Additive SNMMs For Binary Treatment

Why Specify an Additive Blip Model?

A structural nested mean model describes how a treatment changes an outcome on average, conditional on a specified history. For a single treatment occasion with a binary treatment, the central contrast is between the outcome under treatment and the outcome under no treatment. An additive model expresses that contrast as a difference in outcome units, such as points on a test or days of symptoms.

The model is a statement about counterfactual outcomes. It does not, by itself, say how to estimate the contrast from observed data or guarantee that the contrast is identified. Those are separate questions. Here the goal is to specify the one-period model precisely, understand its parameter, and see what it implies for conditional means.

Definition: Let $A\in\{0,1\}$ be treatment at one occasion, let $Y(a)$ be the outcome that would be observed under treatment level $a$, and let $H$ denote measured information available before treatment. An additive structural nested mean model specifies the conditional mean blip $$ \gamma(a,H;\psi)=E\{Y(a)-Y(0)\mid H\}=a\,\psi^{T}f(H), $$ where $f(H)$ is a specified vector of functions of the pre-treatment history and $\psi$ is a parameter vector. The conditional mean blip is the mean difference between treatment level $a$ and the reference level $0$, among units with history $H$.

Because $A$ is binary, this model has only two treatment levels to describe. When $a=0$, the blip is zero by definition of the reference treatment. When $a=1$, the conditional mean difference is $\psi^{T}f(H)$. The simplest version takes $f(H)=1$, giving a single constant parameter $\psi$:

E\{Y(1)-Y(0)\mid H\}=\psi.

This constant-blip version says that the mean treatment contrast is the same across values of $H$. It does not say that every individual’s counterfactual difference is $\psi$. Individual treatment effects may differ; the restriction is on the conditional mean difference. More generally, the specified function $f(H)$ determines how the model allows the conditional mean blip to vary with the measured history.

From the Blip to Conditional Counterfactual Means

The model specifies a contrast, not the level of the outcome under reference treatment. Define the reference conditional mean as $\mu_0(H)=E\{Y(0)\mid H\}$. Adding the reference mean to the blip gives the conditional mean under either treatment level.

\begin{aligned} E\{Y(a)\mid H\} &=E\{Y(0)\mid H\}+E\{Y(a)-Y(0)\mid H\}\\ &=\mu_0(H)+a\,\psi^{T}f(H). \end{aligned}

This short derivation uses only the linearity of conditional expectation and the model definition. In particular, it shows why the additive model is on the outcome’s mean scale: the treatment term is added to the reference conditional mean. The model does not require a particular distribution for $Y(a)$, such as a normal distribution, and it does not specify the variance of the outcome.

The reference mean $\mu_0(H)$ may vary with history even when the blip is constant. For example, people with different baseline severity may have different mean outcomes under no treatment, while sharing the same modeled mean treatment difference. Keeping the baseline outcome level and the treatment blip conceptually separate is essential: the blip is a contrast between counterfactual means, not the mean outcome itself.

Worked Example: Interpreting a Constant Blip

Suppose a hypothetical one-period model uses an outcome measured in symptom-free days over a fixed follow-up period. The additive blip model is $E\{Y(1)-Y(0)\mid H\}=\psi$, with $\psi=2.4$ days.

For any specified pre-treatment history $H=h$, the model says $$ E\{Y(1)-Y(0)\mid H=h\}=2.4\text{ days}. $$ Thus, among units with that history, the mean number of symptom-free days under treatment is 2.4 days greater than the mean under no treatment.

If the hypothetical reference mean for one history is $\mu_0(h)=18.0$ days, then $$ E\{Y(1)\mid H=h\}=18.0+2.4=20.4\text{ days}. $$ The 20.4 days is the treatment conditional mean; the 2.4 days is the modeled mean contrast. Neither number states that every individual gains 2.4 symptom-free days.

Checking a Constant-Blip Specification

For a constant-blip model, the conditional mean difference should be the same at each value of $H$ in the model’s scope. The following hypothetical counterfactual means illustrate this restriction. They are stipulated population quantities for explaining the model, not observed treatment-group averages.

Pre-treatment history	$E\{Y(0)\mid H\}$ (score points)	$E\{Y(1)\mid H\}$ (score points)	Difference (score points)
$H=h_1$	60	64	4
$H=h_2$	70	74	4

Worked Example: Verifying a Constant Mean Blip

Use the hypothetical conditional means in the table to find the treatment blip for each history and assess whether they agree with a constant additive model.

