The argument in the predict.lm function
Head function predict.lm
predict.lm <- function (object, newdata, se.fit = FALSE, scale = NULL, df = Inf,
interval = c("none", "confidence", "prediction"), level = 0.95,
type = c("response", "terms"), terms = NULL, na.action = na.pass,
pred.var = res.var/weights, weights = 1, ...)
I am having trouble understanding how an argument is evaluated pred.var = res.var/weights
.
I know it means the variance of the residuals, but it is passed as a variable, as opposed to a character string, where it can subsequently be read and translated.
The reference for this feature is not explicit for this issue, which I think may be related to how R works, and not so much with this particular feature.
Somehow, the argument pred.var
will deflect from the remainders by default - sd(fit$residuals)^2
- but I can't see how.
source to share
What you see here is the R mechanism for lazy evaluation of function arguments. Take a look at a simplified example:
lazy_arg <- function(x, y = z) {
z <- sum(x * x)
x / y
}
lazy_arg(1:5, y = 10)
#[1] 0.1 0.2 0.3 0.4 0.5
lazy_arg(1:5)
#[1] 0.01818182 0.03636364 0.05454545 0.07272727 0.09090909
The second call will obviously not work in languages that evaluate arguments immediately after the call. Instead, R simply stores the "recipe" y = z
and evaluates it only when it is actually used y
. Of course, when I write a function like this, I have to make sure it is z
correctly defined before using it y
, otherwise I would provide a great opportunity to shoot myself in the foot:
bad_arg <- function(x, y = z) {
if (runif(1) > 0.5) z <- 1
x / y
}
set.seed(112)
z <- 1e5
bad_arg(1:5)
#[1] 1e-05 2e-05 3e-05 4e-05 5e-05
bad_arg(1:5)
#[1] 1 2 3 4 5
If you are wondering why this happened, then how does R variable lookup work (in short, an unintentional top-level collision). So flexibility brings value here.
However, predict.lm
this is quite handy in case , as it provides a sensible default in place, and it is actually calculated much later and depends on other arguments.
For more information and options, see Hadley Wickham's "Advanced R" Lazy Assessment: http://adv-r.had.co.nz/Functions.html
source to share