Estimator Alignment for Treatment-Policy Questions

Why Alignment Is More Than an Analysis Label

A trial can describe its primary analysis as “intention-to-treat” and still leave an important question unanswered: does the calculation estimate the treatment-policy question? The answer depends on what treatment groups are compared, whether the relevant outcome is collected after intercurrent events, and how the observed outcomes are combined. An analysis label alone does not establish alignment.

A treatment-policy question asks about outcomes under assignment to each treatment strategy, allowing events such as treatment discontinuation, rescue medication, or treatment switching to occur as they do under that strategy. The outcome after an event remains part of the question if it is defined and collected. This is different from asking what outcomes would have been if the event had not occurred.

Definition: Let $A$ denote randomized assignment, with $A=1$ for the experimental strategy and $A=0$ for the control strategy. Let $Y^a$ be the endpoint outcome that would be observed under assignment to strategy $a$, including the consequences of intercurrent events occurring under that strategy. For a population mean-difference target, the treatment-policy estimand is $$ \Delta_{\mathrm{TP}} = E[Y^1]-E[Y^0]. $$ An estimator aligns with this target when its data, analysis population, and calculated contrast target these two assignment-strategy means and their specified difference.

This definition requires the endpoint to be meaningful under both strategies, including after relevant events. It also requires a clearly specified population, endpoint, time point, and contrast scale. The examples below use a difference in mean endpoint scores, with lower scores indicating better health. The same alignment questions apply to other population summaries, but the calculation must then match the chosen scale.

A useful first distinction is between an intercurrent event and missing outcome data. Discontinuation is an event that may occur while the endpoint remains defined. Failure to observe the endpoint after discontinuation is missingness. A treatment-policy question includes the post-event outcome; it does not, by itself, make an unobserved outcome available or specify how to estimate it.

A Three-Part Alignment Check

The following checks separate the scientific target from the operational choices made in an analysis.

Check treatment assignment.
Does the comparison use the randomized groups as assigned? Reassigning participants by treatment received, adherence, or event status changes the comparison and may select groups based on post-randomization information.

Check post-event outcome collection.
Is the endpoint measured at the target time even when an intercurrent event has occurred? If follow-up stops at the event, the treatment-policy outcome is not directly observed for those participants.

Check the calculated contrast.
Does the analysis combine outcomes into the target population summaries and compare them on the specified scale? Excluding post-event observations or using an event-defined subset can yield a different estimand.

When every randomized participant has a defined and observed endpoint, a direct estimator for the mean-difference target is the difference between the sample means in the randomized groups:

\widehat{\Delta}_{\mathrm{TP}}=\bar{Y}_{A=1}-\bar{Y}_{A=0}.

This calculation includes post-event outcomes and retains each participant in the group to which they were assigned. Its interpretation is a comparison of assignment strategies, not a comparison of people who did and did not experience an event.

Why the Assignment-Group Contrast Works

The basic justification is randomized assignment with complete endpoint observation. Suppose $N$ participants are randomized, and $Y_i^1$ and $Y_i^0$ are the potential endpoint outcomes for participant $i$ under the two strategies. Under consistency, the observed outcome for a participant assigned to $a$ is $Y_i^a$. Under complete randomization with fixed group sizes, assignment is independent of these potential outcomes, and every participant has the same probability of assignment to each arm. If outcomes are observed for everyone, the expected mean in each randomized group, over the randomization, equals the corresponding finite-population mean.

E[\bar{Y}_{A=1}\mid \{Y_i^1,Y_i^0\}_{i=1}^N] =\frac{1}{N}\sum_{i=1}^{N}Y_i^1, \qquad E[\bar{Y}_{A=0}\mid \{Y_i^1,Y_i^0\}_{i=1}^N] =\frac{1}{N}\sum_{i=1}^{N}Y_i^0.

