📊 statistics

Structural Nested Mean Models Under Parallel Trends Assumptions

This paper bridges time-varying Difference-in-Differences and Structural Nested Mean Models by demonstrating that SNMMs are identifiable under conditional parallel trends assumptions, thereby enabling researchers to estimate a broader range of causal effects, including time-varying heterogeneity and dynamic treatment strategies, while providing methods for sensitivity analysis and optimal regime estimation.

Original authors: Zach Shahn, Oliver Dukes, Meghana Shamsunder, David Richardson, Eric Tchetgen Tchetgen, James Robins

Published 2026-07-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zach Shahn, Oliver Dukes, Meghana Shamsunder, David Richardson, Eric Tchetgen Tchetgen, James Robins

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to figure out the true cause-and-effect relationship between two things: a treatment (like a new policy, a flood, or a temperature change) and an outcome (like insurance rates, crop yields, or test scores).

Usually, to solve this mystery, you need a "control group" (people who didn't get the treatment) that looks exactly like the "treatment group" in every way, except for the treatment itself. In the world of statistics, this is called the "No Unobserved Confounding" assumption. It's like saying, "The only reason these two groups are different is because one got the medicine and the other didn't."

However, in the real world, people and places are messy. There are hidden factors (confounders) we can't measure that might influence both who gets the treatment and what happens later.

This paper introduces a new way to solve this mystery by combining two existing detective tools: Difference-in-Differences (DiD) and Structural Nested Mean Models (SNMMs).

Here is the breakdown of what they did, using simple analogies:

1. The Old Tools vs. The New Hybrid

The "Parallel Trends" Tool (DiD): This is the classic detective method. It assumes that if the treatment hadn't happened, the treatment group and the control group would have followed the same path (trend) over time. Think of two runners on a track. If they are running side-by-side at the same speed, and suddenly one gets a jetpack (the treatment), you can measure the jetpack's effect by seeing how much faster they get compared to the other runner.
- The Problem: This tool is usually limited. It can only tell you the average effect of starting a treatment. It struggles if the treatment changes value repeatedly (like a thermostat going up and down) or if you want to know how the effect changes based on specific details (like "Does the jetpack work better for heavy runners or light runners?").
The "SNMM" Tool: This is a more complex, flexible tool that can answer all those detailed questions (heterogeneity, sustained effects, etc.).
- The Problem: Historically, this tool required the strict "No Unobserved Confounding" assumption. It needed to be sure there were no hidden factors messing things up.

The Paper's Big Breakthrough:
The authors showed that you can use the flexible SNMM tool even when you only have the weaker "Parallel Trends" assumption. They proved that if the "runners" would have stayed on the same path without the treatment, you can use the flexible SNMM to answer much deeper questions than the standard DiD method allows.

2. What Can You Do With This New Super-Tool?

Because this new method is so flexible, the paper claims it allows researchers to ask questions that were previously impossible under "Parallel Trends" assumptions. Here are the specific "detective cases" they can now solve:

The "Who Benefits Most?" Question (Heterogeneity):
- Analogy: Instead of just asking "Did the jetpack work?", you can ask, "Did the jetpack work better for runners who were already tired (high unemployment) versus those who were fresh?"
- Paper Example: They analyzed Medicaid expansion and found it reduced insurance rates more in counties with higher unemployment.
The "One-Time Shock" Question (Blip Effects):
- Analogy: Imagine a runner gets a single push (a flood) and then stops. Did that single push change their speed forever, or did they go back to normal?
- Paper Example: They studied how a single flood affected flood insurance take-up. They found a huge spike in insurance right after the flood, which then faded away as people "forgot" the danger.
The "Keep It Going" Question (Sustained Interventions):
- Analogy: What if the runner keeps the jetpack on for the whole race? What if the temperature keeps rising?
- Paper Example: They looked at "Growing Degree Days" (a measure of heat). They calculated what happens to crop yields if the heat stays high for years, rather than just looking at a one-time change.
The "What If We Stop Other Things?" Question (Controlled Direct Effects):
- Analogy: If the jetpack works, but only because it also accidentally turned on the air conditioning, what is the effect of the jetpack alone?
- Paper Example: They looked at Medicaid expansion while assuming minimum wage didn't change. This isolates the effect of Medicaid from other economic changes.
The "Continuous Scale" Question:
- Analogy: Most DiD tools work like a light switch (On/Off). This new tool works like a dimmer switch. It can handle treatments that are numbers (like temperature or income) where everyone starts at a different level.
- Paper Example: Crop yields and temperature don't have a "zero" baseline that everyone shares. This method handles that messiness.

3. The "Optimal Strategy" (The Coach's Playbook)

The paper also touches on a very advanced idea: Finding the Best Playbook.
If you assume that the "Parallel Trends" rule holds true for every possible strategy a coach could call, and that hidden factors don't change how the players react to those strategies, you can mathematically calculate the optimal dynamic treatment regime.

