Negative regression co-efficients in SEM
Posted on
LearnSEM8
Joined: 08/14/2012
Forums
Hi!
I am new to Structural Equation Modeling but know that it is a confirmatory technique and one has to specify his\her own model and check if the data is supportive of the same. However, I have run into some illogical results using some very logical linkages in my model. In my model, one of the most logical linkages has a low negative regression weight which will be difficult to explain to the client. In such cases, would anyone recommend fixing the regression weight for this linkage to a low positive value (make the results face-valid)?
Thanks in advance!
diagram, net effects, mirror images
The negative weight may be settable to zero without significant loss of fit. Try that and see.
You could also set the lbound to -.0001 and that will force the model to build a positive path if it wants to explain covariance on that path. Again: see what happens to fit.
Often there are equivalent mirror solutions: so everything is opposite in sign to what you are expecting. That just means that what you are thinking of as, say, "schooling" is being modelled as "lack of schooling" - the same scale reversed.
Because SEM generates net effects across the full set of paths between variables, you can get negative residuals when a cause has positive paths through other variables.
Two possibilities come to mind in that case: Either this cause has two factors buried inside it - the other paths are accounting for its positive effects leaving genuine negative effects left over, or else there is a better fitting model.
All of this is much easier to think about with a model, so try and include a drawing of your model.
Log in or register to post comments
In reply to diagram, net effects, mirror images by tbates
Negative regression co-efficients in SEM
Thanks! I tried setting the path co-efficient to 0 and it worked well.
Log in or register to post comments