R lm on each factor level

Question

I'm hoping to find some guidance on both the how-to-code level and on the how-to-solve level in R. I have a simple data frame as follows:

str(pf)
   'data.frame':  1536 obs of 4 variables:
   $ dt : POSIXct 
   $ pfRet : num 
   $ Src : Factor w/6 Levels "ALPA" , "OMEGA", "GAMMA" , ..
   $ ret : num

I simply want to see lm(pfRet ~ ret ) for each of the 6 different factors and compare them. I would love to be able to get to an output showing sorted by the p value for the lm for each factor level. I am trying to ask the question, for each factor ( ALPHA, OMEGA, GAMA, et al), which one best explains the relationship between pfRet and ret?

From a programming perspective there's probably a way to get a vector of the 6 factor levels and then apply lm over the frame. But I couldn't get it to work right. I tried something along the lines of:

GenLm = function(x) { 
  y=pf[which[(pf$Src==x,)]
  return(lm(y$pfRet ~ y$ret )) }

testList = data.frame( Src = as.character(unique( pf$Src ) ) )

testList$lm = apply(  ... ?? )

I tried various combinations of apply, ddply, etc. I just couldn't get it into an easy format where an output could be generated showing the fits for each variable.

Thanks, Josh From a higher level perspective, is this the right way to go about it? This is something surely that R users deal with all of the time.

also see results of using `dplyr`, e.g. http://stackoverflow.com/questions/24376765/dplyrdo-requires-named-function — Ben Bolker, Jan 31 '16 at 03:34

score 2 · Accepted Answer · answered Jan 31 '16 at 03:36

2

lmList, in nlme or lme4 packages, will do this out of the box (they're basically equivalent, except that the lme4 one can handle GLMs as well as linear models, but might be slightly less robust in some cases)

For example:

library("lme4")
mm <- lmList(Reaction~Days|Subject,sleepstudy)
summary(mm)

answered Jan 31 '16 at 03:36

Ben Bolker

211,554
25
370
453

Guys, thanks so much. That is very helpful. On the how-to-program it front that answers it. – JoshK Jan 31 '16 at 15:40
Ty. On the conceptual side, is the right way is just to dcast the data and do a multiple regression? If you have two factors and they are pretty close to each-other would it give a partial wt on two of the independent vars? eg, if you were regressing `taxi.fare~miles.traveled+car.time`, let's say that miles traveled was the best fit in terms of residual error and p value. But time.in.car would probably be close. So what's the right approach? Would a multiple regression be the right way to do this? Or should I do it the way that I was, with separate regressions on each variable? – JoshK Jan 31 '16 at 15:48

R lm on each factor level

1 Answers1