Nonsignificant mediation analysis

#1
Hello there,

first post in this forum. I hope this is the correct subforum, as I have a question regarding psychology statistics, but I am using a regression analysis.

I tested a mediation hypothesis with SPSS using linear regression. I did not find this hypothesized mediation, but I still have problems interpreting the results.
Variable A (Predictor 1) predicts C (dependent variable) significantly, as does variable B (hypothesized mediator/Predictor 2). The thing is that A does not predict B significantly, so there is no mediation, but I have two main effects.

My problem is that if I use a model with both predictors, both effects diminish (as seen in either the p-value or the B). It seems like there is an overlap between both predictors. I also did a correlation analysis to check for that, and Pearson is not significant (p = 0.065), but Spearman and Kendall are significant.

What do I do now? How do I describe the result? I am not sure, if B is a confounder in this case and if I can somehow correct for the overlapping effect.

Thanks in advance!

Edit: Btw, we were told that we could use the Baron & Kenny procedure and do not need to use newer methods, although they might be better
 
Last edited:

Karabiner

TS Contributor
#2
It would be helpful perhaps if you could give us some information.
What is the study about (topic, research question), what was
measured, how large is your sample size, what were the actual
regression weights and p-values ("not significantly" could mean
anything from 0.050 to 1.0), and how large is the bivariate
correlation between the predictors?

With kind regards

K.
 
#3
It is a group decision task. N = 289.

Hypothesis: HEXACO Agreeableness has a positive effect on group decision quality (performance). This effect is mediated by task cohesion.

Reported beta coefficients are unstandardized

At first, a regression analysis of Agreeableness on task cohesion was conducted. This analyzes did not yield significant results (p > 0.05, B = .391). This already reveals that task cohesion is not a mediator between agreeableness and group decision quality. Afterwards, a regression analysis of task cohesion on decision quality was performed.
This analysis showed significant results (p < 0.05, B = 745.591).
To analyze the effects of agreeableness on decision quality, a regression analysis of these two variables was performed, with a significant result (p < 0.05, B = 1244.65). Thus, there are direct effects of both agreeableness and task cohesion on group decision quality.
In a second model, task cohesion was added (thus both agreeableness and task cohesion predicting decision quality). The effect of agreeableness on decision quality declined in this model (p < 0.05, B = 992.01). This shows that there is some overlap between the effects of agreeableness and task cohesion on group decision quality. The effect of task cohesion also declined (p < 0.05, B = 645.47).

And as I said, there also seems to be correlation between agreeableness and task cohesion. Pearson p = 0.065, Kendall p = 0.017, Spearman p = 0.015. But I am not even sure how useful this correlation analysis is.

Unfortunately, I was never taught Durbin-Watson or casewise diagnostics. I just did some residual plots and I did not see anything outstanding, but then again I do not know for what exactly I have to check.
 

rogojel

TS Contributor
#4
hi,
can you check your variance inflation factors? They should be quite high in this case (e.g. above 5). You might want to look at the confidence intervals of the coefficients instead of the point estimates- are the confidence intervals much different? Centering the IVs could possibly help a bit to reduce the VIFs but as long as the effects are significant and the confidence intervals consistent (e.g the model with 2 IVs has larger CIs that contain the ones from the models with 1 IV) the results would not be contradictory.

regards
 
#5
No, the VIFs are equal to 1 in the model with 1 IV and 1.041 in the model with 2IVs. The confidence intervals seem to be roughly the same as the point estimates. However, the confidence interval in the model with 2 IVs is smaller (by that I mean both lower bound and upper bound is lower than in the first model; the width is roughly the same) than in the model with 1 IV.

Regards
 

rogojel

TS Contributor
#6
Interesting. A VIF so close to 1 means that there is not much collinearity between the two factors (AFAIK a VIF for a model with 1 IV makles no sense anyway). Any chance of an interaction between the two IVs?
regards
 
#7
Wait, not much collinearity between the two IVs? But doesn't a significant correlation (tested by Spearman's rho) indicate collinearity?
 
#8
Concerning moderation/interaction: I just standardized both predictors and tested the product of them using multiple linear regression and there does not seem to be a moderation effect (p > 0.25).
 
#9
Okay. Am I correct, if I interpret that there is neither a moderation nor a mediation. And that due to the low VIF and the fact that Pearson's correlation coefficient between the two IVs is not significant, there does not seem to be much collinearity between them. However, because Spearman's rho is significant there seems to be some other form of association (because Spearman's rho and Pearson's correlation coefficient do not measure the exact same type of relationship), which also explains the diminished beta coefficients in the model with two IVs compared to the models with only one IV, although this relationship is not high enough to change anything concerning significance?
 

rogojel

TS Contributor
#10
Wait, not much collinearity between the two IVs? But doesn't a significant correlation (tested by Spearman's rho) indicate collinearity?
The VIF is calculated as 1/(1-R^2) with R-squared being obtained by regressing one IV on the other. A VIF close to 1 means that the R-squared of the regression is pretty much 0.
 

rogojel

TS Contributor
#11
Okay. Am I correct, if I interpret that there is neither a moderation nor a mediation. And that due to the low VIF and the fact that Pearson's correlation coefficient between the two IVs is not significant, there does not seem to be much collinearity between them. However, because Spearman's rho is significant there seems to be some other form of association (because Spearman's rho and Pearson's correlation coefficient do not measure the exact same type of relationship), which also explains the diminished beta coefficients in the model with two IVs compared to the models with only one IV, although this relationship is not high enough to change anything concerning significance?
I think that would be right. In general the multiple regression is a better model anyway, better describing the relationships of the data.

regatds