Logistic Coefficients as Odds Ratios

Why Logistic Coefficients Need Translation

A logistic-regression coefficient is measured on the log-odds scale. Its sign indicates the direction of association, but the coefficient itself is not a percentage change in odds. Exponentiating the coefficient converts it to an odds ratio, a multiplicative measure that is often easier to interpret.

This interpretation is useful when communicating a fitted model: instead of saying that a predictor changes the log-odds by a certain amount, we can state the factor by which the odds are multiplied for a clearly specified change in that predictor. The interpretation must still be tied to the predictor's units and the model's other terms.

Definition: In a logistic-regression model for a binary response \(Y\), let \(p(\mathbf{x})=P(Y=1\mid\mathbf{x})\), where \(\mathbf{x}\) is the predictor vector. The odds of \(Y=1\) at \(\mathbf{x}\) are \(p(\mathbf{x})/[1-p(\mathbf{x})]\). If the model has linear predictor \(\eta(\mathbf{x})=\beta_0+\sum_j\beta_jx_j\), then \(\log\{p(\mathbf{x})/[1-p(\mathbf{x})]\}=\eta(\mathbf{x})\). For a change in predictors from \(\mathbf{x}\) to \(\mathbf{x}'\), the odds ratio is the odds at \(\mathbf{x}'\) divided by the odds at \(\mathbf{x}\).

From a Coefficient to an Odds Ratio

For a logistic model with no interaction involving \(x_j\), increase \(x_j\) by one unit while holding all other predictors fixed. The linear predictor changes by \(\beta_j\). Since the logit is the logarithm of the odds, exponentiating this change gives the ratio of the new odds to the original odds.

\begin{aligned} \eta(\mathbf{x}')-\eta(\mathbf{x})&=\beta_j,\\ \log\!\left(\frac{\operatorname{odds}(\mathbf{x}')}{\operatorname{odds}(\mathbf{x})}\right)&=\beta_j,\\ \frac{\operatorname{odds}(\mathbf{x}')}{\operatorname{odds}(\mathbf{x})}&=\exp(\beta_j). \end{aligned}

This is the key justification: the coefficient is a difference in log-odds, and exponentiation turns a difference of logarithms into a ratio of odds. For an estimated coefficient \(\hat{\beta}_j\), the estimated odds ratio is \(\exp(\hat{\beta}_j)\).

Result: In a logistic model whose linear predictor is additive in \(x_j\), a one-unit increase in \(x_j\), with other predictors fixed, multiplies the odds of \(Y=1\) by \(\exp(\beta_j)\). A change of \(k\) units multiplies the odds by \(\exp(k\beta_j)\). For a binary predictor coded 0 or 1, changing it from 0 to 1 multiplies the odds by \(\exp(\beta_j)\). These formulas do not apply unchanged when the effect varies with another predictor through an interaction term.

An odds ratio greater than 1 means the odds are higher for the specified change; an odds ratio between 0 and 1 means they are lower. The proportional change in odds is \((\text{odds ratio}-1)\times100\%\). Thus, an odds ratio of 1.20 means 20% higher odds, while an odds ratio of 0.70 means 30% lower odds. Neither statement is a percentage-point change in probability.

Worked Examples

Worked Example: A Multi-Unit Change in a Continuous Predictor

A hypothetical insurer models whether a policyholder renews, with \(Y=1\) for renewal and \(Y=0\) otherwise. The fitted model is Bernoulli conditional on the predictors, with \(P(Y=1\mid\mathbf{x})=\operatorname{expit}(\eta)\), where \(\operatorname{expit}(\eta)=1/(1+\exp(-\eta))\). The linear predictor includes months with the insurer, \(x\), with estimated coefficient \(\hat{\beta}_x=0.08\). Assume the model contains no interaction involving \(x\). What is the estimated odds multiplier for a six-month increase, holding other predictors fixed, and what is its proportional change in odds?

Method. For a \(k\)-unit increase in a continuous predictor with an additive effect, use \(\exp(k\hat{\beta}_x)\). Here \(k=6\) months, so the change in the linear predictor is \(6(0.08)=0.48\).

The odds multiplier is \(\exp(0.48)=1.6161\), to four decimal places. The proportional change is \((1.6161-1)\times100\%=61.61\%\), to two decimal places. Therefore, the model estimates 61.61% higher renewal odds for a six-month increase, with the other predictors held fixed.

