Bayesian Mixture of Latent Class Analysis Models with the Telescoping Sampler

In telescope: Bayesian Mixtures with an Unknown Number of Components

Introduction

In this vignette we fit a Bayesian mixture where each component distribution is a latent class analysis (LCA) model and where a prior on the number of components $K$ is specified. We use a synthetic multivariate binary data set which contains 30 binary variables and is composed of three groups. Within each group the variables are dependent. For the detailed simulation setting see @Malsiner+Gruen+Fruehwirth:2024. We use the prior specification and the telescoping sampler for performing MCMC sampling as proposed in @Malsiner+Gruen+Fruehwirth:2024.

First, we load the package.

library("telescope")

Data

We read in the data and extract the variables used for clustering in y as well as the group memberships in variable z.

data("SimData", package = "telescope")
y <- as.matrix(SimData[, 1:30])
z <- SimData[, 31]

There are r ncol(y) clustering variables for r nrow(y) observations.

dim(y)

The frequency table of the group membership variable indicates that the three groups are similar in size.

table(z)

We store the dimension of the data set and determine the number of categories of each variable.

N <- nrow(y)
r <- ncol(y)
cat <- apply(y, 2, max)

Model specification

The following data model and hierarchical prior structure are specified for modeling the multivariate categorical observations $\mathbf{y}_1,\ldots,\mathbf{y}_N$:

\begin{aligned} p(\mathbf{y}i | K, \boldsymbol{\Theta}_K, \boldsymbol{\eta}_K) &= \sum{k=1}^{K} \eta_k p_k(\mathbf{y}i | \boldsymbol{\theta}_k)\qquad i=1,\ldots,N\ % p_k(\mathbf{y}_i | \boldsymbol{\theta}_k) &= \sum{l=1}^{L} w_{kl} \prod_{j=1}^r\prod_{d=1}^{D_j} \pi_{kl,jd}^{I{y_{ij} = d}}\qquad i=1,\ldots,N\ % K \sim p(K)&\ % \boldsymbol{\eta} \sim \mathcal{D}(e_0)& \qquad \text{with either } e_0 \text{ fixed, or } e_0\sim p(e_0) \text {, or } e_0=\frac{\alpha}{K}, \text{ with } \alpha \text{ fixed or } \alpha \sim p(\alpha),\ % \mathbf{w} \sim \mathcal{D}(d_0)&\qquad \text{with } d_0 \text{ fixed}\ % \boldsymbol{\pi}{kl,j} \sim \mathcal{D}(\boldsymbol{\alpha}{k,j})& \qquad \text{where }\boldsymbol{\alpha}{k,j} = \boldsymbol{\mu}{k,j} \phi_{k,j} + \alpha_{00}\mathbf{1}, \quad l=1,\ldots,L, \quad k=1,\ldots,K, \quad j=1,\ldots,r\ % \boldsymbol{\mu}{k,j} \sim \mathcal{D}(a\mu)& \qquad k=1,\ldots,K, \quad j=1,\ldots,r\ % \phi_{k,j} | b_{\phi_j} \sim \mathcal{G}^{-1}(a_\phi, b_{\phi_j})& \qquad k=1,\ldots,K, \quad j=1,\ldots,r\ b_{\phi_j} \sim \mathcal{G}(c_\phi, d_\phi)& \qquad j =1,\ldots,r. \end{aligned} Note that the parameters of the Dirichlet prior on the component and class weights are called $e_0$ and $d_0$ respectively. For more details see also @Malsiner+Gruen+Fruehwirth:2024.

Specification of the MCMC simulation and prior parameters

For MCMC sampling we need to specify Mmax, the maximum number of iterations, thin, the thinning imposed to reduce auto-correlation in the chain by only recording every thined observation, and burnin, the number of burn-in iterations not recorded.

Mmax <- 300
thin <- 1
burnin <- 100

The specifications of Mmax and thin imply M, the number of recorded observations.

M <- Mmax/thin

We specify with Kmax the maximum number of components possible during sampling. Kinit denotes the initial number of filled components.

Kmax <- 50  
Kinit <- 10

We specify L the number of subcomponents (classes) forming each component, i.e., the number of components in the LCA model.

L <- 2

This number is assumed to be fixed and the same across all components.

For the mixture weights on the higher level we use a dynamic specification.

G <- "MixDynamic"

For a static specific one would need to use "MixStatic".

