Stable version: CRAN page - Package NEWS (including version changes)
Development version: Development page - Development package NEWSThe BayesFactor
package enables the computation of Bayes
factors in standard designs, such as one- and two- sample designs, ANOVA
designs, and regression. The Bayes factors are based on work spread
across several papers. This document is designed to show users how to
compute Bayes factors using the package by example. It is not designed
to present the models used in the comparisons in detail; for that, see
the BayesFactor
help and especially the references listed
in this manual. Complete references are given at the end of this document.
If you need help or think you’ve found a bug, please use the links at
the top of this document to contact the developers. When asking a
question or reporting a bug, please send example code and data, the
exact errors you’re seeing (a cut-and-paste from the R console will
work) and instructions for reproducing it. Also, report the output of
BFInfo()
and sessionInfo()
, and let us know
what operating system you’re running.
The BayesFactor
package must be installed and loaded
before it can be used. Installing the package can be done in several
ways and will not be covered here. Once it is installed, use the
library
function to load it:
This command will make the BayesFactor
package ready to
use.
The table below lists some of the functions in the
BayesFactor
package that will be demonstrated in this
manual. For more complete help on the use of these functions, see the
corresponding help()
page in R.
Function | Description |
---|---|
ttestBF |
Bayes factors for one- and two- sample designs |
anovaBF |
Bayes factors comparing many ANOVA models |
regressionBF |
Bayes factors comparing many linear regression models |
generalTestBF |
Bayes factors for all restrictions on a full model (0.9.4+) |
lmBF |
Bayes factors for specific linear models (ANOVA or regression) |
correlationBF |
Bayes factors for linear correlations |
proportionBF |
Bayes factors for tests of single proportions |
contingencyTableBF |
Bayes factors for contingency tables |
posterior |
Sample from the posterior distribution of the numerator of a Bayes factor object |
recompute |
Recompute a Bayes factor or MCMC chain, possibly increasing the precision of the estimate |
compare |
Compare two models; typically used to compare two models in
BayesFactor MCMC objects |
The t test section below has examples showing how to manipulate Bayes
factor objects, but all these functions will work with Bayes factors
generated from any function in the BayesFactor
package.
Function | Description |
---|---|
/ |
Divide two Bayes factor objects to create new model comparisons, or
invert with 1/ |
t |
“Flip” (transpose) a Bayes factor object |
c |
Concatenate two Bayes factor objects together, assuming they have the same denominator |
[ |
Use indexing to select a subset of the Bayes factors |
plot |
plot a Bayes factor object |
sort |
Sort a Bayes factor object |
is.na |
Determine whether a Bayes factor object contains missing values |
head ,tail |
Return the n highest or lowest Bayes factor in an
object |
max , min |
Return the highest or lowest Bayes factor in an object |
which.max ,which.min |
Return the index of the highest or lowest Bayes factor |
as.vector |
Convert to a simple vector (denominator will be lost!) |
as.data.frame |
Convert to data.frame (denominator will be lost!) |
The ttestBF
function is used to obtain Bayes factors
corresponding to tests of a single sample’s mean, or tests that two
independent samples have the same mean.
We use the sleep
data set in R to demonstrate a
one-sample t test. This is a paired design; for details about the data
set, see ?sleep
. One way of analyzing these data is to
compute difference scores by subtracting a participant’s score in one
condition from their score in the other:
data(sleep)
## Compute difference scores
diffScores = sleep$extra[1:10] - sleep$extra[11:20]
## Traditional two-tailed t test
t.test(diffScores)
##
## One Sample t-test
##
## data: diffScores
## t = -4, df = 9, p-value = 0.003
## alternative hypothesis: true mean is not equal to 0
## 95 percent confidence interval:
## -2.46 -0.70
## sample estimates:
## mean of x
## -1.58
We can do a Bayesian version of this analysis using the
ttestBF
function, which performs the “JZS” t test described
by Rouder, Speckman, Sun, Morey, and Iverson
(2009). In this model, the true standardized difference \(\delta=(\mu-\mu_0)/\sigma_\epsilon\) is
assumed to be 0 under the null hypothesis, and \(\text{Cauchy}(\text{scale}=r)\) under the
alternative. The default \(r\) scale in
BayesFactor
for t tests is \(\sqrt{2}/2\). See ?ttestBF
for
more details.
bf = ttestBF(x = diffScores)
## Equivalently:
## bf = ttestBF(x = sleep$extra[1:10],y=sleep$extra[11:20], paired=TRUE)
bf
## Bayes factor analysis
## --------------
## [1] Alt., r=0.707 : 17.3 ±0%
##
## Against denominator:
## Null, mu = 0
## ---
## Bayes factor type: BFoneSample, JZS
The bf
object contains the Bayes factor, and shows the
numerator and denominator models for the Bayes factor comparison. In our
case, the Bayes factor for the comparison of the alternative versus the
null is 17.259. After the Bayes factor is a proportional error estimate
on the Bayes factor.
There are a number of operations we can perform on our Bayes factor, such as taking the reciprocal:
## Bayes factor analysis
## --------------
## [1] Null, mu=0 : 0.0579 ±0%
##
## Against denominator:
## Alternative, r = 0.707106781186548, mu =/= 0
## ---
## Bayes factor type: BFoneSample, JZS
or sampling from the posterior of the numerator model:
##
## Iterations = 1:1000
## Thinning interval = 1
## Number of chains = 1
## Sample size per chain = 1000
##
## 1. Empirical mean and standard deviation for each variable,
## plus standard error of the mean:
##
## Mean SD Naive SE Time-series SE
## mu -1.42 0.436 0.0138 0.0154
## sig2 2.02 1.157 0.0366 0.0395
## delta -1.11 0.427 0.0135 0.0162
## g 6.26 58.623 1.8538 1.8538
##
## 2. Quantiles for each variable:
##
## 2.5% 25% 50% 75% 97.5%
## mu -2.289 -1.705 -1.43 -1.141 -0.597
## sig2 0.744 1.270 1.69 2.446 5.223
## delta -1.973 -1.383 -1.08 -0.813 -0.347
## g 0.176 0.592 1.13 2.928 33.734
The posterior
function returns a object of type
BFmcmc
, which inherits the methods of the mcmc
class from the coda
package. We can thus use summary
, plot
,
and other useful methods on the result of posterior
. If we
were unhappy with the number of iterations we sampled for
chains
, we can recompute
with more iterations,
and then plot
the results:
Directional hypotheses can also be tested with ttestBF
(Morey & Rouder, 2011). The argument
nullInterval
can be passed as a vector of length 2, and
defines an interval to compare to the point null. If null interval is
defined, two Bayes factors are returned: the Bayes factor of
the null interval against the alternative, and the Bayes factor of the
complement of the interval to the point null.
Suppose, for instance, we wanted to test the one-sided hypotheses
that \(\delta<0\) versus the point
null. We set nullInterval
to c(-Inf,0)
:
## Bayes factor analysis
## --------------
## [1] Alt., r=0.707 -Inf<d<0 : 34.4 ±0%
## [2] Alt., r=0.707 !(-Inf<d<0) : 0.101 ±0%
##
## Against denominator:
## Null, mu = 0
## ---
## Bayes factor type: BFoneSample, JZS
We may not be interested in tests against the point null. If we are
interested in the Bayes factor test that \(\delta<0\) versus \(\delta>0\) we can compute it using the
result above. Since the object contains two Bayes factors, both with the
same denominator, and \[
\left.\frac{A}{C}\middle/\frac{B}{C}\right. = \frac{A}{B},
\] we can divide the two Bayes factors in bfInferval
to obtain the desired test:
## Bayes factor analysis
## --------------
## [1] Alt., r=0.707 -Inf<d<0 : 341 ±0%
##
## Against denominator:
## Alternative, r = 0.707106781186548, mu =/= 0 !(-Inf<d<0)
## ---
## Bayes factor type: BFoneSample, JZS
The Bayes factor is about 340.
When we have multiple Bayes factors that all have the same
denominator, we can concatenate them into one object using the
c
function. Since bf
and
bfInterval
both share the point null denominator, we can do
this:
## Bayes factor analysis
## --------------
## [1] Alt., r=0.707 : 17.3 ±0%
## [2] Alt., r=0.707 -Inf<d<0 : 34.4 ±0%
## [3] Alt., r=0.707 !(-Inf<d<0) : 0.101 ±0%
##
## Against denominator:
## Null, mu = 0
## ---
## Bayes factor type: BFoneSample, JZS
The object allbf
now contains three Bayes factors, all
of which share the same denominator. If you try to concatenate Bayes
factors that do not share the same denominator,
BayesFactor
will return an error.