For $H=h_1$, subtract the reference-treatment mean from the treated mean: $$ E\{Y(1)-Y(0)\mid H=h_1\}=64-60=4\text{ score points}. $$ For $H=h_2$, the corresponding calculation is $$ E\{Y(1)-Y(0)\mid H=h_2\}=74-70=4\text{ score points}. $$

The two conditional mean differences are equal, so these stipulated means are consistent with the constant-blip value $\psi=4$ score points for these histories. The reference means differ by 10 points, but that does not conflict with a constant blip: the model restricts the treatment contrast, not the baseline outcome level.

If the two differences had been unequal, they would not both be represented by one constant $\psi$ at those histories. That mismatch would indicate that the constant-blip specification does not describe those conditional means exactly. It would not, on its own, show why the mismatch occurred or determine which alternative model should be used.

Worked One-Period Model in Mean Form

A useful way to check the meaning of the model is to write down both conditional potential-outcome means. Consider a hypothetical outcome measured in blood-pressure points and the simple model $E\{Y(1)-Y(0)\mid H\}=\psi$, with $\psi=-5$ points. Suppose the reference conditional mean is 142 points for one history and 130 points for another.

Worked Example: Writing Both Conditional Means

For the first history, $H=h_1$, suppose $\mu_0(h_1)=142$ points. The reference-treatment mean is $$ E\{Y(0)\mid H=h_1\}=142\text{ points}. $$ The model gives the treatment mean by adding the blip: $$ E\{Y(1)\mid H=h_1\}=142+(-5)=137\text{ points}. $$ The contrast is $137-142=-5$ points, as required.

For the second history, $H=h_2$, suppose $\mu_0(h_2)=130$ points. Then $$ E\{Y(1)\mid H=h_2\}=130+(-5)=125\text{ points}, $$ and the contrast is $125-130=-5$ points. The reference means differ across histories, while the modeled additive treatment contrast remains the same.

The negative sign indicates that the conditional mean outcome under treatment is 5 blood-pressure points lower than under no treatment. This is an interpretation of the stipulated causal model, not evidence that treatment actually causes this difference in a particular population.

What the Model Does—and Does Not—Claim

A structural nested mean model is not merely a regression equation for the observed outcome. Its parameter describes a counterfactual contrast. To interpret an estimated coefficient causally in an application, additional conditions must connect the counterfactual means to the observed data. The model specification alone does not provide those conditions.

For instance, comparing observed mean outcomes among treated and untreated people with the same recorded $H$ need not yield the blip. Treatment groups may differ in other prognostic ways, and their observed outcomes reveal different counterfactuals for different people. Thus, even if a regression of observed $Y$ on $A$ and $H$ produces a treatment coefficient, that coefficient is not automatically the SNMM parameter. Establishing when and how the data identify the causal contrast requires additional assumptions and methods.

The one-period additive model also imposes substantive restrictions. The outcome must have a meaningful additive scale; the modeled contrast is expressed in that scale’s units. The chosen $f(H)$ must encode the intended pattern of conditional mean contrasts. Finally, the model concerns conditional means only: it does not imply equal variances, identical outcome distributions, or identical individual effects across treatment levels.

Key takeaway: In a one-period additive SNMM for binary treatment, the blip is $E\{Y(1)-Y(0)\mid H\}$. A constant model assigns the same conditional mean difference to each history; the conditional means under treatment are obtained by adding that difference to the reference conditional mean. This model statement is distinct from the assumptions needed to identify or estimate its parameter.

Common Mistake

Wrong reasoning: “The observed treated group had an average outcome 4 points higher than the observed untreated group, so the additive SNMM blip is 4 points.”

Correction: That calculation is an observed group difference. The blip is a conditional contrast between counterfactual means, $E\{Y(1)-Y(0)\mid H\}$. The observed group difference does not generally equal this contrast merely because treatment is binary or because an additive regression was fitted. A causal interpretation requires additional identification conditions; the additive SNMM specifies the target contrast and its form, not those conditions.

Check Your Understanding

Use the model definition and examples above to answer these questions.

In a constant additive SNMM with $\psi=3$ kilograms, what does the parameter say about the conditional mean contrast?
If $E\{Y(0)\mid H=h\}=25$ units and the constant blip is $-2$ units, what is $E\{Y(1)\mid H=h\}$?
Two histories have conditional mean differences of 1 and 5 score points. Can one constant-blip model represent both differences exactly?
Does a constant conditional mean blip imply that every individual has the same treatment effect? Explain.
Why is an observed treated-versus-untreated mean difference not automatically the SNMM blip?