-
Notifications
You must be signed in to change notification settings - Fork 1
/
latentFieldModel.R
290 lines (220 loc) · 12.3 KB
/
latentFieldModel.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
288
289
290
#' latentFieldModel: function to define latent field models from dbViewR object
#'
#' @param db dbViewR object with valid column names for INLA model.
#' Smoothing only makes sense by age, location, and time. Factor variables cannot be smoothed!
#' @param shp sf object with GEOID shapes (all higher levels assume iid and not local smoothing)
#' @param family non-standard family override (default = NULL).
#' @param neighborGraph non-standard neighbor graph (default = NULL)
#'
#' @return modelDefinition object for modelTrainR, list with fields
#' type = latentField
#' family : as input
#' formula : model definition for INLA
#' inputData : inla-prepared db$observed data
#' neighborGraph : as input or derived from shp during formula construction
#'
#' @import INLA
#' @import dbViewR
#'
#' @export
#' @examples
#' return h1n1pdm incidence model by time and location
#' modelDefinition <- smoothModel(db = dbViewR::selectFromDB(), shp = dbViewR::masterSpatialDB())
#'
latentFieldModel <- function(db = dbViewR::selectFromDB(), shp = dbViewR::masterSpatialDB(), family = NULL, neighborGraph = NULL){
#INLA data frame that may get augmented columns we don't need to see when we're done
inputData <- db$observedData
# identify intended family
if(is.null(family)){
if (all(inputData$n == inputData$positive)){
family = 'poisson'
} else if (any(inputData$n > inputData$positive)){
family = 'binomial'
} else if (any(inputData$n < inputData$positive)){
return('n < positive !!! invald db$observedData.')
}
}
# construct priors
hyper=list()
hyper$global <- list(prec = list( prior = "pc.prec", param = 1/10, alpha = 0.01))
hyper$local <- list(prec = list( prior = "pc.prec", param = 1/100, alpha = 0.01))
hyper$age <- list(prec = list( prior = "pc.prec", param = 1/100, alpha = 0.01))
# unlike smoothing model, we only replicate latent fields across pathogens, but treat all other factors as fixed effects
# find pathogen types
if('pathogen' %in% names(db$observedData)){
levelSet <- levels(as.factor(inputData$pathogen))
numLevels <- length(levelSet)
validLatentFieldColumns <- c('pathogen')
} else {
return('error! must provide "pathogen" column.')
}
# set family across all levels
family <- rep(family,numLevels)
# build outcome matrix and replicate list for multiple likelihoods
outcome <- matrix(NA,nrow(inputData),numLevels)
replicateIdx <- matrix(NA,nrow(inputData),1)
for( k in levelSet){
idx <- inputData$pathogen %in% k
count <- which(levelSet %in% k)
outcome[idx, count] <- inputData$positive[idx]
replicateIdx[idx]<-count
}
# initialize formula for each level
if (numLevels>1){
outcomeStr <- paste('cbind(',paste(paste('outcome',1:numLevels,sep='.'),sep='',collapse=', '),')',sep='',collapse = '')
formula <- as.formula(paste(outcomeStr,'~','pathogen - 1 + catchment',sep=' '))
} else {
formula <- outcome ~ 1 + catchment
}
# factors as fixed effects, assuming no interaction terms
validFactorNames <- c('samplingLocation','fluShot','sex','hasFever','hasCough','hasMyalgia')
factorIdx <- names(db$observedData) %in% validFactorNames
for(COLUMN in names(db$observedData)[factorIdx]){
