What a GLM Coefficient Does—and Does Not—Say
A generalized linear model (GLM) can summarize how an outcome varies with one predictor while accounting for other variables in the fitted model. That summary is useful in a predictive report, but it is easy to overstate. A coefficient describes a modeled relationship in the observed data, conditional on the model’s other predictors. By itself, it does not show what would happen if someone intervened to change a predictor.
The word effect is sometimes used informally for a coefficient or its transformed value. In a report, that word can sound causal. Prefer precise language such as “associated with higher fitted odds, holding the listed variables fixed,” and specify what the comparison means.
A GLM specifies a response distribution from the exponential-family class and a link between its conditional mean and predictors. For the usual likelihood-based interpretation, observations are modeled as independent conditional on their predictors, and the chosen response distribution and mean structure must be adequate for the analysis. These are modeling assumptions, not evidence that a predictor causes the response. A correctly computed coefficient can still reflect confounding, selection, measurement choices, or model misspecification.
For reporting, identify four items before interpreting a number: the response being modeled, the link function, the predictor’s unit or reference group, and the other predictors held fixed. If there is an interaction or nonlinear term, also state the values or groups at which the comparison applies.
Logistic Regression: Odds Ratios Are Not Probability Ratios
For a binary response \(Y_i\), logistic regression models the probability \(p_i=P(Y_i=1\mid X_i)\) using the logit link. The fitted model has the form below. Conditional on the predictors, a Bernoulli response has mean \(p_i\) and variance \(p_i(1-p_i)\); the logistic model assumes that the log odds follow the stated linear predictor.
For a one-unit increase in \(x_j\), with other predictors fixed and no interaction involving \(x_j\), the fitted log odds change by \(\beta_j\). Exponentiating gives the odds ratio \(e^{\beta_j}\). It compares the fitted odds at the two predictor values; it is not generally the ratio of their probabilities.
Worked Example: Reporting a Logistic Coefficient and a Probability
A hypothetical insurer fits a logistic model for whether a policy has at least one claim in the next 12 months. Let \(x\) be the number of five-year age increments above age 40, and let \(z\) indicate whether the policy has a feature that the model includes as a control. The fitted model is \(\operatorname{logit}(p)=-2.1972+0.18x+0.25z\). For policies with \(z=0\), calculate and report the fitted odds ratio for a one-increment increase in age, and compare the fitted probabilities at \(x=0\) and \(x=1\).
Method. Interpret the coefficient on the log-odds scale, exponentiate it for the odds ratio, then calculate probabilities using the inverse-logit transformation. The comparison holds \(z\) fixed at zero.
Step 1: Calculate the odds ratio. The age coefficient is \(0.18\) per five-year increment. Therefore,
The fitted odds are about \(1.197\) times as large, or \(19.72\%\) higher, for a five-year increment, conditional on the model’s other predictors. This is a statement about odds, not a \(19.72\%\) increase in probability.
Step 2: Find the probability at \(x=0,z=0\). The fitted logit is \(-2.1972\), so the fitted odds are \(e^{-2.1972}\), approximately \(0.1111\). Converting odds to probability gives \(0.1111/(1+0.1111)=0.1000\), or \(10.00\%\).
Step 3: Find the probability at \(x=1,z=0\). The fitted logit is \(-2.1972+0.18=-2.0172\). Equivalently, multiply the original odds by the odds ratio: \(0.1111(1.197217)=0.1330\). The probability is \(0.1330/(1+0.1330)=0.1174\), or \(11.74\%\).
Step 4: State the probability-scale comparison. The fitted probability difference is \(0.1174-0.1000=0.0174\), or about \(1.74\) percentage points. For this profile, report that the five-year age increment is associated with higher fitted odds and a fitted probability increase from \(10.00\%\) to \(11.74\%\), holding the included control fixed. Do not describe the increment as causing a claim.
Check. The probability increases by much less than \(19.72\%\) of its starting value: \(10.00\%\times1.197217\) is not the new probability. The calculated odds, \(0.1111\) and \(0.1330\), have the required ratio of approximately \(1.197\). The common mistake avoided is treating an odds ratio as a probability ratio.
