Non-parametric causal effects based on longitudinal modified treatment policies

• Fixed-level dose-response interventions may be infeasible, while positivity violations are likely for continuous or multivalued exposures and worsen with repeated measurements.Source: §1 · pp. 2–3
• The curve cannot achieve n^1/2-consistency because its parameter is not pathwise differentiable.Source: §1 · p. 3
• LMTPs define exposure changes as functions of observed treatment values, yielding an expected mean-response change without requiring linearity in the causal model.Source: §1 · p. 3
• The authors express the extended g-formula through sequential regressions, introduce LMTP stochastic interventions that draw from the LMTP distribution, and show they require the dynamic-intervention sequential randomization assumption.Source: §1 · p. 4
• The authors propose four estimators; with prespecified parametric nuisance models they are asymptotically Gaussian, while two additional estimators address inference with data-adaptive nuisance estimates.Source: §1 · pp. 4–5
• The derived LMTP efficient influence function assumes the policy does not depend on the data-generating mechanism and, under differentiability and invertibility conditions, enables multiply robust estimation.Source: §1 · p. 5
• The authors develop a targeted minimum loss-based estimator that is multiply robust consistent across τ + 1 configurations of consistent nuisance estimation.Source: §1 · p. 5
• The proposed SDR estimator generalizes Luedtke et al. (2017) and Rotnitzky et al. (2017) using multiply robust unbiased transformations.Source: §1 · p. 5
• The SDR estimator is expected to be consistent under 2^τ configurations of consistent nuisance-parameter estimation.Source: §1 · p. 6
• Cross-fitting is used to obtain root-n convergence while avoiding entropy conditions that data-adaptive nuisance estimators may violate.Source: §1 · p. 6
• Because the LMTP parameter generalizes dynamic interventions in longitudinal studies, the methods and software can also estimate those parameters.Source: §1 · p. 6

2 Notation and definition of causal effects

• LMTP effects contrast counterfactual outcome distributions under user-defined functions that map each exposure and history to a modified exposure.The modified exposure is assigned recursively across time, and the effects are contrasts under different functions.Source: §2 · p. 7
• A threshold LMTP sets exercise to 30 minutes when a participant’s natural value is below 30 and otherwise leaves it unchanged.The example considers the effect of exercising at least 30 minutes daily on coronary heart disease risk.Source: §2 · p. 8
• A shift LMTP can increase a continuous exposure by a user-given amount when feasible, illustrated by raising P/F ratio by 50 units for patients with acute respiratory failure.A multiplicative shift depending on history is another option.Source: §2 · pp. 8–9
• LMTP stochastic interventions draw exposures from the history-conditional distribution of the modified exposure, representing a population-level shift rather than individual exposure modification.Their causal effects contrast counterfactual outcome distributions under different modification functions.Source: §2 · p. 9
• Randomized regimes are permitted when their randomizers are independent across units and of the exogenous variables, with a distribution that does not depend on P.The randomizer can be included in the time-varying covariates.Source: §2 · pp. 9–10
• For binary exposures, an incremental propensity score intervention draws treatment from a Bernoulli distribution with a shifted propensity score.Source: §2 · p. 10
• The subtle but important difference between LMTP and LMTP-SI leads to different assumptions for identification.The paper says these assumptions are discussed in the next section.Source: §2 · p. 10

3 Identification of causal effects

• Identification requires positivity for the policy-shifted treatment and either standard or strong sequential randomization, with the latter imposing an additional independence condition.Positivity requires shifted treatment-history pairs to be supported in the data and may be enforced by defining the policy using conditional support; randomization holds when the relevant common causes are recorded in the history.Source: §3 · p. 11
• The identification theorem expresses θlmtp and θlmtp-si as functions of the observed-data distribution using sequential regressions.The sequential-regression expression is an alternative form of the extended g-formula of Robins et al. (2004) and Richardson and Robins (2013).Source: §3 · pp. 11–12
• The framework accommodates loss-to-follow-up, survival analysis, and missing exposures, including hypothetical no-loss interventions under monotone censoring.Censoring can be included as a treatment component; under monotone loss-to-follow-up, later covariates, exposures, and outcomes are degenerate after censoring.Source: §3 · p. 12
• For time-to-event analysis, the outcome can indicate being event-free and the censoring component can indicate remaining uncensored, imposing restrictions on the regression functions.The outcome may be truncated at an earlier time point, and the functions m_t and r_t must respect restrictions implied by these variable definitions.Source: §3 · p. 12

