outcome regression, and they're adding some weight stabilization and truncation to keep things consistent.

Meng: So it's trying to fix the instability of the per-decision IPW estimator, which they call pd-IPW, that they say is unbiased but highly variable. What does this mean for us when we try to run these models at scale?

Lalam: It means they are trying to make sure that no matter if you're wrong about the treatment model or the outcome model, your final estimate of the mean excursion effect is still valid.

Tom: Exactly, and they claim this DR-EMEE estimator proves double robustness, which means it stays consistent if either your treatment model or your outcome model is specified correctly. That sounds pretty solid.

Jane: But they also focus on showing that this new method has asymptotic efficiency, meaning as the sample size gets bigger, it gets as good as the best possible estimator can get. How does that compare to what we usually have?

Lu: The paper shows that this approach achieves higher relative efficiency in simulation studies compared to standard methods like IPW and EMEE in balanced conditions. They even show a finite-sample variance correction is needed for small sample sizes, which they define as n over n minus p sigma squared.

Meng: That's interesting about the variance correction because I've seen how much standard sandwich variance estimators can underestimate variability when you only have a few clusters, and this seems to address that practical problem. What about when the treatment assignment probabilities are really extreme?

Lalam: They show that weight stabilization and truncation across different randomization probabilities progressively reduce variance and eliminate those heavy tails in the estimates. This is important because in mHealth trials, you often have very rare treatments, like a probability of zero point one or zero point nine.

Tom: It’s about making sure those extreme weights don't blow up the estimate, which they say standard IPW does. The simulation results show DR-EMEE consistently achieves higher relative efficiency and lower RMSE across all tested scenarios, including those with extreme randomization probabilities.

Jane: So what does this actually mean for someone who just listens to the show? It means that when you're looking at data from things like heart rate monitors or activity trackers, this method gives you a more reliable estimate of how much a specific intervention actually causes an effect.

Lu: For the person listening, it means they can trust the numbers they get from these complex mobile health trials even when the study design is tricky and the sample size is modest. They aren't just getting an estimate; they're getting one that has been proven to be robust under different kinds of model misspecification.

Meng: From a deployment standpoint, if we were trying to use this in a production setting, we’d want to know what breaks at scale and how much computational overhead this stabilization adds compared to just running the raw pd-IPW.

Lalam: The paper shows that even with these added steps of stabilization and truncation, DR-EMEE produces stable estimates with substantially narrower intervals than both IPW and EMEE on datasets like HeartSteps I and PAMAP2.

Tom: So to wrap up this part of the paper, it's about giving us a principled framework, the DR-EMEE estimator, for estimating causal effects in mobile health that is valid even in small samples or under extreme treatment probabilities. It’s a new way to handle those messy real-world trial designs.

Jane: The authors Jinho Cha and Eunchan Cha are focused on proving this consistency and showing its efficiency gains over existing doubly robust methods, especially where they've had less work done before in the micro-randomized trial setting. It's a big step for applying these powerful causal inference tools to digital health.

Lu: The implication here is that we can start designing adaptive interventions more reliably because we have a method that handles the inherent uncertainty of sequential decision-making better than what we had previously.

Meng: It moves us from just having an estimate to having a statistically sound foundation for making decisions about how those mobile health tools are actually working in practice.

Lalam: The paper establishes DR-EMEE as a reliable and scalable framework for advancing causal inference in digital health, which is what I’m pretty sure this work is really about.

Conclusion: Tom: So we're wrapping up our look at "Doubly Robust Estimation with Stabilized Weights for Binary Proximal Outcomes in Micro-Randomized Trials."

Jane: Basically, this paper is about a new way to estimate effects in mobile health studies that have been randomized frequently, which they call micro-randomized trials.

Lu: The authors are focusing on the doubly robust estimation method, specifically how it handles those binary outcomes you see over short time windows.

Meng: It tackles the problem of instability that standard inverse probability weighting estimators run into when your sample size is small or your treatment probabilities are very extreme.

Lalam: They introduce this DR-EMEE estimator, which they claim gives you consistent results even if you're slightly off on either the treatment model or the outcome model.

Tom: So, what does that mean in practice for someone trying to use these mobile health interventions? It means a more reliable estimate of whether an intervention actually works.

