How to Interpret glm Output in R

Spread the love

Generalized Linear Models (GLM) are a class of models that generalize linear regression, allowing for response variables that have error distribution models other than a normal distribution. In R, the glm() function provides a way to fit generalized linear models. After fitting a GLM, interpreting the output is crucial to understanding the relationships between variables and making predictions. In this article, we’ll delve into interpreting the output of glm() in R.


  1. Introduction to GLM
  2. Understanding the Output Structure
  3. Coefficients and Their Interpretation
  4. Goodness-of-fit Measures
  5. Model Diagnostics
  6. Residuals and Leverage
  7. Final Thoughts

1. Introduction to GLM

GLM is an extension of the ordinary linear regression (OLR) that allows for non-normal distribution of residuals (errors). GLMs comprise three elements:

  • Random Component: Specifies the probability distribution for the response variable (e.g., normal, Poisson, binomial).
  • Systematic Component: Linear predictor, a linear combination of the predictors.
  • Link Function: A function that connects the expected value of the response to the systematic component.

2. Understanding the Output Structure

When you run the glm() function and call the summary() function in R, you will receive an output that typically comprises:

  • Call: Recap of the model formula.
  • Deviance Residuals: Summary statistics of the residuals.
  • Coefficients: Information about the model coefficients including estimates, standard error, z-values, and p-values.
  • (Dispersion parameter): For binomial and Poisson, it can give insights into over- or underdispersion.
  • Null and Residual Deviance: Useful for comparing models.
  • AIC (Akaike’s Information Criterion): A measure of the goodness of fit of the model.

3. Coefficients and Their Interpretation

The coefficients table is perhaps the most frequently inspected part of the output:

  • Estimate: The estimated change in the response for a one-unit change in the predictor, keeping all other predictors constant.
  • Std. Error: The standard error of the coefficient estimate. It measures the variability of the estimate.
  • z-value: It’s the ratio of the estimate to the standard error.
  • Pr(>|z|): The two-tailed p-value. It tests the null hypothesis that the coefficient is equal to zero (no effect).

Note: The interpretation of the coefficients depends heavily on the link function and the distribution of the response variable.

4. Goodness-of-fit Measures

Null Deviance and Residual Deviance:

  • Null Deviance: Reflects the difference between a model with only the intercept and a saturated model.
  • Residual Deviance: Reflects the difference between the observed data and the model’s predictions. A large decrease in deviance when moving from the null to the model indicates your predictors are improving the model’s fit.

AIC: AIC stands for Akaike Information Criterion. It penalizes adding uninformative variables to the model, so a smaller AIC suggests a better-fitting model.

5. Model Diagnostics

  • Overdispersion/Underdispersion: For Poisson and Binomial models, the dispersion parameter can be examined. If it’s significantly greater than 1, the model may be overdispersed.
  • Link Test: To verify if the chosen link function is appropriate.
  • Leverage and Influence Plots: To identify influential observations.

6. Residuals and Leverage

Examining residuals is an essential part of diagnosing the model.

  • Deviance Residuals: Ideally, they should be roughly normally distributed for a well-fitting model.
  • Leverage: High leverage points can unduly influence the model, so they need to be checked and potentially addressed.

7. Final Thoughts

Interpreting a GLM output in R requires a combination of statistical knowledge and domain expertise. While the above guidelines provide a roadmap, always ensure that the model makes sense in the context of the problem at hand. Moreover, remember that statistical significance does not always equate to practical significance.

In conclusion, GLMs offer a flexible framework for modeling various types of response variables. The key is to interpret the output thoughtfully, using the guidelines above, but also by integrating domain-specific knowledge and further diagnostic tools.

Posted in RTagged

Leave a Reply