brm-Methods: Functions for the Beta Item Response Model

Description

Functions for simulating and estimating the Beta item response model (Noel & Dauvier, 2007). brm.sim can be used for simulating the model, brm.irf computes the item response function. The Beta item response model is estimated as a discrete version to enable estimation in standard IRT software like mirt or TAM packages.

Usage

# simulating the beta item response model
brm.sim(theta, delta, tau, K=NULL)
# computing the item response function of the beta item response model
brm.irf( Theta, delta, tau, ncat, thdim=1, eps=1E-10 )

Value

A simulated dataset of item responses if brm.sim is applied.

A matrix of item response probabilities if brm.irf is applied.

Arguments

theta: Ability vector of $\theta$ values
delta: Vector of item difficulty parameters
tau: Vector item dispersion parameters
K: Number of discretized categories. The default is NULL which means that the simulated item responses are real number values between 0 and 1. If an integer K chosen, then values are discretized such that values of 0, 1, ..., $K$-1 arise.
Theta: Matrix of the ability vector $\bold{\theta}$
ncat: Number of categories
thdim: Theta dimension in the matrix Theta on which the item loads.
eps: Nuisance parameter which stabilize probabilities.

Details

The discrete version of the beta item response model is defined as follows. Assume that for item $i$ there are $K$ categories resulting in values $k=0,1,\dots,K-1$. Each value $k$ is associated with a corresponding the transformed value in $[0,1]$, namely $ q (k)=1/(2 \cdot K), 1/(2 \cdot K) + 1/K, \ldots, 1 - 1/(2 \cdot K) $. The item response model is defined as $$ P( X_{pi}=x_{pi} | \theta_p) \propto q( x_{pi} )^{ m_{pi} - 1 } [ 1- q( x_{pi} ) ]^{ n_{pi} - 1 } $$ This density is a discrete version of a Beta distribution with shape parameters $m_{pi}$ and $n_{pi}$. These parameters are defined as $$ m_{pi}=\mathrm{exp} \left[ ( \theta_p - \delta_i + \tau_i ) / 2 \right] \qquad \mbox{and} \qquad n_{pi}=\mathrm{exp} \left[ ( - \theta_p + \delta_i + \tau_i ) / 2 \right] $$

The item response function can also be formulated as $$ \mathrm{log} \left[ P( X_{pi}=x_{pi} | \theta_p) \right] \propto ( m_{pi} - 1 ) \cdot \mathrm{log} [ q( x_{pi} ) ] + ( n_{pi} - 1 ) \cdot \mathrm{log} [ 1- q( x_{pi} ) ] $$

The item parameters can be reparameterized as $ a_{i}=\mathrm{exp} \left[ ( - \delta_i + \tau_i ) / 2 \right]$ and $ b_{i}=\mathrm{exp} \left[ ( \delta_i + \tau_i ) / 2 \right]$.

Then, the original item parameters can be retrieved by $\tau_i=\mathrm{log} ( a_i b_i)$ and $\delta_i=\mathrm{log} ( b_i / a_i)$. Using $ \gamma _p=\mathrm{exp} ( \theta_p / 2) $, we obtain

$$ \mathrm{log} \left[ P( X_{pi}=x_{pi} | \theta_p) \right] \propto a_{i} \gamma_p \cdot \mathrm{log} [ q( x_{pi} ) ] + b_i / \gamma_p \cdot \mathrm{log} [ 1- q( x_{pi} ) ] - \left[ \mathrm{log} q( x_{pi} ) + \mathrm{log} [ 1- q( x_{pi} ) ] \right] $$

This formulation enables the specification of the Beta item response model as a structured latent class model (see TAM::tam.mml.3pl; Example 1).

See Smithson and Verkuilen (2006) for motivations for treating continuous indicators not as normally distributed variables.

References

Gruen, B., Kosmidis, I., & Zeileis, A. (2012). Extended Beta regression in R: Shaken, stirred, mixed, and partitioned. Journal of Statistical Software, 48(11), 1-25. tools:::Rd_expr_doi("10.18637/jss.v048.i11")

Noel, Y., & Dauvier, B. (2007). A beta item response model for continuous bounded responses. Applied Psychological Measurement, 31(1), 47-73. tools:::Rd_expr_doi("10.1177/0146621605287691")

Smithson, M., & Verkuilen, J. (2006). A better lemon squeezer? Maximum-likelihood regression with beta-distributed dependent variables. Psychological Methods, 11(1), 54-71. doi: 10.1037/1082-989X.11.1.54

Examples

Run this code

#############################################################################
# EXAMPLE 1: Simulated data beta response model
#############################################################################

