How to Interpret Logistic Regression Results

3 min read

Logistic regression models a binary outcome, such as whether a participant completed a program or whether an application was approved. Its coefficients describe changes in log odds. Exponentiating a coefficient produces an odds ratio, which is often easier to report—but also easy to misinterpret.

Before interpreting the output, identify which outcome category the software models, how predictors are coded, and what comparison the model estimates. A polished table cannot compensate for uncertainty about those basics.

Start with the outcome and reference categories

Write down the event coded as 1, the reference category for each categorical predictor, and the units for continuous predictors. If completion is the event, an odds ratio above 1 indicates higher modeled odds of completion for the specified comparison, conditional on the other variables in the model.

If your software instead models noncompletion, that interpretation changes. Reversing the outcome reverses the coefficient signs and reciprocates the odds ratios in the corresponding binary logistic model.

Distinguish probability, odds, and odds ratios

Probability is the chance of the event. Odds equal probability divided by one minus probability. An odds ratio compares two sets of odds; it is not a ratio of probabilities.

Consider a fictional example in which one group has a modeled completion probability of .20 and another has .333, with all other predictors fixed at specified values. Their odds are .20/.80 = .25 and approximately .333/.667 = .50. The odds ratio is approximately 2, but the probability is about 1.67 times as large, with an absolute difference of about 13.3 percentage points.

Calling that result “twice as likely to complete” blurs the distinction. Report “twice the odds” and, when useful, provide the corresponding model-based probabilities and absolute difference.

Explain the unit of change

For a continuous predictor without an interaction or nonlinear term, exp(b) is the odds ratio for a one-unit increase, holding the other predictors fixed. If age is measured in years, an odds ratio of 1.03 applies to one year. For a ten-year contrast under that same linear logit specification, the odds ratio is 1.03 to the tenth power, approximately 1.34.

That calculation does not imply a constant change in probability at every age. The probability change depends on the starting probability and the other predictors. With interactions, splines, or polynomial terms, a single coefficient may no longer describe the contrast you want. Calculate predictions or contrasts from the full model.

Use predicted probabilities with a stated method

Model-based probabilities can make results more accessible, but explain how you obtained them. Predictions for a specified profile—such as a particular age and study site—differ from averaging predictions over the observed sample.

For an adjusted group comparison, one option is to predict each participant’s outcome under each group setting and average the predictions over the same covariate distribution. Identify that target population and include uncertainty for the probabilities and their contrast. Do not present predictions for an artificial “average person” as the observed event rates.

Check the model before writing the conclusion

Review sparse outcome-by-predictor cells and warnings about complete or quasi-complete separation. Huge estimates or standard errors can indicate unstable estimation.
Check the functional form of continuous predictors on the logit scale; logistic regression does not require the raw predictors to be normally distributed.
Account for repeated or clustered observations through an appropriate model or inference method.
Document missing-data handling and consider sensitivity analyses when the conclusions depend on uncertain choices.

For prediction work, evaluate calibration and discrimination using a suitable validation design. A statistically significant coefficient is not evidence that a model predicts well on new data.

Report the uncertainty and the limits

Report the outcome coding, sample size and event count, predictor definitions, adjustment variables, odds ratios with confidence intervals, and the method used for any predicted probabilities. Discuss the magnitude and precision rather than reducing the result to whether p is below .05.

An adjusted odds ratio is conditional on the modeled covariates. Adding covariates can change it even without confounding, a property called noncollapsibility. A change between unadjusted and adjusted odds ratios does not by itself prove confounding. An observational association also does not establish a causal effect.

Use the free analysis-planning worksheet to specify your event, comparison, and reporting scale. If your model contains product terms, read the interaction-effects guide. DASS supports researchers who need help reviewing a model or explaining its results.