https://github.com/cran/bayestestR
Raw File
Tip revision: 601edcd8d1306ebea478657861c54fa9c69a52f3 authored by Dominique Makowski on 31 May 2021, 05:40 UTC
version 0.10.0
Tip revision: 601edcd
bayes_factors.Rmd
---
title: "Bayes Factors"
output: 
  rmarkdown::html_vignette:
    toc: true
    toc_depth: 2
    fig_width: 10.08
    fig_height: 6
tags: [r, bayesian, bayes factors]
vignette: >
  \usepackage[utf8]{inputenc}
  %\VignetteIndexEntry{Bayes Factors}
  %\VignetteEngine{knitr::rmarkdown}
editor_options: 
  chunk_output_type: console
bibliography: bibliography.bib
csl: apa.csl
---

This vignette can be referred to by citing the following:

- Makowski, D., Ben-Shachar, M. S., \& Lüdecke, D. (2019). *bayestestR: Describing Effects and their Uncertainty, Existence and Significance within the Bayesian Framework*. Journal of Open Source Software, 4(40), 1541. https://doi.org/10.21105/joss.01541

- Makowski, D., Ben-Shachar, M. S., Chen, S. H. A., \& Lüdecke, D. (2019). *Indices of Effect Existence and Significance in the Bayesian Framework*. Retrieved from [10.3389/fpsyg.2019.02767](https://doi.org/10.3389/fpsyg.2019.02767)

---

```{r setup, include=FALSE}
library(knitr)
options(knitr.kable.NA = "", digits = 2)
knitr::opts_chunk$set(
  echo = TRUE,
  comment = ">",
  out.width = "100%",
  message = FALSE,
  warning = FALSE,
  dpi = 150
)

pkgs <- c(
  "rstanarm", "BayesFactor", "emmeans", "logspline", "lme4", "ggplot2",
  "see", "insight", "emmeans", "knitr", "effectsize", "bayestestR"
)
if (!all(sapply(pkgs, require, quietly = TRUE, character.only = TRUE))) {
  knitr::opts_chunk$set(eval = FALSE)
}

set.seed(4)

if (require("ggplot2") && require("see")) {
  theme_set(theme_modern())
}
```

The adoption of the Bayesian framework for applied statistics, especially in the
social and psychological sciences, seems to be developing in two distinct
directions. One of the key topics marking their separation is their opinion
about the **Bayes factor**. In short, one school of thought (e.g., the *Amsterdam school*, led by [E. J. Wagenmakers](https://www.bayesianspectacles.org/))
advocate its use, and emphasize its qualities as a statistical index, while
another point to its limits and prefer, instead, the precise description of
posterior distributions (using
[CIs](https://easystats.github.io/bayestestR/reference/hdi.html),
[ROPEs](https://easystats.github.io/bayestestR/reference/rope.html), etc.).

The `bayestestR` package does **not** take a side in this debate, and offers tools to carry out analysis irrespective of the school you subscribe to. Instead, it strongly supports the notion of an *informed choice*:

**discover the methods, learn about them, understand them, try them, and decide for yourself**.

Having said that, here's an introduction to Bayes factors :)


# The Bayes Factor

**Bayes Factors (BFs) are indices of *relative* evidence of one "model" over another**. 

In their role as a hypothesis testing index, they are to Bayesian framework what
a $p$-value is to the **classical/frequentist framework**. In significance-based
testing, $p$-values are used to assess how unlikely are the observed data if the
**null hypothesis** were true, while in the **Bayesian model selection framework**,
Bayes factors assess evidence for different models, each model corresponding to
a specific hypothesis.

According to Bayes' theorem, we can update prior probabilities of some model $M$
($P(M)$) to posterior probabilities ($P(M|D)$) after observing some datum $D$ by
accounting for the probability of observing that datum given the model
($P(D|M)$, also known as the *likelihood*):

$$
P(M|D) = \frac{P(D|M)\times P(M)}{P(D)}
$$

Using this equation, we can compare the probability-odds of two models:

$$
\underbrace{\frac{P(M_1|D)}{P(M_2|D)}}_{\text{Posterior Odds}} = 
\underbrace{\frac{P(D|M_1)}{P(D|M_2)}}_{\text{Likelihood Ratio}} 
\times
\underbrace{\frac{P(M_1)}{P(M_2)}}_{\text{Prior Odds}}
$$

Where the *likelihood ratio* (the middle term) is the *Bayes factor* - it is the ***factor*** by which some **prior odds** have been updated after observing the data to **posterior odds**.

Thus, Bayes factors can be calculated in two ways:

- As a ratio quantifying **the relative probability of the observed data under each of the two models**. (In some contexts, these probabilities are also called *marginal likelihoods*.)

$$
BF_{12}=\frac{P(D|M_1)}{P(D|M_2)}
$$

- As **the degree of shift in prior beliefs** about the relative credibility of
two models (since they can be computed by dividing posterior odds by prior
odds).

$$
BF_{12}=\frac{Posterior~Odds_{12}}{Prior~Odds_{12}}
$$

Here we provide functions for computing Bayes factors in two different contexts: 

- **testing single parameters (coefficients) within a model** 
- **comparing statistical models themselves**

# Testing Models' Parameters with Bayes Factors {#bayesfactor_parameters}

A **Bayes factor for a single parameter** can be used to answer the question:

