-
Notifications
You must be signed in to change notification settings - Fork 0
/
getMercer.R
287 lines (251 loc) · 14.6 KB
/
getMercer.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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
# load direct estimates
# load("resultsDirectNaive.RData")
# source Mercer et al. code
source("mercer.R")
library(INLA)
# load a different 1 of these depending on whether a cluster effect should be included
# in the simulation of the data or not (tausq is the cluster effect variance)
getMercer = function(tausq=.1^2, test=FALSE, margVar=0.15^2, gamma=-1, strictPrior=FALSE, effRange=150) {
rangeText = ifelse(effRange == 150, "", "Range50")
if(!test)
load(paste0("resultsDirectNaiveBeta-1.75margVar", round(margVar, 4), "tausq", round(tausq, 4), "gamma", round(gamma, 4),
"HHoldVar0urbanOverSamplefrac0", rangeText, ".RData"))
else
load(paste0("resultsDirectNaiveBeta-1.75margVar", round(margVar, 4), "tausq", round(tausq, 4), "gamma", round(gamma, 4),
"HHoldVar0urbanOverSamplefrac0Test", rangeText, ".RData"))
# number of simulation scenarios
n = 100
mercerSRS = merceroverSamp = list()
mercerSRSParEst = matrix(nrow=5, ncol=n)
mercerSRSParSD = matrix(nrow=5, ncol=n)
mercerSRSPar10 = matrix(nrow=5, ncol=n)
mercerSRSPar50 = matrix(nrow=5, ncol=n)
mercerSRSPar90 = matrix(nrow=5, ncol=n)
merceroverSampParEst = matrix(nrow=5, ncol=n)
merceroverSampParSD = matrix(nrow=5, ncol=n)
merceroverSampPar10 = matrix(nrow=5, ncol=n)
merceroverSampPar50 = matrix(nrow=5, ncol=n)
merceroverSampPar90 = matrix(nrow=5, ncol=n)
parNames = c("Intercept", "BYM2 Phi", "BYM2 Tot. Var", "BYM2 Spatial Var", "BYM2 iid Var")
rownames(mercerSRSParEst) = parNames
rownames(mercerSRSParSD) = parNames
rownames(mercerSRSPar10) = parNames
rownames(mercerSRSPar50) = parNames
rownames(mercerSRSPar90) = parNames
rownames(merceroverSampParEst) = parNames
rownames(merceroverSampParSD) = parNames
rownames(merceroverSampPar10) = parNames
rownames(merceroverSampPar50) = parNames
rownames(merceroverSampPar90) = parNames
for(i in 1:n){
print(i)
# get predictive summary statistics
tmpoverSamp = mercer_u1m2(directEstoverSamp[[i]]$logit.est, directEstoverSamp[[i]]$var.est,
graph.path = "Kenyaadm1.graph")
resoverSamp= data.frame(admin1=directEstoverSamp[[i]]$admin1,
u1m.mercer=expit(tmpoverSamp$summary.linear.predictor$mean),
lower.mercer=tmpoverSamp$summary.linear.predictor$"0.1quant",
upper.mercer=tmpoverSamp$summary.linear.predictor$"0.9quant",
logit.est.mercer=tmpoverSamp$summary.linear.predictor$mean,
var.est.mercer=(tmpoverSamp$summary.linear.predictor$sd)^2)
merceroverSamp[[i]] = resoverSamp
tmpSRS = mercer_u1m2(directEstSRS[[i]]$logit.est, directEstSRS[[i]]$var.est,
graph.path = "Kenyaadm1.graph", strictPrior=strictPrior)
resSRS = data.frame(admin1=directEstSRS[[i]]$admin1,
u1m.mercer=expit(tmpSRS$summary.linear.predictor$mean),
lower.mercer=tmpSRS$summary.linear.predictor$"0.1quant",
upper.mercer=tmpSRS$summary.linear.predictor$"0.9quant",
logit.est.mercer=tmpSRS$summary.linear.predictor$mean,
var.est.mercer=(tmpSRS$summary.linear.predictor$sd)^2)
mercerSRS[[i]] = resSRS
## include parameter estimates in the tables (SRS)
# intercept
mercerSRSParEst[1, i] = tmpSRS$summary.fixed[,1]
