Re: st: SEM becomes unidentified when introducing single item control variables
Date
Tue, 15 Jan 2013 19:45:06 +0100
Hi:
The model is undefined. You need to set constraints linking the single
indicator (e.g,. x1) of the latent (X), as follows:
sem (y <- X) (X ->x1@1), reliability(x1 .80)
Where reliability < 1 > 0, is your theoretical constraint of how much
true variance x1 captures.
See "help sem reliability"
If course, if you set x1 = 1 you are assuming that x1 is perfect
indicator of X.
HTH,
J.
__________________________________________
Prof. John Antonakis
Faculty of Business and Economics
Department of Organizational Behavior
University of Lausanne
Internef #618
CH-1015 Lausanne-Dorigny
Switzerland
Tel ++41 (0)21 692-3438
Fax ++41 (0)21 692-3305
http://www.hec.unil.ch/people/jantonakis
Associate Editor
The Leadership Quarterly
__________________________________________
On 15.01.2013 15:21, Johannes Kotte wrote:
Dear fellow researchers,
I would be grateful for advice with the following problem: I have
created a very simple SEM (let's call it 'model 1') that works fine
(see below for code). It contains a latent dependent variable called
PRAXREL and a latent independent variable called BKA. Moreover, it
contains latent control variables called KVSENIOR and KVL. As I said,
model 1 works fine (identified, good fit).
However, the model becomes problematic when I introduce single-item
latent variables (CV1, CV2, CV3, CV4) as control variables ('model2').
In this case Stata iterates forever saying “not concave”.
WHAT COULD BE THE REASON? I tried many different setups of the model
(incl. constraining the path coefficients of the CV to 1 or setting
the reliability of the CV to 0.9 or 0.5) but none of them really
worked unless I delete at least some of the CVs.
The following might be interesting: (i) If I let Stata iterate 15
times and take a look at the output, I find that sometimes the
standard errors of CV1, CV2, CV3 and CV4 are extremely high. (ii)
Moreover, I found that pairwise correlation of the variables shows
that they are mostly correlated - at least at the 10% level, sometimes
even 1%. Might there be a collinearity problem?