> "Given the observed data, has the null hypothesis of an absence of an effect
become more or less credible?"


```{r deathsticks_fig, echo=FALSE, fig.cap="Bayesian analysis of the Students' (1908) Sleep data set.", fig.align='center', out.width="80%"}
knitr::include_graphics("https://github.com/easystats/easystats/raw/master/man/figures/bayestestR/deathsticks.jpg")
```

Let's use the Students' (1908) Sleep data set (`data("sleep")`). The data comes
from a study in which participants were administered a drug and the researchers
assessed the extra hours of sleep that participants slept afterwards. We will
try answering the following research question using Bayes factors:

> **Given the observed data, has the hypothesis that the drug (the effect of `group`) has no effect on the numbers of hours of extra sleep (variable `extra`) become more of less credible?**

```{r sleep_boxplot, echo=FALSE}
library(ggplot2)

ggplot(sleep, aes(x = group, y = extra, fill = group)) +
  geom_boxplot() +
  theme_classic() +
  theme(legend.position = "none")
```

The **boxplot** suggests that the second group has a higher number of hours of
extra sleep. *By how much?*

Let's fit a simple [Bayesian linear model](https://easystats.github.io/bayestestR/articles/example1.html), with a
prior of $b_{group} \sim N(0, 3)$ (i.e. the prior follows a Gaussian/normal
distribution with $mean = 0$ and $SD = 3$), using
`rstanarm` package:

```{r rstanarm_model, eval = FALSE}
set.seed(123)
library(rstanarm)

model <- stan_glm(
  formula = extra ~ group,
  data = sleep,
  prior = normal(0, 3, autoscale = FALSE)
)
```

```{r, echo=FALSE}
model <- stan_glm(
  formula = extra ~ group,
  data = sleep,
  prior = normal(0, 3, autoscale = FALSE),
  refresh = 0
)
```

### Testing against a null-*region*

One way of operationlizing the null-hypothesis is by setting a null region, such
that an effect that falls within this interval would be *practically* equivalent
to the null [@kruschke2010believe]. In our case, that means defining a range of
effects we would consider equal to the drug having no effect at all. We can then
compute the prior probability of the drug's effect falling *within this null-region*, 
and the prior probability of the drug's effect falling *outside the null-region*
to get our *prior odds*. Say any effect smaller than an hour of extra sleep is
practically equivalent to being no effect at all, we would define our prior odds
as:

$$
\frac
{P(b_{drug} \in [-1, 1])}
{P(b_{drug} \notin [-1, 1])}
$$

Given our prior has a normal distribution centered at 0 hours with a scale (an
SD) of 3 hours, our priors would look like this:

```{r, echo=FALSE}
null <- c(-1, 1)
xrange <- c(-10, 10)

x_vals <- seq(xrange[1], xrange[2], length.out = 400)
d_vals <- dnorm(x_vals, sd = 3)
in_null <- null[1] < x_vals & x_vals < null[2]
range_groups <- rep(0, length(x_vals))
range_groups[!in_null & x_vals < 0] <- -1
range_groups[!in_null & x_vals > 0] <- 1

ggplot(mapping = aes(x_vals, d_vals, fill = in_null, group = range_groups)) +
  geom_area(color = "black", size = 1) +
  scale_fill_flat(name = "", labels = c("Alternative", "Null")) +
  labs(x = "Drug effect", y = "Density") +
  coord_cartesian(ylim = c(0, 0.45)) +
  theme_modern() +
  theme(legend.position = c(0.2, 0.8))

pnull <- diff(pnorm(null, sd = 2.5))
prior_odds <- (1 - pnull) / pnull
```

and the prior odds would be 2.2.

By looking at the posterior distribution, can now compute the posterior probability of the drug's effect falling *within the null-region*, and the posterior probability of the drug's effect falling *outside the null-region* to get our *posterior odds*:

$$
\frac
{P(b_{drug} \in [-1,1] | Data)}
{P(b_{drug} \notin [-1,1] | Data)}
$$

```{r rstanarm_fit, echo=FALSE}
library(bayestestR)
model_prior <- unupdate(model)
posterior <- insight::get_parameters(model)$group2
prior <- insight::get_parameters(model_prior)$group2

f_post <- logspline::logspline(posterior)

d_vals_post <- logspline::dlogspline(x_vals, f_post)

ggplot(mapping = aes(x_vals, d_vals_post, fill = in_null, group = range_groups)) +
  geom_area(color = "black", size = 1) +
  scale_fill_flat(name = "", labels = c("Alternative", "Null")) +
  labs(x = "Drug effect", y = "Density") +
  coord_cartesian(ylim = c(0, 0.45)) +
  theme_modern() +
  theme(legend.position = c(0.2, 0.8))

My_first_BF <- bayesfactor_parameters(model, null = c(-1, 1), prior = model_prior)

BF <- My_first_BF$BF[2]
post_odds <- prior_odds * BF

med_post <- point_estimate(posterior)$Median
```

We can see that the center of the posterior distribution has shifted away from 0
(to ~1.5). Likewise, the posterior odds are 2, which seems to favor **the effect being non-null**. **But**, does this mean the data support the alternative over
the null? Hard to say, since even before the data were observed, the priors
already favored the alternative - so we need to take our priors into account
here!

