library(BayesFactor) options(BFprogress = FALSE) bfversion = BFInfo() session = sessionInfo()[[1]] rversion = paste(session$version.string," on ",session$platform,sep="") options(markdown.HTML.stylesheet = 'extra/manual.css') library(knitr) opts_chunk$set(dpi = 200, out.width = "67%") options(digits=3) require(graphics) set.seed(2)

This R markdown file runs a series of tests to ensure that the BayesFactor package is giving correct answers, and can gracefully handle probable input.

library(arm) library(lme4)

First we generate some data.

# Number of participants N <- 20 sig2 <- 1 sig2ID <- 1 # 3x3x3 design, with participant as random factor effects <- expand.grid(A = c("A1","A2","A3"), B = c("B1","B2","B3"), C = c("C1","C2","C3"), ID = paste("Sub",1:N,sep="") ) Xdata <- model.matrix(~ A*B*C + ID, data=effects) beta <- matrix(c(50, -.2,.2, 0,0, .1,-.1, rnorm(N-1,0,sqrt(sig2ID)), 0,0,0,0, -.1,.1,.1,-.1, 0,0,0,0, 0,0,0,0,0,0,0,0), ncol=1) effects$y = rnorm(Xdata%*%beta,Xdata%*%beta,sqrt(sig2))

# Typical repeated measures ANOVA summary(fullaov <- aov(y ~ A*B*C + Error(ID/(A*B*C)),data=effects))

We can plot the data with standard errors:

mns <- tapply(effects$y,list(effects$A,effects$B,effects$C),mean) stderr = sqrt((sum(resid(fullaov[[3]])^2)/fullaov[[3]]$df.resid)/N) par(mfrow=c(1,3),cex=1.1) for(i in 1:3){ matplot(mns[,,i],xaxt='n',typ='b',xlab="A",main=paste("C",i), ylim=range(mns)+c(-1,1)*stderr,ylab="y") axis(1,at=1:3,lab=1:3) segments(1:3 + mns[,,i]*0,mns[,,i] + stderr,1:3 + mns[,,i]*0,mns[,,i] - stderr,col=rgb(0,0,0,.3)) }

Compute the Bayes factors, while testing the Laplace approximation

t.is = system.time(bfs.is <- anovaBF(y ~ A*B*C + ID, data = effects, whichRandom="ID") ) t.la = system.time(bfs.la <- anovaBF(y ~ A*B*C + ID, data = effects, whichRandom="ID", method = "laplace") )

t.is t.la plot(log(extractBF(sort(bfs.is))$bf),log(extractBF(sort(bfs.la))$bf), xlab="Default Sampler",ylab="Laplace approximation", pch=21,bg=rgb(0,0,1,.2),col="black",asp=TRUE,cex=1.2) abline(0,1) bfs.is

We can use samples from the posterior distribution to compare `BayesFactor`

with `lmer`

and `arm`

.

chains <- lmBF(y ~ A + B + C + ID, data=effects, whichRandom = "ID", posterior=TRUE, iterations=10000) lmerObj <- lmer(y ~ A + B + C + (1|ID), data=effects) # Use arm function sim() to sample from posterior chainsLmer = sim(lmerObj,n.sims=10000)

Compare estimates of variance

BF.sig2 <- chains[,colnames(chains)=="sig2"] AG.sig2 <- (chainsLmer@sigma)^2 qqplot(log(BF.sig2),log(AG.sig2),pch=21,bg=rgb(0,0,1,.2), col=NULL,asp=TRUE,cex=1,xlab="BayesFactor samples", ylab="arm samples",main="Posterior samples of\nerror variance") abline(0,1)

Compare estimates of participant effects:

AG.raneff <- chainsLmer@ranef$ID[,,1] BF.raneff <- chains[,grep('ID-',colnames(chains),fixed='TRUE')] plot(colMeans(BF.raneff),colMeans(AG.raneff),pch=21,bg=rgb(0,0,1,.2),col="black",asp=TRUE,cex=1.2,xlab="BayesFactor estimate",ylab="arm estimate",main="Random effect posterior means") abline(0,1)

Compare estimates of fixed effects:

AG.fixeff <- chainsLmer@fixef BF.fixeff <- chains[,1:10] # Adjust AG results from reference cell to sum to 0 Z = c(1, 1/3, 1/3, 1/3, 1/3, 1/3, 1/3, 0, -1/3, -1/3, 0, 0, 0, 0, 0, 2/3, -1/3, 0, 0, 0, 0, 0, -1/3, 2/3, 0, 0, 0, 0, 0, 0, 0, -1/3, -1/3, 0, 0, 0, 0, 0, 2/3, -1/3, 0, 0, 0, 0, 0, -1/3, 2/3, 0, 0, 0, 0, 0, 0, 0, -1/3, -1/3, 0, 0, 0, 0, 0, 2/3, -1/3, 0, 0, 0, 0, 0, -1/3, 2/3) dim(Z) = c(7,10) Z = t(Z) AG.fixeff2 = t(Z%*%t(AG.fixeff)) ## Our grand mean has heavier tails qqplot(BF.fixeff[,1],AG.fixeff2[,1],pch=21,bg=rgb(0,0,1,.2),col=NULL,asp=TRUE,cex=1,xlab="BayesFactor estimate",ylab="arm estimate",main="Grand mean posterior samples") abline(0,1) plot(colMeans(BF.fixeff[,-1]),colMeans(AG.fixeff2[,-1]),pch=21,bg=rgb(0,0,1,.2),col="black",asp=TRUE,cex=1.2,xlab="BayesFactor estimate",ylab="arm estimate",main="Fixed effect posterior means") abline(0,1) ## Compare posterior standard deviations BFsd = apply(BF.fixeff[,-1],2,sd) AGsd = apply(AG.fixeff2[,-1],2,sd) plot(sort(AGsd/BFsd),pch=21,bg=rgb(0,0,1,.2),col="black",cex=1.2,ylab="Ratio of posterior standard deviations (arm/BF)",xlab="Fixed effect index") ## AG estimates are slightly larger, consistent with sig2 estimates ## probably due to prior