For the dynamic setup, we specify the gamma distribution $G(1, 2)$ as the prior on alpha.

priorOnAlpha <- priorOnAlpha_spec("gam_1_2")

We need to select the prior on K on the higher level. We specify the prior $K-1 \sim BNB(1, 4, 3)$ as suggested in @Malsiner+Gruen+Fruehwirth:2024.

priorOnK <- priorOnK_spec("BNB_143")

For the lower level, the standard Dirichlet prior with parameter equal to 1 is specified for the class weights within each component.

d0 <- 1  

Now, we specify the component-specific priors. We use a symmetric Dirichlet prior for the occurrence probabilities $\boldsymbol{\pi}_{k,j} \sim \mathcal{D}(a_0)$ for each variable with $a_0=1$ inducing a uniform prior.

a_mu <- rep(20, sum(cat)) 
mu_0 <- matrix(rep(rep(1/cat, cat), Kinit),
               byrow = TRUE, nrow = Kinit)

c_phi <- 30
d_phi <- 1
b_phi <- rep(10, r)

a_phi <- rep(1, r)  
phi_0 <- matrix(cat, Kinit, r, byrow = TRUE)

## regularizing constant
a_00 <- 0.05    

## proposal standard deviations for MH steps for sampling mu and phi, and regularizing constant eps to bound the Dirichlet proposal mu away from the boundary of the simplex
s_mu <- 2  
s_phi <- 2 
eps <- 0.01

We obtain an initial partition S_0 and initial component weights eta_0 using kmeans().

set.seed(1234)
cl_y <- kmeans(y, centers = Kinit, nstart = 100, iter.max = 50)
S_0 <- cl_y$cluster
eta_0 <- cl_y$size/N

Within each component we assign observations to classes and store the class memberships in I_0.

I_0 <- rep(1L, N)
if (L > 1) {
  for (k in 1:Kinit) {
    cl_size <- sum(S_0 == k)
    I_0[S_0 == k] <- rep(1:L, length.out = cl_size)
  }
} 

We determine initial values pi_0 for the occurrence probabilities pi_klj_0 based on initial partitions S_0 and I_0.

index <- c(0, cumsum(cat))
low <- (index + 1)[-length(index)]
up <- index[-1]

pi_km <- array(NA_real_, dim = c(Kinit, L, sum(cat)))
rownames(pi_km) <- paste0("k_", 1:Kinit)
for (k in 1:Kinit) {
    for (l in 1:L) {
        index <- (S_0 == k) & (I_0 == l)
        for (j in 1:r) {
            pi_km[k, l, low[j]:up[j]] <- tabulate(y[index, j], cat[j]) / sum(index)
        }
    }
}
pi_0 <- pi_km 

MCMC sampling

Using this prior specification as well as initialization and MCMC settings, we draw samples from the posterior using the telescoping sampler.

The first two arguments of the sampling function are the data (y is the data matrix where categories are coded as numbers) followed by the initial partitions (S_0, I_0) and the initial parameter values for component-specific occurrence probabilities and weights (eta_0,pi_0, mu_0, phi_0). The next arguments correspond to the hyperparameters of the prior setup (a_00, a_mu, a_phi, b_phi, c_phi, d_phi). Then the setting for the MCMC sampling is specified using M, burnin, thin and Kmax as well as the standard deviations for the proposals in the Metropolis-Hastings steps when sampling $\mu$ and $\phi$ (s_mu, s_phi). Finally the prior specification for the component weights and the prior on the number of components are given (G, priorOnAlpha, d0, priorOnK).

result <- sampleLCAMixture(y, S_0, L, 
                           pi_0, eta_0, mu_0, phi_0,
                           a_00, a_mu, a_phi, b_phi, c_phi, d_phi,
                           M, burnin, thin, Kmax,
                           s_mu, s_phi, eps,
                           G, priorOnAlpha, d0, priorOnK)

The sampling function returns a named list where the sampled parameters and latent variables are contained. The list includes the weights Eta, the assignments S, the number of components K, the number of filled components Kplus, the number of observations Nk and Nl assigned to components and classes within components respectively, the acceptance rate in the Metropolis-Hastings step when sampling $\alpha$ or $e_0$, $\alpha$ and $e_0$, the central component occurrence probabilities Mu, the precisions Phi, parameters b_phi, the acceptance rate in the Metropolis-Hastings step when sampling $\boldsymbol{\mu}{k,j}$ or $\boldsymbol{\phi}{k,j}$.

