Pharmacogenomic Outcomes And Estimands

Why Pharmacogenomic Outcomes Need Explicit Estimands

A genetic association with a drug outcome is not interpretable from the outcome label alone. “Response,” “side effect,” “dose,” and “stopping treatment” can each refer to several different measurements. A response could mean symptom improvement at a fixed visit or improvement at any time; a side effect could mean a particular event or any event; dose could mean the initial dose, the stable dose, or the dose required to reach a target. Changing the definition can change the scientific question and the answer.

An estimand is the precise target quantity an analysis aims to describe or estimate. For pharmacogenomic questions, it should identify the target population, the treatment strategies being contrasted, the genetic subgroup, the outcome definition, the time horizon, and the summary measure. These choices are part of the scientific question, not details to fill in after looking at the data.

For a potential-outcomes description, let \(Y^a\) denote the outcome that would be observed under treatment strategy \(a\). Let \(G=g\) denote membership in a clearly defined genetic subgroup, and let \(S=1\) denote membership in the target population. A mean contrast, for example, can be written \(E(Y^{a_1}-Y^{a_0}\mid S=1,G=g)\). This notation describes a target quantity; it does not by itself establish that the quantity can be identified from a particular data set.

The treatment strategies must be specific enough to distinguish what is being compared. “Drug A versus usual care” is incomplete if usual care varies substantially. One might instead compare Drug A at a specified starting dose with Drug B under a specified titration plan. Similarly, a genetic subgroup must be operationally defined—for example, carriers of at least one prespecified allele at a named locus, using a stated genotype-calling rule. “Genetically susceptible patients” is not a reproducible subgroup definition.

Constructing an Estimand Before Choosing a Model

A practical way to define an estimand is to write its components in words before writing an equation or selecting a statistical model. In particular, name the people to whom the result is intended to apply, not merely the people who happened to be included in the analysis. State the treatment strategies from a common time origin, specify how genotype defines the subgroup, define what is measured, and state when it is measured or over what period it is accumulated.

The subgroup-specific contrast in this tutorial compares treatment strategies within one defined genetic subgroup. It is not, by itself, a comparison of treatment effects between genetic subgroups. Those are distinct questions, and a result from one subgroup alone does not establish that genotype modifies the treatment effect.

Defining Efficacy and Toxicity Outcomes

An efficacy endpoint should state what counts as benefit and how it is assessed. For a binary endpoint, “clinical response” needs a rule, such as meeting a prespecified symptom-reduction threshold at week 12. For a continuous endpoint, specify the instrument, scoring direction, and visit. A fixed-time endpoint and an “ever responded during follow-up” endpoint answer different questions. If a measurement requires survival or continued participation to be observed, the estimand must also say how outcomes after death, withdrawal, or treatment discontinuation are represented.

Toxicity endpoints need the same care. “Any adverse event” can combine events that differ in severity and clinical meaning. An estimand might instead concern the risk of a named adverse reaction by day 30, or the mean change in a laboratory measure at week 8. For a risk, define whether a participant counts as having the outcome if the event occurs at any time through the horizon, and specify whether the event must be judged treatment-related. The definition should be applied consistently across treatment strategies.

A toxicity estimand can also use a fixed-horizon risk, but the event and time window should reflect the clinical question. The same arithmetic may be used for a risk difference, while its direction is interpreted according to whether the event is harmful. The risk difference must be described with the event, subgroup, comparator, and horizon, rather than merely reported as a number.

Making Dose-Requirement Outcomes Well Defined

“Dose requirement” is especially easy to leave ambiguous. It may refer to a dose needed to reach a prespecified therapeutic target under a titration algorithm, but that target, the titration steps, and the assessment schedule must be defined. A stable dose at a visit is not automatically the dose that would be required under another treatment strategy or another titration procedure. Nor is a dose measured only among people who responded an outcome for the full eligible population: response is affected by treatment, so selecting only responders changes who is being described.

A defensible dose outcome specifies a rule that assigns a value for every person in the target population. For example, define the outcome as the protocol-assigned daily dose at week 8 under a fixed titration algorithm; assign the maximum protocol dose if the target has not been reached, and assign the last dose before stopping treatment if discontinuation occurs before week 8. This endpoint measures dose under that specified algorithm, including its rules for nonattainment and stopping. It should not be presented as a universal biological dose requirement.

Discontinuation as an Outcome

Discontinuation is a treatment outcome when the question concerns whether or when patients stop a drug. Define whether the endpoint is stopping for any reason or stopping for a specified reason, such as toxicity or lack of efficacy. Specify the time origin and horizon—for example, all-cause discontinuation by 12 weeks after assignment. State how temporary interruptions, switching, and restarting are classified. If the study records only time to stopping, the estimand should also specify the summary measure used to describe that time.

If death or another event prevents the discontinuation of interest from occurring, the estimand must state how that event is handled. For example, an all-cause discontinuation endpoint can count a composite that includes death, while a question about discontinuation before death needs a different definition. Such decisions affect the meaning of the reported risk; treating people with incomplete follow-up as if they had not discontinued is not a definition of the outcome.

Why the Summary Scale and Assumptions Matter

A risk difference, risk ratio, mean difference, and time-to-event summary are not interchangeable descriptions. They answer different questions and can lead to different impressions, particularly when outcome risks differ across populations. The estimand must name the scale before results are interpreted. For example, a 16-percentage-point remission difference is an absolute contrast; it is not a 16% relative increase.

A brief justification shows how a binary estimand leads to a risk-difference calculation. For a fixed horizon \(\tau\), define \(Y_\tau^a=1\) when the specified event has occurred by \(\tau\) under strategy \(a\), and \(Y_\tau^a=0\) otherwise. Then the risk under that strategy in subgroup \(g\) is \(P(Y_\tau^a=1\mid S=1,G=g)\). Subtracting the two strategy-specific risks gives

In a randomized trial, the observed event proportion in each assigned arm and subgroup estimates the corresponding strategy-specific risk when assignment is randomized within the relevant population and outcomes are ascertained according to the endpoint definition. Interpreting this as the intended causal contrast also depends on the trial’s treatment strategies and follow-up matching the estimand. In observational data, randomization cannot be presumed; whether a causal contrast is identified requires additional assumptions and methods. Defining the estimand is necessary in either setting, but does not itself solve confounding or missing outcome data.

Common Pitfalls

• Using an outcome label without a measurement rule. “Efficacy” is not a complete endpoint. Define the response threshold or measurement, visit, and time horizon.
• Leaving the treatment strategy vague. “Drug versus control” can conceal different doses, titration procedures, or follow-up. Name the strategies being compared.
• Analyzing dose only among responders without saying so. This describes a selected post-treatment group, not dose outcomes for all eligible patients. Define a dose outcome and rules for nonresponse and stopping if the target population is broader.
• Calling a subgroup-specific effect a genotype interaction. A contrast within one genetic subgroup does not show that the contrast differs in another subgroup.
• Interpreting a discontinuation risk as a toxicity risk. All-cause stopping can reflect many reasons. Use a reason-specific endpoint when that is the question.
• Reporting a number without its units or horizon. “The risk difference was −0.14” is incomplete unless the event, treatment order, subgroup, and time period are stated.