E-Value (Sensitivity Analysis for Unmeasured Confounding)

What it is

— The E-value is a sensitivity-analysis metric introduced by VanderWeele and Ding (Annals of Internal Medicine, 2017) for quantifying how robust an observed exposure-outcome association is to unmeasured (residual) confounding. It is the minimum strength of association — expressed as a risk ratio — that an unmeasured confounder would need to have with both the exposure and the outcome, over and above the measured covariates, to reduce the observed point estimate to the null; a companion E-value is computed for the confidence-limit closest to the null. Unlike a checklist with numbered items, the E-value is a single, transparent number that any reader can interpret without distributional assumptions about the unmeasured confounder. It is not a regulatory checklist or a risk-of-bias instrument in the EQUATOR/Cochrane sense; rather it is a methods standard that reporting guidelines and agencies increasingly expect within the sensitivity-analysis section of an observational study. ENCePP methodological standards, HARPER/STaRT-RWE protocol templates, and high-impact journals all now treat a quantitative bias analysis for unmeasured confounding (of which the E-value is the most common form) as part of a complete, decision-grade report. There is no committee that "maintains" the E-value; it is maintained as a literature — the original statement paper, the technical-considerations follow-up, clinician-facing explanations, and the freely available `EValue` R package and web calculator (Mathur et al., 2018).

When to use

— Report an E-value whenever an observational (non-interventional) effect estimate is offered for a causal or comparative-effectiveness interpretation and unmeasured confounding is a live threat: cohort, case-control, new-user, active-comparator new-user, and target-trial-emulation studies in claims, EHR, registry, or linked data. It belongs in FDA/EMA RWE submissions, HTA/payer dossiers, and peer-reviewed manuscripts as the headline robustness statement accompanying the primary adjusted estimate (and key subgroups). Decision rule for which sensitivity tool: use the E-value as the default, communicable summary when you have a single binary/rate/HR-type contrast and want a threshold readers can benchmark against known confounder strengths. Escalate beyond a bare E-value to fuller quantitative bias analysis — bias formulas with specified confounder prevalence, bounding factors, probabilistic/Monte Carlo bias analysis, or negative-control / empirical calibration — when you can credibly parameterize the suspected confounder, when multiple biases act jointly, or when a regulator/HTA reviewer asks for a quantified rather than threshold-style assessment. The E-value is the floor of good practice, not the ceiling.

What it requires

— As a reporting standard the E-value enforces, in the study's sensitivity analysis, that authors: (1) compute the E-value for the point estimate and for the confidence limit nearest the null, and report both — a large point-estimate E-value with a confidence limit E-value near 1 signals fragility; (2) compute it on the correct scale (risk ratio; approximate conversions are needed for odds ratios with common outcomes, hazard ratios, and standardized mean differences, and the approximation should be acknowledged); (3) interpret the number against plausible real-world confounder strengths — i.e., name a measured confounder of comparable strength and argue whether an unmeasured one of that magnitude could plausibly exist after the design and adjustment already applied; and (4) tie the E-value back to the design choices that reduced confounding in the first place — active-comparator/new-user restriction, time-zero alignment, fit-for-purpose data, and the propensity-score/covariate strategy. In RWE specifically, the E-value presupposes that the measured confounding has already been handled credibly and that the estimand, intercurrent-event handling, and outcome/exposure phenotypes are sound; it speaks only to the residual gap.

When NOT to use — limitations and common misapplications

— (a) It is not a quality score and not a risk-of-bias tool. A large E-value does not certify a study as unbiased; it only describes robustness to one bias (unmeasured confounding) conditional on everything else being correct. (b) It does not address other biases — selection bias, immortal-time bias, misclassification of exposure or outcome, missing data, or measurement error are outside its scope; reporting an E-value while ignoring these is checklist-as-theater. (c) It is not a license to claim causation — computing an E-value on a hopelessly confounded claims comparison does not make the contrast causal. (d) Misreading the threshold: a common error is reporting only the point-estimate E-value and omitting the confidence-limit E-value, overstating robustness; another is treating the E-value as the probability of an explanatory confounder rather than the strength one would need. (e) Scale errors: applying the risk-ratio formula to a hazard ratio or odds ratio from a common outcome without the appropriate transformation inflates or deflates the value. (f) Critique to disclose: the E-value has been criticized (notably by Ioannidis and colleagues) for being easy to report mechanically and for inviting overinterpretation when divorced from substantive argument about what the unmeasured confounder actually is; the defensible practice, set out in VanderWeele, Ding, and Mathur's technical-considerations paper, is to pair the number with a named, plausible confounder and the design that already constrains it. Do not substitute an E-value for the harder design work (negative controls, ACNU, hdPS); use it to summarize what residual risk remains after that work.

How it maps to this catalog

— The computational mechanics, formulas, scale conversions, and the `EValue` R package live in the implementing concept e-value-sensitivity-analysis; this guideline entry is the when/why/report-it layer that points to it. Broader bias quantification — bounding factors, probabilistic bias analysis, multiple-bias modeling — is in quantitative-bias-analysis-toolkit-rwe, the natural escalation when a bare E-value is insufficient. The E-value only earns its keep on top of credible measured-confounding control: active-comparator-new-user (removing confounding by indication and immortal time at the design stage) and high-dimensional-propensity-score-hdps-rwe (proxy-based adjustment) are the prerequisites it summarizes residual risk against, and empirical-calibration-negative-controls-rwe / negative-control-outcomes-rwe are complementary empirical checks that triangulate with it. The estimand the E-value qualifies should be defined via estimands-ate-att-intercurrent-events-rwe; the exposure/outcome it depends on must be validly defined via diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe; the data must be appropriate per fit-for-purpose-data-assessment-rwe; cohort losses that could masquerade as confounding belong to attrition-and-loss-to-follow-up-rwe; and the question itself should be framed with picots-framework-rwe. For a claims/EHR/registry applied note: after building an ACNU cohort in claims (see claims-analysis), balancing on an hdPS, and estimating, say, an adjusted HR of 0.70 for a comparative-effectiveness outcome, report the E-value for the HR and for its upper confidence limit, then state explicitly which unmeasured claims-invisible factor (frailty, disease severity, smoking, over-the-counter use) of that strength would be needed and whether the active-comparator design plus hdPS proxies plausibly already capture it.