For some work I'm doing, I need to apply PCA to data containing missing values. So I dug up some old code that I had around, which implements Stock and Watson's EM algorithm (detailed in the 2002 JBES article). I'm attaching a script with the function and an example if anyone's interested. Of course, it'd be nice to make this procedure more readily available to gretl users. I see four (not mutually exclusive) ways of doing so: 1) modify our existing "pca" command to handle missing values; 2) modify our existing "princomp" function to handle missing values; 3) create a small self-contained function package; 4) integrate the code into the existing "staticfactor" function package. Which one do you this is best?

For some work I'm doing, I need to apply PCA to data containing missing values. So I dug up some old code that I had around, which implements Stock and Watson's EM algorithm (detailed in the 2002 JBES article). I'm attaching a script with the function and an example if anyone's interested. Of course, it'd be nice to make this procedure more readily available to gretl users. I see four (not mutually exclusive) ways of doing so: 1) modify our existing "pca" command to handle missing values; 2) modify our existing "princomp" function to handle missing values; 3) create a small self-contained function package; 4) integrate the code into the existing "staticfactor" function package. Which one do you this is best?

On 14/11/2024 16:48, Cottrell, Allin wrote: >> >> Of course, it'd be nice to make this procedure more readily available to >> gretl users. I see four (not mutually exclusive) ways of doing so: >> >> 1) modify our existing "pca" command to handle missing values; >> 2) modify our existing "princomp" function to handle missing values; >> 3) create a small self-contained function package; >> 4) integrate the code into the existing "staticfactor" function package. >> >> Which one do you think is best? > > Not sure about "best", but it seems it wouldn't be super-difficult to > extend pca and princomp() to handle NAs in the manner of your hansl > function. That'd be my preference too. That said, extending princomp() would be rather easy. The "pca" command, on the other hand, also reports some additional statistics (basically, the eigenvalues) that are not readily available with the EM algorithm, so this would require some more thought. PS: in my previous message, I typed "this" instead of "think". I'm getting worse with age, sorry.