Modelling Binary Outcomes: From Odds Ratios to Clinical Probabilities

Logistic regression is one of the most widely used modeling methods in clinical and epidemiological research when the outcome has two categories. It can examine outcomes such as disease present versus absent, response versus no response, or complication versus no complication. Yet reading coefficients from software output is not enough for a defensible interpretation. Variable coding, the meaning of odds ratios, the precision of estimates, calibration, discrimination, and validation should be considered together.

What does logistic regression estimate?

Unlike linear regression, logistic regression does not directly model a continuous outcome. It models the logit of the event probability. In practice, the exponentiated regression coefficient is reported as an odds ratio (OR). Odds are the probability that an event occurs divided by the probability that it does not occur. An odds ratio is therefore not identical to a risk ratio. When an outcome is common, the OR may appear farther from 1 than the corresponding risk ratio.

Define the event and reference categories first

The direction of every interpretation depends on which outcome category is coded as the event. If complication is coded 1 and no complication is coded 0, the model describes the odds of complication. Reversing the coding reverses the direction of the coefficients. Reference categories must also be stated for categorical predictors. A statement such as “OR=0.60 for treatment B” is incomplete unless readers know that treatment A is the reference group.

How should an odds ratio be interpreted?

An OR of 1 indicates no association within the fitted model. Values above 1 indicate higher event odds, while values below 1 indicate lower event odds. If the OR for age is 1.04, each one-year increase in age is associated with approximately 4% higher event odds, holding the other variables constant. If age is entered per 10 years, the same interpretation applies to a 10-year increase. The unit and scaling of every continuous predictor should therefore be reported.

An OR of 0.70 indicates 30% lower event odds relative to the reference category. It should not automatically be written as a 30% reduction in risk. When the outcome is frequent, predicted probabilities, absolute risk differences, or marginal effects may communicate clinical impact more clearly than the OR alone.

Interpret the 95% confidence interval with the p-value

A 95% confidence interval conveys the uncertainty of the estimate. If the interval for an OR includes 1, the result is not statistically significant at the conventional 5% level. For example, OR=1.80 with a 95% CI of 1.20–2.70 suggests higher odds with reasonably good precision. OR=2.10 with a 95% CI of 0.85–5.20 looks large, but the data remain compatible with no association and with a very strong association.

Wide intervals may result from a small sample, few events, sparse categories, or strong relationships among predictors. “Not statistically significant” does not prove that there is no effect. It may indicate that the available data are too imprecise to distinguish a clinically important effect from little or no effect.

Crude versus adjusted odds ratios

A crude OR comes from a model containing the predictor of interest without adjustment for other covariates. An adjusted OR describes the association while holding the other variables in the model constant. A substantial difference between crude and adjusted estimates may reflect confounding, mediation, interaction, or correlation among predictors. Adjustment does not by itself establish causality. Causal interpretation also depends on study design and a scientifically justified covariate strategy.

Check the functional form of continuous predictors

Standard logistic regression assumes that a continuous predictor has a linear relationship with the log-odds of the outcome. The effects of age, body mass index, or a laboratory value may not be constant over their full range. Categorizing continuous variables merely for convenience discards information and can create artificial thresholds. Graphical assessment, the Box–Tidwell approach, or flexible functions such as restricted cubic splines can be used to investigate nonlinearity.

When are interaction terms needed?

An interaction exists when the association of one variable with the outcome differs according to another variable. A treatment effect, for example, may vary by age or baseline disease severity. If an interaction is present, isolated main-effect coefficients can be misleading. Stratum-specific estimates or predicted probabilities should be presented. Interaction analyses are most credible when prespecified and supported by an adequate number of observations and events.

Model performance is not a single goodness-of-fit test

Calibration

Calibration assesses agreement between predicted probabilities and observed event frequencies. A calibration plot, calibration intercept, and calibration slope provide useful information. The Hosmer–Lemeshow test is sensitive to sample size and grouping choices. A p-value above 0.05 does not prove perfect fit, and a p-value below 0.05 does not by itself make a model unusable.

Discrimination and AUC

The area under the ROC curve (AUC), also called the c-statistic, measures how well the model ranks participants with and without the event. An AUC of 0.50 represents chance-level ranking, whereas 1.00 represents perfect separation. A high AUC does not guarantee good calibration. Sensitivity, specificity, predictive values, and the consequences of a chosen decision threshold should also be evaluated when the model will support clinical decisions.

Pseudo-R² and overall fit

Pseudo-R² measures such as Cox–Snell and Nagelkerke R² should not be interpreted like the R² from linear regression. They can help compare models and summarize aspects of fit, but there is no universal cutoff that defines a good clinical model. Likelihood-ratio tests, AIC, and BIC may also be useful when interpreted for the stated objective and the models being compared.

Variable selection and the risk of overfitting

Selecting predictors only because they have small univariable p-values can omit important confounders and produce unstable results. Core covariates should be chosen using clinical knowledge, prior evidence, and the research question. Automated stepwise procedures can yield optimistic estimates and select different variables in slightly different samples. When the number of events is limited, penalized regression, data reduction, or a smaller prespecified model may be more appropriate.

Missing data and internal validation

Complete-case analysis can waste information and introduce bias depending on why observations are missing. Researchers should report the amount and pattern of missingness and consider multiple imputation when its assumptions are reasonable. Performance measured only in the development data is typically optimistic. Bootstrap validation or cross-validation can provide a more realistic estimate of overfitting and expected performance in new participants.

How to report logistic regression results

The methods section should describe outcome coding, reference categories, predictor scaling, the scientific basis for covariate selection, assessment of nonlinear effects and interactions, the missing-data strategy, and the validation method. Results should provide the coefficient or OR, 95% confidence interval, and p-value for each predictor. Calibration and discrimination should be reported as separate properties.

Example: “After adjustment for age, sex, and baseline disease severity, the treatment group had lower odds of complications than the control group (adjusted OR=0.62; 95% CI: 0.41–0.94; p=0.024). The model had an AUC of 0.77, with a bootstrap-corrected AUC of 0.75. The calibration plot showed no major systematic deviation.”

Common interpretation errors

• Describing an OR as if it were a risk ratio or an absolute probability difference.

• Failing to report the event definition, reference category, or measurement unit.

• Focusing on the p-value while ignoring the confidence interval.

• Treating the Hosmer–Lemeshow test as the sole measure of model quality.

• Equating high AUC with clinical usefulness.

• Generalizing apparent performance from development data to new populations.

• Replacing scientific reasoning with data-driven variable selection.

Conclusion

Reliable logistic regression interpretation requires more than reading an OR and a p-value. Coding, uncertainty, functional form, confounding, interaction, calibration, discrimination, and validation all shape the final conclusion. Considering these elements together produces results that are more scientifically transparent and clinically useful. Biostatist provides biostatistical consulting for model development, SPSS or R output interpretation, validation, and publication-ready reporting.

Frequently Asked Questions

Are an odds ratio and a risk ratio the same?

No. They can be similar for rare outcomes, but the OR generally moves farther from 1 than the risk ratio as the outcome becomes more common.

What does it mean when the confidence interval includes 1?

The estimate is not statistically significant at the conventional 5% level. The range of clinically plausible effects should still be examined.

Does a high AUC mean the model is good?

Not necessarily. AUC measures discrimination. Calibration, clinical utility, and performance in independent data must be evaluated separately.