Sensitivity Analysis

In causal inference, sensitivity analysis asks how robust a study's conclusion is to violations of its untestable assumptions, most often the assumption that there is no unmeasured confounding. Rather than treating a point estimate as final, it quantifies how strong a hidden bias would need to be to explain away, or substantially change, the observed effect.

Definition

Sensitivity analysis in causal inference is a set of methods that quantify how much an unmeasured bias, typically unmeasured confounding, would have to be present to materially alter or nullify an estimated causal effect.

Scope

This topic covers the rationale for sensitivity and quantitative bias analysis in observational research, the notion of an unmeasured confounder strong enough to overturn a result, and summary measures such as the E-value. It is a methodological reference, not clinical guidance.

Core questions

• How robust is a causal estimate to unmeasured confounding?
• How strong would a hidden bias need to be to explain away the result?
• How can the impact of untestable assumptions be communicated to readers?

Key concepts

• Unmeasured confounding
• Quantitative bias analysis
• E-value
• Bias bounds
• Robustness of causal estimates

Mechanisms

Sensitivity analysis acknowledges that key identification assumptions cannot be verified from the data and instead asks what magnitude of violation would change the conclusion. Ding and VanderWeele (ding-vanderweele-2016) derived bounds showing how strongly an unmeasured confounder, in its associations with both exposure and outcome, would have to act to account for an observed association, without requiring assumptions about the confounder's distribution. Building on this, the E-value (vanderweele-ding-2017) summarises the result as the minimum strength of association that an unmeasured confounder would need with both exposure and outcome to fully explain away the estimate; larger E-values indicate greater robustness. Such analyses complement careful adjustment, which itself must avoid introducing bias through inappropriate covariate choice (schisterman-2009).

Clinical relevance

Sensitivity analysis tells readers how much confidence an observational causal claim can bear, which is central to appraising evidence on treatments and exposures. It describes the robustness of evidence and is not a basis for individual diagnostic or treatment decisions.

Epidemiology

Quantitative bias analysis and the E-value are increasingly reported alongside primary estimates in observational epidemiology and comparative-effectiveness research, where unmeasured confounding is a persistent concern. The E-value in particular has been widely adopted as a concise robustness summary (vanderweele-ding-2017).

History

Formal sensitivity analysis for unmeasured confounding has roots in mid-twentieth-century debates over observational causal claims, and modern quantitative bias analysis was systematised in epidemiology textbooks (rothman-2008). The assumption-light bounds of Ding and VanderWeele (ding-vanderweele-2016) and the subsequent E-value (vanderweele-ding-2017) made such analyses simple enough for routine reporting.

Debates

How should the E-value be interpreted and used?

The E-value offers a convenient robustness summary, but commentators caution it is not a substitute for substantive reasoning about which confounders are plausible and how strong they could realistically be.

Key figures

• Tyler VanderWeele
• Peng Ding
• Sander Greenland
• Timothy Lash

Seminal works

• ding-vanderweele-2016
• vanderweele-ding-2017

Frequently asked questions

What does an E-value tell you?

It reports the minimum strength of association an unmeasured confounder would need with both the exposure and the outcome to fully explain away the observed effect; a larger E-value means the result is more robust to such confounding.

Does sensitivity analysis prove a result is correct?

No. It does not confirm causation; it quantifies how robust a finding is to specified unmeasured biases, leaving the plausibility of those biases to substantive judgement.