Deriving G-Estimation Moment Conditions

Why Blipping Leads to Moment Conditions

G-estimation uses a treatment-effect model to transform the observed outcome into an outcome with the modeled treatment effects removed. At the true blip parameters, the transformed outcome has no remaining conditional mean association with the treatment decision that was just removed, once the measured history is held fixed. This conditional restriction can be turned into unconditional equations that are useful for estimating the blip parameters.

The distinction between conditional mean independence and full independence matters. A structural nested mean model specifies conditional means of treatment-effect contrasts; by itself, it does not say that the entire distribution of the blipped-down outcome is the same across treatment levels. The moment equations derived here require the mean restriction, not that stronger distributional claim.

Setup and Recursive Blipping

Consider treatment decisions $A_1,\ldots,A_K$ and a final outcome $Y$. Let $H_t$ denote the measured history available immediately before $A_t$, including earlier treatments and covariates. Write $Y^{a_t,\ldots,a_K}$ for the potential outcome under the specified treatment sequence from time $t$ onward, with the history before $t$ held fixed. Consistency means the observed outcome equals the potential outcome under the observed treatment sequence.

Sequential exchangeability requires that, at each $t$, the treatment $A_t$ be conditionally independent of the relevant future potential outcomes given $H_t$. Positivity requires that treatment levels used in the blip contrasts have positive conditional probability for the histories to which those contrasts apply. These assumptions concern treatment assignment and counterfactual outcomes; neither alone specifies the blip model.

Definition: A structural nested mean model specifies treatment-effect blips $\gamma_t(H_t,A_t;\psi)$, indexed by parameter $\psi$, for the conditional mean effect of treatment at time $t$ after later treatment effects have been removed. For a chosen reference treatment coded $0$, define the recursively blipped-down outcomes by $U_{K+1}(\psi)=Y$ and $$ U_t(\psi)=U_{t+1}(\psi)-\gamma_t(H_t,A_t;\psi), \qquad t=K,\ldots,1. $$ The blip at the reference level is conventionally $\gamma_t(H_t,0;\psi)=0$. Thus $U_t(\psi)$ removes the modeled blips from time $t$ onward.

This recursion is a definition of the transformed observed outcome. Its interpretation at the true parameter depends on the SNMM and the causal assumptions: the later blips are removed first, and then the current treatment blip is removed. In an SNMM formulated through conditional mean contrasts, the resulting restriction is a statement about conditional expectations.

From Exchangeability to the Conditional Mean Restriction

Theorem: Suppose consistency and sequential exchangeability hold, and the blip functions at $\psi_0$ correctly specify the conditional mean treatment contrasts at each decision, after later blips have been removed. Suppose also that the conditional expectations below exist. Then, for each $t$, $$ E[U_t(\psi_0)\mid H_t,A_t]=E[U_t(\psi_0)\mid H_t] $$ almost surely. In particular, the conditional mean of $U_t(\psi_0)$ does not vary with $A_t$ after conditioning on $H_t$.

To see the role of the assumptions, consider the transformed outcome just before removing the blip at time $t$, namely $U_{t+1}(\psi_0)$. By the recursive construction and correct specification of later blips, its conditional mean contrast across values of $A_t$, given $H_t$, is the time-$t$ blip. Sequential exchangeability and consistency justify interpreting that contrast causally: within levels of $H_t$, the treatment groups do not differ in the relevant future potential outcomes before treatment. Correct specification says, for each supported treatment value $a$, $$ E[U_{t+1}(\psi_0)\mid H_t,A_t=a] - E[U_{t+1}(\psi_0)\mid H_t,A_t=0] = \gamma_t(H_t,a;\psi_0). $$ Since $\gamma_t(H_t,0;\psi_0)=0$, subtracting $\gamma_t(H_t,A_t;\psi_0)$ from $U_{t+1}(\psi_0)$ removes this conditional mean contrast. Therefore the conditional mean of $U_t(\psi_0)$ is the same at each supported treatment value, which gives the theorem. $\square$

This argument does not prove that the conditional distributions of $U_t(\psi_0)$ are equal across treatment levels. It proves equality of their conditional means. If a stronger individual-level model makes the blipped-down outcome equal to a counterfactual outcome unaffected by $A_t$, sequential exchangeability can support full conditional independence. That stronger conclusion is not part of the mean model stated here.