mercerSRSParSD[1, i] = tmpSRS$summary.fixed[,2]
mercerSRSPar10[1, i] = tmpSRS$summary.fixed[,3]
mercerSRSPar50[1, i] = tmpSRS$summary.fixed[,4]
mercerSRSPar90[1, i] = tmpSRS$summary.fixed[,5]
# BYM2 hyperparameter phi
mercerSRSParEst[2, i] = tmpSRS$summary.hyperpar[2,1]
mercerSRSParSD[2, i] = tmpSRS$summary.hyperpar[2,2]
mercerSRSPar10[2, i] = tmpSRS$summary.hyperpar[2,3]
mercerSRSPar50[2, i] = tmpSRS$summary.hyperpar[2,4]
mercerSRSPar90[2, i] = tmpSRS$summary.hyperpar[2,5]
## transformed hyperparameters
# sample the hyperparameters, using the marginals to improve the sampling
out = inla.hyperpar.sample(1000, tmpSRS, improve.marginals=TRUE)
transformFunction = function(x) {c(1/x[1], 1/x[1]*x[2], 1/x[1]*(1-x[2]))}
transformedOut = apply(out, 1, transformFunction)
# now calculate the summary statistics of the transformed BYM2 hyperparameters
mercerSRSParEst[3:5, i] = rowMeans(transformedOut[1:3,])
mercerSRSParSD[3:5, i] = apply(transformedOut[1:3,], 1, sd)
mercerSRSPar10[3:5, i] = apply(transformedOut[1:3,], 1, quantile, probs=.1)
mercerSRSPar50[3:5, i] = apply(transformedOut[1:3,], 1, quantile, probs=.5)
mercerSRSPar90[3:5, i] = apply(transformedOut[1:3,], 1, quantile, probs=.9)
## include parameter estimates in the tables (urban oversampled)
# intercept
merceroverSampParEst[1, i] = tmpoverSamp$summary.fixed[,1]
merceroverSampParSD[1, i] = tmpoverSamp$summary.fixed[,2]
merceroverSampPar10[1, i] = tmpoverSamp$summary.fixed[,3]
merceroverSampPar50[1, i] = tmpoverSamp$summary.fixed[,4]
merceroverSampPar90[1, i] = tmpoverSamp$summary.fixed[,5]
# BYM2 hyperparameter phi
merceroverSampParEst[2, i] = tmpoverSamp$summary.hyperpar[2,1]
merceroverSampParSD[2, i] = tmpoverSamp$summary.hyperpar[2,2]
merceroverSampPar10[2, i] = tmpoverSamp$summary.hyperpar[2,3]
merceroverSampPar50[2, i] = tmpoverSamp$summary.hyperpar[2,4]
merceroverSampPar90[2, i] = tmpoverSamp$summary.hyperpar[2,5]
## transformed hyperparameters
# sample the hyperparameters, using the marginals to improve the sampling
out = inla.hyperpar.sample(1000, tmpoverSamp, improve.marginals=TRUE)
transformFunction = function(x) {c(1/x[1], 1/x[1]*x[2], 1/x[1]*(1-x[2]))}
transformedOut = apply(out, 1, transformFunction)
# now calculate the summary statistics of the transformed BYM2 hyperparameters
merceroverSampParEst[3:5, i] = rowMeans(transformedOut[1:3,])
merceroverSampParSD[3:5, i] = apply(transformedOut[1:3,], 1, sd)
merceroverSampPar10[3:5, i] = apply(transformedOut[1:3,], 1, quantile, probs=.1)
merceroverSampPar50[3:5, i] = apply(transformedOut[1:3,], 1, quantile, probs=.5)
merceroverSampPar90[3:5, i] = apply(transformedOut[1:3,], 1, quantile, probs=.9)
}
# generate summary statistics about the parameters
mercerSRSPar = data.frame(list(Est=rowMeans(mercerSRSParEst),
SD=rowMeans(mercerSRSParSD),
Q10=rowMeans(mercerSRSPar10),
Q50=rowMeans(mercerSRSPar50),
Q90=rowMeans(mercerSRSPar90)))
merceroverSampPar = data.frame(list(Est=rowMeans(merceroverSampParEst),
SD=rowMeans(merceroverSampParSD),
Q10=rowMeans(merceroverSampPar10),
Q50=rowMeans(merceroverSampPar50),
Q90=rowMeans(merceroverSampPar90)))
strictPriorText = ifelse(strictPrior, "strictPrior", "")
if(!test)
