Translating GLM Effects Into Plain Language

Why the Link Function Matters

A GLM coefficient is not automatically a percentage change in the outcome. Its meaning depends on the model’s link function, the predictor’s units, and the values of the other predictors. A report that says “the outcome rises by 30%” may be accurate for a log-link model, but wrong for a logistic model if the coefficient describes an increase in odds rather than probability.

The goal is to preserve the model’s actual comparison when translating a coefficient: identify what changes, what is held fixed, and how the predicted outcome changes. These statements describe model-based associations and predictions. Unless a study design and analysis support causal inference, they should not say that changing a predictor will cause an outcome to change. This skill is aligned with the SOA Exam PA (Predictive Analytics) syllabus.

Definition: A generalized linear model (GLM) relates the conditional mean of an outcome, given predictors, to a linear predictor through a link function. Write the conditional mean as $\mu(\mathbf{x})=E[Y\mid\mathbf{x}]$, and the linear predictor as $\eta(\mathbf{x})=\beta_0+\beta_1x_1+\cdots+\beta_px_p$. The model specifies $g(\mu(\mathbf{x}))=\eta(\mathbf{x})$, where $g$ is the link function. A coefficient’s interpretation depends on this link, the predictor’s scale, and any interactions in the linear predictor.

For the common links below, a one-unit predictor change has a different interpretation. The comparison assumes that other included predictors remain fixed and that the model form applies over the values being compared.

Model and link	Effect of a one-unit increase in $x_j$	Plain-language scale
Logistic regression: $g(\mu)=\log\{\mu/(1-\mu)\}$	Odds are multiplied by $e^{\beta_j}$, when there is no interaction involving $x_j$.	Odds, not probability
Log link: $g(\mu)=\log(\mu)$	Predicted mean is multiplied by $e^{\beta_j}$, when there is no interaction involving $x_j$.	Relative change in the mean
Identity link: $g(\mu)=\mu$	Predicted mean changes by $\beta_j$, when there is no interaction involving $x_j$.	Absolute change in the outcome’s units

From a Logistic Coefficient to a Probability

For a binary outcome, logistic regression models the log odds of the event. If the linear predictor increases by $\beta_j$ when $x_j$ increases by one unit, exponentiating the coefficient gives an odds ratio. To communicate a predicted probability change, however, the starting probability is also needed. The same odds ratio can correspond to different probability changes at different starting probabilities.

\log\left(\frac{p(\mathbf{x})}{1-p(\mathbf{x})}\right)=\eta(\mathbf{x}), \qquad \frac{\text{odds after}}{\text{odds before}}=e^{\beta_j}

Here $p(\mathbf{x})$ is the conditional probability of the event. The odds are $p/(1-p)$. The odds ratio does not state the ratio of the two probabilities, and it does not by itself give their difference in percentage points.

Justification: Suppose the model has no interaction involving $x_j$, and compare two predictor settings that differ only by one unit in $x_j$. Their linear predictors differ by $\beta_j$. Since the logistic link makes the linear predictor the log odds, subtracting the two log odds gives $\beta_j$. Exponentiating both sides gives the ratio of the odds as $e^{\beta_j}$. To recover a probability from odds $o$, use $p=o/(1+o)$. Thus a probability comparison requires the starting odds, or equivalently the starting probability.

Worked Example: A Logistic Coefficient and a Baseline Probability

In a hypothetical renewal study, a logistic model predicts whether a policy renews. For the same values of the other predictors, the predicted renewal probability is $0.20$ for a customer with a one-unit lower value of a service-contact measure. The fitted coefficient for that measure is $\log(1.5)$ per unit. Translate a one-unit increase into plain language, including the new predicted probability. Name the method and the mistake it avoids.

Method. Convert the coefficient to an odds ratio, apply it to the baseline odds, and convert the resulting odds back to a probability.