When you have a Bayes factor object with several numerators, there are several interesting ways to manipulate them. For instance, we can plot the Bayes factor object to obtain a graphical representation of the Bayes factors:
We can also divide a Bayes factor object by itself — or by a subset of itself — to obtain pairwise comparisons:
## denominator
## numerator Alt., r=0.707 Alt., r=0.707 -Inf<d<0
## Alt., r=0.707 1.00000 0.50146
## Alt., r=0.707 -Inf<d<0 1.99416 1.00000
## Alt., r=0.707 !(-Inf<d<0) 0.00584 0.00293
## denominator
## numerator Alt., r=0.707 !(-Inf<d<0)
## Alt., r=0.707 171
## Alt., r=0.707 -Inf<d<0 341
## Alt., r=0.707 !(-Inf<d<0) 1
The resulting object is of type BFBayesFactorList
, and
is a list of Bayes factor comparisons all of the same numerators
compared to different denominators. The resulting matrix can be
subsetted to return individual Bayes factor objects, or new
BFBayesFactorList
s:
## Bayes factor analysis
## --------------
## [1] Alt., r=0.707 : 0.501 ±0%
## [2] Alt., r=0.707 -Inf<d<0 : 1 ±0%
## [3] Alt., r=0.707 !(-Inf<d<0) : 0.00293 ±0%
##
## Against denominator:
## Alternative, r = 0.707106781186548, mu =/= 0 -Inf<d<0
## ---
## Bayes factor type: BFoneSample, JZS
## denominator
## numerator Alt., r=0.707 Alt., r=0.707 -Inf<d<0 Alt., r=0.707 !(-Inf<d<0)
## Alt., r=0.707 1 0.501 171
and they can also be transposed:
## denominator
## numerator Alt., r=0.707 Alt., r=0.707 -Inf<d<0
## Alt., r=0.707 1.00000 0.50146
## Alt., r=0.707 -Inf<d<0 1.99416 1.00000
## Alt., r=0.707 !(-Inf<d<0) 0.00584 0.00293
## denominator
## numerator Alt., r=0.707 Alt., r=0.707 -Inf<d<0
## Alt., r=0.707 1.00 0.501
## Alt., r=0.707 -Inf<d<0 1.99 1.000
## denominator
## numerator Alt., r=0.707 !(-Inf<d<0)
## Alt., r=0.707 171
## Alt., r=0.707 -Inf<d<0 341
If these values are desired in matrix form, the
as.matrix
function can be used to obtain a matrix.
The ttestBF
function can also be used to compute Bayes
factors in the two sample case as well. We use the chickwts
data set to demonstrate the two-sample t test. The chickwts
data set has six groups, but we reduce it to two for the
demonstration.
data(chickwts)
## Restrict to two groups
chickwts = chickwts[chickwts$feed %in% c("horsebean","linseed"),]
## Drop unused factor levels
chickwts$feed = factor(chickwts$feed)
## Plot data
plot(weight ~ feed, data = chickwts, main = "Chick weights")
Chick weight appears to be affected by the feed type.
##
## Two Sample t-test
##
## data: weight by feed
## t = -3, df = 20, p-value = 0.008
## alternative hypothesis: true difference in means between group horsebean and group linseed is not equal to 0
## 95 percent confidence interval:
## -100.2 -16.9
## sample estimates:
## mean in group horsebean mean in group linseed
## 160 219
We can also compute the corresponding Bayes factor. There are two
ways of specifying a two-sample test: the formula interface and through
the x
and y
arguments. We show the formula
interface here:
## Bayes factor analysis
## --------------
## [1] Alt., r=0.707 : 5.98 ±0%
##
## Against denominator:
## Null, mu1-mu2 = 0
## ---
## Bayes factor type: BFindepSample, JZS
As before, we can sample from the posterior distribution for the numerator model:
Note that the samples assume an (equivalent) ANOVA model; see
?ttestBF
and for notes on the differences in interpretation
of the \(r\) scale parameter between
the two models.
Rouder and Morey (2011; link) discuss a meta-analytic extension of the \(t\) test, whereby multiple \(t\) statistics, along with their corresponding sample sizes, are combined in a single meta-analytic analysis. The \(t\) statistics are assumed to arise from a a common effect size \(\delta\). The prior for the effect size \(\delta\) is the same as that for the \(t\) tests described above.
The meta.ttestBF
function is used to perform
meta-analytic \(t\) tests. It requires
as input a vector of \(t\) statistics,
and one or two vectors of sample sizes (arguments n1
and
n2
). For a set of one-sample \(t\) statistics, n1
should be
provided; for two-sample analyses, both n1
and
n2
should be provided.
As an example, we will replicate the analysis of Rouder & Morey (2011), using \(t\) statistics from Bem (2010; see Rouder & Morey for reference). We begin by defining the one-sample \(t\) statistics and sample sizes:
## Bem's t statistics from four selected experiments
t = c(-.15, 2.39, 2.42, 2.43)
N = c(100, 150, 97, 99)
Rouder and Morey opted for a one-sided analysis, and used an \(r\) scale parameter of 1 (instead of the
current default in BayesFactor
of \(\sqrt{2}/2\)).
## Bayes factor analysis
## --------------
## [1] Alt., r=1 0<d<Inf : 38.7 ±0%
## [2] Alt., r=1 !(0<d<Inf) : 0.00803 ±0%
##
## Against denominator:
## Null, d = 0
## ---
## Bayes factor type: BFmetat, JZS
Notice that as above, the analysis yields a Bayes factor for our selected interval against the null, as well as the Bayes factor for the complement of the interval against the null.
We can also sample from the posterior distribution of the
standardized effect size \(\delta\), as
above, using the posterior
function:
## Do analysis again, without nullInterval restriction
bf = meta.ttestBF(t=t, n1=N, rscale=1)
## Obtain posterior samples
chains = posterior(bf, iterations = 10000)
## Independent-candidate M-H acceptance rate: 98%
Notice that the posterior samples will respect the
nullInterval
argument if given; in order to get
unrestricted samples, perform an analysis with no interval restriction
and pass it to the posterior
function.
See ?meta.ttestBF
for more information.
The BayesFactor
package has two main functions that
allow the comparison of models with factors as predictors (ANOVA):
anovaBF
, which computes several model estimates at once,
and lmBF
, which computes one comparison at a time. We begin
by demonstrating a 3x2 fixed-effect ANOVA using the
ToothGrowth
data set. For details about the data set, see
?ToothGrowth
.
The ToothGrowth
data set contains three columns:
len
, the dependent variable, each of which is the length of
a guinea pig’s tooth after treatment with Vitamin C; supp
,
which is the supplement type (orange juice or ascorbic acid); and
dose
, which is the amount of Vitamin C administered.
data(ToothGrowth)
## Example plot from ?ToothGrowth
coplot(len ~ dose | supp, data = ToothGrowth, panel = panel.smooth,
xlab = "ToothGrowth data: length vs dose, given type of supplement")
## Treat dose as a factor
ToothGrowth$dose = factor(ToothGrowth$dose)
levels(ToothGrowth$dose) = c("Low", "Medium", "High")
summary(aov(len ~ supp*dose, data=ToothGrowth))
## Df Sum Sq Mean Sq F value Pr(>F)
## supp 1 205 205 15.57 0.00023 ***
## dose 2 2426 1213 92.00 < 2e-16 ***
## supp:dose 2 108 54 4.11 0.02186 *
## Residuals 54 712 13
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
There appears to be a large effect of the dosage, a small effect of
the supplement type, and perhaps a hint of an interaction. The
anovaBF
function will compute the Bayes factors of all
models against the intercept-only model; by default, it will choose the
subset of all models in which which an interaction can only be included
if all constituent effects or interactions are included (argument
whichModels
is set to withmain
, indicating
that interactions can only enter in with their main effects). However,
this setting can be changed, as we will demonstrate. First, we show the
default behavior.
## Bayes factor analysis
## --------------
## [1] supp : 1.2 ±0.01%
## [2] dose : 4.98e+12 ±0%
## [3] supp + dose : 2.92e+14 ±1.58%
## [4] supp + dose + supp:dose : 7.44e+14 ±1.01%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
The function will build the requested models from the terms included in the right-hand side of the formula; we could have specified the sum of the two terms, and we would have gotten the same models.
The Bayes factor analysis is consistent with the classical ANOVA
analysis; the favored model is the full model, with both main effects
and the two-way interaction. Suppose we were interested in comparing the
two main-effects model and the full model to the dose
-only
model. We could use indexing and division, along with the
plot
function, to see a graphical representation of these
comparisons:
The model with the main effect of supp
and the
supp:dose
interaction is preferred quite strongly over the
dose
-only model.
There are a number of other options for how to select subsets of
models to test. The whichModels
argument to
anovaBF
controls which subsets are tested. As described
previously, the default is withmain
, where interactions are
only allowed if all constituent sub-effects are included. The other
three options currently available are all
, which tests all
models; top
, which includes the full model and all models
that can be formed by removing one interaction or main effect; and
bottom
, which adds single effects one at a time to the null
model.
