Sequential Blip-Down for Counterfactual Outcomes

Why Remove Treatment Effects in Reverse Time?

In a longitudinal study, a final outcome reflects treatment decisions made at several times, often alongside covariates that change during follow-up. A structural nested mean model describes treatment effects through blip functions: functions that represent the contribution of a treatment decision to a later outcome, relative to a reference treatment. Sequential blip-down applies those functions to the observed outcome, working backward from the last treatment decision to the first.

The resulting transformed outcome is useful because it removes the modeled contributions of the observed treatments. But the interpretation needs care. Under a sufficiently strong individual-level structural model, it equals the individual outcome under the reference treatment history. Under a mean model alone, it is generally a pseudo-outcome; the transformation does not by itself reveal any person's unobserved counterfactual outcome.

Definition and Recursion

Consider $K$ treatment decisions, indexed by $t=1,\ldots,K$. Let $A_t$ be treatment at time $t$, and let $H_t$ denote the measured history available immediately before that decision. The history may include baseline covariates, prior treatments, and covariates measured before time $t$. Let $Y$ be the outcome measured after the final treatment decision. Write $\bar A=(A_1,\ldots,A_K)$ for the observed treatment history, and take the reference history to be all-zero treatment, $\bar 0=(0,\ldots,0)$.

Definition: A blip function $\gamma_t(H_t,A_t;\psi)$ specifies the modeled contribution of treatment $A_t$ at time $t$, given the history $H_t$, relative to reference treatment $A_t=0$. It is normalized so that $\gamma_t(H_t,0;\psi)=0$. The parameter $\psi$ indexes the blip model. Given these functions, define the sequentially blipped outcomes by $$ U_{K+1}(\psi)=Y,\qquad U_t(\psi)=U_{t+1}(\psi)-\gamma_t(H_t,A_t;\psi),\quad t=K,\ldots,1. $$ The final transformed outcome $U_1(\psi)$ has had the modeled blips at all observed treatment decisions subtracted.

The order matters for understanding the construction: first remove the last decision's blip, then the preceding decision's blip, and continue backward. For the additive recursion shown here, the final arithmetic is a sum of subtractions. The recursive order makes explicit which treatment decision each blip belongs to; it does not license replacing a blip's history with a different, unobserved history.

What the Recursion Establishes

The first result is an algebraic identity. It requires only that the blip values are defined at the observed histories and that the recursion is carried out as stated. It is not, by itself, a causal identification result.

Result: For any fixed $\psi$ and defined blip functions, $$ U_1(\psi)=Y-\sum_{t=1}^{K}\gamma_t(H_t,A_t;\psi). $$

Derivation. The recursion gives $U_K=Y-\gamma_K(H_K,A_K;\psi)$. Substituting this into the equation for $U_{K-1}$ gives $U_{K-1}=Y-\gamma_K(H_K,A_K;\psi)-\gamma_{K-1}(H_{K-1},A_{K-1};\psi)$. Continuing the substitution one time at a time yields the stated sum. Equivalently, summing the recursion equations gives

\sum_{t=1}^{K}\{U_t(\psi)-U_{t+1}(\psi)\} =-\sum_{t=1}^{K}\gamma_t(H_t,A_t;\psi).

The left side telescopes to $U_1(\psi)-U_{K+1}(\psi)=U_1(\psi)-Y$, which proves the identity. $\square$

To call this transformed outcome the individual counterfactual $Y^{\bar 0}$ requires more than the algebra. One sufficient, strong assumption is an individual-level additive structural model in which the observed outcome decomposes into the reference outcome plus the treatment blips evaluated at the observed histories.

Result: Suppose, for each individual, the structural equation $$ Y=Y^{\bar 0}+\sum_{t=1}^{K}\gamma_t(H_t,A_t;\psi_0) $$ holds, with defined blips and a true parameter $\psi_0$. Then $U_1(\psi_0)=Y^{\bar 0}$.

