Dear All:
I need advice about efficient looping/vectorization. I am trying to
bootstrap a regression model with one lag of the dependent variable in
the RHS. Specifically, let error^b_(t) be the bootstrapped error of
the regression y_(t) = gamma y_(t-1) + beta x +error_(t) at time (t),
y_(t) is the original dependent variable, and y^b_(t) the bootstraped
y_(t) using parameter estimates gamma and beta. My basic procedure is
like this:
1. Get the first y^b value using y_(1):
y^b_(2) = gamma y_(1) + beta x_(2) + error_(2).b
2. Get the other y^bs:
y^b_(3) = gamma y^b_(2) + beta x_(3) + error_(3).b
?.
?.
?y^b_(T) = gamma y^b_(t-1) + beta x_(T) + error_(4).b
however, my approach that uses a loop similar to the one below, is
extremely slow. In my actual situation I am dealing with observations
indexed over time and cross-sections, however, I thought that it was
simpler to ask my question considering only one source of variation.
Let's suppose that the dataset look like this:
Please, for my question to make any sense, imagine that my bootstraped
errors are identical to the 'res' above, although of course, in
practice they will not be; I just want to keep things simple here:
I first get the first value of y^b_(2) using y_(1):
d$y.b[2]<-with(d, gamma*y[1] + beta*x[2] ?+ res[2]) #y^b_(2) pretending that res[2] is a resampled error.
And then, fill the rest of the values using this loop:
for(.t in 3:length(d$time)){
? ? ? d$y.b[c(.t)]<-gamma*d$y.b[c(.t-1)] + beta*d$x[c(.t)] ?+ d$res[c(.t)] #pretending that res are resampled errors.
? ? ? }
My problem is that this becomes very slow --- painfully slow when I am
actually botstraping --- when I have several observations or more than
one index (hence more looping levels) in the dependent variable. I
tried to use lapply, but it seems to me that it is not adequate when
you have a recursive situation such the one above. Any hint?
I appreciate any help,
Nelson Villoria