parade: Generate dataset for a diagnostic parade

In janhove/cannonball: Tools for Teaching Statistics

Generate dataset for a diagnostic parade

Description

This function generates a parade (= lineup() in the nullabor package) that hides the observations, fitted values, and residuals of a statistical model you want to diagnose among the observations, fitted values, and residuals of a number similar models that were fitted on simulated outcome data. The sets of simulated outcome data are generated from the original model so that this model's assumptions are literally true for the simulated data. The 'tibble' (dataframe) created by this can be used to draw panels of diagnostic plots (see examples).

Usage

parade(model, full_data = NULL, size = 20)

Arguments

model	The name of the statistical model you want to diagnose. Currently only lm(), gam() (from the mgcv package) and lmer() (from the lme4 package) models are supported. For the lmer() models, only residual diagnostics are supported; support for BLUP ('random effects') diagnostics is still lacking.
full_data	By default, the output will only include variables that are part of the model. If you want to include all the variables that are present in the dataframe on which the model was fitted, supply this dataframe's name to full_data.
size	The number of simulated and actual datasets that the parade will contain. This defaults to 20, meaning that the actual dataset will be hidden among 19 simulated datasets.

Value

A tibble containing predictors, outcomes, fitted values and residuals for both the real dataset and simulated datasets.

Transformed predictors

If you want to include transformed predictors in the model call (e.g., log(x)), transform the predictor before using it in the model call (see examples).

This function relies on augment in the broom package. Since augment() cannot handle model calls with poly() or ns(), parade() can't handle these, either. (For 'lmer' models, the augment function in the broom.mixed package is used.)

Examples

# A simple regression model
m <- lm(mpg ~ disp, data = mtcars)

# Generate parade and check linearity
my_parade <- parade(m)
my_parade
lin_plot(my_parade)
reveal(my_parade)

# Regenerate parade and check constant variance
my_parade <- parade(m)
var_plot(my_parade)
reveal(my_parade)

# Regenerate parade and check normality
my_parade <- parade(m)
norm_qq(my_parade)
norm_hist(my_parade)
norm_hist(my_parade, bins = 10)
reveal(my_parade)

# If you want to include all predictors in the dataset in the parade:
my_parade <- parade(m, full_data = mtcars)
my_parade

# If you want to generate a parade with 50 instead of 20 plots:
my_parade <- parade(m, size = 50)
norm_qq(my_parade)

# The function also works for generalised additive models fitted with mgcv:
library(mgcv)
m.gam <- gam(mpg ~ s(disp) + wt + s(qsec), data = mtcars)
my_parade <- parade(m.gam)
lin_plot(my_parade)
my_parade <- parade(m.gam)
norm_qq(my_parade)

m.gam <- gam(mpg ~ te(disp, qsec) + wt, data = mtcars)
my_parade <- parade(m.gam)
lin_plot(my_parade)

# And has some limited support for lmer() models (from the lme4 package)
library(lme4)
m.lmer <- lmer(Reaction ~ Days + (Days|Subject), data = sleepstudy)
my_parade <- parade(m.lmer)
norm_hist(my_parade, bins = 15)
# Support for diagnosing the BLUPs would be nice.

# Transformed predictors:
# This won't work:
# m <- lm(mpg ~ log2(disp), data = mtcars)
# my_parade <- parade(m)

# This will:
mtcars$log2.disp <- log2(mtcars$disp)
m <- lm(mpg ~ log2.disp, data = mtcars)
my_parade <- parade(m)

janhove/cannonball documentation built on Feb. 19, 2025, 5:13 a.m.