formula <- as.formula(paste(as.character(formula)[2],'~',paste(as.character(formula)[3],COLUMN,sep='+')))
}
# latent fields
for(COLUMN in names(inputData)[!(names(inputData) %in% c('positive','n'))]){
if(COLUMN == 'timeRow'){
#INLA needs one column per random effect
inputData$timeRow_rw2 <- inputData$timeRow
inputData$timeRow_IID <- inputData$timeRow
formula <- update(formula, ~ . + f(timeRow_rw2, model='rw2', hyper=modelDefinition$hyper$global, replicate=replicateIdx) +
f(timeRow_IID, model='iid', hyper=modelDefinition$hyper$local, replicate=replicateIdx, constr = TRUE) )
validLatentFieldColumns <- c(validLatentFieldColumns,'timeRow_rw2','timeRow_IID')
}
if(COLUMN == 'ageRow'){
inputData$ageRow_rw2 <- inputData$ageRow
inputData$ageRow_IID <- inputData$ageRow
formula <- update(formula, ~ . + f(ageRow_rw2, model='rw2', hyper=modelDefinition$hyper$age, replicate=replicateIdx) +
f(ageRow_IID, model='iid', hyper=modelDefinition$hyper$local, replicate=replicateIdx, constr = TRUE) )
validLatentFieldColumns <- c(validLatentFieldColumns,'ageRow_rw2','ageRow_IID')
}
if(COLUMN %in% c('PUMA5CE')){
inputData$PUMA5CERow <- match(inputData$PUMA5CE,unique(inputData$PUMA5CE))
if('timeRow' %in% names(inputData)){
inputData$timeRow_PUMA5CE <- inputData$timeRow
formula <- update(formula, ~ . + f(PUMA5CERow, model='iid', hyper=modelDefinition$local, constr = TRUE, replicate=replicateIdx,
group = timeRow_PUMA5CE, control.group=list(model="rw2")))
validLatentFieldColumns <- c(validLatentFieldColumns,'PUMA5CERow','timeRow_PUMA5CE')
} else {
formula <- update(formula, ~ . + f(PUMA5CERow, model='iid', hyper=modelDefinition$hyper$global, replicate=replicateIdx))
validLatentFieldColumns <- c(validLatentFieldColumns,'PUMA5CERow')
}
}
if(COLUMN %in% c('CRA_NAME')){
inputData$CRA_NAMERow <- match(inputData$CRA_NAME,unique(inputData$CRA_NAME))
if('timeRow' %in% names(inputData)){
inputData$timeRow_CRA_NAME <- inputData$timeRow
formula <- update(formula, ~ . + f(CRA_NAMERow, model='iid', hyper=modelDefinition$local, constr = TRUE, replicate=replicateIdx,
group = timeRow_CRA_NAME, control.group=list(model="rw2")))
validLatentFieldColumns <- c(validLatentFieldColumns,'CRA_NAMERow','timeRow_CRA_NAME')
} else {
formula <- update(formula, ~ . + f(CRA_NAMERow, model='iid', hyper=modelDefinition$hyper$global, replicate=replicateIdx))
validLatentFieldColumns <- c(validLatentFieldColumns,'CRA_NAMERow')
}
}
if(COLUMN %in% c('NEIGHBORHOOD_DISTRICT_NAME')){
inputData$NEIGHBORHOOD_DISTRICT_NAMERow <- match(inputData$NEIGHBORHOOD_DISTRICT_NAME,unique(inputData$NEIGHBORHOOD_DISTRICT_NAME))
if('timeRow' %in% names(inputData)){
inputData$timeRow_NEIGHBORHOOD_DISTRICT_NAME <- inputData$timeRow
formula <- update(formula, ~ . + f(NEIGHBORHOOD_DISTRICT_NAMERow, model='iid', hyper=modelDefinition$local, constr = TRUE, replicate=replicateIdx,
group = timeRow_NEIGHBORHOOD_DISTRICT_NAME, control.group=list(model="rw2")))
validLatentFieldColumns <- c(validLatentFieldColumns,'NEIGHBORHOOD_DISTRICT_NAMERow','timeRow_NEIGHBORHOOD_DISTRICT_NAME')
} else {
formula <- update(formula, ~ . + f(NEIGHBORHOOD_DISTRICT_NAMERow, model='iid', hyper=modelDefinition$hyper$global, replicate=replicateIdx))
validLatentFieldColumns <- c(validLatentFieldColumns,'NEIGHBORHOOD_DISTRICT_NAMERow')