The argument whichModels='all'
should be used with
caution: a three-way ANOVA model will contain \(2^{2^3-1}-1 = 127\) model comparisons; a
four-way ANOVA, \(2^{2^4-1}-1 = 32767\)
models, and a five-way ANOVA just over 2.1 billion models. Depending on
the speed of your computer, a four-way ANOVA may take several hours to a
day, but a five-way ANOVA is probably not feasible.
One alternative is whichModels='top'
, which reduces the
number of comparisons to \(2^k-1\),
where \(k\) is the number of factors,
which is manageable. In orthogonal designs, one can construct tests of
each main effect or interaction by comparing the full model to the model
with all effects except the one of interest:
## Bayes factor top-down analysis
## --------------
## When effect is omitted from supp + dose + supp:dose , BF is...
## [1] Omit dose:supp : 0.385 ±3.32%
## [2] Omit dose : 7.11e-16 ±12.2%
## [3] Omit supp : 0.011 ±4.17%
##
## Against denominator:
## len ~ supp + dose + supp:dose
## ---
## Bayes factor type: BFlinearModel, JZS
Note that all of the Bayes factors are less than 1, indicating that removing any effect from the full model is deleterious.
Another way we can reduce the number of models tested is simply to
test only specific models of interest. In the example above, for
instance, we might want to compare the model with the interaction to the
model with only the main effects, if our effect of interest was the
interaction. We can do this with the lmBF
function.
bfMainEffects = lmBF(len ~ supp + dose, data = ToothGrowth)
bfInteraction = lmBF(len ~ supp + dose + supp:dose, data = ToothGrowth)
## Compare the two models
bf = bfInteraction / bfMainEffects
bf
## Bayes factor analysis
## --------------
## [1] supp + dose + supp:dose : 2.79 ±2.51%
##
## Against denominator:
## len ~ supp + dose
## ---
## Bayes factor type: BFlinearModel, JZS
The model with the interaction effect is preferred by a factor of about 3.
Suppose that we were unhappy with the ~2.5% proportional error on the
Bayes factor bf
. anovaBF
and lmBF
use Monte Carlo integration to estimate the Bayes factors. The default
number of Monte Carlo samples is 10,000 but this can be increased. We
could use the recompute
to reduce the error. The
recompute
function performs the sampling required to build
the Bayes factor object again:
## Bayes factor analysis
## --------------
## [1] supp + dose + supp:dose : 2.74 ±0.51%
##
## Against denominator:
## len ~ supp + dose
## ---
## Bayes factor type: BFlinearModel, JZS
The proportional error is now below 1%.
As before, we can use MCMC methods to estimate parameters through the
posterior
function:
## Sample from the posterior of the full model
chains = posterior(bfInteraction, iterations = 10000)
## 1:13 are the only "interesting" parameters
summary(chains[,1:13])
##
## Iterations = 1:10000
## Thinning interval = 1
## Number of chains = 1
## Sample size per chain = 10000
##
## 1. Empirical mean and standard deviation for each variable,
## plus standard error of the mean:
##
## Mean SD Naive SE Time-series SE
## mu 18.819 0.487 0.00487 0.00487
## supp-OJ 1.679 0.488 0.00488 0.00549
## supp-VC -1.679 0.488 0.00488 0.00549
## dose-Low -8.069 0.683 0.00683 0.00705
## dose-Medium 0.910 0.680 0.00680 0.00666
## dose-High 7.159 0.684 0.00684 0.00710
## supp:dose-OJ.&.Low 0.562 0.603 0.00603 0.00616
## supp:dose-OJ.&.Medium 0.822 0.621 0.00621 0.00723
## supp:dose-OJ.&.High -1.384 0.663 0.00663 0.00833
## supp:dose-VC.&.Low -0.562 0.603 0.00603 0.00616
## supp:dose-VC.&.Medium -0.822 0.621 0.00621 0.00723
## supp:dose-VC.&.High 1.384 0.663 0.00663 0.00833
## sig2 14.039 2.772 0.02772 0.03260
##
## 2. Quantiles for each variable:
##
## 2.5% 25% 50% 75% 97.5%
## mu 17.880 18.492 18.817 19.142 19.765
## supp-OJ 0.719 1.356 1.679 2.002 2.647
## supp-VC -2.647 -2.002 -1.679 -1.356 -0.719
## dose-Low -9.421 -8.516 -8.073 -7.608 -6.746
## dose-Medium -0.418 0.457 0.904 1.370 2.234
## dose-High 5.808 6.696 7.167 7.615 8.516
## supp:dose-OJ.&.Low -0.613 0.160 0.552 0.945 1.766
## supp:dose-OJ.&.Medium -0.354 0.404 0.801 1.223 2.093
## supp:dose-OJ.&.High -2.740 -1.825 -1.365 -0.920 -0.154
## supp:dose-VC.&.Low -1.766 -0.945 -0.552 -0.160 0.613
## supp:dose-VC.&.Medium -2.093 -1.223 -0.801 -0.404 0.354
## supp:dose-VC.&.High 0.154 0.920 1.365 1.825 2.740
## sig2 9.625 12.062 13.692 15.615 20.373
And we can plot the posteriors of some selected effects:
In order to demonstrate the analysis of mixed models using
BayesFactor
, we will load the puzzles
data
set, which is part of the BayesFactor
package. See
?puzzles
for details. The data set consists of four
columns: RT
the dependent variable, which is the number of
seconds that it took to complete a puzzle; ID
which is a
participant identifier; and shape
and color
,
which are two factors that describe the type of puzzle solved.
shape
and color
each have two levels, and each
of 12 participants completed puzzles within combination of
shape
and color
. The design is thus 2x2
factorial within-subjects.
We first load the data, then perform a traditional within-subjects ANOVA.
(Code for plot omitted) Individual circles joined by lines show
participants; red squares/lines show the means and within-subject
standard errors. From the plot, there appear to be main effects of
color
and shape, but no interaction.
##
## Error: ID
## Df Sum Sq Mean Sq F value Pr(>F)
## Residuals 11 226 20.6
##
## Error: ID:shape
## Df Sum Sq Mean Sq F value Pr(>F)
## shape 1 12.0 12.00 7.54 0.019 *
## Residuals 11 17.5 1.59
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Error: ID:color
## Df Sum Sq Mean Sq F value Pr(>F)
## color 1 12.0 12.00 13.9 0.0033 **
## Residuals 11 9.5 0.86
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Error: ID:shape:color
## Df Sum Sq Mean Sq F value Pr(>F)
## shape:color 1 0.0 0.00 0 1
## Residuals 11 30.5 2.77
The classical ANOVA appears to corroborate the impression from the
plot. In order to compute the Bayes factor, we must tell
anovaBF
that ID
is an additive effect on top
of the other effects (as is typically assumed) and is a random factor.
The anovaBF
call below shows how this is done:
We alert anovaBF
to the random factor using the
whichRandom
argument. whichRandom
should
contain a character vector with the names of all random factors in it.
All other factors are assumed to be fixed. The anovaBF
will
find all the fixed effects in the formula, and compute the Bayes factor
for the subset of combinations determined by the
whichModels
argument (see the previous section). Note that
anovaBF
does not test random factors; they are assumed to
be nuisance factors. The null model in a test with random factors is not
the intercept-only model; it is the model containing the random effects.
The Bayes factor object bf
thus now contains Bayes factors
comparing various combinations of the fixed effects and an additive
effect of ID
against a denominator containing only
ID
:
## Bayes factor analysis
## --------------
## [1] shape + ID : 2.81 ±0.91%
## [2] color + ID : 2.81 ±0.83%
## [3] shape + color + ID : 11.9 ±3%
## [4] shape + color + shape:color + ID : 4.23 ±2.2%
##
## Against denominator:
## RT ~ ID
## ---
## Bayes factor type: BFlinearModel, JZS
The main effects model is preferred against all models. We can plot the Bayes factor object to obtain a graphical representation of the model comparisons:
Because the anovaBF
function does not test random
factors, we must use lmBF
to build such tests. Doing so is
straightforward. Suppose that we wished to test the random effect
ID
in the puzzles
example. We might compare
the full model shape + color + shape:color + ID
to the same
model without ID
:
## Bayes factor analysis
## --------------
## [1] shape * color : 0.143 ±1.14%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
But notice that the denominator model is the intercept-only model;
the denominator in the previous analysis was the ID
only
model. We need to compare the model with no ID
effect to
the model with only ID
:
## Bayes factor analysis
## --------------
## [1] shape * color : 1.28e-06 ±1.14%
##
## Against denominator:
## RT ~ ID
## ---
## Bayes factor type: BFlinearModel, JZS
Since our bf
object and bf2
object now have
the same denominator, we can concatenate them into one Bayes factor
object:
and we can compare them by dividing:
## Bayes factor analysis
## --------------
## [1] shape + color + shape:color + ID : 3307085 ±2.48%
##
## Against denominator:
## RT ~ shape * color
## ---
## Bayes factor type: BFlinearModel, JZS
The model with ID
is preferred by a factor of over 1
million, which is not surprising.