4 Optimality theory

• The threshold intervention in Example 1 is not pathwise differentiable, so its effect cannot be estimated at the n^1/2 rate in the nonparametric model.The parameter includes the effect of a static intervention setting a continuous exposure to δ, for which nonparametric efficient estimation theory is unavailable.Source: §4 · p. 13
• The paper avoids this limitation by restricting to interventions satisfying piecewise smooth invertibility for continuous exposures, enabling change-of-variable calculations.The exposure support may be partitioned into subintervals on which the intervention has an inverse with a derivative.Source: §4 · p. 14
• With discrete exposures or continuous exposures satisfying Assumption 4, and a policy d that does not depend on P, the EIF resembles that for dynamic regimes.This structure permits multiply robust estimation and generalization of existing dynamic-regime estimators.Source: §4 · pp. 14–15
• The EIF approximation supports consistency when, at each time point, either the treatment mechanism or outcome regression is consistently estimated.The property is notable because the parameter depends on both quantities at every time point, and the EIF also supports estimation under slow nuisance-estimation rates.Source: §4 · pp. 16–17

5 Estimation and statistical inference

• Data-adaptive regression methods, including Super Learner, are advocated for conditional expectations; density-ratio estimation can be recast as classification using 2n observations.Super Learner combines algorithms from a user-given library with weights chosen to minimize cross-validated prediction error.Source: §5 · p. 17
• The TMLE is locally efficient, n^1/2-consistent, and τ + 1-multiply robust under assumptions, while the sequential regression estimator has the additional property of 2τ-multiply robust consistency.Both are presented as alternatives to substitution and IPW estimators that generally cannot achieve n^1/2-consistency with data-adaptive nuisance estimation in a non-parametric model.Source: §5 · p. 18

5.1 Substitution and inverse probability weighted estimators

• Substitution and IPW estimators are asymptotically linear and n^1/2-consistent under correctly specified parametric models, but generally require parametric-rate nuisance estimation for that consistency.With data-adaptive estimation of high-dimensional regressions, this rate requirement is generally unattainable, making the asymptotic distribution unknown and inference difficult.Source: §5.1 · pp. 18–19

5.2 Targeted minimum loss-based estimator

• The TMLE uses sample splitting and cross-fitting, then tilts preliminary outcome-regression estimates toward a solution of the efficient influence-function estimating equation.Its n^1/2-consistency requires consistent estimation of all nuisance parameters at the theorem’s stated rates.Source: §5.2 · pp. 19, 20, 22
• TMLE is consistent under τ + 1 nuisance-consistency configurations, including configurations where treatment-ratio estimates are consistent through time k and outcome-regression estimates thereafter.This result assumes the conditions of Theorem 2.Source: §5.2 · p. 22

tions

• The sequential regression estimator is 2τ-multiply robust: at each time, either the treatment-ratio estimate or the transformed outcome-regression estimate may be consistent.It uses transformation (6) to estimate each m_t without requiring later m_s estimates to be consistent.Source: §5.2 · p. 23, tions · p. 25
• TMLE and sequential regression require the same nuisance-estimation rates for n^1/2-consistency, and the latter offers no apparent asymptotic advantage.The sequential estimator’s weak-convergence result supports Wald-type confidence intervals using an empirical-variance standard error.Source: tions · pp. 25–26
• The proposed TMLE and sequential regression estimators do not provide multiply robust confidence regions or p-values, and sequential-regression estimates can fall outside the parameter bounds.The latter risk may be especially relevant near positivity violations, when the density ratio can be highly variable; truncation can address this issue.Source: tions · p. 26