Proof. Apply the algebraic identity at $\psi_0$:

U_1(\psi_0) =Y-\sum_{t=1}^{K}\gamma_t(H_t,A_t;\psi_0) =Y^{\bar 0}+\sum_{t=1}^{K}\gamma_t(H_t,A_t;\psi_0) -\sum_{t=1}^{K}\gamma_t(H_t,A_t;\psi_0) =Y^{\bar 0}.

This is an individual-level equality because the assumed structural equation is an individual-level equality. A structural nested mean model, which specifies mean treatment contrasts rather than this individual decomposition, does not generally imply it. Identification of blip parameters also needs causal assumptions; that is a separate question from performing the transformation. $\square$

Worked Example: Two Binary Treatment Decisions

Suppose $K=2$, the observed outcome is $Y=18$, and the observed history has $A_1=1$, $A_2=1$, and $L_1=2$. Consider blips $\gamma_1(H_1,A_1;\psi)=2A_1$ and $\gamma_2(H_2,A_2;\psi)=A_2(3+L_1)$. Both are zero when their treatment argument is zero. Compute the sequentially blipped outcome.

Start at the end of follow-up: $U_3=Y=18$. At time 2, the blip is $1(3+2)=5$, so $U_2=18-5=13$. At time 1, the blip is $2(1)=2$, so $U_1=13-2=11$. The sum identity checks the arithmetic: $18-5-2=11$.

If the individual-level structural equation in the result holds, this value is $Y^{(0,0)}=11$ for this individual. Without that assumption, the justified conclusion is only that the specified blips transform the observed outcome to 11; the calculation alone does not establish the unobserved individual counterfactual.

Reference Treatments and the Meaning of a Blip

The zero-treatment reference is a convenient normalization, not a requirement that treatment be naturally coded zero. For a different reference value $a_t^*$, define the blip relative to that value as $\gamma_t^*(H_t,A_t;\psi)=\gamma_t(H_t,A_t;\psi)-\gamma_t(H_t,a_t^*;\psi)$. Then $\gamma_t^*(H_t,a_t^*;\psi)=0$, and the corresponding recursion subtracts $\gamma_t^*(H_t,A_t;\psi)$ at each time. This formula assumes the structural decomposition is parameterized using these contrasts at the histories used in the recursion.

Worked Example: Blipping Down to a Nonzero Reference

At one treatment decision, suppose the original blip is $\gamma(a)=1.5a$, the observed treatment is $a=2$, the reference treatment is $a^*=1$, and the observed outcome is $Y=10$. Calculate the outcome transformed relative to treatment 1.

The observed-treatment blip is $\gamma(2)=1.5(2)=3$, while the reference-treatment blip is $\gamma(1)=1.5(1)=1.5$. The contrast to subtract is $3-1.5=1.5$, not 3. Thus the reference-relative transformed value is $10-1.5=8.5$. If the individual structural decomposition is $Y=Y^{(1)}+\{\gamma(2)-\gamma(1)\}$, then $Y^{(1)}=8.5$. Subtracting the full observed-treatment blip would instead give 7, which corresponds to removing treatment relative to zero, not relative to one.

Errors, Diagnostics, and Failure Modes

In practice, blips are usually evaluated using an estimated parameter $\hat\psi$. The recursion still produces a number, but a numerical result is not evidence that the blip model is correct. The next result makes one limitation especially clear: errors in separate blips accumulate in the final transformed outcome.

Result: Let $\hat\gamma_t$ be fitted blips, and define the pointwise errors $e_t=\hat\gamma_t(H_t,A_t)-\gamma_t(H_t,A_t)$. Then the fitted transformed outcome satisfies $$ \hat U_1-U_1=-\sum_{t=1}^{K}e_t, \qquad |\hat U_1-U_1|\leq\sum_{t=1}^{K}|e_t|. $$

Proof. Apply the sum identity first to the fitted blips and then to the true blips:

\hat U_1-U_1 =\left(Y-\sum_{t=1}^{K}\hat\gamma_t\right) -\left(Y-\sum_{t=1}^{K}\gamma_t\right) =-\sum_{t=1}^{K}(\hat\gamma_t-\gamma_t) =-\sum_{t=1}^{K}e_t.

