🚀 go-pugleaf

In <4791092@mtsg.ubc.ca> John.B.Collins@MTSG.UBC.CA writes:

>Flame as you wish, but you didn't explain what's wrong with factoring
>a matrix of dichotomous data.  They are NOT in the same class as the other
>examples you listed (which I agree would be nonsensical to factor).  Bearing
>in mind that SAS, SPSS and other packages yield correlations among dichotomous
>variables which reduce to Pearson Corrs, what then is the problem?
>John

It's not wrong to factor phi correlations, but it may mislead you if
you are not very alert to the fact that the maximum value phi can
attain is limited by the marginal distributions.   Variables at the
opposite extremes of the range (e.g. with mean scores 0.95 and 0.05)
will never appear to be highly correlated either to your eye or to a
factor analysis program.   The net result tends to be factors that
relate to the means of the variables, not to underlying dimensionality
(whatever that might mean).   The phi/phimax correction that is
sometimes done typically makes the matrix non-Gramian, which takes us
back to negative eigenvalues.
-- 
Steve Blinkhorn <steve@prd.co.uk>