Styjun

Arguments

• data: an optional data frame containing the variables in the model. By default the variables are taken from the environment which randomForestis called from




• subset: an index vector indicating which rows should be used. (NOTE: If given, this argument must be named.)

My questions:

1. Why is data argument "optional"? If data is optional, where does the training data come from? And what exactly is the meaning of "By default the variables are taken from the environment which randomForestis called from"?




2. Why do we need the subset parameter? Let's say, we have the iris data set. If I want to use the first 100 rows as the training data set, I just select training_data <- iris[1:100,]. Why bother? What's the benefit of using subset?

1. 
   

   This is not an uncommon methodology, and certainly not unique to randomForests.

   
   
   

   mpg <- mtcars$mpg
   disp <- mtcars$disp
   lm(mpg~disp)
   # Call:
   # lm(formula = mpg ~ disp)
   # Coefficients:
   # (Intercept) disp 
   # 29.59985 -0.04122 
   

   
   
   

   So when lm (in this case) is attempting to resolve the variables referenced in the formula mpg~disp, it looks at data if provided, then in the calling environment. Further example:

   
   
   

   rm(mpg,disp)
   mpg2 <- mtcars$mpg
   lm(mpg2~disp)
   # Error in eval(predvars, data, env) : object 'disp' not found
   lm(mpg2~disp, data=mtcars)
   # Call:
   # lm(formula = mpg2 ~ disp, data = mtcars)
   # Coefficients:
   # (Intercept) disp 
   # 29.59985 -0.04122 
   

   
   
   

   (Notice that mpg2 is not in mtcars, so this used both methods for finding the data. I don't use this functionality, preferring the resilient step of providing all data in the call; it is not difficult to think of examples where reproducibility suffers if this is not the case.

   
   


2. 
   

   Similarly, many similar functions (including lm) allow this subset= argument, so the fact that randomForests includes it is consistent. I believe it is merely a convenience argument, as the following are roughly equivalent:

   
   
   

   lm(mpg~disp, data=mtcars, subset= cyl==4)
   
   lm(mpg~disp, data=mtcars[mtcars$cyl == 4,])
   
   mt <- mtcars[ mtcars$cyl == 4, ]
   lm(mpg~disp, data=mt)
   

   
   
   

   The use of subset allows slightly simpler referencing (cyl versus mtcars$cyl), and its utility is compounded when the number of referenced variables increases (i.e., for "code golf" purposes). But this could also be done with other mechanisms such as with, so ... mostly personal preference.

   
   

Edit: as joran pointed out, randomForest (and others but notably not lm) can be called with either a formula, which is where you'd typically use the data argument, or by specifying the predictor/response arguments separately with the arguments x and y, as in the following examples taken from ?randomForest (ignore the other arguments being inconsistent):

iris.rf <- randomForest(Species ~ ., data=iris, importance=TRUE, proximity=TRUE)
iris.rrf <- randomForest(iris[-1], iris[[1]], ntree=101, proximity=TRUE, oob.prox=FALSE)