Deriving Unconditional Moments

The conditional restriction implies a family of unconditional equations. The integrability condition is essential: conditional expectations and products must exist, rather than merely looking formally well-defined.

Theorem: Fix $t$. Suppose $E[|U_t(\psi_0)|]<\infty$, and let $q(H_t,A_t)$ be a measurable function such that $E[q(H_t,A_t)\mid H_t]=0$ almost surely and $E[|q(H_t,A_t)U_t(\psi_0)|]<\infty$. Then $$ E[q(H_t,A_t)U_t(\psi_0)]=0. $$ A sufficient, but not necessary, integrability condition is that $q$ be bounded and $U_t(\psi_0)$ integrable.

Proof. Let $m(H_t)=E[U_t(\psi_0)\mid H_t]$, which is finite almost surely under the stated integrability condition. By the conditional mean restriction, $E[U_t(\psi_0)\mid H_t,A_t]=m(H_t)$. The product integrability assumption allows the conditional expectation and iterated-expectation steps below. Hence

\begin{aligned} E[q(H_t,A_t)U_t(\psi_0)] &=E\!\left[q(H_t,A_t)E\{U_t(\psi_0)\mid H_t,A_t\}\right]\\ &=E[q(H_t,A_t)m(H_t)]\\ &=E\!\left[m(H_t)E\{q(H_t,A_t)\mid H_t\}\right]\\ &=0. \end{aligned}

The second use of iterated expectation is justified because the product is integrable and $m(H_t)$ is the conditional expectation of an integrable random variable; the displayed terms are therefore well-defined under the stated hypotheses. If $q$ is bounded, then $E[|qU_t|]\leq \|q\|_\infty E[|U_t|]<\infty$. This proves both the moment and the stated sufficient condition. $\square$

For a binary treatment, a common choice is $q(H_t,A_t)=A_t-e_t(H_t)$, where $e_t(H_t)=E[A_t\mid H_t]$. This is bounded when $A_t$ is binary, has conditional mean zero by definition, and yields a treatment-residual moment. Other choices are possible, but each must satisfy the conditional mean-zero and product-integrability requirements.

Worked Examples

Worked Example: A Correct Blip Gives a Zero Moment

At one decision, let $A$ be binary with $P(A=1)=0.4$. Suppose $Y=U+2A$, where $U$ has mean 1 in both treatment groups and is integrable. The true additive blip is $\gamma(A;\psi_0)=2A$, so $U(\psi_0)=Y-2A=U$. Use $q=A-0.4$.

First, $E[q]=E[A]-0.4=0.4-0.4=0$. The conditional mean of $U$ given $A$ is 1, so $$ E[qU]=E\{qE[U\mid A]\}=E[q]=0. $$ Equivalently, $E[(A-0.4)(Y-2A)]=0$. The calculation uses only the equality of the conditional means; the distribution of $U$ could still differ between treatment groups.

Worked Example: Mean Independence Is Not Full Independence

Let $H$ be constant and $P(A=1)=0.4$. Set $U\mid A=0$ equal to 0 or 2 with probability $1/2$ each, and $U\mid A=1$ equal to $-1$ or 3 with probability $1/2$ each. Both conditional means equal 1, so $E[U\mid A]=1=E[U]$. With $q=A-0.4$, $$ E[qU]=P(A=1)(0.6)(1)+P(A=0)(-0.4)(1)=0.24-0.24=0. $$

Nevertheless, the distributions differ: for example, $P(U=0\mid A=0)=1/2$, whereas $P(U=0\mid A=1)=0$. Thus $U$ and $A$ are not independent. This example verifies that a valid g-estimation moment follows from conditional mean independence, not from full independence.

Worked Example: A Wrong Blip Leaves a Nonzero Moment

Let $A$ be binary with $P(A=1)=0.5$, and let $U$ be independent of $A$ with $E[U]=3$ and finite absolute mean. Generate $Y=U+2A$, so the true blip parameter in $\gamma(A;\psi)=\psi A$ is $\psi_0=2$. For a candidate $\psi$, the transformed outcome is $Y-\psi A=U+(2-\psi)A$. Take $q=A-0.5$.

