E-Value Sensitivity Analysis
- Introduced by
- Tyler J. VanderWeele & Peng Ding
- Year
- 2017
- Type
- Assumption-free sensitivity analysis for unmeasured confounding
- Family
- Causal inference / sensitivity analysis
What does it do?
The E-value, introduced by Tyler VanderWeele and Peng Ding in 2017, is a simple, assumption-free way to quantify how robust an observational association is to unmeasured confounding.
It answers a single, sharply posed question: how strong would an unmeasured confounder have to be — in its association with both the exposure and the outcome — to fully explain away the observed effect? The larger the E-value, the more powerful a hidden confounder would need to be, and so the more robust the finding. The method rests on the bounding factor derived by Ding and VanderWeele in their 2016 'Sensitivity analysis without assumptions,' which holds regardless of the distribution or number of unmeasured confounders. Because it requires only the point estimate and confidence limit on the risk-ratio scale and no untestable bias parameters, the E-value has become a routine reporting standard in observational epidemiology, including social epidemiology where unmeasured confounding is pervasive.
The intuition behind the method
Every observational study faces the worry: maybe the association isn't causal, maybe some unmeasured factor drives both the exposure and the outcome. Traditional sensitivity analyses ask you to guess that confounder's properties — its prevalence, its effect — and then see what happens, which is hard because the confounder is by definition unmeasured. The E-value flips the question into something you can actually reason about. Instead of asking 'what if the confounder looks like this?', it asks 'what is the minimum strength a confounder would need to undo my result?' and reports that single number. If the answer is, say, 3.5, then only an unmeasured confounder associated with both exposure and outcome by a risk ratio of at least 3.5 (above and beyond all the variables you did adjust for) could reduce the effect to nothing. You then judge whether such a strong hidden factor is plausible, often by comparing it to how strongly your measured confounders are associated with the outcome. A large E-value relative to those means the finding is hard to explain away.
History
Sensitivity analysis for unmeasured confounding has a long history, from Cornfield's conditions in the 1959 smoking-and-lung-cancer debate to Rosenbaum's bounds and elaborate bias-parameter models, but these typically required assumptions about the confounder or specialized expertise that limited routine use. Peng Ding and Tyler VanderWeele's 2016 Epidemiology paper 'Sensitivity analysis without assumptions' derived a sharp bounding factor that holds for any unmeasured confounder, free of distributional assumptions. Building directly on that bound, VanderWeele and Ding's 2017 Annals of Internal Medicine paper distilled it into the single, easy-to-compute E-value and argued for its routine reporting. The simplicity and generality of the idea led to rapid adoption: the E-value is now widely reported across epidemiology and medicine, supported by the EValue R package and online calculators, and has become a near-standard component of observational study reporting.
When to use it
Use the E-value whenever you report an observational exposure-outcome association and want a transparent, standardized statement of how vulnerable it is to unmeasured confounding — which is essentially always in observational social epidemiology. It is ideal as a default sensitivity analysis because it needs only the effect estimate and its confidence limit on (or convertible to) the risk-ratio scale and demands no guesses about the unmeasured confounder. It is well suited to communicating robustness to non-technical audiences, since it reduces the confounding worry to one interpretable number. It is most natural for binary or rate outcomes and ratio measures; for other scales the conversion introduces approximation that should be acknowledged. Do not treat a large E-value as proof of causality — it bounds confounding bias only, not selection bias, measurement error, or reverse causation — and pair it with subject-matter judgment about plausible confounders.
Limitations and considerations
- Addresses unmeasured confounding only; it says nothing about selection bias, measurement error, or reverse causation.
- Defined on the risk-ratio scale, so odds ratios, hazard ratios, and continuous effects require approximate conversions.
- Assumes the two confounder associations are equal at their joint maximum, a worst-case framing that can be conservative.
- A large E-value does not prove causality; plausibility of a confounder of that strength still requires substantive judgment.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Assumption-free: the underlying bound holds regardless of the unmeasured confounder's distribution, type, or number.
- Requires no untestable sensitivity parameters — only the effect estimate and confidence limit on the risk-ratio scale.
- Reduces the confounding question to a single interpretable number that is easy to report and to communicate.
- Provides E-values for both the point estimate and the confidence limit, separating robustness of magnitude from robustness of significance.
Applications
- Reporting the robustness of an observational exposure-outcome association to unmeasured confounding as a standard sensitivity analysis.
- Assessing whether a protective or harmful social-exposure effect could be explained by an unmeasured socioeconomic or behavioral confounder.
- Benchmarking the required confounder strength against the strongest measured confounders to judge plausibility.
- Strengthening causal interpretation in meta-analyses and systematic reviews of observational studies.
- Communicating confounding vulnerability to clinicians and policymakers with a single interpretable number.
How it works
How it works
The analysis starts from the study's effect estimate expressed as a risk ratio. The E-value is defined on the risk-ratio scale, so other measures are first converted: odds ratios from common outcomes and hazard ratios are approximated as risk ratios, and standardized mean differences are mapped to an approximate risk ratio using VanderWeele and Ding's conversion. Both the point estimate and the confidence limit closest to the null are carried forward, since the method produces an E-value for each. Effects below the null are inverted so the computation always works with a risk ratio above one.