Analogy: Instead of just saying "Run fast," the tool can tell a coach: "If the player is tired and it's raining, switch to defense. If they are fresh and it's sunny, switch to offense."
Paper Claim: They show how to estimate this "best strategy" using their new method, which is a big deal for fields like Reinforcement Learning (AI that learns by doing).

4. The "What If We're Wrong?" Check (Sensitivity Analysis)

The authors know that "Parallel Trends" is a strong assumption. What if the runners weren't actually on the same path to begin with?

They provide a Sensitivity Analysis tool. This is like a "stress test." You can tell the computer, "Okay, let's assume the hidden factors were messing with the trends by this much." The tool then recalculates the results to show you a range of possibilities. This helps researchers see how robust their conclusions are if their assumptions are slightly off.

Summary

In short, this paper takes a flexible, powerful statistical model (SNMM) and teaches it how to work with the "Parallel Trends" assumption (usually reserved for simpler models).

The result? Researchers can now use a single, robust method to answer complex questions about:

Who benefits most?
What happens after a single event vs. a long-term change?
How to handle treatments that are continuous numbers (not just Yes/No).
How to find the best possible strategy for treating patients or populations over time.

They tested this on real-world data regarding Medicaid expansion, floods, and crop yields, proving it works in practice.

Technical Summary: Structural Nested Mean Models Under Parallel Trends Assumptions

1. Problem Statement

The paper addresses a gap in causal inference methodology for estimating time-varying treatment effects on repeated continuous outcomes. Two dominant frameworks exist but operate in isolation:

Time-varying Difference-in-Differences (DiD): Widely used in economics and social sciences, particularly in "staggered adoption" settings. Standard DiD methods typically rely on the "parallel trends" assumption but are often limited to estimating marginal effects of treatment initiation (e.g., Average Treatment Effects on the Treated, ATT) and struggle with time-varying effect heterogeneity, continuous treatments without a shared baseline, or sustained interventions where treatment values change repeatedly.
Structural Nested Mean Models (SNMMs): A flexible framework developed by Robins and colleagues for modeling time-varying effect heterogeneity and complex dynamic treatment regimes. Historically, SNMMs have been identified only under the strong assumption of "no unobserved confounding" (sequential exchangeability), which is often difficult to justify in observational settings.

The central problem is that SNMMs offer a richer set of causal estimands (e.g., effects of sustained interventions, controlled direct effects, dynamic regimes) but lack identification under the more plausible "parallel trends" assumptions typically used to justify DiD methods. Conversely, DiD methods lack the flexibility to answer complex substantive questions modeled by SNMMs.

2. Methodology

Core Framework

The authors link these two approaches by demonstrating that SNMMs can be nonparametrically identified under a Time-Varying Conditional Parallel Trends assumption.

Notation: The paper considers a cohort observed from time $0$ to $K$ . At each time $m$ , data includes covariates $Z_m$ , treatment $A_m$ , and outcome $Y_m$ . $\bar{L}_m$ represents a summary of covariate and outcome history up to time $m$ .
Causal Estimands: The paper focuses on blip functions $\gamma^*_{mk}(\bar{l}_m, \bar{a}_m)$ $γ_{mk}^{*} (\overset{ˉ}{l}_{m}, \overset{a}{ˉ}_{m})$ , which represent the conditional average effect at time $k$ $k$ of receiving treatment $a_m$ $a_{m}$ at time $m$ $m$ and then following a specific regime $g$ $g$ , compared to following $g$ $g$ at time $m$ $m$ and thereafter.
- Special cases include initiation effects (where $g$ is the "untreated" regime $\bar{0}$ ), sustained intervention effects (where $g$ is the "sustain previous treatment" regime), and controlled direct effects (using multidimensional treatments).
Identification Assumption: The authors propose a Time-Varying Conditional Parallel Trends assumption. For a target regime $g$ $g$ , the assumption states that, conditional on observed history $\bar{L}_m$ $\overset{ˉ}{L}_{m}$ and treatment history up to $m-1$ $m - 1$ , the expected counterfactual outcome trend from $m$ $m$ onwards does not depend on whether the treatment actually received at $m$ $m$ ( $A_m$ $A_{m}$ ) would have been assigned under $g$ $g$ .
- Formally: $E[Y_k(\bar{a}_{m-1}, g_m) - Y_{k-1}(\bar{a}_{m-1}, g_m) | \bar{A}_m = \bar{a}_m, \bar{L}_m] = E[Y_k(\bar{a}_{m-1}, g_m) - Y_{k-1}(\bar{a}_{m-1}, g_m) | \bar{A}_m = (\bar{a}_{m-1}, g_m(\bar{L}_m, \bar{a}_{m-1})), \bar{L}_m]$ .
- Crucially, this assumption conditions on the observed treatment history up to $m-1$ and covariate history, avoiding the conditioning on future treatment (e.g., "never treated" groups) that often renders standard DiD assumptions implausible in the presence of time-varying confounders.