Finally, for post-processing the draws, parameter values corresponding to the mode of the non-normalized posterior nonnormpost_mode and the functional P_k which are weighted component occurrence probabilities are returned. These values can be extracted for post-processing.

Eta <- result$Eta
S <- result$S
K <- result$K
Kplus <- result$Kplus

Nk <- result$Nk
Nl <- result$Nl
acc <- result$acc
e0 <- result$e0
alpha <- result$alpha

Mu <- result$Mu
Phi <- result$Phi
B_phi <- result$B_phi
acc_mu <- result$acc_mu
acc_phi <- result$acc_phi
nonnormpost_mode <- result$nonnormpost_mode
Pi_k <- result$Pi_k

Convergence diagnostics of the run

We inspect the acceptance rate when sampling $\alpha$ and the trace plot of the sampled alpha:

acc <- sum(acc)/M
acc
plot(1:length(alpha), alpha, type = "l", 
     ylim = c(0, max(alpha)), 
     xlab = "iterations", ylab = expression(alpha))

We further assess the distribution of the sampled $\alpha$ by inspecting a histogram as well determining the mean and some quantiles.

hist(alpha, freq = FALSE, breaks = 50)
mean(alpha)
quantile(alpha, probs = c(0.25, 0.5, 0.75))

In addition we plot the histogram of the induced $e_0$ as well as their mean and some quantiles.

hist(e0, freq = FALSE, breaks = 50)
mean(e0)
quantile(e0, probs = c(0.25, 0.5, 0.75))

We inspect the sampled component mean occurrence distributions mu_kj.

The acceptance rate for the first component and the first variable is given by:

k <- 1  # component
j <- 1  # variable
sum(acc_mu[, k, j])/M

The posterior distributions of the occurrence probabilities may be visualized using a box plot. We create this box plot for the first component, all variables and the second category.

boxplot(Mu[, k, seq(2, 2*r, 2)], xlab = "mu_j")

In addition a trace plot may be used to inspect the sampled values for $\mu_{kjd}$. We do so for the first component, the first variable and the second category.

j <- 1  # variable
d <- 2  # category
k <- 1  # component
plot(Mu[, k, low[j]+(d-1)], type = "l", 
     ylab = paste0("mu_jkd, j=", j, ",d=", d," (k=", k, ")"))

In the same way, we inspect the sampled precision parameter phi_kj based on acceptance rate, box plots and trace plots.

k <- 1  # component
j <- 1  # variable
sum(acc_phi[, k, j])/M

boxplot(Phi[, k, ], xlab = "phi_j",
        ylim = quantile(Phi[, k, ], c(0, 0.95)))

j <- 3  # variable
k <- 1  # component
plot(Phi[, k, j], type = "l",
     ylab = paste0("phi_jkd, j=", j, ", (k=", k, ")"))

To further assess convergence, we also inspect the trace plots for the number of components $K$ and the number of filled components $K_+$.

plot(K, type = "l", ylim = c(0, max(K)), 
     xlab = "iteration", main = "", ylab = "count",
     lwd = 0.5, col = "grey")
points(Kplus, type = "l", col = "red3",
       lwd = 2, lty = 1)
legend("topright", legend = c("K", "K+"), col = c("grey", "red3"),
       lwd = 2)

The number of observations Nl assigned to subcomponents $l$ of component $k$ can also be inspected, e.g., for the third component.

k <- 3   # component
matplot(Nl[seq(100, M, 10), ((k-1)*L+1):(k*L)],
        type = "l", ylab = "Nl")

Identification of the mixture model

Step 1: Estimate $K_+$ and $K$

We determine the posterior distribution of the number of filled components $K_+$, approximated using the telescoping sampler. We visualize the distribution of $K_+$ using a bar plot.

p_Kplus <- tabulate(Kplus, nbins = max(Kplus))/M 
barplot(p_Kplus/sum(p_Kplus), xlab = expression(K["+"]), names = 1:length(p_Kplus),
        ylab = expression("p(" ~ K["+"] ~ "|" ~ bold(y) ~ ")"))

The distribution of $K_+$ is also characterized using the 1st and 3rd quartile as well as the median.

quantile(Kplus, probs = c(0.25, 0.5, 0.75))

We obtain a point estimate for $K_+$ by taking the mode and determine the number of MCMC draws where exactly $K_+$ components were filled.

