Why Blip Interpretation Requires Care
A blip parameter can be easy to misread. It may describe a treatment contrast for a particular visit and a particular history, yet be reported as if it were the effect for everyone. Or an analyst may subtract a treatment effect from the outcome at the wrong visit, leaving treatment effects in the quantity that was meant to remove them. A third error is to estimate a history-specific blip where the data contain no treatment variation within that history.
These are not merely issues of wording. They change the causal quantity being discussed or the information the estimator can use. This tutorial uses \(A_t\) for treatment at visit \(t\), \(L_t\) for covariates measured before that visit’s treatment, and \(H_t\) for the observed history just before \(A_t\), including relevant prior treatments and covariates. The outcome is \(Y\).
For clarity, consider a binary treatment with reference value 0. Write \(\gamma_t(1,H_t)\) for the conditional blip comparing treatment 1 with treatment 0 at visit \(t\), with the model’s future-treatment convention held fixed. Its units are the units of \(Y\). The exact interpretation depends on the structural nested model’s definition of the counterfactual contrast; the visit and history are essential parts of that definition.
Assumptions Behind the Interpretation
A blip model specifies contrasts, but causal interpretation and g-estimation also require assumptions about the treatment process and the model. The following conditions describe the basic setting used here; more specialized designs may require additional conditions.
- Consistency: for a person whose observed treatment history equals a specified history, the observed outcome agrees with the corresponding potential outcome.
- Sequential exchangeability: conditional on the measured history \(H_t\), treatment \(A_t\) is independent of the relevant future potential outcomes. In particular, there is no unmeasured confounding of treatment and those outcomes after conditioning on \(H_t\).
- Positivity: within histories for which a contrast is to be identified, each treatment level being compared has positive probability. For binary treatment, \(0<P(A_t=1\mid H_t=h)<1\) at the histories \(h\) of interest.
- Correct blip specification: the chosen blip function adequately represents the conditional treatment contrasts, including the needed history dependence and visit-specific structure.
Sequential exchangeability does not mean that treatment is unconfounded without adjustment. Rather, it says that the measured history is sufficient for the relevant treatment decision at each visit. Positivity is distinct: even with adequate confounder measurement, an effect cannot be learned from a history at which one treatment option never occurs. Correct specification is also distinct from both assumptions; a well-fitting treatment model cannot repair a blip function that assigns the wrong effect to a visit.
Error 1: Calling a Conditional Blip a Population-Average Effect
A blip may vary with \(H_t\). When it does, the value at one history is not generally the average effect for a target population. For a binary treatment and an additive outcome contrast, a population-average contrast can be obtained by averaging the conditional blip over the target distribution of histories, if the relevant causal contrast and identification assumptions support that averaging.
The target distribution matters. Averaging over the histories of people in a particular study, over histories in an eligible population, or over histories under a specified treatment regime can yield different summaries. The conditional function and the averaging distribution should therefore be reported separately.
Worked Example: Averaging a History-Dependent Blip
Suppose a correctly specified model for a single visit gives the conditional mean difference in a hypothetical outcome score, in points, as \(\gamma(1,L)=2+4L\), where \(L\) is a binary pre-treatment characteristic. In a hypothetical target population, 25% have \(L=1\) and 75% have \(L=0\). Assume the treatment contrast is identified at both histories and the model’s additive contrast applies to this target population.
For people with \(L=0\), the modeled contrast is \(2+4(0)=2\) points. For people with \(L=1\), it is \(2+4(1)=6\) points. The population-average contrast is therefore
Thus, 6 points is the conditional contrast for the \(L=1\) group, not the target population average. The average is 3.00 points because the two history-specific contrasts are weighted by their prevalence in the stated target population. If that target population had a different distribution of \(L\), the average could differ even though the conditional blip model stayed the same.
This averaging is not a license to interpret any fitted coefficient as a marginal effect. For example, if the blip includes treatment-by-history interaction, the coefficient on treatment alone is only one part of the conditional contrast. The contrast for each history must first be computed from the full blip function and then averaged over the target history distribution.
Error 2: Blipping Down Treatment at the Wrong Visit
In a longitudinal model, each treatment visit has its own place in the outcome decomposition. Blipping down means subtracting the modeled contribution associated with the treatment at the relevant visit, in the order required by the structural nested model. It does not mean subtracting whichever estimated treatment effect is easiest to access, nor does it mean applying one visit’s effect to another visit’s treatment.
For a simple additive two-visit illustration, suppose the conditional mean outcome follows the hypothetical structural model \(Y=b+2A_1+5A_2\), with no effect modification or other treatment terms. The first-visit blip is 2 outcome units per treatment increment, and the second-visit blip is 5 units. Here \(b\) can vary across histories, but the treatment contributions shown are fixed. In this simple model, removing both contributions gives \(U=Y-2A_1-5A_2\).