Let's compute the Bayes factor as the change from the prior odds to the
posterior odds: $BF_{10} = Odds_{posterior} / Odds_{prior} = 0.9$! This BF
indicates that the data provide 1/0.9 = 1.1 times more evidence for the effect
of the drug being practically nothing than it does for the drug having some
clinically significant effect. Thus, although the center of distribution has
shifted away from 0, and the posterior distribution seems to favor a non-null
effect of the drug, it seems that given the observed data, the probability mass
has *overall* shifted closer to the null interval, making the values in the null
interval more probable! [see *Non-overlapping Hypotheses* in
@morey2011bayesinterval]

All of this can be achieved with the function `bayesfactor_parameters()`, which
computes a Bayes factor for each of the model's parameters:

```{r, eval=FALSE}
My_first_BF <- bayesfactor_parameters(model, null = c(-1, 1))
My_first_BF
```

```{r, echo=FALSE}
print(My_first_BF)
```

We can also plot using the `see` package:

```{r}
library(see)
plot(My_first_BF)
```

Note that **interpretation guides** for Bayes factors can be found in the `effectsize` package:

```{r}
effectsize::interpret_bf(exp(My_first_BF$log_BF[2]), include_value = TRUE)
```


### Testing against the *point*-null (0)

> **What if we don't know what region would be practically equivalent to 0?**

Or if we just want the null to be exactly zero? Not a problem - as the width of
null region shrinks to a point, the change from the prior probability to the
posterior probability of the null can be estimated by comparing the density of
the null value between the two distributions.^[Note that as the width of null
interval shrinks to zero, the prior probability and posterior probability of the
alternative tends towards 1.00.] This ratio is called the **Savage-Dickey ratio**, 
and has the added benefit of also being an approximation of a Bayes factor
comparing the estimated model against the a model in which the parameter of
interest has been restricted to a point-null:

> "[...] the Bayes factor for $H_0$ versus $H_1$ could be obtained by
analytically integrating out the model parameter $\theta$. However, the Bayes
factor may likewise be obtained by only considering $H_1$, and dividing the
height of the posterior for $\theta$ by the height of the prior for $\theta$, at
the point of interest." [@wagenmakers2010bayesian]

```{r, eval=FALSE}
My_second_BF <- bayesfactor_parameters(model, null = 0)
My_second_BF
```

```{r, echo=FALSE}
My_second_BF <- bayesfactor_parameters(
  data.frame(group2 = posterior),
  data.frame(group2 = prior),
  null = 0
)

print(My_second_BF)
```

```{r}
plot(My_second_BF)
```

### Directional hypotheses

We can also compute Bayes factors for directional hypotheses ("one sided"), if we have a prior hypotheses about the direction of the effect. This can be done by setting an *order restriction* on the prior distribution (which results in an order restriction on the posterior distribution) of the alternative [@morey2014simple]. For example, if we have a prior hypothesis that *the drug has a positive effect on the number of sleep hours*, the alternative will be restricted to the region to the right of the null (point or interval):

```{r savagedickey_one_sided, eval=FALSE}
test_group2_right <- bayesfactor_parameters(model, direction = ">")
test_group2_right
```

```{r prior_n_post_plot_one_sided, echo=FALSE}
test_group2_right <- bayesfactor_parameters(
  data.frame(group2 = posterior),
  data.frame(group2 = prior),
  null = 0,
  direction = ">"
)

BF <- test_group2_right$BF

print(test_group2_right)
```

```{r}
plot(test_group2_right)
```

As we can see, given that we have an *a priori* assumption about the direction of the effect (that the effect is positive), **the presence of an effect is 2.8 times more likely than the absence of an effect**, given the observed data (or that the data are 2.8 time more probable under $H_1$ than $H_0$). This indicates that, given the observed data, and a priori hypothesis, the posterior mass has shifted away from the null value, giving some evidence against the null (note that a Bayes factor of 2.8 is still considered quite [weak evidence](https://easystats.github.io/effectsize/reference/interpret_bf.html)).

Thanks to the flexibility of Bayesian framework, it is also possible to compute
a Bayes factor for **dividing** hypotheses - that is, for a null and alternative
that are *complementary*, opposing one-sided hypotheses [@morey2014simple].

For example, above we compared an alternative of $H_A$: *the drug has a positive effects* to the null $H_0$: *the drug has no effect*. But we can also compare instead the same alternative to its *complementary* hypothesis: $H_{-A}$: *the drug has a negative effects*. 

```{r inteval_div, eval=FALSE}
test_group2_dividing <- bayesfactor_parameters(model, null = c(-Inf, 0))
test_group2_dividing
```

```{r inteval_div2, echo=FALSE}
test_group2_dividing <- bayesfactor_parameters(
  data.frame(group2 = posterior),
  data.frame(group2 = prior),
  null = c(-Inf, 0)
)

print(test_group2_dividing)
```

```{r}
plot(test_group2_dividing)
```

We can see that this test produces even stronger (more conclusive) evidence than the one-sided vs. point-null test! And indeed, as a rule of thumb, the more specific the two hypotheses are, and the more distinct they are from one another, the more *power* our Bayes factor has! ^[For more, see [this talk by Richard D. Morey, minute 48](https://philstatwars.files.wordpress.com/2020/09/richard_presentation.mp4)]

