Structural Nested Mean Models (SNMMs)

• SNMMs are semiparametric causal models that define treatment effects via conditional contrasts (blip functions) to capture dynamic responses in longitudinal data.
• They use G-estimation with nuisance models, achieving double robustness under sequential ignorability to address time-varying confounding.
• SNMM frameworks extend to various applications—such as HIV treatment timing and policy evaluation—by modeling causal heterogeneity and optimizing efficiency.

Searching arXiv for recent and foundational papers on Structural Nested Mean Models. Structural nested mean models (SNMMs) are semiparametric causal models for longitudinal data with time-varying treatment and time-varying confounding. Rather than modeling the observed outcome process directly, an SNMM parameterizes conditional causal contrasts—often called blip effects—between potential outcomes under different treatment histories, and then uses blipped-down or mimicking outcomes together with G-estimation to identify and estimate those contrasts. Within the broader class of structural nested models, SNMMs are the mean-model analogue of structural nested distribution models, with related survival versions given by structural nested cumulative failure time models and structural nested failure time models (Vansteelandt et al., 2015).

1. Conceptual framework and historical position

Structural nested models and G-estimation were proposed by James Robins over two decades ago as approaches to modeling and estimating the joint effects of a sequence of treatments or exposures (Vansteelandt et al., 2015). In that framework, SNMMs model the mean outcome, whereas structural nested distribution models target the entire outcome distribution. For survival outcomes, the corresponding special cases are structural nested cumulative failure time models and structural nested failure time models (Vansteelandt et al., 2015).

At the point-treatment level, an SNMM is defined through a conditional causal contrast among subjects who actually received treatment level $a$: $$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$ Here $g(\cdot)$ is a known link and $\gamma^*(l,a;\psi)$ is a known blip function, smooth in $\psi$, with $\gamma^*(l,0;\psi)=0$ (Vansteelandt et al., 2015). In this formulation, the parameter often represents an effect in the treated, possibly modified by covariates $L$, rather than a population-average effect in the marginal structural model sense (Vansteelandt et al., 2015).

For time-varying treatments, the model becomes recursive. The blip at time $m$ compares potential outcomes under histories $(\overline{a}_m,0)$ and $(\overline{a}_{m-1},0)$: $$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$0 The term “nested” reflects that the effect at time $$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$1 is defined within the subgroup determined by past covariates and treatments (Vansteelandt et al., 2015).

This structure distinguishes SNMMs from conventional outcome regression. Standard regression models something like $$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$2 directly, whereas SNMMs model causal increments between counterfactual trajectories (Shahn et al., 2022). The distinction is especially consequential when intermediate covariates are affected by earlier treatment, because naive conditioning on such variables can block mediation pathways or induce collider bias (Vansteelandt et al., 2015, Wodtke et al., 2018).

2. Blip functions, coarse SNMMs, and blipped-down outcomes

A central device in SNMMs is the blip-down transformation: remove the hypothesized treatment effect from the observed trajectory and study the transformed quantity as if it were generated under a less-treated or untreated regime. For repeated measures with identity link, one representation is

$$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$3

which recursively peels off later treatment effects (Vansteelandt et al., 2015). Under the true parameter, the transformed outcome should satisfy conditional mean restrictions relative to current treatment.

A widely used tractable subclass is the coarse structural nested mean model. In a treatment-initiation setting with discrete, coarsened decision times, the model targets the effect of initiating treatment at month $$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$4 and then following up to month $$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$5: $$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$6 Here $$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$7 is the potential outcome if treatment began at month $$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$8, $$g \bigl\{E\bigl(Y^a \mid L=l,A=a\bigr) \bigr\} - g \bigl\{E\bigl(Y^0 \mid L=l,A=a\bigr) \bigr\} = \gamma^*(l,a;\psi^*).$$9 is the potential outcome if treatment were never started, $g(\cdot)$0 is treatment initiation time, and $g(\cdot)$1 is covariate history up to $g(\cdot)$2 (Yang et al., 2015). In one HIV application, the treatment effect model was specialized to

$g(\cdot)$3

so that the effect accumulates with duration $g(\cdot)$4 and depends linearly on initiation month $g(\cdot)$5 (Yang et al., 2015).

