If the issue is the interaction between moderators, that would involve a three-variable interaction between the independent variable and both moderators. That would be useful for testing hypotheses such as the whether the effect of the independent variable required both of the moderators to have high scores.

In a regression analysis, you would add this as a further model after testing the model with the pair of two-variable interactions. 

The process of testing moderation is known in the literature. In the attached recent published articles you can follow the same procedure in testing your models and figure out the moderation effect.

In the article "For fathers raising..." ther are different statistical approaches used to test the moderation effect.

Good luck

Both your models imply that Perceived Financial Security (PFS) and Social Support (SS) are totally independent constructs.  If it were the case, both of your models would be equally good in capturing the independent influence on your outcome variable (which I would rather call Unhappiness).  

In reality, I guess you would expect PFS and SS to have at least some relationship among them, because when we have friends to rely upon (SS), a financial hardship is not that scary (higher PFS), even if we do not intend to borrow money from them, but just ask for advice in financial decisions.

So what I would do is first try to correlate SS and PFS, as well as look into the actual questionnaire items, correlate everything on the item level, essentially do factor analysis, and also perhaps ask independent raters to tell you which of the SS items also predict PFS and vice versa.  

If there is a non-trivial correlation between SS and PFS, you will notice in your Model A that when you look only at SS as a moderator, it will steal some of the moderating influence of PFS -- to the degree that SS can disguise as PFS in the questionnaires.  The opposite will equally be true in the second part of the model.

In this sense, I would go with Model B and do two regression analyses: (1) enter SS first and PFS second, and (2) vice versa.  The difference in the predictive strength between two analyses will essentially be shared influence of the two variables that your model cannot assign to either SS or PFS.  And then, in your discussion section, you may argue that since SS and PFS are related to each other theoretically, this shared influence is essentially the measure of construct overlap in your model of the phenomenon.

Hope this helps!

****

Having read what I have just written, I would add that you may actually have a two-tier structure with SS moderating your outcome either directly, or via PFS, because SS logically predicts at least some of PFS, although not vice versa.  But then again, some people have a lot of friends when they are rich, but not when they become poor, so the reverse relationship may be true as well.  I guess both of these things will fall under your shared influence that you cannot pull apart statistically, unless you have an actual or quasi-experiment.

Hello, I  prefer model B because  you can eliminate variance among respondents if you ask two different groups of respondents. Anyhow, when you test the moderating effect you will do it sequentially just like model A. I  share a tutorial link from Statwiki that you can watch the video. It is a good resource for my research.

Good luck

http://statwiki.kolobkreations.com/wiki/Structural_Equation_Modeling#Interaction

Dear Professor Muayyad Ahmad

thank you so much for your respond. However, i apologize for writing this, but your respond does not answer my question.

i know how to test moderation effect, i will be using a simple multiple regression method and will use an interaction between moderator and predictor towards the outcome variable.

My question was, are Model A and Model B the same model? Because that could lead to a different number of participant. 

Thank you very much Professor Muayyad Ahmad hope you will still reply...

Best Regards

Dear Oleksiy (Alex) Chadyuk.

Thank you for your elaborate answer, it exceed the answer that i was looking for.

Basing from your answer, i will choose model B also, since it will give me less participant. This is due the the fact that my participant is kind of difficult to get. 

Also thank you for your suggestion in analyzing the data, i am aware now that there is a possibility that the two moderators are related to each other. And because of your explanation, i now know how to do the discussion if the relationship between the moderator exist.

Again, thank you...

Best Regard

Dear Didik Wahjudi

Thank you very much for your answer. However i still confuse about your answer.

You said that you prefer model B, but also wrote about different participants. Please be aware that model B is the one that does not have different set of participant. 

And i do also know that the increasing the number of participants will eliminate the variance..

So can you please explain it again which model would you choose? I apologize for not understanding your answer.

However, thank you for your file it could help me understand more about the moderation effect.

Thank you and Hope you still willing to reply

Best regards

Dear Wisnu,

I stated that I recommended model B in which you could eliminate variance between groups. If you use two groups of participants, there is big possibility that the two groups generate differences or variance. In that case, you cannot differentiate whether the variance comes from the moderators or the groups. In addition, I prefer to use SEM instead of multiple regression because it is more robust according to the article in the link below.

I hope it clarifies my answer. Good luck.

http://www.ncbi.nlm.nih.gov/pubmed/11885063

Dear Didik Wahjudi

Thank you very much for the clarification. I still have to learn a lot since im new to this moderation study. Furthermore, a lot of people has recommended me to use SEM i will still look at it, since its also new to me. My statistic class did not include this method. My plan was to use simple multiple regression with interaction effect between moderator and variable counted as new variable. However, i will still look at it.

Again thank you for the suggestion and have a nice day.

Best regard

Willibrordus, I agree that Structure Equation Model (SEM) is the best way to test moderation and mediation. SEM eliminates the problem of multiple comparison by assessing the inter-relationships among all variables. By using two distinct models with different participants you cannot eliminate variance differences attributed to the cluster of participants, as Didik mentioned.

I don't think that a full SEM analysis is necessary in this case. Model A basically implies two different interaction effects, which can be included in the same regression equation. The question of which variable is a "better" moderator" can be examined via effect size. If necessary, the models can be compared directly by systematically removing each interaction from the equation and examining the change in R-sq.

David, I see your point. The issue may be the interaction between moderators. How could a regression model test that without multiple comparisons?

If the issue is the interaction between moderators, that would involve a three-variable interaction between the independent variable and both moderators. That would be useful for testing hypotheses such as the whether the effect of the independent variable required both of the moderators to have high scores.

In a regression analysis, you would add this as a further model after testing the model with the pair of two-variable interactions. 

Dear Joan Vaccaro and David L Morgan

Thank you very much for your information and suggestion. I will read further about this topic.

Have a nice day

:)

One word of caution - this sort of moderator analysis tends to have very low power. Green's rules of thumb depend on very assumptions about standardise effect size - which are approximate at the best of times. However, for a study with measured/observed variables the moderator effects almost always have small standardised effects unless you have unusual distributions (ones with most values at the extremes and few in the centre). Things are somewhat better with manipulated variables (e.g., factors in designed experiments) because you can design the study to exclusively sample the extremes. For measured/observed variables you may want to look at an extreme groups analysis where you over-sample the extremes (at least if an interaction is your main focus).