Common Errors in Blip-Down and G-Estimation

Why Blip-Down Errors Matter

Blip-down transforms an outcome by subtracting modeled treatment effects, then uses the transformed outcome to find parameter values compatible with the treatment assignment process. The logic is easy to misstate: in a mean structural nested model, a blip describes a conditional mean contrast. Subtracting it from a person’s observed outcome does not generally recover that person’s counterfactual outcome under no treatment. Confusing these statements can lead to incorrect derivations, invalid moment equations, and overconfident interpretations.

Three errors deserve particular attention: removing treatment effects in the wrong temporal order, conditioning on variables measured after the treatment whose effect is being estimated, and treating a blip parameter as though it were automatically an individual effect or a marginal effect. The earlier material on g-estimation provides the broader method; here the focus is on checking its logic and interpretation.

Conditional-Mean Blips Are Not Individual Transformations

Definition: Let $Y$ be an outcome, $A$ a binary treatment, and $H$ information measured before treatment. An additive conditional-mean blip model specifies $$ E[Y\mid H,A=1]-E[Y\mid H,A=0]=\gamma(H;\psi_0). $$ Here $\gamma(H;\psi_0)$ is the modeled treatment contrast in conditional means. It does not, without an additional individual-level model, assert that $Y^1-Y^0=\gamma(H;\psi_0)$ for every person.

The distinction is consequential. Suppose everyone has $Y^0=0$, while under treatment $Y^1$ equals 0 or 2 with equal probability. The conditional mean contrast is $1$. For a person with $Y^1=0$, subtracting that mean blip from the observed treated outcome gives $0-1=-1$, not the person’s $Y^0=0$. The subtraction can still be useful for a population moment; it is not a person-specific counterfactual reconstruction.

Worked Example: A Valid Mean Moment Despite Individual Heterogeneity

Let $A$ be randomized with $P(A=1)=0.5$. Let $Y^0=0$ for everyone, and let $Y^1$ be 0 or 2 with equal probability, independently of $A$. The conditional-mean blip is $\psi_0=E[Y^1-Y^0]=1$. Consider $U(\psi)=Y-\psi A$.

First check the two treatment-specific means: $E[Y\mid A=0]=0$ and $E[Y\mid A=1]=1$. Thus the model $E[Y\mid A]=m+\psi_0A$ holds with $m=0$ and $\psi_0=1$.

For the treatment-residual moment, the randomization probability is $p=0.5$. At the true parameter, $$ E[(A-p)U(\psi_0)] =E[(A-0.5)(Y-A)]. $$

If $A=0$, then $Y=0$, so the product is 0. If $A=1$, then $E[Y-1\mid A=1]=0$, so its conditional expected product is also 0. Therefore the overall expectation is 0. Yet for a treated person with $Y=0$, $U(1)=-1$, which is not that person’s $Y^0$. The moment works because of the conditional-mean restriction, not because blip-down recovers each individual counterfactual.

Deriving the Treatment-Residual Moment

For one binary treatment decision, suppose consistency and conditional exchangeability hold: $(Y^0,Y^1)$ is independent of $A$ given $H$. Assume positivity, $0<P(A=1\mid H)<1$ almost surely, and a correctly specified additive conditional-mean model $E[Y\mid H,A]=m(H)+\psi_0 A$. Define $p(H)=P(A=1\mid H)$ and $U(\psi)=Y-\psi A$.

Theorem: Under those assumptions, $$ E[(A-p(H))U(\psi_0)]=0. $$ If $E[\operatorname{Var}(A\mid H)]>0$, the population moment $E[(A-p(H))U(\psi)]=0$ has the unique root $\psi=\psi_0$.

Proof. The conditional-mean model gives $E[U(\psi_0)\mid H,A]=m(H)$. Conditional on $H$, therefore, $$ E[(A-p(H))U(\psi_0)\mid H] =E[(A-p(H))m(H)\mid H] =m(H)E[A-p(H)\mid H]=0, $$ because $E[A\mid H]=p(H)$. Taking expectations proves the moment equation.

