floor of n observations in number generators

I couldnt find a previous posting on this in the archives, maybe it has
already been mentioned.

If you use a calculation to generate n observations in random number
generators and you don't round to the nearest integer you may be
generating n-1 numbers not n numbers as you thought depending on the
storage precision of the calculation. 

e.g.

m <- 1000
pi0 <- 0.9
length(rnorm(m * (1-pi0)))

[1] 99  # Should be 100

options(digits=16)
m * (1-pi0)

[1] 99.99999999999997

identical(m*(1-pi0), 100)

[1] FALSE

Random number generation generates the floor of n observations, this
feature occurs on R-1.8.1 on linux Redhat8, and winXP (also on Unix
SPlus 3.4) 
for probably all of the random number generators.

e.g.

length(rnorm(m*(1-pi0),mean=0,sd=1))

[1] 99

length(rpois(m*(1-pi0),lambda=1))

[1] 99

length(rbeta(m*(1-pi0),shape=1, shape2=2))

[1] 99

length(rbinom(m*(1-pi0),size=1, prob=0.5))

[1] 99

length(runif(m*(1-pi0),min=0, max=1))

[1] 99


marcus

Marcus Davy
Bioinformatics


______________________________________________________
The contents of this e-mail are privileged and/or confidenti...{{dropped}}