[R] please help with estimation of true correlations and reliabilities

[John Fox]

Dear Jens,

One of the nice features of maximum-likelihood factor analysis, implemented
in factanal(), is that you can test for the number of common factors.

More generally, the estimated uniquenesses will depend upon the number of
common factors. Using sem() requires a stronger initial specification -- not
only the number of common factors, but a prior specification of which
variables load on each. If you make the model too general, it won't be
identified.

Furthermore, although I suggested that you use the uniquenesses to assess
unreliability, the justification for this is clearest in models in which
each variable loads on just one factor. I guess it would help to know a bit
more about the application.

Regards,
 John

> -----Original Message-----
> From: Jens Oehlschlägel [mailto:joehl at gmx.de] 
> Sent: Friday, May 14, 2004 8:52 AM
> To: jfox at mcmaster.ca
> Cc: r-help at hypatia.math.ethz.ch
> Subject: Re: [R] please help with estimation of true 
> correlations and reliabilities
> 
> 
> Dear John,
> 
> I checked for the uniquenesses from factanal() and found out 
> that these estimates need an a priori assumption about the 
> number of latent factors (and they depend strongly on that 
> assumption). What I want is an estimate of reliablity 
> independent of such an assumption.  Is it possible to specify 
> such "structure" to sem or another SEM software? (no 
> assumption about number of factors, no assumptions about 
> which variables group together).
> 
> Best regards
> 
> 
> Jens Oehlschlägel
>