Deviance and dev-dev plots

Ported from multinma::plot.nma_dic() (Phillippo et al. 2020). With a single dic() object the plot shows each data point's contribution to the posterior mean deviance; points contributing much more than the rest are fitted poorly. With two dic() objects it draws the dev-dev plot: points below the line of equality fit better under the second model, points above it fit better under the first.

Usage

# S3 method for class 'cpaic_dic'
plot(x, y = NULL, ..., labels = c("Model 1", "Model 2"))

Arguments

x

A cpaic_dic object from dic().

y

An optional second cpaic_dic object, for a dev-dev plot.

...

Unused.

labels

Model names used on the axes of a dev-dev plot.

Value

A ggplot object.

Details

dic() stores the posterior mean deviance per data point, not the residual deviance. The two differ by the saturated log likelihood, which for the binomial and Poisson families is a function of the data alone: it is the same under both models, so it shifts both axes equally and the line of equality keeps its meaning. For a Gaussian likelihood the saturated term also involves the model's own sigma, so two Gaussian models with very different residual variance are shifted by different amounts; read that comparison with care.

For posterior uncertainty and for the leverage plot, which need the saturated model explicitly, call plot_leverage() on the fitted model itself.

See also

dic(), plot_leverage()

Examples

if (FALSE) {
plot(dic(fit_fe), dic(fit_re), labels = c("Fixed", "Random"))
}