Subtracting the second equality from the first gives the finite-population average treatment-policy contrast. This argument relies on randomized assignment, consistency between assigned strategy and observed outcome, and an outcome defined and observed at the target time. It also presumes that one participant’s assignment does not alter another participant’s outcome. In a trial with missing endpoints, the observed group means need not satisfy this result: additional assumptions and a suitable missing-data analysis are needed.

Worked Example: Including Outcomes After Discontinuation

In a hypothetical trial, six participants are assigned to each strategy. All have their endpoint measured at the target visit, including participants who discontinued assigned treatment. Lower scores are better.

In the experimental group, four participants who continued treatment have a mean score of 4 points, and two who discontinued have a mean score of 10 points. In the control group, five who continued have a mean of 7 points, and one who discontinued has a mean of 13 points. The event-specific summaries are used only to check the arithmetic; the target contrast is by randomized assignment.

The experimental assignment-group mean is $$ \bar{Y}_{A=1}=\frac{4(4)+2(10)}{6} =\frac{36}{6}=6.00\text{ points}. $$ The control assignment-group mean is $$ \bar{Y}_{A=0}=\frac{5(7)+1(13)}{6} =\frac{48}{6}=8.00\text{ points}. $$ Therefore, $$ \widehat{\Delta}_{\mathrm{TP}}=6.00-8.00=-2.00\text{ points}. $$

With complete endpoint collection, assignment-group analysis, and a mean-difference target, the calculation aligns with the treatment-policy question. In this hypothetical sample, the experimental assignment strategy has a mean endpoint score 2.00 points lower than the control strategy. This is a sample estimate; it does not by itself quantify statistical uncertainty.

Weights and the Contrast Actually Calculated

A treatment-policy mean uses each arm’s own randomized-group composition. In particular, the frequency of an event may differ between arms, and that difference is part of the outcomes under the assigned strategies. Pooling event categories with the wrong weights, or comparing only a selected event-defined group, can change the target.

Worked Example: Reconstructing the Arm Means

Consider a hypothetical trial with all endpoint outcomes observed. In the experimental group, 10 participants continued treatment and had a mean score of 10 points; 20 discontinued and had a mean of 20 points. In the control group, 10 continued and had a mean of 12 points; 60 discontinued and had a mean of 22 points. The randomized groups therefore contain 30 and 70 participants, respectively.

Weight each event-category mean by its count within its own randomized arm. The experimental mean is $$ \bar{Y}_{A=1} =\frac{10(10)+20(20)}{30} =\frac{100+400}{30} =\frac{500}{30} =16.67\text{ points}. $$ The control mean is $$ \bar{Y}_{A=0} =\frac{10(12)+60(22)}{70} =\frac{120+1{,}320}{70} =\frac{1{,}440}{70} =20.57\text{ points}. $$ Thus the assignment-group contrast is $$ \widehat{\Delta}_{\mathrm{TP}}=16.67-20.57=-3.905\text{ points}, $$ using unrounded arm means for the subtraction.

The estimate is a mean difference between the two assigned strategies in this hypothetical sample. The weighting preserves each arm’s own event distribution. Replacing those weights with a common event distribution would answer a different standardization question; it would not be the unadjusted treatment-policy contrast calculated here.

What Changes When Outcomes Are Missing or Excluded?

If outcome collection stops at discontinuation, the analysis cannot simply treat that missing endpoint as if it were the event itself. The treatment-policy target still concerns the endpoint after discontinuation, if that endpoint remains defined. A complete-case mean uses only participants with observed endpoints; unless the missingness mechanism and analysis assumptions justify it, that mean need not estimate the randomized arm’s target mean. The earlier material on missing data and missingness after treatment discontinuation develops the relevant missing-data reasoning. For alignment, the immediate diagnostic is to identify which outcomes are absent and what assumptions connect the observed data to the target.

Likewise, an analysis that excludes post-event observations may use randomized labels but still fail the outcome-collection or contrast check. For example, an analysis limited to participants who continued treatment does not automatically estimate an effect in a common, pre-treatment-defined population. Continuation happens after randomization and may depend on prognosis and treatment response.