Jane: It means for people relying on heart rate monitors or activity trackers, they can trust the numbers better than before, especially when the study design is complex.

Lu: The paper shows that this method has better efficiency in simulations compared to older techniques like EMEE and IPW when things are balanced.

Meng: I'm interested in how this affects deployment; if we're shipping models into production, we need to know what the computational cost is for using these stabilized weights.

Tom: That’s a fair point, Meng, because the stabilization adds some complexity to the math and requires more careful tuning.

Jane: But it seems like that extra effort pays off in terms of getting a more accurate picture of reality, which is what matters for policy and clinical decisions.

Lu: It moves us closer to having a principled framework for digital health research where we can be more confident in our causal claims even with limited data.

Lalam: They also point out that they've got a trick for small sample sizes, like using n over n minus p sigma squared to fix the variance estimates when the number of clusters is low.

Tom: So, it’s not just about getting a better estimate; it’s about making sure that estimate doesn't have wildly unstable confidence intervals when you're dealing with small numbers.

Jane: Exactly, and that makes a huge difference for real-world data where you don't always get huge samples.

Lu: It also opens the door to thinking about how this framework might integrate with adaptive treatment allocation strategies down the road.

Meng: I'll have to see if those integration points are actually practical to implement without making the system too heavy for a production environment.

Jinho Cha, Eunchan Cha

Department of Computer Programming, Gwinnett Technical College · Department of Biology, Georgia Institute of Technology

stat.ME, math.ST, stat.TH

Submitted: 2025-10-09

Updated: 2026-10-27

Comments: arXiv admin note: This submission has been withdrawn because it does not meet arXiv's research content quality standards

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 93/100

The gist: Micro-randomized trials (MRTs) are increasingly used to evaluate mobile health interventions with binary proximal outcomes because standard inverse probability weighting (IPW) estimators are unbiased…

Key concepts

Micro-Randomized Trials (MRTs)
MRTs are an experimental design for mobile health studies where participants are randomized to treatment or control at each decision point. This allows researchers to observe short-term outcomes and estimate the causal effect of treatment delivery based on context and availability, which is crucial for adaptive interventions.
Doubly Robust (DR) Estimation
This statistical method combines inverse probability weighting (IPW) with outcome regression. It achieves consistency if either the treatment model or the outcome model is correctly specified, offering a robust way to estimate causal effects when one of the two models might be wrong.
Excursion Effect
The excursion effect is a key measure in MRTs that estimates how much the treatment delivery changes an outcome over a short window. Reliable estimation of this effect is vital for designing adaptive interventions and informing digital health policies.
Weight Stabilization and Truncation
These techniques are used to stabilize the highly variable weights produced by inverse probability weighting, especially in small samples. Truncation further reduces variance by limiting the influence of extreme weights, leading to more stable and reliable estimates.

Terminology

Summary

Micro-randomized trials (MRTs) are increasingly used to evaluate mobile health interventions with binary proximal outcomes because standard inverse probability weighting (IPW) estimators are unbiased but unstable in small samples or under extreme randomization. p. 0

Where the paper says it

edu FOR PUBLISHER ONLY Submitted 30 September ; Revised 15 October ; revised version accepted 25 October ABSTRACT Micro-randomized trials (MRTs) are increasingly used to evaluate mobile health interventions with binary proximal outcomes. Standard inverse probability weighting (IPW) estimators are unbiased but unstable in small samples or under extreme randomization.

Motivation and Context

Micro-randomized trials (MRTs) have emerged as a central experimental design for evaluating just-in-time adaptive interventions (JITAIs) in mobile health (mHealth) studies. At each decision time, participants are randomized to treatment or control, and proximal outcomes are observed over short windows. p. 0

Where the paper says it

, 2018, Qian et al. At each decision time, participants are randomized to treatment or control, and proximal outcomes are observed over short windows.

This sequential structure enables the estimation of excursion effects, the causal effect of treatment delivery conditional on availability and context. p. 0
Where the paper says it

At each decision time, participants are randomized to treatment or control, and proximal outcomes are observed over short windows. This sequential structure enables the estimation of excursion effects, the causal effect of treatment delivery conditional on availability and context [Liao et al.

