Help with my hierarchal regression

#1
Hi Guys

I'm accomplishing a hierarchal regression for my thesis and I'm a bit stumped. My hierarchal regression is trying to assess the link between exposure to lead, mercury and polychlorinated biphenyl and their impact on attention. I've got three variables of prenatal exposure for those three contaminants (continuous variables) and three variables of postnatal exposure for those contaminants (continuous variables). I also have the level of education (categorical variable), participant gender (categorical variable), the level of selenium pre and postnatal (continuous variable) and SES (categorical) and age (continuous variable). In the first stage of SPSS I input the Gender, the SES and age and selenium. Then in the second step I enter all the measures of contaminants exposure (pre and postnatal). When I look a the output it only have periods in the output tables and there are no numbers. Is it because I have explained all the variance and there isn't anything else to explain? If this is not the answer, can someone please explain to me what is the issue? Is a hierarchal regression meant more to look at single variables individually and how much variance they can explain instead of looking at many variables together?

Thanks

Lewis
 
#2
You have the right idea, but I think you might have set it up differently in SPSS than would be typical. Gender is dichotomous, so would be used in logisitic regression. If you include sex I would include it as a covariate.

Also, why did you put everything together in the first step? Just wondering, because it may skew your results. I would encourage you to separate your predictors. Here is what I would do:

XStep 1: SES, Age
XStep 2: Selenium
XStep 3: Lead
XStep 4: Mercury
XStep 5: Polycholorinated biphenyl
Y: Attention span
Covariate: Gender

This would provide you with your 'covariates' at the beginning (SES, Age), and then the predictive nature of each variable over and above that. You results would read something to the effect of 'In Step 1, SES and Age did not have a significant effect (stats). In Step 2, selenium was entered into the regression and was a significant predictor of attention span (stats). Selenium accounted for X% of variance in attention span over and above that of SES and Age. In step 3, lead was significant (stats), and accounted for X% of variance over and above that of selenium....' and so forth. You would also state gender was included as an actual covariate; your Step 1 variables are covariates (sort of), but they are held statistically-constant as your regression goes.

An alternate way of doing this would be to start with a correlation table, see which variables are correlated, and then build your regression from there. You could drop everything into one block in SPSS and whichever predictors are significant you can re-run the analysis with just those and then go from there. It's a bit of a rough method, but it could save time in the end. Reporting it might be messy.
 

hlsmith

Omega Contributor
#3
By hierarchial it seems you mean a kind of stepwise regression not multilevel regression.

I would use your background knowledge to select order based on assumed association. May also help to diagram the relationship between variables.
 

rogojel

TS Contributor
#4
hi,
why not use a full model and then do a model selection, based on any of the available approaches? I would not do any pre-selection of variables or of the order they enter the model, the risk of bias is just too high imho.

regards