Any model that is a combination of fixed and random factors,
including interations between fixed and random factors, can be
constructed and tested with lmBF
. anovaBF
is
designed to be a convenience function as is therefore somewhat limited
in flexibility with respect to the models types it can test; however,
because random effects are often nuisance effects, we believe
anovaBF
will be sufficient for most researchers’ use.
Model comparison in multiple linear regression using
BayesFactor
is done via the approach of Liang, Paulo, Molina, Clyde, and Berger (2008).
Further discussion can be found in Rouder
& Morey (in press). To demonstrate Bayes factor model comparison
in a linear regression context, we use the attitude
data
set in R. See ?attitude
. The attitude
consists
of the dependent variable rating
, along with 6 predictors.
We can use BayesFactor
to compute the Bayes factors for
many models simultaneously, or single Bayes factors against the model
containing no predictors.
data(attitude)
## Traditional multiple regression analysis
lmObj = lm(rating ~ ., data = attitude)
summary(lmObj)
##
## Call:
## lm(formula = rating ~ ., data = attitude)
##
## Residuals:
## Min 1Q Median 3Q Max
## -10.942 -4.356 0.316 5.543 11.599
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 10.7871 11.5893 0.93 0.3616
## complaints 0.6132 0.1610 3.81 0.0009 ***
## privileges -0.0731 0.1357 -0.54 0.5956
## learning 0.3203 0.1685 1.90 0.0699 .
## raises 0.0817 0.2215 0.37 0.7155
## critical 0.0384 0.1470 0.26 0.7963
## advance -0.2171 0.1782 -1.22 0.2356
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 7.07 on 23 degrees of freedom
## Multiple R-squared: 0.733, Adjusted R-squared: 0.663
## F-statistic: 10.5 on 6 and 23 DF, p-value: 1.24e-05
The period (.
) is shorthand for all remaining columns,
besides rating
. The predictors complaints
and
learning
appear most stongly related to the dependent
variable, especially complaints
. In order to compute the
Bayes factors for many model comparisons at onces, we use the
regressionBF
function. The most obvious set of all model
comparisons is all possible additive models, which is returned by
default:
## [1] 63
The object bf
now contains \(2^p-1\), or 63, model comparisons. Large
numbers of comparisons can get unweildy, so we can use the functions
built into R to manipulate the Bayes factor object.
## Bayes factor analysis
## --------------
## [1] privileges + learning + raises + critical + advance : 51 ±0%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
## Bayes factor analysis
## --------------
## [1] complaints : 417939 ±0.01%
## [2] complaints + learning : 207272 ±0%
## [3] complaints + learning + advance : 88042 ±0%
## [4] complaints + raises : 77499 ±0%
## [5] complaints + privileges : 75015 ±0%
## [6] complaints + advance : 72760 ±0%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
## Bayes factor analysis
## --------------
## [1] privileges + critical + advance : 0.645 ±0%
## [2] critical : 0.449 ±0%
## [3] advance : 0.447 ±0%
## [4] critical + advance : 0.239 ±0.01%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
## complaints
## 1
## Bayes factor analysis
## --------------
## [1] complaints : 1 ±0%
## [2] complaints + learning : 0.496 ±0.01%
## [3] complaints + learning + advance : 0.211 ±0.01%
## [4] complaints + raises : 0.185 ±0.01%
## [5] complaints + privileges : 0.179 ±0.01%
## [6] complaints + advance : 0.174 ±0.01%
##
## Against denominator:
## rating ~ complaints
## ---
## Bayes factor type: BFlinearModel, JZS
The model preferred by Bayes factor is the
complaints
-only model, followed by the
complaints + learning
model, as might have been expected by
the classical analysis.
We might also be interested in comparing the most complex model to
all models that can be formed by removing a single covariate, or,
similarly, comparing the intercept-only model to all models that can be
formed by added a covariate. These comparisons can be done by setting
the whichModels
argument to 'top'
and
'bottom'
, respectively. For example, for testing against
the most complex model:
bf = regressionBF(rating ~ ., data = attitude, whichModels = "top")
## The seventh model is the most complex
bf
## Bayes factor top-down analysis
## --------------
## When effect is omitted from complaints + privileges + learning + raises + critical + advance , BF is...
## [1] Omit advance : 1.73 ±0%
## [2] Omit critical : 3.23 ±0%
## [3] Omit raises : 3.13 ±0%
## [4] Omit learning : 0.727 ±0%
## [5] Omit privileges : 2.92 ±0%
## [6] Omit complaints : 0.0231 ±0%
##
## Against denominator:
## rating ~ complaints + privileges + learning + raises + critical + advance
## ---
## Bayes factor type: BFlinearModel, JZS
With all other covariates in the model, the model containing
complaints
is preferred to the model not containing
complaints
by a factor of almost 80. The model containing
learning
, is only barely favored to the one without (a
factor of about 1.3).
A similar “bottom-up” test can be done, by setting
whichModels
to 'bottom'
.
The mismatch between the tests of all models, the “top-down” test, and the “bottom-up” test shows that the covariates share variance with one another. As always, whether these tests are interpretable or useful will depend on the data at hand.
In cases where it is desired to only compare a small number of
models, the lmBF
function can be used. Consider the case
that we wish to compare the model containing only
complaints
to the model containing complaints
and learning
:
complaintsOnlyBf = lmBF(rating ~ complaints, data = attitude)
complaintsLearningBf = lmBF(rating ~ complaints + learning, data = attitude)
## Compare the two models
complaintsOnlyBf / complaintsLearningBf
## Bayes factor analysis
## --------------
## [1] complaints : 2.02 ±0.01%
##
## Against denominator:
## rating ~ complaints + learning
## ---
## Bayes factor type: BFlinearModel, JZS
The complaints
-only model is slightly preferred.
As with the other Bayes factors, it is possible to sample from the
posterior distribution of a particular model under consideration. If we
wanted to sample from the posterior distribution of the
complaints + learning
model, we could use the
posterior
function:
##
## Iterations = 1:10000
## Thinning interval = 1
## Number of chains = 1
## Sample size per chain = 10000
##
## 1. Empirical mean and standard deviation for each variable,
## plus standard error of the mean:
##
## Mean SD Naive SE Time-series SE
## mu 64.619 1.305 0.01305 0.01305
## complaints 0.609 0.127 0.00127 0.00133
## learning 0.201 0.139 0.00139 0.00139
## sig2 52.019 15.768 0.15768 0.19387
## g 1.972 7.349 0.07349 0.07611
##
## 2. Quantiles for each variable:
##
## 2.5% 25% 50% 75% 97.5%
## mu 62.0398 63.758 64.627 65.487 67.171
## complaints 0.3572 0.526 0.610 0.694 0.856
## learning -0.0749 0.110 0.202 0.293 0.472
## sig2 29.6847 41.015 49.551 59.708 90.126
## g 0.1672 0.457 0.840 1.679 9.252
Compare these to the corresponding results from the classical regression analysis:
##
## Call:
## lm(formula = rating ~ complaints + learning, data = attitude)
##
## Residuals:
## Min 1Q Median 3Q Max
## -11.56 -5.73 0.67 6.53 10.36
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 9.871 7.061 1.40 0.17
## complaints 0.644 0.118 5.43 9.6e-06 ***
## learning 0.211 0.134 1.57 0.13
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 6.82 on 27 degrees of freedom
## Multiple R-squared: 0.708, Adjusted R-squared: 0.686
## F-statistic: 32.7 on 2 and 27 DF, p-value: 6.06e-08
The results are quite similar, apart from the intercept. This is due
to the Bayesian model centering the covariates before analysis, so the
mu
parameter is the mean of \(y\) rather than the expected value of the
response variable when all uncentered covariates are equal to 0.