Log Links: Rate Ratios and the Exposure Offset
For a count outcome, a Poisson GLM with a log link models the logarithm of the conditional mean. If observations have different exposure amounts \(e_i\), an offset \(\log(e_i)\) can account for exposure, with its coefficient fixed at one. Under the conventional Poisson model, counts are conditionally independent and the conditional variance equals the conditional mean. The rate interpretation is conditional on this model specification.
For a one-unit increase in \(x_i\), \(e^{\beta_1}\) is the ratio of modeled rates, holding other included predictors fixed. If exposure is unchanged, it is also the ratio of modeled expected counts. When exposure differs, compare rates or calculate expected counts using each exposure amount; do not attribute a count difference to the predictor alone.
Worked Example: Interpreting a Rate Ratio With Different Exposure
A hypothetical portfolio model predicts claim counts using a Poisson GLM with a log link and an exposure offset. For a reference profile, the fitted rate is \(0.0800\) claims per policy-year. A one-unit increase in a recorded driving-score variable has coefficient \(-0.22\). Calculate the rate ratio and the fitted expected claim counts for exposure of \(1.5\) policy-years at the reference score and one unit higher.
Method. Exponentiate the log-rate coefficient, then multiply each fitted rate by its exposure. This keeps the rate comparison separate from the exposure comparison.
Step 1: Calculate the rate ratio.
The model associates a one-unit higher score with a fitted rate \(0.8025\) times as large, or about \(19.75\%\) lower, conditional on the model’s other variables. A coefficient below zero gives a ratio below one because exponentiation preserves the direction of the change.
Step 2: Calculate expected claims at the reference score. The expected count is rate multiplied by exposure: \(0.0800\times1.5=0.1200\) claims.
Step 3: Calculate the fitted rate and expected count one score unit higher. The new rate is \(0.0800\times0.802519=0.0642015\) claims per policy-year. The expected count is \(0.0642015\times1.5=0.0963023\) claims.
Step 4: Compare and check. The expected-count ratio is \(0.0963023/0.1200=0.802519\), the same rate ratio because exposure is the same in both calculations. The difference is \(0.0963023-0.1200=-0.0236977\) expected claims. Report the fitted rate association, not a causal reduction. The common mistake avoided is comparing counts while overlooking unequal exposure; if the exposures differed, the count ratio would also include their ratio.
Interactions Change What a Coefficient Means
A model with an interaction allows the association for one predictor to vary with another predictor. In a logistic model with a binary predictor \(x\), a binary group indicator \(z\), and their product \(xz\), the coefficient on \(x\) alone is the log-odds comparison when \(z=0\). When \(z=1\), the comparison uses the sum of the \(x\) coefficient and the interaction coefficient. Reporting the main-effect coefficient as a single association for all groups would omit this qualification.
Worked Example: Reporting Group-Specific Logistic Associations
A hypothetical service provider models whether a customer renews a contract. Let \(x=1\) indicate enrollment in an optional reminder program, and \(z=1\) indicate a remote service area. The fitted model is \(\operatorname{logit}(p)=-2+0.30x+0.50z-0.40xz\). Calculate the odds ratio and fitted probability comparison for enrollment within each service-area group.
Method. Substitute each group value into the model. For each comparison, calculate the change in logit, exponentiate for the odds ratio, and apply the inverse-logit to obtain probabilities. The model describes conditional associations; the enrollment indicator is not assumed to be randomly assigned.
Step 1: Calculate the comparison for \(z=0\). With \(z=0\), the enrollment log-odds difference is \(0.30\), so the odds ratio is \(e^{0.30}=1.349859\), approximately \(1.350\). At \(x=0\), the logit is \(-2\); the probability is \(1/(1+e^2)=0.1192\). At \(x=1\), the logit is \(-2+0.30=-1.70\); the probability is \(1/(1+e^{1.70})=0.1545\). The fitted probability difference is \(0.1545-0.1192=0.0353\), or \(3.53\) percentage points.
Step 2: Calculate the comparison for \(z=1\). The enrollment log-odds difference is \(0.30-0.40=-0.10\), so the odds ratio is \(e^{-0.10}=0.904837\), approximately \(0.905\). At \(x=0\), the logit is \(-2+0.50=-1.50\); the probability is \(1/(1+e^{1.50})=0.1824\). At \(x=1\), the logit is \(-2+0.50+0.30-0.40=-1.60\); the probability is \(1/(1+e^{1.60})=0.1680\). The fitted probability difference is \(0.1680-0.1824=-0.0144\), or \(-1.44\) percentage points.