Thanks to the transitivity of Bayes factors, we can also use
`bayesfactor_parameters()` to compare even more types of hypotheses, with some
trickery. For example:

$$
\underbrace{BF_{0<b<1\text{ vs. }b=0}}_{\text{range vs. point}}
= 
\underbrace{BF_{b<0\text{ vs. }b=0}}_{\text{directional vs. point}} 
/
\underbrace{BF_{b<0\text{ vs. }0<b<1}}_{\text{directional vs. range}} 
$$

**NOTE**: See the *Testing Contrasts* appendix below.

### Support intervals {#si}

So far we've seen that Bayes factors quantify relative support between competing hypotheses. However, we can also ask:

> **Upon observing the data, the credibility of which of the parameter’s values has increased (or decreased)?**

For example, we've seen that the point null has become somewhat less credible
after observing the data, but we might also ask which values have **gained**
credibility given the observed data?. The resulting range of values is called
**the support interval** as it indicates which values are supported by the data
[@wagenmakers2018SI]. We can do this by once again comparing the prior and
posterior distributions and checking where the posterior densities are higher
than the prior densities.

In `bayestestR`, this can be achieved with the `si()` function:

```{r}
my_first_si <- si(
  posterior = data.frame(group2 = posterior),
  prior = data.frame(group2 = prior),
  BF = 1
)

print(my_first_si)
```

The argument `BF = 1` indicates that we want the interval to contain values that
have gained support by a factor of at least 1 (that is, any support at all).

Visually, we can see that the credibility of all the values within this interval
has increased (and likewise the credibility of all the values outside this
interval has decreased):

```{r}
plot(my_first_si)
```

We can also see the this support interval (just barely) excludes the point null
(0) - whose credibility we've already seen has decreased by the observed data.
This emphasizes the relationship between the support interval and the Bayes
factor:

> "The interpretation of such intervals would be analogous to how a frequentist
confidence interval contains all the parameter values that would not have been
rejected if tested at level $\alpha$. For instance, a BF = 1/3 support interval
encloses all values of theta for which the updating factor is not stronger than
3 against." [@wagenmakers2018SI]

Thus, the choice of BF (the level of support the interval should indicate)
depends on what we want our interval to represent:

- A $BF = 1$ contains values whose credibility has merely not decreased by
observing the data.
- A $BF > 1$ contains values who received more impressive support from the data.  
- A $BF < 1$ contains values whose credibility has *not* been impressively
decreased by observing the data. Testing against values outside this interval
will produce a Bayes factor larger than $1/BF$ in support of the alternative.

# Comparing Models using Bayes Factors {#bayesfactor_models}

Bayes factors can also be used to compare statistical **models**. In this
statistical context, they answer the following question:

> **Under which model are the observed data more probable?**

In other words, which model is more likely to have produced the observed data?
This is usually done by comparing the marginal likelihoods of two models. In
such a case, the Bayes factor is a measure of the **relative** evidence for one
model over the other.

Let's use Bayes factors for model comparison to find a model that best describes
the length of an iris' sepal using the `iris` data set.

### For Bayesian models (`brms` and `rstanarm`)

**Note: In order to compute Bayes factors for Bayesian models, non-default arguments must be added upon fitting:**
  
  - `brmsfit` models **must** have been fitted with `save_pars = save_pars(all = TRUE)`
  - `stanreg` models **must** have been fitted with a defined `diagnostic_file`.
  
Let's first fit 5 Bayesian regressions with `brms` to predict `Sepal.Length`:

```{r brms_disp, eval = FALSE}
library(brms)

# intercept only model
m0 <- brm(Sepal.Length ~ 1, data = iris, 
          save_pars = save_pars(all = TRUE),  refresh = 0, backend = "rstan")

# main effects models
m1 <- brm(Sepal.Length ~ Petal.Length, data = iris, 
          save_pars = save_pars(all = TRUE), refresh = 0, backend = "rstan")

m2 <- brm(Sepal.Length ~ Species, data = iris,
          save_pars = save_pars(all = TRUE), refresh = 0, backend = "rstan")

# additive main effects model
m3 <- brm(Sepal.Length ~ Species + Petal.Length, data = iris,
          save_pars = save_pars(all = TRUE), refresh = 0, backend = "rstan")

# full model
m4 <- brm(Sepal.Length ~ Species * Petal.Length, data = iris,
          save_pars = save_pars(all = TRUE), refresh = 0, backend = "rstan")
```

We can now compare these models with the `bayesfactor_models()` function, using
the `denominator` argument to specify the model against which the rest of the
models will be compared (in this case, the intercept-only model):

```{r brms_models_disp, eval = FALSE}
library(bayestestR)

comparison <- bayesfactor_models(m1, m2, m3, m4, denominator = m0)
comparison
```

```{r, echo = FALSE}
comparison <- structure(
  list(Model = c("Petal.Length", "Species", "Species + Petal.Length", "Species * Petal.Length", "1"), 
       log_BF = c(102.600193981361, 68.5488807613052, 128.665504913167, 128.919384882682, 0)),
  class = c("bayesfactor_models", "see_bayesfactor_models", "data.frame"), 
  row.names = c("m1", "m2", "m3", "m4", "m0"),
  denominator = 5L,
  BF_method = "marginal likelihoods (bridgesampling)",
  unsupported_models = FALSE)
print(comparison)
```


We can see that the full model is the best model - with $BF_{\text{m0}}=9\times 10^{55}$ compared to the null (intercept only).