Reliable estimation of excursion effects is crucial for informing the design of adaptive interventions, clinical decision support, and digital health policy. p. 0
Where the paper says it

, 2020, Nahum-Shani et al. Reliable estimation of excursion effects is crucial for informing the design of adaptive interventions, clinical decision support, and digital health policy [Walton et al.

Gaps in Existing Literature

Doubly robust (DR) methods, which combine IPW with outcome regression, achieve consistency if either the treatment or outcome model is correctly specified. However, in the MRT context, methodological development has been more limited. p. 1

Where the paper says it

, 2018, Kennedy, 2022, Xu et al. However, in the MRT context, methodological development has been more limited.

The per-decision IPW (pd-IPW) estimator is unbiased but highly variable, and the estimated mean excursion effect (EMEE) improves efficiency but is not doubly robust and remains sensitive to model misspecification. p. 1
Where the paper says it

, 2022] is unbiased but highly variable, and the estimated mean excursion effect (EMEE) [Liao et al. , 2020] improves efficiency but is not doubly robust and remains sensitive to model misspecification.

To our knowledge, no prior work has formally established a doubly robust excursion effect estimator for MRTs under binary proximal outcomes, nor characterized its small-sample behavior. p. 1
Where the paper says it

, 2011], their integration with DR estimators in MRTs with binary proximal outcomes has not been rigorously developed or theoretically analyzed. To our knowledge, no prior work has formally established a doubly robust excursion effect estimator for MRTs under binary proximal outcomes, nor characterized its small-sample behavior.

Proposed Methodology

The proposed DR-EMEE estimator integrates the strengths of inverse probability weighting and outcome regression while mitigating instability through weight stabilization and truncation. The estimator is formulated as an augmented IPW form, which identifies the efficient influence function whenever either the treatment model or the outcome model is correctly specified. p. 4

Where the paper says it

, 1994, Bang and Robins, 2005, Tsiatis, 2006], and it identifies the efficient influence function whenever either the treatment model pt(H) or the outcome model ma(H) is correctly specified. Estimating Equation Structure The DR-EMEE estimator (ˆα, βˆ) solves the sample moment condition Uˆ(α, β) = 1 n Xn i=1 XT t=1 With Yit,∆ − mα(Hit) − AitS ⊤ itβ i g(Hit)(Ait − p˜t(Sit)) = 0, (3.

The estimating equation solves a sample moment condition that ensures double robustness by preserving consistency if either the treatment model or outcome model is correct. p. 4
Where the paper says it

, 1994, Bang and Robins, 2005, Tsiatis, 2006], and it identifies the efficient influence function whenever either the treatment model pt(H) or the outcome model ma(H) is correctly specified. Estimating Equation Structure The DR-EMEE estimator (ˆα, βˆ) solves the sample moment condition Uˆ(α, β) = 1 n Xn i=1 XT t=1 With Yit,∆ − mα(Hit) − AitS ⊤ itβ i g(Hit)(Ait − p˜t(Sit)) = 0, (3.

Simulation and Empirical Validation

In simulation studies, DR-EMEE consistently achieved the smallest RMSE, the highest relative efficiency, and coverage close to nominal across all settings. Empirical applications to HeartSteps I, PAMAP2, and mHealth datasets confirmed these advantages in practice. p. 0

Where the paper says it

In simulations, DR-EMEE reduces root mean squared error, improves coverage, and achieves up to two-fold efficiency gains over IPW and 5–10% over EMEE. Applications to HeartSteps, PAMAP2, and mHealth datasets confirm stable and efficient inference across both randomized and observational settings.

The results show that DR-EMEE consistently achieved higher relative efficiency (RE) and lower RMSE across all tested conditions. p. 3
Where the paper says it

By combining stabilized and truncated per-decision weighting with outcome augmentation, DR-EMEE integrates the strengths of variance reduction and double robustness into a single estimator. This ensures small-sample stability, protection against model misspecification, and asymptotic efficiency—properties not jointly achieved by prior estimators.

