Skip to content

predict() for bootstrapped model coefficients

2 messages · Sascha Vieweg

Sascha Vieweg
# 
I run a multinomial regression on a data set with an outcome that 
has three values. First, I build an initial model, b.mod. Then I 
run a loop to bootstrap the coefficients. For the initial model, 
using "predict()", I can print the wrong/false predictions table. 
But how do I get this table (that is, the predictions) for the 
bootstrapped model? Thanks for hints, *S*

df <- data.frame(
   "y"=factor(sample(LETTERS[1:3], 200, repl=T)),
   "a"=rnorm(200),
   "b"=rnorm(200)
)
library(nnet)
b.mod <- multinom(y ~ ., data=df, trace=F)
i <- 1; k <- 50
c.mat.1 <- matrix(nrow=k, ncol=length(names(df)), 
dimnames=list(NULL, c("cons", names(df)[-1])))
c.mat.2 <- matrix(nrow=k, ncol=length(names(df)), 
dimnames=list(NULL, c("cons", names(df)[-1])))
set.seed(123)
while(i <= k){
   l <- sample(1:length(df[, 1]), replace=T)
   m <- update(b.mod, . ~ ., data=df[l, ], trace=F)
   c.mat.1[i, ] <- coef(m)[1, ]
   c.mat.2[i, ] <- coef(m)[2, ]
   i <- i + 1
}
(coefs1 <- apply(c.mat.1, 2, mean))
(coefs2 <- apply(c.mat.2, 2, mean))
summary(predict(b.mod)==model.frame(b.mod)$y)
1 day later
Sascha Vieweg
#
On 11-01-20 17:05, Sascha Vieweg wrote:

            

      
      
      
    

        
I give it a second trial to poll some ideas. I *mean* that predict 
uses information from the model object, here b.mod. My idea was 
now to "fake" an object and fill it with relevant data from my 
bootstrap to run predict on that faked object:

f.b.mod <- b.mod
f.b.mod$n <- 14:16 # senseless, anyway
f.b.mod[1:3]

What I don't know, and don't get from the predict() source code, 
and thus requesting help upon is, which parts of the model object 
does predict use and how can I replace them with my loop data? 
Thanks for hints, *S*