G-Estimation of Structural Nested Models

G-estimation of structural nested models is a semiparametric longitudinal causal inference methodology in epidemiology, biostatistics, and public health introduced by American epidemiologist and biostatistician James M. Robins between 1989 and 1994. Together with the parametric G-formula (1986) and marginal structural models (2000), structural nested models (SNMs)—principally structural nested mean models (SNMMs) for continuous or binary outcomes and structural nested failure time models (SNFTMs) for time-to-event survival outcomes—form the third foundational pillar of Robins's G-methods for estimating the causal effects of time-varying exposures in the presence of treatment–confounder feedback.

Whereas marginal structural models parameterize the marginal counterfactual expectation E[Y^ā] unconditional on time-varying confounders L̄_K and rely on inverse probability of treatment weighting (IPTW), structural nested models parameterize the conditional causal blip—the incremental causal effect of receiving a final "blip" of treatment at visit m followed by control (zero treatment) thereafter, conditional on the patient's observed treatment and covariate history (Ā_m, L̄_m) up to visit m. Robins's G-estimation procedure estimates the structural blip parameter vector ψ₀ by computing, for each candidate parameter ψ, the imputed untreated potential outcome H_m(ψ) (conceptually "peeling back" the causal effect of treatments received from time m onward) and finding the unique value ψ̂ for which H_m(ψ̂) is statistically independent of actual treatment assignment A_m given past history (Ā_m−1, L̄_m).

Effect Modification by Time-Varying Covariates and Positivity Violations

In longitudinal cohort studies and randomized controlled trials with treatment switching, a time-varying biomarker L_m (such as CD4⁺ T-lymphocyte count in HIV infection or absolute neutrophil count during chemotherapy) simultaneously acts as a time-dependent confounder via confounding by indication (L_m → A_m and L_m → Y) and as a mediator affected by prior therapy (A_m−1 → L_m). Standard regression models conditioning on L_m induce overadjustment and collider-stratification bias (related to the Table 2 fallacy and immortal time bias).

Although both the parametric G-formula and marginal structural models resolve treatment–confounder feedback, three structural capabilities motivated Robins to develop structural nested models and G-estimation:

1. Direct Estimation of Time-Varying Effect Modification: Clinicians frequently need to know whether the causal benefit of administering treatment A_m at visit m is modified by the patient's current biomarker value L_m at visit m. A marginal structural model E[Y^ā | L₀] averages over post-baseline covariates L₁, …, L_K and cannot include L_m (m ≥ 1) in its conditioning set without re-introducing post-treatment conditioning bias. Structural nested models condition on L̄_m directly inside the blip function at visit m.
2. Immunity to the G-Null Paradox: Unlike the parametric G-formula—where main-effects parametric models for f(l_m | ·) and E[Y | ·] are incompatible with the sharp causal null hypothesis whenever A_m−1 → L_m → Y—a structural nested model encodes the sharp causal null hypothesis directly via ψ₀ = 0.
3. Consistency Under Deterministic Positivity Violations: When clinical safety protocols mandate withholding treatment (P(A_m = 1 | L̄_m, Ā_m−1) = 0) whenever a toxicity biomarker L_m crosses a threshold, IPTW estimators of marginal structural models break down because the inverse weight 1 / P(A_m | L̄_m) is infinite. G-estimation does not divide by treatment probabilities; patients in deterministic strata where Var(A_m | L̄_m, Ā_m−1) = 0 receive zero weight (A_m − E[A_m | L̄_m, Ā_m−1] = 0) at visit m while still contributing at visits where treatment varies.

Counterfactual Notation for "Last Blip of Treatment" Regimes

Consider a cohort of i = 1, …, n individuals followed across discrete visits m = 0, 1, …, K, with covariate history L̄_m = (L₀, …, L_m), treatment history Ā_m = (A₀, …, A_m), and end-of-study outcome Y measured at time K + 1. For any visit m ∈ {0, …, K} and treatment prefix ā_m = (a₀, …, a_m−1, a_m), define two counterfactual future regimes from visit m through K:

• Regime (ā_m, 0): The individual follows history ā_m through visit m (receiving a_m at visit m), and receives zero treatment at all subsequent visits m + 1, …, K: (a₀, …, a_m−1, a_m, 0, …, 0).
• Regime (ā_m−1, 0): The individual follows history ā_m−1 through visit m − 1, and receives zero treatment from visit m onward: (a₀, …, a_m−1, 0, 0, …, 0).

