📊 statistics

Modified treatment policies that depend on the natural history of treatment

This paper develops targeted learning estimators for longitudinal modified treatment policies that depend on the history of natural treatment values, enabling $\sqrt{n}$ -consistent causal inference for complex interventions like delaying treatment initiation, and applies the method to assess the impact of postponing risky pain treatment on opioid use disorder.

Original authors: Iván Díaz, Nicholas T. Williams, Paweł Morzywołek, Kara E. Rudolph

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Iván Díaz, Nicholas T. Williams, Paweł Morzywołek, Kara E. Rudolph

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 trying to understand the best way to navigate a complex journey, like driving through a city with changing traffic lights, road closures, and detours. In the world of medical research, scientists often want to know: "What would happen if we changed the rules of the road?"

This paper introduces a new, more flexible set of rules for asking that question, specifically when the "rules" depend on the driver's entire history, not just what they are doing right this second.

Here is a breakdown of the paper's core ideas using simple analogies:

1. The Old Way vs. The New Way

The Old Way (Contemporaneous Policies):
Imagine a traffic cop who only looks at your car right now. If you are speeding, they tell you to slow down. If you are driving slowly, they tell you to speed up. They don't care if you were speeding five minutes ago or if you've been driving safely for an hour. They only react to the immediate moment.

In the paper: Previous methods could only handle interventions that depend on the treatment a patient is receiving at that exact moment.

The New Way (History-Dependent Policies):
Now, imagine a smart navigation system that remembers your entire trip. It knows you were speeding earlier, so even if you are currently driving slowly, it might say, "You were speeding before, so let's keep you at this speed for a bit longer to be safe." Or, it might say, "You were supposed to take a turn 10 minutes ago, but you missed it; let's adjust the route to get you there later."

In the paper: This paper develops a new method to handle interventions that depend on the history of natural treatment. For example, "Delay the start of a treatment by one month" requires knowing when the treatment would have naturally started in the past, not just what is happening now.

2. The "Ghost" of the Past

To make these calculations work, the authors use a concept called the "Natural Value of Treatment."
Think of this as a "Ghost Driver."

The Real Driver: The patient who actually received the treatment.
The Ghost Driver: The patient who received no intervention, just following their natural path.
The Problem: If you change the Real Driver's route today, the Ghost Driver's past route might look different in a standard model.
The Solution: This paper creates a special "Augmented Data" system. Imagine a video game where you can save multiple versions of the game state. The authors build a system that keeps track of both the "Real" path and the "Ghost" path simultaneously, allowing them to calculate what would happen if they shifted the Ghost's timeline without losing track of the history.

3. The "Recipe" for the Answer

The paper provides a new mathematical "recipe" (called a Sequential Regression Formula) to calculate the results of these complex scenarios.

The Analogy: Imagine trying to bake a cake where the amount of sugar you add at step 3 depends on how much flour you would have added at step 1, even though you already changed the flour at step 1.
The Innovation: Standard recipes break here. The authors created a new "Augmented Recipe" that allows you to look back at the original ingredients (the natural history) while you are mixing the new batter (the intervention). This ensures the math stays accurate even when the rules are complex.

4. The "Double-Check" System

To make sure their new recipe doesn't fail, the authors use two advanced statistical tools (called TMLE and SDR).

The Analogy: Think of these as two different types of safety nets.
- Net A (TMLE): Very careful and strict. It guarantees the answer stays within realistic bounds (like ensuring a probability doesn't go above 100%), even if the data is messy.
- Net B (SDR): Very fast and efficient. It gives a great answer if the data is good, but if the data is tricky, it might occasionally give a result that seems impossible (like a probability of 127%).
The Benefit: The paper shows that both nets work well, provided the underlying data models are reasonably accurate. They allow researchers to use powerful machine learning tools to make these predictions without needing to perfectly guess every single detail of the data beforehand.

5. Real-World Test: The Pain Treatment Delay

The authors tested their new method on a real-world scenario involving 12,745 patients with back pain (lumbar spinal stenosis).

The Question: What happens if we delay a "risky" pain treatment by one month?
The Challenge: To answer this, you can't just look at the patient today. You have to know: "If we hadn't intervened, when would they have naturally started this risky treatment?" If they naturally started it on Day 1, delaying it means moving it to Day 30. If they naturally started it on Day 29, delaying it means moving it to Day 60.
The Result: Using their new "History-Dependent" method, they estimated that delaying the risky treatment slightly increased the chance of the patient not developing an opioid use disorder (though the confidence interval was wide, meaning the result wasn't statistically definitive in this specific sample).
Key Takeaway: The old methods couldn't have answered this specific question because they couldn't "see" the natural start date once the intervention changed the timeline. The new method could.

Summary

This paper is about upgrading the "traffic cop" of medical research. Instead of just reacting to what is happening right now, the new method allows researchers to simulate policies that depend on a patient's entire medical history. It provides the mathematical tools (the "Augmented Recipe" and "Double-Check Nets") to do this safely and accurately, specifically for scenarios like delaying treatments or adjusting doses based on past behavior.

Technical Summary: Modified Treatment Policies Dependent on the Natural History of Treatment

Problem Statement
Longitudinal Modified Treatment Policies (LMTPs) provide a flexible framework for defining and estimating causal effects in real-world settings involving continuous exposures, grace periods, or treatment delays. Previous work, specifically Díaz et al. (2023), established targeted machine learning estimators for LMTPs where the intervention at time $t$ depends solely on the contemporaneous natural value of treatment (i.e., the treatment value that would have been observed at time $t$ had the intervention been stopped immediately prior).