Conclusion and Future Directions

The work introduced the doubly-robust estimated mean excursion effect (DR-EMEE) as a principled framework for estimating causal effects in mobile health. This method delivers valid inference even in small samples and under extreme treatment probabilities. p. 3

Where the paper says it

, 2018, Kennedy, 2020, 2022, Rotnitzky and Smucler, 2020, Xu et al. Yet, despite this extensive progress, no single method simultaneously addresses the three central challenges of MRTs with binary proximal outcomes: (i) small-sample instability due to extreme treatment probabilities, (ii) sensitivity to misspecification of nuisance models, and (iii) efficiency loss relative to semiparametric benchmarks.

Future directions include integration with adaptive treatment allocation and reinforcement learning algorithms and extension to multivariate or longitudinal proximal outcomes. p. 1
Where the paper says it

Section 7 concludes with implications, limitations, and directions for future research. Excursion Effects in MRTs Micro-randomized trials (MRTs) have become a widely adopted design for evaluating just-in-time adaptive interventions (JITAIs) in mobile health, behavioral science, and chronic disease management [Collins, 2014, Klasnja et al.

The study establishes DR-EMEE as a reliable and scalable framework for advancing causal inference in digital health. p. 3
Where the paper says it

This ensures small-sample stability, protection against model misspecification, and asymptotic efficiency—properties not jointly achieved by prior estimators. Thus, DR-EMEE not only advances the methodological frontier of causal inference but also strengthens the applied toolkit for analyzing MRTs and designing evidence-based digital health interventions.

Additional Insights

One unexpected but important insight from our empirical analyses is that the performance of variance estimators plays a decisive role in small-sample settings. Standard sandwich variance estimators can severely underestimate variability when the number of clusters is small. p. 8

Where the paper says it

, 2018, Kennedy, 2020, 2022] and can be satisfied using cross-fitting and flexible ML learners. Small-sample correction In finite samples, the naive sandwich variance estimator tends to underestimate the true variance.

The study demonstrated that reliable inference in mobile health trials requires variance estimation strategies tailored to small-sample designs. p. 0
Where the paper says it

Applications to HeartSteps, PAMAP2, and mHealth datasets confirm stable and efficient inference across both randomized and observational settings. KEYWORDS: Doubly robust; micro-randomized trial; binary proximal outcome; weight stabilization; causal inference; mHealth INTRODUCTION Motivation and Context Micro-randomized trials (MRTs) have emerged as a central experimental design for evaluating just-in-time adaptive interventions (JITAIs) in mobile health (mHealth) studies [Klasnja et al.

Practical Note

The corrected sandwich variance estimator, defined as n/n−p Σbsand, improves small-sample coverage while remaining consistent as n → ∞. This adjustment mirrors the classical regression correction for small-sample covariance estimation. p. 8

Where the paper says it

, 2018, Kennedy, 2020, 2022] and can be satisfied using cross-fitting and flexible ML learners. Small-sample correction In finite samples, the naive sandwich variance estimator tends to underestimate the true variance.

The findings highlight that stabilized DR-EMEE can serve as a practical and theoretically grounded solution for small-sample, weak-overlap sensor data settings. p. 1
Where the paper says it

(iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting. (iv) We empirically validate the proposed estimator through extensive simulation studies and real-data applications, including the Drink Less and HeartSteps I MRTs, as well as wearable sensor datasets such as PAMAP2 and mHealth.

Relation to Existing Methods

Recent advances in causal inference such as overlap weighting (Li et al., 2018), targeted maximum likelihood estimation (TMLE) (van der Laan and Rubin, 2006), and double machine learning (DML) (Chernozhukov et al. p. 2

Where the paper says it

, 2002, Bang and Robins, 2005, Funk et al. Targeted maximum likelihood estimation (TMLE) further improved efficiency and provided a general template for constructing DR estimators in complex settings [van der Laan and Rubin, 2006, van der Laan and Rose, 2011, Tsiatis, 2006].

, 2018) have established powerful frameworks for achieving efficiency and robustness in large-sample settings. p. 1
Where the paper says it

(ii) We introduce a stabilized and truncated weighting scheme that improves finite-sample stability, particularly under small sample sizes and extreme randomization probabilities that are common in practice. (iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting.

