Hello M. F.,

I wasn't very clear about how to interpret your query, so this response may not be as helpful as if I better understood your study.

Part of the answer depends on what is being tested: Are you looking for significance of individual path coefficient estimates (e.g., different from zero), or are you looking at indices of overall model-data fit?

If the former, then usual regression-based answers could include: (a) suppression effects; (b) multicollinearity; (c) sample size; (d) excess of missing/imputed data points; and/or (e) model mis-specification that changes as you add portions rather than look at subsets in isolation.

If the latter, then the comparisons are not "apples to apples," as the overall model becomes more complex (and more relationships are to be estimated) as you add portions to the analysis.

As is, portions of the model are nearly saturated, whereas a few others are not; it's possible that this, too, could have some impact on what your analyses reveal.

Good luck with your work.

Any tips about what can I do to improve GFI ? Now I have all other indices satisfactory but can't have a good value for GFI even doing the modifications proposed by amos

Hi M.F. Marquetto

your model is extremely complex. So complex that you probably won't be able to fit it in one phase of a study. To show that, I re-generated your model with the browser software dagitty.net wich allows you to draw a path diagram and print the "testable implications" -- which are in total evaluated by the model chisquare test and described by the fit indices. The lower part of this post shows the list of these restrictions. The signs "⊥" mean "independence" (i.e., zero correlation) and the bar "|" is a "condition"-sign. For instance, your model assumes/predicts that A ⊥ H | G, which means that A does not correlate with H when you control G. The rules are simple path tracing, path blocking/d-separation (google for these terms).

Even leaving out the downstream part (G, H, I, J) would still imply a complex model, however, with some chance to get it done. But the relationships between C to F are still complex. Ideally, such a huge model could/would be developed and tested via successive steps of primary studies, each with some exploratory and confirmatory part. Perhaps starting with an upstream model, in which some of the mediators are collapsed/excluded could be a start.

A different issue (beyond fit) is in howfar even a fitting model would support YOUR model. Take for instance the triangular part including C, D, and E. Every one of the involved effects (e.g., from C to D) is "shielded" -- that is has at least one common cause affecting both. This implies that different causal directions cannot be differentiated. In simple words: you could turn the arrows and you would not notice it. That is, even if a model has testable implications that doe not mean that all parts are tested. There are always equivalent models and inspecting which parts have which equivalent structures should be a routine assessment in any SEM study. This can easily done by using the TETRAD software (http://www.phil.cmu.edu/tetrad/ ). I included some references but I attached them to not completely blow this post :)

I know that this is not what you wanted to hear. But if SEM gets often difficult then because IT IS difficult.

Best,

Holger

- - - - - - - - - - - - - - - - - - - - - -

• B ⊥ I | G
• B ⊥ I | C, D, E
• B ⊥ A
• B ⊥ C
• B ⊥ F
• B ⊥ G | C, D, E
• B ⊥ H | G
• B ⊥ H | C, D, E
• B ⊥ J | G
• B ⊥ J | C, D, E
• I ⊥ A | C, D, E
• I ⊥ A | G
• I ⊥ C | G
• I ⊥ D | G
• I ⊥ E | G
• I ⊥ F | C, D, E
• I ⊥ F | G
• I ⊥ H | G
• I ⊥ J | G
• A ⊥ G | C, D, E
• A ⊥ H | G
• A ⊥ H | C, D, E
• A ⊥ J | G
• A ⊥ J | C, D, E
• C ⊥ F | A
• C ⊥ H | G
• C ⊥ J | G
• D ⊥ H | G
• D ⊥ J | G
• E ⊥ H | G
• E ⊥ J | G
• F ⊥ G | C, D, E
• F ⊥ H | G
• F ⊥ H | C, D, E
• F ⊥ J | G
• F ⊥ J | C, D, E
• H ⊥ J | G