5.4 Density ratio estimation via classification

• A possible strategy for estimating the required density ratio is to estimate the treatment density and plug it into the ratio’s definition.The passage introduces this approach but does not provide further details about classification.Source: §5.4 · p. 27
• The proposed alternative recasts density-ratio estimation as classification in an augmented dataset.The motivation is that data-adaptive estimators for high-dimensional conditional densities are scarce.Source: §5.4 · p. 27
• A classifier can estimate the density ratio by distinguishing duplicated records assigned the observed versus intervened exposure.The augmented data include an indicator for whether a record corresponds to the intervention.Source: §5.4 · pp. 27–28
• Because augmented rows are not i.i.d., perform cross-validation over original observation indices i = 1, ..., n, rather than augmented pairs λ = 0, 1; i = 1, ..., n, to ensure independencies required for estimator properties such as asymptotic normality.Source: §5.4 · p. 28

6 Illustrative application

• The intervention raised the P/F ratio by 50 among patients with acute respiratory failure (P/F < 300), beginning at invasive ventilation; the authors described this increase as feasible and meaningful.The cohort comprised 10,044 intubated ICU patients, and the outcome was 14-day survival.Source: §6 · pp. 28–29
• The analysis adjusted for age, sex, race, Elixhauser comorbidities, pneumonia, daily ventilation status, and SOFA score without its pulmonary component; discharge triggered censoring because later vital status was unknown.Source: §6 · p. 29
• Substitution, IPW, TMLE, and SDR used Super Learner models for outcome and treatment components, conditioning on complete histories without Markov assumptions.Candidate learners included MARS, extreme gradient boosting, LASSO logistic regression, and simple logistic regression.Source: §6 · p. 30
• 14-day survival was 88.3% (95% CI 87.5, 89.0%) without intervention versus 89.8% (95% CI 88.0, 91.5%) under the policy, a 1.5% increase (95% CI -0.1, 3%).Estimates used SDR to adjust for right censoring.Source: §6 · p. 30
• Substitution, TMLE, and SDR produced comparable results, whereas IPW estimated a 22% survival increase, attributed to variable density-ratio weights rather than positivity violations.The weights had a 5.5% coefficient of variation, and their maximum value was 97.7; the authors attributed the IPW instability to weight variability, not positivity violations.Source: §6 · p. 30

• The paper is titled “Non-parametric causal effects based on longitudinal modified.”Source: Supplementary Materials for · p. 37
• The listed authors are Iván Díaz, Nicholas Williams, Katherine L. Hoffman, and Edward J. Schenck.Source: Supplementary Materials for · p. 37

1 Simulation study

• The study evaluated four nuisance-estimation scenarios, ranging from consistent estimation of all nuisance parameters to inconsistency of all estimators.It also examined mixed consistency across time points for m_t and r_t estimators.Source: §1 · p. 38
• With all nuisance estimators consistent, TMLE and SDR had vanishing n^1/2-scaled bias and achieved the non-parametric efficiency bound.These results are shown in Figure 1, panels E and I.Source: §1 · p. 39, Figure 1 · p. 39
• When only Lemma 4 conditions, but not Lemma 2 conditions, were satisfied, only the SDR estimator seemed consistent, illustrating the additional benefits of 2τ-multiple robustness.Source: §1 · p. 39
• Substitution seemed consistent when all m_t estimators were consistent but lacked n^1/2-rate bias convergence; IPW seemed consistent when all r_t estimators were consistent.IPW was n^1/2-consistent in that setting, though this is not generally expected; it was inconsistent when any r_t estimator was inconsistent, and all estimators were inconsistent when all models were misspecified.Source: §1 · p. 39
• With correctly specified models, the EIF empirical standard deviation seemed to underestimate standard error, yielding slightly below-nominal coverage for TMLE and SDR.In scenarios 2–4, n^1/2-bias produced zero asymptotic coverage for both estimators, with worse TMLE performance in scenario 3.Source: §1 · p. 40

