Question 7 of 9 Intermediate Troubleshooting
How would you explain the difference between a statistical estimand and the estimator used to calculate it?
Short answer
The estimand is the precisely defined quantity or treatment effect we want to know. The estimator is the mathematical or computational procedure used to estimate it from data.
Why
An estimand specifies population, treatment or comparison, endpoint, summary and handling of intercurrent events. Two estimators may target the same estimand under different assumptions. Conversely, an apparently familiar estimator can target the wrong scientific question when missingness or treatment switching changes interpretation.
How to work it through
- Define the scientific decision and target population before selecting a model.
- Specify treatment conditions, variable, population summary and intercurrent-event strategies.
- Then choose an estimator whose assumptions and data collection align with that target, report uncertainty and conduct sensitivity analyses that keep the estimand fixed while varying assumptions.
An example
A trial wants the treatment effect regardless of subsequent therapy changes. That treatment-policy estimand is not automatically answered by an analysis censoring everyone at treatment discontinuation. A model that explicitly conditions on discontinuation may instead estimate a different effect.
Where the answer stops holding
In laboratory studies the estimand might be mean treatment effect across donors, not mean across all cells. Neither a p-value nor the coefficient name defines an estimand. Multiple clinically meaningful estimands can coexist.
A strong answer covers
- The estimand is the precisely defined quantity or treatment effect we want to know
- The estimator is the mathematical or computational procedure used to estimate it from data
- Define the scientific decision and target population before selecting a model
Common mistakes
- Equating estimand with sample mean
- Treating a sensitivity analysis for a different target as a robustness check for the primary estimand
The interviewer might ask next
Could two different estimators legitimately answer the same clinical question?