statistical modelling SAS vs R
Bill Szkotnicki wrote:
Hello, Recently I have been reading a lot of material about statistical modeling using R. There seems to be conflicting opinions about what the best approach is between the SAS community and the R community. 1) In R one might start with a model that has all possible effects of interest in it and then simplify by eliminating/adding insignificant effects using a stepwise procedure. 2) In SAS one may starts with a "reasonable" model and look at type 3 SS's to test hypotheses and report LSMEANS. This can be done in R too I think. Does anyone have current opinions about this? I know it's been discussed before but I would be very interested in hearing about the advantages and pitfalls of both approaches. Bill
You'll get lots of opinions about this. Both R and SAS can be abused terribly, and both approaches you mentioned have major problems if you use P-values to specify models. Better and more replicable results can be obtained using modern shrinkage methods and being more liberal with inclusion of variables, or by using Bayesian model averaging. Note that LSMEANS and Type III tests are SAS concoctions and that if you have interactions in the model, type III tests have been criticized. If there are no interactions, type Type III = Type II. Frank Harrell
Frank E Harrell Jr Professor and Chair School of Medicine
Department of Biostatistics Vanderbilt University