Strange behavior with fitted lm objects

["David Kane - GEODE <David Kane" <a296180@mica.fmr.com>]

Hi,

I see some strange behavior when generating the fitted values from an lm
object. Consider:
> version
Version 6.0 Release 1 for Sun SPARC, SunOS 5.6 : 2000 
> set.seed(9)
> tempDF <- data.frame( y = rnorm(6), f = as.factor(paste("level", rep(1:2,3), 
> sep = "")))
> tempDF
           y      f 
1 -0.2009222 level1
2  0.4997202 level2
3  0.3564666 level1
4  0.4703904 level2
5 -0.4061262 level1
6 -1.6078728 level2
> fitted(lm(tempDF$y ~ tempDF$f))
           1          2           3          4           5          6 
 -0.08352725 -0.2125874 -0.08352725 -0.2125874 -0.08352725 -0.2125874

At this point, everything looks OK. The fitted values are the mean of the
response at the level of each factor.

> tapply(tempDF$y, tempDF$f, mean)
      level1     level2 
 -0.08352725 -0.2125874

Unfortunately, the results are not *identical* within each level. Consider:

> unique(fitted(lm(tempDF$y ~ tempDF$f)))
[1] -0.08352725 -0.21258740 -0.08352725 -0.21258740 -0.21258740
> fitted(lm(tempDF$y ~ tempDF$f))[1] == fitted(lm(tempDF$y ~ tempDF$f))[3]
 1 
 F
> options(digits = 15)
> fitted(lm(tempDF$y ~ tempDF$f))[1:3]
                   1                  2                   3 
 -0.0835272529974701 -0.212587401068558 -0.0835272529974702


Any insights on this problem would be much appreciated.

Thanks,

Dave Kane

-- 
David Kane
Geode Capital Management
617-563-0122
david.d.kane@fmr.com