[Rd] Bug in poly() (PR#11243)

[russell-lenth at uiowa.edu]

Full_Name: Russell Lenth
Version: 2.6.2
OS: Windows XP Pro
Submission from: (NULL) (128.255.132.36)


The poly() function allows a higher-degree polynomial than it should, when
raw=FALSE.

For example, consider 5 distinct 'x' values, each repeated twice.  we can fit a
polynomial of degree 8:

=====
R> x = rep(1:5, 2)
R> y = rnorm(10)
R> lm(y ~ poly(x, 8))

Call:
lm(formula = y ~ poly(x, 8))

Coefficients:
(Intercept)  poly(x, 8)1  poly(x, 8)2  poly(x, 8)3  poly(x, 8)4  poly(x, 8)5  
   -0.07790      0.58768      0.30147     -0.18237      0.64779     -0.04444  
poly(x, 8)6  poly(x, 8)7  poly(x, 8)8  
   -0.22694     -0.07500      0.17235  
=====

If we specify 'raw=TRUE' in the same example, we are limited (as we should) to a
degree-4 polynomial:

=====
R> lm(y ~ poly(x, 8, raw=TRUE))

Call:
lm(formula = y ~ poly(x, 8, raw = TRUE))

Coefficients:
            (Intercept)  poly(x, 8, raw = TRUE)1  poly(x, 8, raw = TRUE)2  
                 7.3958                 -14.0150                   8.2784  
poly(x, 8, raw = TRUE)3  poly(x, 8, raw = TRUE)4  poly(x, 8, raw = TRUE)5  
                -1.9502                   0.1597                       NA  
poly(x, 8, raw = TRUE)6  poly(x, 8, raw = TRUE)7  poly(x, 8, raw = TRUE)8  
                     NA                       NA                       NA  
=====

We do get an error with an even higher degree:

=====
R> lm(y ~ poly(x, 10))
Error in poly(x, 10) : 'degree' must be less than number of points
=====

Looking at the code for poly(), I think the problem is that it should check the
'rank' result from its call to qr().

[The above results are using Windows XP Pro, but I verified this bug under Linux
as well (x86_64, dual core).  It seems pretty obvious to me that this bug is not
platform-dependent.]