Now sure what to do now (4 variables, no interaction, significant correlated)

#1
Hello everybody - I am in need of some statistical help as I'm a bit confused as to what to do next. I'll give you a rundown of my situation:

I've conducted a survey measuring 4 variables.

VAR1 - VAR2 - VAR3 - VAR4

All questions loading on each variable were scored on a 1-7 Likert scale, and there were 46 total questions. There were no missing values.

I hypothesised that VAR2 and VAR3 would positively predict VAR4 while VAR1 would negatively predict VAR4.

Initial bivariate correlation analysis (using Pearson R) has failed to reject the null hypothesis - significant positive correlations were found between VAR2~VAR4 and VAR3~VAR4. A significant negative correlation was found between VAR1~VAR4. Further, VAR2~VAR3 were positively significantly correlated.

I then thought that VAR1 could be interacting with VAR2 and VAR3 to moderate the effects of VAR2 and VAR3 on VAR4. So I standardized the results and created interaction terms of:

zVAR1 * zVAR2 - and - zVAR1 * zVAR3

I then performed a linear regression with VAR4 as the dependent variable, VAR1/VAR2 in block 1, zVAR1 * zVAR2 in block 2 and again for VAR1/VAR3 and zVAR1 * zVAR3.

There was no significant beta score indicating that VAR1 does not significantly moderate the effects of VAR2/VAR3 on VAR4.

Now I am at a loss as to where to go with my data.

I originally hypothesised that VAR1 would affect the effects of VAR2/3 on VAR4, but having failed to find a significant interaction effect, I'm not sure where to go now.

What kind of statistical test would be useful to my data now? ANOVA? Multiple regression?

Thanks. My statistical knowlege is quite severely limited and I'm running out of time.

Your help is greatly appreciated.
 
#2
Could anyone indicate a direction on where to go now?

What is even worse is that my DV is predicted by 3 IV's that are all intercorrelated. I think this is called mutlicollinearity.

I have no idea what type of regression technique to use to correct for this.
 
Last edited:

Masteras

TS Contributor
#3
if i understood correct you want to predict var4 using var1-3, right? but var1-3 are correlated and in the regression no significant terms are present. you said you have 46 questions in total, the rest 42? furthermore, the pairwise correlations for these 4 variables are they too low? in what levels?
 
#4
Let me clarify:

There were 46 questions total in my questionnaire. 9 on VAR1, 11 on VAR2, 12 on VAR3 and 14 on VAR4.

I want to predict VAR4 using VAR 1-3.

VAR1 is significantly negatively correlated with VAR4. (.01)
VAR 2/3 are significantly positively correlated with VAR4 (.01 and .05)

My VIF scores are acceptable, indicating that collinearity is not obscuring my results.

When I do a linear regression analysis, I only get a significant coefficient with VAR1 (<.000) but VAR2/3 are not significant.

I thought that VAR1 might be moderating the effects of VAR2/3 on VAR4. So I standardized scores, created an interaction term (VAR1*VAR2 & VAR1*VAR3) and ran heirarchical linear regression twice, once for each term. (Block 1: VAR1/2; Block 2:VAR1*VAR2 etc.)

However, that did not reveal my interaction term to be significant.

I don't know where to take my data now.
 
#5
You see, here is something even more interesting.

I ran the following analysis (linear regression) separately with VAR 4 as the DV:


(ONE IV Regression)
+ IV: VAR 1 ---------- Significant

+ IV: VAR 2 ---------- Significant

+ IV: VAR 3 ---------- Significant

(TWO IV REGRESSION)
+ IVs: VAR 1 ---------- Significant
VAR 2 ---------- Not Significant

+IVs: VAR 1 ---------- Significant
VAR 3 ---------- Not Significant

+IVs: VAR 2 ----------- Not Significant
VAR 3 ----------- Significant

And of course when I run all 3 IV's only VAR 1 is significant.


I understand what this is telling me - it is telling me that from all the VAR's, VAR 1 is the only one which can predict VAR 4 with significance if ALL VARs are accounted for.

But I just can't believe that VAR 2 / 3 are not significant as my whole hypothesis was based around the fact that VAR 1 will interact with VAR's 2/3. I didn't think my VAR 1 would have such a huge impact on predicting VAR4 that it would simply overpower VAR2/3.

You're right - I may simply not have found the results that I hypothesised I would. But so much previous literature has stated to the contrary and I'm trying to find a way to manipulate this data that will at least allow me to gleam something other than the VAR1------>VAR4 relationship.

Further, my interaction terms not being significant really confuses me.

Am I simply missing some other relationship in my data?
 
Last edited:

Masteras

TS Contributor
#6
Ok let's take this slowly. the var IVs each contains 9 11 and 12 quesions. in order to make the var1 did you add the reponses for the 9 questions?
The literature is right, but perhaps you respondents answered accidentally, i.e. they just filled in the questionnaire. As for the significance of var1 when the var2 and var3 are present yes you are right.
 
#7
Ok let's take this slowly. the var IVs each contains 9 11 and 12 quesions. in order to make the var1 did you add the reponses for the 9 questions?
The literature is right, but perhaps you respondents answered accidentally, i.e. they just filled in the questionnaire. As for the significance of var1 when the var2 and var3 are present yes you are right.
Yes I did add them all up and reverse code and the smaller details.

After being stuck for a while, I eventually figured it out. It was a stupid situation. I had missed a vital relationship - that VAR 3 was mediating (bootstrapping method - 5000 samples on my sample (N=144)) the effects of VAR1/2 on VAR4.

When interchanged Mediator / DV, I had obvious reverse-causal-effects as the mediator VAR3 was "causing" or predicting VAR4 regardless of its relationship with VAR1 and VAR2.

I suppose, though contrary to the literature, I am going to have to soldier on down the mediation path.

However, I have gleamed insight into the design of my study and have found a flaw in an assumption of a logical hierarchical order between the VARs (which are personality variables). In fact, the assumption makes sense with the mediation result but was out of line with previous literature.

So I suppose it's backtracking a bit...

Thanks for all the help and patience you provided. It was very much appreciated.

I understand my mistakes here may sound frustratingly elementary. I do beg your pardon.
 

Masteras

TS Contributor
#8
o no do not beg anybody's pardon. it was my pleasure helping you. if you need anything else again, we are here. cu and good luck with the rest