Another confounding question. -ANCOVA and perhaps Regression

#1
Hi. I'm trying to analyze the relationship between child abuse and personality disorders.
I have two scales which have their own subscales, continious variables.

I want to research the relationship of different types of child abuse (emotional, physical, and sexual abuse and neglect) on 15 different Personality Disorders.
B/c different types of child abuse are correlated with each other and because gender is correlated with sexual abuse and some personality disorders, I decided to use gender and the child abuse scales as confounding variables.

I can also code all these scales on a YES/NO variable. When a scale is above a certain predetermined cutoff point.

I want to research two questions:
"Does having been the victim of a certain type of child abuse(categorical) have an effect on the level of PD (Continious) when controlled for the other types of child abuse and gender?"
For example, does having been the victiom of emotinal abuse have an effect on Narcissistic PD symptom level when controlled for gender and all the other 4 types of child abuse?
Here I would do ANCOVA, correct?

Here is an example output:
Am I correct on the following:
On the Test of Between-Subjects Effects table of the results of ANCOVA, gender (one of my confounding variables has a sig level of 0,015 and physical abuse (another confounding variable) has a sig of 0,046 while the remaining three types of abuse (the last confounding variables) have sig values above 0,05.
My IV, emotional abuse coded as a 1/0 variable has a sig level of 0,067.
Is it safe to say that in this test the only variables that actually has an effect on my DV are the two confounding variables?

edit: My main question is can I use a continious variable for my DV? Which test can I use to see the relationship between a continous DV and a continious IV while controlling for continious and categorial confounding variables?

And my "second" question is can I do a correlation version of this? For example, what test should I use to answer the following questions:
"What is the relationship between emotional abuse and x PD when controlled for gender and the other types of abuse; specifically, how much of the variance in the IV is explained by the DV and each of the confounding variables and which of these relationships are significant?"

Thanks a lot for any help!
 
Last edited:

Karabiner

TS Contributor
#2
Do you want to know how to do this using the SPSS software?
Seemingly, general linear model (GLM) / univariate can be used
here (if sample size is sufficient).

With kind regards

Karabiner
 
#4
I read somewhere that in doing a univariate GLM, "one of the assumptions is that the confounding variables should not be strongly correlated to one another."

Is this true? Because mine are :)
 

Karabiner

TS Contributor
#5
If the confounders are indeed strongly correlated, then you do not need all of them, since one ore more are redundant.

With kind regards

Karabiner
 

noetsi

Fortran must die
#7
As an aside you usually don't call variables confounds. A confound is something that is related to your dependent variable and also to one of your predictors normally and is distorting the relationship. When you control for that effect by placing a variable in a model it is normally called a control variable not a confound (unless biology is different than social science in that regard). If you are going to write this up you need to address that issue in the wording you use or you are going to confuse your audience.

Its not good practice to include variables in your model you don't need. It will inflate the R square results artificially, may lead to Multicolinearity, and I think it also weakens the statistical power (although this would probably only matter when sample size is small relative to the number of variables estimated). There is a "rule" of Parsimony in statistics that says you keep your model as simple as you can. Adding variables you don't need obviously violates that rule.
 
#8
Great. Thanks for the response :)
I wrote a response to your helpful response on my other thread..
I'm working on Child Abuse(IV), Personality Disorders(DV) and gender(confound). Gender is shown to be related to both IV and DV.
Basically I don't care about the regression model. All I want is the correlation coefficient controlled for other variables such as gender.
Also, I have 5 subscales of Child abuse (physical abuse, emotional abuse, etc.). They're well correlated with each other (r=0,151 to 0,61).
Therefore when I'm analyzing the relationship between emotional abuse and personality disorders, the other 4 types of child abuse will be related to both emotional abuse and PD. So "keep" them in the analysis? Is it reasonable to put keep in quotes?
I'm doing a regression instead of partial correlation b/c gender, something I want to control for, is categorically coded.
 

noetsi

Fortran must die
#9
The regression model's effect size [slope] between child abuse and personality disorders is the correlation coefficient essentially controlling for gender {I am not sure that is formally true, but its how you normally evaluate the relationship between a regressor and a predictor controlling for another variable}.

I am not expert at all in this, but I would think if you had five highly related subscales you might want to create a variable or factor that collapses them into one variable and run your analysis against that not the subscales. I think they do that in psychological analysis.
 
#10
I'm interested in the specific effect of one of the five very related IV's. They are all related and they all predict the DV. I think it makes sense to keep them separate. I saw other research that did that, too. Better to be stingy with possible effects, on my commentary/interpretation.
Thanks very much for the help.