However, this restriction precludes the estimation of policies where the intervention rule depends on the history of natural treatment values prior to time $t$ . Important scientific questions, such as delaying the initiation of a treatment or capping dose escalation relative to a clinician's natural trajectory, require access to past natural treatment values that may be overwritten by the intervention itself. Standard LMTP formulations lose access to these original natural values once an intervention modifies the treatment path, rendering such "delay" or "trajectory modification" policies unidentifiable under the existing framework. This paper addresses the gap in developing targeted learning estimators for the general case where interventions depend on the full natural history of treatment.

Methodology
The paper proposes a generalization of LMTPs, termed Generalized Longitudinal Modified Treatment Policies (G-LMTPs), where the intervention $A^d_t$ at time $t$ is a function of the history of natural treatment values $\bar{A}_t(\bar{A}^d_{t-1})$ and the history of covariates $H_t(\bar{A}^d_{t-1})$ .

Identification and Representation:
The authors demonstrate that the standard sequential regression representation (iterated conditional expectations) used for contemporaneous LMTPs fails for G-LMTPs because the conditioning set and the intervention function involve conflicting values of past natural treatments. To resolve this, they introduce an augmented-data version of the sequential regression formula. This approach decouples the values used for conditioning (integrating over the observed data distribution) from the values used in the intervention function (tracking the hypothetical natural history). This is formalized in Proposition 1, which defines a recursive algorithm involving augmented variables $\bar{s}_t$ to track possible natural treatment histories.
Efficiency Theory:
The paper derives the Efficient Influence Function (EIF) for the resulting nonparametric functional. Theorem 2 provides a conditional von Mises expansion, characterizing the error of the regression estimators. This expansion reveals a "multiply robust" structure: the estimator remains unbiased if either the outcome regression models or the treatment mechanism models are correctly specified at each step, provided specific rate conditions are met.
Estimation Algorithms:
Two estimators are developed based on the EIF and the augmented sequential regression:
- Targeted Minimum Loss-Based Estimator (TMLE): A substitution estimator that iteratively updates preliminary regression estimates to solve the efficient influence function estimating equation. It utilizes sample splitting to avoid Donsker conditions and ensures predictions remain within the parameter space.
- Sequential Doubly Robust (SDR) Estimator: A regression-based estimator that uses the multiply robust unbiased transformation (the pseudo-outcome $\phi_{t+1}$ ) derived from the EIF. It regresses these pseudo-outcomes on covariates sequentially. The authors note that while SDR is simpler and may offer better finite-sample bias properties, it is not a substitution estimator and can produce estimates outside the parameter space (e.g., probabilities $>1$ ) if the true value is near a boundary.

Key Contributions

Generalization of LMTPs: The paper formally defines and provides identification theorems for G-LMTPs, extending the scope of causal inference to policies dependent on the natural history of treatment.
Augmented Sequential Regression: It introduces a novel computational strategy using augmented datasets to represent the identification formula, allowing for the application of standard machine learning libraries to complex, history-dependent interventions.
Robust Estimators: The development of TMLE and SDR estimators that achieve $\sqrt{n}$ -consistency and asymptotic normality under standard doubly robust rate assumptions, even when nuisance parameters (outcome and treatment regressions) are estimated using flexible, data-adaptive machine learning methods.
Application to Delay Interventions: The paper explicitly demonstrates how G-LMTPs solve the "delay" problem (e.g., delaying a risky pain treatment) which was previously intractable under the contemporaneous-only framework.

Results

Simulation Study: Monte Carlo simulations were conducted with sample sizes ranging from 250 to 10,000. Both TMLE and SDR estimators demonstrated decreasing bias and mean squared error (MSE) as sample size increased. The confidence interval coverage was generally close to or slightly conservative relative to the nominal 95% level. The study highlighted a trade-off in finite samples where SDR showed slightly lower variance but potentially higher bias compared to TMLE in specific scenarios.
Motivating Application: The methods were applied to a cohort of 12,745 Medicaid-insured patients with lumbar spinal stenosis to estimate the effect of delaying risky pain treatments (opioids, gabapentinoids, benzodiazepines) by one month on the 12-month incidence of Opioid Use Disorder (OUD) or overdose.
- The observed 12-month probability of remaining free from OUD/overdose was 0.963.
- Under the hypothetical one-month delay policy, the TMLE estimated a probability of 0.986 (95% CI: 0.915–1).
- The SDR estimator produced a point estimate of 1.27, illustrating the known limitation of non-substitution estimators when the true parameter lies near the boundary of the parameter space.

Significance and Claims
The paper claims to extend the class of estimable causal effects to include interventions that modify the timing or trajectory of treatment decisions based on natural history. By providing a rigorous identification strategy and efficient estimators, the authors enable the evaluation of complex policy questions—such as delaying treatment initiation or capping dose escalation—that were previously unaddressable by standard LMTP methods. The authors maintain a modest stance, noting that while the current work focuses on discrete exposures, future research is required to extend these methods to continuous, multivariate, or mixed treatment settings without explicit enumeration of treatment histories. They also identify future directions including scalable algorithms for long follow-up and diagnostics for support violations induced by history-dependent policies.