The transformed outcome in the coarse-SNMM setting is often written as

$g(\cdot)$6

or, for end-of-study outcomes,

$g(\cdot)$7

These are designed to mimic untreated potential outcomes when the blip model is correctly specified (Yang et al., 2015, Lok et al., 2021). In general DiD-SNMM notation, the analogous blipped-down quantity is

$g(\cdot)$8

with $g(\cdot)$9, so the blip functions determine counterfactual mean trajectories under a regime $\gamma^*(l,a;\psi)$0 (Shahn et al., 2022).

These constructions clarify the scientific meaning of SNMM parameters. They do not describe the full observed-data mean surface. They parameterize the effect of removing, adding, or modifying one component of treatment history, conditional on the observed history up to that time (Vansteelandt et al., 2015, Shahn et al., 2022).

3. Identification assumptions and G-estimation

Identification of SNMM parameters classically relies on no unmeasured confounding. For point treatments this is

$\gamma^*(l,a;\psi)$1

and for sequential treatments it becomes

$\gamma^*(l,a;\psi)$2

This is sequential ignorability or exchangeability (Vansteelandt et al., 2015). Together with consistency and positivity, it implies that the correctly blipped-down outcome behaves as though it were mean-independent of current treatment conditional on observed history (Vansteelandt et al., 2015, Lok et al., 2021).

That property yields the G-estimating equations. A canonical form is

$\gamma^*(l,a;\psi)$3

the sequential analogue of the point-treatment estimating equation (Vansteelandt et al., 2015). In coarse SNMMs for treatment initiation, one such estimating function is

$\gamma^*(l,a;\psi)$4

where $\gamma^*(l,a;\psi)$5 models treatment initiation and the conditional mean of $\gamma^*(l,a;\psi)$6 is a nuisance regression (Yang et al., 2015).

SNMM estimation therefore requires nuisance models for the treatment mechanism and for the transformed outcome. In the overview formulation, one specifies a treatment distribution $\gamma^*(l,a;\psi)$7 and a model for the distribution or mean of $\gamma^*(l,a;\psi)$8, or their sequential analogues (Vansteelandt et al., 2015). In the coarse-SNMM HIV setting, the treatment initiation model was a pooled logistic regression for $\gamma^*(l,a;\psi)$9, and the nuisance outcome model was the conditional mean of the blipped-down outcome (Yang et al., 2015).

A major property of G-estimation in this setting is double robustness. For linear and loglinear structural mean models and structural nested distribution models, the G-estimator is consistent if either the treatment model $\psi$0 or the outcome/transformed-outcome model $\psi$1 is correctly specified, provided the structural model and ignorability are correct (Vansteelandt et al., 2015). The same pattern appears in coarse SNMMs: if the treatment effect model is correctly specified, the estimator is consistent if either the treatment initiation model or the nuisance regression outcome model is correct (Yang et al., 2015, Lok et al., 2021).

A recurrent clarification is that this protection does not extend to misspecification of the blip model itself. The treatment effect model is the crucial scientific component; if it is wrong, causal effect estimates can be biased even when nuisance models are correct (Yang et al., 2015).

4. Efficiency, optimal weighting, and model checking

Within the broad class of unbiased estimating equations, SNMM estimators can differ substantially in precision. For coarse SNMMs, centered estimating equations improve efficiency over uncentered ones. In one formulation,

$\psi$2

and $\psi$3 is the projection of the corresponding uncentered estimating function onto the orthocomplement of a nuisance subspace, implying that the centered estimator is at least as efficient (Lok et al., 2021).

Under additional assumptions there is an explicit optimal choice of instrument or weight. In the longitudinal-outcome setting, the optimal $\psi$4 solves a conditional covariance equation, and under homoscedasticity the weight simplifies to an inverse-variance form (Lok et al., 2021). A related optimal instrument in the goodness-of-fit paper is

$\psi$5

which is analogous to optimal instruments in generalized method of moments (Yang et al., 2015).

