Interpretation help

IK31

New Member
#1
Hi all,

I'm really struggling to interpret some of my results. I am exploring whether lie detection abilities can predict victimisation in autistic vs. neurotypical samples.

Firstly, I conducted a correlation analysis between the independent variables to check for multicollinearity. It suggested that the variables correlate differently between the diagnostic groups. In the autistic group, I found no correlations between the variables while in the neurotypical group, some of the variables correlated. Why do they differ?

I then conducted a hierarchical linear regression. For the neurotypical group I found that lie detection (and mentalising ability) were significant unique predictors of victimisation. However, none of the four models for the autistic group were significant with no significant unique predictors. How can I explain this difference?

Further, when I collapsed the two groups into one regression, none of the models were significant yet lie detection ability was a significant unique predictor of victimisation. How can I explain this disparity?

Thanks a lot in advance!
 

Karabiner

TS Contributor
#2
Firstly, I conducted a correlation analysis between the independent variables to check for multicollinearity. It suggested that the variables correlate differently between the diagnostic groups.
Why did you do it separately for each group?
In the autistic group, I found no correlations between the variables
r=0.00 for all pairs of variables? That is astounding.
How large are your sample sizes, by the way?
I then conducted a hierarchical linear regression. For the neurotypical group I found that lie detection (and mentalising ability) were significant unique predictors of victimisation. However, none of the four models for the autistic group were significant with no significant unique predictors. How can I explain this difference?
You cannot simply compare significant with nonsignificant results from separate analyses.
For example, the same coefficient can be statistically significant in a larger sample,
while non-significant in a small sample. As Andrew Gelman puts it "the difference between
significant and non-significant is itself not significant"

If you want to find out whether a predictor behaves differently between samples, then
you have to analyse the whole sample and include the group*predictor interaction. Or,
in case of correlation coefficients, there are tests for their direct comparison.

With kind regards

Karabiner
 
Last edited: