Blip-Down Transformations And Treatment-Free Outcomes

Why Transform an Outcome?

With treatment at multiple times, an observed outcome reflects the history of treatments actually received. A structural nested mean model represents treatment effects through a sequence of blips: causal increments associated with treatment at each time, after later treatment effects have been removed. Blip-down transformations use those increments to construct an outcome with the modeled treatment effects subtracted.

The resulting quantity is useful because it aims to remove the contribution of treatment from the outcome while retaining its other variation. But a transformed value is not automatically the outcome that the same person would have had under no treatment. It is an algebraically constructed pseudo-outcome; its causal interpretation depends on the model and identification assumptions.

The histories \(\bar A_k=(A_1,\ldots,A_k)\) and \(\bar L_k=(L_1,\ldots,L_k)\) include treatment and covariate information through time \(k\). Thus, the blip at time \(k\) can depend on the treatment and covariate history up to that time. The blip is a conditional mean contrast, not necessarily a fixed effect shared by everyone or an individual-level treatment effect.

The Backward Transformation

Suppose there are \(K\) treatment times. Start with the observed outcome, then remove the modeled blip at the latest time. From that result, remove the blip at the preceding time, and continue backward. Write \(H_{K+1}(\psi)=Y\), and define the successive transformed outcomes as follows:

The order matters conceptually. The blip at time \(k\) describes the effect of that treatment on an outcome from which the modeled effects of later treatments have already been removed. It is therefore applied to the outcome produced at the preceding backward step, not used as a substitute for the entire transformation.

Repeated substitution gives a compact expression for the final transformation:

This equation makes the arithmetic transparent: each modeled blip is subtracted once. It does not, by itself, establish a causal interpretation. That interpretation requires the functions to describe the relevant causal mean increments and requires conditions that let those increments be identified from observed data.

Why Subtracting the Blips Can Remove Treatment Effects

The central idea is that a correctly specified blip accounts for the conditional mean change associated with the current treatment, given the history, on an outcome already adjusted for later treatment effects. Subtracting that increment should remove the corresponding treatment-related difference in conditional means.

More formally, let \(H_{k+1}(\psi)\) be the outcome after later blips have been removed. At the true model value \(\psi_0\), the structural mean model represents the conditional mean contrast between setting \(A_k=a\) and setting \(A_k=0\), for the same prior history, by the blip function. In schematic notation, this contrast is:

When the observed blip is subtracted, the conditional mean contrast between treatment \(a\) and reference treatment \(0\) is removed:

This is a conditional-mean statement, not a claim that the transformed values are identical across people or treatment groups. Individuals can differ in their covariates and in outcome variation not explained by the blip. Applying the argument backward at each treatment time motivates the final transformation toward the reference-treatment outcome.

The assumptions have distinct roles. Consistency links an observed outcome to the potential outcome under the treatment history actually received. Sequential exchangeability requires that, conditional on the measured history available before each treatment, treatment assignment carries no further information about the relevant potential outcomes. Positivity requires that each treatment value being compared has positive probability in histories for which that contrast is to be interpreted. Correct specification of the blip functions is also needed for the stated transformation to remove the modeled contrasts. If these conditions fail, the algebra still produces numbers, but the intended causal interpretation may not hold.

Worked Transformations and Their Interpretation

What the Transformation Does—and Does Not—Do

Blip-down transformations matter because they turn a model for sequential treatment effects into an outcome whose mean is intended to correspond to a reference treatment history. This offers a way to express causal assumptions in terms of how outcome means change with treatment at each time. It also makes explicit why later treatment effects must be addressed before earlier ones.

The transformation does not erase all differences among people. It subtracts the modeled conditional mean increments; residual variation, covariate-related variation, and model error remain. Nor does it automatically solve confounding: the blip values have a causal interpretation only under suitable identification conditions. In particular, if important time-varying confounders are not measured, subtracting a fitted blip does not repair that omission.

A frequent mistake is to treat a transformed value as the individual’s actual no-treatment outcome. The calculation is tempting: it starts from that person’s observed outcome and subtracts effects, so the result can look like a recovered counterfactual. But a mean blip is a conditional mean contrast, not generally that person’s realized individual effect. The correction is to call the result a pseudo-outcome and distinguish its value for one record from the population mean it may represent under the structural model and identification assumptions.