Using total score from multi-scale instrument in structural equation modeling

Question

I wonder if you can help me to get the right answer to a question about structural equation modeling. Imagine someone trying to validate a questionnaire using component factor analysis that highlighted 5 subscales. Is it right to use the total score of the questionnaire in a model created by structural equation modeling? That is, would it be correct to consider that the questionnaire is 'validated' and use its total score? Would you think that this model is right or valid or should the subscales be considered as observed variables and total score as latent variable in the model?

Answer 1

(This is a very late answer)

Is it right to use the total score of the questionnaire in a model created by structural equation modeling? That is, would it be correct to consider that the questionnaire is 'validated' and use its total score?

With commonly used scales that have previously been validated in your studied population, it is usually not necessary to validate the factor structure in your data. You might consider doing confirmatory factor analysis to check factorial invariance if you are repeatedly using the scale over time, or if you want to find out whether it's doing the same things in different subgroups of your sample (e.g. males and females). However, if the scale is not previously validated in your studied population, or not previously validated at all, you will probably have to go all the way with EFA, CFA, and scale validation.

should the subscales be considered as observed variables and total score as latent variable in the model?

If your scale is validated, and you choose to use subscales or total scores (I assume you derive these from some scoring procedures rather than SEM based on your data) to build subsequent SEM to test your theory, they will all be observed variables. An alternative to using the total score as observed variable, is to use the 5 subscale scores to build a measurement model that gives you a latent variable that represents the underlying construct the scale was supposed to measure. This would be conceptually similar to the total score (e.g. if you have a scale that measures depression, both could indicate overall levels of depressive symptoms), except that if you use the total score, you assume no measurement error (as observed variables do), whilst in SEM style, the use of latent variable will incorporate measurement error.