There are some doubts about the reliability when using sum scores (means) of respondents answers in scientific measurement in social sciences. I agree that summing responses in previously developed scales could lead to unproper results or just too superficial results.
The major difference in SEM (Structural Equation Modeling) is that the loadings from the indicators in the latent variable are different that in sum scores.
What if a scientist theoretically justify the construction of a scale that some indicators (questionnaire responses) are summed?
It would be regarded as a methodological mistake?