Causal Inference with Temporal Carryover Effects

18 concepts15 papers1 proposal

In classical A/B testing, we rely on a convenient fiction: the system resets between observations. Under the Stable Unit Treatment Value Assumption (SUTVA), an intervention applied to a unit today is assumed to have an immediate, isolated impact that does not bleed into tomorrow or affect other units.

In real-world dynamic systems, however, platforms have memory. If a ride-sharing platform runs an experiment that lowers prices to attract more drivers, those drivers do not instantly vanish when the experiment ends; they remain on the road, altering the marketplace's baseline state for the next interval. This phenomenon is known as temporal carryover or temporal interference. When today's treatment changes tomorrow's starting point, standard statistical tools like $t$-tests break down, often yielding severely biased estimates or showing near-zero power to detect true long-term impacts.

A rapidly growing body of research addresses this challenge by reimagining experimental design and causal inference through the lens of dynamic systems, time-series analysis, and reinforcement learning.

The Switchback Dilemma: Bias vs. Variance

To experiment in environments with high spatial or temporal interference (like ride-sharing networks or congested routing systems), practitioners frequently turn to switchback experiments. Instead of randomizing by user, a switchback design toggles the entire system between treatment and control states over sequential time intervals.

As Spang et al. demonstrate, the validity of a switchback experiment hinges on the assumption that the system state resets between intervals. If a treatment induces long-term changes—such as altering network cache states or shifting user habits—the effects bleed into subsequent control periods, contaminating the counterfactual baseline.

This creates a fundamental tension analyzed by Chin et al. and Wen et al.:

  • High-frequency switching: Toggling treatments rapidly (e.g., every 30 minutes) yields many data points and helps cancel out auto-correlated environmental noise, minimizing estimator variance. However, it maximizes carryover bias, because a large percentage of the time is spent in "transition states" contaminated by the previous period's policy.
  • Low-frequency switching: Toggling treatments slowly (e.g., every 4 hours) allows the system enough time to settle into its true steady state, minimizing carryover bias. However, it drastically reduces the number of independent data points, leading to high estimator variance.
High-Frequency:  | Tx | Co | Tx | Co | Tx | Co |  --> Low Variance, High Carryover Bias
Low-Frequency:   |   Treatment   |    Control    |  --> High Variance, Low Carryover Bias

To navigate this trade-off, Wen et al. provide a rigorous framework to optimize switchback schedules by explicitly modeling the interplay between carryover duration and reward autocorrelations, allowing engineers to choose the exact switching frequency that minimizes Mean Squared Error (MSE).

Modeling Experiments as Markov Decision Processes

If a treatment alters future states, then an experiment is not just a sequence of independent trials—it is a trajectory through a state space. Recognizing this, researchers have reframed causal inference in sequential settings using the mathematics of Markov Decision Processes (MDPs) and Reinforcement Learning (RL).

In a seminal paper, Shi et al. introduced an RL framework for A/B testing that treats the assignment of treatments over time as a policy within an MDP. By modeling how the system transitions from state to state, the authors decouple the immediate proximal reward from the long-term cumulative effect. This framework allows for valid sequential testing and early stopping, even when purely delayed effects would render a standard $t$-test powerless.

Building on this state-space view, Glynn et al. tackle adaptive experimental designs where a policy's current actions influence the future state of the environment. They utilize a Maximum Likelihood approach to adaptively optimize data collection while systematically correcting for the fact that the terminal state left by a predecessor policy biases the performance of its successor.

Taming Variance with Truncation and History

While RL methods offer an elegant way to capture long-term dynamics, full reinforcement learning estimators often suffer from high variance due to the compounding uncertainty of long trajectories. Recent work focuses on finding a middle ground between the high bias of myopic averages and the high variance of full RL.

Truncated Policy Gradients

To balance this tradeoff, Johari et al. introduced the Truncated Policy Gradient (TPG) estimator. Instead of trying to model an infinite-horizon steady state, TPG replaces immediate rewards with short-horizon trajectories. By truncating the evaluation window to a fixed number of steps, the estimator captures the vast majority of the carryover effect while keeping variance tightly bounded, providing a robust tool for nonstationary environments.

History-Aware Allocation

Traditional experimental designs are often Markovian, deciding the next treatment allocation based only on the current step. However, Wu et al. proved that when carryover effects are present, an optimal design must consider the entire data history. They introduced a Transformer Reinforcement Learning (TRL) approach that ingests the full sequence of past allocations and outcomes to dynamically assign treatments, directly minimizing the estimator's mean squared error and significantly outperforming traditional designs.

Markovian Design:     [Current State] ----------> [Next Allocation]
History-Aware (TRL):  [Full Sequence History] --> [Next Allocation] (Optimizes MSE)

Specialized Designs for Diverse Domains

The challenges of temporal carryover manifest differently depending on the structure of the data and the application domain, prompting specialized methodologies:

  • Micro-randomized Trials (MRTs): In mobile health (mHealth), wearables deliver digital nudges to users repeatedly over time. Liao et al. established foundational sample size and power calculation frameworks for MRTs, explicitly accounting for how a current nudge might habituate a user and reduce their receptivity to future treatments.
  • Staggered Rollouts: When a new feature is rolled out gradually across different markets or cohorts, Xiong et al. developed a Precision-Guided Adaptive Experiment (PGAE) framework. This approach optimizes the timing of when units enter treatment to maximize the precision of both immediate and long-term effect estimates, reducing sample size requirements by over 50%.
  • Auto-correlated Policies: In many business settings, treatments are naturally auto-correlated (e.g., if a customer is put on a premium tier, they stay there for multiple periods). Battocchi et al. introduced a "Dynamic Adjusted Surrogate Index" to peel away the bias caused by these sequential dependencies, allowing researchers to forecast long-term impacts using only short-term data.
  • Network and Panel Interference: When temporal carryover intersects with social or physical network spillovers, the math becomes doubly complex. Bojinov et al. demonstrated that carryover effects break the temporal independence required for simplified network interference models, necessitating highly restrictive structural assumptions or specialized longitudinal designs to maintain valid inference. Similarly, in time-series cross-sectional data, Liu et al. showed that long-lasting carryover effects can contaminate counterfactual baselines, requiring careful adjustment of counterfactual prediction estimators.

By moving away from static assumptions and treating time as a continuous, stateful thread, this body of research provides the mathematical foundation necessary to run clean, high-powered experiments in a fundamentally interconnected world.

Key Foundational Work

Sample size calculations for micro‐randomized trials in mHealth (2015) — contributed 1 concept elements to this direction.
Optimal Experimental Design for Staggered Rollouts (2019) — 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_temporal_carryover_effects_ab79359f,
  title = {temporal carryover effects},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-carryover-effects/rdr_ab79359fd19240a7b91bb3213afeacd4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_temporal_carryover_effects_ab79359f_v1,
  title = {temporal carryover effects},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-carryover-effects/rdr_ab79359fd19240a7b91bb3213afeacd4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake