Shows the values a model predicts as one focal predictor varies, holding the other predictors at typical values (the mean for numeric predictors, the most frequent level for factors). A confidence band (numeric predictor) or confidence intervals (factor predictor) convey uncertainty.
Usage
effects_plot(
model,
predictor,
conf_level = 0.95,
n = 100,
rug = TRUE,
colour = depictr_brand(),
title = NULL,
x_lab = NULL,
y_lab = NULL
)Arguments
- model
A fitted model (
lm,glm,merMod, ...).- predictor
Name of the focal predictor (string): a variable in the model's formula, or a model-frame name such as
"log(rainfall)", which stands for the one variable it transforms. A term built from several variables, such as"I(rainfall * fertiliser)", is refused.- conf_level
Confidence level for the interval, a single number strictly between 0 and 1.
- n
Number of points across the range of a numeric predictor.
- rug
Whether to add a rug of the observed predictor values (numeric predictors).
- colour
Colour for the line/points and band. Defaults to the depictr brand blue.
- title, x_lab, y_lab
Title and axis labels.
Value
A ggplot2::ggplot object.
Details
The prediction grid holds the variables the formula names, on their own
scale, and stats::predict() re-applies each transformation to it. A model
with terms such as log(rainfall), poly(fertiliser, 2), splines::ns(),
scale(), factor() or I() is therefore predicted as fitted, and the
x-axis shows the raw variable. Typical values and the range of the focal
predictor are taken over the rows the model was fitted to. A numeric
variable that enters only through factor() is held at its most frequent
value. A variable that enters only through a transformation is recovered
from the data named in the model's call, so such a model fitted to
workspace vectors without data = is refused. A term computed from
statistics of its rows, such as I(x - mean(x)), is refused too, since
predict() would compute it afresh over the grid. Centre such a variable in
the data, or use scale(), which predict() re-applies as fitted.
An offset, whether an offset() term or the offset argument, is held at
its mean over the rows the model was fitted to, on the link scale. For a
Poisson rate model with offset(log(exposure)), the line is the expected
count at the geometric-mean exposure. The offset stays at its mean when the
focal predictor enters it too, as in log(exposure) + offset(log(exposure)),
so the line then shows the predictor's effect beyond the offset. The y-axis
label ends in "at the mean offset".
Predictions and standard errors come from stats::predict(); glm
predictions are formed on the link scale and back-transformed, so a binomial
model shows predicted probabilities. The band uses the multiplier of the
model's own tests, as tidy_estimates() does. An lm, and a glm whose
dispersion is estimated, get t on the residual degrees of freedom, so a
Gaussian glm draws the band of the same lm. Binomial and Poisson fits
and mixed models get z. Mixed models fitted with
lme4::lmer()/lme4::glmer() are supported too: predictions use only the
fixed effects (as re.form = NA does) and standard errors come from the
fixed-effect design matrix and vcov(). Works with lm, glm and
merMod; other model classes are attempted on a best-effort basis.