Kplus_hat <- which.max(p_Kplus)
Kplus_hat
M0 <- sum(Kplus == Kplus_hat)
M0          

We also determine the posterior distribution of the number of components $K$ directly drawn using the telescoping sampler.

p_K <- tabulate(K, nbins = max(K))/M
quantile(K, probs = c(0.25, 0.5, 0.75))

barplot(p_K/sum(p_K), names = 1:length(p_K), xlab = "K", 
        ylab = expression("p(" ~ K ~ "|" ~ bold(y) ~ ")"))
which.max(tabulate(K, nbins = max(K)))   

Step 2: Extracting the draws with exactly $\hat{K}_+$ non-empty components

First we select those draws where the number of filled groups was exactly $\hat{K}_+$:

index <- Kplus == Kplus_hat
Nk[is.na(Nk)] <- 0
Nk_Kplus <- (Nk[Kplus == Kplus_hat, ] > 0)

In the following we extract the cluster occurrence probabilities, the data cluster sizes and the cluster assignments for the draws where exactly $\hat{K}_+$ components were filled.

index <- Kplus == Kplus_hat
Nk[is.na(Nk)] <- 0
Nk_Kplus <- (Nk[Kplus == Kplus_hat, ] > 0)

Pi_k_inter <- Pi_k[index,,] 
Pi_k_Kplus <- array(0, dim = c(M0, Kplus_hat, sum(cat)))  
for (j in 1:sum(cat)) {
  Pi_k_Kplus[, , j] <- Pi_k_inter[, , j][Nk_Kplus]
}

Mu_inter <- Mu[index, , ]       
Mu_Kplus <- array(0, dim = c(M0, Kplus_hat, sum(cat)))  
for (j in 1:sum(cat)) {
  Mu_Kplus[, , j] <- Mu_inter[, , j][Nk_Kplus]
}

Phi_inter <- Phi[index, , ]
Phi_Kplus <- array(0, dim = c(M0, Kplus_hat, r))
for (j in 1:r) {
  Phi_Kplus[, , j] <- Phi_inter[, , j][Nk_Kplus]
}

Eta_inter <- Eta[index, ]
Eta_Kplus <- matrix(Eta_inter[Nk_Kplus], ncol = Kplus_hat)

v <- which(index)
S_Kplus <- matrix(0, M0, N)
for (i in seq_along(v)) {
  m <- v[i]
  perm_S <- rep(0, Kmax)
  perm_S[Nk[m, ] != 0] <- 1:Kplus_hat
  S_Kplus[i, ] <- perm_S[S[m, ]]
}

Step 3: Clustering and relabeling of the MCMC draws

For model identification, we cluster the draws of the cluster location probabilities Pi_k_Kplus of exactly $\hat{K}_+$ filled components in the point process representation using $k$-means clustering. We use the obtained unique labeling of the draws to reorder the sampled parameters Mu_Kplus, Phi_Kplus, Eta_Kplus and the allocations S_Kplus.

Func_init <- nonnormpost_mode[[Kplus_hat]]$pi_k
identified_Kplus <- identifyLCAMixture(
    Pi_k_Kplus, Mu_Kplus, Phi_Kplus, Eta_Kplus, S_Kplus, Func_init)

We inspect the non-permutation rate to assess how well separated the data clusters are and thus how easily one can obtain a suitable relabeling of the draws. Low values of the non-permutation rate, i.e., close to zero, indicate that the solution can be easily identified pointing to a good clustering solution being obtained.

identified_Kplus$non_perm_rate

Step 4: Estimating data cluster specific parameters and determining the final partition

The relabeled draws are also returned which can be used to determine posterior mean values for data cluster specific parameters.

colMeans(identified_Kplus$Eta)

The final partition is obtained by assigning each observation to the group where it was assigned most frequently during sampling.

z_sp <- apply(identified_Kplus$S, 2,
              function(x) which.max(tabulate(x, Kplus_hat)))
table(z_sp)
library("mclust")
classError(z_sp, z)$errorRate
adjustedRandIndex(z, z_sp)

Step 4: Visualizing the posterior component occurrence and precision distributions mu_k and phi_j.

k <- 1  # component
boxplot(identified_Kplus$Mu[, k, seq(2, 2*r, 2)],
        ylab = "mu_j")
boxplot(identified_Kplus$Phi[, k, ], ylab = "phi_j",
        ylim = quantile(identified_Kplus$Phi[, k, ], c(0, 0.95)))

References

telescope documentation built on April 4, 2025, 2:40 a.m.