[R] Etimating time to run an analysis?

Ben Bolker bbolker at gmail.com
Wed Nov 27 22:14:35 CET 2013


Erika Barthelmess <barthelmess <at> stlawu.edu> writes:

> Hi everyone,
> 
> I'm new to this list and have searched R help prior for an answer 
> to this question, without luck.  If I'm
> posting in error, please forgive.
> 
> I'm thinking about using package MuMIn to do multimodel inference 
> with logistic regression.  I have many
> (25) possible predictors and am curious if there is a way to 
> estimate how long the dredge command might take
> to run?
> 
> Any suggestions most welcome.
> 
> Thanks,
> erika

  This is likely to be a bad idea.  With 25 predictors you have 2^25 =
33 million candidate models (you can think of an array of models, each
predictor is either present or absent in each model -- that makes this
a set of 25-digit binary strings ...).  (If this doesn't make sense,
convince yourself by writing out the number of possible models for a
1-parameter (2), 2-parameter (4), and 3-parameter (8) model, and do
the extrapolation.) So pick a model of intermediate complexity, run
it, see how long it takes, and multiply that by 33 million ...  (if
each model takes about one second to fit, the analysis will take
about a year to run).

  You might want to look into penalized regression approaches
(e.g. see the glmnet package), which are a much more efficient
approach to this type of problem.



More information about the R-help mailing list