MeasurementYCat: Sampler for Partial Mediation Model with Multiple Categorical...

In arashl1364/BFMediate: Mediation Analysis Using a Combination of Bayesian Estimation Methods, Latent Variable Models, and Bayes Factors

Sampler for Partial Mediation Model with Multiple Categorical Indicator for the DV

Description

Estimates a partial mediation model with multiple categorical indicator for the dependent variable

Usage

MeasurementYCat(Data,Prior,R)

Arguments

Data	list(X, M, y_star)
Prior	list(A_M,A_Y)
R	number of MCMC iterations, default = 10000

Details

Model

M = \beta_{0M} + X\beta_1 + U_M (eq.1)
Y = \beta_{0Y} + M\beta_2 + X\beta_3 + U_Y (eq.2)

Indicator equations

y^*_1 = Y + U_{y^*_1}
\tilde{y}_1 = OrdProbit(y^*_1 ,C_{y_1})
y^*_2 = \tau_{20} + Y + U_{y^*_2}
\tilde{y}_2 = OrdProbit(y^*_2 ,C_{y_2})
...
y^*_L = \tau_{L0} + Y + U_{y^*_L}
\tilde{y}_L = OrdProbit(y^*_L ,C_{y_L})

Prior specification:

\beta_{0M} \sim N(0,100), \beta_{0Y} \sim N(0,100)
\beta_1 \sim N(0,100), \beta_2 \sim N(0,100), \beta_3 \sim N(0,1)
\sigma^2_M \sim Inv\chi^2(\nu,\nu S) , where \nu=1 and S=3.
\tau_{20}, ..., \tau_{L0} \sim N(0,100)
\sigma^2_{y*_1}, ..., \sigma^2_{y*_L} \sim Inv\chi^2(\nu,\nu S)
C*_{y_1}, ..., C*_{y_L} \sim N(0,I)

Note: C*_{y_1}, ..., C*_{y_L} are untransformed cutoffs, which are then exponentially transformed to impose sign and order constraint on them. Subjective prior values for the error variances are \nu=1, S=3.

Argument Details

Data = list(X, M, y_star)

X(N x 1)

treatment variable vector

M(N x 1)

mediator vector

y_star(N x Y_ind)

dependent variable indicators' matrix

Prior = list(A_M,A_Y) [optional]

A_M

vector of coefficients' prior variances of eq.1, default = rep(100,2)

A_Y

vector of coefficients' prior variances of eq.2, default = c(100,100,1)

Value

beta_M(R X 2)

matrix of eq.1 coefficients' draws

beta_Y(R X 3)

matrix of eq.2 coefficients' draws

tau(Y_ind X 2 X R)

array of indicator coefficients' draws. Each slice is one draw, where rows represent the indicator equation and columns the coefficients. All Slope coefficients as well as intercept of the first equation are fixed to 1 and 0 respectively.

ssq_y_star(R X Y_ind)

Matrix of indicator equations' coefficients' error variance draws

ssq_M(R X 1)

vector of eq.1 error variance draws

cutoff_Y (Y_ind X k_Y X R)

array of discretized dependent variable indicators' cutoff values.

Ydraw(R X N)

Matrix of the augmented latent dependent variable

mu_draw

vector of means indexing MCMC draws of the direct effect (used in BFSD to compute Bayes factor)

var_draw

vector of variance indexing MCMC draws of the direct effect (used in BFSD to compute Bayes factor)

Examples

set.seed(60)
SimMeasurementYCat = function(X, beta_M, beta_Y, sigma_M, cutoff_Y, Y_ind, tau, ssq_y_star){

  nobs = dim(X)[1]
  y_tilde = y_star = matrix(double(nobs*Y_ind), ncol = Y_ind)

  M = beta_M[1] + beta_M[2] * X + rnorm(nobs) * sigma_M
  Y = beta_Y[1] + beta_Y[2] * M + beta_Y[3] * X + rnorm(nobs)

  for(i in 1: Y_ind){
    y_star[,i] = tau[i] + Y + sqrt(ssq_y_star[i])*rnorm(nobs);
    y_tilde[,i] = cut(y_star[,i], br = cutoff_Y[i,], right=TRUE, include.lowest = TRUE, labels = FALSE)
  }

  return(list(Y = Y, M = M, y_tilde = y_tilde, X = X,
              beta_M = beta_M,
              k_Y=dim(cutoff_Y)[2]-1, beta_Y = beta_Y, tau = tau,
             ssq_y_star = ssq_y_star, y_star = y_star, cutoff_Y = cutoff_Y,
              Y_ind=Y_ind))
}

Y_ind = 2
Ycut = 8
nobs = 1000
X=as.matrix(runif(nobs,min=0, max=1))
beta_M = c(.5,1)
beta_Y = c(1, 4, 2)
sigma_M = 1^.5
ssq_y_star = c(.5,.7)
tau = c(0,-.5)   #first intercept should always be 0
cutoff_Y =  matrix(c(-100, 0, 1.6, 2, 2.2, 3.3, 6,  100,
                     -100, 0, 1, 2, 3, 4, 5, 100) ,ncol= Ycut, byrow = TRUE)
DataYCat = SimMeasurementYCat(X, beta_M, beta_Y, sigma_M, cutoff_Y, Y_ind, tau, ssq_y_star)

#estimation
A_M = c(100,100); #Prior variance for beta_0M, beta_1
A_Y = c(100,100,1) #Prior variance for beta_0Y, beta_2, beta_3
Prior = list(A_M = A_M, A_Y = A_Y)
Ycut = max(as.matrix(DataYCat$y_tilde)[,1]) +1
Data = list(X=cbind(rep(1,length(DataYCat$X)),DataYCat$M,DataYCat$X), y = as.matrix(DataYCat$y_tilde),
            k=Ycut-1, Y_ind=dim(as.matrix(DataYCat$y_tilde))[2])
out = MeasurementYCat(Data=Data, Prior=Prior, R=10000)

#results
colMeans(out$beta_M)
colMeans(out$beta_Y)
apply(out$cutoff_Y,c(1,2),FUN = mean)

arashl1364/BFMediate documentation built on Oct. 11, 2023, 5:54 p.m.