insilico: Implement InSilicoVA methods

Description Usage Arguments Details Value Author(s) References See Also Examples

View source: R/insilico_standard.r

Description

This function implements InSilicoVA model. The InSilicoVA model is fitted with MCMC implemented in Java. For more detail, see the paper on https://arxiv.org/abs/1411.3042.

Usage

 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
insilico(
  data,
  data.type = c("WHO2012", "WHO2016")[1],
  sci = NULL,
  isNumeric = FALSE,
  updateCondProb = TRUE,
  keepProbbase.level = TRUE,
  CondProb = NULL,
  CondProbNum = NULL,
  datacheck = TRUE,
  datacheck.missing = TRUE,
  warning.write = FALSE,
  directory = NULL,
  external.sep = TRUE,
  Nsim = 4000,
  thin = 10,
  burnin = 2000,
  auto.length = TRUE,
  conv.csmf = 0.02,
  jump.scale = 0.1,
  levels.prior = NULL,
  levels.strength = 1,
  trunc.min = 1e-04,
  trunc.max = 0.9999,
  subpop = NULL,
  java_option = "-Xmx1g",
  seed = 1,
  phy.code = NULL,
  phy.cat = NULL,
  phy.unknown = NULL,
  phy.external = NULL,
  phy.debias = NULL,
  exclude.impossible.cause = c("subset2", "subset", "all", "InterVA", "none")[1],
  impossible.combination = NULL,
  no.is.missing = FALSE,
  indiv.CI = NULL,
  groupcode = FALSE,
  ...
)

Arguments

data

The original data to be used. It is suggested to use similar input as InterVA4, with the first column being death IDs and 245 symptoms. The only difference in input is InsilicoVA takes three levels: “present”, “absent”, and “missing (no data)”. Similar to InterVA software, “present” symptoms takes value “Y”; “absent” symptoms take take value “NA” or “”. For missing symptoms, e.g., questions not asked or answered in the original interview, corrupted data, etc., the input should be coded by “.” to distinguish from “absent” category. The order of the columns does not matter as long as the column names are correct. It can also include more unused columns than the standard InterVA4 input. But the first column should be the death ID. For example input data format, see RandomVA1 and RandomVA2.

data.type

Type of questionnaire. “WHO2012” corresponds to the standard input of InterVA4, and “WHO2016” corresponds to the standard input of InterVA5.

sci

A data frame that contains the symptom-cause-information (aka Probbase) that InterVA uses to assign a cause of death.

isNumeric

Indicator if the input is already in numeric form. If the input is coded numerically such that 1 for “present”, 0 for “absent”, and -1 for “missing”, this indicator could be set to True to avoid conversion to standard InterVA format.

updateCondProb

Logical indicator. If FALSE, then fit InSilicoVA model without re-estimating conditional probabilities.

keepProbbase.level

Logical indicator when updateCondProb is FALSE. If TRUE, then only estimate the InterVA's conditional probability interpretation table; if FALSE, estimate the whole conditional probability matrix. Default to TRUE.

CondProb

Customized conditional probability matrix to use.It should be strict the same configuration as InterVA-4 software. That is, it should be a matrix of 245 rows of symptoms and 60 columns of causes, arranged in the same order as in InterVA-4 specification. The elements in the matrix should be the conditional probability of corresponding symptom given the corresponding cause, represented in alphabetic form indicating levels. For example input, see condprob

CondProbNum

Customized conditional probability matrix to use if specified fully by numerical values between 0 and 1. If it is specified, re-estimation of conditional probabilities will not be performed, i.e., updateCondProb will be set to FALSE.

datacheck

Logical indicator for whether to check the data satisfying InterVA rules. Default set to be TRUE. If warning.write is set to true, the inconsistent input will be logged in file warning_insilico.txt and errorlog_insilico.txt. It's strongly suggested to be set to TRUE.

datacheck.missing

Logical indicator for whether to perform data check before deleting complete missing symptoms. Default to TRUE.

warning.write

Logical indicator for whether to save the changes made to data input by datacheck. If set to TRUE, the changes will be logged in file warning_insilico.txt and errorlog_insilico.txt in current working directory.

directory

The directory to store the output from. It should be an valid existing directory or a folder to be created.

external.sep

Logical indicator for whether to separate out external causes first. Default set to be TRUE. If set to TRUE, the algorithm will estimate external causes, e.g., traffic accident, accidental fall, suicide, etc., by checking the corresponding indicator only without considering other medical symptoms. It is strongly suggested to set to be TRUE.

