Don't do it. Stepwise models are no substitute for a hypothesis. And if you are aiming at a pragmatic decision rule, then allow a knowledge of the workflow to influence you, prioritising data that are readily available and of high quality.

See further here: http://www.stata.com/support/faqs/statistics/stepwise-regression-problems/

Confession: I am one of the authors of the FAQ

I do it with stepAIC function. let's consider you have some model "mymodel".

then stepAIC(mymodel) will give you all steps of forward-backward selection with AIC criterion. If you need onla backward, you can choose stepAIC(mymodel, direction="backward")

http://stat.ethz.ch/R-manual/R-patched/library/MASS/html/stepAIC.html

Hi Oleg,

I did it and obtained that error message

Warning in nobs.default(object, use.fallback = TRUE) :

no 'nobs' method is available

Error in UseMethod("extractAIC") :

no applicable method for 'extractAIC' applied to an object of class "formula"

Thank you very much

it must be something like

mymodel

seems to work. However, I am wandering how one can select the F to enter and F to remove to adjust te stingency of the selection? Also, why the model should be binomial?

Thank you

Julien

I've (more or less) the same "problem"... I want to do a stepwise DA and I found package klaR (http://cran.r-project.org/web/packages/klaR/index.html) that has a function specifically for that purpose. I'm in the process of reading the documentation (which is not the easy to find) so... for now I cannot be of further help.

Regards, Eduardo.

Julien,

What kind of model are you fitting? If you are fitting a linear model, you can use the lm(y ~ x1 + x2+ ... ) command instead of the `glm' function suggested by Oleg. The stepAIC function in MASS package does covariate selection on the basis of AIC or BIC. The default is AIC. If you want to use a p-value based criterion, you need to figure out a corresponding value for the penalty parameter `k' in the stepAIC function. Suppose that you want to use a p-value of 0.1 for selection. Now, the chi-squared deviate corresponding to this p-value is 2.706. So, you would set k = 2.706 in the stepAIC function. I hope this is helpful.

My model is linear I think. I want to assess the effect of 6 conditions on the output containing 152 dependent variables.

Using lm to fit the model, I obtain this following error message.

Warning in model.response(mf, "numeric") :

using type="numeric" with a factor response will be ignored

Error in lm.fit(x, y, offset = offset, singular.ok = singular.ok, ...) :

invalid to change the storage mode of a factor

For te K, I got it, I think. And the default value is 2.

thank you to all

Don't do it. Stepwise models are no substitute for a hypothesis. And if you are aiming at a pragmatic decision rule, then allow a knowledge of the workflow to influence you, prioritising data that are readily available and of high quality.

See further here: http://www.stata.com/support/faqs/statistics/stepwise-regression-problems/

Confession: I am one of the authors of the FAQ

More arguments against stepwise model selection http://www.nesug.org/proceedings/nesug07/sa/sa07.pdf.

A penalized model is a better alternative (e.g. ridge or LASSO). The R pakage penalizedLDA may be useful http://cran.r-project.org/web/packages/penalizedLDA/index.html

Hello all,

For those who argue against the use of stepwise tests, please suggest an anlternative method to select discriminating variables among a large number (which we easily get in the era of high throughput).

cheers

Julien.

Julien, read the linked texts - there are alternatives given.

My suggestion is using L1 regularization as alternative to stepwise variable selection. You can fit a generalized linear model via penalized maximum likelihood using the lasso penalty in R (glmnet package http://cran.r-project.org/web/packages/glmnet/index.html).