csmf.diag {InSilicoVA}

Convergence test for fitted InSilico model

Description

Produce convergence test for CSMFs from fitted "insilico" objects.

Usage

csmf.diag(
  csmf,
  conv.csmf = 0.02,
  test = c("gelman", "heidel")[2],
  verbose = TRUE,
  autoburnin = FALSE,
  which.sub = NULL,
  ...
)

Arguments

csmf

It could be either fitted "insilico" object, a list of fitted "insilico" object from different chains of the same length but different starting values, i.e., different seed. Or it could be the matrix of CSMF obtained by insilico, or the list of matrices of CSMF. All CSMF could contain more than one subpopulations, but should be in the same format and order. And notice if the raw CSMF is used instead of the "insilico" object, external causes might need to be removed manually by user is external.sep is TRUE when fitting the model.

conv.csmf

The minimum mean CSMF to be checked. Default to be 0.02, which means any causes with mean CSMF lower than 0.02 will not be tested.

test

Type of test. Currently supporting Gelman and Rubin's test (test = "gelman") for multi-chain test, and Heidelberger and Welch's test (test = "heidel") for single-chain test.

verbose

Logical indicator to return the test detail instead of one logical outcome for Heidelberger and Welch's test. Default to be TRUE.

autoburnin

Logical indicator of whether to omit the first half of the chain as burn in. Default to be FALSE since insilico return only the iterations after burnin by default.

which.sub

the name of the sub-population to test when there are multiple in the fitted object.

...

Arguments to be passed to heidel.diag or gelman.diag

Details

The tests are performed using heidel.diag and gelman.diag functions in coda package. The function takes either one or a list of output from insilico function, or only the iteration by CSMF matrix. Usually in practice, many causes with very tiny CSMF are hard to converge based on standard tests, thus it is suggested to check convergence for only causes with mean CSMF over certain threshold by setting proper conv.csmf.

Note for Gelman and Rubin's test, all chains should have the same length. If the chains are sampled with automatically length determination, they might not be comparable by this test.

Author(s)

Zehang Li, Tyler McCormick, Sam Clark

Maintainer: Zehang Li <[e-mail]>

References

Tyler H. McCormick, Zehang R. Li, Clara Calvert, Amelia C. Crampin, Kathleen Kahn and Samuel J. Clark Probabilistic cause-of-death assignment using verbal autopsies, Journal of the American Statistical Association (2016), 111(515):1036-1049.

Gelman, Andrew, and Donald B. Rubin. Inference from iterative simulation using multiple sequences. Statistical science (1992): 457-472.

Brooks, Stephen P., and Andrew Gelman. General methods for monitoring convergence of iterative simulations. Journal of computational and graphical statistics 7.4 (1998): 434-455.

Heidelberger, Philip, and Peter D. Welch. A spectral method for confidence interval generation and run length control in simulations. Communications of the ACM 24.4 (1981): 233-245.

Heidelberger, Philip, and Peter D. Welch. Simulation run length control in the presence of an initial transient. Operations Research 31.6 (1983): 1109-1144.

Schruben, Lee W. Detecting initialization bias in simulation output. Operations Research 30.3 (1982): 569-590.

See Also

insilico, summary.insilico

Examples

# load sample data together with sub-population list
data(RandomVA2)
## Not run: 
# extract InterVA style input data
data <- RandomVA2
# extract sub-population information. 
subpop <- RandomVA2$sex

# run without sub-population
fit1a<- insilico( data, subpop = NULL, 
              Nsim = 400, burnin = 200, thin = 10 , seed = 1, 
              auto.length = FALSE)
fit1b<- insilico( data, subpop = NULL,  
              Nsim = 400, burnin = 200, thin = 10 , seed = 2, 
              auto.length = FALSE)
fit1c<- insilico( data, subpop = NULL,  
              Nsim = 400, burnin = 200, thin = 10 , seed = 3, 
              auto.length = FALSE)
# single chain check
csmf.diag(fit1a)

# multiple chains check
csmf.diag(list(fit1a, fit1b, fit1c), test = "gelman")


# with sub-populations
fit2a<- insilico( data, subpop = subpop,  
              Nsim = 400, burnin = 200, thin = 10 , seed = 1, 
              auto.length = FALSE)
fit2b<- insilico( data, subpop = subpop,  
              Nsim = 400, burnin = 200, thin = 10 , seed = 2, 
              auto.length = FALSE)
fit2c<- insilico( data, subpop = subpop,   
              Nsim = 400, burnin = 200, thin = 10 , seed = 3, 
              auto.length = FALSE)
# single chain check
csmf.diag(fit2a)

# multiple chains check
csmf.diag(list(fit2a, fit2b, fit2c), test = "gelman", which.sub = "Men")

## End(Not run)