Tuesday, November 18, 2008

Hierarchical Stepwise Regression

What is a hierarchical stepwise regression? Is it as just as "bad and evil" as stepwise regressions? How does it relate to forward and backward entry?

Thank you.

4 comments:

Kris said...

This may be a better question for Mari, but I will give it a shot. I think the issue that is present with stepwise regression is that it is not particularly based upon a model but simply includes predictors based on their goodness of fit as decided by a computer program. Hierarchical stepwise regression is then the imposition of the researcher in terms of the sequencing of the predictors. The issue here is that stepwise regression is motivated by a lot of data with a lot of possible predictors and no underlying theory or model of analysis (Cohen, et al. 2003) From what I can tell, hierarchical stepwise regression is simply including some logical imposition of the entry or selection of variables and yet the computer program still continues to apply variables to an analysis until no more variables are available that contribute unique variance. So yes, I would conjecture that Mari would say it is almost as evil, but at least the researcher is playing some role. The more hierarchical this becomes, the less it is stepwise and just becomes multiple regression.

Mari said...

Let me try to clear this up.

Hierarchical (with or without the term stepwise after it) implies that you as the researcher determined the order of entry.

Stepwise--by itself--implies that SPSS determined the order of entry.

Thus, hierarchical (where you entered order of variables) = good. Stepwise (where computer entered order of variables) = bad.

Jenny Le said...

I think both are just as bad. Hierarchical regression can lead towards subjective models. Stepwise regression might lead to exclusion of effects that might have been important when variables would have been arranged differently. Anderson (2008) proposes a method in which one first have to make proper hypothesis about the data and than make all the possible models that hypothetically make sense. One can then compare these models and select the model with the best fit. If there is no single best model one can calculate an averaged model based on all the models using model weights. The book I refer to is named "Model Based Inference in the Life Sciences". It is a really good read and it will change your idea about the current use of statistics.

Jenny Le said...

I think both are just as bad. Hierarchical regression can lead towards subjective models. Stepwise regression might lead to exclusion of effects that might have been important when variables would have been arranged differently. Anderson (2008) proposes a method in which one first have to make proper hypothesis about the data and than make all the possible models that hypothetically make sense. One can then compare these models and select the model with the best fit. If there is no single best model one can calculate an averaged model based on all the models using model weights. The book I refer to is named "Model Based Inference in the Life Sciences". It is a really good read and it will change your idea about the current use of statistics.