fitglme error says data is not full rank, but it is.

Hi,

I'm trying to fit a glme to this data (there are over 600 rows).

The line of code looks like this:

glme = fitglme(TABLE,'X ~ A + B + C + D + E + F + G + H + I + J + K + (1|ID)','Distribution','Binomial')

It returns the following error:

Fixed Effects design matrix X must be of full column rank.

However, when I check the rank of the matrix, it is full rank:

>> test=table2array(TABLE);
>> rank(test)
ans =
    13
>> size(test,2)
ans =
    13

So I'm quite puzzled, it is the first time I try a glme using Matlab instead of R...

Answer 1

[Jeff Miller]

I'm guessing: (1) your ID column is a unique identifier rather than an actual numerical predictor value, and (2) you only have one row for each value of ID.

If those two guesses are right, then you should omit the (1|ID) term from the model. That term tells fitglme to set up dummy variables to code that column, and the predictor matrix with these dummy variables is not full rank if you only have one row for each value of ID.

HTH

[Luis J Gilarranz]

Both of your two guesses are correct.

However, omitting (1|ID) and removing ID from the table didin't work. I've also checked that the table doesn't have any other variable other than the ones I'm using, and ther's no NaN values.

[Jeff Miller]

Oh, sorry that wasn't helpful, and now I can't really think what the problem might be. Speculating wildly, it might be informative to look at the correlation matrix of columns A-K, and/or a principal components analysis of those columns. Perhaps fitglme checks rank with a different tolerance than rank's default and the predictor matrix is near enough to singular for fitglme to complain.

[Luis J Gilarranz]

That seems to be the issue. It works now. Thanks a lot.

[Haidong Zhao]

Hi Luis

I have a similar problem when running fitglme. how do you please fix thi problem? Thanks.