http://www.universalrejection.org/

Hi Dr. Anna Gellert,

Thanks for the website.  I found the mentioned journal to be amusing and this journal is a little different from what I was looking for.  

There have been many research papers on the dangers of using automatic variable selection in multiple linear regression such as step-wise regression with backward, forward elimination, best-subset, etc.  I remember reading a paper which stated that this journal will automatically reject papers as a formal policy if they use step-wise regression, etc .

Thanks 

Sangdon

Multiple regression seems to be something f the past that is considered old-fashioned in a way. So I guess it would be difficult ot publish a study using multiple regression in a good journal....

Hi Mohamed,

I would not say that MLR is old-fashioned.  It is just that the automatic variable selection procedure in MLR or in any methods including artificial neural networks should be careful.  Many regressions in the generalized linear model (including, logistic regression, Poisson regression in general linear model) have the same issue: automatic selection of predictors should be careful. 

Sangdon

My research used stepwise multiple regression. I aslo want to know the same answer.

Hi Yan

I found the journal names

1. Annals of Internal Medicine (http://annals.org/public/authorsinfo.aspx )

2. Stepwise Regression and Stepwise Discriminant Analysis Need Not Apply here: A Guidelines Editorial, EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT, AUGUST 1995

(http://www.researchgate.net/publication/247648213)

It is well researched and known the danger of using stepwise regression "automatically".  

We need to be very careful about the difference between two types of models: 

(1) a model to "predict" Y from X, and  

(2) a model to find out how X "causes" Y

If we just want to predict the Y from the X, the use of stepwise is not a problem at all (even with severe multicollinearity).  As an example, if the subscription of Wallstreet Journal can predict the vehicle sales very well somehow, we will monitor the subscription very carefully, however, if we want to increase the vehicle sales, the subscription of Wallstreet Journal is not a good variable to work on.   

Understanding causation is a difficult subject and needs to understand (1) data type (observational vs. experimental data) AND (2) purpose of modeling (prediction vs. causation).   Mathematically or statistically, understanding of "collinearity" also known as "ill-conditioning" is the key to understand.

Sangdon

Article Stepwise Regression and Stepwise Discriminant Analysis Need ...

I found these helpful

https://annals.org/aim/pages/AuthorInformationStatisticsOnly

Article Sun GW, Shook TL, Kay GLInappropriate use of bivariable anal...