Index: The Book of Statistical ProofsModel SelectionBayesian model selectionBayes factor ▷ Computation using encompassing prior method

Theorem: Consider two models $m_1$ and $m_e$, where $m_1$ is nested within an encompassing model $m_e$ via an inequality constraint on some parameter $\theta$, and $\theta$ is unconstrained under $m_e$. Then, the Bayes factor is

\[\label{eq:bf-ep} \text{BF}_{1e} = \frac{c}{d} = \frac{1/d}{1/c}\]

where $1/d$ and $1/c$ represent the proportions of the posterior and prior of the encompassing model, respectively, that are in agreement with the inequality constraint imposed by the nested model $m_1$.

Proof: Consider first that for any model $m_1$ on data $y$ with parameter $\theta$, Bayes’ theorem implies

\[\label{eq:bayesth} p(\theta \mid y,m_1) = \frac{p(y \mid \theta,m_1) \cdot p(\theta \mid m_1)}{p(y \mid m_1)}.\]

Rearranging equation \eqref{eq:bayesth} allows us to write the marginal likelihood for $y$ under $m_1$ as

\[\label{eq:marginal} p(y \mid m_1) = \frac{p(y \mid \theta,m_1) \cdot p(\theta \mid m_1)}{p(\theta \mid y,m_1)}.\]

Taking the ratio of the marginal likelihoods for $m_1$ and the encompassing model $m_e$ yields the following Bayes factor:

\[\label{eq:bayesfactor} \text{BF}_{1e} = \frac{p(y \mid \theta,m_1) \cdot p(\theta \mid m_1) / p(\theta \mid y,m_1)}{p(y \mid \theta,m_e) \cdot p(\theta \mid m_e) / p(\theta \mid y,m_e)}.\]

Now, both the constrained model $m_1$ and the encompassing model $m_e$ contain the same parameter vector $\theta$. Choose a specific value of $\theta$, say $\theta’$, that exists in the support of both models $m_1$ and $m_e$ (we can do this, because $m_1$ is nested within $m_e$). Then, for this parameter value $\theta’$, we have $p(y \mid \theta’,m_1)=p(y \mid \theta’,m_e)$, so the expression for the Bayes factor in equation \eqref{eq:bayesfactor} reduces to an expression involving only the priors and posteriors for $\theta’$ under $m_1$ and $m_e$:

\[\label{eq:bayesfactor2} \text{BF}_{1e} = \frac{p(\theta' \mid m_1) / p(\theta' \mid y,m_1)}{p(\theta' \mid m_e) / p(\theta' \mid y,m_e)}.\]

Because $m_1$ is nested within $m_e$ via an inequality constraint, the prior $p(\theta’ \mid m_1)$ is simply a truncation of the encompassing prior $p(\theta’ \mid m_e)$. Thus, we can express $p(\theta’ \mid m_1)$ in terms of the encompassing prior $p(\theta’ \mid m_e)$ by multiplying the encompassing prior by an indicator function over $m_1$ and then normalizing the resulting product. That is,

\[\label{eq:normalize} \begin{split} p(\theta' \mid m_1) & = \frac{p(\theta' \mid m_e) \cdot I_{\theta' \in m_1}}{\int p(\theta' \mid m_e) \cdot I_{\theta' \in m_1} \, \mathrm{d}\theta'}\\ & = \Biggl(\frac{I_{\theta' \in m_1}}{\int p(\theta' \mid m_e) \cdot I_{\theta' \in m_1} \, \mathrm{d}\theta'}\Biggr) \cdot p(\theta' \mid m_e), \end{split}\]

where $I_{\theta’ \in m_1}$ is an indicator function. For parameters $\theta’ \in m_1$, this indicator function is identically equal to 1, so the expression in parentheses reduces to a constant, say $c$, allowing us to write the prior as

\[\label{eq:prior} p(\theta' \mid m_1) = c \cdot p(\theta' \mid m_e).\]

By similar reasoning, we can write the posterior as

\[\label{eq:posterior} p(\theta' \mid y,m_1) = \Biggl(\frac{I_{\theta' \in m_1}}{\int p(\theta' \mid y,m_e) \cdot I_{\theta' \in m_1} \, \mathrm{d}\theta'}\Biggr)\cdot p(\theta' \mid y,m_e) = d \cdot p(\theta' \mid y,m_e).\]

Plugging \eqref{eq:prior} and \eqref{eq:posterior} into \eqref{eq:bayesfactor2}, this gives us

\[\label{eq:bayesfactor3} \text{BF}_{1e} = \frac{c \cdot p(\theta' \mid m_e) / d \cdot p(\theta' \mid y,m_e)}{p(\theta' \mid m_e) / p(\theta' \mid y,m_e)} = \frac{c}{d} = \frac{1/d}{1/c},\]

which completes the proof. Note that by definition, $1/d$ represents the proportion of the posterior distribution for $\theta$ under the encompassing model $m_e$ that agrees with the constraints imposed by $m_1$. Similarly, $1/c$ represents the proportion of the prior distribution for $\theta$ under the encompassing model $m_e$ that agrees with the constraints imposed by $m_1$.

Sources:

Metadata: ID: P157 | shortcut: bf-ep | author: tomfaulkenberry | date: 2020-09-02, 12:00.