Worked Example: Accumulation of Blip-Model Error

Suppose there are three treatment decisions. At the observed histories, the true blips are $2, -1,$ and $4$, while the fitted blips are $2.5, -0.5,$ and $3$. Let $Y=20$. Calculate both transformed outcomes and verify the error bound.

Using the true blips gives $U_1=20-2-(-1)-4=15$. Using the fitted blips gives $\hat U_1=20-2.5-(-0.5)-3=15$. The net errors are $e_1=0.5$, $e_2=0.5$, and $e_3=-1$, so $\hat U_1-U_1=0$, agreeing with $-\sum_t e_t=-(0.5+0.5-1)=0$. The bound is also satisfied: $0\leq |0.5|+|0.5|+|-1|=2$. The exact cancellation in this example is possible, but it should not be expected in general.

Useful checks include verifying that every reference-treatment blip is zero, evaluating the recursion in reverse time, and independently checking the final sum. Inspect blip predictions across histories and treatment values: implausible magnitudes, abrupt extrapolations, or a failure to reflect known effect modification may signal model problems. If fitted blips are near zero for all histories despite substantial observed treatment variation, that is a prompt to examine the model and its assumptions, not proof that treatment has no effect.

Using the wrong history. A blip is evaluated at the history specified by its model. Replacing an observed covariate history with a reference-regime covariate history requires additional counterfactual quantities and is not part of this arithmetic recursion.
Confusing a pseudo-outcome with an observed outcome. $U_1$ is calculated from the observed outcome and modeled blips. It is not directly observed, and mean-model assumptions alone do not make it an individual's true $Y^{\bar 0}$.
Ignoring accumulated error. Small errors at several times may add in the same direction. A good fit at only the final decision does not guarantee an accurate fully blipped outcome.
Overlooking support and model limitations. A fitted blip may be extrapolated to histories or treatment levels with little data. The recursion cannot repair unsupported comparisons, unmeasured confounding, or a misspecified blip model.

How Blip-Down Fits Among g-Methods

Sequential blip-down and the parametric g-formula address longitudinal interventions from different directions. The g-formula models outcome and covariate transitions under treatment strategies and uses those models to calculate regime-specific outcomes. Blip-down instead transforms an outcome using specified treatment-effect functions. In marginal structural models, inverse-probability weighting is another approach to estimating causal effects under its own modeling and identification conditions. These approaches are not interchangeable computations, and a blip-down transformation alone does not supply the identification assumptions required for a causal conclusion.

The recursion is especially useful for making a proposed structural nested model concrete: it shows exactly what outcome would remain after removing the modeled treatment contributions. Its limitations are equally important. Incorrect blip functions can produce misleading transformed outcomes; individual counterfactual recovery requires a stronger model than mean treatment-effect specification; and causal interpretation depends on assumptions about treatment assignment and measured histories. The transformation is a construction, not a diagnostic test that proves those assumptions.

Check Your Understanding

Work through these questions, then compare with the solutions.

With $Y=14$, two observed treatment blips of 3 and 2, what is $U_1$?
Why does the recursion alone not prove that $U_1=Y^{\bar 0}$?
If a blip is $\gamma(a)=4a$, the observed treatment is 3, and the reference is 1, what contrast is subtracted?
If two fitted blip errors are $0.4$ and $-0.7$, what is $\hat U_1-U_1$?

Solutions. (1) $U_1=14-3-2=9$. (2) The recursion is an algebraic transformation; individual counterfactual equality additionally requires an individual-level structural decomposition (or other sufficient assumptions). A mean model alone does not assert that equality for each person. (3) The contrast is $\gamma(3)-\gamma(1)=12-4=8$. (4) The difference is $-\{0.4+(-0.7)\}=0.3$, by the blip-error identity.