From research question to estimand
Start with the decision the answer should inform
A topic such as readmission is not yet a question. Decide whether the aim describes frequency, measures an association, predicts an individual's outcome or compares outcomes under interventions. An adjusted association between a biomarker and readmission describes a relationship in the specified population; it does not by itself estimate what would happen if an intervention changed that biomarker. These purposes can use similar variables while requiring different designs and interpretations.
When to use it
Do this for each primary and secondary aim before finalizing data collection or analysis. A structured clinical question is a useful start, but it can still leave the population, effect measure or treatment changes ambiguous.
When not to force one template
The ICH estimand framework addresses treatment effects in clinical trials. Its discipline is useful more broadly, but descriptive, diagnostic and prediction aims need quantities suited to their purposes; they should not acquire a treatment comparison merely to fill a template.
Write down the target
- Population: whose outcomes should the answer describe, including eligibility and setting?
- Conditions: which intervention strategies or exposure conditions are compared, when a comparison is relevant?
- Outcome: what is measured, from what starting point and over which horizon?
- Summary: a risk difference, mean difference, rate or other explicitly named quantity?
- Events after treatment starts: how do discontinuation, switching or death affect the question and whether its outcome remains defined?
Distinguish the target from its calculation
The estimand is the quantity sought. An estimator is a method for estimating it; the estimate is the number obtained from data. Different estimators may target the same quantity under different assumptions. A logistic coefficient, for example, is not automatically a population risk difference.
Causal assumptions remain necessary
For an observational treatment comparison, specify the trial you are trying to emulate and what the available data can support. Comparable treatment groups require adequate control of confounding, meaningful treatment definitions and treatment options represented across relevant patient groups. A detailed target does not make those assumptions true.
Clinical example
A team asks whether offering a discharge call reduces 30-day readmission. One possible target is the difference in 30-day readmission risk among eligible adults assigned to an offer versus usual discharge, regardless of whether the call is completed. Restricting analysis to people who answered the phone instead changes the comparison. Death also needs an explicit competing-event strategy; it cannot silently become a missing readmission outcome.
What to watch out for
Choosing a regression model does not define the question. A precise estimand still needs a design, measurements and assumptions that can identify it; changing the population, horizon or handling of treatment changes can change the answer being sought.
Before using this in your workspace
Write a plain-language target for each aim, then check whether the required variables and follow-up exist. Record unresolved choices and missing-data assumptions with that aim. If feasibility requires changing the target, document the scientific consequence before selecting a new analysis.
- The estimands framework: a primer on the ICH E9(R1) addendum (opens in a new tab) · BMJ · doi 10.1136/bmj-2023-076316
- Estimands in practice: Bridging the gap between study objectives and statistical analysis (opens in a new tab) · Pharm Stat · doi 10.1002/pst.2056
- Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kwv254