Written and maintained by CASRAI Editorial Board
Last updated
An odds ratio (OR) of 2.0 does not mean “twice as likely” — it means the odds of the outcome are twice as large, and odds are not the same thing as probability (risk). This distinction sounds pedantic until you see how far the two diverge numerically: for a common outcome, an OR of 3.5 can correspond to a real risk ratio of only 2.0. Reporting “3.5 times more likely” when the true risk increase is 2-fold is not a rounding error — it is a different, larger claim than the data support. This guide gives the correct interpretation language, works a numeric example that shows exactly where OR and RR pull apart, and covers the below-1 direction confusion and the adjusted-versus-crude reading that trip up even careful readers of a regression table.
A Worked Example First
Suppose a study compares a complication rate between two treatment groups, 100 patients each:
| Complication | No complication | Total | |
|---|---|---|---|
| Treatment A | 60 | 40 | 100 |
| Treatment B | 30 | 70 | 100 |
Two different summary numbers can be computed from this same table, and they answer two different questions:
- Risk (probability) of complication: Group A = 60/100 = 0.60; Group B = 30/100 = 0.30. Risk ratio = 0.60 / 0.30 = 2.0 — Group A patients were twice as likely to have a complication.
- Odds of complication: Group A = 60/40 = 1.5; Group B = 30/70 ≈ 0.4286. Odds ratio = 1.5 / 0.4286 ≈ 3.5 — Group A’s odds of complication were 3.5 times Group B’s odds.
Both numbers are correct descriptions of the same data. The sentence “patients in Group A were 3.5 times more likely to have a complication” is not one of them — that sentence describes the risk ratio, and the actual risk ratio here is 2.0. The OR of 3.5 overstates the true risk increase by 75% when read as if it were a risk ratio. This is exactly why the wrong-phrasing trap matters: “X times more likely” is risk-ratio language, and substituting an odds ratio into it silently inflates whatever effect you’re describing.
Odds Are Not Probability — The Language That Stays Correct
Odds and probability are related but not interchangeable: odds = p / (1 − p), where p is the probability of the outcome. At low probabilities (a rare outcome), odds and probability are numerically close, which is why the confusion so often goes unnoticed — and why it becomes dangerous exactly when the outcome is common, as in the worked example above. The language that stays accurate for an odds ratio:
- Correct: “The odds of the outcome were 2.0 times higher in Group A” or “Group A had twice the odds of the outcome.”
- Correct: “The odds ratio was 2.0 (95% CI, x–y).”
- Incorrect: “Group A patients were 2.0 times more likely to have the outcome” — “more likely” is probability language, and it is only a valid restatement of an odds ratio when the outcome is rare (conventionally under about 10% in both groups). Outside that range, say “odds,” not “likely” or “risk.”
This is the same reason a hazard ratio cannot be read as “more likely” either — both are ratios of a different underlying quantity (a rate, or odds) being casually substituted for a ratio of risk. The fix in both cases is the same: name the quantity the ratio is actually a ratio of.
When an Odds Ratio Does Approximate a Risk Ratio
The “rare disease assumption” is the specific, named condition under which reading an OR as if it were an RR is defensible as an approximation rather than an error. When the outcome’s probability is low in both compared groups (commonly cited as roughly under 10%), (1 − p) in the odds formula is close to 1 for both groups, so odds ≈ probability and OR ≈ RR. This is also the reason case-control studies report an odds ratio rather than a risk ratio in the first place: a case-control design fixes the ratio of cases to controls by the researcher’s sampling choice rather than by true population prevalence, so a valid risk ratio cannot be calculated directly from the sample at all — odds ratio is the only measure of association case-control data can legitimately produce, and under the rare-disease assumption it closely approximates the risk ratio a cohort study of the same exposure would have shown. See CASRAI’s case-control vs. cohort study comparison for the full design distinction. Cohort and randomized-trial designs, by contrast, can report either measure directly — and when the outcome is common, as in the worked example above, they diverge exactly as shown, which is why choosing (and correctly labeling) the right one matters.
The Below-1 Direction Confusion
An OR below 1.0 indicates lower odds of the outcome in the group being compared to the reference group — a protective or negative association, not “no effect.” The confusion that trips up readers here runs two directions:
- Misreading the magnitude. An OR of 0.5 means the odds are half those of the reference group — it does not mean “50% lower risk.” Converting an OR below 1 to a percentage change in odds is (1 − OR) × 100, e.g., OR = 0.68 → a 32% reduction in odds, not in probability of the outcome.
