-
Notifications
You must be signed in to change notification settings - Fork 0
/
multiposteriorchineseAFT.R
265 lines (197 loc) · 11.4 KB
/
multiposteriorchineseAFT.R
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
multiposteriorchineseAFT = function(c,Y1,Y2,D1,D2,That, K, r, si,sig2.dat,gmmx1, gmmx2, regy1, regy2) {
numclust <- table(factor(c, levels = 1:K))
activeclass<- which(numclust!=0)
gmmx1 <- gmmx1
gmmx2 <- gmmx2
regy1 <- regy1
regy2 <- regy2
ctemp <- c
Ytemp1 <- Y1
Ytemp2 <- Y2
Ytemp1.scaled <- matrix(NA, nrow = N, ncol = D1)
Ytemp2.scaled <- matrix(NA, nrow = N, ncol = D2)
for ( i in 1:length(activeclass)) {
clust <- which(c == activeclass[i])
if (length(clust)==1){
Ytemp1.scaled[clust,1:D1] <- matrix(0, nrow =1, ncol =D1)
Ytemp2.scaled[clust,1:D2] <- matrix(0, nrow =1, ncol =D2)
} else {
Ytemp1.scaled[clust,1:D1] <- scale(Ytemp1[clust,1:D1], center = TRUE, scale = TRUE)
Ytemp2.scaled[clust,1:D2] <- scale(Ytemp2[clust,1:D2], center = TRUE, scale = TRUE)
}
}
## This can't be parallelized !!!!!
for(l in 1:N) {
temp <- ctemp[l]
cminus <- ctemp
cminus[l] <- NA
## The table function helps converting the data point specific indicator variables to class specific indicator variables
g <- table(factor(cminus, levels = 1:K))
active <- which(g!=0)
### Length of the number of active clusters
kminus <- length(active)
## Two Auxilary Variables
## The name of the auxilary variables are taken to be one and two more than the maximum value in the already active cluster set
active <- append(active, max(active)+1)
active <- append(active, max(active)+1)
## If the observation was singelton (i.e no other point was associated with it then we assign to kminus +1 parameter)
if(length(which(cminus==temp))==0)
{
## The kminus+1 parameter gets the value of the temporary variable
ctemp[l] <- active[kminus+1]
gmmx1$mu[active[kminus+1],1:D1] <- gmmx1$mu[temp,1:D1]
gmmx1$S[active[kminus]+1,1:D1,1:D1] <- gmmx1$S[temp,1:D1,1:D1]
regy1$beta0[active[kminus+1]] <- regy1$beta0[temp]
regy1$betahat[active[kminus+1], 1:D1] <- regy1$betahat[temp, 1:D1]
regy1$sigma2[active[kminus+1]] <- regy1$sigma2[temp]
regy1$lambda2[active[kminus+1]] <- regy1$lambda2[temp]
regy1$tau2[active[kminus+1], 1:D1] <- regy1$tau2[temp, 1:D1]
gmmx2$mu[active[kminus+1],1:D2] <- gmmx2$mu[temp,1:D2]
gmmx2$S2[active[kminus]+1,1:D2,1:D2] <- gmmx2$S2[temp,1:D2,1:D2]
regy2$beta0[active[kminus+1]] <- regy2$beta0[temp]
regy2$betahat[active[kminus+1], 1:D2] <- regy2$betahat[temp, 1:D2]
regy2$sigma2[active[kminus+1]] <- regy2$sigma2[temp]
regy2$lambda2[active[kminus+1]] <- regy2$lambda2[temp]
regy2$tau2[active[kminus+1], 1:D2] <- regy2$tau2[temp, 1:D2]
## Also the second auxilary variable should be drawn from the prior distribution
source('priordraw.R')
priorone1 <- NA
priorone2 <- NA
##### priorone1 <- priordraw(beta, gmmx1$W, gmmx1$epsilon, ro, r, si,N,D1, sig2.dat)
repeat
{
priorone1 <- priordraw(beta, gmmx1$W, gmmx1$epsilon, gmmx1$ro, r, si,N,D1, sig2.dat)
res <- try(chol(priorone1$Sigma), silent = TRUE)
if (class(res) != "try-error"){
break
}
}
gmmx1$mu[active[kminus+2],1:D1] <- priorone1$mu
gmmx1$S[active[kminus+2],1:D1,1:D1] <- priorone1$Sigma
regy1$beta0[active[kminus+2]] <- priorone1$beta0
regy1$sigma2[active[kminus+2]] <- priorone1$sigma2
regy1$betahat[active[kminus+2],1:D1] <- priorone1$betahat
regy1$lambda2[active[kminus+2]] <- priorone1$lambda2
regy1$tau2[active[kminus+2], 1:D1] <- priorone1$tau2
# priorone2 <- priordraw(beta, gmmx2$W, gmmx2$epsilon, ro, r, si,N,D2, sig2.dat)
repeat
