--- title: "Lecture 10 - Key" output: pdf_document --- ```{r setup, include=FALSE} knitr::opts_chunk$set(echo = TRUE) library(dplyr) ``` ### Bayesian GLMs and Metropolis-Hastings Algorithm We have seen that with conjugate or semi-conjugate prior distributions the Gibbs sampler can be used to sample from the posterior distribution. In situations, such as Generalized Linear Models (GLMs), where conjugate prior distributions are not available the Gibbs sampler cannot be used. \vfill #### Poisson Regression Model Example. (Hoff p. 171) A sample was taken of 52 female sparrows to study their reproductive habits. The number of offspring and the age of the sparrow are recorded. The response (number of offspring) is a non-negative integer. A logical distribution for this data is the Poisson distribution. The conditional expection from a Poisson regression model, as a function of age, can be formulated as: \vfill \vfill \vfill Formallly, the Poisson regression model can be expressed as: \vfill In a generalized linear model there are three defining components: \begin{enumerate} \item \item \item \end{enumerate} \vfill Note it is important to consider the restriction of the Poisson distribution, namely that the mean and variance are the same. __Q:__ If this is not appropriate, what do we do? \vfill \newpage One option is to consider the negative binomial distribution. The natural parameterization of the negative binomial distribution has one parameter corresponding to the probability of a sucess and the other as the number of successes until stopping. However, it can be parameterized such that one parameter represents the mean of the distribution and the other as the variance of the distribution. \vfill Up until now, we haven't talked about anything inherently Bayesian. So what do we need to think about this model from a Bayesian perspective? \vfill \vfill \vfill \vfill \vfill Metropolis-Hastings, coming soon... \subsubsection*{Logistic Regression Model} Another common GLM is logistic regression, which is used for modeling binary outcomes. The logistic regression model uses the logistic link function where: \vfill \vfill \vfill \vfill Let's consider the structure of the posterior distribution as a function of the prior for $\tilde{\beta}.$ \begin{eqnarray*} p(\tilde{\beta}|X,y_i) &\propto & p(y_i|\tilde{x}_i,\tilde{\beta}) \times p(\tilde{\beta})\\ &\propto&\\ &&\\ &&\\ \end{eqnarray*} \vfill \vfill \newpage Note, often Bayesians will use the probit link that enables the use of a Gibbs sampler on a latent variable, where $p_i = \Phi(\tilde{\beta}^T \tilde{x}_i)$, where $\Phi(.)$ is the cumulative distribution function of a standard normal random variable. \vfill #### Metropolis Algorithm Before considering the sampling for the GLMs detailed above, first consider a generic case. Assume we have a sampling model $p(y|\theta)$ and a prior distribution $p(\theta).$ In general: \begin{equation*} p(\theta|y) = \frac{p(y|\theta) p(\theta)}{\int p(y|\theta) p(\theta) d\theta} \end{equation*} is hard to sample from due to the integration. However, given samples from $p(\theta|y)$ we can compute functions of the posterior distribution. \vfill The challenge is drawing samples from $p(\theta|y)$, so far we have done this using Monte Carlo and Markov Chain Monte Carlo procedures (Gibbs sampler). \vfill The goal is to create a sampler where the empirical distribution of the samples corresponds to the true samples. In otherwords, \vfill \vfill \vfill {\bf Q:} So how do we design a sampler to meet this requirement? \vfill \newpage Assume the current value of $\theta$ is denoted as $\theta^{(s)}$, suppose we propose a new value $\theta^*$. The question is should we include this in our samples, or in otherwords should we move from $\theta^{(s)}$ to $\theta^*$? \vfill Ideally we would evaluate the ratio of: \begin{equation*} r = \end{equation*} however in many cases computing this value is difficult due to the required integration. Fortunately, it is not necessary as: \begin{equation*} r = \left(\frac{p(y|\theta^*) p(\theta^*)}{p(y)}\right) \times \left(\frac{p(y)}{p(y|\theta^{(s)} p(\theta^{(s)}} \right) = \frac{p(y|\theta^*)p(\theta^*)}{p(y|\theta^{(s)}) p(\theta^{(s)})}. \end{equation*} \vfill Now the question is, what do we do when: - $r \geq 1$. \vfill - $r<1$. \vfill This intuitive algorithm that we have devised is known as the Metroplis Algorithm, which is a special case of the Metropolis-Hastings algorithm where the proposal distribution (to select $\theta^*$) is symmetric. \vfill \newpage Formally, the Metropolis algorithm follows as: 1. Sample $\theta^*|\theta^{(s)} \sim J(\theta^*|\theta^{(s)})$. Typically $J(.)$ is a random walk function such as $J(\theta^*|\theta^{(s)}) \sim N(\theta^{(s)}, \gamma^2)$, where $\gamma$ is thought of as the step size. The symmetric requirement means $J(\theta^*|\theta^{(s)}) = J(\theta^{(s)}|\theta^*)$, this restriction is not necessary in a slightly more complicated algorithm (Metropolis-Hastings). \vfill 2. Compute the acceptance ratio: \begin{equation*} r = \frac{p(\theta^*|y)}{p(\theta^{(s)})}= \frac{p(y|\theta^*)p(\theta^*)}{p(y|\theta^{(s)}) p(\theta^{(s)})}. \end{equation*} \vfill 3. Let \[ \theta^{(s+1)} = \begin{cases} \theta^* & \text{with probability min (r,1)} \\ \theta^{(s)} & \text{with probability 1 0 min (r,1)} \\ \end{cases} \] In practice to complete step 3, sample $u \sim Unif(0,1)$ and set $\theta^{(s+1)}=\theta^*$ if $u < r$. \vfill ##### Metropolis for Poisson - with R code Note: in many cases we need to consider $\log(r)$ as the likelihood calculations can easily end up being numerically zero. ```{r, eval= F} set.seed(11112018) library(mnormt) # rmnorm # Simulate Data n <- 1000 p <- 3 beta.true <- c(2,.5,.5) X <- matrix(c(rep(1,n),rnorm(n*2)),nrow=n,ncol=p) theta <- exp(X %*% beta.true) y <- rpois(n,theta) hist(y,breaks='FD') ``` ```{r, eval = F} # Run Metropolis Algorithm num.mcmc <- 10000 step.size <- .0001 accept.ratio <- rep(0,num.mcmc) beta.mcmc <- matrix(0,num.mcmc,p) beta.prior.var <- diag(p) * 100 # b ~n(0,100*I) for (i in 2:num.mcmc){ beta.star <- beta.mcmc[i-1,] + rmnorm(1,0,step.size * diag(p)) #compute r theta.current <- exp(X %*% beta.mcmc[i-1,]) theta.star <- exp(X %*% beta.star) log.p.current <- sum(dpois(y,theta.current,log=T)) + dmnorm(beta.mcmc[i-1,],0,beta.prior.var,log=T) log.p.star <- sum(dpois(y,theta.star,log=T)) + dmnorm(beta.star,0,beta.prior.var,log=T) log.r <- log.p.star - log.p.current if (log(runif(1)) < log.r){ beta.mcmc[i,] <- beta.star accept.ratio[i] <- 1 } else{ beta.mcmc[i,] <- beta.mcmc[i-1,] } } #mean(accept.ratio) #colMeans(beta.mcmc) plot(beta.mcmc[,1],type='l') abline(h=beta.true[1],lwd=2,col='gray') plot(beta.mcmc[,2],type='l') abline(h=beta.true[2],lwd=2,col='gray') plot(beta.mcmc[,3],type='l') abline(h=beta.true[3],lwd=2,col='gray') ``` Note the step size in as important consideration in a Metropolis-Hastings Algorithm. If the proposal is too large, the algorithm will tend to stay in one location for a large number of iterations as $\tilde{\beta}^*$ will be unattractive. If the step size is too small, virtually all of the proposals will be accepted, but the sampler will not efficiently explore the space. These can be seen visually