Discussion: View Thread

This message has been cross posted to the following eGroups: Young Professionals Group and Statistical Consulting Section .
-------------------------------------------
Dear All,

I am fitting GEE model to my data. Let's say Y is my dependent variable and X1, X2, X3, X4, X5, X6 are independent variables in the model.

• If I use X1as the sole predictor it is significant , similar is the case with X2 and X3 and the model QIC value for each of these models is 574
• In the model with X1-X6 only one of the three (X1,X2, X3) is significant (p-value 0.03) and there is another lets say X5 (0.04) which is significant with QIC =584.
• However, if I keep X1, X2, X3 and X5 nothing is significant, QIC =577.

-------------------------------------------
[Tasneem] [Zaihra]
[Post Doctoral Fellow]
[McGill University]
-------------------------------------------

OK, you centered the variables. That does not eliminate collinierity, just reduces certain kinds of collinierity. Have you examined the VIF for the variables? Have you examined the covariance pattern for X1, X2, X3? It certainly sounds like a collinierity problem of some kind.

-------------------------------------------
Paul Thompson
Director, Methodology and Data Analysis Center
Sanford Research/USD
-------------------------------------------

In addition to collinearity, you might consider the scientific relationships among these 6 covariates, as in any modeling project. For example: What happens in a model with just X1, X2, and X3, all of which are individually statistically significant? What is the relationship between X5 and the two variables among X1, X2, X3 that are not significant when you fit the model with all 6 independent variables? How do X4 and X6 relate to these variables?

-------------------------------------------
Alicia Toledano
President
Biostatistics Consulting, LLC
-------------------------------------------







Original Message:
Sent: 11-04-2013 17:13
From: Paul Thompson
Subject: Variable Selection for GEE model

OK, you centered the variables. That does not eliminate collinierity, just reduces certain kinds of collinierity. Have you examined the VIF for the variables? Have you examined the covariance pattern for X1, X2, X3? It certainly sounds like a collinierity problem of some kind.

-------------------------------------------
Paul Thompson
Director, Methodology and Data Analysis Center
Sanford Research/USD
-------------------------------------------

Full disclosure: I'm fairly suspicious of automated or algorithmic approaches to variable selection.
I don't believe centering can "fix" a confound or collinearity...except under vary particular circumstances.

There are scenarios where one algorithm or another happens to be structured in a way that appropriately gets you through the phenomenon on which you're working...that is, the algorithms ARE taking a particular approach, envisioned by their creator(s). When you choose the algorithm, you choose their approach.

I tend to think of it as parking your car on a hill, putting it in neutral, and getting out. You may not be "driving", but you're still responsible for understanding the predictable events you set in motion...and for that car landing in someone's living room.

I tend to approach the problem more mechanistically. What are X1-X6? What is Y?

Are X1-X3 imperfect measures of the same construct (simple: use average, complex: latent trait)
Do X1-X3 represent items in a potential causal chain? (simple: use the one most proximal to Y, complex: path/SEM/causal)
...and so on.


-------------------------------------------
Jason T. Machan
Director, Lifespan Biostatistics Core,
Lifespan Hospital System
Research Scientist, Biostatistics, Research
Rhode Island Hospital
Assistant Professor, Departments of Orthopaedics and Surgery
The Warren Alpert Medical School, Brown University
Director Biostatistics Externship, Adjunct Assistant Professor, Department of Psychology
University of Rhode Island

-------------------------------------------

Centering the X's especially helps if you are dealing with polynomial terms in the X's.  I don't see much value in centering the y's, indeed it may change the shape of the outcome dramatically if the y is non-normal and require you to use a different model.
Clearly what you have is co-linear variables. You ought to start with the most important variable
(clinically or statistically - often clinically is critical, depending on what you want to do).  and then step up
(you may have to do it by hand - by hand stepping down is usually easier.)
It is important to realize that because of the co-linearity you will probably have many equivalent models that predict
the results equally well.  There is no unique model, so you need to take into account which of them are the most relevant and work from there.
  I would recommend Fitzmaurice, Laird and Ware's book (2nd ed) as a reference.
You don't say what distribution Y has, which is critical. 
I also don't use GEEs that much any more, GLMMs allow better parameterizations, especially for the variances and covariances.

Ray

-------------------------------------------
Raymond Hoffmann
Professor of Biostatistics in Pediatrics
Medical College of Wisconsin
-------------------------------------------







Original Message:
Sent: 11-05-2013 11:50
From: Jason Machan
Subject: Variable Selection for GEE model

Full disclosure: I'm fairly suspicious of automated or algorithmic approaches to variable selection.
I don't believe centering can "fix" a confound or collinearity...except under vary particular circumstances.

There are scenarios where one algorithm or another happens to be structured in a way that appropriately gets you through the phenomenon on which you're working...that is, the algorithms ARE taking a particular approach, envisioned by their creator(s). When you choose the algorithm, you choose their approach.

I tend to think of it as parking your car on a hill, putting it in neutral, and getting out. You may not be "driving", but you're still responsible for understanding the predictable events you set in motion...and for that car landing in someone's living room.

I tend to approach the problem more mechanistically. What are X1-X6? What is Y?

Are X1-X3 imperfect measures of the same construct (simple: use average, complex: latent trait)
Do X1-X3 represent items in a potential causal chain? (simple: use the one most proximal to Y, complex: path/SEM/causal)
...and so on.


-------------------------------------------
Jason T. Machan
Director, Lifespan Biostatistics Core,
Lifespan Hospital System
Research Scientist, Biostatistics, Research
Rhode Island Hospital
Assistant Professor, Departments of Orthopaedics and Surgery
The Warren Alpert Medical School, Brown University
Director Biostatistics Externship, Adjunct Assistant Professor, Department of Psychology
University of Rhode Island

-------------------------------------------