23 Feb, 2016

Hypothesis Testing


Classical Setting

$H_0: \theta \in \Theta_0$ vs. $H_1: \theta \in \Theta_1$

  $H_0$ is True $H_0$ is False
Reject $H_0$ Type I error Correct
Fail to Reject $H_0$ Correct Type II error

Decision Theoretic Approach

\[L(\theta,\delta) = \begin{cases} \delta, & \text{if } \theta \in \Theta_0 \\ 1-\delta,& \text{if } \theta \in \Theta_1 \end{cases}\]

for $\delta = {0,1}$.

Risk:

Frequentist Version

\[R(\theta,\delta) = \int~L(\theta,\delta(x)) f(x|\theta)~dx\] \[\begin{cases} P(\delta(x)=1) & \theta\in\Theta_0 \text{ Type I error} \\ P(\delta(x)=0) & \theta\in\Theta_1 \text{ Type II error} \end{cases}\]

Bayesian Version

\[\int ~ L(\theta,\delta(x)) \pi(\theta|x) ~ dx = \begin{cases} \pi(\theta\in\Theta_1|x), & \delta=0 ~ \text{ (Don't reject) }\\ \pi(\theta\in\Theta_0|x), & \delta=1 ~ \text{ (reject) } \end{cases}\]

Bayes Factors

The posterior odds ratio is given by

\[\frac{p(\Theta_0|x)}{p(\Theta_1|x)} = \frac{p(x|H_0)}{p(x|H_1)} \frac{p(H_0)}{p(H_1)}\]

The factor $B_{01} = \frac{p(x\v H_0)}{p(x \v H_1)}$ updates the prior odds ratio to the posterior odds ratio. This is known as the Bayes Factor in favr of $H_0$.

When $H_0$ and $H_1$ are both simple hypotheses, the BF corresponds to the likelihood ratio $B_{01}=\frac{p(x\v\theta_0)}{p(x\v\theta_1)}$. In general, if $g_i(\theta)$ is the prior for $\theta$ under $H_i$, then

\[B_{01} = \frac{\int_{\Theta_0}~p(x|\theta)g_0(\theta)~d\theta}{\int_{\Theta_1}~p(x|\theta)g_1(\theta)~d\theta} = \frac{m_0(x)}{m_1(x)}\]
$\log_{10} BF(H_1,H_0)$ $BF(H_1,H_0)$ Evidence against $H_0$
0-.5 1-3.2 Not work more that mentioning
.5-1 3.2-10 Substantial
1-2 10-100 strong
$> 2$ $> 100$ Decisive

Bayes factors cannot be reconciled with the p-values in two-sided tests.

$B_{ij}(x) = \frac{m_i(x)}{m_j(x)}$, $p(M_k|x) = \p{\sum_{i=1}^m \frac{p(M_i)}{p(M_k)}B_{jk}}^{-1}$, $m_k(x) = \int_{\Theta_k}~p_k(x\v\theta_k)p_k(\theta_k)~d\theta_k$

If the priors $p_k(\theta_k)$ are defined up to a proportionality constant $c_k$, then the BF will depend on the ratio of such constants. If the priors are proper densities, then we have

\[c_k^{-1} = \int_{\Theta_k} p_k(\theta_k)d\theta_k\]

and then the BF is uniquely defined. For improper priors, not defined.