#*** (1) Simulation of the beta response model
# Table 3 (p. 65) of Noel and Dauvier (2007)
delta <- c( -.942, -.649, -.603, -.398, -.379, .523, .649, .781, .907 )
tau <- c( .382, .166, 1.799, .615, 2.092, 1.988, 1.899, 1.439, 1.057 )
K <- 5        # number of categories for discretization
N <- 500        # number of persons
I <- length(delta) # number of items

set.seed(865)
theta <- stats::rnorm( N )
dat <- sirt::brm.sim( theta=theta, delta=delta, tau=tau, K=K)
psych::describe(dat)

#*** (2) some preliminaries for estimation of the model in mirt
#*** define a mirt function
library(mirt)
Theta <- matrix( seq( -4, 4, len=21), ncol=1 )

# compute item response function
ii <- 1     # item ii=1
b1 <- sirt::brm.irf( Theta=Theta, delta=delta[ii], tau=tau[ii],  ncat=K )
# plot item response functions
graphics::matplot( Theta[,1], b1, type="l" )

#*** defining the beta item response function for estimation in mirt
par <- c( 0, 1,  1)
names(par) <- c( "delta", "tau","thdim")
est <- c( TRUE, TRUE, FALSE )
names(est) <- names(par)
brm.icc <- function( par, Theta, ncat ){
     delta <- par[1]
     tau <- par[2]
     thdim <- par[3]
     probs <- sirt::brm.irf( Theta=Theta, delta=delta, tau=tau,  ncat=ncat,
            thdim=thdim)
     return(probs)
            }
name <- "brm"
# create item response function
brm.itemfct <- mirt::createItem(name, par=par, est=est, P=brm.icc)
#*** define model in mirt
mirtmodel <- mirt::mirt.model("
           F1=1-9
            " )
itemtype <- rep("brm", I )
customItems <- list("brm"=brm.itemfct)

# define parameters to be estimated
mod1.pars <- mirt::mirt(dat, mirtmodel, itemtype=itemtype,
                   customItems=customItems, pars="values")

if (FALSE) {
#*** (3) estimate beta item response model in mirt
mod1 <- mirt::mirt(dat,mirtmodel, itemtype=itemtype, customItems=customItems,
               pars=mod1.pars, verbose=TRUE  )
# model summaries
print(mod1)
summary(mod1)
coef(mod1)
# estimated coefficients and comparison with simulated data
cbind( sirt::mirt.wrapper.coef( mod1 )$coef, delta, tau )
mirt.wrapper.itemplot(mod1,ask=TRUE)

#---------------------------
# estimate beta item response model in TAM
library(TAM)

# define the skill space: standard normal distribution
TP <- 21                   # number of theta points
theta.k <- diag(TP)
theta.vec <-  seq( -6,6, len=TP)
d1 <- stats::dnorm(theta.vec)
d1 <- d1 / sum(d1)
delta.designmatrix <- matrix( log(d1), ncol=1 )
delta.fixed <- cbind( 1, 1, 1 )

# define design matrix E
E <- array(0, dim=c(I,K,TP,2*I + 1) )
dimnames(E)[[1]] <- items <- colnames(dat)
dimnames(E)[[4]] <- c( paste0( rep( items, each=2 ),
        rep( c("_a","_b" ), I) ), "one" )
for (ii in 1:I){
    for (kk in 1:K){
      for (tt in 1:TP){
        qk <- (2*(kk-1)+1)/(2*K)
        gammap <- exp( theta.vec[tt] / 2 )
        E[ii, kk, tt, 2*(ii-1) + 1 ] <- gammap * log( qk )
        E[ii, kk, tt, 2*(ii-1) + 2 ] <- 1 / gammap * log( 1 - qk )
        E[ii, kk, tt, 2*I+1 ] <- - log(qk) - log( 1 - qk )
                    }
            }
        }
gammaslope.fixed <- cbind( 2*I+1, 1 )
gammaslope <- exp( rep(0,2*I+1) )

# estimate model in TAM
mod2 <- TAM::tam.mml.3pl(resp=dat, E=E,control=list(maxiter=100),
              skillspace="discrete", delta.designmatrix=delta.designmatrix,
              delta.fixed=delta.fixed, theta.k=theta.k, gammaslope=gammaslope,
              gammaslope.fixed=gammaslope.fixed, notA=TRUE )
summary(mod2)

# extract original tau and delta parameters
m1 <- matrix( mod2$gammaslope[1:(2*I) ], ncol=2, byrow=TRUE )
m1 <- as.data.frame(m1)
colnames(m1) <- c("a","b")
m1$delta.TAM <- log( m1$b / m1$a)
m1$tau.TAM <- log( m1$a * m1$b )

# compare estimated parameter
m2 <- cbind( sirt::mirt.wrapper.coef( mod1 )$coef, delta, tau )[,-1]
colnames(m2) <- c(  "delta.mirt", "tau.mirt", "thdim","delta.true","tau.true"   )
m2 <- cbind(m1,m2)
round( m2, 3 )
}

Run the code above in your browser using DataLab