nplcm {baker}

Fit nested partially-latent class models (highest-level wrapper function)

Description

Uses JAGS (OSX or Windows) operating system for Bayesian posterior inference (see README file for an instruction to install JAGS). If running JAGS on windows, please go to control panel to add the directory to JAGS into ENVIRONMENTAL VARIABLE.

Usage

nplcm(data_nplcm, model_options, mcmc_options)

Arguments

data_nplcm

Cases are on top of controls in the rows of diagnostic test results and the covariate matrix. This is assumed by baker to automatically write model files (.bug).

• Mobs A list of measurements of distinct qualities (Bronze-, Silver, and Gold-Standard: MBS,MSS,MGS). The elements of the list should include MBS, MSS, and MGS. If any of the component is not available, please specify it as, e.g., MGS=NULL (effectively deleting MGS from Mobs).

  • MBS a list of data frame of bronze-standard (BrS) measurements. For each data frame (referred to as a 'slice'), rows are subjects, columns are causative agents (e.g., pathogen species). We use list here to accommodate the possibility of multiple sets of BrS data. They have imperfect sensitivity/specificity (e.g. nasopharyngeal polymerase chain reaction - NPPCR).

  • MSS a list of data frame of silver-standard (SS) measurements. Rows are subjects, columns are causative agents measured in specimen (e.g. blood culture). These measurements have perfect specificity but imperfect sensitivity.

  • MGS a list of data frame of gold-standard (GS) measurements. Rows are subject, columns are measured causative agents These measurements have perfect sensitivity and specificity.

• Y Vector of disease status: 1 for case, 0 for control.

• X Covariate matrix. A subset of columns are primary covariates in cause-specific- case-fraction (CSCF) functions and hence must be available for cases, and another subset are covariates that are available in the cases and the controls. The two sets of covariates may be identical, overlapping or completely different. In general, this is not the design matrix for regression models, because for enrollment date in a study which may have non-linear effect, basis expansion is often needed for approximation.

model_options

A list of model options: likelihood and prior.

use_measurements

A vector of characters strings; can be one or more from "BrS", "SS", "GS".

likelihood

cause_list

The vector of causes (NB: specify);

k_subclass

The number of nested subclasses in each disease class (one of case classes or the control class; the same k_subclass is assumed for each class) and each slice of BrS measurements. 1 for conditional independence; larger than 1 for conditional dependence. It is only available for BrS measurements. It is a vector of length equal to the number of slices of BrS measurements;

Eti_formula

Formula for etiology regressions. You can use s_date_Eti() to specify the design matrix for R format enrollment date; it will produce natural cubic spline basis. Specify ~ 1 if no regression is intended.

FPR_formula

formula for false positive rates (FPR) regressions; see formula(). You can use s_date_FPR() to specify part of the design matrix for R format enrollment date; it will produce penalized-spline basis (based on B-splines). Specify ~ 1 if no regression is intended. (NB: If effect="fixed", dm_Rdate_FPR() will just specify a design matrix with appropriately standardized dates.)

prior

Eti_prior

Description of etiology prior (e.g., overall_uniform - all hyperparameters are 1; or ⁠0_1⁠ - all hyperparameters are 0.1);

TPR_prior

Description of priors for the measurements (e.g., informative vs non-informative). Its length should be the same as use_measurements above. Please see examples for how to specify. The package can also handle multiple slices of BrS, SS data, so separate specification of the TPR priors are needed.

mcmc_options

A list of Markov chain Monte Carlo (MCMC) options.

• debugstatus Logical - whether to pause WinBUGS after it finishes model fitting; (NB: is this obsolete? Test.)

• n.chains Number of MCMC chains;

• n.burnin Number of burn-in iterations;

• n.thin To keep every other n.thin samples after burn-in period;

• individual.pred TRUE to perform individual prediction (Icat variables in the .bug file); FALSE otherwise;

• ppd TRUE to simulate new data (XXX.new variables in the .bug file) from the posterior predictive distribution (ppd); FALSE otherwise;

• get.pEti TRUE for getting posterior samples of individual etiologic fractions; FALSE otherwise. For non-regression, or regression models with all discrete predictors, by default this is TRUE, so no need to specify this entry. It is only relevant for regression models with non-discrete covariates. Because individuals have distinct CSCFs at their specific covariate values, it's easier to just store the posterior samples of the regression coefficients and reconstruct the pies afterwards, rather than storing them through JAGS.

