Comparing Logistic Odds Changes With Probability Changes

Why Odds Changes and Probability Changes Differ

A logistic regression coefficient has a direct interpretation on the odds scale: for a one-unit increase in a predictor, with the other predictors held fixed, the odds are multiplied by a constant. That does not mean the probability is multiplied by that same constant, or that it changes by a fixed number of percentage points. The probability change depends on the risk before the predictor changes.

This distinction matters when translating model output into an actuarial or operational description. An odds ratio can be the same for two hypothetical groups even though the corresponding increases in fitted probability differ. The calculations below focus on that translation and are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.

Definition: In a logistic regression for a binary outcome $Y$, the conditional distribution is $Y\mid \mathbf{x}\sim\operatorname{Bernoulli}(p)$, where $p=P(Y=1\mid\mathbf{x})$. The model specifies the logit of the conditional probability as a linear predictor: $$ \log\left(\frac{p}{1-p}\right)=\eta=\beta_0+\beta_1x_1+\cdots+\beta_kx_k. $$ The odds are $p/(1-p)$, and the fitted probability corresponding to a linear predictor $\eta$ is $p=1/(1+e^{-\eta})$.

For a predictor $x_j$ with coefficient $\beta_j$, increasing $x_j$ by one unit while holding the other predictors fixed increases $\eta$ by $\beta_j$. Exponentiating the change in log-odds gives the odds ratio $\theta=e^{\beta_j}$. This interpretation assumes the model has a constant coefficient for that predictor over the profiles being compared. If the predictor's effect varies with another predictor, a single fixed odds ratio does not describe every such comparison.

From Baseline Risk to Updated Probability

Let $p_0$ be the fitted probability before the one-unit predictor increase, and let $p_1$ be the fitted probability after it. If the odds ratio is $\theta$, first convert $p_0$ to odds, multiply those odds by $\theta$, and convert the result back to a probability. The probability change is $p_1-p_0$, not $\theta p_0-p_0$.

\frac{p_1}{1-p_1} =\theta\frac{p_0}{1-p_0} \qquad\Longrightarrow\qquad p_1=\frac{\theta p_0}{1-p_0+\theta p_0}.

The denominator in the expression for $p_1$ comes from converting odds $O_1$ to probability: $p_1=O_1/(1+O_1)$. Substituting $O_1=\theta p_0/(1-p_0)$ and multiplying numerator and denominator by $1-p_0$ gives the displayed result.

Result: For baseline probability $p_0$ and fixed odds ratio $\theta$, the change in fitted probability is $$ p_1-p_0 =\frac{p_0(1-p_0)(\theta-1)}{1-p_0+\theta p_0}. $$ Thus, when $\theta>1$, the fitted probability increases; when $0<\theta<1$, it decreases. The size of that change depends on $p_0$ as well as on $\theta$.

To verify the change formula, subtract $p_0$ from the expression for $p_1$ using a common denominator:

\begin{aligned} p_1-p_0 &=\frac{\theta p_0-p_0(1-p_0+\theta p_0)} {1-p_0+\theta p_0}\\ &=\frac{\theta p_0-p_0+p_0^2-\theta p_0^2} {1-p_0+\theta p_0}\\ &=\frac{p_0(1-p_0)(\theta-1)} {1-p_0+\theta p_0}. \end{aligned}

The factors $p_0$ and $1-p_0$ help explain why a fixed odds ratio does not imply a fixed probability change. As the baseline probability approaches zero or one, the absolute change approaches zero for a fixed finite $\theta$. Between those extremes, the change can be larger. The exact values should be calculated from the conversion formula rather than inferred from the odds ratio alone.

Worked Example: One Odds Ratio at Two Baseline Risks

In a hypothetical binary-outcome model, $Y\mid\mathbf{x}\sim\operatorname{Bernoulli}(p)$. A one-unit increase in predictor $x$ has coefficient $\ln 2$, so its odds ratio is $e^{\ln 2}=2$. For one profile, the baseline linear predictor is $\eta_0=\ln(1/4)$; for another, it is $\eta_0=\ln 4$. Calculate the fitted probabilities before and after the increase and compare the probability changes. Treat all other predictor contributions as fixed within each profile.

Method. Convert each baseline linear predictor to a baseline probability. Then add $\ln 2$ to the linear predictor, convert again, and subtract the baseline probability. This method avoids treating the odds ratio as a probability multiplier.

First profile. The baseline odds are $e^{\ln(1/4)}=1/4$, so $p_0=(1/4)/(1+1/4)=1/5=0.2000$. After the increase, the linear predictor is $\ln(1/4)+\ln 2=\ln(1/2)$. The updated odds are $1/2$, giving $p_1=(1/2)/(1+1/2)=1/3=0.3333$. The change is $p_1-p_0=1/3-1/5=2/15=0.1333$.

Second profile. The baseline odds are $e^{\ln 4}=4$, so $p_0=4/(1+4)=4/5=0.8000$. After the increase, the linear predictor is $\ln 4+\ln 2=\ln 8$. The updated odds are $8$, giving $p_1=8/(1+8)=8/9=0.8889$. The change is $p_1-p_0=8/9-4/5=4/45=0.0889$.

Check. Both profiles have odds multiplied by exactly 2, but their probability increases are 0.1333 and 0.0889. The second profile starts closer to one, so its probability has less room to increase. Applying the factor 2 directly to either baseline probability would give an incorrect result.

Worked Probability Changes

A useful exam procedure is to keep the odds calculation and the probability calculation separate. The following examples use hypothetical Bernoulli outcomes; the stated probabilities are fitted conditional probabilities, not claims about any real population.

