as.character(as.item(..., missing.values = 0, ...)) returns NA instead of a string

[bixiou]

Since an update in memisc (somewhere between 0.99.22 and 0.99.28), the following code returns an error:

foo <- as.item(c(0,1,1,-1), missing.values= 0, labels = structure(c(-1,0,1), names=c('Yes', 'PNR', 'No')))
bar <- as.factor(as.character(foo))
bar <- relevel(bar, 'PNR')

Before the update (I had 0.99.22), we had: foo = c('PNR', 'Yes', 'Yes', 'No') while now we have foo = c(NA, 'Yes', 'Yes', 'No').
This is a weird behavior because the point of memisc is to distinguish missing values from NA.
Can you revert to the previous behavior?

[bixiou]

PS: I corrected a typo in my post. With the typo there was no error. Now, there is the error.

[melff]

This is not a bug, but a feature. Values equal to 0 are translated into NA by as.character() because 0 is marked as a missing value. Such an automatic translation into NA is the whole point of the "item" class, because otherwise there is no way in R to distinguish between "non-NA" missing values and NA.

If you want to retain "PNR" as a valid factor level, you have to declare 0 as valid.
E.g. by

missing.values(foo) <- NULL
bar <- as.factor(foo)
bar <- relevel(bar, 'PNR')

or

valid.values(foo) <- -1:1
bar <- as.factor(foo)
bar <- relevel(bar, 'PNR')

or

bar <- as.factor(include.missings(foo))
bar <- relevel(bar, 'PNR')

If in earlier versions of memisc as.character() did not take into account the missingness status of certain values then that was a bug.

[bixiou]

Well, this is a pity because this "bug" was really convenient: it allowed me to run regressions without dropping data ('PNR' were considered as a response category, which it is) while allowing me to treat these 'PNR' as missing in the analysis (i.e. have
a special status for descriptive stats, graphs...).
So I'll keep using memisc 0.99.22 which does exactly what I need.

[melff]

I think

bar <- as.factor(as.character(include.missings(foo)))

or

bar <- as.factor(as.character(foo,include.missings=TRUE))

will do what you want it to do with newer versions of memisc.

[bixiou]

In this particular case, yes, because I convert the variable in a factor anyway. But I noticed other parts of the code where the update causes issues, namely regressions. I'd need to transform each item variable to a factor in my regressions with the new memisc version. Easier (and more compact code) to keep the old one.