save(merceroverSamp, mercerSRS, mercerSRSPar, merceroverSampPar, file=paste0("resultsMercerBeta-1.75margVar", round(margVar, 4), "tausq", round(tausq, 4), "gamma", round(gamma, 4),
"HHoldVar0urbanOverSamplefrac0", strictPriorText, rangeText, ".RData"))
else
save(merceroverSamp, mercerSRS, mercerSRSPar, merceroverSampPar, file=paste0("resultsMercerBeta-1.75margVar", round(margVar, 4), "tausq", round(tausq, 4), "gamma", round(gamma, 4),
"HHoldVar0urbanOverSamplefrac0", strictPriorText, "Test", rangeText, ".RData"))
invisible(NULL)
}
# leave one county of data out at a time in order to validate the mercer model versus county level direct estimates
validateMercer = function(dat=ed, directLogitEsts, directLogitVars, directVars,
counties=sort(unique(poppc$County))) {
# fit the full model once, calculating certain validation scores
print("Fitting full model")
modelFit = mercer_u1m2(directLogitEsts, directLogitVars,
graph.path = "Kenyaadm1.graph", doValidation = TRUE)
cpo = modelFit$cpo$cpo
cpoFailure = modelFit$cpo$failure
dic = modelFit$dic$dic
waic = modelFit$waic$waic
countyPredsInSample = data.frame(admin1=counties,
est.mercer=expit(modelFit$summary.linear.predictor$mean),
lower.mercer=modelFit$summary.linear.predictor$"0.1quant",
upper.mercer=modelFit$summary.linear.predictor$"0.9quant",
logit.est.mercer=modelFit$summary.linear.predictor$mean,
var.est.mercer=(modelFit$summary.linear.predictor$sd)^2)
# now do leave one county out across validation
for(i in 1:length(counties)) {
if(i %% 10 == 1)
print(paste0("Fitting model with data from county ", i, "/", length(counties), " left out"))
thisCountyName = counties[i]
# fit model, get all predictions for each areal level and each posterior sample
fit = mercer_u1m2(directLogitEsts, directLogitVars,
graph.path = "Kenyaadm1.graph",
previousResult=modelFit, predCountyI=i)
# get predictive distribution for the left out county
res = data.frame(admin1=counties,
est.mercer=expit(fit$summary.linear.predictor$mean),
lower.mercer=fit$summary.linear.predictor$"0.1quant",
upper.mercer=fit$summary.linear.predictor$"0.9quant",
logit.est.mercer=fit$summary.linear.predictor$mean,
var.est.mercer=(fit$summary.linear.predictor$sd)^2)
thisCountyPreds = res[i,]
if(i == 1) {
countyPreds = thisCountyPreds
} else {
countyPreds = rbind(countyPreds, thisCountyPreds)
}
}
# don't use weights here since were calculating scoring rules for clusters rather than counties
# # calculate weights further predictions based on inverse direct estimate variances, calculate validation scores
# # don't use binomial variation here since our data is not binomial
# weights = 1 / directVars
# weights = weights / sum(weights)
# calculate scoring rules for leaving out county
countyI = match(dat$admin1, counties)
theseScores = getValidationScores(dat$y / dat$n,
countyPreds$logit.est.mercer[countyI], countyPreds$var.est.mercer[countyI],
usePearson=FALSE, n=dat$n,
urbanVec=dat$urban, filterType="leftOutCounty")
# calculate in sample scoring rules
theseScoresInSample = getValidationScores(dat$y / dat$n,
countyPredsInSample$logit.est.mercer[countyI], countyPredsInSample$var.est.mercer[countyI],
usePearson=FALSE, n=dat$n,
urbanVec=dat$urban, filterType="inSample")
