Logistic Regression: Interpreting and Reporting Odds Ratios

When your outcome has two categories — pass/fail, diseased/healthy, purchased/did not — the right tool is logistic regression. Ordinary least squares (OLS) breaks down here: it produces 'probability' predictions that spill outside the 0–1 range, its errors cannot be normally distributed, and the variance is not constant. Yet the odds ratio (OR) that logistic regression produces is one of the most frequently misread statistics in theses. This guide builds the full interpretive chain, from the odds–probability distinction to an APA reporting template.

Probability and odds are not the same thing

Here lies the classic confusion. Probability is the chance an event occurs; odds are the probability of occurrence divided by the probability of non-occurrence: odds = P / (1 − P). Worked example: if a student's probability of passing an exam is 0.80, the odds of passing are 0.80 / 0.20 = 4 (four to one). If the probability is 0.20, the odds are 0.20 / 0.80 = 0.25. This is why the sentence 'OR = 2 means the probability doubles' is wrong: it is the odds that double, and the change in probability depends on the baseline, with the gap widening as P grows. Logistic regression fits the linear model not to the probability but to the natural logarithm of the odds (the logit link function); exponentiating the coefficients (exp(B)) yields the odds ratios.

Reading odds ratios in logistic regression output

For continuous predictors, the OR tells you by what factor the odds change for a one-unit increase: OR = 1.04 for age means each additional year raises the odds of the event by 4 per cent — and the unit (one year or ten?) must always be stated in the interpretation. For categorical predictors, the OR compares a category with the reference category: with 'male' as reference, OR = 1.85 for women means women's odds are 85 per cent higher than men's — the number is uninterpretable unless the reference category is named. OR > 1 indicates a relationship that raises the odds, OR < 1 one that lowers them; OR = 0.50 is a 50 per cent reduction in odds. A handy conversion for comparing directions: the inverse of 0.50 is 1 / 0.50 = 2.00. To pin the definition down: an odds ratio is simply one group's odds divided by another's — if women's odds of passing are 4 and men's are 2, the OR is 2.00.

If the confidence interval crosses 1, the effect is not statistically significant: an OR of 1.27 with 95% CI [0.92, 1.76] is not evidence of a relationship, however 'close' the p value may look. Conversely, an interval lying entirely above or entirely below 1 indicates a significant effect. The chart below shows how a typical coefficient table is visualised around the OR = 1 reference line.

Evaluating the model: fit, explanation, discrimination

There is no single 'model fit' index in logistic regression; a good report answers three questions at once: does the model fit the data, how much do the predictors explain, and how well does the model discriminate between the groups?

• −2 Log-Likelihood and the omnibus chi-square: tests whether the model improves significantly on the intercept-only model; this is the opening sentence of the results.
• Hosmer–Lemeshow test: p > 0.05 means fit is not rejected; but the test is sensitive to the number of groups and flags trivial departures in large samples. Do not rely on it as the sole evidence of fit — support it with a calibration assessment.
• Nagelkerke pseudo-R-squared: not the same animal as the OLS R-squared; present it as a rough summary of relative explanatory strength, never as 'variance explained'.
• Classification accuracy and ROC/AUC: the accuracy percentage depends on the 0.50 cut-off and the group base rates; AUC measures discrimination independently of any cut-off. Rough guide: 0.70–0.80 acceptable, 0.80–0.90 good discrimination.

The sample size rule and the model's limits

The key to sample planning in logistic regression is the events-per-variable (EPV) rule: aim for at least 10–20 events from the less frequent outcome category per predictor. In a sample of 300 with only 45 events, the model can safely carry no more than about 2–4 predictors; as EPV falls, the coefficients inflate and the confidence intervals become unreliable. Beyond EPV, two technical checks remain: linearity of continuous predictors with the logit, and multicollinearity (VIF). And if a predictor classifies the outcome perfectly (complete separation), the coefficients run off to infinity — consider collapsing categories or a penalised estimation approach. When the outcome has more than two unordered categories, multinomial logistic regression applies; with ordered categories, ordinal logistic regression uses the same interpretive logic. For formatting the write-up itself, our APA 7 reporting guide covers the conventions in detail.

An APA-style reporting template

Report the model first, then the coefficients: "The model was statistically significant, chi-square(3, N = 412) = 48.67, p < .001; Nagelkerke R-squared = .18. Smoking significantly increased the odds of the outcome, B = 0.96, SE = 0.21, Wald = 20.90, p < .001, OR = 2.60, 95% CI [1.72, 3.93]." The table should give B, standard error, Wald, p, OR and the 95% confidence interval for every predictor, with reference categories and coding stated beneath it. The cut-off, the classification table and the AUC can each be summarised in a single sentence of text.

An odds ratio speaks about odds, not probability; a report that confuses the two loses the reader's trust in its first sentence.

Frequently Asked Questions