Hierarchical Regression

#1
Hey Everyone,

I'm certainly stressing out in regards to my MSc Thesis right now and I could certainly use your help.
I am testing a IS success model. It consists of 9 constructs. I did a multiple regression analysis for all expected significant correlations (according to theory). The whole model and the regression analysis look like this:
model.png

However, after this went fine (according to my opinion) I now need to do a hierarchical regression analysis (i think, to see what the effect of the mediators are or something?). I haven't slept last night due to working on my thesis. I am starting to hallucinate. What does the following hierarchical regression table tell me about the full model?

Screenshot 2019-06-11 at 22.23.41.png

Furthermore, did I test all 4 models using this method? I really need some help, so thanks in advance for even reading this!

Cheers,
A desperate student
 

noetsi

Fortran must die
#2
Are you talking about multilevel regression or regression that test whether you should add new variables to your model? I think the later, but I am not sure. I know nothing of your theory at all and am just commenting on your method (which I have only seen in class never outside it).

It looks to me that the second block is not significant suggesting they add little to the model. But the third block is significant suggesting you should add them. But some in the third block are in the 2nd block which suggest inconsistent results. I would try rearrange the blocks so that Itcomp and infonet as well as innovReg is added in the 2nd block. The third block would add Servqul, infoqual and sysqual.

"A significant F-change means that the variables added in that step signficantly improved the prediction."
http://web.pdx.edu/~newsomj/pa551/lectur20.htm

That said using this approach is not common any more as far as I know. You usually base your variables on theory or something like LASSO (an empirical test of what variables to add to the model).
 
#3
Wow, thanks so much for your reply. It makes more sense to start with usage and use satisfaction, adding innovationres, ITcomp and infointensity in the second block, and the Qual variables in the third block, when looking at the full model.

I did base my variables on theory, the full model represents the theory I used. I tested this model by doing several multiple linear regressions.
Can I just fill in the coefficients and significant levels of the (sub) models in the full model? Is it a bad think, looking at the full model, that the third block with the Qual variables do not predict net benefits? Is it a bad thing that the second block with innovationres, ITcomp and infointensity do predict benefits?

My theory is that those variables influence net benefits through user satisfaction and actual usage.

Screenshot 2019-06-12 at 13.17.14.png Screenshot 2019-06-12 at 13.17.18.png
 

noetsi

Fortran must die
#4
Remember I don't know your theory and am not a statistician (I am a data analyst who uses statistics) so take what I say with a grain of salt. :p Well several.

'Can I just fill in the coefficients and significant levels of the (sub) models in the full model?'

I am not sure what this means. The common way to compare models that are nested (you have one model with all the variables and then models that have some, but not all the variables in that model) is to do an F test. This shows that.

https://www.stat.ncsu.edu/people/bloomfield/courses/st370/Slides/MandR-ch12-sec02-06.pdf

If the models are not nested, you would compare the effect of two variables on Y to a model with 3 other variables on the same Y with AIC. This is not telling you, as far as I know about any one variables contribution.