The baseline probability is $0.20$, so the baseline odds are $\frac{0.20}{1-0.20}=\frac{0.20}{0.80}=0.25$. The one-unit increase multiplies those odds by $e^{\log(1.5)}=1.5$. The new odds are $0.25(1.5)=0.375$.

Convert those odds to a probability: $\frac{0.375}{1+0.375}=\frac{0.375}{1.375}=\frac{3}{11}=0.2727$, to four decimal places. The predicted probability therefore rises from $0.2000$ to $0.2727$, an increase of $0.0727$, or 7.27 percentage points.

Plain-language statement: Holding the other modeled characteristics fixed, a one-unit increase in the measure is associated with 1.5 times the renewal odds. At this baseline, the model’s predicted renewal probability increases from 20.00% to 27.27%, a rise of 7.27 percentage points.

Check: The odds increased by 50%, from 0.25 to 0.375. The probability increased by 7.27 percentage points, not by 50%. Saying “the probability rises by 50%” would incorrectly treat an odds ratio as a probability ratio.

Log Links and Identity Links

In a log-link model, exponentiating a coefficient gives a multiplicative effect on the conditional mean, when the predictor has no interaction with other terms. A coefficient of $\log(1.25)$, for example, corresponds to a 25% increase in the predicted mean for a one-unit increase in the predictor. The absolute change in the mean depends on its starting value.

In an identity-link model, the coefficient gives an additive change in the conditional mean in the outcome’s units. If the coefficient is $-80$ dollars per 100 units of a predictor, a 200-unit increase corresponds to a predicted decrease of $160$ dollars, provided the model has no relevant interaction. Do not convert an additive coefficient into a percentage without specifying a baseline and calculating that percentage separately.

Worked Example: A Log-Link Effect on an Expected Count

A hypothetical Poisson GLM predicts the number of service claims for a fixed one-year exposure. At a mileage value of zero hundreds of miles, the predicted mean is 2.4 claims. The model has a log link and a coefficient of $\log(1.25)$ for mileage measured in hundreds of miles. Compare the predicted mean at 200 additional miles with the starting mean. State the method and the mistake it avoids.

Method. Use the log-link coefficient as a multiplicative effect per predictor unit. Since 200 miles equals two units of 100 miles, apply the multiplier twice.

The predicted mean after two units is $2.4e^{2\log(1.25)}=2.4(1.25)^2$. Now, $(1.25)^2=1.5625$, so the new predicted mean is $2.4(1.5625)=3.75$ claims.

The mean difference is $3.75-2.4=1.35$ claims. The relative increase is $\frac{3.75-2.4}{2.4}=\frac{1.35}{2.4}=0.5625$, or 56.25%.

Plain-language statement: For the same one-year exposure and the same values of the other modeled predictors, 200 additional miles are associated with a predicted mean of 3.75 claims rather than 2.4, a 56.25% increase in the predicted mean.

Check: Applying the one-unit 25% increase only once would give $2.4(1.25)=3.0$, which is incorrect for a two-unit change. The effect compounds: $1.25^2=1.5625$. The prediction is a mean count, not a claim count guaranteed for an individual policy.

Worked Example: An Additive Effect in Outcome Units

A hypothetical identity-link GLM predicts repair cost in dollars. For a vehicle with a predictor value of zero, the predicted mean repair cost is $1,200. The fitted coefficient for vehicle age, measured in years, is $85 per year. For the same values of all other predictors, compare the predicted mean for a vehicle that is three years older. State the method and the mistake it avoids.

Method. Under the identity link, multiply the coefficient by the predictor change and add that change to the starting predicted mean.

The predictor change is three years, so the change in predicted mean is $3(85)=\$255$. The new predicted mean is $\$1{,}200+\$255=\$1{,}455$.

Plain-language statement: Holding the other modeled characteristics fixed, a three-year difference in vehicle age is associated with a $255 higher predicted mean repair cost, from $1,200 to $1,455.

