[R] Mixture of Distributions

David Winsemius dwinsemius at comcast.net
Tue Mar 4 05:20:36 CET 2008


(Ted Harding) <Ted.Harding at manchester.ac.uk> wrote in
news:XFMail.080221222007.Ted.Harding at manchester.ac.uk: 

> I just realised I made a bad mistaqke (see below)
> 
> On 21-Feb-08 21:39:56, Ted Harding wrote:
>> On 21-Feb-08 20:58:25, Evgenia wrote:
>>> 
>>> Dear R users, 
>>> I would like to sample from a mixture distribution
>>> p1*f1+p2*f2+p3*f3 with f1,f2,f3 three different forms
>>> of distributions. I know that in the case of two
>>> distributions I have to sample the mixture compoment
>>> membership.
>>> 
>>> How can I weight my three distributions with their
>>> respective probabilities?
>>> 
>>> Evgenia
>> 
>> There are several ways in which you could write code
>> to do this, but all amount to the fact that each value
>> sampled from such a mixture is obtained by
>> a) choose between f1, f2, f3 at random according to the
>>    probabilities p1, p2, p3 (it is assumed p1+p2+p3=1).
>> b) sample 1 value from whichever of f1, f2, f3 was chosen.
>> 
>> Suggestion:
>> 
>> Suppose the functions rf1(n), rf2(n), rf3(n) respectively
>> sample n values from f1, f2 and f3.
> 
> S0 <- cbind(rf1(n),rf2(n),rf3(n))
> ix <- sample(c(1,2,3),n,3,prob=c(p1,p2,p3),replace=TRUE)
> 
> So far so good!
> 
> S  <- S0[,ix]
> 
> But this will not do what I intended (i.e. select element ix[1]
> from the first row os S0, ix[2] from the second row, and so on).
> 
> Instead, it will return an nxn matrix with n rows, and n columns
> which are copies of columns of S0 in the order selected by ix!
> 
> The following will do the correct thing, though there must
> be a neater way of doing it! The example below has been
> corrected.
> 
> S <- S0[,1]*(ix==1) + S0[,2]*(ix==2) + S0[,3]*(ix==3)

I discovered a shorter method (and I even think I might understand why 
it works.) Try:

 S0[col(S0)==ix]

R loops through the S0 matrix and finds only those entries where the 
column numbers match the ix[] values. I would have thought that the 
first row and  ix[1]th column entry should be first but the machinery 
seems to be working differently. The user should realize that all the 
column==1 hits will occur first

#-----toy code-------
rf1 <- function(n){rnorm(n,0,1)} ## normal distribution with mean=0
rf2 <- function(n){rnorm(n,4,1)} ## normal distribution with mean=4
rf3 <- function(n){rnorm(n,9,1)} ## normal distribution with mean=9
S0 <- cbind(rf1(10),rf2(10),rf3(10))
p1 <- 0.2; p2 <- 0.3; p3 <-0.5
set.seed<-37
ix <- sample(c(1,2,3),10,prob=c(p1,p2,p3),replace=TRUE)
#> ix
# [1] 2 3 1 3 1 2 1 2 2 1
S0
#----------
 S0
            [,1]     [,2]      [,3]
 [1,]  0.5143426 3.823013  9.666210
 [2,]  0.7044110 6.095287 10.276792
 [3,] -0.3597926 5.848588  8.801378
 [4,]  0.1346979 5.518667  8.357181
 [5,] -0.6856507 2.882090  9.893790
 [6,]  0.5882977 4.864201  8.708929
 [7,] -0.4241657 2.891672  8.056254
 [8,]  1.6243569 2.981563 11.684422
 [9,] -0.3072410 4.379916 10.679470
[10,] -0.6289604 3.119195  7.216593
#--------------------------

S0[col(S0)==ix]
# [1]  0.1346979 -0.3072410  3.8230128  5.8485880  4.8642010  2.9815631
# [7] 10.2767916  9.8937904  8.0562542  7.2165927

R seems to be looping through the columns "looking" for the TRUE values 
in the col(S0)==ix matrix.

col(S0)==ix
       [,1]  [,2]  [,3]
 [1,] FALSE  TRUE FALSE
 [2,] FALSE FALSE  TRUE
 [3,] FALSE  TRUE FALSE
 [4,]  TRUE FALSE FALSE
 [5,] FALSE FALSE  TRUE
 [6,] FALSE  TRUE FALSE
 [7,] FALSE FALSE  TRUE
 [8,] FALSE  TRUE FALSE
 [9,]  TRUE FALSE FALSE
[10,] FALSE FALSE  TRUE

-- 
Regards;
David Winsemius



More information about the R-help mailing list