The estimand
What question does this help answer?
It answers "what exactly are we measuring", which sounds obvious until two analyses of the same trial disagree and it turns out they were estimating different quantities all along.
What it is made of
Five components: the POPULATION the estimate applies to; the TREATMENT conditions being compared; the OUTCOME variable; how INTERCURRENT EVENTS are handled, meaning things like death, treatment discontinuation or rescue medication that happen after randomization and affect interpretation; and the POPULATION-LEVEL SUMMARY, such as a risk difference or a hazard ratio.
What an estimand is made of, drawn
Why intercurrent events are the hard part
The other four components are usually stated somewhere. Intercurrent events are the ones that get left implicit, and they are where two reasonable analysts diverge. A treatment-policy strategy asks what happens if we assign this treatment, regardless of what people do next. A while-on-treatment strategy asks what happens while people actually take it. A composite strategy treats the event itself as part of the outcome. These are different clinical questions with different answers, and choosing between them is not a statistical detail.
Why Mallard types it
Every plan Mallard writes carries a TYPED estimand, and everything downstream derives from it: the recommended analysis must estimate the named summary measure, the sample size method must size that measure, and the exhibits must present it. The app checks those mechanically, and where the sizing method contradicts the estimand it WITHHOLDS the headline sample size rather than printing a number that answers a different question.
Estimand and target of inference are not the same as the analysis
A common misreading is that the estimand is just the model. It is not. The model is one way to estimate the quantity; the estimand is the quantity. Two different models can target the same estimand, and one model can target different estimands depending on how intercurrent events are handled.
Key assumptions and design requirements
- The estimand is specified in the protocol, before the data exist.
- Intercurrent events are ENUMERATED and a strategy is named for each, not left to the analyst.
- The population summary measure is stated, because a risk difference and an odds ratio are different estimands of the same comparison.
- The main analysis targets the stated estimand, and sensitivity analyses vary the assumptions rather than the question.
- For an observational study the same discipline applies, with the additional burden of stating the causal assumptions that make the estimate interpretable.
What to watch out for
An estimand that is written down after the analysis is a description, not a specification, and it cannot constrain anything. The commonest failure is leaving intercurrent events unaddressed, so the reported effect silently mixes several different questions depending on how dropouts, deaths and treatment switches happened to be handled.
Clinical example
Question: Does a new inhaler reduce exacerbations? Two estimands: the effect of ASSIGNING the inhaler regardless of adherence or later switching (treatment policy), and the effect WHILE PATIENTS ARE TAKING IT (while on treatment). The second will usually look better. Which one a formulary committee needs is a clinical judgement, and reporting one while implying the other is the failure this framework exists to prevent.
- Estimands: bringing clarity and focus to research questions in clinical trials (opens in a new tab) · BMJ Open · doi 10.1136/bmjopen-2021-052953
- Overview of clinical study designs (opens in a new tab) · Clin Exp Emerg Med · doi 10.15441/ceem.23.036