Why logistic coefficients are exponentiated

Binary logistic regression models the log odds of an outcome. If p is the modeled probability, odds are p/(1−p), and log odds are ln[p/(1−p)]. A coefficient β is the change in log odds associated with a one-unit increase in its predictor while the other model variables are held fixed. Exponentiating removes the logarithm: exp(β) is the multiplicative change in odds.

For β = 0.405465, exp(β) is approximately 1.5. A one-unit increase is associated with 1.5 times the modeled odds, or 50% higher odds, conditional on the model. For a two-unit change, the log-odds change is 2β and the odds ratio is exp(2β) = 1.5² = 2.25. Multiplying the one-unit percentage by two would be wrong because odds ratios compound multiplicatively.

Worked example with a baseline probability

Assume a fitted model reports β = 0.405465 with standard error 0.15. The one-unit odds ratio is 1.5000. A two-unit contrast produces a log-odds change of 0.8109 and an odds ratio of 2.2500. The coefficient-to-standard-error ratio is about 2.7031, a Wald z statistic under the usual large-sample model.

At a 20% baseline probability, the baseline odds are 0.20/0.80 = 0.25. Multiplying those odds by 2.25 gives 0.5625. Converting odds back to probability gives 0.5625/1.5625 = 0.36, or 36%. The absolute probability change is 16 percentage points and the relative probability change is 80%.

This example shows why an odds ratio is not a probability ratio. Multiplying 20% by 2.25 would give 45%, which does not follow the logistic transformation. The probability impact of the same odds ratio depends on baseline risk. Starting near 1% produces a much smaller percentage-point change than starting near 50%.

Confidence interval calculation

The calculator forms a large-sample Wald interval for β: β ± z × SE. At 95%, z is approximately 1.959964. The default coefficient interval is about 0.1115 to 0.6995. Exponentiating its endpoints gives the one-unit odds-ratio interval, approximately 1.1178 to 2.0128.

For a Δx-unit contrast, both the coefficient effect and its standard error scale by the absolute size of the contrast. The ordered log-odds limits are exponentiated to obtain the contrast OR interval. If Δx is negative, the comparison direction reverses and the interval endpoints are reordered automatically.

Wald intervals may perform poorly with small samples, rare outcomes, separation, very large coefficients, or unstable models. Profile-likelihood or penalized methods can be preferable. Use the interval method reported by the fitted model rather than mixing a coefficient from one procedure with a standard error from another.

Odds, probability, and risk ratio

At low probabilities, odds and probability are numerically close, so an odds ratio can resemble a risk ratio. That approximation weakens as outcomes become common. Report “odds” rather than “risk” unless the model and estimand justify a risk ratio.

An absolute probability change needs a baseline probability and values for the model’s other predictors. The scenario in this calculator holds those covariates fixed implicitly. It does not average over a population distribution and is not a marginal effect unless the modeling workflow defines it that way.

Continuous and categorical predictors

For a continuous predictor, define a meaningful unit. An OR per one dollar, one year, or one biomarker unit may be awkward or misleading. A 10-unit contrast is exp(10β), provided the model is linear in that predictor on the log-odds scale. Centering or rescaling the predictor changes the coefficient’s numerical value but not predictions.

For a binary indicator coded 0 and 1, Δx = 1 compares the group coded 1 with the reference group coded 0. Inverting direction negates β and returns the reciprocal odds ratio. For a multi-level categorical variable, each displayed coefficient compares one level with the chosen reference; changing the reference requires refitting or combining coefficients and their covariance.

Interaction terms make the effect of one predictor depend on another. A single main-effect coefficient may then describe the odds ratio only when the interacting variable equals zero. Polynomial and spline terms likewise require several coefficients for a meaningful contrast. This one-coefficient calculator should not be used to reconstruct those effects without the complete covariance matrix.

Model assumptions and limitations

Logistic regression assumes the specified conditional log-odds structure is appropriate. Relevant issues include independent observations or correctly modeled clusters, a suitable functional form for continuous predictors, adequate sample information, and no perfect separation. Multicollinearity can inflate standard errors even when predictions appear stable.

An adjusted odds ratio is conditional on the variables included in the model. It can differ from an unadjusted two-by-two table odds ratio, and odds ratios are noncollapsible: adjusted and unadjusted values can differ even without conventional confounding. State the covariate set and model specification when reporting.

Association is not causation. Controlling for measured covariates does not automatically remove confounding, and controlling for mediators or colliders can introduce bias. Causal language needs a defensible design and estimand.

Health-use boundary: A probability scenario from a published coefficient is not an individualized diagnosis or treatment recommendation. Clinical prediction requires the full validated model, intercept, covariates, calibration, and target-population evidence.

Outcome coding and reference direction

Software models one category as the event. If “disease present” is coded 1, a positive coefficient raises modeled odds of disease. If the outcome is recoded so “disease absent” is the event, the coefficient changes sign and the OR becomes its reciprocal. Neither coding is inherently correct; the report must name it.

Predictor reference coding behaves similarly. An indicator OR for urban versus rural is the reciprocal of rural versus urban, but its standard error on the log scale stays the same. The calculator’s invert button reverses both β and Δx to demonstrate a logically equivalent contrast; when interpreting an actual model, reverse only the component needed to match the stated comparison.

For a categorical predictor with more than two levels, one coefficient cannot compare every pair. If the model uses A as reference and reports coefficients for B and C, the B-versus-C log odds ratio is βB − βC. Its standard error requires both variances and their covariance. Simply subtracting the two printed standard errors is invalid.

Why adjusted and unadjusted ORs can differ

Confounding is one reason an adjusted odds ratio changes, but it is not the only reason. The odds ratio is noncollapsible: a conditional OR from a correctly specified logistic model can differ from a marginal population OR even when the added covariate is not a confounder. This mathematical property is stronger when outcomes are common.

Do not automatically describe a coefficient as “corrected” because more covariates were added. Adjustment changes the estimand and can improve or harm causal interpretation depending on the variables. Compare models by design and purpose, not by choosing the OR closest to a preferred conclusion.

When copying output, confirm whether the software prints β, exp(β), a standard error, or an already exponentiated interval. Exponentiating an odds ratio a second time produces nonsense. The default β of 0.405465 is the natural logarithm of 1.5; entering 1.5 as β would instead produce OR ≈ 4.48.

Reporting checklist and sanity checks

1. State the outcome coding, especially which value counts as the event.
2. Name the predictor, its coding, unit, and the chosen Δx contrast.
3. Report β or the OR with its interval and the interval method.
4. List adjustment variables and important interactions.
5. If showing probabilities, state the baseline scenario and whether estimates are conditional or averaged.
6. Separate odds, risks, percentage changes, and percentage-point changes.

Sanity checks are simple. β = 0 must give OR = 1. Negating β must produce the reciprocal OR. A zero predictor change must give OR = 1 and leave the scenario probability unchanged. An OR above one increases probability for any baseline strictly between zero and one; an OR below one decreases it.

Frequently asked questions

References

CDC Principles of Epidemiology — measures of association and odds ratios

UCLA Statistical Consulting — interpreting logistic-regression odds ratios

NIST/SEMATECH e-Handbook — binomial proportion context