Standard or Hierarchical Regression?

#1
I would like to predict creativity (DV) through involvement (IV), after controlling for team size and gender.

Questions
1) Do I use hierarchical multiple regression or a standard multiple regression model?
2) A standard model in which all predictors are entered simultaneously, would not allow me to determine how much each variable adds to the model, right?
3) Would entering the variables stepwise in a standard OLS regression be an option?
 

spunky

Doesn't actually exist
#4
1) Do I use hierarchical multiple regression or a standard multiple regression model?
it seems like hierarchical regression is what you're after. i'm guessing you would enter team size and gender in block 1, then involvement in block 2 and check whether there is a significant increase in the R2

2) A standard model in which all predictors are entered simultaneously, would not allow me to determine how much each variable adds to the model, right?
yes, you're right. to see how much each variable contributes to the model you would need to use slightly more elaborate techniques like Budescu's Dominance Analysis or the Pratt Index that allow you to see how much each variable contributes to the explanatory power of the model through an R2-decomposition (that is assuming you don't have weird things going on like multicollinearity, suppression, missing data, etc.)


3) Would entering the variables stepwise in a standard OLS regression be an option?
not sure what you mean here? stepwise as in stepwise regression? in general, computer-generated methods are a bad idea when you're working with explanatory (as opposed to predictive) models.
 

spunky

Doesn't actually exist
#6
I thought this was about multilevel analysis. Now I am not sure :(
So you may have data clustered within teams?

thing is raudenbush & bryk are dumb so they used a dumb name to re-name something that already existed and make everything confusing.

in R&B parlance, multilevel models are called Hierarchical Linear Models (HLM) but people started using the name Multi Level Models instead because HLM is also the name of the software they sell so it sounds like they came up with the technique and the software to analyze it (which means lotsa free publicity and $$$).

Hierarchical Regression is just the idea that you enter predictors sequentially in "hierarchies" so that you can look at the change in R2 and see whether your predictors, as set, contribute to the explanatory power of the model.

MultiLevelModels already had the name by which everybody else who is not in the social sciences knows them: Linear Mixed Effects Models or Linear Mixed Models. my guess is that R&B chose this word "hierarchical" to send the idea that there are "levels" in the data... even though there is no level of anything. there are just fixed effects and random effects.
 

noetsi

Fortran must die
#7
I agree it is confusing although it appears common given how often the same methods have different names. I suspect this is particularly likely when it gets moved from one discipline to another by a few people.

It all goes back to an old argument of mine when I was still in academics. They need to formal organizations in place that set naming conventions and nomeclatures. Rather than let individuals make up terms themself.

There are levels in the data sort of, one set of variables is embedded in another.
 
#8
Here comes my problem, I want to use 1 DV with some IV Like X and x1, x2, x3 and x4 are the part of X and one other Variable PR is also there.

Now my question is, Do I use all the variables in one stance or I go for 3 blocks or Hierarchical Regression Model.

Kindly suggest me the best.

Thanks
 

noetsi

Fortran must die
#9
The primary point of hiearchical regression is that it is theory based. That is the variables are entered in blocks based on preexisting theory. So the answer to your question is, what does the theory say should be added in what order? If you don't have a theory, why are you using hiearchical regression?

I am not sure I understand what you are doing still but if you are using IV where certain of the IV are part of the other IV that is not a great idea. Multicollinearity would be a major issue.
 
#10
I got your point. but when I used squared of each variable, the multicollinearty problem was solved. I am trying to use the original variable, but the problem of normality occur, that is why I am confused. Can you tell me about any regression model which does not require normality. I need a non-parametric model
 
Last edited:

noetsi

Fortran must die
#11
There are a lot of alternatives, before you do that was the residuals non-normal or were the univariate distributions of the IV and DV non-normal. Only the normality of the residuals has to be normal a point which often gets lost. In some cases you can transform the data as well.

However you approach this hiearchical regression is not going to address normality [unless you use it to drop variables out of the model which is not the ideal way to use this].
 

noetsi

Fortran must die
#15
I am not sure how a PM would help.:) Given that I am hardly an expert myself I always prefer to post here so that other might correct results.

Hiearchical regression as I understand it is useful for two reasons. First, to test theory, and second to determine if variables add predictive validity to a model. Since I rarely test theory, or have any, I rarely if ever use it. I have never heard that it addresses issues of normality and I can't imagine how it would.

It would be useful if you post more information on your regression here, particularly the residuals from your regression run.
 
#16
Ok I share the results here. These are original Variables

Model R R Square Adj R2 Std. Error
1 0.433a 0.187 0.174 4.90
2 0.514b 0.264 0.249 4.67
3 0.514c 0.264 0.244 4.69
 
#17
a. Predictors: (Constant), PHI, ECO, ETH, LEG
b. Predictors: (Constant), PHI, ECO, ETH, LEG, PMR
C. Predictors: (Constant), PHI, ECO, ETH, LEG, PMR, Age(y), Gender
d. DV: TSat
 
#18
ANOVA
Model Sum of Sq df Mean Sq F Sig
1 Reg 1413.9 4 353.48 14.68 0.00
Res 6140.2 255 24.0

2 Reg 1993.6 5 398.7 18.21 0.0
Res 5560.5 254 21.9

3 Reg 1995.8 7 285.1 12.9 0.0
Res 5558.3 252 22.05
 
#19
Residual Stats
Min Max Mean Std. Dev N
Predicted Value 15.75 35.58 29.9 2.77 260
Std. Pred Value -5.1 2.0 0.0 1.0 260
Stand. Error 0.40 1.65 0.79 0.22
of Pred. Value

Adj. Pred. Value 16.85 35.6 29.9 2.75
Residual -18.32 10.6 0.00 2.75
Std.Residual -3.9 2.2 0.0 4.6
Stud. Residual -3.98 2.3 0.0 0.98
 

noetsi

Fortran must die
#20
What I need to see is your actual residuals, these are a graph that shows how the points were different than those actually fited by the model. That is how you know if you have violations of the assumptions such as normality. You fit these to a historgram or better yet QQ plot to see if the data is normally distributed.