R/rasch.va.R

Defines functions rasch.va

Documented in rasch.va

## File Name: rasch.va.R
## File Version: 0.13


# estimating the Rasch model with variational approximation
rasch.va <- function( dat, globconv=.001, maxiter=1000)
{
    # data preparation
    N <- nrow(dat)
    I <- ncol(dat)
    dat2 <- dat
    dat2.resp <- 1 - is.na(dat)
    dat2[ dat2.resp==0 ] <- 0
    # initial values
    b <- - stats::qlogis( colMeans( dat, na.rm=T ) )
    sig2 <- 1
    mu.i <- rep(0,N)
    mu.i <- .8*stats::qnorm( rowMeans( ( dat + .1 ) / 1.2 ) )
    sigma2.i <- rep( sig2, N )
    disp <- "...........................................................\n"
    iter <- 0
    conv <- 1000

    #***************
    # begin algorithm
    while( (globconv < conv)  &    ( iter < maxiter )    ){
        cat(disp)
        cat("Iteration", iter+1, "   ", paste( Sys.time() ), "\n" )
        b0 <- b
        # compute xsi.ij: formula (18)
        xsi.ij <- matrix( b^2, N, I, byrow=T) -       # (x'beta)^2
                    2 * matrix( b, N, I, byrow=T ) * mu.i +  # 2nd term
                    mu.i^2 + sigma2.i
        xsi.ij <- sqrt( xsi.ij )
        lam.xsi.ij <- tanh( xsi.ij / 2 ) / ( 4*xsi.ij )
        # update beta
        t1 <- colSums( 2 * lam.xsi.ij  * dat2.resp )
        b <- colSums( - ( dat2 - .5 ) * dat2.resp + 2 * lam.xsi.ij * mu.i ) / t1
        # update sigma2
        sig2 <- mean( mu.i^2 + sigma2.i )
        # update sigma2.i
        sigma2.i <- 1 / ( 2 * rowSums( lam.xsi.ij ) + 1 / sig2  )
        # update mu.i
        mu.i <- rowSums( ( dat2 - 1/2 + 2*lam.xsi.ij * matrix( b, N, I, byrow=TRUE ) )
                                * dat2.resp  ) * sigma2.i
        # display progress
        conv <- max( abs(b-b0))
        iter <- iter+1
        cat( paste( "    Maximum b parameter change=",
                paste( round(max(abs(b-b0)),6), collapse=" " ), "\n", sep=""))
        cat( paste( "    SD Trait=",
                paste( round(sqrt(sig2),3), collapse=" " ), "\n", sep=""))
        utils::flush.console()
    }
    item <- data.frame("item"=colnames(dat), "b"=b )
    res <- list(sig=sqrt(sig2), item=item, xsi.ij=xsi.ij, mu.i=mu.i, sigma2.i=sigma2.i)
    return(res)
}
alexanderrobitzsch/sirt documentation built on March 18, 2024, 1:29 p.m.