# Bayes factor

Short description: A statistical factor used to compare competing hypotheses

The Bayes factor is the ratio of two marginal likelihoods, that is the likelihoods of two statistical models integrated over the prior probabilities of their parameters. The aim of the Bayes factor is to quantify the support for one model over another and therefore assist in model selection. Oftentimes the models in question have a common set of parameters, but this is not necessary. By extension, the Bayes factor can be used for hypothesis testing in the Bayesian framework, and under simple hypotheses (e.g., two specific parameter values) is in fact equal to the likelihood-ratio test in frequentist statistics. In contrast with null hypothesis significance testing, however, Bayes factors support evaluation of evidence in favor of a null hypothesis, rather than only allowing the null to be rejected or not rejected.

Although conceptually simple, the computation of the Bayes factor can be challenging depending on the complexity of the model and the hypotheses. For certain special cases, simplified expressions are available; for instance, the Savage–Dickey density ratio in the case of a precise (equality constrained) hypothesis against an unrestricted alternative. Another approximation, derived by applying Laplace's method to the integrated likelihoods, is known as the Bayesian information criterion (BIC); in large data sets the Bayes factor will approach the BIC as the influence of the priors wanes. In small data sets, priors generally matter and may not be improper since the Bayes factor will be undefined if either of the two integrals in its ratio is not finite.

## Definition

The Bayes factor is a likelihood ratio of the marginal likelihood of two competing hypotheses, usually a null and an alternative.

The posterior probability $\displaystyle{ \Pr(M|D) }$ of a model M given data D is given by Bayes' theorem:

$\displaystyle{ \Pr(M|D) = \frac{\Pr(D|M)\Pr(M)}{\Pr(D)}. }$

The key data-dependent term $\displaystyle{ \Pr(D|M) }$ represents the probability that some data are produced under the assumption of the model M; evaluating it correctly is the key to Bayesian model comparison.

Given a model selection problem in which we have to choose between two models on the basis of observed data D, the plausibility of the two different models M1 and M2, parametrised by model parameter vectors $\displaystyle{ \theta_1 }$ and $\displaystyle{ \theta_2 }$, is assessed by the Bayes factor K given by

$\displaystyle{ K = \frac{\Pr(D|M_1)}{\Pr(D|M_2)} = \frac{\int \Pr(\theta_1|M_1)\Pr(D|\theta_1,M_1)\,d\theta_1} {\int \Pr(\theta_2|M_2)\Pr(D|\theta_2,M_2)\,d\theta_2} = \frac{\frac{\Pr(M_1|D)\Pr(D)}{\Pr(M_1)}}{\frac{\Pr(M_2|D)\Pr(D)}{\Pr(M_2)}} = \frac{\Pr(M_1|D)}{\Pr(M_2|D)}\frac{\Pr(M_2)}{\Pr(M_1)}. }$

When the two models have equal prior probability, so that $\displaystyle{ \Pr(M_1) = \Pr(M_2) }$, the Bayes factor is equal to the ratio of the posterior probabilities of M1 and M2. If instead of the Bayes factor integral, the likelihood corresponding to the maximum likelihood estimate of the parameter for each statistical model is used, then the test becomes a classical likelihood-ratio test. Unlike a likelihood-ratio test, this Bayesian model comparison does not depend on any single set of parameters, as it integrates over all parameters in each model (with respect to the respective priors). However, an advantage of the use of Bayes factors is that it automatically, and quite naturally, includes a penalty for including too much model structure. It thus guards against overfitting. For models where an explicit version of the likelihood is not available or too costly to evaluate numerically, approximate Bayesian computation can be used for model selection in a Bayesian framework, with the caveat that approximate-Bayesian estimates of Bayes factors are often biased.

Other approaches are:

## Interpretation

A value of K > 1 means that M1 is more strongly supported by the data under consideration than M2. Note that classical hypothesis testing gives one hypothesis (or model) preferred status (the 'null hypothesis'), and only considers evidence against it. Harold Jeffreys gave a scale for interpretation of K:

Template:Alternating rows table style="text-align: center; margin-left: auto; margin-right: auto; border: none;" ! K !! dHart !! bits !! Strength of evidence |- | < 100 || < 0 || < 0 || Negative (supports M2) |- | 100 to 101/2 || 0 to 5 || 0 to 1.6 || Barely worth mentioning |- | 101/2 to 101 || 5 to 10 || 1.6 to 3.3 || Substantial |- | 101 to 103/2 || 10 to 15 || 3.3 to 5.0 || Strong |- | 103/2 to 102 || 15 to 20 || 5.0 to 6.6 || Very strong |- | > 102 || > 20 || > 6.6 || Decisive |- |}

The second column gives the corresponding weights of evidence in decihartleys (also known as decibans); bits are added in the third column for clarity. According to I. J. Good a change in a weight of evidence of 1 deciban or 1/3 of a bit (i.e. a change in an odds ratio from evens to about 5:4) is about as finely as humans can reasonably perceive their degree of belief in a hypothesis in everyday use.

An alternative table, widely cited, is provided by Kass and Raftery (1995):

Template:Alternating rows table style="text-align: center; margin-left: auto; margin-right: auto; border: none;" ! log10 K !! K !! Strength of evidence |- | 0 to 1/2 || 1 to 3.2 || Not worth more than a bare mention |- | 1/2 to 1 || 3.2 to 10 || Substantial |- | 1 to 2 || 10 to 100 || Strong |- | > 2 || > 100 || Decisive |- |}

## Example

Suppose we have a random variable that produces either a success or a failure. We want to compare a model M1 where the probability of success is q = ​12, and another model M2 where q is unknown and we take a prior distribution for q that is uniform on [0,1]. We take a sample of 200, and find 115 successes and 85 failures. The likelihood can be calculated according to the binomial distribution:

$\displaystyle{ {{200 \choose 115}q^{115}(1-q)^{85}}. }$

Thus we have for M1

$\displaystyle{ P(X=115 \mid M_1)={200 \choose 115}\left({1 \over 2}\right)^{200} \approx 0.006 }$

whereas for M2 we have

$\displaystyle{ P(X=115 \mid M_2) = \int_{0}^1{200 \choose 115}q^{115}(1-q)^{85}dq = {1 \over 201} \approx 0.005 }$

The ratio is then 1.2, which is "barely worth mentioning" even if it points very slightly towards M1.

A frequentist hypothesis test of M1 (here considered as a null hypothesis) would have produced a very different result. Such a test says that M1 should be rejected at the 5% significance level, since the probability of getting 115 or more successes from a sample of 200 if q = ​12 is 0.02, and as a two-tailed test of getting a figure as extreme as or more extreme than 115 is 0.04. Note that 115 is more than two standard deviations away from 100. Thus, whereas a frequentist hypothesis test would yield significant results at the 5% significance level, the Bayes factor hardly considers this to be an extreme result. Note, however, that a non-uniform prior (for example one that reflects the fact that you expect the number of success and failures to be of the same order of magnitude) could result in a Bayes factor that is more in agreement with the frequentist hypothesis test.

A classical likelihood-ratio test would have found the maximum likelihood estimate for q, namely 115200 = 0.575, whence

$\displaystyle{ \textstyle P(X=115 \mid M_2) = {{200 \choose 115}q^{115}(1-q)^{85}} \approx 0.06 }$

(rather than averaging over all possible q). That gives a likelihood ratio of 0.1 and points towards M2.

M2 is a more complex model than M1 because it has a free parameter which allows it to model the data more closely. The ability of Bayes factors to take this into account is a reason why Bayesian inference has been put forward as a theoretical justification for and generalisation of Occam's razor, reducing Type I errors.

On the other hand, the modern method of relative likelihood takes into account the number of free parameters in the models, unlike the classical likelihood ratio. The relative likelihood method could be applied as follows. Model M1 has 0 parameters, and so its AIC value is 2·0 − 2·ln(0.005956) = 10.2467. Model M2 has 1 parameter, and so its AIC value is 2·1 − 2·ln(0.056991) = 7.7297. Hence M1 is about exp((7.7297 − 10.2467)/2) = 0.284 times as probable as M2 to minimize the information loss. Thus M2 is slightly preferred, but M1 cannot be excluded.