At the final visit m = K, (ā_K, 0) equals the full history ā_K, whereas before baseline (m = 0), (ā₋₁, 0) is the never-treated counterfactual trajectory 0̄ = (0, 0, …, 0) with potential outcome Y^0̄.

Additive and Multiplicative Causal Blip Functions

An additive structural nested mean model (SNMM) specifies a parametric model γ_m(ā_m, l̄_m; ψ₀) for the causal blip function at each visit m = 0, …, K—the average causal effect of receiving treatment a_m at visit m versus 0 at visit m, assuming zero treatment after visit m, among subjects with history (Ā_m = ā_m, L̄_m = l̄_m):

By definition, γ_m((ā_m−1, 0), l̄_m; ψ) = 0 for all ψ. A linear blip specification allowing effect modification by current biomarker L_m and prior treatment A_m−1 takes the form:

Here, ψ₁ is the main effect of a dose at visit m, ψ₂ measures effect modification by the time-varying confounder L_m, and ψ₃ captures interaction with prior treatment A_m−1. For positive outcomes or counts, Robins (1994) also defined multiplicative (log-linear) SNMMs parameterizing log(E[Y^(ā_m,0) | Ā_m = ā_m, L̄_m = l̄_m] / E[Y^(ā_m−1,0) | Ā_m = ā_m, L̄_m = l̄_m]) = γ_m(ā_m, l̄_m; ψ₀).

Telescoping Sum and Imputed Untreated Potential Outcomes H_m(ψ)

Because the observed outcome Y equals Y^{(Ā_K, 0)} under consistency, the difference between Y and the counterfactual outcome Y^{(Ā_m−1, 0)} telescopes into the sum of causal blips from k = m to K. For any candidate parameter vector ψ, define the imputed blip-stripped outcome at visit m:

Intuitively, H_m(ψ) starts with the observed outcome Y and subtracts ("peels back") the modeled causal effects of all treatments received at visits m, m + 1, …, K. Under consistency and sequential conditional exchangeability, when evaluated at the true causal parameter ψ = ψ₀, H_m(ψ₀) has the same conditional expectation given (Ā_m, L̄_m) as the true counterfactual outcome Y^{(Ā_m−1, 0)}:

Sequential Exchangeability and the G-Estimating Equation

Under sequential conditional exchangeability (Y^{(ā_m−1, 0)} ⊥ A_m | Ā_m−1, L̄_m), the conditional mean of H_m(ψ₀) does not depend on A_m given (Ā_m−1, L̄_m):

Consequently, for any vector-valued index function d_m(L̄_m, Ā_m−1) of dimension dim(ψ), the conditional covariance between H_m(ψ₀) and the centered treatment residual A_m − E[A_m | Ā_m−1, L̄_m] is zero. Summing over all n individuals and all K + 1 visits yields Robins's unbiased G-estimating equation:

where ê_m(L̄_m, Ā_m−1) = Ê[A_m | Ā_m−1, L̄_m] is the fitted propensity score at visit m. When the blip function γ_m(·; ψ) = ψ^⊤ Z_m is linear in ψ, H_i,m(ψ) is linear in ψ, yielding a closed-form linear-algebraic solution for ψ̂_G. Subtracting an outcome-regression projection E[H_m(ψ) | Ā_m−1, L̄_m] from H_i,m(ψ) makes the G-estimator doubly robust and locally semiparametric efficient, analogous to targeted maximum likelihood estimation.

Counterfactual Untreated Survival Time and the Accelerated Failure Time Link

In survival analysis—where researchers compare survival distributions using the Kaplan–Meier estimator, Cox proportional hazards model, or accelerated failure time model—Robins and Tsiatis (1991) and Robins (1992) developed structural nested failure time models (SNFTMs). Let T_i denote subject i's observed failure time under continuous-time treatment trajectory {A_i(u) : 0 ≤ u ≤ T_i}, and let T_i^0̄ denote their untreated baseline potential survival time under A(u) = 0 for all u ≥ 0.

Under an accelerated failure time SNFTM, treatment A_i(u) expands or contracts biological survival time by exp(ψ₀ A_i(u)), linking T_i^0̄ to T_i via the path integral:

When ψ₀ < 0, active treatment A(u) = 1 prolongs survival by factor exp(−ψ₀) > 1; when ψ₀ = 0, treatment has zero causal effect (T_i(0) = T_i). Because T_i(ψ₀) = T_i^0̄ satisfies sequential exchangeability T_i(ψ₀) ⊥ A_i(t) | Ā_i(t⁻), L̄_i(t), G-estimation searches over a grid of candidate values ψ to find ψ̂ where T_i(ψ) has zero association with treatment initiation.