Nsim

Number of iterations to run. Default to be 4000.

thin

Proportion of thinning for storing parameters. For example, if thin = k, the output parameters will only be saved every k iterations. Default to be 10

burnin

Number of iterations as burn-in period. Parameters sampled in burn-in period will not be saved.

auto.length

Logical indicator of whether to automatically increase chain length if convergence not reached.

conv.csmf

Minimum CSMF value to check for convergence if auto.length is set to TRUE. For example, under the default value 0.02, all causes with mean CSMF at least 0.02 will be checked for convergence.

jump.scale

The scale of Metropolis proposal in the Normal model. Default to be 0.1.

levels.prior

Vector of prior expectation of conditional probability levels. They do not have to be scaled. The algorithm internally calibrate the scale to the working scale through levels.strength. If NULL the algorithm will use InterVA table as prior.

levels.strength

Scaling factor for the strength of prior beliefs in the conditional probability levels. Larger value constrain the posterior estimates to be closer to prior expectation. Defult value 1 scales levels.prior to a suggested scale that works empirically.

trunc.min

Minimum possible value for estimated conditional probability table. Default to be 0.0001

trunc.max

Maximum possible value for estimated conditional probability table. Default to be 0.9999

subpop

This could be the column name of the variable in data that is to be used as sub-population indicator, or a list of column names if more than one variable are to be used. Or it could be a vector of sub-population assignments of the same length of death records. It could be numerical indicators or character vectors of names.

java_option

Option to initialize java JVM. Default to “-Xmx1g”, which sets the maximum heap size to be 1GB. If R produces “java.lang.OutOfMemoryError: Java heap space” error message, consider increasing heap size using this option, or one of the following: (1) decreasing Nsim, (2) increasing thin, or (3) disabling auto.length.

seed

Seed used for initializing sampler. The algorithm will produce the same outcome with the same seed in each machine.

phy.code

A matrix of physician assigned cause distribution. The physician assigned causes need not be the same as the list of causes used in InSilicoVA and InterVA-4. The cause list used could be a higher level aggregation of the InSilicoVA causes. See phy.cat for more detail. The first column of phy.code should be death ID that could be matched to the symptom dataset, the following columns are the probabilities of each cause category used by physicians.

phy.cat

A two column matrix describing the correspondence between InSilicoVA causes and the physician assigned causes. Note each InSilicoVA cause (see causetext) could only correspond to one physician assigned cause. See SampleCategory for an example. 'Unknown' category should not be included in this matrix.

phy.unknown

The name of the physician assigned cause that correspond to unknown COD.

phy.external

The name of the physician assigned cause that correspond to external causes. This will only be used if external.sep is set to TRUE. In that case, all external causes should be grouped together, as they are assigned deterministically by the corresponding symptoms.

phy.debias

Fitted object from physician coding debias function (see physician_debias) that overwrites phy.code.

exclude.impossible.cause

option to exclude impossible causes at the individual level. The following rules are implemented: subset: Causes with 0 probability given the age group and gender of the observation, according to the InterVA conditional probabilities, are removed; subset2: In addition to the same rules as subset, also remove Prematurity for baby born during at least 37 weeks of pregnancy, remove Birth asphyxia for baby not born during at least 37 weeks of pregnancy, and remove pregnancy-related deaths for deaths without pregnancy; all: Causes with 0 probability given any symptom of the observation, according to the InterVA conditional probabilities, are removed; interVA: Causes with 0 probability given any positive indicators according to the InterVA conditional probabilities, are removed; and none: no causes are removed. subset2 is the default.

impossible.combination

matrix indicating additional impossible symptom-cause combinations in addition to the ones specified by exlcude.impossible.cause. It should be a matrix of three columns, where each row is one rule of impossible combination. In each row, the first column specify the name of the symptom (as in the input data column names, e.g., "i079o"), the second column specify the name of the cause (as in the probbase column name, i.e., "b_0101"), and the third column specifying either 0 or 1, indicating the cause is impossible when the symptom takes the specified value.

no.is.missing

logical indicator to treat all absence of symptoms as missing. Default to FALSE. If set to TRUE, InSilicoVA will perform calculations similar to InterVA-4 w.r.t treating absent symptoms. It is highly recommended to set this argument to FALSE.

indiv.CI

credible interval for individual probabilities. If set to NULL, individual COD distributions will not be calculated to accelerate model fitting time. See get.indiv for details of updating the C.I. later after fitting the model.

groupcode

logical indicator of including the group code in the output causes

...

not used

Details

For Windows user, this function will produce a popup window showing the progress. For Mac and Unix user, this function will print progress messages on the console. Special notice for users using default R GUI for mac, the output will not be printed on console while the function is running, and will only be printed out after it is completed. Thus if you use a Mac, we suggest using either RStudio for mac, or running R from terminal.

