Why “Rare” Matters When Interpreting an Odds Ratio
A logistic regression coefficient is often reported as an odds ratio. When outcomes are uncommon, that odds ratio may be close to a risk ratio, which can make interpretation more intuitive. But “uncommon” is not a universal cutoff: the accuracy also depends on the size and direction of the odds ratio. If the outcome is not rare in the comparison group, calling an odds ratio a risk ratio can substantially misstate the effect.
The distinction matters in practical settings such as a hypothetical insurer comparing the probability of a claim across two groups. If the estimated odds ratio is 2, it does not automatically mean that one group has twice the probability of a claim. The baseline probability is needed to determine the comparison probability and the risk ratio.
The Exact Relationship
Suppose the odds ratio \(\theta\) compares two groups at the same values of any other predictors in the model. Assume \(0<p_0<1\), and that the odds ratio is positive. Starting with the definition of the odds ratio, the comparison-group odds are the baseline odds multiplied by \(\theta\). Converting those odds back to a probability gives the exact formula below.
The risk ratio follows by dividing \(p_1\) by \(p_0\). The resulting expression shows exactly when it is close to the odds ratio.
Why the Approximation Works
The relationship can be justified directly from the definition. The comparison odds are \(\theta p_0/(1-p_0)\). A probability is odds divided by one plus odds, so substituting the comparison odds gives \(p_1\). Dividing by \(p_0\) then produces the exact risk-ratio formula. The approximation comes from its denominator: if \((\theta-1)p_0\) is close to zero, then \(1+(\theta-1)p_0\) is close to 1, and the risk ratio is close to \(\theta\).
For a harmful association with \(\theta>1\), the denominator \(1+(\theta-1)p_0\) exceeds 1. Therefore the risk ratio is less than the odds ratio: treating the odds ratio as the risk ratio overstates the relative risk. For a protective association with \(0<\theta<1\), the denominator is less than 1, so the risk ratio exceeds the odds ratio. In that case, equating them understates the risk ratio.
In logistic regression, a coefficient \(\beta\) for a one-unit predictor change corresponds to an odds ratio \(\theta=\exp(\beta)\), holding other predictors fixed. The formula above then compares the fitted probabilities at those same predictor values. The resulting probability ratio is conditional on those values; do not silently substitute a group-average baseline risk from a different population.
Record the baseline risk \(p_0\) and the odds ratio \(\theta\), noting which group is the reference.
Calculate \(p_1=\theta p_0/(1-p_0+\theta p_0)\), rather than multiplying \(p_0\) directly by the odds ratio.
Calculate \(p_1/p_0\) and assess whether \(|(\theta-1)p_0|\) is small enough for the approximation being used.
Worked Example: A Moderate Odds Ratio With a Rare Outcome
In a hypothetical portfolio, the annual probability of a specified claim is 1.5% for the reference group. A logistic model estimates an odds ratio of 2.4 for another group, at the same values of the other predictors. Find the comparison-group probability and risk ratio, and assess the rare-outcome approximation. Round probabilities to four decimal places and ratios to three decimal places.
Method. Use the exact odds-to-probability conversion, then compare the exact risk ratio with the odds ratio.
The baseline risk is \(p_0=0.015\), and \(\theta=2.4\). The numerator of the comparison-risk formula is \(2.4(0.015)=0.036\). Its denominator is \(1-0.015+2.4(0.015)=0.985+0.036=1.021\). Thus \(p_1=0.036/1.021=0.0352595\), or 0.0353 after rounding.
The exact risk ratio is \(0.0352595/0.015=2.350636\), or 2.351. As a check, the direct formula gives \(2.4/\{1+(2.4-1)(0.015)\}=2.4/1.021=2.350636\). The odds ratio of 2.4 is a reasonable approximation to the risk ratio of 2.351 here, but is not exactly equal to it.
Check and common mistake avoided. Multiplying the baseline risk by 2.4 gives 0.036, not the exact comparison risk of 0.0353. That shortcut applies an odds multiplier as if it were a probability multiplier. The exact conversion accounts for the fact that odds use \(1-p\) in their denominator.
Worked Example: How a Nonrare Outcome Changes the Interpretation
A hypothetical service plan has a 18% annual probability of a specified event in its reference group. The estimated odds ratio for a second group is 2.5. Calculate the second group’s probability and risk ratio, then quantify the error from interpreting 2.5 as the risk ratio. Round probabilities to four decimal places and ratios to three decimal places.
Method. Convert the odds ratio to a probability before calculating the risk ratio. Measure the relative overstatement by comparing the naive value with the exact risk ratio.