General linear models: mixing continuous and
categorical covariates ——– The anovaBF
and
regressionBF
functions are convenience functions designed
to test several hypotheses of a particular type at once. Neither
function allows the mixing of continuous and categorical covariates. If
it is desired to test a model including both kinds of covariates,
lmBF
function must be used. We will continue the
ToothGrowth
example, this time without converting
dose
to a categorical variable. Instead, we will model the
logarithm of the dose.
data(ToothGrowth)
# model log2 of dose instead of dose directly
ToothGrowth$dose = log2(ToothGrowth$dose)
# Classical analysis for comparison
lmToothGrowth <- lm(len ~ supp + dose + supp:dose, data=ToothGrowth)
summary(lmToothGrowth)
##
## Call:
## lm(formula = len ~ supp + dose + supp:dose, data = ToothGrowth)
##
## Residuals:
## Min 1Q Median 3Q Max
## -7.543 -2.492 -0.503 2.712 7.857
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 20.663 0.679 30.43 < 2e-16 ***
## suppVC -3.700 0.960 -3.85 0.0003 ***
## dose 6.415 0.832 7.71 2.3e-10 ***
## suppVC:dose 2.665 1.176 2.27 0.0274 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 3.72 on 56 degrees of freedom
## Multiple R-squared: 0.776, Adjusted R-squared: 0.764
## F-statistic: 64.5 on 3 and 56 DF, p-value: <2e-16
The classical analysis, presented for comparison, reveals extremely
low p values for the effects of the supplement type and of the dose, but
the interaction p value is more moderate, at about 0.03. We can use the
lmBF
function to compute the Bayes factors for all models
of interest against the null model, which in this case is the
intercept-only model. We then concatenate them into a single Bayes
factor object for convenience.
full <- lmBF(len ~ supp + dose + supp:dose, data=ToothGrowth)
noInteraction <- lmBF(len ~ supp + dose, data=ToothGrowth)
onlyDose <- lmBF(len ~ dose, data=ToothGrowth)
onlySupp <- lmBF(len ~ supp, data=ToothGrowth)
allBFs <- c(full, noInteraction, onlyDose, onlySupp)
allBFs
## Bayes factor analysis
## --------------
## [1] supp + dose + supp:dose : 1.58e+15 ±1.08%
## [2] supp + dose : 1.49e+15 ±1.33%
## [3] dose : 2.77e+13 ±0.01%
## [4] supp : 1.2 ±0.01%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
The highest two Bayes factors belong to the full model and the model with no interaction. We can directly compute the Bayes factor for the simpler model with no interaction against the full model:
## Bayes factor analysis
## --------------
## [1] supp + dose + supp:dose : 1.06 ±1.72%
##
## Against denominator:
## len ~ supp + dose
## ---
## Bayes factor type: BFlinearModel, JZS
The evidence here is clearly equivocal. We can also use the
posterior
function to compute parameter estimates.
chainsFull <- posterior(full, iterations = 10000)
# summary of the "interesting" parameters
summary(chainsFull[,1:7])
##
## Iterations = 1:10000
## Thinning interval = 1
## Number of chains = 1
## Sample size per chain = 10000
##
## 1. Empirical mean and standard deviation for each variable,
## plus standard error of the mean:
##
## Mean SD Naive SE Time-series SE
## mu 18.81 0.500 0.00500 0.00500
## supp-OJ 1.68 0.501 0.00501 0.00546
## supp-VC -1.68 0.501 0.00501 0.00546
## dose-dose 7.62 0.609 0.00609 0.00627
## supp:dose-OJ.&.dose -1.32 0.596 0.00596 0.00596
## supp:dose-VC.&.dose 1.32 0.596 0.00596 0.00596
## sig2 14.70 2.900 0.02900 0.03299
##
## 2. Quantiles for each variable:
##
## 2.5% 25% 50% 75% 97.5%
## mu 17.830 18.481 18.82 19.145 19.784
## supp-OJ 0.711 1.345 1.68 2.019 2.664
## supp-VC -2.664 -2.019 -1.68 -1.345 -0.711
## dose-dose 6.428 7.209 7.61 8.029 8.816
## supp:dose-OJ.&.dose -2.503 -1.713 -1.32 -0.923 -0.117
## supp:dose-VC.&.dose 0.117 0.923 1.32 1.713 2.503
## sig2 10.113 12.627 14.35 16.376 21.180
The left panel of the figure below shows the data and linear fits. The green points represent guinea pigs given the orange juice supplement (OJ); red points represent guinea pigs given the vitamin C supplement. The solid lines show the posterior means from the Bayesian model; the dashed lines show the classical least-squares fit when applied to each supplement separately. The fits are quite close.
Because the no-interaction model fares so well against the
interaction model, it may be instructive to examine the fit of the
no-interaction model. We sample from the no-interaction model with the
posterior
function:
chainsNoInt <- posterior(noInteraction, iterations = 10000)
# summary of the "interesting" parameters
summary(chainsNoInt[,1:5])
##
## Iterations = 1:10000
## Thinning interval = 1
## Number of chains = 1
## Sample size per chain = 10000
##
## 1. Empirical mean and standard deviation for each variable,
## plus standard error of the mean:
##
## Mean SD Naive SE Time-series SE
## mu 18.81 0.508 0.00508 0.00508
## supp-OJ 1.67 0.511 0.00511 0.00568
## supp-VC -1.67 0.511 0.00511 0.00568
## dose-dose 7.66 0.631 0.00631 0.00631
## sig2 15.67 3.060 0.03060 0.03395
##
## 2. Quantiles for each variable:
##
## 2.5% 25% 50% 75% 97.5%
## mu 17.830 18.47 18.81 19.15 19.812
## supp-OJ 0.692 1.32 1.67 2.01 2.692
## supp-VC -2.692 -2.01 -1.67 -1.32 -0.692
## dose-dose 6.445 7.23 7.65 8.07 8.921
## sig2 10.832 13.50 15.26 17.41 22.690
The right panel of the figure above shows the fit of the no-interaction model to the data. This model appears to account for the data satisfactorily. Though the moderate p value of the classical result might lead us to reject the no-interaction model, the Bayes factor and the visual fit appear to agree that the evidence is equivocal at best.
We have now analyzed the ToothGrowth
data using both
ANOVA (with dose
as a factor) and regression (with
dose
as a continuous covariate). We may wish to compare the
two approaches. We first create a column of the data with
dose
as a factor, then use anovaBF
:
ToothGrowth$doseAsFactor <- factor(ToothGrowth$dose)
levels(ToothGrowth$doseAsFactor) <- c(.5,1,2)
aovBFs <- anovaBF(len ~ doseAsFactor + supp + doseAsFactor:supp, data = ToothGrowth)
Because all models we’ve considered are compared to the null
intercept-only model, we can concatenate the aovBFs
object
with the Bayes factors we previously computed in this section:
allBFs <- c(aovBFs, full, noInteraction, onlyDose)
## eliminate the supp-only model, since it performs so badly
allBFs <- allBFs[-1]
## Compare to best model
allBFs / max(allBFs)
## Bayes factor analysis
## --------------
## [1] doseAsFactor : 0.00315 ±1.08%
## [2] supp + doseAsFactor : 0.182 ±2.51%
## [3] supp + doseAsFactor + supp:doseAsFactor : 0.488 ±2.02%
## [4] supp + dose + supp:dose : 1 ±0%
## [5] supp + dose : 0.939 ±1.72%
## [6] dose : 0.0175 ±1.08%
##
## Against denominator:
## len ~ supp + dose + supp:dose
## ---
## Bayes factor type: BFlinearModel, JZS
Two of the models score essentially equally well in terms of Bayes
factors: supp + dose + supp:dose
and
supp + dose
, suggesting that the interaction adds little.
The Bayes factors where dose is treated as a factor are all worse than
when dose is treated as a continuous covariate. This is likely due to a
the added flexibility allowed by including more parameters. Plotting the
Bayes factors shows how large the differences are:
Ly, Verhagen, and Wagenmakers (2015; link)
present a Bayes factor test for linear correlation. The
BayesFactor
package allows the computing of the Bayes
factor and sampling from the posterior of the Bayes factor. Note that
the model and priors are somewhat different from those used in the
linear regression models presented above; further discussion can be
found in Ly et al.
We demonstrate the use of the correlationBF
function
using Fisher’s iris
data set built into R
. See
the help (?iris
in R) for more details. We will focus on
the correlation between Sepal.Length
and
Sepal.Width
.
First, we create a scatterplot.
plot(Sepal.Width ~ Sepal.Length, data = iris)
abline(lm(Sepal.Width ~ Sepal.Length, data = iris), col = "red")
There does not appear to be a substantial correlation between these
two variables. We can compute a classical test of the correlation using
R
’s cor.test
function:
##
## Pearson's product-moment correlation
##
## data: iris$Sepal.Width and iris$Sepal.Length
## t = -1, df = 148, p-value = 0.2
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
## -0.2727 0.0435
## sample estimates:
## cor
## -0.118
The \(p\) value is nonsignificant at typical \(\alpha\) levels, and the point estimate is not terribly impressive at -0.12.
To compute the corresponding Bayes factor test, we use the
correlationBF
function (note the default prior scale).