Check: The model specifies an additive change of $85 per year. It does not say that cost rises by 85% per year. Describing the result as an 85% increase would confuse the coefficient’s outcome units with a relative change.

Reference Categories and Interactions

For a categorical predictor, the coefficient compares a named category with the reference category, holding other modeled predictors fixed. State both categories in the report. If the model includes an interaction, the effect of one predictor depends on the value of the other predictor. In that case, a main-effect coefficient describes the comparison only at the reference value of the interacting predictor; it is not a universal effect.

For example, in a logistic model with a binary predictor $x$, a binary group indicator $G$, and their interaction, the linear predictor includes $\beta_xx+\beta_GG+\beta_{xG}xG$. The odds ratio for changing $x$ from zero to one is $e^{\beta_x}$ when $G=0$, but is $e^{\beta_x+\beta_{xG}}$ when $G=1$. A clear interpretation must identify the group and, when a probability is reported, calculate probabilities at the relevant predictor settings.

Worked Example: Translating an Interaction Without Overgeneralizing

A hypothetical logistic model predicts whether a customer renews a service contract. Let $x=1$ indicate that a service reminder was recorded and $G=1$ indicate a new customer; $G=0$ is the reference group of established customers. The model is $\eta=\log(0.25)+\log(2)x+\log(3)G+\log(0.5)xG$. Calculate and compare predicted probabilities for all four combinations. Explain the interaction and state the method.

Method. Calculate the linear predictor for each combination, exponentiate it to obtain odds, and convert odds to probability using $p=o/(1+o)$. This avoids treating a main effect as if it applied equally to both groups.

For established customers with no reminder, $x=0,G=0$, so $\eta=\log(0.25)$. The odds are $0.25$, and the probability is $0.25/(1+0.25)=0.20$.

For established customers with a reminder, $x=1,G=0$, so $\eta=\log(0.25)+\log(2)=\log(0.50)$. The odds are $0.50$, and the probability is $0.50/1.50=1/3=0.3333$, to four decimal places.

For new customers with no reminder, $x=0,G=1$, so $\eta=\log(0.25)+\log(3)=\log(0.75)$. The odds are $0.75$, and the probability is $0.75/1.75=3/7=0.4286$, to four decimal places.

For new customers with a reminder, $x=1,G=1$, all four terms apply: $\eta=\log(0.25)+\log(2)+\log(3)+\log(0.5)=\log(0.75)$. The odds are again $0.75$, and the probability is $3/7=0.4286$.

Plain-language statement: Among established customers, the reminder indicator is associated with odds multiplied by $e^{\log(2)}=2$; the predicted probability is 20.00% without the indicator and 33.33% with it. Among new customers, the odds multiplier is $e^{\log(2)+\log(0.5)}=1$, and the predicted probability is 42.86% at either indicator value. These are model-based comparisons, not evidence that reminders caused renewal.

Check: The coefficient $\log(2)$ is the reminder effect only for the reference group $G=0$. Reporting “the reminder doubles renewal odds for all customers” ignores the interaction term and is incorrect.

A Reliable Translation Process

Before writing a sentence about a GLM effect, identify the outcome, link, predictor scale, and comparison. Then decide whether the audience needs an effect on the model scale (such as odds) or a prediction on the outcome scale (such as probability). If you give predicted outcomes, state the settings used for the comparison; a probability or mean is conditional on the predictor values supplied to the model.

Name the outcome and the predictor change.
Specify the event or quantity predicted, the direction and size of the predictor change, and the predictor’s units.

Identify the link and comparison.
Determine whether the coefficient describes an additive mean change, a multiplicative mean change, or a change in log odds. Name any reference category.

Check for interactions.
If an interaction is present, calculate or describe the effect at the relevant value of the interacting predictor rather than presenting one coefficient as universal.

Translate and qualify.
Use outcome-scale predictions when useful, state what is held fixed, and describe an association rather than a causal effect unless causal interpretation is justified.