Due to the transitive property of Bayes factors, we can easily change the
reference model to the main effects model:

```{r update_models1}
update(comparison, reference = 3)
```

As we can see, though the full model is the best, there is hardly any evidence
that it is preferable to the main effects model.

We can also change the reference model to the `Species` model:

```{r update_models2}
update(comparison, reference = 2)
```

Notice that, in the Bayesian framework the compared models *do not* need to be
nested models, as happened here when we compared the `Petal.Length`-only model
to the `Species`-only model (something that cannot be done in the frequentist
framework, where compared models must be nested in one another).

We can also get a matrix of Bayes factors of all the pairwise model comparisons:

```{r}
as.matrix(comparison)
```

**NOTE:** In order to correctly and precisely estimate Bayes Factors, you always need the 4 P's: **P**roper **P**riors ^[[Robert, 2016](https://doi.org/10.1016/j.jmp.2015.08.002); [Kass & Raftery, 1993](https://doi.org/10.1080/01621459.1995.10476572); [Fernández, Ley, & Steel, 2001](https://doi.org/10.1016/S0304-4076(00)00076-2)], and a **P**lentiful **P**osterior ^[[Gronau, Singmann, & Wagenmakers, 2017](https://arxiv.org/abs/1710.08162)].

### For Frequentist models via the BIC approximation 

It is also possible to compute Bayes factors for the comparison of frequentist
models. This is done by comparing BIC measures, allowing a Bayesian comparison
of non-nested frequentist models [@wagenmakers2007practical].

Let's try it out on some **linear mixed-effects models**:

```{r lme4_models}
library(lme4)

# define models with increasing complexity
m0 <- lmer(Sepal.Length ~ (1 | Species), data = iris)
m1 <- lmer(Sepal.Length ~ Petal.Length + (1 | Species), data = iris)
m2 <- lmer(Sepal.Length ~ Petal.Length + (Petal.Length | Species), data = iris)
m3 <- lmer(Sepal.Length ~ Petal.Length + Petal.Width + (Petal.Length | Species), data = iris)
m4 <- lmer(Sepal.Length ~ Petal.Length * Petal.Width + (Petal.Length | Species), data = iris)

# model comparison
bayesfactor_models(m1, m2, m3, m4, denominator = m0)
```

### Order restricted models {#bayesfactor_restricted}

As stated above when discussing one-sided hypothesis tests, we can create new
models by imposing order restrictions on a given model. For example, consider
the following model, in which we predict the length of an iris' sepal from the
length of its petal, as well as from its species, with priors:
- $b_{petal} \sim N(0,2)$ 
- $b_{versicolors}\ \&\  b_{virginica} \sim N(0,1.2)$

```{r}
iris_model <- stan_glm(Sepal.Length ~ Species + Petal.Length,
  data = iris,
  prior = normal(0, c(2, 1.2, 1.2), autoscale = FALSE),
  refresh = 0
)
```

These priors are **unrestricted** - that is, **all values** between $-\infty$
and $\infty$ of all parameters in the model have some non-zero credibility (no
matter how small; this is true for both the prior and posterior distribution).
Subsequently, *a priori* the ordering of the parameters relating to the iris
species can have any ordering, such that *a priori* setosa can have larger
sepals than virginica, but it is also possible for virginica to have larger
sepals than setosa!

Does it make sense to let our priors cover all of these possibilities? That
depends on our *prior* knowledge or hypotheses. For example, even a novice
botanist will assume that it is unlikely that petal length will be *negatively*
associated with sepal length - an iris with longer petals is likely larger, and
thus will also have a longer sepal. And an expert botanist will perhaps assume
that setosas have smaller sepals than both versicolors and virginica. 

These priors can be formulated as **restricted** priors [@morey_2015_blog;
@morey2011bayesinterval]:

1. The novice botanist: $b_{petal} > 0$
2. The expert botanist: $b_{versicolors} > 0\ \&\ b_{virginica} > 0$

By testing these restrictions on prior and posterior samples, we can see how the
probabilities of the restricted distributions change after observing the data.
This can be achieved with `bayesfactor_restricted()`, that compute a Bayes
factor for these restricted model vs the unrestricted model. Let's first specify
these restrictions as logical conditions:

```{r}
botanist_hypotheses <- c(
  "Petal.Length > 0",
  "(Speciesversicolor > 0) & (Speciesvirginica > 0)"
)
```

Let's test these hypotheses:

```{r}
model_prior <- unupdate(iris_model)

botanist_BFs <- bayesfactor_restricted(
  posterior = iris_model,
  prior = model_prior,
  hypothesis = botanist_hypotheses
)

print(botanist_BFs)
```

We can see that the novice botanist's hypothesis gets a Bayes factor of ~2,
indicating the data provides twice as much evidence for a model in which petal
length is restricted to be positively associated with sepal length than for a
model with not such restriction.

What about our expert botanist? He seems to have failed miserably, with a BF
favoring the *unrestricted* model many many times over ($BF\gg1,000$). How is
this possible? It seems that when *controlling for petal length*, versicolor and
virginica actually have shorter sepals!

```{r plot_iris, echo=FALSE}
ggplot(iris, aes(Petal.Length, Sepal.Length, color = Species)) +
  geom_point() +
  scale_color_flat() +
  theme(legend.position = c(0.2, 0.8))
```

Note that these Bayes factors compare the restricted model to the unrestricted
model. If we wanted to compare the restricted model to the null model, we could
use the transitive property of Bayes factors like so:

$$
BF_{\text{restricted vs. NULL}} = \frac
{BF_{\text{restricted vs. un-restricted}}}
{BF_{\text{un-restricted vs NULL}}}
$$

**Because these restrictions are on the prior distribution, they are only appropriate for testing pre-planned (*a priori*) hypotheses, and should not be used for any post hoc comparisons [@morey_2015_blog].**

**NOTE**: See the *Specifying Correct Priors for Factors with More Than 2 Levels* appendix below.

# Bayesian Model Averaging

In the previous section, we discussed the direct comparison of two models to
determine if an effect is supported by the data. However, in many cases there
are too many models to consider, or perhaps it is not straightforward which
models we should compare to determine if an effect is supported by the data. For
such cases, we can use Bayesian model averaging (BMA) to determine the support
provided by the data for a parameter or term across many models.

### Inclusion Bayes factors {#bayesfactor_inclusion}

Inclusion Bayes factors answer the question:

> **Are the observed data more probable under models with a particular predictor, than they are under models without that particular predictor?**

In other words, on average, are models with predictor $X$ more likely to have
produced the observed data than models without predictor $X$?^[A model without
predictor $X$ can be thought of as a model in which the parameter(s) of the
predictor have been restricted to a null-point of 0.]

Since each model has a prior probability, it is possible to sum the prior
probability of all models that include a predictor of interest (the *prior inclusion probability*), and of all models that do not include that predictor
(the *prior exclusion probability*). After the data are observed, and each model
is assigned a posterior probability, we can similarly consider the sums of the
posterior models' probabilities to obtain the *posterior inclusion probability*
and the *posterior exclusion probability*. Once again, the change from prior
inclusion odds to the posterior inclusion odds is the **Inclusion Bayes factor**
["$BF_{Inclusion}$"; @clyde2011bayesian].

Lets use the `brms` example from above:

```{r inclusion_brms}
bayesfactor_inclusion(comparison)
```

If we examine the interaction term's inclusion Bayes factor, we can see that
across all 5 models, a model with the interaction term (`Species:Petal.Length`)
is *on average* 5 times more likely than a model without the interaction term.
Note that `Species`, a factor represented in the model with several parameters,
gets a *single* Bayes factor - inclusion Bayes factors are given **per predictor**!

We can also compare only matched models - such that averaging is done only
across models that (1) do not include any interactions with the predictor of
interest; (2) for interaction predictors, averaging is done only across models
that contain the main effects from which the interaction predictor is comprised
(see explanation for why you might want to do this
[here](https://www.cogsci.nl/blog/interpreting-bayesian-repeated-measures-in-jasp)).

```{r inclusion_brms2}
bayesfactor_inclusion(comparison, match_models = TRUE)
```

### Comparison with JASP

`bayesfactor_inclusion()` is meant to provide Bayes Factors per predictor,
similar to JASP's *Effects* option. 

Let's compare the two:

1. **Across all models**:

```{r JASP_all}
library(BayesFactor)
data(ToothGrowth)
ToothGrowth$dose <- as.factor(ToothGrowth$dose)

BF_ToothGrowth <- anovaBF(len ~ dose * supp, ToothGrowth, progress = FALSE)

bayesfactor_inclusion(BF_ToothGrowth)
```

```{r JASP_all_fig, echo=FALSE}
knitr::include_graphics("https://github.com/easystats/easystats/raw/master/man/figures/bayestestR/JASP1.PNG")
```

2. **Across matched models**:

```{r JASP_matched}
bayesfactor_inclusion(BF_ToothGrowth, match_models = TRUE)
```


```{r JASP_matched_fig, echo=FALSE}
knitr::include_graphics("https://github.com/easystats/easystats/raw/master/man/figures/bayestestR/JASP2.PNG")
```

3. **With Nuisance Effects**:

We'll add `dose` to the null model in JASP, and do the same in `R`:

```{r JASP_Nuisance}
BF_ToothGrowth_against_dose <- BF_ToothGrowth[3:4] / BF_ToothGrowth[2] # OR:
# update(bayesfactor_models(BF_ToothGrowth),
#        subset = c(4, 5),
#        reference = 3)
BF_ToothGrowth_against_dose


bayesfactor_inclusion(BF_ToothGrowth_against_dose)
```

```{r JASP_Nuisance_fig, echo=FALSE}
knitr::include_graphics("https://github.com/easystats/easystats/raw/master/man/figures/bayestestR/JASP3.PNG")
```

## Averaging posteriors {#weighted_posteriors}

Similar to how we can average evidence for a predictor across models, we can
also average the **posterior estimate** across models. This is useful in
situations where Bayes factors seem to support a null effect, yet the *HDI* for
the alternative excludes the null value (also see `si()` described above).

For example, looking at Motor *Trend Car Road Tests* (`data(mtcars)`), we would
naturally predict miles/gallon (`mpg`) from transition type (`am`) and weight
(`wt`), but what about number of carburetors (`carb`)? Is this a good predictor?

We can determine this by comparing the following models:

```{r}
mod <- stan_glm(mpg ~ wt + am,
  data = mtcars,
  prior = normal(0, c(10, 10), autoscale = FALSE),
  diagnostic_file = file.path(tempdir(), "df1.csv"),
  refresh = 0
)

mod_carb <- stan_glm(mpg ~ wt + am + carb,
  data = mtcars,
  prior = normal(0, c(10, 10, 20), autoscale = FALSE),
  diagnostic_file = file.path(tempdir(), "df0.csv"),
  refresh = 0
)

BF_carb <- bayesfactor_models(mod_carb, denominator = mod, verbose = FALSE)
BF <- BF_carb$BF[1]
print(BF_carb)
```

It seems that the model without `carb` as a predictor is $1/BF=1.2$ times more
likely than the model *with* `carb` as a predictor. We might then assume that in
the latter model, the `HDI` will include the point-null value of 0 effect, to also
indicate the credibility of the null in the posterior. However, this is not the
case:

```{r}
hdi(mod_carb, ci = .95)
```

How can this be? By estimating the HDI of the effect for `carb` in the full
model, we are acting under the assumption that this model is correct. However,
as we've just seen, both models are practically tied. If this is the case **why limit our estimation of the effect just to one model?** [@van2019cautionary].

Using Bayesian Model Averaging, we can combine the posteriors samples from
several models, weighted by the models' marginal likelihood (done via the
`bayesfactor_models()` function). If some parameter is part of some of the
models but is missing from others, it is assumed to be fixed a 0 (which can also
be seen as a method of applying shrinkage to our estimates). This results in a
posterior distribution across several models, which we can now treat like any
posterior distribution, and estimate the HDI. 

In `bayestestR`, we can do this with the `weighted_posteriors()` function:

```{r}
BMA_draws <- weighted_posteriors(mod, mod_carb)

BMA_hdi <- hdi(BMA_draws, ci = .95)
BMA_hdi

plot(BMA_hdi)
```

We can see that across both models under consideration, the posterior of the
`carb` effect is almost equally weighted between the alternative model and the
null model - as represented by about half of the posterior mass concentrated at
0 - which makes sense as both models were almost equally supported by the data.
We can also see that across both models, that now **the HDI does contain 0**.
Thus we have resolved the conflict between the Bayes factor and the HDI
[@rouder2018bayesian]!

**Note**: Parameters might play different roles across different models. 

For example, the parameter `A` plays a different role in the model `Y ~ A + B`
(where it is a *main* effect) than it does in the model `Y ~ A + B + A:B` (where
it is a *simple* effect). In many cases centering of predictors (mean subtracting
for continuous variables, and effects coding via `contr.sum` or orthonormal
coding via
[`contr.orthonorm`](https://easystats.github.io/bayestestR/reference/contr.orthonorm.html)
for factors) can in some cases reduce this issue.

# Appendices

## Testing contrasts (with `emmeans` / `modelbased`)

Besides testing parameter `bayesfactor_parameters()` can be used to test any
estimate based on the prior and posterior distribution of the estimate. One way
to achieve this is with a mix of `bayesfactor_parameters()` +
[**`emmeans`**](https://cran.r-project.org/package=emmeans) to [test Bayesian
contrasts](https://easystats.github.io/blog/posts/bayestestr_emmeans/).

For example, in the `sleep` example from above, we can estimate the group means
and the difference between them:

```{r, echo=FALSE}
set.seed(1)
```

```{r}
library(emmeans)

groups <- emmeans(model, ~group)
group_diff <- pairs(groups)

(groups_all <- rbind(groups, group_diff))

# pass the original model via prior
bayesfactor_parameters(groups_all, prior = model)
```

That is strong evidence for the mean of group 1 being 0, and for group 2 for not
being 0, but hardly any evidence for the difference between them being not 0.
Conflict? Uncertainty? That is the Bayesian way!

We can also use the `easystats`' [**`modelbased`**](https://cran.r-project.org/package=modelbased) package to compute Bayes factors for contrasts:

```{r, echo=FALSE}
set.seed(1)
```

```{r, eval=FALSE}
library(modelbased)

estimate_contrasts(model, test = "bf")
```

**NOTE**: See the *Specifying Correct Priors for Factors with More Than 2 Levels* section below.

## Specifying correct priors for factors {#contr_bayes}

This section introduces the biased priors obtained when using the common *effects* factor coding (`contr.sum`) or dummy factor coding (`contr.treatment`), and the solution of using orthonormal factor coding (`contr.orthonorm`) [as outlined in @rouder2012default, section 7.2]. 

**Special care should be taken when working with factors with 3 or more levels**.

### Contrasts (and marginal means)

The *effects* factor coding commonly used in factorial analysis carries a hidden
bias when it is applies to Bayesian priors. For example, if we want to test all
pairwise differences between 3 levels of the same factor, we would expect all *a priori* differences to have the same distribution, but...

