Odds Ratio (OR)

Odds ratio (OR) is a measure of association between an exposure (or treatment) and an outcome, expressed as the ratio of the odds of the outcome in one group to the odds of the outcome in another. It is the standard association measure in case-control studies, where true population incidence cannot be observed, and it is also produced directly by logistic regression, which makes it the default effect measure for any binary-outcome model with covariates.

Odds vs. probability

Odds and probability (risk) answer related but different questions. If an event has probability p, its odds are p / (1−p). A probability of 0.75 (75%) corresponds to odds of 0.75/0.25 = 3, usually written “3 to 1.” Odds and probability move together in the same direction but at different rates — they diverge most sharply as probability approaches 1, which is exactly why an odds ratio built from two probabilities close to the extremes (very common or very rare outcomes) behaves differently from a simple risk ratio built from the probabilities themselves.

Formula

	Outcome present	Outcome absent
Exposed / Group 1	A	B
Unexposed / Group 2	C	D

Odds in Group 1 = A/B. Odds in Group 2 = C/D. OR = (A/B) ÷ (C/D) = (A×D) / (B×C) — the “cross-product” formula most references use, algebraically identical to dividing the two odds. This calculation requires only the four cell counts, not knowledge of how many people were originally “at risk” in a defined population — which is why OR, unlike relative risk, can be validly computed from a case-control sample where the case:control ratio was fixed by study design rather than observed.

Worked example

A case-control study enrolls 150 patients with a rare vascular condition and 150 matched controls, then asks about a history of a specific occupational exposure.

	Exposed	Not exposed	Total
Cases	60	90	150
Controls	30	120	150

Odds of exposure among cases = 60/90 ≈ 0.667. Odds of exposure among controls = 30/120 = 0.25. OR = 0.667 / 0.25 ≈ 2.67 — equivalently, (60×120)/(90×30) = 7,200/2,700 ≈ 2.67. Because the condition is rare in the general population, this OR is a reasonable approximation of the relative risk a cohort study of the same exposure would show — that approximation breaks down for common outcomes.

Interpreting the value

• OR = 1.0 — no association between exposure and outcome.
• OR > 1.0 — exposure associated with higher odds of the outcome (e.g. OR = 2.67 means the odds were 2.67 times higher).
• OR < 1.0 — exposure associated with lower odds (a protective association).

The critical interpretation trap: an OR should be described as “X times the odds,” never as “X times more likely” or “X times the risk” — those are relative-risk claims. The two statements coincide closely only when the outcome is rare (well under 10% prevalence); for common outcomes, OR systematically overstates the effect relative to RR, sometimes substantially. See How to Interpret an Odds Ratio for a full worked comparison of how far OR and RR can diverge on the same data, and Confidence Interval for an Odds Ratio for how to report precision alongside the point estimate.

Odds ratio vs. relative risk

Cohort studies and randomized trials observe a population’s true incidence, so they can validly report relative risk. Case-control studies fix the case:control ratio by design rather than observing true incidence, so relative risk cannot be validly calculated from them — odds ratio is used instead, and under the rare-disease assumption it closely approximates the relative risk a cohort study would find. Odds ratio is also the effect measure logistic regression produces natively for a binary outcome, which is why it remains the standard reporting metric even in some cohort analyses that use logistic rather than log-binomial or Poisson regression. See Odds Ratio vs. Risk Ratio for the full side-by-side comparison and Relative Risk (RR) for the risk-ratio definition this term is most often confused with.

Related odds-ratio metrics

• Diagnostic odds ratio — a single summary statistic for a diagnostic test’s overall discriminative ability, built from sensitivity and specificity rather than an exposure-outcome table.
• Confidence interval for an odds ratio — how to report precision, and why a CI crossing 1.0 signals a non-significant result.
• Mantel-Haenszel method — pooling odds ratios across strata or studies while adjusting for a confounder.
• Logistic regression (logit model) — the regression framework that produces adjusted odds ratios directly from its coefficients.

Frequently asked questions

Can an odds ratio be calculated from a cohort study?

Yes — OR can be calculated from any 2×2 table, including one from a cohort study. It is the case-control design that specifically requires OR rather than RR, because RR cannot be validly derived from case-control sampling; cohort data supports either measure, and RR is generally preferred there because it has a more directly interpretable “risk” meaning.

Why does OR overstate the effect for common outcomes?

Because odds and probability diverge as probability moves away from small values. For a rare outcome, odds ≈ probability, so OR ≈ RR. As the outcome becomes common, the denominator of the odds calculation (1 − probability) shrinks, inflating the odds and therefore the OR relative to the underlying risk ratio — an OR of 3.5 can correspond to a real risk ratio of only 2.0 for a sufficiently common outcome.

Is an OR below 1 the same as a “negative” effect?

An OR below 1.0 indicates a protective or risk-reducing association, not a mathematically negative number — OR is a ratio and cannot be negative. An OR of 0.5 means the odds are half as large in the reference group’s comparison, a genuine reduction in odds, described correctly as “50% lower odds,” not “a 50% reduction in risk” unless the outcome is rare enough for OR and RR to coincide.