2 Identification (Theorem 1)

• Under Assumptions 1 and 2, the counterfactual mean is identified by applying Kennedy (2019)’s result and substituting the specified policy.The supplied passage refers to the substitution as “(-1)” but does not provide the full identifying expression.Source: §2 · p. 40
• The theorem’s definition follows from recursively evaluating the integrals, using iterated expectation and the stated assumptions.The proof attributes equalities to iterated expectation, the definition of Y(d), Lemma 1, and Assumptions 1 and 3.Source: §2 · pp. 41–42

3 Efficient influence functions (Theorem 2)

• The proof assumes d is smooth and invertible over the range of A; under Assumption 4, it partitions that range.The passage states that the proof under Assumption 4 follows by partitioning the range of A as in the assumption.Source: §3 · p. 42
• To conjecture the EIF, the proof first considers a discrete measure ν, where integrals can be written as sums.It then uses Lemma 2 in expressing the non-parametric quantity.Source: §3 · p. 42
• The Delta-method derivation for a substitution estimator yields influence function φ1(Z; η) after rearranging terms.The proof applies the result to the substitution estimator and its class of functions.Source: §3 · p. 43
• The proof verifies that φ1(Z; η) satisfies the EIF condition using the expansion implied by Lemma 1.The stated result follows after applying the expansion and its subsequent identity.Source: §3 · pp. 44–45

4 Sequential double robustness (Lemma 1)

• The lemma’s proof applies the stated relation recursively for s = t + 1, …, τ.The supplied passage does not include the relation itself.Source: §4 · p. 45
• The relation follows from the change-of-variable formula under Assumption 4.Source: §4 · p. 45

5 Second order representation for SDR (Lemma 3)

• The proof denotes the sequence by ˇη = (ˆr1, ˇm1, . . . ,ˆrτ, ˇmτ).Source: §5 · p. 46
• Substituting equation (7) into equation (6), then iterating the equality for t = τ − 1, . . . , 0, yields the representation.The proof suppresses function arguments when they are clear from context.Source: §5 · p. 46
• The proof uses the lemma’s assumption to establish a bound involving a constant C1.The supplied passage does not include the bound’s expression.Source: §5 · p. 46
• The final step follows from the law of iterated expectation and the Cauchy–Schwarz inequality.Source: §5 · p. 46

6 Asymptotic Normality of TMLE (Theorem 3)

• The TMLE solves the efficient influence function estimating equation, which the proof uses to begin its expansion.The proof writes the influence function as φη(Z) = φ1(Z; η).Source: §6 · p. 47
• The proof makes explicit that the estimated influence function depends on the full sample through the generalized linear tilting-model parameters.It then conditions on the training data and invokes bracketing-entropy results to establish Rn,1 = oP(1).Source: §6 · pp. 47–48
• Conditional on Tj, Rn,1 = oP(1); Lemma 1, the Cauchy-Schwartz inequality, and the theorem’s assumptions yield Rn,2 = oP(1), concluding the proof of the theorem.Source: §6 · p. 48

7 Other results

• Lemma 2 states a result under Assumption 4, and its proof derives it by recursively applying a change-of-variable formula.The supplied passage identifies gd as defined in equation (3) of the main document but does not include the lemma’s formula.Source: §7 · p. 48

8 Supplemental Figures

• Figure 3 displays average cross-fitting-fold learner weights for estimating rt and mt at each time point in the illustrative ICU study.The study concerns P/F ratios and 28-day mortality.Source: Figure 3 · p. 49
• Figure 4 shows a histogram of log10(Q14) in the illustrative example.The accompanying passage begins describing values of Q14 but is truncated, so it does not support further characterization of the histogram.Source: Figure 4 · p. 50, §8 · p. 50