Nuisance-Model Correctness Patterns and Robustness

Why Correctness Patterns Matter

Longitudinal TMLE uses several nuisance models: outcome regressions that predict later outcomes from observed histories, treatment models that describe treatment assignment, and censoring models that describe continued observation. A fitted model may be correct at some stages and misspecified at others. The relevant question is therefore not simply whether “the nuisance models” are correct, but which parts are correct, at which stages, and whether that pattern meets a robustness condition.

The sequential double-robustness conditions give a structured answer. They do not mean that arbitrary errors cancel, nor do they guarantee good finite-sample behavior. They describe sufficient patterns under which the population target of a properly targeted estimator remains correct, subject to consistency, positivity, and suitable targeting and estimation conditions.

Stage-Specific Notation and the Robustness Condition

Consider treatment decisions at times \(t=0,\ldots,K\), a fixed regime \(d=(d_0,\ldots,d_K)\), and a final outcome \(Y\). Let \(H_t\) be the observed history immediately before the treatment decision at time \(t\). Let \(C_{t+1}=1\) indicate that the outcome and history needed for the next stage are observed. Write \(g_t(a\mid H_t)\) for the conditional treatment probability and \(c_{t+1}(H_t,A_t)\) for the relevant conditional probability of remaining observed. The regime-specific treatment probability at stage \(t\) is \(g_t(d_t(H_t)\mid H_t)\).

Define the sequential outcome regressions by \(Q_{K+1}=Y\) and, recursively, \(Q_t(h_t)=E[Q_{t+1}\mid H_t=h_t,A_t=d_t(h_t),C_{t+1}=1]\). Thus \(Q_t\) is the conditional mean of the next continuation outcome among those following the regime and observed at that stage. The regime-specific mean is \(E[Q_0(H_0)]\). Hats denote fitted nuisance functions. “Correct” below means that the fitted function converges to the corresponding true conditional function on the relevant histories; it does not merely mean that a model has a plausible form.

For compactness, let \(p_t\) denote the stage-specific mechanism factor that enters the cumulative clever covariate: it includes the regime treatment probability and the relevant observation probability. The cumulative factor through stage \(t\) is the product of these factors over stages \(0,\ldots,t\). Correctness of this cumulative mechanism means that the factors used in that product consistently represent the true treatment and censoring probabilities needed to transport the stage-\(t\) residual to the target regime. A simple, stronger way to ensure it is for every treatment and censoring factor in the product to be correct.

This formulation separates two ways to protect a stage. A correct \(Q_t\) supplies the appropriate conditional continuation mean directly. Alternatively, correct mechanisms allow the residual information among people who followed the regime and remained observed to be transported with the appropriate weights. In an implementation, the treatment and censoring components should be checked separately even when their product is the object entering a weight: a product that happens to look plausible does not establish that either component is well estimated or stable.

Why the Residual Correction Has a Plus Sign

The sign of the correction is easiest to verify in a one-stage setting. Let \(Q^*(H)\) be an initial outcome prediction for a fixed treatment value \(a\), and let \(g(a\mid H)\) be the true treatment probability. Suppose the treatment mechanism is correctly used, positivity holds, and the residual is defined as observed outcome minus initial prediction. Then the weighted residual has population mean \(E[I(A=a)\{Y-Q^*(H)\}/g(a\mid H)]\). Conditioning on \(H\) shows exactly what this term adds to the standardized initial prediction.

Here \(Q_0(a,H)=E[Y\mid A=a,H]\). Averaging over \(H\), the initial prediction plus the weighted residual equals the standardized true conditional mean. This calculation explains why the correction is added: the residual is observed outcome minus prediction, so a negative residual moves the prediction down.

Worked Patterns Across Two Stages

For two treatment decisions, \(t=0,1\), the condition can be organized as a checklist. At stage 0, protection is supplied by \(Q_0\), or by the mechanism through stage 0. At stage 1, protection is supplied by \(Q_1\), or by the cumulative mechanism through stages 0 and 1. In each case the treatment and censoring factors required in the corresponding cumulative mechanism must be correct.

Reading a Correctness Table

The following table gives a compact way to assess typical patterns. “Mechanism through \(t\)” means the cumulative treatment and censoring factors required for the stage-\(t\) clever covariate. A check mark indicates that the corresponding stage has at least one of the two stated protections.

In practice, make one row per stage and label the actual nuisance fits: \(Q_t\), treatment probability, and censoring probability. Then record whether the relevant cumulative mechanism is correct, not just whether one isolated treatment model is correct. If the condition fails at any stage, report that the robustness guarantee is unavailable rather than asserting that the estimate is necessarily wrong.

Assumptions, Diagnostics, and Failure Modes

The robustness statements rely on the usual identification and estimation requirements. Consistency and the specified treatment regime must be well defined; conditional exchangeability must hold given the recorded history; and treatment and observation probabilities must be positive on histories relevant to the target regime. Nuisance estimates must converge in the sense required by the targeted estimator, and the targeting step must actually solve its specified score equations closely enough. For asymptotic inference, additional rate and empirical-process conditions are needed; correctness patterns alone do not establish a valid standard error.

Diagnostics should focus on whether the assumptions and fitted components are credible. Examine treatment and censoring probabilities by stage, especially their lower tails; inspect effective sample sizes under cumulative weights; compare observed histories represented in the target regime with the full cohort; and check whether the fluctuation converged and its score residual is small. Large or erratic cumulative weights can signal weak positivity or unstable mechanism fits. They do not establish that a particular nuisance model is wrong, but they warn that the intended transport may be poorly supported by the data.

A second failure mode is treating a model as “correct” because it includes familiar covariates or has good predictive performance. Correctness here concerns the relevant conditional function, not a favorable fit statistic. A treatment classifier can predict assignment well while misestimating probabilities in the histories that matter for the target regime. Similarly, a high-performing outcome model can still miss important conditional structure.

How This Differs From Nearby Approaches

The g-formula approach obtains a regime-specific mean by iterating outcome regressions, so its robustness primarily depends on those regressions being correct. Inverse-probability weighting instead relies on correctly estimated treatment and censoring mechanisms and can be sensitive to extreme weights. Longitudinal TMLE combines outcome regression with mechanism-based targeting: the stage-specific correctness patterns explain why it can remain consistent under certain forms of misspecification on either side. This is not immunity to misspecification; if the relevant outcome regression and cumulative mechanism are both wrong at a stage, the condition gives no protection there.

The condition is also not a promise of equal efficiency across patterns. Correct nuisance models can improve precision, while very small probabilities can make estimates unstable even when the relevant model is correctly specified. And when flexible nuisance estimation is used, good cross-validated prediction alone does not settle the rate conditions required for asymptotic linearity.