Sensitivity Analysis for Hidden Bias (Rosenbaum Bounds / E-value)
Sensitivity Analysis for Hidden Bias in Observational Studies (Rosenbaum Bounds / E-value) · Also known as: Rosenbaum bounds, E-value, hidden bias sensitivity analysis, unmeasured confounding sensitivity, Duyarlılık Analizi — Gizli Yanlılık (Rosenbaum / E-value)
Sensitivity analysis for hidden bias is a family of methods that quantify how strongly an unmeasured confounder would have to operate before it could overturn a causal conclusion drawn from observational data. It was crystallised by Paul Rosenbaum's sensitivity bounds (2002) and extended by VanderWeele and Ding's E-value (2017).
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use sensitivity analysis after a primary observational causal estimate (such as a matched comparison or regression adjustment) to gauge how robust that estimate is to unmeasured confounding. It is appropriate when adequate adjustment for observed covariates has already been done, with at least about 50 observations so the reference effect size can be estimated reliably. It complements rather than replaces identification: it does not remove bias but reports how much hidden bias the conclusion can tolerate. Interpreting the Γ threshold or E-value requires domain knowledge — they are not purely statistical cut-offs.
Strengths & limitations
- Directly quantifies the robustness of a causal finding to unmeasured confounding instead of assuming such confounding away.
- The E-value is a single, scale-free number that is easy to report and compare across studies.
- Applies on top of an existing primary analysis (matching, regression) without requiring a new identification strategy.
- It does not correct bias; it only describes how much hidden bias would change the conclusion.
- Below about 50 observations the reference effect size cannot be estimated reliably, so the analysis is uninformative.
- Translating a Γ threshold or E-value into 'plausible' or 'implausible' requires substantive domain judgement, not statistics alone.
Frequently asked
What does the E-value actually measure?
It is the minimum association strength, on the risk-ratio scale, that an unmeasured confounder would need to have with both the treatment and the outcome to fully explain away the observed effect. A larger E-value means the finding is harder to overturn.
What is the Rosenbaum Γ parameter?
Γ bounds how much two units matched on observed covariates could still differ in their odds of receiving treatment because of an unmeasured factor. Γ = 1 means no hidden bias; raising Γ shows the point at which the inference would change.
Does sensitivity analysis remove confounding?
No. It does not correct or eliminate bias. It only quantifies how strong a hidden confounder would have to be to alter the conclusion, so robust findings can be distinguished from fragile ones.
What if the E-value comes out small?
A small E-value means even a weak unmeasured confounder could nullify the result, so the primary finding is fragile. This signals that a stronger identification strategy, such as instrumental-variable estimation, may be needed.
Sources
- Rosenbaum, P. R. (2002). Observational Studies (2nd ed.). Springer. ISBN: 978-0387989679
- VanderWeele, T. J. & Ding, P. (2017). Sensitivity Analysis in Observational Research: Introducing the E-Value. Annals of Internal Medicine, 167(4), 268-274. DOI: 10.7326/M16-2607 ↗
How to cite this page
ScholarGate. (2026, June 1). Sensitivity Analysis for Hidden Bias in Observational Studies (Rosenbaum Bounds / E-value). ScholarGate. https://scholargate.app/en/causal-inference/sensitivity-analysis-observational
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Frontdoor AdjustmentCausal inference↔ compare
- Local Average Treatment EffectCausal inference↔ compare
- Placebo TestsCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare
- Two-Stage Least Squares (2SLS)Causal inference↔ compare