R help - Apr 2006 - Function to approximate complex integral

I am writing a small function to approximate an integral that cannot be
evaluated in closed form. I am partially successful at this point and am
experiencing one small, albeit important problem. Here is part of my
function below.

This is a psychometric problem for dichotomously scored test items where
x is a vector of 1s or 0s denoting whether the respondent answered the
item correctly (1) or otherwise (0), b is a vector of item difficulties,
and theta is an ability estimate for the individual.

rasch <- function(b,theta){
   1 / ( 1 + exp(b - theta))
   }

The function rasch gives the probability of a correct response to item i
conditional on theta, the individuals ability estimate

myfun <- function(x, b, theta){
   sum(rasch(b, theta)^x * ( 1 - rasch(b,theta) )^(1-x) * dnorm(theta))
   }

This is the likelihood function assuming the data are Bernoulli
distributed multiplied by a population distribution.

Now, when x,b, and theta are of equal length the function works fine as
below
x <- c(1,1,0)
b <-   c(-2,-1,0)
t <-   c(-2,-1.5,-1)
> myfun(x,b,t)
[1] 0.2527884

However, I want theta to be a vector of discrete values that will be
larger than both x and b, something like 

t <- seq(-5, 0, by = .01)

However, this gives me the error 
> myfun(x,b,t)
 Warning messages:
1: longer object length
        is not a multiple of shorter object length in: b - theta 

So, for the problem above, I want item 1 to be evaluated at theta 1
through theta q and then item 2 is evaluated at theta 1 and through
theta q and so forth for each item.

Can anyone recommend a way for me to modify my function above to
accomplish this?

Harold


	[[alternative HTML version deleted]]

On 4/19/06, Doran, Harold <HDoran at air.org> wrote:
> I am writing a small function to approximate an integral that cannot be
> evaluated in closed form. I am partially successful at this point and am
> experiencing one small, albeit important problem. Here is part of my
> function below.
> This is a psychometric problem for dichotomously scored test items where
> x is a vector of 1s or 0s denoting whether the respondent answered the
> item correctly (1) or otherwise (0), b is a vector of item difficulties,
> and theta is an ability estimate for the individual.
>
> rasch <- function(b,theta){
>    1 / ( 1 + exp(b - theta))
>    }
>
> The function rasch gives the probability of a correct response to item i
> conditional on theta, the individuals ability estimate
>
> myfun <- function(x, b, theta){
>    sum(rasch(b, theta)^x * ( 1 - rasch(b,theta) )^(1-x) * dnorm(theta))
>    }
>
> This is the likelihood function assuming the data are Bernoulli
> distributed multiplied by a population distribution.
>
> Now, when x,b, and theta are of equal length the function works fine as
> below
> x <- c(1,1,0)
> b <-   c(-2,-1,0)
> t <-   c(-2,-1.5,-1)
> > myfun(x,b,t)
> [1] 0.2527884
>
> However, I want theta to be a vector of discrete values that will be
> larger than both x and b, something like
>
> t <- seq(-5, 0, by = .01)
>
> However, this gives me the error
> > myfun(x,b,t)
>
>  Warning messages:
> 1: longer object length
>         is not a multiple of shorter object length in: b - theta
>
> So, for the problem above, I want item 1 to be evaluated at theta 1
> through theta q and then item 2 is evaluated at theta 1 and through
> theta q and so forth for each item.
>
> Can anyone recommend a way for me to modify my function above to
> accomplish this?
I believe (partially based on our off-list discussions of this model)
that you want to use the function "outer", as in

rasch <- function(b, theta) 1/(1 - exp(outer(theta, b, "-")))

If length(theta) == m and length(b) == n then this returns the m by n
matrix of probabilities of subject i correctly answering question j.