However, these methods rely heavily on either sufficient overlap or asymptotic approximations, and their performance in small-sample, weak-overlap sensor datasets remains less well understood. p. 1
Where the paper says it

(iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting. (iv) We empirically validate the proposed estimator through extensive simulation studies and real-data applications, including the Drink Less and HeartSteps I MRTs, as well as wearable sensor datasets such as PAMAP2 and mHealth.

The empirical results suggest that under precisely this challenging regime, stabilized DR-EMEE provides a more reliable alternative.

Software and Reproducibility

Software in the form of R code, together with a sample input data set and complete documentation is available on request from the corresponding author. By uniting theoretical rigor with empirical validation across trial-like and observational settings, this work establishes DR-EMEE as a reliable and scalable framework for advancing causal inference in digital health. p. 0

Where the paper says it

Applications to HeartSteps, PAMAP2, and mHealth datasets confirm stable and efficient inference across both randomized and observational settings. KEYWORDS: Doubly robust; micro-randomized trial; binary proximal outcome; weight stabilization; causal inference; mHealth INTRODUCTION Motivation and Context Micro-randomized trials (MRTs) have emerged as a central experimental design for evaluating just-in-time adaptive interventions (JITAIs) in mobile health (mHealth) studies [Klasnja et al.

References

The paper references numerous foundational works in causal inference, including Robins et al. (1994), Hernán and Robins (2020), and Chernozhukov et al. p. 0

Where the paper says it

Under these conditions, classical inverse probability weighting (IPW) [Robins et al. , 1994, Hern´an and Robins, 2020] yields unbiased but highly variable estimators, leading to unstable inference and undercoverage of confidence intervals.

(2018). The list includes key papers on propensity score weighting, doubly robust estimation, and machine learning applications in causal inference. p. 0
Where the paper says it

Applications to HeartSteps, PAMAP2, and mHealth datasets confirm stable and efficient inference across both randomized and observational settings. KEYWORDS: Doubly robust; micro-randomized trial; binary proximal outcome; weight stabilization; causal inference; mHealth INTRODUCTION Motivation and Context Micro-randomized trials (MRTs) have emerged as a central experimental design for evaluating just-in-time adaptive interventions (JITAIs) in mobile health (mHealth) studies [Klasnja et al.

The references cover topics ranging from foundational causal theory to recent advances in digital health interventions. p. 3
Where the paper says it

This ensures small-sample stability, protection against model misspecification, and asymptotic efficiency—properties not jointly achieved by prior estimators. Thus, DR-EMEE not only advances the methodological frontier of causal inference but also strengthens the applied toolkit for analyzing MRTs and designing evidence-based digital health interventions.

The references also include works on small-sample adjustments for wald-type tests using sandwich estimators. p. 8
Where the paper says it

Small-sample corrections of this type are standard in generalized estimating equations and survey sampling [Kauermann and Carroll, 2001, Mancl and DeRouen, 2001, Fay and Graubard, 2001]. HeartSteps I: Small-sample Robustness In the HeartSteps I MRT analysis, we observed that the naive variance estimator substantially under-covered, whereas the corrected sandwich variance restored coverage close to the nominal level.

Appendix

This appendix provides proofs of all theoretical results stated in the main text, organized according to the order of appearance. It includes extended simulation studies, methodological comparisons, and proofs for identification, consistency, asymptotic normality, truncation analysis, and efficiency. p. 1

Where the paper says it

(iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting. (iv) We empirically validate the proposed estimator through extensive simulation studies and real-data applications, including the Drink Less and HeartSteps I MRTs, as well as wearable sensor datasets such as PAMAP2 and mHealth.

The appendix also details real data applications like PAMAP2 and mHealth datasets. p. 1
Where the paper says it

(iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting. (iv) We empirically validate the proposed estimator through extensive simulation studies and real-data applications, including the Drink Less and HeartSteps I MRTs, as well as wearable sensor datasets such as PAMAP2 and mHealth.

The appendix provides comprehensive reproducibility details for all simulation code and processed datasets. p. 1
Where the paper says it

(iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting. (iv) We empirically validate the proposed estimator through extensive simulation studies and real-data applications, including the Drink Less and HeartSteps I MRTs, as well as wearable sensor datasets such as PAMAP2 and mHealth.

