Distal Causal Excursion Effects: Modeling Long-Term Effects of Time-Varying Treatments in Micro-Randomized Trials

Tianchen Qian · 2025

Read Distal Causal Excursion Effects: Modeling Long-Term Effects of Time-Varying Treatments in Micro-Randomized Trials by Tianchen Qian (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-13500.txt.

Introduction

Qian builds a new causal measure for trials that nudge people many times, such as phone prompts to walk more. In these micro-randomized trials, each person is randomized hundreds of times, and the real goal is long-term habit change, not the next hour's steps. Existing causal tools only reach the short-term outcome. Qian's 2025 paper defines a distal excursion effect that contrasts long-term outcomes under two prompt policies.

Two estimators are offered, both guarded against a wrong outcome model, with theory for their behavior in large samples. Simulations check the pair. Applied to a real activity prompt trial, the method finds early prompts shaped long-term habits more than late ones. The author offers the toolkit for digital health science.

Catalog record

CREATORTianchen Qian
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE71 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-13500

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Distal Causal Excursion Effects: Modeling Long-Term Effects of Time-Varying Treatments in Micro-Randomized Trials Tianchen Qian Abstract Micro-randomized trials (MRTs) play a crucial role in optimizing digital interventions. In an MRT, each participant is sequentially randomized among treatment options hundreds of times. While the interventions tested in MRTs target short-term behavioral responses (proximal outcomes), their ultimate goal is to drive long-term behavior change (distal outcomes). However, existing causal inference methods, such as the causal excursion effect, are limited to proximal outcomes, making it challenging to quantify the long-term impact of interventions. To address this gap, we introduce the distal causal excursion effect (DCEE), a novel estimand that quantifies the long-term effect of time-varying treatments. The DCEE contrasts distal outcomes under two excursion policies while marginalizing over most treatment assignments, enabling a parsimonious and interpretable causal model even with a large number of decision points. We propose two estimators for the DCEE—one with cross-fitting and one without—both robust to misspecification of the outcome model. We establish their asymptotic properties and validate their performance through simulations. We apply our method to the HeartSteps MRT to assess the impact of activity prompts on long-term habit formation. Our findings suggest that prompts delivered earlier in the study have a stronger long-term effect than those delivered later, underscoring the importance of intervention timing in behavior change. This work provides the critically needed toolkit for scientists working on digital interventions to assess long-term causal effects using MRT data. Keywords: causal excursion effect; distal outcome; micro-randomized trial; mobile health; time-varying treatment 1 Introduction Micro-randomized trials (MRTs) are widely used to develop and optimize digital adaptive interventions ( Klasnja et al., 2015 ; Liao et al., 2016 ) , with applications in health, education, and information systems ( Walton et al., 2018 ; Breitwieser et al., 2024 ; Pieper et al., 2024 ) . In an MRT, each participant is sequentially randomized among treatment options (e.g., receiving or not receiving a push notification encouraging behavioral change). These interventions directly target proximal outcomes (short-term behavioral responses, such as step count in the next 30 minutes), but ultimately aim to influence distal outcomes (long-term behaviors and habit formation). To analyze MRT data, the causal excursion effect (CEE) was introduced to model the effect of time-varying treatments on proximal outcomes ( Boruvka et al., 2018 ) . It has become the standard for primary and secondary analyses in MRTs ( Klasnja et al., 2019 ; Liu et al., 2023 ) . However, this estimand does not address distal outcomes, leaving key scientific questions unanswered. For example, in a physical activity study ( Klasnja et al., 2015 ) , do activity prompts lead to sustained habit formation? Current practice relies on behavioral theory to assume long-term effects, highlighting the need for statistical methods to infer them from MRT data. To address this critical need, we propose the distal causal excursion effect (DCEE) to quantify the impact of time-varying treatments on a distal outcome measured at the end of an MRT. The DCEE contrasts distal outcomes under