Ask Mallard
Analysis method

E-value

5 min read

An E-value expresses the strength of unmeasured confounding that could explain away an observed risk-ratio association.

Fit

When to use it, and when not to

Use it when

  • You have an adjusted risk ratio and want a clear sensitivity measure for unmeasured confounding.

Not when

  • Your main concern is missing-data bias or selection bias. Choose a sensitivity analysis aimed at that bias. Read: Sensitivity analyses
  • Your effect is a mean difference. Use a sensitivity method for that scale rather than inserting it in the risk-ratio formula. Read: Sensitivity analyses

Before you start

What you need

Decide these first
  1. An adjusted estimate on the risk-ratio scalesummaryMeasureRecord the estimate and what covariates were controlled.
  2. The confidence intervalThe limit nearest no effect measures how much confounding could move the interval to include no effect.
  3. Plausible unmeasured confoundersCompare the bound with credible strengths of their links to exposure and outcome.
It rests on these
  1. The formula receives a risk ratio.How to check: Do not insert a common-outcome odds ratio or hazard ratio directly. Use justified scale conversions with their assumptions.
  2. Unmeasured confounding is the bias being examined.How to check: The bound concerns links beyond measured covariates. Other biases require separate analyses.
  3. The two confounder links may have different strengths.How to check: The E-value gives an equal-strength benchmark. A weaker link can be offset by a stronger other link.
  4. Choose a missing-data method because of a stated assumption.How to check: Missing completely at random (MCAR) means gaps depend on neither observed nor unseen data. Missing at random (MAR) means recorded data explain gaps. Under MAR, suitable multiple imputation may help; it is not inherently better. Missing not at random (MNAR) means unseen values still predict gaps. Under plausible MNAR, add a delta or tipping-point sensitivity analysis. Dropping people with missing exposure can cause selection bias.

Plain language

How it works

For a risk ratio R greater than 1, the E-value is R + √(R × (R − 1)). For a protective ratio, invert it first. This bound describes joint exposure-confounder and confounder-outcome associations beyond measured covariates. It does not estimate those associations.

Include uncertainty and context

Repeat the calculation for the confidence limit closest to 1. If the interval includes 1, its E-value is 1. For a protective interval wholly under 1, invert its upper limit. Report the effect, interval and both E-values. Judge plausibility against clinical knowledge rather than a universal cutoff.

The bias being challenged
A causal diagram with three nodes. A confounder has arrows into both the exposure and the outcome, and the exposure has an arrow into the outcome labeled the effect of interest. The path from exposure back through the confounder to the outcome is not the effect of interest.Confounding: a common cause of exposure and outcomeEffect of interestConfounderExposureOutcomeExposure and outcome are associated through the confounder even with no effect at all. Adjusting for theconfounder closes that path.
Conceptual diagram. It shows the structure of the idea, not data from any study.

Worked example

A worked example

Invented for illustration
Setting
A cohort reports an adjusted risk ratio for an adverse outcome.
Question
How strong would equal confounder links need to be to explain the result?
Approach
Suppose the risk ratio is 2.00, with a reported confidence interval from 1.25 to 3.20.
Result
The estimate’s E-value is 3.41. The confidence limit’s E-value is 1.81.
What it means
Equal confounder links of 3.41 on the risk-ratio scale could explain the estimate. Links of 1.81 could move the interval to include no effect.

In practice

Recent studies PubMed search · not vetted

Searching PubMed…

Watch out

Common mistakes

The one to remember

An E-value is a bound for unmeasured confounding, not proof of causality or the probability that a finding is true. It does not address selection bias, measurement error or missing-data bias.

  • Reporting only the estimate’s E-value.Instead: Also report the E-value for the confidence limit nearest no effect.
  • Treating a large E-value as evidence that no confounder exists.Instead: Discuss concrete candidate confounders and plausible strengths.
  • Requiring both links to exceed the E-value in every case.Instead: The benchmark uses equal strengths. Unequal links can also explain the association.

Inside Mallard

How Mallard uses this

Plans do not link to this method yet. You can find it in Learn and the Atlas.

Code example in Mallard: E-value and a simple quantitative bias analysis for unmeasured confounding (e-value-quantitative-bias).
Atlas node analysis.e_value, the id a plan cue records.
Code for this method
Loading…

Related Learn pages

Related designs and methods

Evidence

Sources & further reading

  1. Sensitivity Analysis in Observational Research: Introducing the E-Value (opens in a new tab) · Ann Intern Med · doi 10.7326/M16-2607
  2. Sensitivity Analysis Without Assumptions (opens in a new tab) · Epidemiology · doi 10.1097/EDE.0000000000000457
Prepared with AI assistance for the owner-approved one-method wave 1. Indexed sources were checked by PMID, title, authors, year and DOI against PubMed on 2026-11-06. Record verification is not methodological approval. VanderWeele and Ding introduce the E-value. Ding and VanderWeele derive the confounding bound, including unequal confounder links. These sources do not make the E-value a test for other biases.
Last source review: 2026-11-06
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.