1 Mar, 2016

More on Hypothesis Testing


BIC = $-2 \log(L(\hat\theta)) + p\log(n)$ AIC = $-2 \log(L(\hat\theta)) + 2p$

We want BIC,AIC to be small.

Compare $M_1,M_2$.

\[S = -.5[BIC(M_1)-BIC(M_2)]\]

We can show that $-2S = -2\log(BF(M_1,M_2))$.

Bayes Factors: Computation

  • Using Laplace approx
  • Using posterior samples

Laplace Approximation

\[\begin{aligned} p(x|H_i)&=\int p_i(x|\theta_i)\pi_i(\theta_i|H_i)d\theta_i &= \int \exp\p{-n\bc{-\frac{1}{n}\log p_i(x|\theta_,H_i) -\frac{1}{n} \log\pi_i(\theta_i|H_i) }} d\theta_i, \end{aligned}\]

where $H_i$ can be a model.

\[BF(H_0,H_1) \approx (2n\pi)^{(p_0-p_1)/2} \frac{p_0(x|\hat\theta_0)\pi(\hat\theta_0)} {p_1(x|\hat\theta_1)\pi(\hat\theta_1)} \sqrt{\frac{\Sigma_0}{\Sigma_1}}\] \[\begin{aligned} p(x|H_i)&=\int p_i(x|\theta_i)\pi_i(\theta_i|H_i)d\theta_i\\ \theta_i^{(1)},...,\theta_i^{(M)} &\sim \pi_i(\theta_i|H_i)\\ \hat p (x|H_i) &= \sum_{i=1}^M p_i(x|\theta_i^{(m)},H_i) / M \end{aligned}\]

If we have posterior samples, and dim($\theta_1$) = dim($\theta_0$) \(\begin{aligned} BF(M_0,M_1) &= \sum_{i=1}^M \frac{p_0(x|\theta^{(m)},M_0)\pi_0(\theta^{(m)})}{p_1(x|\theta^{(m)},M_1)\pi_1(\theta^{(m)})} / M \end{aligned}\)

Case II: Nested (dimension of parameters not the same)

Example: \(\begin{array}{rcl} M0: & y_i&= \beta_0 + \beta_1 x_i + \epsilon_i\\ M1: & y_i&= \beta_0^* + \beta_1^* x_i + \beta_2^* z_i+ \epsilon_i \end{array}\)

Int this case it is also possible to obtain an approximation to the BF using samples from the posterior distribution of $M_1$. The approximation will have the form

\[\sum_{m=1}^M r(\theta^{(m)})\]

Exercise: find $r(\theta^{(m)})$

Hierarchical Models

Putting priors on parameters.