The chains could be set to run automatically longer. If set auto.length to be TRUE, the chain will assess convergence after finishing the length K chain input by user using Heidelberger and Welch's convergence diagnostic. If convergence is not reached, the chain will run another K iterations and use the first K iterations as burn-in. If the chain is still not converged after 2K iterations, it will proceed to another 2K iterations and again use the first 2K iterations as burn-in. If convergence is still not reached by the end, it will not double the length again to avoid heavy memory use. A warning will be given in that case. The extended chains will be thinned in the same way.

For more detail of model specification, see the paper on https://arxiv.org/abs/1411.3042.

Value

id

A vector of death ID. Note the order of the ID is in general different from the input file. See report for organizing the report.

data

Cleaned numerical data.

indiv.prob

Matrix of individual mean cause of death distribution. Each row corresponds to one death with the corresponding ID.

csmf

Matrix of CSMF vector at each iterations after burn-in and thinning. Each column corresponds to one cause.

conditional.probs

If the model is estimated with keepProbbase.level = TRUE, this value gives a matrix of each conditional probability at each level at each iterations. Each column corresponds to one level of probability. If keepProbbase.level = FALSE, this value gives a three-dimensional array. If updateCondProb = FALSE, the value will be set to NULL. See report for more analysis.

missing.symptoms

Vector of symptoms missing from all input data.

external

Logical indicator of whether the model is fitted with external causes separated calculated.

impossible.causes

Impossible cause-symptom pairs, if any.

indiv.CI

The posterior credible interval to compute for individual COD probability distributions. If set to NULL, only the posterior mean of the individual COD probabilities will be produced. Default to be 0.95.

indiv.prob.median

median probability of each cause of death for each individual death.

indiv.prob.lower

lower CI bound for the probability of each cause of death for each individual death.

indiv.prob.upper

upper CI bound for the probability of each cause of death for each individual death.

errors

Logs of deleted observations and reasons of deletion.

Author(s)

Zehang Li, Tyler McCormick, Sam Clark

Maintainer: Zehang Li <lizehang@uw.edu>

References

Tyler H. McCormick, Zehang R. Li, Clara Calvert, Amelia C. Crampin, Kathleen Kahn and Samuel J. Clark(2014) Probabilistic cause-of-death assignment using verbal autopsies, https://arxiv.org/abs/1411.3042
Working paper no. 147, Center for Statistics and the Social Sciences, University of Washington

See Also

plot.insilico, summary.insilico, physician_debias

Examples

  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
## Not run: 
data(RandomVA1) 
fit0<- insilico(RandomVA1, subpop = NULL,  
                Nsim = 20, burnin = 10, thin = 1 , seed = 1,
			 auto.length = FALSE)
summary(fit0)
summary(fit0, id = "d199")

##
## Scenario 1: standard input without sub-population specification
##
fit1<- insilico(RandomVA1, subpop = NULL,  
              Nsim = 1000, burnin = 500, thin = 10 , seed = 1,
		   auto.length = FALSE)
summary(fit1)
plot(fit1)

##
## Scenario 2: standard input with sub-population specification
##
data(RandomVA2)
fit2<- insilico(RandomVA2, subpop = list("sex"),  
              Nsim = 1000, burnin = 500, thin = 10 , seed = 1,
		   auto.length = FALSE)
summary(fit2)
plot(fit2, type = "compare")
plot(fit2, which.sub = "Men")

##
## Scenario 3: standard input with multiple sub-population specification
##
fit3<- insilico(RandomVA2, subpop = list("sex", "age"),  
              Nsim = 1000, burnin = 500, thin = 10 , seed = 1,
		   auto.length = FALSE)
summary(fit3)

##
## Scenario 3: standard input with multiple sub-population specification
##
fit3<- insilico(RandomVA2, subpop = list("sex", "age"),  
              Nsim = 1000, burnin = 500, thin = 10 , seed = 1,
		   auto.length = FALSE)
summary(fit3)

