validate: Tests correctness of Bayesian Model-Fitting Software

In BayesValidate: BayesValidate Package

Description

Inputs functions to generate and analyze data. Compares output from these functions to test that the model-fitting software works correctly.

Usage

validate(generate.param, generate.param.inputs = NULL, generate.data, 
	generate.data.inputs = NULL, analyze.data, analyze.data.inputs = NULL, 
	n.rep = 20, n.batch = NULL, params.batch = NULL, print.reps = FALSE)

Arguments

generate.param	Function for generating parameters from prior distribution Should output a vector of parameters. Function should look like generate.param <- function() {...} or generate.param <- function(generate.param.inputs) {...}
generate.param.inputs	Inputs to the function generate.param
generate.data	Function for generating data given parameters. Should take as input the output from generate.param. Should output data matrix to be analyzed. Function should look like generate.data <- function(theta.true) {...} or generate.data <- function(theta.true, generate.data.inputs) {...}
generate.data.inputs	Inputs to the function generate.data (in addition to theta.true)
analyze.data	Function for generating sample from posterior distribution. Should take as input the output from generate.data and generate.param. Should output a matrix of parameters, each row is one parameter draw. Function should look like analyze.data <- function(data.rep,theta.true) {...} or analyze.data <- function(data.rep,theta.true, analyze.data.inputs) {...}. ! This is the software being tested !!
analyze.data.inputs	Inputs to the function analyze.data (in addition to data.rep and theta.true)
n.rep	Number of replications to be performed, default is 20.
n.batch	Lengths of parameter batches. A parameter batch might consist of, for example, all person-level means in a hierarchical model or any group of parameters that is sampled in a loop. Must sum to n.param (length of parameter vector) and correspond to the order of parameters as they are output from analyze.data. For example, if there are 5 total parameters with the first two in one batch and the next three in another batch, use n.batch=c(2,3)
params.batch	Names of parameter batches, used as the y axis in the output plot. Must have length equal to the number of batches. Can consist of text (e.g., params.batch=c("alpha","beta")) or an expression (e.g., params.batch=expression(alpha,beta)). Not used if n.batch not provided.
print.reps	Indcator of whether or not to print the replication number, default is FALSE

Details

Validate tests whether software developed to fit a specific Bayesian model works properly, capitalizing on properties of Bayesian posterior distributions. The validation method involves repeatedly generating parameters and data from the model to be fit and then fitting the same model to these simulated data (i.e., generating a sample from the posterior distribution). For all scalar parameters, the quantile of the "true" parameter value with respect to its posterior distribution should follow a uniform distribution if the software is written correctly. Testing that the software works amounts to testing that these quantiles are uniformly distributed. For each scalar parameter, the function gives a p-value for a test that its quantiles are uniformly distributed.

Value

p.vals	p-values for "unbatched" parameters. Null if all batches consist of only one parameter.
p.batch	p-values for the means of batched parameters. Null if n.batch not provided
adj.min.p	The smallest of p.batch (or, if p.batch=NULL, the smallest of p.vals), with the Bonferroni correction for multiple comparisons applied.

Author(s)

Samantha Cook [e-mail]

References

http://www.stat.columbia.edu/~cook/Cook_Software_Validation.pdf

See Also

quant

Examples

set.seed(314)

##functions for generating parameters mu, sigma^2 from their prior distribution
rinvchisq <- function(n,v,s){
        alpha <- v/2
        beta <- alpha*s;
        draws <- 1/rgamma(n,alpha,beta)
        return(draws)}

generate.param <- function(hyper){
	mu.0 <- hyper[1]
	kappa.0 <- hyper[2]
	nu.0 <- hyper[3]
	sigsq.0 <- hyper[4]
	sigsq <- rinvchisq(1, nu.0, sigsq.0)
	mu <- rnorm(1, mu.0, sqrt(sigsq/kappa.0))
	return(c(sigsq,mu))}

##generate normal data with mean mu, variance sigma^2, sample size n
generate.data <- function(params,n){
	y <- rnorm(n,params[2],sqrt(params[1]))
	return(y)}