and as a product of \vfill Consider three figures below for an example of what happens as this varies \begin{figure}[ht!] \begin{center} \includegraphics[width=\textwidth]{Metropolis.png} \caption{Trace plots for a step size that is too large, too small, and just right. } \end{center} \end{figure} \newpage ##### Metropolis-Hastings Example. Assume we observe data from a negative-binomial distribution where the probability of a success (or failure) $p$ is known. Use the following parametrization, \begin{equation} Pr(X=x) = \binom{x+r-1}{x} (1-p)^{r} p^x \end{equation} where $r$ is the number of successes, $x$ is the number of failures, and $p$ is the probability of failure. The goal is to make inferences about $r$. \vfill Assume that the parents favorite Halloween candy are Reese's peanut butter cups. Unbeknownst to their children they have decided to continue visiting homes until $r$ more Reese's peanut butter cups have been obtained. In this example the probability of visiting a home and not getting a peanut butter cup (a failure) is $p$. The child is allowed to trick-or-treat until receiving $r$ Reese's peanut butter cups. \vfill Luckily for you the child keeps meticulous records and has recorded the number of homes visited in the last 4 years that did not have Reese's peanut butter cups. \vfill - Consider using a Metropolis algorithm to learn the value of $r$. What will you use as a proposal distribution $J(\theta^*|\theta)$? \vfill - Is your proposal distribution symmetric? In other words, does the $Pr(\theta^* \rightarrow \theta) = Pr(\theta \rightarrow \theta^*)$ for all $\theta^*,\theta$? \vfill - Assume, you have devised a non-symmetric proposal, where: \begin{equation*} \frac{J(1|0)}{J(0|1)} \approx 2. \end{equation*} In other words, you are twice as likely to propose a move from 0 to 1 than from 1 to 0. This could be due to a random step proposal near the end of the support for $r$. What implications do you suppose this has on the posterior probabilities $\left(Pr(r=1|x) \text{ and } Pr(r=0|x)\right)$ using the standard Metropolis algorithm, with an acceptance ratio of proportional to $min(1,\alpha)$ where : \begin{equation*} \alpha = \frac{p(\theta^*|y)}{p(\theta^{(s)})}= \frac{p(y|\theta^*)p(\theta^*)}{p(y|\theta^{(s)}) p(\theta^{(s)})}. \end{equation*} \newpage The Metropolis-Hastings algorithm permits non-symmetric proposals, \vfill \vfill __Q:__ What does the second term in the acceptance ratio do? Consider the case from above where $\frac{J(1|0)}{J(0|1)} \approx 2$ what impact does this have on moving from $\theta^{(s)}=1$ to $\theta^* = 0$ ? \vfill \vfill It is obvious that the Metropolis algorithm is a special case of Metropolis-Hastings, but how about a Gibbs sampler? Assume we are interested in moving from $\theta_1^{(s)}$ to $\theta_1^*$. \begin{eqnarray*} r &=& \frac{\pi(\theta^*_1,\theta_2^{(s)})}{\pi(\theta^{(s)}_1,\theta_2^{(s)})} \times \frac{J(\theta_1^{(s)}|\theta_1^*,\theta_2^{(s)})}{J(\theta_1^*|\theta_1^{(s)},\theta_2^{(s)})} \\ \text{the proposal is the full conditional} &=& \frac{\pi(\theta^*_1,\theta_2^{(s)})}{\pi(\theta^{(s)}_1,\theta_2^{(s)})} \times \frac{\pi(\theta_1^{(s)}|\theta_2^{(s)})}{\pi(\theta_1^*|\theta_2^{(s)})}\\ &=& \frac{\pi(\theta^*_1|\theta_2^{(s)}) \pi(\theta_2^{(s)})}{\pi(\theta^{(s)}_1|\theta_2^{(s)})\pi(\theta_2^{(s)})} \times \frac{\pi(\theta_1^{(s)}|\theta_2^{(s)})}{\pi(\theta_1^*|\theta_2^{(s)})}\\ &=& \frac{\pi(\theta_2^{(s)})}{\pi(\theta_2^{(s)})} = 1 \end{eqnarray*} So the Gibbs sampler is a very specific Metropolis-Hastings algorithm where the acceptance probability is always 1. ##### MCMC Theory