##
## Scenario 5 - 7 are special situations rarely needed in practice,
##   but included here for completeness. 
##   The below examples use no sub-population or physician codes, 
##   but specifying sub-population is still possible as in Scenario 2 - 4.
## 

##
## Scenario 5: skipping re-estimation of conditional probabilities
##
# Though in practice the need for this situation is very unlikely, 
# use only the default conditional probabilities without re-estimation
fit5<- insilico(RandomVA1, subpop = NULL,  
              Nsim = 1000, burnin = 500, thin = 10 , seed = 1,
              updateCondProb = FALSE, 
		   auto.length = FALSE) 
summary(fit5)

##
## Scenario 6: modify default conditional probability matrix
##
# Load the default conditional probability matrix 
data(condprob)
# The conditional probabilities are given in levels such as I, A+, A, A-, etc.
condprob[1:5, 1:5]
# To modify certain cells 
new_cond_prob <- condprob
new_cond_prob["elder", "HIV/AIDS related death"] <- "C"
# or equivalently
new_cond_prob[1, 3] <- "C"

fit6<- insilico(RandomVA1, subpop = NULL,  
              Nsim = 1000, burnin = 500, thin = 10 , seed = 1,
              CondProb = new_cond_prob, 
		   auto.length = FALSE) 
# note: compare this with fit1 above to see the change induced 
#  by changing Pr(elder | HIV) from "C+" to "C".
summary(fit6)

##
## Scenario 7: modify default numerical values in conditional probabilities directly
##
# Load the default conditional probability matrix 
data(condprobnum)
# The conditional probabilities are given in numerical values in this dataset
condprobnum[1:5, 1:5]
# To modify certain cells, into any numerical values you want 
new_cond_prob_num <- condprobnum
new_cond_prob_num["elder", "HIV/AIDS related death"] <- 0.004
# or equivalently
new_cond_prob_num[1, 3] <- 0.005

fit7<- insilico(RandomVA1, subpop = NULL,  
              Nsim = 1000, burnin = 500, thin = 10 , seed = 1,
              CondProbNum = new_cond_prob_num, 
		   auto.length = FALSE) 
# note: compare this with fit1, fit5, and fit6
summary(fit7)

##
## Scenario 8: physician coding
## see also the examples in physician_debias() function section
##
# Load sample input for physicians
data(RandomPhysician)
# The symptom section looks the same as standard input
head(RandomPhysician[, 1:5])
# At the end of file, including a few more columns of physician id and coded cause
head(RandomPhysician[, 245:250])

# load Cause Grouping (if physician-coded causes are in larger categories)
data(SampleCategory)
head(SampleCategory)

# existing doctor codes in the sample dataset
doctors <- paste0("doc", c(1:15))
causelist <- c("Communicable", "TB/AIDS", "Maternal",
               "NCD", "External", "Unknown")
phydebias <- physician_debias(RandomPhysician, 
phy.id = c("rev1", "rev2"), phy.code = c("code1", "code2"), 
phylist = doctors, causelist = causelist, 
tol = 0.0001, max.itr = 100)

fit8 <- insilico(RandomVA1, subpop = NULL,  
              Nsim = 1000, burnin = 500, thin = 10 , seed = 1,
              phy.debias = phydebias,
              phy.cat = SampleCategory, 
              phy.external = "External", phy.unknown = "Unknown",
		   auto.length = FALSE) 
summary(fit8)

# example to fit WHO2016 data
data(RandomVA5)
fit1a <- insilico(RandomVA5, data.type="WHO2016", subpop = NULL,  
              Nsim = 1000, burnin = 500, thin = 10 , seed = 1,
		   auto.length = FALSE)
summary(fit1a)
plot(fit1)

# example to change directory for error files
fit1b <- insilico(RandomVA5[1:50, ], data.type="WHO2016", 
					Nsim = 1000, burnin = 500, thin = 10 , 
					seed = 1, warning.write = T, auto.length=F)
fit1c <- insilico(RandomVA5[1:50, ], data.type="WHO2016", 
					Nsim = 1000, burnin = 500, thin = 10 , 
					seed = 1, warning.write = T, 
					directory = "fit1b_errorfolder", auto.length=F)

 # similarly for WHO 2012 version
fit1<- insilico(RandomVA1, subpop = NULL,  
              Nsim = 1000, burnin = 500, thin = 10 , seed = 1,
		         auto.length = FALSE, 
				 warning.write = T, directory = "fit1_errorfolder")


## End(Not run)

InSilicoVA documentation built on Aug. 2, 2021, 5:08 p.m.