Worked Example: The Effect of Excluding Post-Event Participants

In another hypothetical trial, 20 participants are assigned to each arm, and every endpoint is collected. In the experimental arm, 12 participants continued treatment and had a mean of 5 points; 8 discontinued and had a mean of 11 points. In the control arm, 18 continued and had a mean of 8 points; 2 discontinued and had a mean of 10 points.

The treatment-policy arm means include both event categories: $$ \bar{Y}_{A=1}=\frac{12(5)+8(11)}{20} =\frac{60+88}{20}=7.40\text{ points}, $$ $$ \bar{Y}_{A=0}=\frac{18(8)+2(10)}{20} =\frac{144+20}{20}=8.20\text{ points}. $$ The assignment-group contrast is $7.40-8.20=-0.80$ points.

If the analyst instead retains only participants who continued, the contrast is $5-8=-3.00$ points. That is an arithmetic comparison among observed continuers, not the treatment-policy contrast. Since continuation is post-randomization and its frequency differs by arm, this restricted comparison can also compare selected groups that differ in prognosis. The two estimates differ because they use different populations and outcomes, not because one is a more precise calculation of the other.

Diagnostics, Alternatives, and Limits

A practical alignment review should make the analysis auditable. Confirm the estimand’s population, endpoint, time point, event strategy, summary measure, and contrast. Then check the randomized-arm counts against the analysis dataset; tabulate event status and endpoint availability by arm; verify whether post-event values contribute; and reproduce the final contrast from the arm-level summaries. Large differences between randomized and analyzed counts, or a sharp drop in follow-up after an event, signal that the direct complete-outcome argument may not apply.

Compare the analysis with the main alternatives by asking what happens to the event and to outcomes after it. A while-on-treatment analysis stops using outcomes after treatment ends. A hypothetical approach targets outcomes under a specified scenario in which the event did not occur. A composite strategy incorporates the event into the outcome definition. Principal-stratum approaches target groups defined by potential event behavior. Each addresses a different question; none becomes a treatment-policy analysis merely by using randomized group labels.

A treatment-policy target does not eliminate all limitations. If the endpoint is undefined after an event, as can occur with death for some measurements, an outcome definition or different estimand is needed rather than a routine missing-data calculation. If post-event outcomes are not collected, identification depends on assumptions that cannot be verified from the observed outcomes alone. Sampling uncertainty also remains even when the estimator aligns perfectly. Alignment is therefore necessary for a relevant answer, but it does not guarantee unbiased inference under missingness, correct implementation, or adequate precision.

Common pitfall: “We analyzed participants by randomized group, so this is necessarily a treatment-policy estimate.” That reasoning checks only assignment. It overlooks whether the endpoint was collected after the event and whether the reported contrast actually includes those outcomes with the correct target-population weights. Check all three components before making the claim.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A randomized trial measures the endpoint only until treatment discontinuation. Does assignment-group analysis alone establish alignment with a treatment-policy target? Why or why not?
In the weighted-mean example, what is the experimental-minus-control treatment-policy contrast to three decimal places?
Why can an analysis restricted to participants who continued assigned treatment differ from the treatment-policy contrast, even if all endpoint values are observed?
Name the three components of the alignment check.
What additional issue arises if the endpoint is not defined after an intercurrent event?

Solutions. (1) No. The analysis uses randomized assignment, but it lacks endpoints after discontinuation. Estimating the target then requires assumptions and an appropriate method for missing outcomes; assignment labels alone are insufficient. (2) The experimental mean is $500/30=16.6667$ points and the control mean is $1{,}440/70=20.5714$ points, so the contrast is $16.6667-20.5714=-3.9048$, or $-3.905$ points. (3) Continuation is post-randomization, so restriction changes the analyzed population and can select different kinds of participants in the two arms. (4) Check treatment assignment, post-event outcome collection, and the contrast actually calculated. (5) The outcome needs a definition or the target must be reconsidered; an undefined value cannot be made available simply by treating it as ordinary missingness.