}
}
# Do we want the option of neighbor smoothing at larger scales?
if(COLUMN == 'GEOID'){
if(exists('shp')){
neighborGraph <- constructAdjacencyNetwork(shp)
inputData$GEOIDRow <- shp$rowID[match(inputData$GEOID,shp$GEOID)]
if('timeRow' %in% names(inputData)){
inputData$timeRow_GEOID <- inputData$timeRow
formula <- update(formula, ~ . + f(GEOIDRow, model='besag', graph=modelDefinition$neighborGraph, constr = TRUE, hyper=modelDefinition$hyper$local, replicate=replicateIdx,
group = timeRow_GEOID, control.group=list(model="rw2")))
validLatentFieldColumns <- c(validLatentFieldColumns,'GEOIDRow','timeRow_GEOID')
} else {
formula <- update(formula, ~ . + f(GEOIDRow, model='bym2', graph=modelDefinition$neighborGraph, constr = TRUE, hyper=modelDefinition$hyper$local, replicate=replicateIdx))
validLatentFieldColumns <- c(validLatentFieldColumns,'GEOIDRow')
}
} else {
inputData$GEOIDRow <- match(inputData$GEOID,unique(inputData$GEOID))
if('timeRow' %in% names(inputData)){
inputData$timeRow_GEOID <- inputData$timeRow
formula <- update(formula, ~ . + f(GEOIDRow, model='iid', graph=modelDefinition$neighborGraph, hyper=modelDefinition$hyper$local, replicate=replicateIdx,
group = timeRow_GEOID, control.group=list(model="rw2")))
validLatentFieldColumns <- c(validLatentFieldColumns,'GEOIDRow','timeRow_GEOID')
} else {
formula <- update(formula, ~ . + f(GEOIDRow, model='iid', graph=modelDefinition$neighborGraph, hyper=modelDefinition$hyper$local, replicate=replicateIdx))
validLatentFieldColumns <- c(validLatentFieldColumns,'GEOIDRow')
}
}
}
}
# linear combination of pathogen and latent fields
lc.data <- data.frame(inputData[,names(inputData) %in% validLatentFieldColumns], replicateIdx = replicateIdx)
lc.data <- lc.data[!duplicated(lc.data),]
# generate list of desired linear combinations # https://groups.google.com/forum/#!topic/r-inla-discussion-group/_e2C2L7Wc30
lcIdx=c()
spentColumn<-rep(FALSE,length(validLatentFieldColumns))
for(COLUMN in validLatentFieldColumns){
if(COLUMN %in% c('pathogen') ){
# need to promote pathogen levels to independent columns! https://groups.google.com/forum/#!topic/r-inla-discussion-group/IaTSakB7qy4
pathogenNames <- paste('pathogen',levelSet,sep='')
} else if (!(COLUMN == 'timeRow_PUMA5CE' )) {
groupIdx<-grepl( paste0('_',gsub('Row','',COLUMN)) ,validLatentFieldColumns) # this nasty thing will get refactored: https://github.com/seattleflu/incidence-mapper/issues/13
if(any(groupIdx & !spentColumn)){ # grouped?
lcIdx[[COLUMN]] <- inla.idx(lc.data[[COLUMN]], group = lc.data[[validLatentFieldColumns[groupIdx]]], replicate = lc.data$replicateIdx)
spentColumn[groupIdx]<-TRUE
} else if(!spentColumn[validLatentFieldColumns %in% COLUMN]) {
lcIdx[[COLUMN]] <- inla.idx(lc.data[[COLUMN]], replicate = lc.data$replicateIdx)
}
}
spentColumn[validLatentFieldColumns %in% COLUMN]<-TRUE
}
# generate list of desired linear combinations # https://groups.google.com/forum/#!topic/r-inla-discussion-group/_e2C2L7Wc30
lc.latentField <- c()
for(k in 1:nrow(lc.data)){
w<-list()
for(n in 1:length(names(lcIdx))){
w[[n]]<-rep(0,nrow(lc.data))
w[[n]][lcIdx[[n]][k]]<-1
}
A <- c(x=1, w)
names(A) <- c(pathogenNames[lc.data$replicateIdx[k]],names(lcIdx))
lc <- inla.make.lincomb(A)
names(lc)<- paste0('latentField',k)
lc.latentField<-c(lc.latentField,lc)
lc.data$latentField[k]<-names(lc)
}
df <- data.frame(outcome = outcome, inputData, replicateIdx)
modelDefinition <- list(type='latentField', family = family, formula = formula, lincomb = lc.latentField,
inputData = df, neighborGraph=neighborGraph, hyper=hyper,
latentFieldData = lc.data, # clean up formatting, but this will be useful for exporting latentField csv
queryList = db$queryList)
return(modelDefinition)
}
#' appendLatentFieldData: internal function for adding model$summary.random to db$observedData from latentFieldModel fit
#'
#' @param model inla model object
#' @param db object from dbViewer with observedData tibble and query
#' @return db with added modeledData tibble
#'
#' @import lubridate
#'
appendLatentFieldData <- function(model,db, family = 'poisson'){
modeledData <- db$observedData
if(family[1] == 'binomial'){
modeledData$fraction <- modeledData$positive/modeledData$n
}
# summary.fitted.values
nCol <- ncol(modeledData)
modeledData[,nCol+1:ncol(model$summary.fitted.values)]<-model$summary.fitted.values
names(modeledData)[nCol+1:ncol(model$summary.fitted.values)]<-paste('fitted.values',names(model$summary.fitted.values),sep='.')
rownames(modeledData)<-c()
# summary.random
# TO-DO
return(modeledData)
}