Table 1

Table 1 summarizes the spectrum of IPW-related methods and their properties, highlighting DR-EMEE as a new proposal. It provides a cross-mapping of all theoretical results to supplementary proof sections and main-text subsections. p. 1

Where the paper says it

(ii) We introduce a stabilized and truncated weighting scheme that improves finite-sample stability, particularly under small sample sizes and extreme randomization probabilities that are common in practice. (iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting.

Table 2

Table 2 presents finite-sample simulation results for n = 100, pt = 0.5, comparing IPW, EMEE, and DR-EMEE. p. 0

Where the paper says it

Standard inverse probability weighting (IPW) estimators are unbiased but unstable in small samples or under extreme randomization. Estimated mean excursion effect (EMEE) improves efficiency but lacks double robustness.

It shows that DR-EMEE achieved substantially improved performance relative to IPW in balanced conditions.

Table 4

Table 4 summarizes the estimated excursion effects across four scenarios (S1–S4) for the PAMAP2 dataset. It shows that DR-EMEE consistently achieved tighter confidence intervals and greater efficiency across S2–S4. p. 0

Where the paper says it

Under these conditions, classical inverse probability weighting (IPW) [Robins et al. , 1994, Hern´an and Robins, 2020] yields unbiased but highly variable estimators, leading to unstable inference and undercoverage of confidence intervals.

Table 7

Table 7 summarizes extended simulation results across n ∈ three sample sizes and pt ∈ three randomization probabilities. It shows that DR-EMEE consistently achieved higher relative efficiency (RE) and lower RMSE across all scenarios. p. 3

Where the paper says it

By combining stabilized and truncated per-decision weighting with outcome augmentation, DR-EMEE integrates the strengths of variance reduction and double robustness into a single estimator. This ensures small-sample stability, protection against model misspecification, and asymptotic efficiency—properties not jointly achieved by prior estimators.

p. 3
Where the paper says it

By combining stabilized and truncated per-decision weighting with outcome augmentation, DR-EMEE integrates the strengths of variance reduction and double robustness into a single estimator. This ensures small-sample stability, protection against model misspecification, and asymptotic efficiency—properties not jointly achieved by prior estimators.

Table 9

Table 9 provides an extended overview of inverse probability weighting (IPW) and doubly-robust estimators, highlighting the distinct contributions of DR-EMEE. It includes a cross-mapping of all theoretical results to supplementary proof sections and main-text subsections. p. 1

Where the paper says it

(ii) We introduce a stabilized and truncated weighting scheme that improves finite-sample stability, particularly under small sample sizes and extreme randomization probabilities that are common in practice. (iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting.

Table 10

Table 10 cross-maps all theoretical results to supplementary proof sections and main-text subsections. It details the proofs for identification, double robustness property, consistency and CLT, small-sample properties, truncation analysis, efficiency comparison, and projection-based improvement. p. 4

Where the paper says it

By combining stabilization/truncation of weights with augmentation, DR-EMEE achieves double robustness, finite-sample stability, and asymptotic efficiency. Influence Function Representation Under standard causal assumptions (consistency, sequential ignorability, positivity), the DR-EMEE estimator can be expressed in an augmented IPW form: τˆDR = Pn h {mˆ 1(H) − mˆ 0(H)} + A pt(H) {Yt,∆ − mˆ 1(H)} − 1−A 1−pt(H) {Yt,∆ − mˆ 0(H)} i , (3.

Table 11

Table 11 summarizes the performance metrics across four key dimensions: bias as the deviation of the Monte Carlo mean of βˆ from the true effect β∗>, accuracy summarized by RMSE, variability assessed by comparing empirical variance to mean estimated SE, interval validity by computing coverage. It also defines relative efficiency (RE) as MSE(βˆIPW). p. 2

Where the paper says it

, 2022], but it inherits the high variance of classical IPW. The estimated mean excursion effect (EMEE) augments pd-IPW with outcome regression, achieving efficiency gains relative to pd-IPW [Liao et al.

