Hello Lucia,

I wasn't quite clear on the functional meaning of 4 levels of psychopathy; are these some sort of ordered scale of dysfunction or mutually exclusive diagnostic categories? (The reason for asking is to help judge whether it makes sense for these to be analyzed separately or not.)

To your query, there's a simple way to answer the question, what is the contribution of variable "X" as an IV when considered among a larger set of IVs for predicting/explaining scores on a specific DV (for your sample)?

1. Run a MLR with ALL IVs involved as predictors.

2. Exclude variable "X", and re-run the MLR.

3. Change in R-squared represents the additional explanatory power attributable to variable "X" when all the other IVs are statistically controlled for.

Repeat steps 2-3 to evaluate the "impact" of other IVs in a similar fashion, being sure to replace the previously omitted IV.

If your research question was, what is the collective impact of gender AND traumatic experiences AND aggressiveness, then the hierarchical approach would likely be more straightforward.

Good luck with your work.

David Morse , thank you for your advise.

MLR seems to be the best option indeed.

Psychopathy is divided in four factors and some of them have been suggested to be linked to self-aggression but not all of them, hence I want to test them separately.

Hi Lucia,

I grasp from your question that you have 4 main independent variables of interest, 3 secondary measured variables that you want to find out if they will play any effects (i.e., gender, traumatic experiences, and aggressiveness), and also you tend to control for 2 variables, age and BDP.

I vote for two options to answer your question:

1. Hierarchical regression

As regards your question about the kind of regression method you should follow, the answer for that ordinarily depends on one’s decision of the order the predictor variables are entered in the model, or on whether you want to build blocks of the variables that are conceptually related for the purpose of computing their overall effect size independent of the other variables, and/or whether there are variables that you want to control for in an individual block.

Just to clarify some issues, hierarchical regression and multiple regression are analogous in that they are both equipped with Forced Entry method which is thought to be the only appropriate method of theory testing (or Enter as it is known in SPSS). In multiple regression, predictors are pooled together in one single block; and therefore, producing one R2 and F-statistic. And one common practice says that significant predictors are entered first (just run the test, see the values, and then re-enter your predictors). In hierarchical regression, on the other hand, you enter your variables in various models based on some theoretical rationale, or for the individuality of reasons given above.

Way back to your question, it shows that you have predictors that you tend to control for, such as the demographic info of the participants (i.e., age, gender), which often go in one block (Block 1), the secondary control variables in another (Block 2), and the other main predictors in a new one (Block 3), for instance -Again it all depends on the sort of reasons you want to group your predictors for. So, you decide the way that applies to your theory.

2. Bivariate correlation and partial correlation before multiple regression

This is a two-step method. In the first step, you run bivariate correlation to screen for the variables that are correlated with the dependent measure, on the grounds that not all your predictors will predict the variance in your dependent variable (sometimes appropriate when sample size is small). Let’s say that you found two predictors that are significantly correlated with the outcome variable, age and traumatic experiences. Now check for collinearity, specifically if these two variables have significant intercorrelations with other independent variables and take note of them.

That being done, run the partial correlation and enter the variables that you want to control for, including age and traumatic experience (a bit tricky stage and you should use your brain about the nature of relationships among variables). Let's assume that a new variable called aggression appeared significant. In order to further explore how these three variables age, traumatic experience, and aggression predict the outcome variable, run multiple regression with these three variables, together with the variables that significantly correlate with them, in order to partial out the shared variance between these variables. This method can generally reduce the excessive number of variables that you may have wanted to enter in a single block, and therefore, circumventing statistical noise as a function of variable redundancy.

My said explanation depends on my recent reading and understanding of two book chapters for Andy Field on correlation and multiple regression besides other materials. So, I advise that you see any other feedback coming around the corner before you construct your final picture.

Best wishes,

Ahmed

You do not need to do any post-hoc tests for multiple linear regression analysis.

From the part correlation column you see how much each IV contributes with (when you square the part correlation you get the %). The IV percentages do not add up to the R2 because they only tell how much each IV uniquely and significantly contributes to the total variance in your DV.

Unless you define your DV as scale variable then you can perform stepwise linear regression and how model improves