0
$\begingroup$

I am new to SEM. My aim is to determine which factors (several intrisic and serveral extrinsic) influence school participation in children with disabilities. Some of my latent variables are scores to quetionnaires, which have a lot of items (around 70 - 80). These items are divided into subscales.

  1. I wanted to know if I can use subscales scores as "observed" variables. The answer appear to be yes. However, I never saw that in a paper. Do you have any references on that?

  2. I also read that there should be at list 3 observed variables for a latent varaibles. Are there references for that as well ? and references stating that using only two can be ok?

  3. I fear that my model will be to much complex and won't converge. If is is the case, is it appropriate to run 2 models (1 for intrisc and the other one for extrinsic faactors). What king of problem this approach could pose?

Any answer and comment on these questions would be much appreciated !

Thanks !

$\endgroup$

1 Answer 1

0
$\begingroup$
  1. You can use subscale scores or item parcels (see the literature on item parceling in SEM) as indicators. However, if the subscales measure distinct, only modestly related factors (i.e., if the overall scale is multidimensional), then your single factor resulting from that strategy may be difficult to interpret, the standardized factor loadings may not be very strong, and the error terms may contain systematic scale-specific variance in addition to measurement error variance.

  2. This recommendation is given in many SEM textbooks. The main reason is probably that a single-factor model with just two indicators and unequal loadings is underidentified unless there are other variables in the model with which the factor is substantially correlated. In other words, a 2-indicator factor can be OK as long as the factor is correlated with at least one other variable in your model ("variable" here could refer to either an observed variable such as age or gender or another latent factor). A model with a single factor, 3 indicators, and unequal loadings is identified "per se" (as long as the indicators of this factor have substantial positive covariances). Also, models with just two indicators per factor appear to be more prone to Heywood cases (improper solutions). That being said, there are also many situations in which models with just 2 indicators per factor work just fine.

  3. You could do that. One downside maybe that with separate models, you could not examine all predictors in a single model and thus not fully study their potential redundancies and/or interactions.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.