Bayesian Iterative Simulation Methods

Apr 27, 2016 rubric

A useful alternative approach to Maximum Likelihood(ML) methods, particularly when the sample size is small, is to include a reasonable prior distribution for the parameters and compute the posterior distribution of the parameters of interest. The posterior distribution for a model with ignorable missingness is

$p (θ ∣ Y_{0}, M) \equiv p (θ ∣ Y_{0}) \propto p (θ) f (Y_{0} ∣ θ),$

where $p (θ)$ is the prior and $f (Y_{0} ∣ θ)$ is the density of the observed data $Y_{0}$ . Simulation from the posterior without iteration can be accomplished if the likelihood can be factored into complete data components, while for general patterns of missing data, Bayesian simulation requires iteration.

Data Augmentation

Data Augmentation(Tanner and Wong (1987)), or DA, is an iterative method of simulating the posteiror distribution of $θ$ that combines features of the Expecation Maximisation(EM) algorithm and Multiple Imputation(MI). Starting with an initial draw $θ_{0}$ from an approximation to the posterior, then given the value $θ_{t}$ at iteration $t$ :

Draw $Y_{1, t + 1}$ with density $p (Y_{1} ∣ Y_{0}, θ_{t})$ (I step).
Draw $θ_{t + 1}$ with density $p (θ ∣ Y_{0}, Y_{1, t + 1})$ (P step).

The procedure is motivated by the fact that the distributions in these two steps are often much easier to draw from than either of the posteriors $p (Y_{1} ∣ Y_{0})$ and $p (θ ∣ Y_{0})$ , or the joint posterior $p (θ, Y_{1} ∣ Y_{0})$ . The procedure can be shown to eventually yield a draw from the joint posterior of $Y_{1}$ and $θ$ given $Y_{0}$ , in the sense that as $t$ tends to infinity this sequence converges to a draw from the joint distribution.

Bivariate Normal Data Example

Suppose having a sample $y_{i} = (y_{1 i}, y_{2 i})$ from a Bivariate Normal distribution for $i = 1, \dots, n$ units, with mean vector $μ = (μ_{1}, μ_{2})$ and $2 \times 2$ covariance matrix $Σ$ . Assume that one group of units has $Y_{1}$ observed and $Y_{2}$ missing, while a second group of units has both variables observed and a third group of units has $Y_{1}$ missing and $Y_{2}$ observed. Under DA methods, each iteration $t$ consists of an I step and a P step. In the first, missing data are replaced with draws from its conditional distribution given the observed data and current values of the parameters (rather then its conditional mean as in the EM algorithm). Because units are conditionally independent given the parameters, each missing $y_{2 i}$ is drawn independently as

$y_{2 i, t + 1} \sim N (β_{20 t} + β_{21 t} y_{1 i}, σ_{2 t}^{2}),$

where $β_{20 t}, β_{21 t}$ and $σ_{2 t}^{2}$ are the $t$ -th iterates of the regression parameters of $Y_{2}$ on $Y_{1}$ . Analogously, each missing $y_{1 i}$ is drawn independently as

$y_{1 i, t + 1} \sim N (β_{10 t} + β_{11 t} y_{2 i}, σ_{1 t}^{2}),$

where $β_{10 t}, β_{11 t}$ and $σ_{1 t}^{2}$ are the $t$ -th iterates of the regression parameters of $Y_{1}$ on $Y_{2}$ . In the second step, these drawn values are treated as if they were the observed values and one draw of the bivariate Normal parameters is made from the complete data posterior. In the limit, the draws are from the joint posterior of the missing values and the parameters. Thus, a run of DA generates both a draw from the posterior predictive distribution of $Y_{1}$ and a draw from the posterior of $θ$ , and the procedure can be run $D$ times to obtain $D$ iid draws from the joint posterior of $θ$ and $Y_{1}$ . Unlike the EM, estimates of the sampling covariance matrix from the filled-in data can be computed without any corrections to the estimated variances because draws from the posterior predictive distribution of the missing values are imputed in the I step of DA, rather than the conditional means as in the E step of EM. The loss of efficiency from imputing draws is limited when the posterior mean from DA is computed over many draws from the posterior.

The Gibbs’ Sampler

The Gibbs’s sampler is an iterative simulation method that is designed to yield draws from the joint posterior distribution in the case of a general pattern of missingness and provides a Bayesian analogous to the Expectation Conditonal Maximisation (ECM) algorithm for ML estimation. The Gibbs’ sampler eventually generates a draw from the distribution $p (x_{1}, \dots, x_{J})$ of a set of $J$ random variables $X_{1}, \dots, X_{J}$ in settings where draws from the joint distribution are hard to compute but draws from the conditional distributions $p (x_{j} ∣ x_{1}, \dots, x_{j - 1}, x_{j + 1}, \dots, x_{J})$ are relatively easy to compute. Initial values $x_{10}, \dots, x_{J 0}$ are chosen in some way and then, given current values of $x_{1 t}, \dots, x_{J t}$ at iteration $t$ , new values are found by drawing from the following sequence of conditional distributions:

$x_{1 t + 1} \sim p (x_{1} ∣ x_{2 t}, \dots, x_{J t}),$

$x_{2 t + 1} \sim p (x_{2} ∣ x_{1 t + 1}, \dots, x_{J t}),$

up to

$x_{J t + 1} \sim p (x_{J} ∣ x_{2 t + 1}, \dots, x_{J - 1 t + 1}) .$

It can be shown that, under general conditions, the sequence of $J$ iterates converges to a draw from the joint posterior of the variables. When $J = 2$ , the Gibbs’ sampler is the same as DA if $x_{1} = Y_{1}$ and $x_{2} = θ$ and the distributions condition on $Y_{0}$ . We can then obtain a draw from the joint posterior of $Y_{1}, θ ∣ Y_{0}$ by applying the Gibbs’ sampler, where at iteration $t$ for the $d$ -th imputed data set:

$Y_{1 t + 1}^{d} \sim p (Y_{1} ∣ Y_{0}, θ_{t}^{d}) and θ_{t + 1}^{d} \sim p (θ ∣ Y_{1 t + 1}^{d}, Y_{0}),$

such that one run of the sampler converges to a draw from the posterior predictive distribution of $Y_{1}$ and a draw from the posterior of $θ$ . The sampler can be run independently $D$ times to generate $D$ iid draws from the approximate joint posterior of $θ$ and $Y_{1}$ . The values of $Y_{1}$ are multiple imputations of the missing values, drawn from their posterior predictive distribution.