Table 12

Table 12 shows the bias–variance–MSE tradeoff curves, highlighting how moderate truncation minimizes MSE, confirming the finite-sample stability gains that motivate the DR-EMEE estimator. It also shows that raw IPW weights are highly unstable when treatment assignment is rare. p. 2

Where the paper says it

By re-weighting observed outcomes according to treatment probabilities, IPW yields unbiased estimators under sequential randomization. However, a long-standing concern is instability: when treatment probabilities are small or highly imbalanced, weights become extremely variable, leading to finite-sample bias, inflated variance, and poor coverage of confidence intervals [Kang and Schafer, 2007, Cole and Hern´an, 2008, Lee et al.

Table 16

Table 16 summarizes the estimates for the mHealth dataset, showing that DR-EMEE produces stable estimates with substantially narrower intervals than IPW and EMEE. It confirms the robustness of DR-EMEE in real-world activity monitoring data. p. 1

Where the paper says it

(iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting. (iv) We empirically validate the proposed estimator through extensive simulation studies and real-data applications, including the Drink Less and HeartSteps I MRTs, as well as wearable sensor datasets such as PAMAP2 and mHealth.

p. 1
Where the paper says it

(iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting. (iv) We empirically validate the proposed estimator through extensive simulation studies and real-data applications, including the Drink Less and HeartSteps I MRTs, as well as wearable sensor datasets such as PAMAP2 and mHealth.

Table 7 reports estimates across various sample sizes and randomization probabilities, showing that DR-EMEE consistently achieved higher relative efficiency (RE) and lower RMSE across all scenarios. It demonstrates the incremental efficiency gain of DR-EMEE over EMEE was uniformly positive. p. 5

Where the paper says it

Proof is deferred to Supplement 9. This efficiency gain is consistent with the general theory of augmented estimators [Bang and Robins, 2005, Funk et al.

p. 5
Where the paper says it

Proof is deferred to Supplement 9. This efficiency gain is consistent with the general theory of augmented estimators [Bang and Robins, 2005, Funk et al.

Figure 6

Figure 6 summarizes the preprocessed PAMAP2 dataset, including subject-level sample counts, outcome and treatment distributions, and heart rate stratification. It illustrates the data structure used for empirical analysis. p. 8

Where the paper says it

HeartSteps I: Small-sample Robustness In the HeartSteps I MRT analysis, we observed that the naive variance estimator substantially under-covered, whereas the corrected sandwich variance restored coverage close to the nominal level. This illustrates the practical importance of the small-sample correction discussed in Proposition 12.

p. 8
Where the paper says it

HeartSteps I: Small-sample Robustness In the HeartSteps I MRT analysis, we observed that the naive variance estimator substantially under-covered, whereas the corrected sandwich variance restored coverage close to the nominal level. This illustrates the practical importance of the small-sample correction discussed in Proposition 12.

Figure 7

Figure 7 presents descriptive summaries of the preprocessed mHealth dataset, including subject-level sample counts and stratification of chest accelerometer measurements by outcome. It illustrates the data structure used for empirical analysis.

Figure 4

Figure 4 shows a forest plot of PAMAP2 dataset results, where DR-EMEE consistently exhibits tighter confidence intervals and reduced variance relative to IPW and EMEE.

Figure 5

Figure 5 presents a forest plot of the mHealth dataset results, where DR-EMEE produces stable estimates with substantially narrower intervals than IPW and EMEE. It confirms the robustness of DR-EMEE in real-world activity monitoring data.

Figure 3

Figure 3 summarizes simulation results across sample sizes and randomization probabilities, showing that DR-EMEE consistently achieved higher relative efficiency and lower RMSE while maintaining valid coverage across all scenarios. It demonstrates the incremental efficiency gain of DR-EMEE over EMEE was uniformly positive.

Figure 2

Figure 2 shows that weight stabilization and truncation across randomization probabilities progressively reduce variance and eliminate heavy tails. The bias–variance–MSE tradeoff curves highlight how moderate truncation minimizes MSE, demonstrating the finite-sample stability advantage of DR-EMEE across diverse scenarios. p. 6

Where the paper says it

Thus the mean squared error (MSE) admits the decomposition MSEτˆ (L,U) = Bias2 τˆ (L,U) + Varτˆ (L,U) , (3. 11) highlighting an explicit bias–variance tradeoff: moderate truncation improves MSE, especially in small samples [Li et al.

