Interpreting Coefficients Under Different Links
You'll learn: How the choice of link function changes whether coefficients describe effects on the original outcome scale or a transformed scale.
The Core Challenge
When you fit a Generalized Linear Model (GLM), the link function transforms your outcome variable. This means your regression coefficients live on the transformed scale, not the original scale you care about. Understanding what a coefficient actually means depends entirely on which link you used.
Think of it like currency exchange: a coefficient might be "5 euros," but to know what that buys in dollars, you need the exchange rate (the inverse link function).
Link-Specific Interpretations
Identity Link (e.g., normal regression): Coefficients are straightforward. A one-unit increase in X changes Y by β units directly. The coefficient is already on the original scale.
Log Link (e.g., Poisson regression): Coefficients represent changes on the log scale. A coefficient β means a one-unit increase in X multiplies the expected outcome by e^β. You must exponentiate to get the multiplicative effect on the original count scale.
Logit Link (e.g., logistic regression): Coefficients represent changes in log-odds. A coefficient β means a one-unit increase in X changes the log-odds by β. Exponentiating gives you an odds ratio, not a probability directly.
Probit Link: Coefficients live on the latent normal scale, making direct interpretation even harder—you typically focus on predicted probabilities instead.
Practical Advice
Always ask: "Does this coefficient describe the effect on my original outcome, or on a transformed version?" For non-identity links, convert coefficients to the original scale when reporting (e.g., "a 23% increase in count" rather than "a coefficient of 0.207").
Key Takeaway: Link functions determine the scale of your coefficients—identity links give direct effects, while log and logit links require transformation (exponentiation) to interpret effects on the original outcome scale.