Hello there,
I have been struggling to wrap my head around which statistical analysis to use for my research question:
I have 17-20 potential outcome variables (subtests on a cognitive battery, haven't decided yet whether I will retain three of them or not).
I have three predictor variables, age, gender and pain. Pain is my primary predictor variable of interest. I would like to know how pain can predict cognition differentially.
I built the model on AMOS (picture included... yikes) and it is saturated.
But because it's not a path per-say, would global fit matter? Can I pursue this analysis regardless and start interpreting the local fit (which paths are significant?) Since the outcome variables are not exogenous. Should I even report global fit?
Further to this, would I report the b, se b, ß, CR and p for each path? And R-squared for each variable?
And finally, is there any way to conduct a hierarchical multivariate multiple regression? Ideally, I would like this to be hierarchical, since I want to know pain's unique variance and adj. R2, but as far as I can see, there's no way I can do that wish so many outcome variables.
The covariances were added by my supervisor because they are all measuring cognition, so it followed that they would all be correlated with one another.
ANY HELP! would be so, so, so, greatly appreciated.
Yikes! Where to start?
There is too much here to deal with in one post, Katie.
The model is "saturated" because you have everything connected to everything else. Think back to your intro-stats classes where you ran simple, and then multiple, regression and there were many more observations than there were variables in your model. This is a similar situation. You need far fewer connections than there are variables.
This is where theory is important. Are there some variables that logically (theoretically) go together -- Where your measures can be seen as physical manifestations of some underlying mental condition? For example, "short-term learning functions", "abstract calculation functions", etc. These might be brought together as latent variables (represented as ovals in AMOS).
Think also about whether reality works as you have described it... Does gender really affect "story memory" directly? Or should gender affect the relationship between pain and "story memory"? If the later, then you're talking about a moderating variable.
There is a great deal more that could be discussed, but you would do well to walk through one of the many online tutorials on AMOS. For example:
http://www.structuralequations.com/AmosTutorials.html
Also, your library probably has a copy of this book:
Byrne, Barbara M. (2016), Structural Equation Modeling With AMOS: Basic Concepts, Applications, and Programming, 3rd Ed. Routledge
In principle, SEM should be similar to MR, so I don´t see why not.
Hope this helps
Matt
Hello Katie,
I have two suggestions:
1. Try canonical correlation; which is multivariate regression, but without the scary picture! Here's a good link for some tips on how to proceed: Article Conducting and Interpreting Canonical Correlation Analysis i...
2. Try factoring the 17-20 variable set into a smaller grouping of latent variables that make sense, then try your model-building.
Good luck with your work!
- Badges
- Science topic
- Similar topics
- Electrochemistry
- Battery