PaperlayerView exact revision on arXiv

Source-linked AI summary

Non-parametric causal effects based on longitudinal modified treatment policies

Iván Díaz, Nicholas Williams, Katherine L. Hoffman, Edward J. Schenck

arXiv:2006.01366v4stat.ME

TL;DR

Deterministic interventions are often impractical and worsen positivity problems for continuous or multi-valued longitudinal exposures. The paper develops LMTPs, derives their identification and efficient influence function, and proposes four estimators. The estimators include efficient and sequentially doubly robust procedures, with robustness and consistency properties established under stated nuisance-estimation configurations.

  • Problem

    Deterministic interventions for continuous or multi-valued exposures may be impractical, while positivity violations and restrictive modeling complicate non-parametric causal-effect estimation.

  • Method

    The paper develops LMTPs, derives a sequential regression identification formula and efficient influence function, and proposes four estimators.

  • Results

    The proposed estimators include a TMLE with τ + 1 multiply robust consistency configurations and a sequentially doubly robust estimator.

  • Takeaways & Limitations

    LMTPs provide a longitudinal framework for estimating causal effects under interventions that modify observed treatment values rather than deterministically setting impractical exposure levels.

  • Takeaways & Limitations

    The paper notes that n1/2-consistency requires consistent estimation of all nuisance parameters, and data-adaptive nuisance estimation complicates inference for IPW and g-computation estimators.

Abstract

from arXiv · show

Most causal inference methods consider counterfactual variables under interventions that set the treatment deterministically. With continuous or multi-valued treatments or exposures, such counterfactuals may be of little practical interest because no feasible intervention can be implemented that would bring them about. Furthermore, violations to the positivity assumption, necessary for identification, are exacerbated with continuous and multi-valued treatments and deterministic interventions. In this paper we propose longitudinal modified treatment policies (LMTPs) as a non-parametric alternative. LMTPs can be designed to guarantee positivity, and yield effects of immediate practical relevance with an interpretation that is familiar to regular users of linear regression adjustment. We study the identification of the LMTP parameter, study properties of the statistical estimand such as the efficient influence function, and propose four different estimators. Two of our estimators are efficient, and one is sequentially doubly robust in the sense that it is consistent if, for each time point, either an outcome regression or a treatment mechanism is consistently estimated. We perform a simulation study to illustrate the properties of the estimators, and present the results of our motivating study on hypoxemia and mortality in Intensive Care Unit (ICU) patients. Software implementing our methods is provided in the form of the open source \texttt{R} package \texttt{lmtp} freely available on GitHub (\url{https://github.com/nt-williams/lmtp}).

1 Introduction

The paper motivates longitudinal modified treatment policies as practical, non-parametric alternatives to deterministic interventions for continuous or multi-valued exposures. It develops identification results and estimators, including efficient and multiply robust procedures.

  • Motivation: Continuous and multi-valued exposures make deterministic interventions difficult to implement and intensify positivity violations, especially longitudinally.Non-parametric dose-response methods also face slower convergence and may require restrictive parametric assumptions.
  • Motivation: Modified treatment policies shift observed exposure values, yielding effects interpretable as expected mean-response changes without assuming causal-model linearity.The paper extends this idea longitudinally as LMTPs.
  • Contributions: A sequential regression formula identifies the LMTP effect and provides an alternative expression of the extended g-formula.The paper also introduces a stochastic LMTP intervention that draws exposures from the LMTP distribution.
  • Implications: LMTP identification uses the same sequential randomization assumption as dynamic interventions, and the methods generalize to estimating dynamic-intervention parameters.The software implementation is provided in the open-source R package lmtp.
  • Estimation: The authors propose four estimators, including extensions of inverse probability weighting and g-computation, plus efficient and multiply robust procedures.The sequentially doubly robust estimator is consistent when, at each time point, either the outcome regression or treatment mechanism is consistently estimated.

2 Notation and definition of causal effects

The paper defines longitudinal causal effects through counterfactual worlds generated by interventions that modify treatment using a user-specified function of treatment and history. It distinguishes deterministic modified treatment policies from stochastic interventions that draw treatment from the induced shifted distribution.

  • Longitudinal modified treatment policies: An LMTP recursively assigns modified exposures over time after removing the observed treatment equation from the structural model.The counterfactual history is built from the assigned exposures and evolving covariate history.
  • Longitudinal modified treatment policies: LMTP causal effects compare counterfactual outcome distributions generated by different user-specified treatment-modification functions.The function d(at, ht) maps treatment and history to a new exposure value.
  • Examples: Threshold LMTPs raise exercise to 30 minutes when observed or natural exercise falls below 30, while leaving higher values unchanged.This regime targets the effect of exercising at least 30 minutes daily on coronary heart disease risk.
  • Examples: Shift LMTPs increase a continuous exposure by a user-given amount when the increase is feasible for the individual’s history.The paper illustrates a 50-unit P/F-ratio increase for patients with acute respiratory failure.
  • LMTP stochastic interventions: LMTP stochastic interventions instead draw each exposure from the distribution induced by the modification function, representing a population-level shift.They need not shift every individual’s exposure to a fixed value.
  • LMTP stochastic interventions: Both LMTPs and LMTP stochastic interventions can accommodate randomized regimes when the randomizer is independent across units and of the exogenous variables.The randomizer’s distribution must also not depend on the observed-data distribution.