Check. A one-month increase has multiplier \(\exp(0.08)=1.0833\). Applying the unrounded one-month multiplier six times gives \(\exp(0.08)^6=\exp(0.48)=1.6161\), as required. Interest and mortality assumptions are not relevant to this logistic-regression example; the response distribution is Bernoulli conditional on the predictors.

Common mistake avoided: Treating \(0.08\) as an 8% increase in odds, or applying the one-month effect only once for a six-month change. The coefficient is a log-odds change per month; the six-month odds ratio is \(\exp(6\times0.08)\).

Worked Example: A Binary Predictor Coded 0 and 1

In a hypothetical study of application completion, let \(Y=1\) indicate completion and \(Y=0\) indicate non-completion. The response is modeled as Bernoulli conditional on the predictors using a logistic link. A binary predictor \(z\) equals 1 if an applicant received a reminder and 0 otherwise. Its estimated coefficient is \(\hat{\beta}_z=-0.40\). Suppose the other predictors are held fixed and there is no interaction involving \(z\). Find the estimated odds ratio for receiving a reminder rather than not receiving one. If the model's odds at \(z=0\), for those fixed predictor values, are 0.75, calculate the corresponding odds at \(z=1\).

Method. The change from \(z=0\) to \(z=1\) changes the linear predictor by \(-0.40(1-0)=-0.40\). The odds ratio is therefore \(\exp(-0.40)=0.6703\), to four decimal places.

The proportional change in odds is \((0.6703-1)\times100\%=-32.97\%\), to two decimal places. This means 32.97% lower modeled odds for \(z=1\) than for \(z=0\), conditional on the other predictors. To calculate the odds for \(z=1\), multiply the \(z=0\) odds by the unrounded odds ratio:

0.75\exp(-0.40)=0.75(0.670320\ldots)=0.502740\ldots\approx0.5027.

Thus, the odds at \(z=1\) are approximately 0.5027. The ratio check is \(0.502740\ldots/0.75=0.670320\ldots\), which agrees with \(\exp(-0.40)\). The example uses a Bernoulli conditional response model; interest and mortality assumptions do not apply.

Common mistake avoided: Reporting a negative odds ratio or saying that the odds fall by 40%. The coefficient is negative, but the odds ratio must be positive because it is an exponential. The fitted change is about 32.97% lower odds, not 40% lower odds.

Worked Example: Combining Changes and Checking Rounding

A hypothetical maintenance provider models whether a customer accepts a service plan (\(Y=1\) for acceptance). The response is Bernoulli conditional on the predictors with a logistic link. The additive linear predictor includes a continuous predictor \(x\), measured in years of prior service, with estimated coefficient \(0.06\), and a binary predictor \(z\), coded 1 for online contact and 0 for in-person contact, with estimated coefficient \(-0.40\). There is no interaction between these predictors. Find the odds ratio for increasing \(x\) by one year while changing \(z\) from 0 to 1, with all other predictors fixed. If the original odds are 2.00, also find the new odds.

Method. The total change in the linear predictor is the sum of the two coefficient contributions:

\Delta\eta=(0.06)(1)+(-0.40)(1-0)=-0.34.

The combined odds ratio is \(\exp(-0.34)=0.711770\ldots\), which rounds to 0.7118 to four decimal places. The estimated odds are therefore 28.82% lower for the specified combined change, since \((0.711770\ldots-1)\times100\%=-28.8229\ldots\%\).

The resulting odds are \(2.00\exp(-0.34)=1.423540\ldots\), or 1.4235 to four decimal places. As a check, this is less than the original odds of 2.00, consistent with a combined odds ratio below 1.

The same result can be checked by multiplying the separate, unrounded multipliers: \(\exp(0.06)\exp(-0.40)=\exp(-0.34)\). The displayed rounded factors are \(\exp(0.06)\approx1.0618\) and \(\exp(-0.40)\approx0.6703\). Multiplying those displayed values gives \(1.0618\times0.6703=0.71172454\), which rounds to 0.7117. Multiplying the unrounded factors gives \(\exp(-0.34)\approx0.7118\). The one-unit difference in the last displayed decimal is due to rounding the separate factors before multiplying.

Common mistake avoided: Adding the coefficients to obtain the change in the linear predictor is correct, but the odds ratio is the exponential of that sum, not the sum of separate odds ratios. Also, multiplying already-rounded factors can produce a different last decimal than exponentiating the combined unrounded coefficient.