Artificial Censoring for Administrative End of Follow-Up

When follow-up ends at an administrative censoring time C_i, even if C_i is independent of T_i^0̄, the observed indicator I(T_i ≤ C_i) depends on treatment history Ā_i whenever ψ₀ ≠ 0. Robins (1992) restored conditional independence via artificial censoring: for any candidate ψ, define the guaranteed observable counterfactual censoring time C_i(ψ) = C_i if ψ ≥ 0 and C_i(ψ) = C_i exp(ψ) if ψ < 0 (for A(u) ∈ [0, 1]). Replacing T_i(ψ) in the G-estimating equation by the artificially censored indicator Δ_i(ψ) = I(T_i(ψ) ≤ C_i(ψ)) and follow-up time X_i(ψ) = min(T_i(ψ), C_i(ψ))—which depend only on (T_i^0̄, C_i) at ψ = ψ₀—yields unbiased tests and confidence intervals (verifiable via bootstrapping). This Rank-Preserving Structural Failure Time Model (RPSFTM) is widely used by health technology assessment agencies to adjust for control-arm treatment crossover in oncology randomized controlled trials.

Why is G-estimation immune to the G-null paradox when the parametric G-formula is not?

Under the sharp causal null hypothesis (Y^ā = Y^0̄ for all ā), the causal blip function γ_m(ā_m, l̄_m; ψ₀) is identically zero at every visit m. Because a structural nested model directly parameterizes γ_m(·; ψ) so that γ_m(·; 0) = 0, setting ψ = 0 represents the sharp null hypothesis without restricting how prior treatment A_m−1 alters L_m.

How is G-estimation used to adjust for treatment crossover in randomized oncology trials?

In a randomized controlled trial where control-arm patients (Z_i = 0) cross over to the experimental drug (A_i(t) = 1) upon disease progression, intention-to-treat underestimates the survival benefit while per-protocol analysis suffers from strong time-varying confounding. A Rank-Preserving Structural Failure Time Model (RPSFTM) (Robins and Tsiatis, 1991) uses baseline randomization Z_i ∈ {0, 1} in place of A_i(t) inside the G-estimating equation: because Z_i is independent of untreated survival time T_i^0̄ by randomization, G-estimation finds the survival acceleration factor ψ̂ for which T_i(ψ̂) has an identical distribution across randomized arms Z = 1 and Z = 0.

Can structural nested mean models be used when the outcome is binary rather than continuous?

Yes. For a binary outcome Y ∈ {0, 1}, a multiplicative (log-linear) structural nested mean model (Robins, 1994) works with G-estimation by defining H_m(ψ) = Y · exp(−∑_k=m^K γ_k(Ā_k, L̄_k; ψ)), preserving non-negativity. Logistic structural nested mean models (parameterizing conditional causal odds ratios) can also be estimated by modeling the association nuisance parameters alongside the blip function.

1. Robins, J. M. (1989). "The Analysis of Randomized and Non-Randomized AIDS Treatment Trials Using a New Approach to Causal Inference in Longitudinal Studies." In L. Sechrest, H. Freeman, & A. Mulley (Eds.), Health Service Research Methodology: A Focus on AIDS (pp. 113–159). Washington, DC: U.S. Public Health Service, National Center for Health Services Research. https://www.hsph.harvard.edu/james-robins/files/2013/03/holly.pdf
2. Robins, J. M., & Tsiatis, A. A. (1991). "Correcting for Non-Compliance in Randomized Trials Using Rank Preserving Structural Failure Time Models." Communications in Statistics—Theory and Methods, 20(8), 2609–2631. https://doi.org/10.1080/03610929108830654
3. Robins, J. M. (1992). "Estimation of the Time-Dependent Accelerated Failure Time Model in the Presence of Confounding Factors." Biometrika, 79(2), 321–334. https://doi.org/10.1093/biomet/79.2.321
4. Robins, J. M. (1994). "Correcting for Non-Compliance in Randomized Trials Using Structural Nested Mean Models." Communications in Statistics—Theory and Methods, 23(8), 2379–2412. https://doi.org/10.1080/03610929408831393
5. Vansteelandt, S., & Joffe, M. (2014). "Structural Nested Models and G-Estimation: The Partially Realized Promise." Statistical Science, 29(4), 707–731. https://doi.org/10.1214/14-STS493