- Misreading the reference direction. Which group is “higher odds” and which is “lower odds” depends entirely on which category was coded as the reference level in the model. An OR of 0.5 for “Group A vs. Group B” is the same underlying association as an OR of 2.0 for “Group B vs. Group A” — always check the stated reference category before interpreting the direction, and always state it explicitly when reporting your own results.
The confidence interval is what determines whether a below-1 (or above-1) OR is distinguishable from no association at all: if the 95% CI spans 1.0, the data are consistent with no difference in odds between groups, regardless of how far the point estimate itself sits from 1.0.
Adjusted vs. Crude: Reading an Odds Ratio in a Regression Table
A regression results table commonly reports two odds ratios for the same predictor, and conflating them is a separate, common misreading:
| Predictor | Crude OR (95% CI) | Adjusted OR (95% CI) | p-value |
|---|---|---|---|
| Treatment A vs. B | 3.50 (2.02–6.06) | 2.15 (1.18–3.92) | 0.012 |
- Crude (unadjusted) OR comes from a bivariate model — the predictor alone against the outcome, with no other variables in the model. It can absorb the effect of confounders that happen to correlate with the predictor.
- Adjusted OR comes from a multivariable model that holds other specified covariates constant. It is the number to report as “independent of [the covariates actually in the model]” — never as “independent of confounding” in general, since a model only adjusts for what it explicitly includes.
In the table above, the drop from 3.50 crude to 2.15 adjusted indicates that part of the crude association was attributable to the covariates added to the model, not to the treatment itself. The sentence a results section would use: “After adjustment for [named covariates], Treatment A was associated with 2.15 times the odds of the outcome compared with Treatment B (adjusted OR 2.15, 95% CI 1.18–3.92, p = 0.012).” Report both numbers when both are available, and always state which covariates the adjusted model actually includes — “adjusted” with no covariate list named is not a complete result.
How to Report an Odds Ratio
A complete odds ratio statement needs four elements: the OR itself, its 95% confidence interval, what the reference category is, and (for a multivariable model) which covariates it’s adjusted for. Template: “[Group/exposure] had [OR value] times the odds of [outcome] compared with [reference group] (OR [value], 95% CI [lower]–[upper][, adjusted for covariates]).” Avoid “times more likely,” “times as likely,” or “% more likely to” unless the rare-disease assumption genuinely applies and you say so — when it doesn’t, those phrasings misstate an odds ratio as a risk ratio, exactly as shown in the worked example above.
Frequently Asked Questions
Is an odds ratio the same as relative risk?
No. Relative risk (risk ratio) compares probabilities directly; an odds ratio compares odds. They converge only when the outcome is rare in both groups being compared (the rare-disease assumption); when the outcome is common, as in the worked example above, the OR is always further from 1.0 than the true RR in the same direction — it overstates the effect if read as a risk ratio.
What does an odds ratio of exactly 1 mean?
No association — the odds of the outcome are identical between the group being compared and the reference group. Whether an OR near but not exactly 1 is distinguishable from “no association” is a question for the confidence interval, not the point estimate alone.
Can an odds ratio be negative?
No. An odds ratio is a ratio of two positive quantities and is always greater than 0. It ranges from 0 (the outcome never occurs in the comparison group when it does occur in the reference group) through 1 (no association) to unbounded positive values. What can be negative is the log-odds coefficient a logistic regression model actually estimates before it’s exponentiated into the odds ratio you read in results tables.
Why do case-control studies report odds ratios instead of relative risk?
Because a case-control design selects participants by outcome status (a fixed number of cases and controls, chosen by the researcher) rather than following an exposed and unexposed population forward in time, true incidence — and therefore a valid risk ratio — cannot be calculated from the sample. The odds ratio is the measure of association the design can legitimately produce, and under the rare-disease assumption it approximates the risk ratio a cohort study would have shown.
Is a bigger sample size enough to trust a wide odds-ratio confidence interval?
No — interval width reflects the precision of the specific estimate (driven by the number of outcome events, not just total sample size), and a wide interval that still excludes 1.0 can be statistically significant while remaining too imprecise to be practically useful. Report the interval alongside the point estimate rather than the point estimate alone, and treat a CI that spans 1.0 as “not distinguishable from no association” regardless of how far the point estimate sits from 1.0.
