The knowledge of independent (IV), dependent (DV) and moderating (MV) variables may also come from observation and experience but should be supported theoretically. If phenomena is new, as in your case, just use imaginative power. Take the IV out of the model and see what happens to DV.  Similarly, take MV out of the equation and see whether the effect of IV on DV increases or decreases.

In practice, I agree it is difficult to determine which one is IV and which one is DV. The problem is even more pronounced in social sciences when dozens of variable simultaneously affect behaviour and no experimental work is possible. However, to make convincing case, you have either to borrow theory from some other neighboring domains or at least produce compelling rational for the explanatory mechanism which exist between/among variables. The following papers may help:

Davis, J. P., Eisenhardt, K. M., & Bingham, C. B. (2007). Developing theory through simulation methods. Academy of Management Review, 32(2), 480-499.

Weick, K. E. (1989). Theory construction as disciplined imagination. Academy of management review, 14(4), 516-531.

Gioia, D. A., & Pitre, E. (1990). Multiparadigm perspectives on theory building. Academy of management review, 15(4), 584-602.

It is not easy to tell. But there is one way, in your data you can apply path coefficient analysis, and by their results you know that how much a variable have effect on the dependent variable also you can tell that how much a variable has effect on dependent variable through another variable.

Read detail about path coefficient analysis.

I hope that will help.

Thanks

I agree with Claudia Abreu Lopes that the distinction between an IV and a moderator variable is a matter of theory. Further, if your theory is about an interaction effect, there is not necessarily any need to define a moderator variable -- you might simply have two IV's that affect each other:

Y = X + Z + X*Z

Thank you for all the answers and suggestions. I'm planning to do the empirical testing part later down the line, so I'll broaden the question a little. 

I'm currently in the qualitative phase of formulating a theoretical model for a new phenomenon. Let's say I already have an idea what's the DV. Hypothetically, I identified two variables which may either be both IV (with or without interaction) or one IV and one MV (no clue which is IV and which is MV) to my intended DV. 

Can I (or should I?) narrow down to one option without any reference (I'm proposing a new variable) nor empirical testing (I don't have the data yet) and only qualitative data (e.g., user comments, interviews)?

In mixed methods research, it is quite common to use what is known as a "sequential exploratory" design (qual --> QUANT), where the inputs from a smaller qualitative study help determine the nature of a larger quantitative study. That qualitative study can both generate hypotheses and help develop measures to test those hypotheses.

The best applications of this design begin by using the needs of the quantitative study to define the goals of the qualitative work that will serve as an input to that later work. In your case, that would involve further specification of the hypotheses involved and efforts to define measurable versions of the relevant variables.

Also, note that what Barron and Kenny labeled as "moderation" is inherently an interaction effect, so if there is no interaction effect in your model, there is no moderation.

Ultimately, no matter what, ultimately after taking into account the interaction effect (moderation), unexplained variance in the dependent variable and homoscedasticity,  the establishment of the independent variable has to be determined by theory.  Review the literature. Mathematical models and abstracted empiricism are not a substitute for a concise and deductive article review.  

There are so many good books about quant. research to help you with this: going to sagepub.com to find these

Dear Imam,

as mentioned above, it should be supported theoretically, also using empirical evidence, by identifying the moderator and the IV. Moreover, I share the approach suggested by Claudia.

Good luck!

It is always a good practice to establish the theoretical foundation to your research instead of guessing using quantitative methods. So, plan well and define the scope of your research to avoid such confusions. Our colleagues advises are numerous and make sense.

It is always a good approach to establish the relationships between constructs on the basis of a theory(s). An experiment should be based on a theory. 

Look at the past researches, see whether there is correlation between variables. If there is, it  is justify the role of IV and DV. Now look at the DV, if this consequence factor can take a role as an IV, and when tested it is well correlate with a new DV, it is possible to be tested as mediating factor. So now the arrangement of correlation will be IV-MV-DV. Now test correlation between IV direct to DV and if the correlation is significant, then proceed to test correlation between IV to MV, then MV to DV. When the result either lower than or higher than the result of direct correlation (IV to DV) and we can generally say that MV is appropriate to be a mediator.

A moderator variable, commonly denoted as just M, is a third variable that affects the strength of the relationship between a dependent and independent variable In correlation, a moderator is a third variable that affects the correlation of two variables.

Thanks for the answer, Eben.

I know what the definition of a moderator variable is. My question is, how to determine which is the moderator, and which is the moderated, with only a limited literature background, and in a way that is defensible in a peer review?