## Bayes factor analysis
## --------------
## [1] Alt., r=0.333 : 0.509 ±0%
##
## Against denominator:
## Null, rho = 0
## ---
## Bayes factor type: BFcorrelation, Jeffreys-beta*
As would be expected from the middling \(p\) value in the classical test, the Bayes factor test shows little evidence either way (about 2 in favor of the null).
If we’d like to estimate the correlation on the assumption that it is
non-zero, we can sample from the posterior distribution using the
posterior
function.
## Independent-candidate M-H acceptance rate: 96%
The important parameter is rho
, the estimate of the true
linear correlation.
##
## Iterations = 1:10000
## Thinning interval = 1
## Number of chains = 1
## Sample size per chain = 10000
##
## 1. Empirical mean and standard deviation for each variable,
## plus standard error of the mean:
##
## Mean SD Naive SE Time-series SE
## rho -0.111 0.0789 0.000789 0.000824
## zeta -0.112 0.0803 0.000803 0.000839
##
## 2. Quantiles for each variable:
##
## 2.5% 25% 50% 75% 97.5%
## rho -0.260 -0.165 -0.111 -0.0576 0.0454
## zeta -0.266 -0.166 -0.112 -0.0577 0.0454
The posterior mean and credible interval for rho
are
very close to the point estimate and confidence interval obtained from
cor.test
.
We can also plot the full posterior distribution, if we like:
The default test for a proportion assumes that all observations were independent with fixed probability \(\pi\). The rule for stopping can be fixed \(N\) (binomial sampling) or a fixed number of successes (negative binomial sampling); unlike a significance test, the Bayes factor does not depend on the stopping rule.
For the Bayes factor test of a single proportion, there are two hypotheses; the null hypothesis assumes that the probability \(\pi\) is a fixed, known value \(p\); under the alternative, the log-odds corresponding to \(\pi\), denoted \(\omega = \log(\pi/(1-\pi))\), has a logistic distribution centered on the log-odds corresponding to the null value \(p\) (denoted \(\omega_0 = \log(p/(1-p))\): \[ \omega \sim \mbox{logistic}(\mbox{mean}=\omega_0, \mbox{scale}=r) \] The default prior \(r\) scale is 1/2. The figure below shows the prior distribution assuming the null hypothesis \(p=0.5\), for the three named prior scale settings \(r\) (“medium”, “wide”, and “ultrawide”). The default is “medium”:
The following example is taken from ?binom.test
, which
cites Conover (1971).
Under (the assumption of) simple Mendelian inheritance, a cross between plants of two particular genotypes produces progeny 1/4 of which are “dwarf” and 3/4 of which are “giant”, respectively. In an experiment to determine if this assumption is reasonable, a cross results in progeny having 243 dwarf and 682 giant plants. If “giant” is taken as success, the null hypothesis is that \(p = 3/4\) and the alternative that \(p \neq 3/4\).
## Bayes factor analysis
## --------------
## [1] Null, p=0.75 : 7.27 ±0.01%
##
## Against denominator:
## Alternative, p0 = 0.75, r = 0.5, p =/= p0
## ---
## Bayes factor type: BFproportion, logistic
The Bayes factor favors the null hypothesis by a factor of about 7 (which is not surprising given that the observed proportion is 73.7%). In contrast, the best we can say about the classical result is that it is not statistically “significant”:
##
## Exact binomial test
##
## data: 682 and 682 + 243
## number of successes = 682, number of trials = 925, p-value = 0.4
## alternative hypothesis: true probability of success is not equal to 0.75
## 95 percent confidence interval:
## 0.708 0.765
## sample estimates:
## probability of success
## 0.737
Using the posterior
function, we can draw samples from
the posterior distribution of the true log odds and true probability and
plot the estimate of the posterior.
## Independent-candidate M-H acceptance rate: 97%
The BayesFactor
package implements versions of Gunel and Dickey’s (1974) contingency table
Bayes factor tests. Bayes factors for contingency tests are computed
using the contingencyTableBF
function. The necessary
arguments are a matrix of cell frequencies and details about the
sampling plan that produced the data.
Here, we provide an example analysis of Hraba
and Grant’s (1970) data, included as part of the
BayesFactor
package as the raceDolls
data set.
71 white children and 89 black children from Lincoln, Nebraska were
offered two dolls, one of whose “race” was the same as the child’s and
one that was different (either white or black). The children were then
asked to select one of the dolls, with prompts such as “Give me the doll
that is a nice doll.” 50 of the 71 white children (70%) selected the
white doll, while 48 of the 89 black children (54%) selected the black
doll. These data are shown in the table below:
White child | Black child | |
---|---|---|
Same-race doll | 50 | 48 |
Different-race doll | 21 | 41 |
We can perform a Bayes factor analysis using the
contingencyTableBF
function:
## Bayes factor analysis
## --------------
## [1] Non-indep. (a=1) : 1.81 ±0%
##
## Against denominator:
## Null, independence, a = 1
## ---
## Bayes factor type: BFcontingencyTable, independent multinomial
Here we used sampleType="indepMulti"
and
fixedMargin="cols"
to specify that the columns are assumed
to be sampled as independent multinomials with their total fixed. See
the help at ?contingencyTableBF
for more details about
possible sampling plans and the priors.
The Bayes factor in favor of the alternative that the factors are not independent is just shy of 2, which is not very much evidence against the null hypothesis. For comparison, consider the results classical chi-square test, with continuity correction:
##
## Pearson's Chi-squared test with Yates' continuity correction
##
## data: raceDolls
## X-squared = 4, df = 1, p-value = 0.05
The classical test is just barely statistically significant.
We can also use the posterior
function to estimate the
difference in probabilities of selecing a doll of the same race between
white and black children, assuming the non-independence alternative:
For the independent multinomial sampling plan, the chains will contain the individual cell probabilities and the marginal column probabilities. We first need to compute the conditional probabilities from the results:
sameRaceGivenWhite = chains[,"pi[1,1]"] / chains[,"pi[*,1]"]
sameRaceGivenBlack = chains[,"pi[1,2]"] / chains[,"pi[*,2]"]
…and then plot the MCMC estimate of the difference:
plot(mcmc(sameRaceGivenWhite - sameRaceGivenBlack), main = "Increase in probability of child picking\nsame race doll (white - black)")
For more information, see ?contingencyTableBF
.
Additional tips and tricks (0.9.4+) ——— In this section, tricks to help save time and memory are described. These tricks work with version BayesFactor version 0.9.4+, unless otherwise indicated.
The convienience functions anovaBF
and
regressionBF
are specifically designed for cetagorical and
continuous covariates respectively, and have limitations that make those
functions easier to use. For instance, anovaBF
cannot
incorporate continuous covariates, and treats random effects as untested
nuissance parameters. The regressionBF
on the other hand,
being strictly for multiple regression, cannot incorporate categorical
covariates. These functions exist for particular purposes, since
guessing what model comparisons a user wants in general is difficult.
The lmBF
function, on the other hand, can handle any model
but is limited to a single model comparison: the specified model against
the intercept-only model.
The generalTestBF
function allows the testing of groups
of models (like anovaBF
and regressionBF
) but
can handle any kind of model (like lmBF
). Users specify a
full model, and generalTestBF
successively removes terms
from that model and tests the resulting submodels. For example, using
the puzzles
data set described above:
data(puzzles)
puzzleGenBF <- generalTestBF(RT ~ shape + color + shape:color + ID, data=puzzles, whichRandom="ID")
puzzleGenBF
## Bayes factor analysis
## --------------
## [1] shape : 0.611 ±0.01%
## [2] color : 0.611 ±0.01%
## [3] shape + color : 0.382 ±0.83%
## [4] shape + color + shape:color : 0.134 ±2.18%
## [5] ID : 111517 ±0%
## [6] shape + ID : 318608 ±1.12%
## [7] color + ID : 312963 ±0.93%
## [8] shape + color + ID : 1320210 ±2.27%
## [9] shape + color + shape:color + ID : 509046 ±7.21%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
The resulting 9 models are the full model, plus the models that can
be built by removing a single term at a time from the full model. By
default, the generalTestBF
function will not eliminate a
term that is involved in a higher-order interaction (for instance, we
will not remove shape
unless the shape:color
interaction is also removed); this behavior can be modified through the
whichModels
argument.
It is often the case that some terms are nuisance terms that we would
like to always keep in the model. For instance, ID
in the
puzzles
data set is a participant effect; we would not
generally consider models without a participant effect to be plausible.
We can use the neverExclude
argument to the function to
specify a set of search terms (technically, extended
regular expressions) that, if matched, will specify that the term is
always to be kept, and never excluded. To keep the ID
term:
puzzleGenBF <- generalTestBF(RT ~ shape + color + shape:color + ID, data=puzzles, whichRandom="ID", neverExclude="ID")
puzzleGenBF
## Bayes factor analysis
## --------------
## [1] shape + ID : 318624 ±1.25%
## [2] color + ID : 315164 ±1.09%
## [3] shape + color + ID : 1287049 ±2.7%
## [4] shape + color + shape:color + ID : 465008 ±1.99%
## [5] ID : 111517 ±0%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
The function now only considers models that contain ID
.