Interpreting Size, Direction, and Scope

The meaning of an odds ratio depends on the change being compared. For a continuous predictor, state the unit explicitly: an odds ratio per dollar is different in scale from one per thousand dollars. If a predictor is measured in units that make a one-unit change uninformative, choose a meaningful change \(k\) and calculate \(\exp(k\hat{\beta})\). The model's linearity in that predictor on the log-odds scale is what permits this multiplication by \(k\).

For a binary predictor coded 0 and 1, the coefficient describes the 1-versus-0 comparison, conditional on the other included predictors. Reversing the comparison reverses the odds ratio: the odds ratio for 0 versus 1 is \(\exp(-\hat{\beta})=1/\exp(\hat{\beta})\). This reciprocal relationship is a useful check when interpreting the direction of a contrast.

The interpretation is conditional on the model specification. In an additive model with no interaction for the predictor, the odds ratio for a given predictor change is the same at fixed values of the other predictors. If an interaction is present, the coefficient alone may not describe the odds ratio for a change: the relevant change in the full linear predictor must include all terms that change. The interpretation also describes a model-based association, not a causal effect merely because the coefficient has been exponentiated.

Identify the comparison.
Specify the starting and ending predictor values, including units and the direction of any binary change.

Calculate the linear-predictor change.
For an additive term, a \(k\)-unit change contributes \(k\hat{\beta}\); for a 0-to-1 change, it contributes \(\hat{\beta}\). Include every changing term if the model has interactions.

Exponentiate and interpret.
Exponentiate the total change to obtain the odds ratio. State whether the odds are higher or lower and give the proportional change in odds if useful.

Check the result.
An exponential is positive; a positive linear-predictor change gives an odds ratio above 1, and a negative change gives an odds ratio below 1.

Common Exam Traps

Calling a coefficient a percent change. If \(\hat{\beta}=0.20\), the odds ratio is \(\exp(0.20)=1.2214\), not 1.20. The corresponding odds increase is about 22.14%, not 20%.
Confusing odds with probability. An odds ratio compares \(p/(1-p)\), not \(p\). It does not directly give a probability difference or a percentage-point change.
Ignoring predictor units. A coefficient per one unit cannot be interpreted as the effect of ten units without scaling the coefficient contribution by ten before exponentiating.
Getting the binary comparison backward. A coefficient's usual 0-to-1 interpretation reverses when the question asks for 0 versus 1; use the reciprocal odds ratio.
Rounding too early. Keep full precision through the calculation and round the final odds ratio or odds. Rounded component multipliers may not reproduce the rounded combined result.

Key takeaway: A logistic coefficient is a change in log-odds. For a specified predictor change, calculate the full change in the linear predictor and exponentiate it to obtain the odds ratio. Interpret that ratio as a multiplier of odds, not as a direct change in probability.

Check Your Understanding

For each problem, assume a Bernoulli response conditional on the predictors with a logistic link. Unless stated otherwise, the model is additive for the predictors in the question and all other predictors are held fixed. Interest and mortality assumptions are not relevant.

A continuous predictor has estimated coefficient \(0.12\) per unit. Find the odds ratio for a two-unit increase and the percentage change in odds. (Answer: \(\exp(0.24)=1.2712\), or 27.12% higher odds.)
A binary predictor coded 0 or 1 has estimated coefficient \(-0.70\). Find the odds ratio for changing from 0 to 1 and state the direction and percentage change in odds. (Answer: \(\exp(-0.70)=0.4966\), or 50.34% lower odds.)
A continuous predictor has estimated coefficient \(0.035\) per unit. Find the odds ratio for a four-unit increase, rounding to four decimal places, and interpret it. (Answer: \(\exp(0.14)=1.1503\), or 15.03% higher odds.)
A continuous predictor \(x\) has coefficient \(0.20\), and a binary predictor \(z\), coded 0 or 1, has coefficient \(-0.50\). Find the combined odds ratio for increasing \(x\) by two units and changing \(z\) from 0 to 1. (Answer: \(\exp(2(0.20)-0.50)=\exp(-0.10)=0.9048\), or 9.52% lower odds.)
A binary predictor has coefficient \(0.30\). Find the odds ratio for the reverse comparison, changing from 1 to 0, and state the percentage change in odds. (Answer: \(\exp(-0.30)=0.7408\), or 25.92% lower odds.)

What You'll Learn