CODEWITHSUNDEEP
rmissing-datanlme
qoheleth
qoheleth

Reputation: 2309

nlme: incorrect factor names with NA?

Basically, I am running nlme::lme with a missing value. The model fits fine with na.action=na.omit, but how come the name of the fitted/residuals/coef all seem to be shifted by a row?

## Generate data ---------------------
X1=gl(2,4)
X2=gl(2,2,8)
Y=rnorm(8)
dat=data.frame(Y=Y,X1=X1,X2=X2)
dat

## missing value -------------
mis.dat=dat
mis.dat[3,"Y"]=NA
mis.dat
> mis.dat
            Y X1 X2
1 -0.06845332  1  1
2  0.89169085  1  1
3          NA  1  2
4  1.88997449  1  2
5  0.95912879  2  1
6 -0.64049400  2  1
7 -0.23354948  2  2
8 -0.66869350  2  2


## Fit model -----------------------
model=nlme::lme(Y~1,random=~1|X1/X2,data=mis.dat,na.action=na.omit)
summary(model)

## Notie the names -------------------
fitted(model)
> fitted(model)
       1/1        1/1        2/1        2/1        2/2        2/2       <NA> 
0.67179438 0.67179438 0.67179439 0.02855517 0.02855517 0.02855517 0.02855517 
attr(,"label")
[1] "Fitted values"

#model$coef$random
#resid(model)

Notice how the names of the fitted value? Shouldn't there be a 1/2 in the 3rd position, and the names after that shift one position to the right, thus eliminating the NA?

Upvotes: 0

Views: 180

Answers (1)

Roland
Roland

Reputation: 132706

You found a minor bug. Look at nlme:::fitted.lme to see what happens:

The fitted values are extracted from the model using 

object[["fitted"]]
      fixed        X1        X2
1 0.6014526 0.1686912 0.1686912
2 0.6014526 0.1686912 0.1686912
4 0.6014526 0.1686912 0.1686912
5 0.6014526 1.0342140 1.0342140
6 0.6014526 1.0342140 1.0342140
7 0.6014526 1.0342140 1.0342140
8 0.6014526 1.0342140 1.0342140

Note how there are 8 fitted values even though observation 3 was omitted from the fit due to the missing y-value and shouldn't be there. The names are then created from 

object[["groups"]]
  X1  X2
1  1 1/1
2  1 1/1
3  1 1/2
4  2 2/1
5  2 2/1
6  2 2/2
7  2 2/2

Note how there are only 7 names. The NA gets introduced, when match is used.

Ultimately the problem is in lme, which should only return 7 fitted values. However, I don't have time to find out how this could be fixed. Feel free to report it.

Upvotes: 1

Related Questions