In some cases — especially when variable names are short, or a term to
be kept is part of an interaction term that can be eliminated — we need
to be careful in specifying search terms using
neverExclude
. For instance, suppose we are interested in
testing the ID:shape
interaction
puzzleGenBF <- generalTestBF(RT ~ shape + color + shape:color + shape:ID + ID, data=puzzles, whichRandom="ID", neverExclude="ID")
puzzleGenBF
## Bayes factor analysis
## --------------
## [1] color + shape + ID + shape:ID : 174461 ±2.88%
## [2] color + color:shape + shape + ID + shape:ID : 63932 ±2.28%
## [3] shape + ID + shape:ID : 28761 ±1.09%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
The shape:ID
interaction is never eliminated, because it
matches the ID
search term from neverExclude
.
Regular
expressions are useful here. There are special characters
representing the beginning and ending of a string (^
and
$
, respectively) that we can use to construct a regular
expression that will match ID
but not
shape:ID
:
puzzleGenBF <- generalTestBF(RT ~ shape + color + shape:color + shape:ID + ID, data=puzzles, whichRandom="ID", neverExclude="^ID$")
puzzleGenBF
## Bayes factor analysis
## --------------
## [1] shape + ID : 310892 ±0.73%
## [2] color + ID : 313151 ±0.98%
## [3] shape + color + ID : 1610764 ±19.8%
## [4] shape + color + shape:color + ID : 454583 ±1.56%
## [5] shape + shape:ID + ID : 28852 ±0.95%
## [6] shape + color + shape:ID + ID : 169313 ±1.25%
## [7] shape + color + shape:color + shape:ID + ID : 63199 ±2.38%
## [8] ID : 111517 ±0%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
The shape:ID
interaction term is now eliminated in some
models, because it does not match "^ID$"
. Multiple terms
may be provided to neverExclude
by providing a character
vector; terms which match any element in the vector will always be
included in model comparisons.
In cases where the default analysis produces many models to compare,
the sampling approach to computing Bayes factors can be time consuming.
The BayesFactor
package identifies situations where
sampling is not needed and thus saves time, but any model in which there
is more than one categorical factor or a mix of categorical and
continuous predictors will require sampling. When a default analysis
produces many models that are not of interest, much of the time spent
sampling may be wasted.
The main functions in the BayesFactor
package include
the noSample
argument which, if true, will prevent
sampling. If a Bayes factor can be computed without sampling, the
package will compute it, returning NA
for Bayes factors
that would require sampling. Continuing using the puzzles
dataset:
puzzleCullBF <- generalTestBF(RT ~ shape + color + shape:color + ID, data=puzzles, whichRandom="ID", noSample=TRUE,whichModels='all')
puzzleCullBF
## Bayes factor analysis
## --------------
## [1] shape : 0.611 ±0.01%
## [2] color : 0.611 ±0.01%
## [3] ID : 111517 ±0%
## [4] shape:color : 0.287 ±0.01%
## [5] shape + color : NA ±NA%
## [6] shape + ID : NA ±NA%
## [7] shape + shape:color : NA ±NA%
## [8] color + ID : NA ±NA%
## [9] color + shape:color : NA ±NA%
## [10] ID + shape:color : NA ±NA%
## [11] shape + color + ID : NA ±NA%
## [12] shape + color + shape:color : NA ±NA%
## [13] shape + ID + shape:color : NA ±NA%
## [14] color + ID + shape:color : NA ±NA%
## [15] shape + color + ID + shape:color : NA ±NA%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
Here we use whichModels='all'
for demonstration, in
order to obtain more possible model comparisons. Notice that several of
the Bayes factors were computable without sampling, and are reported.
The others have missing values, because the Bayes factor would have
required sampling to compute.
For now, we can separate the missing and non-missing Bayes factors in
separate variables. This is made easy by the is.na
method
for BayesFactor objects:
## Bayes factor analysis
## --------------
## [1] shape + color : NA ±NA%
## [2] shape + ID : NA ±NA%
## [3] shape + shape:color : NA ±NA%
## [4] color + ID : NA ±NA%
## [5] color + shape:color : NA ±NA%
## [6] ID + shape:color : NA ±NA%
## [7] shape + color + ID : NA ±NA%
## [8] shape + color + shape:color : NA ±NA%
## [9] shape + ID + shape:color : NA ±NA%
## [10] color + ID + shape:color : NA ±NA%
## [11] shape + color + ID + shape:color : NA ±NA%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
The variable missing
now contains all models for which
we lack a Bayes factor. At this point, we decide which of the Bayes
factors we would like to compute. We can do this in any way we like: we
could simple specify a subset, like missing[1:3]
or we
could do something more complicated. Here, we will include based on the
model formula, using the R function grepl
(?grepl).
Suppose we only wanted models that did not include both
shape
and color
. First, we obtain the names of
the models in missing
, and then test the names to see if
they match our restriction with grepl
. We can use the
result to restrict the models to compare to only those of interest.
# get the names of the numerator models
missingModels = names(missing)$numerator
# search them to see if they contain "shape" or "color" -
# results are logical vectors
containsShape = grepl("shape",missingModels)
containsColor = grepl("color",missingModels)
# anything that does not contain "shape" and "color"
containsOnlyOne = !(containsShape & containsColor)
# restrict missing to only those of interest
missingOfInterest = missing[containsOnlyOne]
missingOfInterest
## Bayes factor analysis
## --------------
## [1] shape + ID : NA ±NA%
## [2] color + ID : NA ±NA%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
We have restricted our set down to 2 items from 11 items. We can now
use recompute
to compute the missing Bayes factors:
# recompute the Bayes factors for the missing models of interest
sampledBayesFactors = recompute(missingOfInterest)
sampledBayesFactors
## Bayes factor analysis
## --------------
## [1] shape + ID : 317471 ±0.95%
## [2] color + ID : 316296 ±2.53%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
# Add them together with our other Bayes factors, already computed:
completeBayesFactors = c(done, sampledBayesFactors)
completeBayesFactors
## Bayes factor analysis
## --------------
## [1] shape : 0.611 ±0.01%
## [2] color : 0.611 ±0.01%
## [3] ID : 111517 ±0%
## [4] shape:color : 0.287 ±0.01%
## [5] shape + ID : 317471 ±0.95%
## [6] color + ID : 316296 ±2.53%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
Note that we’re still left with one model that contains both
shape
and color
, because it was computed
without sampling. Assuming that we were not interested in any model
containing both shape
and color
, however, we
may have saved considerable time by not sampling to estimate their Bayes
factors.
The noSample
argument will also work with the sampling
of posteriors. This is especially useful, for instance, if one would
like to know what order the MCMC chain results will be output in before
sampling.
Modern computer systems, which have many gigabytes of RAM, contain
sufficient memory to perform analyses of moderate scale using the
BayesFactor
package. Some systems — particularly older
32-bit systems — are limited in the amount of memory that can address.
Posterior sampling can create output that is hundreds of megabytes in
size. If a user conducts several of these analyses, R may not have
sufficient memory to store the results.
Consider, for instance, an analysis with 100 participants, 100 items, and two fixed effects with 3 levels each. We include all main effects in the model, as well as all two-way interactions (excluding the participant by item interaction). This results in 619 parameters. Because each number stored in an MCMC chain uses 8 bytes of memory, each iterations of the chain uses 8*619=4952 bytes. If a user then requests a 100,000 iteration MCMC chain — a large, but not unreasonably, sized MCMC chain — the resulting object will use about 500Mb of memory. This is most of the memory available to the default installation of R on a 32-bit Windows system. Even if a computer has a lot of memory, many of the parameters may not be interesting to the analyst. The participant and item effects, for instance, may be nuisance variation. If a user is not interested in the estimates, it is a waste of memory to include them in the MCMC chain.
The BayesFactor
package includes several methods for
reducing the size of MCMC chains: column filtering and chain thinning.
Column filtering ensures that certain parameters do not appear in the
output; thinning reduces the length of MCMC chains by only keeping some
of the iterations.
