Comments on: R-Sessions 32: Forward.lmer: Basic stepwise function for mixed effects in R

By: Gaiarrido

Gaiarrido — Mon, 08 Jun 2015 17:15:59 +0000

Fantastic!!
Thanks

By: Rense Nieuwenhuis

Rense Nieuwenhuis — Mon, 15 Jul 2013 12:46:13 +0000

Hi Nick,

thank you for this excellent recommendation. For those interested, here’s the link to the MuMln-package manual: http://cran.r-project.org/web/packages/MuMIn/MuMIn.pdf

By: Nick Isaac

Nick Isaac — Mon, 15 Jul 2013 11:22:59 +0000

The MuMIn package now works well for mixed-effects models. I would recommend this over stepwise regression.

By: Rense Nieuwenhuis

Rense Nieuwenhuis — Thu, 11 Jul 2013 12:01:08 +0000

Hi Allie,

I’m afraid I don’t really have any examples to share. I wrote this code to help a colleague out, and it really was tailor-made to his needs. It is good to hear that you find it helpful!

By: Allie B.

Allie B. — Fri, 05 Jul 2013 23:40:27 +0000

Hi Rense,

Thanks for sharing this function! I was wondering if you have any example data/code available using this function? I’m trying to run it on my own and it would be really helpful to work through an example as I run into errors.

Thanks

By: Rense Nieuwenhuis

Rense Nieuwenhuis — Wed, 20 Jun 2012 17:11:44 +0000

Hi Frank,

thanks for using this function. I am aware of one journal article that used this function (and later this article was incorporated in a dissertation). The paper is by Jochem Tolsma, Tom van der Meer and Maurice Gesthuizen, and was published in Acta Politica. It is titled “The impact of neighbourhood and municipality characteristics on social cohesion in the Netherlands”.

You can find the article here: http://www.jtolsma.nl/uploads/Tolsma%202009%20AP.pdf

By: Frank

Frank — Fri, 27 Apr 2012 21:25:11 +0000

He Rense! Thanks for sharing this wonderful function. I’ve found it to be quite useful. Are you aware of any papers that have used and cited it?

By: Rick Vinod

Rick Vinod — Fri, 24 Sep 2010 15:46:18 +0000

would your function work for cox proportional hazard
models? I am fitting one of those with 3 sets of possible
dependent variables and some 100 possible regressors.
stepwise seems to be needed.

By: Alban Guillaumet

Alban Guillaumet — Mon, 02 Nov 2009 20:08:35 +0000

Hi Rense,

very helpful indeed.

I did not go through the entire code, but I think I have noticed a bug so I wanted to let you know

when writing the logs for if(sig.level != FALSE), your anova calculation for log.p[log.step] comes after you already replaces model.basis by model[[j]], so you do not get the appropriate p.value.
I suggest you store anova(model.basis, models[[j]]) before you apply model.basis length(blocks)) max.iter = length(blocks)

for(i in 1:max.iter){ # 2a

models = list()
aic = numeric(length(blocks))

for(j in 1 : length(blocks)){ # 3a

models[[j]] = update(model.basis, as.formula(paste(“. ~ . + “, blocks[j])))

LL = logLik(models[[j]]) ; df = as.numeric(attr(LL, “df”))
aic[j] = as.numeric(-2* LL) + 2*df

} # 3b

LL_mb = logLik(model.basis) ; df_mb = as.numeric(attr(LL_mb, “df”))
aic_mb = as.numeric(-2* LL_mb) + 2*df_mb

j = order(aic, decreasing=FALSE)[1]

if(aic[j] 0) log.df = data.frame(log.step = 1:log.step, log.block = log.block, log.AIC.bef = log.AIC.bef, log.AIC.aft = log.AIC.aft) else log.df = data.frame(log.step = NA, log.block = log.block, log.AIC.bef = log.AIC.bef, log.AIC.aft = log.AIC.aft)
return(log.df)} # 1b

By: Nick Isaac

Nick Isaac — Fri, 16 Oct 2009 16:51:50 +0000

Dear Rense,

Thanks for putting this out – I’ve been thinking for a while how I’d code this. I should’ve realised someone would have done it already.

I’m intending to use the simpler ‘lowest AIC’ criterion to choose the best model. I will start modifying the code next week. Do let me know if you’ve done this already, but if not I will let you have the result.

Best wishes, Nick