Predictor effect plots display the response surface of a complex
regression model (such as a linear or generalized linear model with
interactions, predictor and response transformations, polynomial
terms, or regression-spline terms) as a series of 2D graphs,
focusing in turn on each predictor in the model and averaging or
conditioning over the other predictors. Adding partial residuals to
these graphs and smoothing the residuals can reveal lack of
fit---that is, failure in the functional form of the model---and
point the way toward improvement of the model. The methods
described are implemented in a very general manner in the effects
package for R.
This talk describes joint work with Sanford Weisberg that appears
in a recent paper: J. Fox and S. Weisberg, "Visualizing Fit and
Lack of Fit in Complex Regression Models with Predictor Effect
Plots and Partial Residuals," Journal of Statistical Software,
87(9): 1-27.