For an arbitrary $\psi$, write $U(\psi)=U(\psi_0)-(\psi-\psi_0)A$. The moment becomes $$ E[(A-p(H))U(\psi)] =-(\psi-\psi_0)E[(A-p(H))A] =-(\psi-\psi_0)E[\operatorname{Var}(A\mid H)]. $$ The last equality follows from $E[(A-p)A\mid H]=E[A^2\mid H]-p^2=\operatorname{Var}(A\mid H)$. If this expectation is positive, the moment vanishes only at $\psi_0$. $\square$

The result depends on the conditional-mean model and the stated treatment probability. It does not require $U(\psi_0)$ to equal an individual counterfactual outcome. In repeated-decision additive mean SNMMs, the corresponding sequential restrictions are applied after subtracting the modeled blips for the relevant decisions. These restrictions concern conditional means of the blipped-down outcomes; they should not be justified by asserting pointwise equality with counterfactual outcomes.

Temporal Order: Remove Later Blips Before Earlier Ones

At decision time $t$, the history $H_t$ includes information available before $A_t$, including earlier treatment and covariate history, but not later variables. In a repeated-decision additive model, a later blip may depend on earlier treatment through that history. Consequently, a blip-down calculation must evaluate each blip at its correctly observed history and remove later-decision blips before interpreting the result as stripped of all treatment effects.

For example, with two decisions, the additive transformation used at the first decision has the form $Y-\gamma_2(H_2,A_2;\psi)-\gamma_1(H_1,A_1;\psi)$. The second-decision blip is evaluated using $H_2$, which includes $A_1$. This is a modeling and conditioning order, not a claim that the transformed individual outcome equals an individual counterfactual.

Worked Example: A Later Blip Depends on Earlier Treatment

Suppose a conditional-mean model has baseline component 10 and blips $\gamma_1(H_1,A_1)=2A_1$ and $\gamma_2(H_2,A_2)=(1+3A_1)A_2$. Consider the treatment history $A_1=1,A_2=1$. The modeled mean outcome is $10+2(1)+(1+3(1))(1)=16$.

Remove the later blip first. Its value is $1+3(1)=4$, because the later history includes $A_1=1$. The remaining modeled mean is $16-4=12$. Removing the first-decision blip then gives $12-2=10$, the baseline component.

A wrong-order calculation might first erase $A_1$ from the history and then evaluate the second blip as if $A_1=0$. That gives a later blip of 1, followed by subtraction of the first blip 2: $16-1-2=13$, not 10. The error is not arithmetic; it is evaluating $\gamma_2$ at the wrong history. In an actual mean SNMM, these subtractions describe the specified mean-scale transformation and do not imply that any individual’s counterfactual outcome has been recovered.

Do Not Condition on Post-Treatment Variables as Histories

A treatment model for decision $t$ conditions on $H_t$, the history available before $A_t$. A variable measured after $A_t$ is not part of that history merely because it predicts the outcome or treatment later. Conditioning on it can change the estimand by blocking part of a treatment effect, or can create bias by opening a noncausal association. In longitudinal settings, a variable affected by earlier treatment may properly enter a later treatment history; that does not make it appropriate to condition on when estimating the earlier treatment’s effect.

Worked Example: A Post-Treatment Collider Creates an Association

Let $A$ and an unmeasured binary variable $U$ be independent Bernoulli variables with probability 0.5. Let the outcome be $Y=U$, so treatment has no causal effect. Now define a post-treatment variable $M=A$ OR $U$. It is a common effect of $A$ and $U$.

Among people with $M=1$ and $A=0$, $U$ must equal 1, so $E[Y\mid M=1,A=0]=1$. Among people with $M=1$ and $A=1$, $U$ remains equally likely to be 0 or 1, so $E[Y\mid M=1,A=1]=0.5$. Conditioning on $M=1$ produces an apparent treatment contrast of $0.5-1=-0.5$, even though $Y$ is unchanged by treatment for every person.

The diagnostic lesson is to establish measurement time before including a variable in a treatment model or blip function. If a proposed covariate occurs after the treatment whose effect is being blipped down, ask whether it is a mediator, a collider, or a variable affected by prior treatment. Its predictive value alone does not justify conditioning on it.

A Blip Parameter Is Not Automatically a Marginal Effect

A blip parameter describes the contrast specified by the structural nested model, conditional on its history and scale. It may be constant across histories, or it may index a blip function that varies with history. Neither case automatically makes the parameter an individual effect or the marginal contrast $E[Y^1-Y^0]$. To obtain a marginal effect, one must define the target population and treatment regimes and average the appropriate conditional contrasts over the relevant history distribution, under assumptions that identify that average.

