Hello Xuechen,

such a huge model is impossible to run as the number of causal assumptions is so huge that the fit will be terrible. So I see two options:

a) either stay with all variables and run an exploratory TETRAD analysis. This will result in a single path diagram that shows the whole class of equivalent models. In this diagram, some pairs of variables will be unconnected--for these you can be quite sure that they have no causal connections. Other will be connected but in an ambiguous way (i.e., you cannot tell based on the data whether x causes y or the reverse or an unobserved confounder affects both). For a minority (depending on the structure), you may find clear symbols in the diagram pointing to a clear direction. This means that this effect emerged in this direction in all models of this equivalence class and that a different connection would have been in contradiction with the data. In most cases, 70% of pairs are totally ambiguous, 20% are partially-ambiguous and only 10% are "oriented".

This exploration is a good starting point for planning further research (i.e., for considering how to nail the ambiguous parts) or at least to inform the literature about the thin ice. The issue of equivalent models is rarely dealt with.

b) You could indeed focus on a few key effects and try to theoretically identify necessary control variables or perhaps even instruments (i.e., variables that predict your IV but are not directly related to the DV.

It should be noted that interestingly, our style of estimating these huge models strongly deviates from practices for instance in econometrics who much more focus on specific effects and try everything to hammer the inherent endogeneity out of it....

HTH

--Holger

You could indeed focus on a few key effects and try to theoretically identify necessary control variables or perhaps even instruments (i.e., variables that predict your IV but are not directly related to the DV.

I agree with Bazar Dzhumaevich Ulugov suggestion

Let me preface my answer by saying that I have served as Associate Editor at several journals in the AIS "Basket of Eight" journal list. Your question sounds like a simple question; however, it's not clear to me whether you are asking a question about "how do I conduct an empirical study using an old, well-validated model like UTAUT that makes a contribution to theory and is publishable?" or whether your question is "how can I get (covariance-based) SEM to work when I have 9 independent variables, plus 7 mediating variables, and (at least) one dependent variable." I think that your question is framed as the latter (e.g., "how can I get SEM to work ...?"), but when you mention looking at previous studies, I believe that your question shifts closer to the first question. So let me answer the first question. The reason that PhD students and faculty do research is in order to learn something new that has not been done before, and preferable something that is interesting and surprising. If you want to be able to publish your empirical study in a good journal, it is NOT sufficient for you to simply show that all (or most) of your hypotheses are supported; rather, it is important for you to study something that contributes theoretically to what we already know in the IS discipline. If you are just re-using an old model that has been well-validated in the past (UTAUT), then what is new about that? One possible answer is to say: "I don't care about publishing my research; I just want to get experience in collecting data and validating a model in SEM." That is an acceptable answer, especially if this is a pre-dissertation project that you are doing for a class assignment, or to complete a requirement in your university's PhD program (for example, some PhD programs require students to complete a project by their 2nd year, which is not necessarily intended for publication).

However, if you want to publish your work in a good journal later, then you need to focus on the question: "what is new and interesting about my study?" Even if you are the first researcher to apply UTAUT to a new technology that has not been studied before, it is unclear whether that is a contribution to theory. You could probably publish it in a lower-quality journal, but not in a medium-to-high quality journal. So, you need to ask (and discuss with your advisor and colleagues) the question of "what makes my study interesting and novel?"

I believe that this may be the reason that, for the previous studies you reviewed, that some authors did NOT include all 9 of the UTAUT predictor variables (because it was not interesting for them to do so).

One final question. You mentioned that you are using SEM. I assume that you are referring to covariance-based SEM (e.g., Lisrel, AMOS, or another covariance-based software package that computes all the variables Goodness of Fit metrics). Many researchers do not understand what these Goodness of Fit metrics actually mean. The fact is that, if your model has most or nearly all of the possible connections between constructs with paths between them, then your model is "guaranteed" to have high Goodness of Fit, so that the fit metrics become nearly meaningless. So, if all 9 of your Independent Variables connect to all 7 of your mediating variables, and then the latter all connect to your Dependent Variable, you will have high Goodness of Fit, by definition. In fact, your SEM model may even be "perfectly identified" so that the software can't even compute the Goodness of Fit metrics.

So, to conclude, if you want to publish your study in a decent journal, I recommend that you discuss with your advisor and colleagues the question of "what makes this study interesting?". Since you are using a model that has been well-validated, then you have a big challenge if you want to show something new. This is a challenge even if you are applying the UTAUT model to a new technology, or even in a different type of user context. Lastly, consider the answer to another question: "what will be my response rate, and how do I know if there is response bias?" (in terms of the types of respondents who do the survey vs. do not do it).

Good luck!

Dear Prof. Mj Gallivan

Thank you so much for the comments and suggestions.

Actually, after presenting my project to the advisor and fellow students, I received feedback almost the same as yours—even the two "real" questions between my expression. Finally, I am now aware that It's not like doing a census to put all potential variables on the stage, but find those new and important ones. Also, there are problems with response rate and Goodness of Fit, as you pointed out.

Instead of merely looking at the model, I am now reconsidering what on earth the story I will tell and impress the readers. I think that's the essential thing to think about before the model.

Your answers about my previous researchers did NOT include all predictor variables, and about Goodness of Fit are highly appreciated. I've heard a lot about "This way is right, and you just follow it. No asking." Seldom do I receive replies on "Why that way doesn't work."

Thanks again for your kindness and patience.

Best Regards,

Xuechen