3 Identification of causal effects

The paper identifies LMTP parameters from the observed-data distribution under positivity and sequential randomization assumptions. Its identification result is expressed through sequential regressions and extends to settings including censoring, survival analysis, and missing exposures.

  • Identification assumptions: Positivity requires every modified treatment-history pair targeted by d to lie within the support of the observed treatment distribution.The intervention can be designed to enforce this condition when conditional-support information is available.
  • Identification assumptions: Standard sequential randomization requires treatment-assignment exogenous variables to be conditionally independent of future covariate exogenous variables given history.Strong sequential randomization additionally conditions independence from future treatment-assignment exogenous variables.
  • Identification result: The identification theorem expresses θ_lmtp and θ_lmtp-si solely as functionals of the observed-data distribution.The result provides the basis for estimating both deterministic and stochastic LMTP effects.
  • Identification result: The identification expression is written as sequential regressions and is an alternative form of the extended g-formula.The framework supports interventions and data structures beyond a simple single-time-point setting.
  • Extensions: The framework handles loss-to-follow-up, survival analysis, and missing exposures through appropriate exposure, censoring, and outcome definitions.Monotone loss-to-follow-up makes later covariates and outcomes degenerate after censoring.

4 Optimality theory

The paper develops efficiency theory for LMTP parameters, including their efficient influence functions, efficiency bounds, and conditions supporting robust estimation. For continuous exposures, piecewise smooth invertibility avoids the non-pathwise-differentiability problem of threshold interventions.

  • The sequential regression formula identifies the effect of an LMTP and underpins its efficiency analysis.
  • Threshold interventions setting a continuous exposure to δ are not generally n^1/2-estimable non-parametrically because the parameter is not pathwise differentiable.The problematic term corresponds to estimating the causal effect of a static intervention at A = δ.
  • The paper instead restricts continuous-exposure interventions to piecewise smooth, invertible policies so change-of-variable calculations apply.The exposure support is partitioned into subintervals where each policy component has an inverse and derivative.
  • Under these conditions, the LMTP efficient influence function has a structure similar to the influence function for dynamic treatment regimes.The policy is assumed not to depend on the data distribution P.
  • The efficient influence function supports locally efficient and multiply robust estimators, including consistency when, at each time point, either the treatment mechanism or outcome regression is consistently estimated.The corresponding first-order approximation uses a second-order remainder term, allowing slower nuisance-parameter convergence rates.
  • For a single time point, the efficiency bound depends on outcome variability, treatment-effect heterogeneity, and how much the intervention changes the exposure density.Larger exposure-density changes in regions of high outcome variability are expected to produce larger efficiency bounds.

5 Estimation and statistical inference

The paper proposes estimators for LMTPs using regression, inverse probability weighting, targeted updating, and sequential regression. The EIF-based estimators achieve efficiency and robustness under weaker nuisance-estimation conditions than substitution and IPW estimators.

  • Density-ratio estimation is recast as classification on 2n observations, allowing flexible classification methods such as Super Learner.Super Learner combines algorithms using weights selected by cross-validated prediction error.
  • Substitution and IPW estimators generally cannot achieve n^1/2-consistency with data-adaptive nuisance estimation in non-parametric models.Their n^1/2-consistency generally requires parametric-rate L2(P) convergence of both outcome and treatment-mechanism estimators.
  • TMLE is locally efficient, n^1/2-consistent, and τ + 1-multiply robust under the stated assumptions.It targets preliminary outcome-regression estimates by solving the efficient influence function estimating equation.
  • EIF-based estimators require only n^1/2-convergence of the second-order regression-bias term for efficiency under weaker conditions.

6 Illustrative application

The illustrative application estimates how increasing P/F ratios by 50 units among patients below 300 affects 14-day survival after invasive ventilation. Across estimators, the estimated improvement is modest, while IPW is unstable because of highly variable density-ratio weights.

  • Study population: The cohort comprises 10,044 intubated ICU patients followed for 14-day survival.The data come from the Weill Cornell Critical carE Database for Advanced Research.
  • Study objective: The study estimates the effect of increasing P/F ratios by 50 among patients with clinically defined acute respiratory failure.The target population has P/F ratio < 300, and the increase is described as clinically feasible and meaningful.
  • Results: Substitution, TMLE, and SDR produced comparable results, whereas IPW estimated a 22% increase in survival.The disagreement is attributed to high variability in the density ratios rather than positivity violations.
  • Estimator behavior: The IPW weights had a coefficient of variation of 5.5%, making the IPW estimator highly variable.The maximum value of the relevant weight was 97.7, and the weight distribution is presented in supplementary Figure 4.

Supplementary Materials for

The supplementary materials accompany the paper on non-parametric causal effects based on longitudinal modified treatment policies. They identify the authors, affiliations, and July 2021 arXiv version.