Assessing Convergence

Assessing convergence of the sequence of draws to the target distribution is more difficult than assessing convergence of an EM-type algorithm because there is no single target quantity to monitor like the maximum value of the likelihood. Methods have been proposed to assess convergence of a single sequence (Geyer (1992)), but a more reliable approach is to simulate $D > 1$ sequences with starting values dispersed throughout the parameter space, and the convergence of all quantities of interest can then be monitored by comparing variation between and within simulated sequences, until the “within” variation roughly equals the “between” variation. The idea is that when the distribution of each simulated sequence is close enough to the distribution of all the sequences mixed together, they can all be approximating the target distribution. Gelman and Rubin (1992) developed an explicit monitoring statistic based on the following idea. For each scalar estimand $ψ$ , label the draws from $D$ parallel sequences as $ψ_{t}^{d}$ , for $t = 1, \dots, T$ iterations and $d = 1, \dots, D$ sequences, and compute the between $B$ and within $\bar{V}$ sequence variances as:

$B = \frac{T}{D - 1} \sum_{d = 1}^{D} ({\bar{ψ}}_{d .} - {\bar{ψ}}_{. .})^{2}, and \bar{V} = \frac{1}{D} \sum_{d = 1}^{D} s_{d}^{2},$

where ${\bar{ψ}}_{d .} = \frac{1}{T} \sum_{t = 1}^{T} ψ_{d t}$ , ${\bar{ψ}}_{. .} = \frac{1}{D} \sum_{d = 1}^{D} {\bar{ψ}}_{d}$ , and $s_{d}^{2} = \frac{1}{T - 1} \sum_{t = 1}^{T} (ψ_{d t} - {\bar{ψ}}_{d .})^{2}$ . We can then estimate the marginal posterior variance of the estimand as

$\hat{V a r} (ψ ∣ Y_{0}) = \frac{T - 1}{T} \hat{V} + \frac{1}{T} B,$

which will overestimate the marginal posterior variance assuming the starting distribution is appropriately over-dispersed but is unbiased under stationarity (starting distribution equals the target distribution). For any finte $T$ , the within variance $\hat{V}$ will underestimate the marginal variance because individual sequences have not had time to range over all the target distribution and should have smaller variance then B. In the limit as $T \to \infty$ the expecation of $\hat{V}$ approaches the marginal variance. These facts suggest monitoring convergence by estimating the factor by which the scale of the current distribution for $ψ$ might be reduced if the simulations were continued. This is the potential scale reduction factor and is estimated by

$\sqrt{\hat{R}} = \sqrt{\frac{\hat{V a r} (ψ ∣ Y_{0})}{\hat{V}}},$

which declines to 1 as $T \to \infty$ . When this quantity is high, there is evidence to proceed the simulations further to improve our inference about the target distribution.

Other Simulation Methods

When draws from the sequence of conditional distributions forming the Gibbs’ sampler are not easy to obtain, other simulation approaches are needed. Among these there are the Sequential Imputation (Kong, Liu, and Wong (1994)), Sampling Imprtance Resampling (Gelfand and Smith (1990)), Rejection Sampling (Von Neumann and others (1951)). One of these alternatives are the Metropolis-Hastings (Metropolis et al. (1953)) algorithms, of which the Gibbs’ sampler is a particular case, which constitute the so-called Markov Chain Monte Carlo (MCMC) algorithms as the sequence of iterates forms a Markov Chain (Gelman et al. (2013)).

References

Gelfand, Alan E, and Adrian FM Smith. 1990. “Sampling-Based Approaches to Calculating Marginal Densities.” Journal of the American Statistical Association 85 (410): 398–409.

Gelman, Andrew, John B Carlin, Hal S Stern, David B Dunson, Aki Vehtari, and Donald B Rubin. 2013. Bayesian Data Analysis. Chapman; Hall/CRC.

Gelman, Andrew, and Donald B Rubin. 1992. “A Single Series from the Gibbs Sampler Provides a False Sense of Security.” Bayesian Statistics 4: 625–31.

Geyer, Charles J. 1992. “Practical Markov Chain Monte Carlo.” Statistical Science, 473–83.

Kong, Augustine, Jun S Liu, and Wing Hung Wong. 1994. “Sequential Imputations and Bayesian Missing Data Problems.” Journal of the American Statistical Association 89 (425): 278–88.

Metropolis, Nicholas, Arianna W Rosenbluth, Marshall N Rosenbluth, Augusta H Teller, and Edward Teller. 1953. “Equation of State Calculations by Fast Computing Machines.” The Journal of Chemical Physics 21 (6): 1087–92.

Tanner, Martin A, and Wing Hung Wong. 1987. “The Calculation of Posterior Distributions by Data Augmentation.” Journal of the American Statistical Association 82 (398): 528–40.

Von Neumann, John, and others. 1951. “The General and Logical Theory of Automata.” 1951, 1–41.

Edit this page

Bayesian Iterative Simulation Methods Bayesian Inference Likelihood Based Methods Ignorable