Unexpected results from "replicate" (compared to saply)

[Paul Johnson <>]

Recall my heteroskedasicity example code "heterosk.R".   The management 
of the Monte Carlo models was a bit tedious.

The following creates a list of regression objects, or at least I wish 
it would.  The result is not a list of 100 regression objects, but 
rather a 12x100 array in which the row represents the component of model 
output and the column represents the model.

-----------------------------
# create datasets with 10000 observations each, conduct
# a regression
createOLS <- function(b0=10, b1=13){
 e <- rnorm(10000)
 x <- 10*rnorm(10000)+100
 y <- b0 + b1*x + 1*x*x *e
 myOLS1 <- lm(y~x)
}

oneHundredRegressions <- replicate(100, createOLS(3, 55))

oneHundredBVectors <- oneHundredRegressions[1,]

-----------------------------

oneHundredRegressions appears as a 2 dimensional array, with one column 
for each run of the model and one row for each element outputted.  
Example for 4 models:

> oneHundredRegressions
             [,1]          [,2]          [,3]          [,4]        
coefficients  Numeric,2     Numeric,2     Numeric,2     Numeric,2   
residuals     Numeric,10000 Numeric,10000 Numeric,10000 Numeric,10000
effects       Numeric,10000 Numeric,10000 Numeric,10000 Numeric,10000
rank          2             2             2             2           
fitted.values Numeric,10000 Numeric,10000 Numeric,10000 Numeric,10000
assign        Integer,2     Integer,2     Integer,2     Integer,2   
qr            List,5        List,5        List,5        List,5      
df.residual   9998          9998          9998          9998        
xlevels       List,0        List,0        List,0        List,0      
call          Expression    Expression    Expression    Expression  
terms         Expression    Expression    Expression    Expression  
model         List,2        List,2        List,2        List,2      

One can obtain the summary output for a particular element by this

summary.lm(oneHundredRegressions[,1])

But
> vcov(oneHundredRegressions[,1])
Error in vcov(oneHundredRegressions[, 1]) :
   no applicable method for "vcov"
> vcov.lm(oneHundredRegressions[,1])
Error: couldn't find function "vcov.lm"
>

On the contrary, I found that if I used the lapply command to replicate, 
then the results were what I wanted.

oneHundredLapply <- lapply(1:100, function(x) createOLS(4,5))

With that, one can ask for particular regressions easily, as in

oneHundredLapply[[1]]

and vcov and hccm work.


After wresting with this for quite a while, I read the help page for 
sapply and replicate and realized that the simplify=T option is default 
for replicate, and that "breaks" the object output into the two 
dimensional array that I don't want. So if I fix my code by running

oneHundredRegressions <- replicate(100, createOLS(3, 55), simplify=F)

The result is the kind I want.  I've put this to use in a revision of my 
"heterosk.R" example program.