two treatment policies representing excursions (i.e., deviations) from the MRT policy. By marginalizing over the stochastic treatment assignments at most decision points, the DCEE enables a parsimonious causal effect model even with many decision points; a similar idea was used in CEE for proximal outcomes. This marginalization addresses a key challenge in MRT analysis: treatments are randomized at hundreds or thousands of decision points. Classical models such as marginal structural models and structural nested models ( Robins, 1999a ) work well with a few treatment occasions but become impractical in MRTs with many decision points unless restrictive and often unrealistic assumptions are imposed ( Rudolph et al., 2022 ) . To estimate the DCEE, we develop two estimators—one with cross-fitting and one without—both robust to outcome model misspecification. We establish their asymptotic properties and validate finite-sample performance through simulations. We apply the method to the HeartSteps MRT to assess the long-term effect of activity suggestions on habit formation at study end. Our analysis reveals that suggestions delivered earlier, despite being farther from the distal outcome, have a larger effect than those delivered later. We discuss the implications for designing future digital interventions. Section 2 defines the DCEE. Section 3 presents the estimators and the asymptotic theory. Section 4 presents the simulation study. Section 5 presents the real data application. Section 6 concludes with discussion. 2 DCEE Definition 2.1 Notation: MRT with a Distal Outcome Consider an MRT with n individuals, each in the trial for T decision points where treatment is randomized. Variables without subscript i refer to a generic individual. Let A_{t} denote the binary treatment at decision point t : A_{t}=1 for treatment, A_{t}=0 for no treatment. Let X_{t} denote observations recorded between decision points t-1 and t . The distal outcome Y , typically the primary health outcome, is measured at study end. The observed data trajectory for one individual is O=(X_{1},A_{1},\ldots,X_{T},A_{T},Y) . Overbars indicate sequences of variables; for example, \bar{A}_{t}=(A_{1},\ldots,A_{t}) . The history up to t (excluding A_{t} ) is H_{t}=(X_{1},A_{1},\ldots,X_{t-1},A_{t-1},X_{t}) . At each t , A_{t} is assigned with randomization probability p_{t}(a|H_{t}):=P(A_{t}=a\mid H_{t}) for a\in\{0,1\} . Data from individuals are assumed independent and identically distributed samples from the distribution P_{0} , and expectations are taken under P_{0} unless stated otherwise. Fig. 1 illustrates the data structure. X_{1} A_{1} X_{2} A_{2} \cdots X_{t} A_{t} \cdots X_{T} A_{T} Y H_{t} Figure 1 : MRT data structure with a distal outcome. To simplify notation, a thick arrow is used to denote arrows from all nodes on the left to all nodes on the right. In designing the MRT, researchers may deem it unsafe or unethical to deliver push notifications at certain times, such as when a participant is driving. At these decision points, the participant is considered ineligible for randomization, and no treatment is delivered. Formally, X_{t} includes an indicator I_{t} , with I_{t}=1 denoting being eligible for randomization at t , and I_{t}=0 otherwise. If I_{t}=0 , then A_{t}=0 deterministically. In the MRT literature, I_{t}=0 is also called being unavailable for treatment ( Boruvka et al., 2018 ) ; here, we use “ineligible” for clarity. For simplicity, we sometimes explicitly write I_{t} together with H_{t} in conditional expectations, even though I_{t} is part of H_{t} : for instance, \mathbb{E}(\cdot|H_{t},I_{t}=1) means \mathbb{E}(\cdot|H_{t}\setminus\{I_{t}\},I_{t}=1) . Let \mathbb{P}_{n} denote the empirical mean over everyone. For any positive integer k , define [k]:=\{1,2,\ldots,k\} . The superscript \star is used to indicate quantities associated with the true data-generating distribution P_{0} . For a vector \alpha and a vector-valued function f(\alpha) , the notation \partial_{\alpha}f(\alpha):=\partial f(\alpha)/\partial\alpha^{T} denotes the Jacobian matrix, where the (i,j) -th entry corresponds to the partial derivative of the i -th entry of f with respect to the j -th entry of \alpha . 2.2 Distal Causal Excursion Effect (DCEE) To define the causal effect, we use the potential outcomes notation ( Rubin, 1974 ; Robins, 1986 ) . Lowercase letters denote instantiations (non-random values) of corresponding random variables; for example, a_{t} is an instantiation of A_{t} . For each individual, let X_{t}(\bar{a}_{t-1}) be the value of X_{t} that would have been observed at t if the individual were assigned treatment sequence \bar{a}_{t-1} . The potential outcome of

Read the complete text