Robustness check
Also known as: Sensitivity analysis
Re-running an analysis under different assumptions to test if the result survives.
What it means
A robustness check assesses whether a key finding persists when the analysis is varied in reasonable ways, such as altering model specifications, sample restrictions, outlier handling, or measurement choices. Its purpose is to show that a conclusion is not an artifact of one particular set of assumptions and to gauge how sensitive the estimate is to them. Closely related sensitivity analyses probe how strong an unmeasured confounder would have to be to overturn a causal claim. While reassuring when results hold, robustness checks can themselves be selectively reported, so their credibility is greatest when the full set of checks is pre-specified or exhaustively shown.
Examples
Confirming a treatment effect remains significant whether or not outliers are excluded and across two alternative model specifications.
Before rolling out a pricing change, an analyst re-runs the A/B test with and without Black Friday week and with bot traffic filtered out — the lift holds every way, so it ships.
A study linking long commutes to poor sleep survives when the authors also control for shift work, restrict to full-time workers, and swap self-reported hours for phone-tracked ones.
First described in Econometrics and applied-statistics practice.