Using independence and $E[qU]=E[q]E[U]=0$, $$ \begin{aligned} E[q(Y-\psi A)] &=E[qU]+(2-\psi)E[(A-0.5)A]\\ &=(2-\psi)\{0.5(1-0.5)\}\\ &=0.25(2-\psi). \end{aligned} $$ At $\psi=2$, the moment is 0. At $\psi=1$, it is $0.25$, not zero. In this simple model the moment has a unique zero at 2; in general, uniqueness requires additional conditions and does not follow merely from moment validity.

Choosing Moments, Diagnostics, and Alternatives

A g-estimation procedure selects one or more treatment-residual functions $q$ and solves sample analogues of the population moments. The derivation establishes why valid choices have expectation zero at the true parameter. It does not establish that a selected set of moments identifies $\psi_0$, nor that a finite-sample solution exists or is numerically stable.

Several checks help reveal problems before interpreting a fitted parameter:

Check treatment support. If a treatment level used to define a blip contrast is absent for some histories, that contrast is not supported there. Sparse support can make the moment weak even when it is formally defined.
Check the transformation. Confirm that later blips are removed before earlier ones and that the reference treatment has zero blip. A sign error in $U_{t+1}-\gamma_t$ changes the moment's target.
Check moment behavior across choices of $q$. Strongly different estimates from reasonable treatment-residual choices can signal poor fit, weak information, or an inadequate treatment or blip specification. A zero sample moment by itself is not a proof of the causal assumptions.
Check numerical and distributional behavior. Extreme transformed outcomes or near-flat moment functions can produce unstable roots. Bounded $q$ protects product integrability when $U_t$ is integrable, but it does not ensure good precision or identification.

G-estimation differs in emphasis from the parametric g-formula and inverse-probability weighting. The g-formula models and averages outcome transitions under interventions; weighting reweights observed data to represent an intervention distribution. G-estimation instead uses a blip model and treatment-residual moments. These approaches rely on different modeling choices and can have different failure modes; none turns unmeasured confounding into measured history.

The main limitation here is the conditional mean model: it cannot establish equality of full conditional distributions, and a misspecified blip can leave residual treatment association. Furthermore, the moment equations can be valid but uninformative about some components of $\psi$. Identification depends on the treatment variation, history distribution, and sensitivity of the chosen moments to the parameter.

Common Mistake

Common mistake: “Because $E[q\mid H_t]=0$, the product moment $E[qU_t]=0$ follows for every integrable $q$ and integrable $U_t$.” This skips a necessary condition: separate integrability of $q$ and $U_t$ does not guarantee integrability of their product. Require $E[|qU_t|]<\infty$, or use a sufficient condition such as bounded $q$ and integrable $U_t$. Also, do not infer full independence from the conditional mean restriction.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

What conditional restriction does the correctly specified blipped-down outcome satisfy at time $t$?
Why does $q(H_t,A_t)=A_t-E[A_t\mid H_t]$ have conditional mean zero?
Is $E[U_t\mid H_t,A_t]=E[U_t\mid H_t]$ enough to conclude $U_t$ and $A_t$ are conditionally independent?
What additional condition makes the product moment derivation valid for a general $q$? Give one sufficient alternative.
Does a zero population moment at $\psi_0$ prove that $\psi_0$ is the unique solution? Why or why not?

Solutions. (1) It satisfies $E[U_t(\psi_0)\mid H_t,A_t]=E[U_t(\psi_0)\mid H_t]$. (2) By the definition $E[A_t-E[A_t\mid H_t]\mid H_t]=E[A_t\mid H_t]-E[A_t\mid H_t]=0$. (3) No; the conditional distributions may differ while their means agree. (4) Require $E[|qU_t|]<\infty$; bounded $q$ and integrable $U_t$ are sufficient. (5) No. Validity establishes that the true parameter is a root, but uniqueness requires additional identification conditions, such as moments that respond distinctly to different parameter values.