• result.folder Path to folder storing the results;

• bugsmodel.dir Path to .bug model files;

• jags.dir Path to where JAGS is installed; if NULL, this will be set to jags.dir="".

Value

A JAGS output result, fitted by function R2jags::jags2() from R2jags. It is an object of class nplcm and bugs. Current implemented models follow the hierarchy below:

• no regression: Fitted by at low level by nplcm_fit_NoReg

• regression: Given disease class (control or a class of cases with the same subset of causative agents):

  • local independence model for BrS measures: Fitted at lower level by

    • nplcm_fit_Reg_NoNest deals with the setting with two sets of covariates, one for CSCF regression and the other for FPR regression. The two sets of covariates may be identical, overlapping or non-overlapping. This function is called when there exists one or more than one discrete covariate among the union of the two covariate sets. The method implemented by this function directly lets FPR depend upon covariates. This is different from Wu and Chen (2021), which let the subclass weights depend upon covariates. We implemented this function for methods comparison.

    • nplcm_fit_Reg_discrete_predictor_NoNest deals with the setting with all discrete covariates for FPRs and CSCFs. The strata defined by the two sets of covariates need not be identical, e.g., as a result of distinct sets of covariates. Again, this is directly to let FPR be stratified by covariates, hence different from Wu and Chen (2020+) We implemented this function for methods comparison.

  • local dependence model for BrS measures: Fitted at lower level by nplcm_fit_Reg_Nest: This is the method introduced in Wu and Chen (2021): CSCF regression + case/control subclass weight regression. It does not provide a specialized function for the setting with all discrete covariates.

Examples

data(data_nplcm_noreg)
cause_list <- LETTERS[1:6]
J.BrS      <- 6
model_options_no_reg <- list(
  likelihood   = list(
    cause_list = cause_list,
    k_subclass = 2,
    Eti_formula = ~-1, # no covariate for the etiology regression
    FPR_formula = list(
      MBS1 =   ~-1)    # no covariate for the subclass weight regression
  ),
  use_measurements = c("BrS"), 
  # use bronze-standard data only for model estimation.
  prior= list(
    Eti_prior = overall_uniform(1,cause_list), 
    # Dirichlet(1,...,1) prior for the etiology.
    TPR_prior  = list(BrS = list(
      info  = "informative", # informative prior for TPRs
      input = "match_range", 
      # specify the informative prior for TPRs by specifying a plausible range.
      val = list(MBS1 = list(up =  list(rep(0.99,J.BrS)), 
                             # upper ranges: matched to 97.5% quantile of a Beta prior
                             low = list(rep(0.55,J.BrS))))
      # lower ranges: matched to 2.5% quantile of a Beta prior
    )
    )
  )
)     


set.seed(1)
# include stratification information in file name:
thedir    <- paste0(tempdir(),"_no_reg")

# create folders to store the model results 
dir.create(thedir, showWarnings = FALSE)
result_folder_no_reg <- file.path(thedir,paste("results",collapse="_"))
thedir <- result_folder_no_reg
dir.create(thedir, showWarnings = FALSE)

# options for MCMC chains:
mcmc_options_no_reg <- list(
  debugstatus = TRUE,
  n.chains = 1,
  n.itermcmc = as.integer(200), 
  n.burnin = as.integer(100), 
  n.thin = 1,
  individual.pred = TRUE, # <- must set to TRUE! <------- NOTE! 
  ppd = FALSE,
  result.folder = thedir,
  bugsmodel.dir = thedir
)

BrS_object_1 <- make_meas_object(patho = LETTERS[1:6], 
                                 specimen = "MBS", test = "1", 
                                 quality = "BrS", cause_list = cause_list)
clean_options <- list(BrS_objects = make_list(BrS_object_1))
# place the nplcm data and cleaning options into the results folder
dput(data_nplcm_noreg,file.path(thedir,"data_nplcm.txt")) 
dput(clean_options, file.path(thedir, "data_clean_options.txt"))

rjags::load.module("glm")

nplcm_noreg <- nplcm(data_nplcm_noreg,model_options_no_reg,mcmc_options_no_reg)