{
priorone2 <- priordraw(beta, gmmx2$W, gmmx2$epsilon, gmmx2$ro, r, si,N,D2, sig2.dat)
res <- try(chol(priorone2$Sigma), silent = TRUE)
if (class(res) != "try-error"){
break
}
}
gmmx2$mu[active[kminus+2],1:D2] <- priorone2$mu
gmmx2$S[active[kminus+2],1:D2,1:D2] <- priorone2$Sigma
regy2$beta0[active[kminus+2]] <- priorone2$beta0
regy2$sigma2[active[kminus+2]] <- priorone2$sigma2
regy2$betahat[active[kminus+2],1:D2] <- priorone2$betahat
regy2$lambda2[active[kminus+2]] <- priorone2$lambda2
regy2$tau2[active[kminus+2], 1:D2] <- priorone2$tau2
## The number of points in the cluster
} else {
priorone1 <- NA
priorone2 <- NA
### Draw the values of two auxilary parameters from Prior Distribution
source('priordraw.R')
#priorone1 <- priordraw(beta, gmmx1$W, gmmx1$epsilon, ro, r, si,N,D1, sig2.dat)
repeat {
priorone1 <- priordraw(beta, gmmx1$W, gmmx1$epsilon, gmmx1$ro, r, si,N,D1, sig2.dat)
res <- try(chol(priorone1$Sigma), silent = TRUE)
if (class(res) != "try-error"){
break
}
}
gmmx1$mu[active[kminus+1],1:D1] <- priorone1$mu
gmmx1$S[active[kminus+1],1:D1,1:D1] <- priorone1$Sigma
regy1$beta0[active[kminus+1]] <- priorone1$beta0
regy1$sigma2[active[kminus+1]] <- priorone1$sigma2
regy1$betahat[active[kminus+1],1:D1] <- priorone1$betahat
regy1$lambda2[active[kminus+1]] <- priorone1$lambda2
regy1$tau2[active[kminus+1], 1:D1] <- priorone1$tau2
# priorone2 <- priordraw(beta, gmmx2$W, gmmx2$epsilon, ro, r, si,N,D2, sig2.dat)
repeat {
priorone2 <- priordraw(beta, gmmx2$W, gmmx2$epsilon, gmmx2$ro, r, si,N, D2, sig2.dat)
res <- try(chol(priorone2$Sigma), silent = TRUE)
if (class(res) != "try-error"){
break
}
}
gmmx2$mu[active[kminus+1],1:D2] <- priorone2$mu
gmmx2$S[active[kminus+1],1:D2,1:D2] <- priorone2$Sigma
regy2$beta0[active[kminus+1]] <- priorone2$beta0
regy2$sigma2[active[kminus+1]] <- priorone2$sigma2
regy2$betahat[active[kminus+1],1:D2] <- priorone2$betahat
regy2$lambda2[active[kminus+1]] <- priorone2$lambda2
regy2$tau2[active[kminus+1], 1:D2] <- priorone2$tau2
source('priordraw.R')
#priorone1 <- priordraw(beta, gmmx1$W, gmmx1$epsilon, ro, r, si,N,D1, sig2.dat)
repeat {
priorone1 <- priordraw(beta, gmmx1$W, gmmx1$epsilon, gmmx1$ro, r, si,N,D1, sig2.dat)
res <- try(chol(priorone1$Sigma),silent = TRUE)
if (class(res) != "try-error"){
break
}
}
gmmx1$mu[active[kminus+2],1:D1] <- priorone1$mu
gmmx1$S[active[kminus+2],1:D1,1:D1] <- priorone1$Sigma
regy1$beta0[active[kminus+2]] <- priorone1$beta0
regy1$sigma2[active[kminus+2]] <- priorone1$sigma2
regy1$betahat[active[kminus+2],1:D1] <- priorone1$betahat
regy1$lambda2[active[kminus+2]] <- priorone1$lambda2
regy1$tau2[active[kminus+2], 1:D1] <- priorone1$tau2
##priorone2 <- priordraw(beta, gmmx2$W, gmmx2$epsilon, ro, r, si,N,D2, sig2.dat)
repeat {
priorone2 <- priordraw(beta, gmmx2$W, gmmx2$epsilon, gmmx2$ro, r, si,N,D2, sig2.dat)
res <- try(chol(priorone2$Sigma), silent = TRUE)
if (class(res) != "try-error"){
break
}
}
gmmx2$mu[active[kminus+2],1:D2] <- priorone2$mu
gmmx2$S[active[kminus+2],1:D2,1:D2] <- priorone2$Sigma
regy2$beta0[active[kminus+2]] <- priorone2$beta0
regy2$sigma2[active[kminus+2]] <- priorone2$sigma2
regy2$betahat[active[kminus+2],1:D2] <- priorone2$betahat
regy2$lambda2[active[kminus+2]] <- priorone2$lambda2
regy2$tau2[active[kminus+2], 1:D2] <- priorone2$tau2
}
#######################################################
posterior <- matrix(NA, nrow = length(active), ncol = 1)
## Calculating the probabalities for drawing the value of c_i from the active classes
for (j in 1:kminus) {
