get_dummy() handling of vectors containing zeros

[guilhemchalancon]

Hi,

I've spotted an issue with get_dummy(), when running plspm() with a model comprising ordinal variables with "0" as one of the categories.

Let's assume we have a variable in a block q such that Xblocks[[q]][,p] is made of numbers between 0 and 4 (i.e. 5 categories).

Basically, get_weights_nonmetric() will fail at handling such variable, stopping at that line: get_ord_scale(Z[,q], Xblocks[[q]][,p], aux_dummy)

The reason is that in that case, aux_dummy lacks 1 column, precisely the column corresponding to the category "0".

So I looked at how dummies are obtained, and found out that get_dummy() is at the root of the problem.

get_dummy <- function (x) 
{
    n = length(x)
    p = max(x, na.rm = TRUE)
    Xdummy = matrix(0, n, p)
    for (k in 1:p) {
        Xdummy[x == k, k] = 1
    }
    if (any(is.na(x))) {
        Xdummy[which(rowSums(Xdummy) == 0), ] <- NA
    }
    Xdummy
}

As you can see, the number of columns p is based on the maximum value of x. But this won't work when 0s are present in x (p would need to be max(x)+1).

Very simply, defining p as p = length(unique(x)) ensures that the function works every time.

Best,
G

[guilhemchalancon]

The same reasoning applies for get_ord_scale where also p is defined as max(x, na.rm=T)

Also, I wonder whether having 0s in categories like I did actually makes sense when computing the ordinal scale, so perhaps a conditional test should be introduced to prevent that users define ordinal variables like that?