We begin by loading required packages...

library(languageR) library(xtable)

...and creating the data set to analyze.

data(primingHeidPrevRT) primingHeidPrevRT$lRTmin1 <- log(primingHeidPrevRT$RTmin1) ###Frequentist lr4 <- lmer(RT ~ Condition + (1|Word)+ (1|Subject) + lRTmin1 + RTtoPrime + ResponseToPrime + ResponseToPrime*RTtoPrime +BaseFrequency ,primingHeidPrevRT) # Get rid rid of some outlying response times INDOL <- which(scale(resid(lr4)) < 2.5) primHeidOL <- primingHeidPrevRT[INDOL,]

The first thing we have to do is center the continuous variables. This is done automatically by lmBF(), as required by Liang et al. (2008). This, of course, changes the definition of the intercept.

# Center continuous variables primHeidOL$BaseFrequency <- primHeidOL$BaseFrequency - mean(primHeidOL$BaseFrequency) primHeidOL$lRTmin1 <- primHeidOL$lRTmin1 - mean(primHeidOL$lRTmin1) primHeidOL$RTtoPrime <- primHeidOL$RTtoPrime - mean(primHeidOL$RTtoPrime)

Now we perform both analyses on the same data, and place the fixed effect estimates for both packages into their own vectors.

# LMER lr4b <- lmer( RT ~ Condition + ResponseToPrime + (1|Word)+ (1|Subject) + lRTmin1 + RTtoPrime + ResponseToPrime*RTtoPrime + BaseFrequency , primHeidOL) # BayesFactor B5out <- lmBF( RT ~ Condition + ResponseToPrime + Word + Subject + lRTmin1 + RTtoPrime + ResponseToPrime*RTtoPrime + BaseFrequency , primHeidOL , whichRandom = c("Word", "Subject"), posterior = TRUE, iteration = 50000,columnFilter=c("Word","Subject")) lmerEff <- fixef(lr4b) bfEff <- colMeans(B5out[,1:10])

`lmer`

uses a "reference cell" parameterization, rather than imposing sum-to-0 constraints. We can tell what the reference cell is by looking at the parameter names.

print(xtable(cbind("lmer fixed effects"=names(lmerEff))), type='html')

Notice what's missing: for the categorical parameters, we are missing `Conditionbaseheid`

and `ResponseToPrimecorrect`

. For the slope parameters, we are missing `ResponseToPrimecorrect:RTtoPrime`

. The missing effects tell us what the reference cells are. Since the reference cell parameterization is just a linear transformation of the sum-to-0 parameterization, we can create a matrix that allows us to move from one to the other. We call this $10 \times 7$ matrix `Z`

. It takes the 7 "reference-cell" parameters from `lmer`

and maps them into the 10 linearly constrained parameters from `lmBF`

.

The first row of `Z`

transforms the intercept (reference cell) to the grand mean (sum-to-0). We have to add half of the two fixed effects back into the intercept. The second and third row divide the totl effect of `Condition`

into two equal parts, one for `baseheid`

and one for `heid`

. Rows four and five do the same for `ResponseToPrime`

.

The slopes that do not enter into interactions are fine as they are; however, `ResponseToPrimecorrect:RTtoPrime`

serves as our reference cell for the `ResponseToPrime:RTtoPrime`

interaction. We treat these slopes analogously to the grand mean; we take `RTtoPrime`

and add half the `ResponseToPrimeincorrect:RTtoPrime`

effect to it, to make it a grand mean slope. The last two rows divide up the `ResponseToPrimeincorrect:RTtoPrime`

effect between `ResponseToPrimeincorrect:RTtoPrime`

and `ResponseToPrimecorrect:RTtoPrime`

.

# Adjust lmer results from reference cell to sum to 0 Z = c(1, 1/2, 1/2, 0, 0, 0, 0, 0, -1/2, 0, 0, 0, 0, 0, 0, 1/2, 0, 0, 0, 0, 0, 0, 0,-1/2, 0, 0, 0, 0, 0, 0, 1/2, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1/2, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, -1/2, 0, 0, 0, 0, 0, 0, 1/2) dim(Z) = c(7,10) Z = t(Z) # Do reparameterization by pre-multimplying the parameter vector by Z reparLmer <- Z %*% matrix(lmerEff,ncol=1) # put results in data.frame for comparison sideBySide <- data.frame(BayesFactor=bfEff,lmer=reparLmer)

We can look at them side by side for comparison:

print(xtable(sideBySide,digits=4), type='html')

...and plot them:

# Notice Bayesian shrinkage par(cex=1.5) plot(sideBySide[-1,],pch=21,bg=rgb(0,0,1,.2),col="black",asp=TRUE,cex=1.2, main="fixed effects\n (excluding grand mean)") abline(0,1, lty=2)

The results are quite close to one another, with a bit of Bayesian shrinkage.

*This document was compiled with version r bfversion of BayesFactor (r rversion).*

**Any scripts or data that you put into this service are public.**

Embedding an R snippet on your website

Add the following code to your website.

For more information on customizing the embed code, read Embedding Snippets.