  • Paper information: The paper is titled “Non-parametric causal effects based on longitudinal modified treatment policies.”
  • Paper information: The authors are Iván Díaz, Nicholas Williams, Katherine L. Hoffman, and Edward J. Schenck.
  • Paper information: The document is identified as arXiv:2006.01366v4, dated 6 July 2021.
  • Affiliations: The listed affiliations are Weill Cornell Medicine’s Division of Biostatistics and Division of Pulmonary & Critical Care Medicine.

1 Simulation study

The simulation evaluates four estimators across nuisance-modeling scenarios, including consistent, partially inconsistent, and fully misspecified models. Results support the efficiency and multiple-robustness properties of TMLE and SDR, while confidence-interval coverage can fail under misspecification.

  • Simulation design: The simulation tests four estimators under four scenarios varying whether treatment-mechanism and outcome-regression nuisance estimators are consistent.The scenarios include all nuisance parameters consistent, complementary partial consistency, and complete inconsistency.
  • Estimator performance: When all estimators are consistent, TMLE and SDR have n^1/2-bias converging to zero and achieve the non-parametric efficiency bound.These results are reported for panels E and I of Figure 1.
  • Estimator performance: IPW appears consistent when all treatment-mechanism estimators are consistent but is inconsistent if any treatment-mechanism estimator is inconsistent.All estimators are inconsistent when all models are misspecified.
  • Inference: The empirical influence-function standard deviation underestimates the standard error, producing slightly below-nominal coverage for TMLE and SDR.Under scenarios 2–4, n^1/2-bias yields confidence intervals with zero asymptotic coverage, with worse TMLE performance in scenario 3.

2 Identification (Theorem 1)

The identification argument expresses the LMTP counterfactual mean through recursive integration over observed longitudinal treatment and covariate distributions. The proof uses iterated expectation, the intervention definition, and assumptions establishing the required counterfactual relationships.

  • Identification: The LMTP counterfactual mean is identified under the paper’s stated assumptions through recursive evaluation of nested integrals.The theorem follows after substituting the identified expressions into the recursive representation.
  • Counterfactual construction: At each time point, counterfactual treatment and covariate variables are defined in a hypothetical world under the intervention history.The proof fixes a time point and constructs the corresponding counterfactual variables recursively.
  • Proof strategy: The proof repeatedly applies the law of iterated expectation and the definition of the counterfactual outcome.Other equalities invoke Lemma 1 and the paper’s assumptions.
  • Assumptions: Under the assumed NPSEM, the relevant counterfactual variable is a deterministic function of the treatment- and covariate-related exogenous variables.

3 Efficient influence functions (Theorem 2)

The theorem characterizes the efficient influence function for the LMTP parameter in a non-parametric model, using a conjecture-and-verification argument based on substitution estimators and change-of-variable identities.

  • The proof defines the parameter as a functional mapping the observed-data distribution P to a real number and assumes the treatment policy d is smooth and invertible.
  • The candidate influence function φ1(Z; η) is verified by showing that it satisfies the defining pathwise-derivative condition for an efficient influence function.
  • The proof first derives the efficient influence function in a model with discrete measure ν, where integrals can be represented as sums.
  • The treatment policy d is treated as fixed and therefore does not need to be estimated.
  • The substitution-estimator argument applies the Delta method to the function class containing the relevant outcome and treatment-mechanism components, yielding the influence function φ1(Z; η).
  • For continuous treatments, the proof uses the change-of-variable formula under Assumption 4, while the discrete-treatment case holds directly.

4 Sequential double robustness (Lemma 1)

Lemma 1 establishes the recursive identity needed for sequential double robustness by repeatedly applying a change-of-variable formula under Assumption 4.

  • The lemma follows from recursive application of the change-of-variable formula across time points s = t + 1, ..., τ.

5 Second order representation for SDR (Lemma 3)

The second-order representation decomposes the estimation remainder into recursively propagated errors and bounds the resulting terms using the lemma, regularity assumptions, and Cauchy–Schwarz.

  • The proof simplifies notation and substitutes equation (7) into equation (6) before applying the resulting equality inductively backward over time.
  • The recursive argument is controlled under the lemma’s assumption, with the bound involving a constant C1.
  • The remainder control follows from the law of iterated expectation and the Cauchy–Schwarz inequality.

6 Asymptotic Normality of TMLE (Theorem 3)

Theorem 3 establishes asymptotic normality for the TMLE by controlling empirical-process and second-order remainder terms, with cross-fitting and entropy conditions supporting the argument.

  • The TMLE solves the efficient influence function estimating equation, which is used in the asymptotic expansion.
  • Cross-fitting fixes the nuisance estimate on each training-data split, allowing empirical-process results to control the corresponding prediction-set term.
  • Entropy and envelope conditions are imposed through bracketing-number arguments for the relevant function classes.
  • Under the theorem’s assumptions, the empirical-process and second-order remainder terms converge to zero in probability.
  • The supplemental figures describe learner weights for estimating rt and mt across 14 time points and show the distribution of IPW estimator weights.
Loading 2006.01366v4…