Figure 3 illustrates that DR-EMEE consistently achieved higher relative efficiency (panel a), lower RMSE (panel b), and coverage closer to nominal (panels c–f) across all scenarios. This demonstrates the incremental efficiency gain of DR-EMEE over EMEE was uniformly positive. p. 5

Where the paper says it

Proof is deferred to Supplement 9. This efficiency gain is consistent with the general theory of augmented estimators [Bang and Robins, 2005, Funk et al.

Figure 4 shows a forest plot of PAMAP2 dataset results, where DR-EMEE consistently exhibits tighter confidence intervals and reduced variance relative to IPW and EMEE. It highlights the robustness in real-world activity monitoring data. p. 1

Where the paper says it

(iii) We establish theoretical guarantees for DR-EMEE, including double robustness, asymptotic normality, and semiparametric efficiency, and extend these results to settings with machine learning-based nuisance estimation via cross-fitting. (iv) We empirically validate the proposed estimator through extensive simulation studies and real-data applications, including the Drink Less and HeartSteps I MRTs, as well as wearable sensor datasets such as PAMAP2 and mHealth.

Table 6

Table 6 reports truncation sensitivity under nonlinear misspecification, showing that DR-EMEE maintains stability across scenarios. It shows that DR-EMEE consistently achieved higher relative efficiency (RE) and lower RMSE across all scenarios.

Table 3

Table 3 compares coverage before and after small-sample sandwich correction, showing that the corrected estimator improves inference reliability in practice. It also shows that DR-EMEE produces stable estimates with substantially narrower intervals than IPW and EMEE. p. 0

Where the paper says it

In simulations, DR-EMEE reduces root mean squared error, improves coverage, and achieves up to two-fold efficiency gains over IPW and 5–10% over EMEE. Applications to HeartSteps, PAMAP2, and mHealth datasets confirm stable and efficient inference across both randomized and observational settings.

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

  1. Stabilized per-decision IPW estimation can be used to estimate excursion effects in mobile health interventions, specifically reducing variance when treatment assignment probabilities are extreme or sample sizes are small. This allows AI systems to make more reliable decisions in real-world settings like wearable sensor data where randomization probabilities might be near zero or one.

  2. The doubly robust estimated mean excursion effect (DR-EMEE) provides a consistent estimator for the true excursion effect parameter even if either the treatment model or the outcome model is misspecified. This means AI systems can be deployed in complex environments where underlying behavior models are imperfect, leading to more trustworthy causal inferences.

  3. The DR-EMEE estimator achieves asymptotic efficiency and maintains coverage close to nominal levels, even under challenging conditions such as small sample sizes and extreme randomization probabilities. This allows AI systems to generate confidence intervals that accurately reflect uncertainty in resource-constrained or highly imbalanced data settings.

  4. Projection-based DR-EMEE2 can be employed to further reduce variance by projecting the influence function onto the score space of the treatment model, resulting in an estimator with weakly smaller variance than DR-EMEE. This offers a path for AI systems to achieve even tighter uncertainty quantification in their causal effect estimates.

  5. Small-sample sandwich correction can be applied to improve coverage properties when estimating excursion effects from small clusters, such as those found in the HeartSteps I trial or mHealth datasets. This ensures that the uncertainty quantification provided by the AI system is reliable and not overly optimistic under limited data conditions.

Abstract

Micro-randomized trials (MRTs) are increasingly used to evaluate mobile health interventions with binary proximal outcomes. Standard inverse probability weighting (IPW) estimators are unbiased but unstable in small samples or under extreme randomization. Estimated mean excursion effect (EMEE) improves efficiency but lacks double robustness. We propose a doubly robust EMEE (DR-EMEE) with stabilized and truncated weights, combining per-decision IPW and outcome regression. We prove double robustness, asymptotic efficiency, and provide finite-sample variance corrections, with extensions to machine learning nuisance estimators. In simulations, DR-EMEE reduces root mean squared error, improves coverage, and achieves up to twofold efficiency gains over IPW and five to ten percent over EMEE. Applications to HeartSteps, PAMAP2, and mHealth datasets confirm stable and efficient inference across both randomized and observational settings.

Related papers