For our example, we will be test all ***prior*** pairwise differences between
the 3 species in the `iris` dataset.

```{r}
df <- iris
contrasts(df$Species)[, ] <- contr.sum(3)

fit_sum <- stan_glm(Sepal.Length ~ Species,
  data = df,
  prior = normal(0, c(1, 1), autoscale = FALSE),
  prior_PD = TRUE, # sample priors
  family = gaussian(),
  refresh = 0
)

pairs_sum <- pairs(emmeans(fit_sum, ~Species))

em_pairs_samples <- as.data.frame(as.matrix(emmeans::as.mcmc.emmGrid(pairs_sum, names = FALSE)))

print(pairs_sum)

ggplot(stack(em_pairs_samples), aes(x = values, fill = ind)) +
  geom_density(size = 1) +
  facet_grid(ind ~ .) +
  labs(x = "prior difference values") +
  theme(legend.position = "none")
```

Notice that, though the prior estimate for all 3 pairwise contrasts is ~0, the
scale or the HDI is much narrower for the prior of the `setosa - versicolor`
contrast!

**What happened???** 

This is caused by an inherent bias in the priors introduced by the *effects* coding (it's even worse with the default treatment coding, because the prior for the intercept is usually drastically different from the effect's parameters). **And since it affects the priors, this bias will also bias the Bayes factors over / understating evidence for some contrasts over others!**

The solution is to use *orthonormal* factor coding, a-la `contr.orthonorm`, which can either specify this factor coding per-factor:

```{r, eval=FALSE}
contrasts(df$Species) <- contr.orthonorm
```

Or you can set it globally:

```{r, eval=FALSE}
options(contrasts = c("contr.orthonorm", "contr.poly"))
```

Let's again estimate the ***prior*** differences:

```{r}
contrasts(df$Species)[, ] <- contr.orthonorm(3)

fit_bayes <- stan_glm(Sepal.Length ~ Species,
  data = df,
  prior = normal(0, c(1, 1), autoscale = FALSE),
  prior_PD = TRUE, # sample priors
  family = gaussian(),
  refresh = 0
)

pairs_bayes <- pairs(emmeans(fit_bayes, ~Species))

em_pairs_samples <- as.data.frame(as.matrix(emmeans::as.mcmc.emmGrid(pairs_bayes, names = FALSE)))

print(pairs_bayes)

ggplot(stack(em_pairs_samples), aes(x = values, fill = ind)) +
  geom_density(size = 1) +
  facet_grid(ind ~ .) +
  labs(x = "prior difference values") +
  theme(legend.position = "none")
```


We can see that using this coding scheme, we have equal priors on all pairwise
contrasts.

### Order restrictions

This bias also affect order restrictions involving 3 or more levels. For
example, if we want to test an order restriction among A, B, and C, the *a
priori* probability of obtaining the order A > C > B is 1/6 (reach back to
*intro to stats* year 1), but...

For our example, we will be interested in the following order restrictions in
the `iris` dataset (each line is a separate restriction):

```{r}
hyp <- c(
  # comparing 2 levels
  "setosa < versicolor",
  "setosa < virginica",
  "versicolor < virginica",

  # comparing 3 (or more) levels
  "setosa    < virginica  & virginica  < versicolor",
  "virginica < setosa     & setosa     < versicolor",
  "setosa    < versicolor & versicolor < virginica"
)
```

With the default factor coding, this looks like this:

```{r, eval=FALSE}
contrasts(df$Species) <- contr.sum
fit_sum <- stan_glm(Sepal.Length ~ Species,
  data = df,
  prior = normal(0, c(1, 1), autoscale = FALSE),
  family = gaussian()
)

em_sum <- emmeans(fit_sum, ~Species) # the posterior marginal means

bayesfactor_restricted(em_sum, fit_sum, hypothesis = hyp)
```

```{r, echo=FALSE}
contrasts(df$Species)[, ] <- contr.sum(3)
fit_sum <- stan_glm(Sepal.Length ~ Species,
  data = df,
  prior = normal(0, c(1, 1), autoscale = FALSE),
  family = gaussian(),
  refresh = 0
)

em_sum <- emmeans(fit_sum, ~Species) # the posterior marginal means

bayesfactor_restricted(em_sum, fit_sum, hypothesis = hyp)
```

***What happened???***   

1. The comparison of 2 levels all have a prior of ~0.5, as expected.  
2. The comparison of 3 levels has different priors, depending on the order restriction - i.e. **some orders are *a priori* more likely than others!!!**  

Again, this is solved by using the *orthonormal* factor coding (from above).

```{r, eval=FALSE}
contrasts(df$Species) <- contr.orthonorm
fit_bayes <- stan_glm(Sepal.Length ~ Species,
  data = df,
  prior = normal(0, c(1, 1), autoscale = FALSE),
  family = gaussian()
)
em_bayes <- emmeans(fit_sum, ~Species) # the posterior marginal means
bayesfactor_restricted(em_bayes, fit_sum, hypothesis = hyp)
```

```{r, echo=FALSE}
contrasts(df$Species)[, ] <- contr.orthonorm(3)
fit_bayes <- stan_glm(Sepal.Length ~ Species,
  data = df,
  prior = normal(0, c(1, 1), autoscale = FALSE),
  family = gaussian(),
  refresh = 0
)
em_bayes <- emmeans(fit_bayes, ~Species) # the posterior marginal means
bayesfactor_restricted(em_bayes, fit_bayes, hypothesis = hyp)
```

### Conclusion

When comparing the results from the two factor coding schemes, we find:  
1. In both cases, the estimated (posterior) means are quite similar (if not identical).  
2. The priors and Bayes factors differ between the two schemes.  
3. Only with `contr.orthonorm`, the prior distribution of the difference or the order of 3 (or more) means is balanced.

# References
back to top