Here \(p_0=0.18\) and \(\theta=2.5\). The numerator is \(2.5(0.18)=0.45\), and the denominator is \(1-0.18+0.45=0.82+0.45=1.27\). Therefore \(p_1=0.45/1.27=0.3543307\), or 0.3543.
The exact risk ratio is \(0.3543307/0.18=1.9685039\), or 1.969. Equivalently, it is \(2.5/\{1+(2.5-1)(0.18)\}=2.5/1.27=1.9685039\). The naive risk-ratio interpretation is 2.5, which is higher than the exact value by \(2.5-1.9685039=0.5314961\). Relative to the exact risk ratio, the overstatement is \(0.5314961/1.9685039=0.27\), or 27%.
Check and common mistake avoided. Multiplying 18% by 2.5 would claim a comparison probability of 45%. The exact probability is about 35.43%. Since the outcome is not rare in the comparison group, treating the odds ratio as a risk ratio materially exaggerates the relative difference.
Worked Example: A Protective Odds Ratio
For a hypothetical coverage feature, the baseline probability of a specified loss is 8%. The feature has an odds ratio of 0.4 relative to the reference group. Calculate the probability and risk ratio for the feature group, and state the direction of the approximation error. Round probabilities to four decimal places and ratios to three decimal places.
Method. Apply the same exact conversion for a protective association; do not assume that an odds ratio below 1 is itself the risk ratio.
Set \(p_0=0.08\) and \(\theta=0.4\). The numerator is \(0.4(0.08)=0.032\), while the denominator is \(1-0.08+0.032=0.92+0.032=0.952\). Thus \(p_1=0.032/0.952=0.0336134\), or 0.0336.
The risk ratio is \(0.0336134/0.08=0.420168\), or 0.420. The direct calculation agrees: \(0.4/\{1+(0.4-1)(0.08)\}=0.4/(1-0.048)=0.4/0.952=0.420168\). The odds ratio is smaller than the risk ratio, so calling 0.4 the risk ratio would overstate the protective reduction in relative risk.
Check and common mistake avoided. Multiplying 8% by 0.4 gives 3.2%, not the exact 3.36%. The direction of the error differs from the harmful-outcome example: for a protective odds ratio, the odds ratio is below the risk ratio.
Common Interpretation Traps
This distinction is aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus. The following checks help keep an interpretation tied to the quantity the model actually estimates.
- Calling an odds ratio a probability multiplier. An odds ratio multiplies \(p/(1-p)\), not \(p\). Use the exact conversion to obtain a probability.
- Assuming “rare” has one fixed cutoff. Approximation accuracy depends on \(|(\theta-1)p_0|\), as well as on the baseline risk. A very large odds ratio can make the comparison risk nonrare even when \(p_0\) is small.
- Ignoring which group is the reference. Reversing the comparison changes the odds ratio to its reciprocal and changes which risk is the baseline.
- Reporting only a relative effect. Even when an odds ratio approximates a risk ratio, the absolute change \(p_1-p_0\) may be important for interpretation.
- Mixing conditional and marginal comparisons. A logistic-model odds ratio for fixed predictor values compares fitted probabilities at those same values. Do not combine it with a baseline risk that represents a different mix of predictor values without specifying the comparison.
Check Your Understanding
For each problem, calculate the requested quantity from the stated baseline risk and odds ratio. Use the exact conversion unless the question specifically asks about approximation.
- A hypothetical group has baseline risk \(p_0=0.03\) and odds ratio \(\theta=1.8\). Find the comparison risk and risk ratio. (Answer: \(p_1=0.0527\), risk ratio \(=1.758\).)
- A feature has baseline risk \(p_0=0.12\) and protective odds ratio \(\theta=0.6\). Find the comparison risk and risk ratio, and state whether the odds ratio is above or below the risk ratio. (Answer: \(p_1=0.0756\), risk ratio \(=0.630\); the odds ratio is below the risk ratio.)
- For baseline risk \(p_0=0.04\) and odds ratio \(\theta=4.5\), calculate the comparison risk and exact risk ratio. (Answer: \(p_1=0.1579\), risk ratio \(=3.947\).)
- A hypothetical event has baseline risk \(p_0=0.02\) and odds ratio \(\theta=6\). Calculate the comparison risk and the percentage by which treating the odds ratio as a risk ratio overstates the exact risk ratio, relative to that exact ratio. (Answer: \(p_1=0.1091\), risk ratio \(=5.455\); overstatement \(=10\%\).)
- For a harmful odds ratio \(\theta=3\), what is the largest baseline risk \(p_0\) for which the odds ratio exceeds the exact risk ratio by no more than 5% of the exact risk ratio? Use the exact formula to determine the condition. (Answer: \(p_0=0.025\), or 2.5%.)