theseScoresInSample$scores = cbind(WAIC=NA, DIC=NA, theseScoresInSample$scores)
theseScoresInSample$allResults = cbind(WAIC=waic, DIC=dic, theseScoresInSample$allResults)
# filter out unwanted results so that we only have MSE
filterI = grepl("CPO", names(theseScoresInSample$scores)) | grepl("CRPS", names(theseScoresInSample$scores)) | grepl("logScore", names(theseScoresInSample$scores))
theseScoresInSample$scores[filterI] = NA
filterI = grepl("CPO", names(theseScores$scores)) | grepl("CRPS", names(theseScores$scores)) | grepl("logScore", names(theseScores$scores))
theseScores$scores[filterI] = NA
# this model is not fit to data at the cluster level, so we cannot compute CPO at the cluster level
theseScoresleaveOutCluster = NULL
# compile scores c("MSE", "CPO", "CRPS", "logScore")
mercerResultsInSample = theseScoresInSample
mercerResultsLeaveOutCounty = theseScores
mercerResults = list(countyPreds=countyPreds, countyPredsInSample=countyPredsInSample, modelFitFull=modelFit,
mercerResultsInSample=mercerResultsInSample,
mercerResultsLeaveOutCounty=mercerResultsLeaveOutCounty,
mercerResultsLeaveOutCluster=NULL)
mercerResults
}
##### the below code appeared to be a copy of the above code, so I commented it out:
# # load a different 1 of these depending on whether a cluster effect should be included
# # in the simulation of the data or not (tausq is the cluster effect variance)
# getMercer = function(tausq=.1^2, test=FALSE) {
# # load("resultsDirectNaiveTausq0.RData")
# if(!test)
# load(paste0("resultsDirectNaiveTausq", round(tausq, 4), ".RData"))
# else
# load(paste0("resultsDirectNaiveTausq", round(tausq, 4), "Test.RData"))
#
# # number of simulation scenarios
# n = 100
# mercerSRS = merceroverSamp = list()
#
# for(i in 1:100){
# print(i)
# tmpoverSamp = mercer_u1m(directEstoverSamp[[i]]$logit.est, directEstoverSamp[[i]]$var.est,
# graph.path = "Kenyaadm1.graph")
# resoverSamp= data.frame(admin1=directEstoverSamp[[i]]$admin1,
# u1m.mercer=expit(tmpoverSamp$summary.linear.predictor$mean),
# lower.mercer=tmpoverSamp$summary.linear.predictor$"0.1quant",
# upper.mercer=tmpoverSamp$summary.linear.predictor$"0.9quant",
# logit.est.mercer=tmpoverSamp$summary.linear.predictor$mean,
# var.est.mercer=(tmpoverSamp$summary.linear.predictor$sd)^2)
# merceroverSamp[[i]] = resoverSamp
#
#
# tmpSRS = mercer_u1m(directEstSRS[[i]]$logit.est, directEstSRS[[i]]$var.est,
# graph.path = "Kenyaadm1.graph")
#
# resSRS = data.frame(admin1=directEstSRS[[i]]$admin1,
# u1m.mercer=expit(tmpSRS$summary.linear.predictor$mean),
# lower.mercer=tmpSRS$summary.linear.predictor$"0.1quant",
# upper.mercer=tmpSRS$summary.linear.predictor$"0.9quant",
# logit.est.mercer=tmpSRS$summary.linear.predictor$mean,
# var.est.mercer=(tmpSRS$summary.linear.predictor$sd)^2)
# mercerSRS[[i]] = resSRS
# }
#
# # save(merceroverSamp, mercerSRS, file="resultsMercer.RData")
# # save(merceroverSamp, mercerSRS, file="resultsMercerTausq0.RData")
# if(!test)
# save(merceroverSamp, mercerSRS, file=paste0("resultsMercerTausq", round(tausq, 4), ".RData"))
# else
# save(merceroverSamp, mercerSRS, file=paste0("resultsMercerTausq", round(tausq, 4), "test.RData"))
#
# invisible(NULL)
# }