SNMM methodology also contains an internal mechanism for assessing misspecification of the treatment effect model. A doubly robust goodness-of-fit test for coarse SNMMs can be constructed from overidentification restrictions. The test statistic is

$\psi$6

and, under the null that the treatment effect model is correctly specified, it has a chi-squared limiting distribution with degrees of freedom equal to the number of overidentification restrictions if either the treatment initiation model or the nuisance regression outcome model is correctly specified (Yang et al., 2015).

Several later developments broadened efficiency and robustness arguments. For continuous-time SNMMs with irregularly spaced observations, semiparametric efficiency theory yields locally efficient estimators, and inverse probability of censoring weighting gives a multiple robustness feature in the presence of dependent censoring (Yang, 2018). For repeated outcomes with unknown effect modifiers, SCAD-penalized G-estimation provides simultaneous modifier selection and causal estimation, with sparsity, asymptotic normality, oracle property, and double robustness under the stated conditions (Jaman et al., 2024).

A practical implication is that SNMM performance depends not only on causal assumptions such as consistency, positivity, and no unmeasured confounding, but also on how well the blip parameterization captures effect heterogeneity. This suggests that efficiency theory and model checking are not peripheral technicalities; they are central to valid inference in structural nested analyses.

5. Major variants and extensions

The SNMM literature now includes several distinct subclasses and identification regimes. The table summarizes representative extensions developed in the cited papers.

Variant	Defining feature	Representative source
Coarse SNMM	Treatment changes at discrete, coarsened times; often models treatment initiation time	(Yang et al., 2015, Lok et al., 2021)
Binary multiplicative SNMM	Reparametrizes nuisance parameters to achieve variation independence	(Wang et al., 2017)
Repeated-outcome SNMM	Session-specific or proximal effects with within-subject correlation	(Jaman et al., 2024)
Continuous-time SNMM	Irregularly spaced observations and martingale no unmeasured confounding	(Yang, 2018)
DiD-SNMM	Identification under conditional parallel trends rather than sequential exchangeability	(Shahn et al., 2022)
MTP-SNMM	Blip effects for modified treatment policies tied to natural treatment values	(Shahn, 26 Sep 2025)
SNMM with interference	Direct and spillover blips under cluster or network interference	(Shahn et al., 2024)
Neural SNMM	Simultaneous learning of blip functions with LSTM or transformer encoders	(Ma et al., 18 Nov 2025)

Binary outcomes present a distinctive modeling issue. For longitudinal binary data, additive SNMMs are variation dependent on each other, whereas multiplicative SNMMs are variation independent of each other (Wang et al., 2017). A reparametrization based on a generalized odds product yields nuisance parameters that are variation independent of the causal parameters, enabling coherent modeling and providing a key building block for flexible doubly robust estimation (Wang et al., 2017). The same work argues that an additive SNMM with binary outcomes does not admit a variation independent parametrization, which explains a longstanding difficulty in estimation and interpretation for binary additive blips (Wang et al., 2017).

Another major line of work replaces no unobserved confounding with conditional parallel trends. In the DiD-SNMM framework, blip functions are identified from invariance of trends in blipped-down outcomes under a reference regime $\psi$7, allowing estimation of heterogeneous initiation effects, lasting effects of a single blip, sustained interventions, controlled direct effects, dynamic policy effects, and optimal regimes under parallel-trends-type assumptions (Shahn et al., 2022). That framework has been extended to settings with cluster or network interference, where the blip can encode direct exposure, indirect exposure, or both (Shahn et al., 2024).

Modified treatment policies generalize the regime itself. In that setting, the intervention maps the observed history and natural treatment value into a modified treatment value $\psi$8, and the SNMM blip

$\psi$9

captures the effect of one final blip of the observational regime before moving onto the modified policy (Shahn, 26 Sep 2025). Under exchangeability assumptions this extension supports Neyman-orthogonal, cross-fitted estimation compatible with machine learning nuisance regressions; under parallel trends assumptions it extends SNMM reasoning to settings with some unobserved confounding (Shahn, 26 Sep 2025).