##generate from the posterior distribution of mu, sigma^2
analyze.data <- function(y,params.true,inputs){
	n <- length(y)
	mu.0 <- inputs[1]
	kappa.0 <- inputs[2]
	nu.0 <- inputs[3]
	sigsq.0 <- inputs[4]
	n.draws <- inputs[5]
	kappa.n <- kappa.0 + n
	mu.n <- (mu.0*kappa.0 + sum(y))/kappa.n
	nu.n <- nu.0 + n
	sigsq.n <- ( nu.0*sigsq.0 + (n-1)*var(y) + 
		kappa.0*n*(mean(y) - mu.0)^2/kappa.n ) / nu.n 

	
	sigsq.post <- rinvchisq(n.draws, nu.n, sigsq.n)

	var.mu.post <- sigsq.post/(kappa.n)
	mu.post <- rnorm(n.draws, mu.n, sqrt(var.mu.post))

	return(cbind(sigsq.post,mu.post))}

##generate from the posterior distribution of mu, sigma^2
##error sampling sigma^2
analyze.data.error1 <- function(y,params.true,inputs){
	n <- length(y)
	mu.0 <- inputs[1]
	kappa.0 <- inputs[2]
	nu.0 <- inputs[3]
	sigsq.0 <- inputs[4]
	n.draws <- inputs[5]
	kappa.n <- kappa.0 + n
	mu.n <- (mu.0*kappa.0 + sum(y))/kappa.n
	nu.n <- nu.0 + n
	sigsq.n <- ( nu.0*sigsq.0 + (n-1)*var(y) + 
		kappa.0*n*(mean(y) - mu.0)^2/kappa.n ) / nu.n 

	sigsq.post <- rinvchisq(n.draws, nu.n, sigsq.0)

	var.mu.post <- sigsq.post/(kappa.n)
	mu.post <- rnorm(n.draws, mu.n, sqrt(var.mu.post))

	return(cbind(sigsq.post,mu.post))}

##generate from the posterior distribution of mu, sigma^2
##error sampling mu
analyze.data.error2 <- function(y,params.true,inputs){
	n <- length(y)
	mu.0 <- inputs[1]
	kappa.0 <- inputs[2]
	nu.0 <- inputs[3]
	sigsq.0 <- inputs[4]
	n.draws <- inputs[5]
	kappa.n <- kappa.0 + n
	mu.n <- (mu.0*kappa.0 + sum(y))/kappa.n
	nu.n <- nu.0 + n
	sigsq.n <- ( nu.0*sigsq.0 + (n-1)*var(y) + 
		kappa.0*n*(mean(y) - mu.0)^2/kappa.n ) / nu.n 

	
	sigsq.post <- rinvchisq(n.draws, nu.n, sigsq.n)

	var.mu.post <- sigsq.post/(kappa.n)
	mu.post <- rnorm(n.draws, mu.n, var.mu.post)

	return(cbind(sigsq.post,mu.post))}

##function inputs
hyper<-c(6,5,5,7)
n<-20
n.draws<-5000

generate.param.inputs<-hyper
generate.data.inputs<-n
analyze.data.inputs<-c(hyper,n.draws)

##run validation function for the three model-fitting functions
tst.0 <- validate(generate.param = generate.param, generate.param.inputs = 
	generate.param.inputs, generate.data = generate.data, 
	generate.data.inputs = generate.data.inputs, analyze.data = 
	analyze.data, analyze.data.inputs = analyze.data.inputs, 
	n.rep = 20, params.batch = expression(sigma^2,mu), n.batch = c(1,1))


tst.1 <- validate(generate.param = generate.param, generate.param.inputs = 
	generate.param.inputs, generate.data = generate.data, 
	generate.data.inputs = generate.data.inputs, analyze.data = 
	analyze.data.error1, analyze.data.inputs = analyze.data.inputs, 
	n.rep = 20, params.batch = expression(sigma^2,mu), n.batch = c(1,1))


tst.2 <- validate(generate.param = generate.param, generate.param.inputs = 
	generate.param.inputs, generate.data = generate.data, 
	generate.data.inputs = generate.data.inputs, analyze.data = 
	analyze.data.error2, analyze.data.inputs = analyze.data.inputs, 
	n.rep = 20, params.batch = expression(sigma^2,mu), n.batch = c(1,1))

BayesValidate documentation built on May 1, 2019, 8:07 p.m.