Inspect Residuals

Plot deviance residuals and the QQ plot to assess model fit.

Deviance residuals summarize how far each observation falls from the model’s fitted value, scaled to the family’s variance function. If the model fits well, deviance residuals should look like Gaussian white noise: symmetric, roughly constant variance, and no systematic pattern with respect to fitted values or covariates.

Fit

Fit a simple GAM to get started.

import whittaker as wk

# Load data and fit GAM
data = wk.load_dataset("mcycle")
model = wk.GAM("accel ~ s(times)").fit(data)

Four-panel diagnostics

The wk.check() function returns all four diagnostic panels as a single stacked chart.

wk.check(model)

The four panels are:

  • QQ plot (top-left): deviance residuals against theoretical Gaussian quantiles. Points should fall close to the diagonal reference line.
  • Residuals vs fitted values (top-right): residuals on the y-axis, fitted values on the x-axis. No trend or funnel shape is ideal.
  • Histogram of residuals (bottom-left): the marginal distribution of deviance residuals. A single symmetric peak centered near zero is the target.
  • Response vs fitted values (bottom-right): observed response against model predictions. Points should scatter evenly around the 45-degree line.

What to look for

Patterns in these panels point to specific problems:

  • S-curve in the QQ plot: the residuals have heavier or lighter tails than a Gaussian, suggesting the wrong response family. Try wk.Gamma() for right-skewed continuous data or wk.Poisson() for counts.
  • Funnel shape in residuals vs fitted: variance grows with the mean (heteroscedasticity). A log or square-root transformation of the response, or switching to a Poisson or Gamma family, can stabilize variance.
  • Bimodal histogram: the data may come from two distinct sub-populations. Consider adding a grouping factor or fitting separate models.
  • Systematic curve in residuals vs fitted: the model is missing a nonlinear relationship. Add an interaction term, increase k, or include an additional smooth for a covariate not yet in the formula.