Worked Example: Keeping Two Visit Effects in Their Places
Suppose the hypothetical outcome under a fixed covariate history has \(b=10\), with \(Y=10+2A_1+5A_2\). Four possible treatment histories give the following conditional mean outcomes:
| \(A_1\) | \(A_2\) | Conditional mean \(Y\) | Correctly blipped-down \(U\) |
|---|---|---|---|
| 0 | 0 | 10 | \(10-2(0)-5(0)=10\) |
| 1 | 0 | 12 | \(12-2(1)-5(0)=10\) |
| 0 | 1 | 15 | \(15-2(0)-5(1)=10\) |
| 1 | 1 | 17 | \(17-2(1)-5(1)=10\) |
The arithmetic shows why the indices matter: the 2-unit contribution is removed for \(A_1\), and the 5-unit contribution is removed for \(A_2\). If an analyst swaps those effects, the history \((A_1,A_2)=(1,0)\) gives \(12-5(1)-2(0)=7\), while \((0,1)\) gives \(15-5(0)-2(1)=13\). The resulting values retain an artificial dependence on which visit was treated. The swapped calculation has not removed the treatment contributions specified by the model.
In a general structural nested model, blip-down quantities and moment equations are defined to respect the time ordering and conditioning history. A useful schematic form at visit \(t\) is a conditional moment involving the treatment residual \(A_t-E(A_t\mid H_t)\) and the appropriately blipped-down outcome for that visit. Under the assumptions above and the correctly specified model, the residual is orthogonal to that blipped-down outcome conditional on \(H_t\). The exact quantity to blip down is model-specific; the simple two-visit arithmetic is not a substitute for checking that definition.
Here \(U_t(\psi)\) denotes the outcome quantity defined by the structural nested model after removing the relevant treatment blips, and \(\psi\) denotes its blip parameters. The equation emphasizes two checks: use the treatment residual for the same visit \(t\), and use the outcome quantity defined for that visit’s estimation step. Treating \(U_t\) as a generic outcome that can be formed by subtracting effects in any order is a common source of error.
Error 3: Ignoring Treatment-History Support
A fitted model can return a numerical estimate even when the observed data offer little or no direct information about a particular history-specific contrast. G-estimation uses treatment variation conditional on history. If treatment is constant within a history, the treatment residual is zero there, and that history contributes no within-history contrast to the moment equation.
Worked Example: No Second-Visit Contrast in One History
Consider a hypothetical study with 100 people who share the same measured history \(H_2=h\) before the second treatment. All 100 received \(A_2=1\); none received \(A_2=0\). The conditional treatment probability in this history is therefore \(P(A_2=1\mid H_2=h)=1\), so positivity fails for the comparison of second-visit treatment 1 with 0 at \(h\).
The treatment residual for each person is \(A_2-E(A_2\mid H_2=h)=1-1=0\). Consequently, for any proposed blipped-down outcome values \(U_2(\psi)\), every contribution to the history-specific moment is
Summing or averaging those contributions cannot distinguish a zero blip from a nonzero blip. The observed data at this history provide no treatment contrast for estimating its second-visit blip. A parametric model that forces the blip at \(h\) to equal the blip at other histories might produce a value there, but that value comes from the model’s restriction and information elsewhere, not from treatment variation among these 100 people.
This example is an extreme case. With very few observations in one treatment arm, the moment may not be exactly zero, but the estimated contrast can still be unstable and heavily dependent on modeling assumptions. Positivity diagnostics should therefore be examined at the level of relevant treatment histories, not only through an overall treatment proportion.
Support is also tied to the estimand. A blip for histories that occur in the target population but lack treatment variation may not be identified without additional restrictions. One defensible response may be to narrow the target population to histories with adequate support; another may be to state the structural extrapolation being imposed. Neither response makes the unsupported contrast nonparametrically learned from those histories.
A Practical Check Before Interpreting an Estimate
The following sequence helps distinguish a correctly interpreted, supported blip from a misleading numerical result.
State the visit, treatment comparison, outcome scale, and history on which the blip is conditioned. Include the reference treatment and any future-treatment convention used by the model.
Report a history-specific contrast as conditional. If a population summary is needed, calculate it by averaging the conditional contrasts over a stated target distribution.
Verify that each treatment contribution is removed from the outcome according to the blip function for that same visit and according to the model’s sequential construction.
Assess measured confounding, treatment variation within histories, and the adequacy of the chosen blip form. Do not interpret numerical convergence as evidence that these conditions hold.
Common Mistakes
- Wrong: “The coefficient is the average treatment effect because it is one number.” Correction: A coefficient may describe a conditional component of a blip model. Calculate the relevant contrast across histories and average over the stated target distribution before calling it a population average.
- Wrong: “The latest treatment effect can be subtracted from either visit.” Correction: Blips are visit-indexed. Use the treatment contribution specified for each visit and follow the structural model’s ordering.
- Wrong: “A converged g-estimation algorithm identifies the effect even if treatment never varies in a history.” Correction: Numerical convergence does not create positivity. Without within-history variation, the treatment-residual moment at that history contains no direct contrast information.
- Wrong: “A model that predicts outcomes well must identify the causal blip.” Correction: Prediction fit does not establish sequential exchangeability, positivity, or correct specification of the causal contrast.
Check Your Understanding
For each question, identify what must be checked before making the causal interpretation.
- A blip is 4 points for one covariate history and 9 points for another. What additional information is needed to report a population-average contrast?
- In a two-visit additive model, the first-visit blip is 3 points and the second-visit blip is 7 points. Which contribution is removed for each treatment when blipping down the outcome?
- Within a particular history, every person receives treatment. What does the treatment residual equal there, and what does that imply for estimating the history-specific blip?
- Why does a numerical estimate from a model that shares a blip across histories not prove that the effect is identified at an unsupported history?
- Name two assumptions, other than correct blip specification, needed for the causal interpretation of the g-estimated blip.