Consider again the puzzles
data set. We begin by
sampling from the MCMC chain of the model with the main effect of
shape
and color
, along with their interaction,
plus a participant effect:
data(puzzles)
# Get MCMC chains corresponding to "full" model
# We prevent sampling so we can see the parameter names
# iterations argument is necessary, but not used
fullModel = lmBF(RT ~ shape + color + shape:color + ID, data = puzzles, noSample=TRUE, posterior = TRUE, iterations=3)
fullModel
## Object of class "mcmc"
## Markov Chain Monte Carlo (MCMC) output:
## Start = 1
## End = 2
## Thinning interval = 1
## mu shape-round shape-square color-color color-monochromatic ID-1 ID-2 ID-3
## [1,] NA NA NA NA NA NA NA NA
## [2,] NA NA NA NA NA NA NA NA
## ID-4 ID-5 ID-6 ID-7 ID-8 ID-9 ID-10 ID-11 ID-12 shape:color-round.&.color
## [1,] NA NA NA NA NA NA NA NA NA NA
## [2,] NA NA NA NA NA NA NA NA NA NA
## shape:color-round.&.monochromatic shape:color-square.&.color
## [1,] NA NA
## [2,] NA NA
## shape:color-square.&.monochromatic sig2 g_shape g_color g_ID g_shape:color
## [1,] NA NA NA NA NA NA
## [2,] NA NA NA NA NA NA
## ---
## Model:
## Type: BFlinearModel, JZS
## RT ~ shape + color + shape:color + ID
## Data types:
## ID : fixed
## shape : fixed
## color : fixed
Notice that the participant effects, which are often regarded as
nuisance, are included in the chain. These parameters double the size of
the MCMC object; if we are not interested in the parameter values, we
could eliminate them from the output for a considerable savings. This
does not mean, however, that the parameters are not estimated; they will
still be used by BayesFactor
, but will not be reported.
To do this, we pass the columnFilter
argument to the
sampler, which surpresses output of any columns that arise from a term
matched by an element in columnFilter.
fullModelFiltered = lmBF(RT ~ shape + color + shape:color + ID, data = puzzles, noSample=TRUE, posterior = TRUE, iterations=3,columnFilter="ID")
fullModelFiltered
## Object of class "mcmc"
## Markov Chain Monte Carlo (MCMC) output:
## Start = 1
## End = 2
## Thinning interval = 1
## mu shape-round shape-square color-color color-monochromatic
## [1,] NA NA NA NA NA
## [2,] NA NA NA NA NA
## shape:color-round.&.color shape:color-round.&.monochromatic
## [1,] NA NA
## [2,] NA NA
## shape:color-square.&.color shape:color-square.&.monochromatic sig2 g_shape
## [1,] NA NA NA NA
## [2,] NA NA NA NA
## g_color g_ID g_shape:color
## [1,] NA NA NA
## [2,] NA NA NA
## ---
## Model:
## Type: BFlinearModel, JZS
## RT ~ shape + color + shape:color + ID
## Data types:
## ID : fixed
## shape : fixed
## color : fixed
Like the neverExclude
argument discussed above, the columnFilter
argument
is a character vector of extended
regular expressions. If a model term is matched by a search term in
columnFilter
, then all columns for term are eliminated from
the MCMC output. Remember that "ID"
will match anything
containing letters ID
; it would, for instance, also
eliminate terms GID
and ID:shape
, if they
existed. See the manual section on
generalTestBF
for details about how to use specific
regular expressions to avoid eliminating columns by accident.
MCMC chains are characterized by the fact that successive iterations are correlated with one another: that is, they are not indepenedent samples from the posterior distribution. To see this, we sample from the posterior of the full model and plot the results:
# Sample 10000 iterations, eliminating ID columns
chains = lmBF(RT ~ shape + color + shape:color + ID, data = puzzles, posterior = TRUE, iterations=10000,columnFilter="ID")
The figure below shows the first 1000 iterations of the MCMC chain for a selected parameter (left), and the autocorrelation function [CRAN / Wikipedia] for the same parameter (right).
The autocorrelation here is minimal, which will be the case in
general for chains from the BayesFactor
package. If we
wanted to reduce it even further, we might consider thinning
the chain: that is, keeping only every \(k\) iterations. Thinning throws away
information, and is generally not necessary or recommended; however, if
memory is at a premium, we might prefer storing nearly independent
samples to storing somewhat dependent samples.
The autocorrelation plot shows that the autocorrelation is reduced to
0 after 2 iterations. To get nearly independent samples, then, we could
thin to every \(k=2\) iterations using
the thin
argument:
chainsThinned = recompute(chains, iterations=20000, thin=2)
# check size of MCMC chain
dim(chainsThinned)
## [1] 10000 26
Notice that we are left with 10,000 iterations, instead of the 20,000 we sampled, because half were thinned. The figure below shows the resulting MCMC chain and autocorrelation functions. The MCMC chain does not visually look very different, because the autocorrelation was minimal in the first place. However, the autocorrelation function no longer shows autocorrelation from one iteration to the next, implying that we have obtained 10,000 nearly independent samples.
Previous to version 0.9.12-2
, it was only possible to
change the priors on a per-effect-type basis; ie, fixed effects all had
the same prior scale, random effects had a different prior scale, and
slopes had third prior scale. As of 0.9.12-2
, it is
possible to change the prior on a per-effect basis for fixed and random
effects (slopes still share a common prior, due to the use of the Liang
et al. hyper-g priors for the slopes). This is accomplished via the
rscaleEffects
argument to lmBF
,
anovaBF
, and generalTestBF
.
The rscaleEffects
argument is a named vector. The names
correspond to the effect you’d for which you’d like to set the prior,
and the value is the prior scale value. Any settings in
rscaleEffects
will override the settings in
rscaleFixed
and rscaleRandom
; if no settings
are found in rscaleEffects
, then the settings in
rscaleFixed
and rscaleRandom
are used.
We can demonstrate using the puzzles
data set. Suppose
we prefer a prior on the color
main effect of \(r=1\), a prior twice as wide as the default
in rscaleFixed
, \(r=.5\).
We set the prior scale for color
using the
rscaleEffects
argument:
newprior.bf = anovaBF(RT ~ shape + color + shape:color + ID, data = puzzles,
whichRandom = "ID",rscaleEffects = c( color = 1 ))
newprior.bf
## Bayes factor analysis
## --------------
## [1] shape + ID : 2.8 ±0.75%
## [2] color + ID : 2.14 ±1.74%
## [3] shape + color + ID : 8.97 ±1.66%
## [4] shape + color + shape:color + ID : 3.36 ±2.18%
##
## Against denominator:
## RT ~ ID
## ---
## Bayes factor type: BFlinearModel, JZS
The other fixed effects, shape
and
shape:color
, retain the prior scale of \(r=.5\) from rscaleFixed
.
Compare these Bayes factors to the ones with in the mixed modeling section above.
Conover, W. J. (1971), Practical nonparametric statistics. New York: John Wiley & Sons. Pages 97–104.
Gunel, E. and Dickey, J. (1974) Bayes Factors for Independence in Contingency Tables. Biometrika, 61, 545-557. (JSTOR)
Hraba, J. and Grant, G. (1970). Black is Beautiful: A reexamination of racial preference and identification. Journal of Personality and Social Psychology, 16, 398-402. psychnet.apa.org
Liang, F. and Paulo, R. and Molina, G. and Clyde, M. A. and Berger, J. O. (2008). Mixtures of g-priors for Bayesian Variable Selection. Journal of the American Statistical Association, 103, pp. 410-423 (Publisher)
Morey, R. D. and Rouder, J. N. (2011). Bayes Factor Approaches for Testing Interval Null Hypotheses. Psychological Methods, 16, pp. 406-419 (Publisher)
Morey, R. D. and Rouder, J. N. and Pratte, M. S. and Speckman, P. L. (2011). Using MCMC chain outputs to efficiently estimate Bayes factors. Journal of Mathematical Psychology, 55, pp. 368-378 (Publisher)
Rouder, J. N. and Morey, R. D. (2013) Default Bayes Factors for Model Selection in Regression, Multivariate Behavioral Research, 47, pp. 877-903 (Publisher)
Rouder, J. N. and Morey, R. D. and Speckman, P. L. and Province, J. M. (2012), Default Bayes Factors for ANOVA Designs. Journal of Mathematical Psychology, 56, pp. 356–374 (Publisher)
Rouder, J. N. and Speckman, P. L. and Sun, D. and Morey, R. D. and Iverson, G. (2009). Bayesian t-tests for accepting and rejecting the null hypothesis. Psychonomic Bulletin and Review, 16, pp. 225-237 (Publisher)
Rouder, J. N. and Morey, R. D. (2011). A Bayes Factor Meta-Analysis of Bem’s ESP Claim. Psychonomic Bulletin & Review 18, pp. 682-689 (Publisher)
Ly, A., Verhagen, A. J. & Wagenmakers, E.-J. (2015). Harold Jeffreys’s Default Bayes Factor Hypothesis Tests: Explanation, Extension, and Application in Psychology. Journal of Mathematical Psychology (Publisher)
This document was compiled with version 0.9.12-4.7 of BayesFactor (R version 4.3.2 (2023-10-31) on aarch64-apple-darwin20).