Odds Ratio vs Risk Ratio: Calculate Both and Write the Evidence Accurately

The denominator changes the meaning of the measure

Risk divides people with an event by everyone in the group. Odds divides people with the event by people without it. A risk ratio compares two risks; an odds ratio compares two odds. They can describe the same table while having different numerical values.

Cochrane's effect-measures guidance explicitly warns against interpreting an odds ratio as a risk ratio when events are common. Neither measure is inherently wrong. The error is changing the quantity while keeping the same number—for example, reporting OR 2.67 as a 2.67-fold risk.

This guide uses fictional groups observed for seven days. The event is reporting at least one episode of sleep disruption during that period, counted once per person. Group A contains 40 people with the event out of 100; group B contains 20 out of 100. Assume complete follow-up and independent people for the arithmetic. These are invented observations, not clinical evidence or a causal comparison of treatments.

Calculate risk and odds side by side

For group A, risk is 40/100 = 0.40, or 40%. Its odds are 40/60, approximately 0.667. The 60 is the number without the event, not the total group size. For group B, risk is 20/100 = 0.20, or 20%, and odds are 20/80 = 0.25.

Comparing A with B, RR = 0.40/0.20 = 2.00. The observed risk is twice as high in A. OR = (40/60)/(20/80) = 2.667. The observed odds are about 2.67 times as high in A. Saying 'the risk is 2.67 times as high' would replace the risk ratio of 2.00 with a different measure.

The absolute risk difference is 40% − 20% = 20 percentage points. This adds the baseline context that neither ratio supplies alone. Forty versus twenty people per hundred is often easier to interpret than the odds ratio by itself, provided the study design actually permits risks to be estimated.

Check the direction and the event before using ordinary language

The comparison order matters. Reversing the groups gives RR B/A = 0.5 and OR B/A = 0.375. Those values describe the same data from the opposite direction. Keep the reference group in the sentence so the reader can tell what 'higher' or 'lower' refers to.

The event also matters. For no sleep disruption, the risks are 60% in A and 80% in B, giving a no-event RR of 0.75. That is not the reciprocal of the event RR of 2.00. Reversing the comparison groups and changing the event to its complement are different operations.

For the event odds, OR A/B is 2.667. For the no-event odds, OR A/B is (60/40)/(80/20) = 0.375, its reciprocal. This mathematical property does not make the risks interchangeable; it is one reason the exact event label must travel with the reported measure.

Avoid the unqualified phrase 'more likely' when it conceals whether the paper reports risk, odds, or hazard. If the source reports an adjusted odds ratio, name it as adjusted odds and state the comparison. A plain-language summary can also explain that it is not a direct ratio of probabilities.

See why rare events make the numbers closer

Use a second invented example with 4 events among 1,000 people in A and 2 among 1,000 in B. Risks are 0.4% and 0.2%, so RR is again 2.00. Odds are 4/996 and 2/998, giving OR = (4 × 998)/(996 × 2), approximately 2.004.

The values are close because, when events are uncommon, the number without the event is close to the total. In the first example, 60 differs substantially from 100; in the second, 996 is close to 1,000. The approximation comes from the denominators, not from a rule that odds and risks mean the same thing.

There is no need to relabel a reported odds ratio just because the outcome is uncommon. Keep its name, interval, and analysis context. Whether an approximation is adequate depends on the actual risks and the purpose of the interpretation. Do not treat a fixed rarity cutoff as permission to rewrite every OR as an RR.

The absolute contrast also changes: the second example differs by 2 per 1,000, while the first differs by 200 per 1,000. Both have RR 2.00. A ratio without baseline frequency can therefore conceal a large difference in the number of people affected.

Convert only when the baseline and analysis support it

For a simple unadjusted comparison with a known comparator risk p0, the corresponding intervention risk from an odds ratio is p1 = OR × p0 / (1 − p0 + OR × p0). The corresponding RR is OR / (1 − p0 + OR × p0). Cochrane's interpretation chapter gives the relationship between relative measures and baseline risk.

In the original fictional table, use the exact OR of 8/3 and p0 = 0.20. The numerator for p1 is (8/3) × 0.20 = 0.5333. The denominator is 0.80 + 0.5333 = 1.3333. Thus p1 = 0.40 and RR = 0.40/0.20 = 2.00, recovering the observed risks. Simply multiplying 0.20 by 2.667 would give 0.5334, the wrong probability.

Do not plug an adjusted logistic-regression OR into this formula with an arbitrary overall baseline and label the answer an adjusted RR. An adjusted OR is conditional on the model and covariates; an appropriate absolute-risk or marginal-risk estimate needs the corresponding model-based analysis. If the paper supplies such estimates, use and describe them as reported.

Study design can also prevent the naive risk calculation. In a case-control study, researchers may select the numbers with and without the outcome. The event fraction in that selected sample is not automatically the population risk. An OR may be estimable under the design while absolute risks require additional information. Do not reconstruct a population risk from the case-control row totals as if they were a followed cohort.

Carry the interval and analysis method into the summary

Ratios are estimates, not exact population facts. For the original invented independent groups, a simple unadjusted log-OR calculation gives an approximate 95% interval of 1.42–5.02. The analogous log-RR calculation gives approximately 1.26–3.17. These intervals belong to different measures; swapping the labels does not make them interchangeable.

For reproducibility, SE(log OR) = sqrt(1/40 + 1/60 + 1/20 + 1/80), and the limits are exp(ln(8/3) ± 1.96 × SE). For the RR, SE(log RR) = sqrt(1/40 − 1/100 + 1/20 − 1/100), and the limits are exp(ln 2 ± 1.96 × SE). These are elementary approximations for the stated independent complete-data setup, not default analyses for every clinical study.

An interval excluding 1 does not establish that group membership caused the event difference. Confounding, selection, measurement, and applicability still matter. The fictional observations above contain no randomized treatment assignment and no information that would justify a causal claim.

For a real evidence summary, retain the reported estimate's identity: adjusted or unadjusted, OR or RR, reference group, event definition, follow-up, interval, and analysis population. If those details cannot be determined from the abstract, check the methods and results rather than filling the gaps from the headline.

Use wording that preserves what the paper actually measured

Incorrect: 'Group A was 2.67 times as likely to report sleep disruption.' In this example that wording reads like a probability ratio, although 2.67 is the odds ratio and the risk ratio is 2.00.

A complete teaching statement is: 'In the fictional seven-day observations, 40/100 people in group A and 20/100 in group B reported at least one episode of sleep disruption. The unadjusted RR was 2.00 (approximate 95% CI 1.26–3.17), while the unadjusted OR was 2.67 (1.42–5.02). The observed risks differed by 20 percentage points. These describe an association in the invented groups, not an established causal effect.'

If your real source reports only an adjusted OR, a suitable structure is: 'After adjustment for [reported covariates], the odds of [event] were [OR] times those in [reference group] over [period], with [interval]. This is an odds comparison; the paper [does or does not] provide corresponding absolute risks.' Use actual source details and do not claim a conversion you have not justified.

The final check is arithmetic and language together. Identify the denominator used for each group, retain the reference group, and make sure the noun after the number is the quantity actually estimated. That small discipline prevents a common but consequential error in nursing evidence writing.

• What event was counted, and over what period?
• Does the measure compare risks, odds, rates, or hazards?
• Which group is the reference, and is the estimate adjusted?
• Does the design permit the absolute risks you are about to report?
• Are the interval, baseline context, and causal limits preserved?