A concise report can still be precise. For example: “For otherwise similar records in the model, the predicted mean is 25% higher for each additional 100 miles.” For logistic regression, avoid substituting “probability” for “odds”; when probability is the decision-relevant quantity, calculate and report it at a stated baseline or pair of predictor settings.

Common Reporting Traps

Calling an odds ratio a probability ratio. An odds ratio of 1.5 means 50% higher odds, not a 50% higher probability. Convert odds to probabilities if that is the quantity the reader needs.
Omitting the predictor unit. A coefficient per year, per $1,000, or per 100 miles describes a different comparison. Match the report to the model’s actual coding and units.
Ignoring the baseline. The same odds ratio can yield different probability changes at different starting probabilities. State the baseline when presenting a probability comparison.
Misreading a main effect with an interaction. A main-effect coefficient applies at the reference value of the interacting predictor. Calculate the combined effect for other values.
Turning a prediction into a causal claim. “The model predicts a higher outcome for records with this characteristic” is not equivalent to “changing this characteristic will increase the outcome.”
Reporting a percentage when the model gives an additive effect. For an identity link, first state the change in outcome units. A relative percentage needs a specified baseline and a separate calculation.

Key takeaway: A trustworthy GLM interpretation preserves the link function, predictor units, reference category, and interaction structure. Report odds as odds, multiplicative mean effects as relative changes, and identity-link effects in outcome units. When translating to predicted outcomes, state the comparison and avoid causal wording the model does not justify.

Check Your Understanding

For each problem, identify the relevant GLM scale before writing the plain-language interpretation.

A logistic model has a coefficient of $\log(0.5)$ per 10-year increase in a predictor. At the lower value, the predicted event probability is $0.40$. Find the probability after a 10-year increase and state both the odds change and the percentage-point change. (Answer: Odds are multiplied by 0.5; probability falls from 0.4000 to 0.2500, a decrease of 15.00 percentage points.)
A Poisson GLM with a log link has a predicted mean of 10 events at $x=0$, and the coefficient for $x$ is $\log(1.08)$ per unit. Find the predicted mean at $x=3$ and the relative increase. (Answer: $10(1.08)^3=12.5971$, or about 12.60 events; the increase is about 25.97%.)
An identity-link model predicts an outcome of 840 units at $x=0$. Its coefficient is $-120$ per 100-unit increase in $x$. Find the prediction at $x=200$, assuming no relevant interaction. (Answer: $840+2(-120)=600$ outcome units.)
A logistic model has a categorical predictor with odds ratio 1.4 for category B compared with reference category A. For otherwise fixed predictors, the event probability in category A is 0.30. Find the predicted probability in category B and the percentage-point difference. (Answer: Category A odds are $0.30/0.70=3/7$; category B odds are $0.6$, giving probability $0.3750$, an increase of 7.50 percentage points.)
A logistic model includes $x$, $G$, and an $xG$ interaction, with coefficients $\beta_x=\log(1.5)$ and $\beta_{xG}=\log(2)$. State the odds ratio for a one-unit increase in $x$ when $G=0$ and when $G=1$. (Answer: When $G=0$, the odds ratio is $e^{\beta_x}=1.5$; when $G=1$, it is $e^{\beta_x+\beta_{xG}}=3.0$.)

Model and link	Effect of a one-unit increase in \(x_j\)	Plain-language scale
Logistic regression: \(g(\mu)=\log\{\mu/(1-\mu)\}\)	Odds are multiplied by \(e^{\beta_j}\), when there is no interaction involving \(x_j\).	Odds, not probability
Log link: \(g(\mu)=\log(\mu)\)	Predicted mean is multiplied by \(e^{\beta_j}\), when there is no interaction involving \(x_j\).	Relative change in the mean
Identity link: \(g(\mu)=\mu\)	Predicted mean changes by \(\beta_j\), when there is no interaction involving \(x_j\).	Absolute change in the outcome’s units

What You'll Learn