We have seen some empirical results that suggest these MCMC algorithms are reasonable, but what about theoretical guarantees? Recall the first (MC) chunk stand for Markov chain. The __Markov property__, in this case, is that each iteration of the sampler is only dependent on on the current values. First we establish some general properties of Markov chains that are useful for convergence of our MCMC algorithms. \vfill 1. __irreducible__. \vfill 2. __aperiodic__. \vfill 3. __recurrent__. \vfill \newpage Recall, our Monte Carlo algorithms use central limit theorem ideas to show convergence of quantities compute from the posterior samples. However, given the dependence in the MCMC samples, we use theory of Markov Chains. \vfill __Ergodic Theorem__. *If $\{x^{(1)}, x^{(2)}, \dots \}$ is an irreducible, aperiodic, and recurrent Markov chain, then there is a unique probability distribution $\pi$ such that as $s \rightarrow \infty$, then* \vfill 1. $Pr(x^{(s)} \in A)$ \vfill 2. $\frac{1}{S} \sum g(x^{(s)})$ \vfill Here $\pi()$ is denoted as the stationary distribution of the Markov Chain and has the following property: if $x^{(s)} \sim \pi$ and $x^{(s+1)}$ comes from that Markov Chain started by $x^{(s)}$ then $Pr(x^{(s+1)} \in A) = \pi(A)$. In other words, if a sample comes from the stationary distribution and used to generate more realizations from the Markov chain, then those appear according to the probabilities of $\pi$. \vfill Now we need to show that $\pi(.) =$ the target distribution, $p_0()$ (joint posterior) for our MCMC examples. To verify this, assume $x^{(s)}$ comes from the target distribution and $x^{(s)}$ is generated from $x^{(s)}$ via the M-H algorithm, then we need to show $Pr(x^{(s+1)}=x)=p_0(x)$. Let $x_a$ and $x_b$ be any two x values. WLOG assume $p_0(x_a) J(x_b|x_a) \geq p_0(x_b) J(x_a|x_b)$. Then for MH, the probability of transitioning from $x^{(s)} = x_a$ to $x^{(s+1)} = x_b$ is equal to the probability of: 1. sampling $x^{(s)} = x_a$ from $p_0$, \vfill 2. proposing $x^{*} = x_b$ from $J(x^*|x^{(s)})$ \vfill 3. accepting $x^{(s+1)} = x_b$. \vfill The probability of these three steps is: \begin{eqnarray*} Pr(x^{(s)} = x_a, x^{(s+1)} = x_b) &=& p_0(x_a) \times J(x_b|x_a) \times \frac{p_0(x_b)J(x_a|x_b)}{p_0(x_b) J(x_b|x_a)}\\ &=& p_0(x_b) J(x_a|x_b). \end{eqnarray*} \vfill \newpage To go the other direction, where $x^{(s)}= x_b$ and $x^{(s+1)} = x_a$ the acceptance probability is 1 (as $p_0(x_a) J(x_b|x_a) \geq p_0(x_b) J(x_a|x_b)$). So $Pr(x^{(s)} = x_b, x^{(s+1)} = x_a) = p_0(x_b) J(x_a|x_b).$ This implies that the joint probability of observing $x^{(s)}$ and $x^{(s+1)}$ is the same for any order of $x_b$ and $x_a$. The final step of the proof is to show that $Pr(x^{(s+1)}=x) = p_0(x)$. \vfill \begin{eqnarray*} Pr(x^{(s+1)} = x) &=& \sum_{x_a} Pr(x^{(s+1)} = x, x^{(s)} - x_a)\\ &=& \sum_{x_a} Pr(x^{(s+1)}=x_a,x^{(s)}=x)\\ &=& Pr(x^{(s)} = x) \end{eqnarray*} Hence as $Pr(x^{(s)}=x) = p_0(x)$ then $Pr(x^{(s+1)}=x)=p_0(x)$. \vfill ##### Metropolis with Gibbs - Bayesian Kriging Often a Gibbs sampler and a more vanilla Metropolis-Hastings style proposal can be used together in the same algorithm. \vfill Recall the Bayesian spatial model: \begin{equation*} y \sim N(X\tilde{\beta}, \sigma^2 H(\phi) + \tau^2 I), \end{equation*} Where $H(\phi)$ is a correlation matrix, such as $h_{ij} = \exp(-d_{ij}/ \phi)$ where $h_{ij}$ is the correlation between sites $i$ and $j$ and $d_{ij}$ is the distance between sites $i$ and $j$. \vfill Sketch out an algorithm to sample the following parameters: $\tilde{\beta}$, $\sigma^2$, $\phi$, and $\tau^2$. \vfill