Teaching Predictions with Confidence and Prediction Intervals

predict20x() gives predictions and intervals for a fitted model. It is an S3 generic, so the same function works for linear models, generalised linear models and tslm fits.

Usage

predict20x(object, ...)

# Default S3 method
predict20x(object, ...)

# S3 method for class 'lm'
predict20x(
  object,
  newdata,
  conf.level = 0.95,
  digits = 3,
  print.out = TRUE,
  cilevel = conf.level,
  ...
)

# S3 method for class 'glm'
predict20x(
  object,
  newdata,
  conf.level = 0.95,
  digits = 3,
  print.out = TRUE,
  cilevel = conf.level,
  type = c("response", "link", "odds"),
  quasit = FALSE,
  ...
)

# S3 method for class 'tslm'
predict20x(object, newdata, ...)

Arguments

object: a fitted model, such as the output from lm or glm.
...: optional arguments that are passed to the model's prediction method, or for other methods to predict20x().
newdata: prediction data frame.
conf.level: confidence level for the intervals.
digits: number of decimal places to print.
print.out: if TRUE, print the prediction table.
cilevel: the older name for conf.level.
type: for a glm fit, the scale of the results: "response" (the default), "link", or "odds" for a logit-link model.
quasit: if TRUE, use a t multiplier rather than a normal multiplier when object is a quasi model.

Value

Invisibly returns an object of class s20xPrediction, which prints as the prediction table. It is a list with components

frame: data frame containing the predictions and intervals.
fit: prediction values, on the link scale for a glm fit.
se.fit: standard errors of predictions.
residual.scale: residual standard deviation, for an lm fit.
df: residual degrees of freedom, for an lm fit.
type: the scale of the results, for a glm fit.
cilevel: confidence level of the interval.

Details

For an lm fit the table holds fitted values, confidence intervals for the mean response, and prediction intervals for new observations. For a glm fit the intervals are confidence intervals, calculated on the link scale and then transformed according to type.

The columns of newdata are matched to the model variables by name when every variable is present. Otherwise they are taken in order, either one column per variable or one column per model term.

This is not a method for predict, so it does not change what predict() does for these models.

Examples


# Zoo data
data(zoo.df)
zoo.df = within(zoo.df, {day.type = factor(day.type)})
zoo.fit = lm(log(attendance) ~ time + sun.yesterday + nice.day + day.type + tv.ads,
             data = zoo.df)
pred.zoo = data.frame(time = 8, sun.yesterday = 10.8, nice.day = 0,
                      day.type = factor(3), tv.ads = 1.181)
predict20x(zoo.fit, pred.zoo)
#>   Predicted  Conf.lower Conf.upper  Pred.lower  Pred.upper
#> 1     6.758       6.567       6.95       5.738       7.778

# Peruvian Indians data
data(peru.df)
peru.fit = lm(BP ~ age + years + I(years^2) + weight + height, data = peru.df)
pred.peru = data.frame(age = 21, years = 2, `I(years^2)` = 2, weight = 71, height = 1629)
predict20x(peru.fit, pred.peru)
#>   Predicted  Conf.lower Conf.upper  Pred.lower  Pred.upper
#> 1   149.437     138.811    160.064     126.924     171.951

# A generalised linear model: predictions on the response scale
peru.glm = glm(BP ~ log(weight), data = peru.df, family = poisson)
predict20x(peru.glm, data.frame(weight = c(60, 70)))
#>   Predicted  Conf.lower Conf.upper
#> 1   124.477     120.711    128.360
#> 2   134.093     129.053    139.331
predict20x(peru.glm, data.frame(weight = c(60, 70)), type = "link")
#>   Predicted  Conf.lower Conf.upper
#> 1     4.824       4.793      4.855
#> 2     4.899       4.860      4.937