Estimation Strategy

The paper provides multiply robust estimators for the parameters of parametric SNMMs:

G-Estimation: The authors construct estimating equations based on the property that, under the parallel trends assumption, a specific "blipped-down" outcome function $H_{mk}(\gamma)$ behaves like the counterfactual outcome in that its conditional trend is independent of $A_m$ .
Doubly Robust (DR) Estimators: The paper proposes cross-fit DR-ML estimators. These estimators remain consistent if either the model for the treatment assignment mechanism (propensity score) or the model for the outcome trend (nuisance function) is correctly specified. This allows for the use of flexible machine learning algorithms to estimate nuisance parameters.
Optimal Regimes: Under the stronger assumption that parallel trends holds for all regimes and that unobserved confounders do not modify effects (effect modification), the authors show how to identify and estimate optimal dynamic treatment regimes (A-learning) using SNMMs.

Sensitivity Analysis

Recognizing that parallel trends is untestable, the authors adapt sensitivity analysis methods for SNMMs. They introduce a bias function $c^g_{mk}$ to quantify deviations from the parallel trends assumption. By specifying a range of plausible bias functions, analysts can produce a range of plausible effect estimates, allowing for a sensitivity analysis of effect heterogeneity and derived quantities.

3. Key Contributions

Identification of SNMMs under Parallel Trends: The primary theoretical contribution is proving that SNMMs, previously known only under no-unobserved-confounding, are identified under a modified parallel trends assumption. This bridges the gap between the DiD and SNMM literatures.
Expanded Estimands: The results allow practitioners to estimate a broader set of causal questions under parallel trends, including:
- Time-varying effect heterogeneity (e.g., how Medicaid expansion effects vary by unemployment rate).
- Lasting effects of a single "blip" of treatment (e.g., the impact of a single flood on future insurance take-up).
- Effects of sustained interventions where treatment values change repeatedly.
- Effects of continuous-valued treatments without a shared baseline (e.g., temperature changes on crop yields).
- Controlled direct effects (e.g., Medicaid expansion effects barring minimum wage changes).
- Optimal dynamic treatment regimes.
Methodological Advancements: The paper provides practical, multiply robust estimators (including cross-fit DR-ML) and a framework for sensitivity analysis that accommodates time-varying covariates and effect heterogeneity.
Critique of Existing DiD Assumptions: The authors argue that their proposed parallel trends assumption is more plausible than standard "never-treated" or "not-yet-treated" assumptions in the presence of time-varying confounders. Standard assumptions often condition on future treatment values, which can induce bias if covariates are time-varying confounders.

4. Results and Applications

The paper validates the methodology through three real-data applications:

Medicaid Expansion (Staggered Adoption): Estimating the effect of Medicaid expansion on county-level uninsurance rates. The SNMM allowed the authors to model how the effect varies with county-level unemployment and population, finding that expansion was more impactful in counties with higher unemployment and lower population.
Floods and Insurance (Blip Effects): Analyzing the lasting impact of a single flood on flood insurance take-up. The model estimated an initial surge in uptake followed by a steep decline, with the surge being larger for more recent floods.
Temperature and Crop Yields (Sustained/Continuous Treatment): Re-analyzing the relationship between "growing degree days" (GDD) and crop profits. Unlike standard DiD, which struggles with continuous treatments lacking a shared baseline and repeated changes, the SNMM estimated the effects of sustained changes in GDD, finding that effects were larger when the increase was sustained longer and in counties with higher pre-increase GDD levels.

5. Significance and Claims

The paper claims that utilizing SNMMs under parallel trends assumptions is a fruitful strategy that expands the scope of causal inference in observational studies.

For DiD Practitioners: It offers a way to address substantive questions (heterogeneity, sustained effects, continuous treatments) that are currently out of reach for standard DiD methods, using similar identifying assumptions.
For SNMM Practitioners: It allows analysts to estimate causal effects under alternative identifying conditions (parallel trends vs. no unobserved confounding), facilitating evidence triangulation.
Theoretical Contribution: The paper situates itself within the reinforcement learning literature by showing how optimal dynamic regimes can be identified under parallel trends (with the additional assumption of no effect modification by unobserved confounders).

The authors are modest regarding the assumptions required for optimal regime identification (noting the strength of the "no effect modification" assumption) and acknowledge that the parallel trends assumption remains untestable, necessitating the proposed sensitivity analysis. The paper concludes that while the conditions for parallel trends in the presence of time-varying confounders are stringent, the proposed framework provides a rigorous and flexible tool for causal inference in complex, time-varying settings.