posterior[j] <- log(g[active[j]] /(N-1+alpha)) + dMVN(x = as.vector(t(Ytemp1[l,])), mean = gmmx1$mu[active[j],1:D1], Q = gmmx1$S[active[j],1:D1,1:D1], log = TRUE) + dMVN(x = as.vector(t(Ytemp2[l,])), mean = gmmx2$mu[active[j],1:D2], Q = gmmx2$S[active[j],1:D2,1:D2], log =TRUE) + dnorm(x = That[l], mean = regy1$beta0[active[j]] + regy1$betahat[active[j],] %*% as.vector(t(Ytemp1.scaled[l,])), sd = sqrt(regy1$sigma2[active[j]]),log = TRUE ) + dnorm(x = That[l], mean = regy2$beta0[active[j]] + regy2$betahat[active[j],] %*% as.vector(t(Ytemp2.scaled[l,])), sd = sqrt(regy2$sigma2[active[j]]), log = TRUE )
}
posterior[kminus+1] <- log((0.5 * alpha) /(N-1+alpha)) + dMVN(x = as.vector(t(Ytemp1[l,])), mean = gmmx1$mu[active[kminus+1],1:D1], Q = gmmx1$S[active[kminus+1],1:D1,1:D1], log = TRUE) + dMVN(x = as.vector(t(Ytemp2[l,])), mean = gmmx2$mu[active[kminus+1],1:D2], Q = gmmx2$S[active[kminus+1],1:D2,1:D2], log = TRUE) + dnorm(x = That[l], mean = regy1$beta0[active[kminus+1]] + regy1$betahat[active[kminus+1],] %*% as.vector(t(Ytemp1.scaled[l,])), sd = sqrt(regy1$sigma2[active[kminus+1]]), log =TRUE ) + dnorm(x = That[l], mean = regy2$beta0[active[kminus+1]] + regy2$betahat[active[kminus+1],] %*% as.vector(t(Ytemp2.scaled[l,])), sd = sqrt(regy2$sigma2[active[kminus+1]]), log =TRUE)
posterior[kminus+2] <- log((0.5 * alpha) /(N-1+alpha)) + dMVN(x = as.vector(t(Ytemp1[l,])), mean = gmmx1$mu[active[kminus+2],1:D1], Q = gmmx1$S[active[kminus+2],1:D1,1:D1], log = TRUE) + dMVN(x = as.vector(t(Ytemp2[l,])), mean = gmmx2$mu[active[kminus+2],1:D2], Q = gmmx2$S[active[kminus+2],1:D2,1:D2], log = TRUE) + dnorm(x = That[l], mean = regy1$beta0[active[kminus+2]] + regy1$betahat[active[kminus+2],] %*% as.vector(t(Ytemp1.scaled[l,])), sd = sqrt(regy1$sigma2[active[kminus+2]]), log =TRUE ) + dnorm(x = That[l], mean = regy2$beta0[active[kminus+2]] + regy2$betahat[active[kminus+2],] %*% as.vector(t(Ytemp2.scaled[l,])), sd = sqrt(regy2$sigma2[active[kminus+2]]), log =TRUE)
# posterior[kminus+1] <- (0.5 * alpha) /(N-1+alpha) * dMVN(as.vector(t(Y[l,1:D])), mean = mu[active[kminus+1],1:D], Q= S[active[kminus+1],1:D,1:D]) * dnorm(x = That[l], mean = beta0[active[kminus+1]] + betahat[active[kminus+1],1:D] %*% as.vector(t(Ytemp[l,1:D])), sd = sqrt(sigma2[active[kminus+1]]) )
#
# posterior[kminus+2] <- (0.5 * alpha) /(N-1+alpha) * dMVN(as.vector(t(Y[l,1:D])), mean = mu[active[kminus+2],1:D], Q = S[active[kminus+2],1:D,1:D]) * dnorm(x = That[l], mean = beta0[active[kminus+2]] + betahat[active[kminus+2],1:D] %*% as.vector(t(Ytemp[l,1:D])), sd = sqrt(sigma2[active[kminus+2]]) )
#
## Calculating the normalization constant for probabilities
post <- exp(posterior)
ctemp[l] <- sample(active, 1, prob= post, replace = TRUE)
}
c <- ctemp
######## Delete those observations that are not associcated with no data point #################3
g <- table(factor(c, levels = 1:K))
inactive <- which(g==0)
for ( i in 1:length(inactive)){
gmmx1$mu[inactive[i],] <- NA
gmmx1$S[inactive[i],1:D1,1:D1] <- NA
gmmx2$mu[inactive[i],] <- NA
gmmx2$S[inactive[i],1:D2,1:D2] <- NA
regy1$beta0[inactive[i]] <- NA
regy2$beta0[inactive[i]] <- NA
regy2$sigma2[inactive[i]] <- NA
regy1$sigma2[inactive[i]] <- NA
regy1$betahat[inactive[i],] <- NA
regy2$betahat[inactive[i],] <- NA
regy1$lambda2[inactive[i]] <- NA
regy2$lambda2[inactive[i]] <- NA
regy1$tau2[inactive[i],] <- NA
regy2$tau2[inactive[i],] <- NA
}
list('c'=c,'gmmx1'=gmmx1,'gmmx2'= gmmx2,'regy1'= regy1,'regy2'= regy2)
}