Recent methodological work also adapts SNMMs to new data structures and learning architectures. A two-step within-person variability score method first removes stable trait factors through structural equation modeling and then applies SNMMs with G-estimation to within-person scores, aiming at within-person causal effects of time-varying continuous treatments (Usami, 2020). DeepBlip, described as the first neural framework for SNMMs, replaces classical function estimation with neural networks, uses a double optimization trick to enable simultaneous learning of all blip functions, and employs a Neyman-orthogonal loss function for robustness to nuisance misspecification (Ma et al., 18 Nov 2025).

6. Applications, comparative position, and unresolved issues

SNMMs are used when treatment is time-varying and confounding is itself affected by prior treatment. Representative applications include initiation of highly active antiretroviral treatment in HIV-positive patients (Yang et al., 2015), optimal estimation of ART timing in recently infected HIV patients (Lok et al., 2021), continuous-time analysis of HAART initiation with irregular visit times (Yang, 2018), repeated hemodiafiltration outcomes and effect modifier selection (Jaman et al., 2024), neighborhood disadvantage and math achievement as well as education and mental health mediation (Wodtke et al., 2018), Medicaid expansion, floods, and sustained temperature changes under DiD-SNMMs (Shahn et al., 2022), COVID mobility under modified treatment policies (Shahn, 26 Sep 2025), sleep duration and depressive symptoms after stable-trait adjustment (Usami, 2020), and weight gain under sulfonylurea versus metformin treatment in an instrumented difference-in-differences setting (Vo et al., 2022).

Relative to competing approaches, SNMMs occupy a specific niche. Ordinary regression can be more efficient if correctly specified, but in the presence of time-varying confounders affected by prior treatment it conditions on post-treatment variables in ways that can introduce collider bias and block mediation pathways (Vansteelandt et al., 2015). Inverse probability weighting and marginal structural models target marginal effects and are often simpler computationally, but they are typically less efficient, especially when the propensity score is near $\gamma^*(l,0;\psi)=0$0 or $\gamma^*(l,0;\psi)=0$1, and they average over effect heterogeneity instead of allowing treatment effects to vary with covariates (Vansteelandt et al., 2015). SNMMs, by contrast, explicitly model effect modification by time-varying confounders and can borrow information across strata (Vansteelandt et al., 2015). Regression-with-residuals implementations of moderately constrained SNMMs were developed partly to address the unrealistic no-effect-moderation assumption in earlier highly constrained SNMMs for marginal effects (Wodtke et al., 2018).

Several common misconceptions recur in the literature. One is that an SNMM is simply an outcome regression written in counterfactual notation. The literature instead treats the SNMM as a model for causal contrasts, with nuisance associations handled separately (Wodtke et al., 2018). A second is that double robustness protects against misspecification of the structural effect model. In coarse SNMMs it protects against misspecification of nuisance models, not against misspecification of $\gamma^*(l,0;\psi)=0$2 itself (Yang et al., 2015, Lok et al., 2021). A third is that all SNMM estimands are marginal or population-average. Many SNMM parameters are conditional effects in the treated or in history-defined subgroups (Vansteelandt et al., 2015).

The main obstacle to broader use of SNMMs has been described as computational rather than conceptual (Vansteelandt et al., 2015). G-estimation requires models for treatment and transformed outcomes, and in some settings—especially survival outcomes with artificial censoring—optimization can be difficult and unstable (Vansteelandt et al., 2015). Other unresolved issues are more structural. Under parallel trends with interference, valid inference in network settings remains an open problem because standard bootstrap and sandwich methods may fail when observations are not independent (Shahn et al., 2024). Under modified treatment policies, orthogonal machine-learning-friendly estimation has been developed for the exchangeability-based construction but not yet for the parallel-trends extension (Shahn, 26 Sep 2025). For binary outcomes, the variation dependence of additive SNMMs remains a fundamental limitation rather than a merely computational nuisance (Wang et al., 2017).

Across these variants, the enduring contribution of SNMMs is their decomposition of longitudinal causal effects into localized blips that can be removed, recombined, or optimized. That decomposition supports conditional effect heterogeneity, dynamic treatment regime analysis, policy evaluation, and formal model criticism in settings where standard regression or weighting approaches are often least satisfactory (Vansteelandt et al., 2015, Shahn et al., 2022, Ma et al., 18 Nov 2025).