[R] Safe prediction does not work for bivariate polynomial terms?

[Xing Zhao]

Hi everyone,

R documents says the safe prediction is implemented, when basis
functions are used, such as poly(), bs(), ns()

This works for univariate basis, but fails in my bivariate polynomial setting.
Can anyone explain the reason?


The following is a small example.

set.seed(731)
x<-runif(300)
y<-runif(300)

f <- function(x, y) {  0.5+2*x+13*x^2-14*y^2+12*x*y+y }

z <- f(x,y)+rnorm(length(x))*0.2

# default orthogonal polynomials basis
mod <- lm (z ~ poly(x,y,degree = 2))

# raw polynomials basis
mod1 <- lm (z ~ poly(x,y,degree = 2, raw = T))

# data points to evaluate, just the first five data
new <- data.frame(x=x[1:5],y= y[1:5])

z[1:5]
[1]  9.796620 10.397851  1.280832  4.028284  4.811709

# two predicted values differ dramatically, and the orthogonal
polynomials basis fails
predict(mod, new)
        1         2         3         4         5
121.46776  40.85002  18.67273  31.82417  20.81673
predict(mod1, new)
        1         2         3         4         5
 9.981091 10.418628  1.223148  4.031664  4.837099

# However, the fitted.values are the same
mod$fitted.values[1:5]
        1         2         3         4         5
 9.981091 10.418628  1.223148  4.031664  4.837099
mod1$fitted.values[1:5]
        1         2         3         4         5
 9.981091 10.418628  1.223148  4.031664  4.837099


Thanks in advance
Xing