SEM - formative indicator

[Bradley Richfield]

Hello everybody,

I am currently trying to fit a measurement model in which my latent variable is not reflective but formative.
The latent variable ("Zusarb") represents the degree to which collaboration with one's colleagues takes place.
The four indicators ("ed1010a", "ed1010b, "ed1010c, "ed1010d") tell us how often certain forms of collaboration take place (never-several times a week on a six-point-scale).
According to Acock(2013) I fixed the error variance for Zusarb to be 0, and fixed one of the indicators to have a loading of 1.

However, when running the syntax, following error message occurs:

. sem (Zusarb <- ed1010b@1 ed1010a ed1010c ed1010d), var(e.Zusarb@0)latent(Zusarb)
model not identified;
no paths from latent variable Zusarb to observed variables
r(503);

I have done research for hours now - also in this forum - but could not find an answer to this problem.
Does anybody have an idea for a solution?


Best regards
Bradley Richfield

[Richard Williams]

The latent variable has to affect at least one variable. Really 2, or there is not much point in having it. You are basically imposing proportionality constraints on the xs, e.g. If the effect of x1 on y1 is twice as large as the effect of x2 on y1, then the effect of x1 on y2 will also be twice as large as the effect of x2 on y2.

[Patrick Abi Nader]

You made a typing error I think.

sem (Zusarb <- ed1010b@1 ed1010a ed1010c ed1010d), var(e.Zusarb@0)latent(Zusarb)

Is supposed to be

sem (Zusarb -> ed1010b@1 ed1010a ed1010c ed1010d), var(e.Zusarb@0)latent(Zusarb)

I am unsure of this part of your code

var(e.Zusarb@0)latent(Zusarb)

what is your aim with it?

[daniel klein]

On serious conceptual flaws of formative models, see the excellent article by Edwards (2011).

Best
Daniel

Edwards, J. R. (2011). The Fallacy of Formative Measurement. Organizational Research Methods, 14, 370-388.

[Bradley Richfield]

Big thanks for your replies!

The part var(e.Zusarb@0) is - according to Acock (2013) - necessary because the error variance for the formative latent variable is supposed to be fixed at 0.

With the option latent(Zusarb) I want to tell Stata that "Zusarb" is actually being the latent variable in the model, not any other measured variable in the data set which begins with a capital letter.

Indeed, I succeeded fitting respectively identifying the model transforming it into a MIMIC model - there were two measured variables in my data of which can justifiably be supposed that they are effected by my latent variable.

Once again - thank you

[Richard Williams]

I've always liked this 1983 paper by Hauser, Tsai and Sewell for its discussion of whether SES is best modeled as a composite or as a factor.

http://citeseerx.ist.psu.edu/viewdoc...=rep1&type=pdf