Worked Example: A Fourfold Odds Ratio at Low and Moderate Risk

A hypothetical model has $Y\mid\mathbf{x}\sim\operatorname{Bernoulli}(p)$. A specified predictor change has odds ratio $\theta=4$. Calculate the updated probability and absolute change when the baseline probabilities are $0.1000$ and $0.5000$.

Method. For each baseline probability, compute $O_0=p_0/(1-p_0)$, set $O_1=4O_0$, and convert to $p_1=O_1/(1+O_1)$. This avoids confusing an odds ratio with a risk ratio.

For $p_0=0.1000$, the baseline odds are $0.1/0.9=1/9$. The updated odds are $4/9$, so the updated probability is $(4/9)/(1+4/9)=4/13=0.3077$. The absolute change is $4/13-1/10=(40-13)/130=27/130=0.2077$.

For $p_0=0.5000$, the baseline odds are $1$. The updated odds are $4$, so the updated probability is $4/(1+4)=0.8000$. The absolute change is $0.8000-0.5000=0.3000$.

Check. In both calculations, updated odds divided by baseline odds equal 4. The probability changes differ: 0.2077 versus 0.3000. The incorrect calculation $4(0.1000)=0.4000$ would not have odds four times the baseline odds.

Worked Example: A Protective Odds Ratio at Two Baseline Risks

For a hypothetical outcome, suppose $Y\mid\mathbf{x}\sim\operatorname{Bernoulli}(p)$, and a predictor change has odds ratio $\theta=0.5$. Calculate the updated probabilities and probability reductions from baseline probabilities $0.4000$ and $0.0500$.

Method. Multiply baseline odds by 0.5, then transform the updated odds back to probability. Since the odds ratio is below one, the updated probability should be lower than the baseline in both cases.

At baseline $p_0=0.4000$, the odds are $0.4/0.6=2/3$. The updated odds are $1/3$, so $p_1=(1/3)/(1+1/3)=1/4=0.2500$. The change is $0.2500-0.4000=-0.1500$, a reduction of 0.1500.

At baseline $p_0=0.0500$, the odds are $0.05/0.95=1/19$. The updated odds are $1/38$, so $p_1=(1/38)/(1+1/38)=1/39=0.0256$. The change is $1/39-1/20=-19/780=-0.0244$, rounded to four decimal places.

Check. The updated probabilities are below their baselines, as required by $\theta=0.5$. The same halving of odds produces different absolute reductions: 0.1500 and 0.0244. The incorrect shortcut $0.4000\times0.5=0.2000$ would halve probability, not odds.

Interpretation and Common Exam Traps

An odds ratio describes a multiplicative change in odds, conditional on the other modeled predictors being held fixed. A probability difference describes an absolute change on the zero-to-one probability scale. These are distinct quantities. For example, an absolute change of 0.10 is a ten-percentage-point increase, not necessarily a 10% relative increase.

The reverse calculation is also useful: if a problem supplies $p_0$ and $p_1$, the implied odds ratio is \[ \theta=\frac{p_1/(1-p_1)}{p_0/(1-p_0)}. \] Checking this ratio can confirm whether a proposed pair of fitted probabilities is consistent with the stated logistic effect.

Avoid interpreting an odds ratio as a causal effect without a suitable study design and assumptions. In a fitted logistic model, it is a conditional association represented by the model. Also check how the predictor is scaled: a coefficient's odds ratio applies to a one-unit increase, while a change of $c$ units has odds ratio $e^{c\beta_j}$ in a model with a constant coefficient for that predictor.

Wrong: “The odds ratio is 2, so the probability doubles.” Correction: Double the baseline odds, then convert those odds to a probability.
Wrong: “The odds ratio is 2, so the probability rises by 0.20.” Correction: Calculate $p_1-p_0$; there is no fixed difference implied by the odds ratio.
Wrong: “An odds ratio below one means the probability reduction is the same for everyone.” Correction: It gives a lower probability for the specified comparison, but the absolute reduction depends on baseline risk.

Key takeaway: For a fixed logistic odds ratio $\theta$, convert baseline probability to odds, multiply by $\theta$, and convert back. Report the resulting probability change separately; it depends on baseline risk.

Check Your Understanding

For each hypothetical Bernoulli outcome, calculate probabilities from odds or a linear predictor as specified. Probabilities in the answers are rounded to four decimal places.

A predictor change has odds ratio 2 and the baseline probability is 0.3000. Find the updated probability and absolute increase. (Answer: $p_1=6/13=0.4615$; increase $=0.1615$.)
A predictor change has odds ratio 0.5 and baseline probability 0.2000. Find the updated probability and absolute change. (Answer: $p_1=1/9=0.1111$; change $=-0.0889$.)
A logistic model has baseline linear predictor $-\ln 3$ and the predictor change adds $\ln 2$. Find both fitted probabilities and their difference. (Answer: $p_0=0.2500$, $p_1=0.4000$; increase $=0.1500$.)
A predictor change has odds ratio 3. Compare the probability increases from baseline probabilities 0.2000 and 0.7000. (Answer: from 0.2000, $p_1=0.4286$, increase $=0.2286$; from 0.7000, $p_1=0.8750$, increase $=0.1750$.)
A predictor has coefficient $\ln(1.5)$. Compare the fitted probability change for baseline probabilities 0.1000 and 0.6000 after a one-unit increase. (Answer: odds ratio $=1.5$; changes are $0.0429$ and $0.0923$, respectively.)

What You'll Learn