BayesFactor (version 0.9.12-4.7)

nWayAOV: Use ANOVA design matrix to compute Bayes factors or sample posterior

Description

Computes a single Bayes factor, or samples from the posterior, for an ANOVA model defined by a design matrix

Usage

nWayAOV(
  y,
  X,
  gMap,
  rscale,
  iterations = 10000,
  progress = getOption("BFprogress", interactive()),
  callback = function(...) as.integer(0),
  gibbs = NULL,
  posterior = FALSE,
  ignoreCols = NULL,
  thin = 1,
  method = "auto",
  continuous = FALSE,
  noSample = FALSE
)

Value

If posterior is FALSE, a vector of length 2 containing the computed log(e) Bayes factor (against the intercept-only null), along with a proportional error estimate on the Bayes factor. Otherwise, an object of class mcmc, containing MCMC samples from the posterior is returned.

Arguments

y

vector of observations

X

design matrix whose number of rows match length(y).

gMap

vector grouping the columns of X (see Details).

rscale

a vector of prior scale(s) of appropriate length (see Details).

iterations

Number of Monte Carlo samples used to estimate Bayes factor or posterior

progress

if TRUE, show progress with a text progress bar

callback

callback function for third-party interfaces

gibbs

will be deprecated. See posterior

posterior

if TRUE, return samples from the posterior using Gibbs sampling, instead of the Bayes factor

ignoreCols

if NULL and posterior=TRUE, all parameter estimates are returned in the MCMC object. If not NULL, a vector of length P-1 (where P is number of columns in the design matrix) giving which effect estimates to ignore in output

thin

MCMC chain to every thin iterations. Default of 1 means no thinning. Only used if posterior=TRUE

method

the integration method (only valid if posterior=FALSE); one of "simple", "importance", "laplace", or "auto"

continuous

either FALSE if no continuous covariates are included, or a logical vector of length equal to number of columns of X indicating which columns of the design matrix represent continuous covariates

noSample

if TRUE, do not sample, instead returning NA. This is intended to be used with functions generating and testing many models at one time, such as anovaBF

Author

Richard D. Morey (), Jeffery N. Rouder ()

Details

This function is not meant to be called by end-users, although technically-minded users can call this function for flexibility beyond what the other functions in this package provide. See lmBF for a user-friendly front-end to this function. Details about the priors can be found in the help for anovaBF and the references therein.

Argument gMap provides a way of grouping columns of the design matrix as a factor; the effects in each group will share a common \(g\) parameter. gMap should be a vector of the same length as the number of nonconstant rows in X. It will contain all integers from 0 to \(N_g-1\), where \(N_g\) is the total number of \(g\) parameters. Each element of gMap specifies the group to which that column belongs.

If all columns belonging to a group are adjacent, struc can instead be used to compactly represent the groupings. struc is a vector of length \(N_g\). Each element specifies the number columns in the group.

The vector rscale should be of length \(N_g\), and contain the prior scales of the standardized effects. See Rouder et al. (2012) for more details and the help for anovaBF for some typical values.

The method used to estimate the Bayes factor depends on the method argument. "simple" is most accurate for small to moderate sample sizes, and uses the Monte Carlo sampling method described in Rouder et al. (2012). "importance" uses an importance sampling algorithm with an importance distribution that is multivariate normal on log(g). "laplace" does not sample, but uses a Laplace approximation to the integral. It is expected to be more accurate for large sample sizes, where MC sampling is slow. If method="auto", then an initial run with both samplers is done, and the sampling method that yields the least-variable samples is chosen. The number of initial test iterations is determined by options(BFpretestIterations).

If posterior samples are requested, the posterior is sampled with a Gibbs sampler.

References

Rouder, J. N., Morey, R. D., Speckman, P. L., Province, J. M., (2012) Default Bayes Factors for ANOVA Designs. Journal of Mathematical Psychology. 56. p. 356-374.

See Also

See lmBF for the user-friendly front end to this function; see regressionBF and anovaBF for testing many regression or ANOVA models simultaneously.

Examples

## Classical example, taken from t.test() example
## Student's sleep data
data(sleep)
plot(extra ~ group, data = sleep)

## traditional ANOVA gives a p value of 0.00283
summary(aov(extra ~ group + Error(ID/group), data = sleep))

## Build design matrix
group.column <- rep(1/c(-sqrt(2),sqrt(2)),each=10)
subject.matrix <- model.matrix(~sleep$ID - 1,data=sleep$ID)
## Note that we include no constant column
X <- cbind(group.column, subject.matrix)

## (log) Bayes factor of full model against grand-mean only model
bf.full <- nWayAOV(y = sleep$extra, X = X, gMap = c(0,rep(1,10)), rscale=c(.5,1))
exp(bf.full[['bf']])

## Compare with lmBF result (should be about the same, give or take 1%)
bf.full2 <- lmBF(extra ~ group + ID, data = sleep, whichRandom = "ID")
bf.full2