Worked Example: Effect Modification Changes the Marginal Contrast

Suppose a baseline covariate $H$ takes values 0 and 1 with equal probability. The conditional mean blip is $\gamma(H,A;\psi)=\psi(H)A$, where $\psi(0)=0$ and $\psi(1)=4$. The conditional treatment contrast is therefore 0 for $H=0$ and 4 for $H=1$.

For a target population with equal proportions in the two strata, the marginal mean contrast is $0.5(0)+0.5(4)=2$. The stratum-specific blip values are 0 and 4; neither equals the marginal contrast. If the population instead has $P(H=1)=0.25$, the marginal contrast is $0.75(0)+0.25(4)=1$. The target-population distribution matters. Thus reporting a blip coefficient as “the average treatment effect” requires additional justification, not just a fitted SNMM.

Diagnostics, Alternatives, and Limits

Several checks help locate these errors before interpreting estimates:

Check the time index. Write down exactly what is known immediately before each $A_t$. Confirm that every component of $H_t$ is available then.
Check each blip’s arguments. For a later decision, verify that its blip is evaluated at the full later history, including earlier treatment when the model specifies it.
Check conditional residual balance. At a proposed parameter, examine whether the blipped-down outcome has residual association with treatment after conditioning on the appropriate pre-treatment history. A persistent pattern can signal a misspecified blip or treatment model.
Check positivity and variation. If treatment is nearly deterministic within histories, the residual moment contains little information. A small or unstable estimated treatment-residual variance warns that the parameter may be weakly identified.
Check the estimand language. State whether a reported quantity is a history-specific blip parameter, a conditional contrast, or a marginal contrast over a named population.

G-estimation uses treatment-residual restrictions implied by the structural nested model. Inverse-probability weighting and outcome-regression approaches are common alternatives, but they organize estimation around different models and may target different summaries. They do not repair a post-treatment conditioning error or turn a conditional blip into a marginal effect automatically. All approaches remain vulnerable to inadequate positivity, unmeasured confounding, and model misspecification. A treatment-residual diagnostic can reveal incompatibility with a fitted model; it cannot prove that all causal assumptions hold.

Common Mistakes

“The blipped-down value is the person’s untreated outcome.” This does not follow from a conditional-mean blip. It requires a stronger individual-level structural model.
“Any outcome predictor belongs in the history.” Only variables available before the treatment decision belong in that decision’s history. Post-treatment variables can mediate or distort the contrast.
“Subtract the blips in any order.” Later blips may depend on earlier treatment. Evaluate them using their specified histories, removing later effects before earlier ones.
“The coefficient is the population-average effect.” A blip parameter has the interpretation supplied by its model. A marginal effect requires explicit averaging over a target population.

Key takeaway: Justify blip-down through conditional-mean restrictions, not an assumed individual counterfactual identity. Use pre-treatment histories, respect the temporal dependence of blip functions, and name the estimand before interpreting a parameter.

Check Your Understanding

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

A mean blip is 3, and a treated person’s observed outcome is 5. Can you conclude that this person’s untreated outcome is 2? Why or why not?
In a two-decision model, the second blip depends on $A_1$. Which value of $A_1$ should be used when evaluating that blip for an observed person?
A variable is measured after $A_1$ but before $A_2$. Can it be part of $H_2$? Does that mean it can be treated as pre-treatment history for estimating the effect of $A_1$?
Two baseline strata have treatment contrasts 1 and 5, and the target population is half in each stratum. What is the marginal contrast? Is it either stratum-specific blip?

Solutions. (1) No. A mean blip of 3 describes a conditional mean contrast; it does not generally identify this person’s counterfactual outcome as $5-3=2$. (2) Use the observed $A_1$ value included in that person’s $H_2$, because the second blip is defined as a function of that history. (3) It can be part of $H_2$ because it precedes $A_2$, but it is post-treatment for $A_1$ and should not be treated as baseline history for the earlier effect. (4) The marginal contrast is $0.5(1)+0.5(5)=3$, which is neither stratum-specific blip, though it is their population-weighted average.