Step 3: Report the comparison in context. In the fitted model, enrollment is associated with higher odds of renewal in the \(z=0\) group and lower odds in the \(z=1\) group, conditional on the model terms. The estimates correspond to fitted probabilities of \(11.92\%\) versus \(15.45\%\) for \(z=0\), and \(18.24\%\) versus \(16.80\%\) for \(z=1\). They do not show that the reminder program causes those changes.
Check. The two odds ratios differ because the interaction coefficient changes the enrollment slope from \(0.30\) to \(0.30-0.40=-0.10\). The common mistake avoided is reporting \(e^{0.30}\) as the enrollment odds ratio for both groups.
Write the Interpretation, Not a Causal Story
The fitted coefficient is a statement about the model’s conditional pattern: it compares records that have the same values of the other included predictors, apart from the predictor being changed. This adjustment does not automatically make the comparison causal. Predictors may be associated with unmeasured factors, recorded after the outcome, or affected by how the study population was selected. The model may also use an inadequate functional form or omit relevant interactions.
A causal claim concerns outcomes under an intervention or exposure assignment, not merely the fitted association among observed records. A GLM fitted for prediction does not, by itself, establish the assumptions needed for a causal interpretation. Unless a suitable study design and causal analysis justify that interpretation, report a conditional association and avoid verbs such as “causes,” “prevents,” “drives,” or “reduces” when they assert intervention effects.
Say whether the response is a probability, count, rate, or another quantity, and identify the link used.
Give the predictor unit or category contrast, the reference group, and the values of any interacting predictors.
For a logit model, distinguish odds ratios from probabilities. For a log-link model, describe a mean or rate ratio and account for exposure.
Say which included predictors are held fixed, and clarify that the result is a fitted association unless causal evidence supports more.
A useful report sentence is: “Among records with the same values of the other listed model predictors, a one-unit increase in \(x\) is associated with an estimated odds ratio of \(r\) for the modeled outcome.” For a log-link model, replace “odds ratio” with “rate ratio” or “mean ratio” as appropriate. If an absolute probability difference matters to the decision, calculate and report fitted probabilities for a clearly specified profile, as in the examples. Those probabilities remain model predictions, not guaranteed outcomes for an individual.
This disciplined wording supports a defensible predictive interpretation: it preserves what the GLM estimates, makes the comparison auditable, and avoids converting an observational association into an unsupported claim about causes or interventions. This reporting approach is aligned with the SOA Exam PA (Predictive Analytics) syllabus.
Check Your Understanding
For each problem, identify the comparison and use language that describes a fitted association rather than a cause.
- A logistic model has a coefficient of \(0.12\) for each additional year of account tenure, with no interaction involving tenure. Calculate the odds ratio for two additional years. (Answer: \(e^{0.24}=1.271\), approximately \(27.1\%\) higher fitted odds, conditional on the other modeled predictors.)
- A binary-outcome model gives fitted probabilities of \(0.20\) and \(0.25\) for two profiles that differ in one predictor only. Calculate the absolute probability difference and explain why it is not an odds ratio. (Answer: \(0.25-0.20=0.05\), or 5 percentage points; the odds ratio is \((0.25/0.75)/(0.20/0.80)=1.333\).)
- A Poisson log-link model has a coefficient of \(0.15\) for a one-unit predictor increase. Calculate the rate ratio and percentage rate change. (Answer: \(e^{0.15}=1.162\), approximately \(16.2\%\) higher fitted rate, conditional on other model terms.)
- A logistic model includes \(0.40x-0.25xz\), where \(x\) is a binary indicator and \(z\) is a binary group indicator. Calculate the odds ratio for changing \(x\) from 0 to 1 in each group. (Answer: for \(z=0\), \(e^{0.40}=1.492\); for \(z=1\), \(e^{0.40-0.25}=e^{0.15}=1.162\).)
- A count model has fitted rate \(0.06\) per exposure-year for a reference profile and a coefficient of \(-0.10\) for a predictor increase. For a profile with 2 exposure-years, calculate the new expected count. (Answer: rate \(0.06e^{-0.10}=0.0543\); expected count \(0.0543(2)=0.1086\).)