Formula
RR_{\text{obs}} = \frac{\Pr(Y=1\mid A=1)}{\Pr(Y=1\mid A=0)};\quad \text{convert OR, HR, or } d \text{ to approximate } RRUnderlying the E-value is the Ding-VanderWeele bounding factor, which states that the maximum amount by which unmeasured confounding can bias a risk ratio is governed by two parameters: the association between the confounder and the exposure and the association between the confounder and the outcome, both on the risk-ratio scale. The true effect, adjusted for the hidden confounder, can be no smaller than the observed effect divided by this bounding factor. Crucially, this inequality is sharp and requires no assumptions about the confounder's distribution, whether it is binary or continuous, or how many confounders there are — which is what makes the approach 'assumption-free.'
Formula
RR_{\text{true}} \ge \frac{RR_{\text{obs}}}{B},\quad B = \frac{RR_{AU}\,RR_{UY}}{RR_{AU}+RR_{UY}-1}The E-value for the point estimate is obtained by finding the smallest value of those two confounder associations — taken to be equal — that would be just large enough to drive the observed risk ratio to the null. Solving the bounding-factor equation for that common value gives the closed-form expression: the observed risk ratio plus the square root of the risk ratio times itself minus one. The result is a single number on the risk-ratio scale: the minimum joint strength of confounder-exposure and confounder-outcome association required to explain away the estimate entirely.
Formula
E\text{-value} = RR_{\text{obs}} + \sqrt{RR_{\text{obs}}\,(RR_{\text{obs}}-1)}Robustness of the conclusion, not just the point estimate, is assessed with an E-value for the confidence interval. The same formula is applied to the confidence limit nearest the null (the lower limit for a risk ratio above one): it gives the minimum confounder strength needed to shift that limit to the null, i.e., to make the result no longer statistically distinguishable from no effect. This E-value is always smaller than the one for the point estimate, and it is often the more policy-relevant figure because it speaks to whether the evidence of an effect — not merely its magnitude — could be undone by confounding.
Formula
E\text{-value}_{\text{CI}} = LL + \sqrt{LL\,(LL-1)},\quad LL = \text{lower confidence limit} \ (>1)Finally the E-value is interpreted in context. A confounder would have to be associated with both exposure and outcome by at least the E-value (on the risk-ratio scale, beyond the measured covariates) to nullify the result; weaker confounders could not. To judge plausibility, analysts compare the E-value to the observed associations of the strongest measured confounders: if no measured variable comes close to the E-value, an unmeasured one of that strength is usually implausible and the finding is robust. The E-value is reported as a complement to, not a replacement for, the estimate and its confidence interval.
Formula
RR_{AU}=RR_{UY}=E\text{-value} \Rightarrow RR_{\text{true}}=1;\quad \text{compare to } RR \text{ of measured confounders}Common pitfalls
- Interpreting the E-value as the actual amount of confounding present rather than the minimum needed to explain the result away.
- Reporting only the point-estimate E-value and ignoring the smaller, often decisive, confidence-interval E-value.
- Treating a large E-value as evidence of causality while overlooking selection bias, measurement error, or reverse causation.
- Converting odds or hazard ratios to risk ratios for rare-then-common outcomes without noting the approximation error.
Frequently asked questions
What exactly does the E-value tell me?
The E-value is the minimum strength of association, on the risk-ratio scale, that an unmeasured confounder would need to have with both the exposure and the outcome — over and above the variables already adjusted for — to fully explain away the observed effect. For example, an E-value of 2.5 means a confounder associated with both exposure and outcome by a risk ratio of 2.5 each could account for the result, but anything weaker could not. It is the threshold a hidden confounder must exceed to undo your finding, not an estimate of confounding that is actually present.
Why is there a separate E-value for the confidence interval?
The point-estimate E-value tells you how strong a confounder must be to move the estimate to the null, but a result can remain significant even after substantial bias. The confidence-interval E-value applies the same formula to the confidence limit nearest the null, giving the minimum confounder strength needed to make the result statistically compatible with no effect. It is always smaller than the point-estimate E-value and is frequently the more important figure, because it speaks to whether the evidence that there is any effect at all could be overturned by unmeasured confounding.
Does a large E-value mean the effect is causal?
No. A large E-value means unmeasured confounding alone would have to be implausibly strong to explain the association away, which strengthens a causal interpretation — but only with respect to confounding. The E-value is silent about other threats to validity such as selection bias, measurement error, and reverse causation, and it does not by itself establish causation. It should be read as one piece of evidence: a quantified statement of robustness to confounding, to be combined with study design, subject-matter knowledge about plausible confounders, and other sensitivity analyses.
Related methods
Sources
VanderWeele, T. J., & Ding, P. (2017). Sensitivity analysis in observational research: introducing the E-value. Annals of Internal Medicine, 167(4), 268-274.
Open DOI referenceCopied.Ding, P., & VanderWeele, T. J. (2016). Sensitivity analysis without assumptions. Epidemiology, 27(3), 368-377.
Open DOI referenceCopied.