Continuous Treatment Blips And Dose-Response Models

Why Model a Continuous-Treatment Blip?

A treatment need not be simply present or absent. A medication may be prescribed at different daily doses, an intervention may vary in intensity, and exposure may accumulate on a continuous scale. In these settings, a causal question often concerns the effect of changing the dose, not merely comparing treatment with no treatment.

A structural nested mean model represents that causal contrast as a blip: the change in mean outcome attributable to a specified dose, relative to a reference dose, among units with the same pre-treatment history. G-estimation uses treatment variation conditional on that history to estimate the blip parameters, after subtracting the modeled causal contribution from the observed outcome.

Defining the Dose Blip

Let \(A\) be a continuous treatment dose, \(H\) the measured history available before treatment, and \(Y(a)\) the potential outcome that would be observed if dose \(a\) were assigned. For this tutorial, use dose zero as the reference. The conditional mean blip at dose \(a\) is the mean causal contrast between dose \(a\) and dose zero, within history \(H\).

A simple model assumes a common linear response per unit of dose:

Here \(\psi\) is the mean change in outcome for a one-unit increase in dose, under this linear model. A curved relationship can instead be represented, for example, by a quadratic blip:

The blip describes a causal contrast, not the absolute mean outcome at dose \(a\). The latter also includes the mean outcome under the reference dose, which may depend on \(H\). A reference dose of zero is useful only when it has a meaningful interpretation; if zero is not a relevant reference, a different reference dose can be used and the blip defined relative to it.

Identification and the Blip-Down Outcome

The causal interpretation of a fitted blip requires assumptions about how treatment was assigned and recorded. Write \(Y(a)\) for the potential outcome at dose \(a\), and let \(H\) contain the measured pre-treatment variables needed to control confounding.

These conditions clarify the role of the g-estimation equation. If the true parameter is \(\psi_0\), subtract the modeled blip at the received dose from the observed outcome:

This is called the blip-down outcome. Under the assumptions above and a correctly specified blip, its conditional mean does not depend further on the received dose once \(H\) is fixed. In words, removing the modeled causal contribution of the received dose leaves an outcome whose mean is the mean outcome under the reference dose.

The right side can vary with \(H\), so \(U(\psi_0)\) need not have the same mean for every person. The key restriction is that, within a history, its conditional mean is not associated with which dose was received. Treatment-residual moment equations express that restriction without requiring a model for the reference-dose mean.

Deriving the G-Estimation Equation

Let \(q(A)\) be a vector of dose functions used by the blip model. For a linear blip, \(q(A)=A\); for the quadratic model above, \(q(A)=(A,A^2)^\mathsf{T}\). Center each component by its conditional expectation given \(H\). Under the true blip, the centered dose functions have zero conditional mean, while the blip-down outcome has a conditional mean that is constant over doses within \(H\).

The argument is a conditional-expectation calculation. Let \(\mu_0(H)=E[Y(0)\mid H]\). First condition the product on \(A,H\); the centered dose term is then fixed, and the conditional mean of the blip-down outcome is \(\mu_0(H)\). Next condition on \(H\). The centered dose term has conditional mean zero, so the entire expectation is zero:

In data, replace expectations with sample averages and estimate the conditional dose means from the treatment model. For a scalar linear blip, the sample equation is a residualized-dose moment. For a blip with multiple dose-basis terms, use one such moment per parameter; the resulting system must contain enough independent treatment variation to identify those parameters.

The conditional expectations must correspond to the components of \(q\). In particular, for a quadratic blip the centered terms are \(A-\widehat{E}[A\mid H]\) and \(A^2-\widehat{E}[A^2\mid H]\); centering \(A\) alone does not supply the second moment equation.

Curved Dose Responses and Dose Changes

With a quadratic blip, two parameters describe the curve relative to dose zero. The first-order coefficient alone is not the effect of a one-unit dose increase at every dose. A dose change from \(a\) to \(a+\delta\) has modeled conditional mean contrast \(\gamma(a+\delta,H;\psi)-\gamma(a,H;\psi)\). For the quadratic model this is \(\psi_1\delta+\psi_2\{(a+\delta)^2-a^2\}\). When the curve is differentiable, its local slope at dose \(a\) is \(\psi_1+2\psi_2a\), in outcome units per dose unit.

Why Residualizing Treatment Matters

When dose assignment depends on measured history, raw treatment-outcome association can reflect both causal dose response and differences in history. The treatment residual compares a person's received dose with the dose expected from their history. Using that residual in the moment equation removes the part predictable from \(H\), which is the source of confounding adjustment in this g-estimation construction.

Interpretation and Common Pitfalls

A fitted blip must be translated into the dose contrast actually asked about. For a linear blip, \(\psi\) is the modeled change per dose unit, so a change of \(\delta\) units has contrast \(\psi\delta\). For a nonlinear blip, calculate the difference in fitted blips at the two doses; do not report a coefficient as if it were a universal per-unit effect. The contrast is conditional on the model and the stated history structure.

• Confusing the blip with the absolute dose-specific mean. The blip is relative to the reference dose. It does not by itself give \(E[Y(a)\mid H]\).
• Using only a linear dose residual for a multi-parameter curve. A quadratic blip requires moments for both \(A\) and \(A^2\), centered conditional on \(H\).
• Reading a nonlinear coefficient as the dose effect everywhere. In a quadratic model, the local slope depends on dose; compute the relevant finite contrast or derivative.
• Assuming residualization guarantees causal identification. A treatment residual does not repair unmeasured confounding, lack of dose support, inconsistent treatment definitions, or a misspecified blip.
• Extrapolating beyond observed dose support. A fitted curve can produce numerical predictions outside the range with adequate treatment variation, but positivity does not justify their causal interpretation.