One-way ANOVA with fixed and random effects from a Bayesian perspective

This blog post is derived from a computer practical session that I ran as part of my new course on Statistics for Big Data, previously discussed. This course covered a lot of material very quickly. In particular, I deferred introducing notions of hierarchical modelling until the Bayesian part of the course, where I feel it is more natural and powerful. However, some of the terminology associated with hierarchical statistical modelling probably seems a bit mysterious to those without a strong background in classical statistical modelling, and so this practical session was intended to clear up some potential confusion. I will analyse a simple one-way Analysis of Variance (ANOVA) model from a Bayesian perspective, making sure to highlight the difference between fixed and random effects in a Bayesian context where everything is random, as well as emphasising the associated identifiability issues. R code is used to illustrate the ideas.

Example scenario

We will consider the body mass index (BMI) of new male undergraduate students at a selection of UK Universities. Let us suppose that our data consist of measurements of (log) BMI for a random sample of 1,000 males at each of 8 Universities. We are interested to know if there are any differences between the Universities. Again, we want to model the process as we would simulate it, so thinking about how we would simulate such data is instructive. We start by assuming that the log BMI is a normal random quantity, and that the variance is common across the Universities in question (this is quite a big assumption, and it is easy to relax). We assume that the mean of this normal distribution is University-specific, but that we do not have strong prior opinions regarding the way in which the Universities differ. That said, we expect that the Universities would not be very different from one another.

Simulating data

A simple simulation of the data with some plausible parameters can be carried out as follows.

set.seed(1)
Z=matrix(rnorm(1000*8,3.1,0.1),nrow=8)
RE=rnorm(8,0,0.01)
X=t(Z+RE)
colnames(X)=paste("Uni",1:8,sep="")
Data=stack(data.frame(X))
boxplot(exp(values)~ind,data=Data,notch=TRUE)

Make sure that you understand exactly what this code is doing before proceeding. The boxplot showing the simulated data is given below.

Frequentist analysis

We will start with a frequentist analysis of the data. The model we would like to fit is

where i is an indicator for the University and j for the individual within a particular University. The “effect”, represents how the ith University differs from the overall mean. We know that this model is not actually identifiable when the model parameters are all treated as “fixed effects”, but R will handle this for us.

> mod=lm(values~ind,data=Data)
> summary(mod)

Call:
lm(formula = values ~ ind, data = Data)

Residuals:
     Min       1Q   Median       3Q      Max 
-0.36846 -0.06778 -0.00069  0.06910  0.38219 

Coefficients:
             Estimate Std. Error t value Pr(>|t|)    
(Intercept)  3.101068   0.003223 962.244  < 2e-16 ***
indUni2     -0.006516   0.004558  -1.430 0.152826    
indUni3     -0.017168   0.004558  -3.767 0.000166 ***
indUni4      0.017916   0.004558   3.931 8.53e-05 ***
indUni5     -0.022838   0.004558  -5.011 5.53e-07 ***
indUni6     -0.001651   0.004558  -0.362 0.717143    
indUni7      0.007935   0.004558   1.741 0.081707 .  
indUni8      0.003373   0.004558   0.740 0.459300    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.1019 on 7992 degrees of freedom
Multiple R-squared:  0.01439,	Adjusted R-squared:  0.01353 
F-statistic: 16.67 on 7 and 7992 DF,  p-value: < 2.2e-16

We see that R has handled the identifiability problem using “treatment contrasts”, dropping the fixed effect for the first university, so that the intercept actually represents the mean value for the first University, and the effects for the other Univeristies represent the differences from the first University. If we would prefer to impose a sum constraint, then we can switch to sum contrasts with

options(contrasts=rep("contr.sum",2))

and then re-fit the model.

> mods=lm(values~ind,data=Data)
> summary(mods)

Call:
lm(formula = values ~ ind, data = Data)

Residuals:
     Min       1Q   Median       3Q      Max 
-0.36846 -0.06778 -0.00069  0.06910  0.38219 

Coefficients:
              Estimate Std. Error  t value Pr(>|t|)    
(Intercept)  3.0986991  0.0011394 2719.558  < 2e-16 ***
ind1         0.0023687  0.0030146    0.786 0.432048    
ind2        -0.0041477  0.0030146   -1.376 0.168905    
ind3        -0.0147997  0.0030146   -4.909 9.32e-07 ***
ind4         0.0202851  0.0030146    6.729 1.83e-11 ***
ind5        -0.0204693  0.0030146   -6.790 1.20e-11 ***
ind6         0.0007175  0.0030146    0.238 0.811889    
ind7         0.0103039  0.0030146    3.418 0.000634 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.1019 on 7992 degrees of freedom
Multiple R-squared:  0.01439,	Adjusted R-squared:  0.01353 
F-statistic: 16.67 on 7 and 7992 DF,  p-value: < 2.2e-16

This has 7 degrees of freedom for the effects, as before, but ensures that the 8 effects sum to precisely zero. This is arguably more interpretable in this case.

Bayesian analysis

We will now analyse the simulated data from a Bayesian perspective, using JAGS.

Fixed effects

All parameters in Bayesian models are uncertain, and therefore random, so there is much confusion regarding the difference between “fixed” and “random” effects in a Bayesian context. For “fixed” effects, our prior captures the idea that we sample the effects independently from a “fixed” (typically vague) prior distribution. We could simply code this up and fit it in JAGS as follows.

require(rjags)
n=dim(X)[1]
p=dim(X)[2]
data=list(X=X,n=n,p=p)
init=list(mu=2,tau=1)
modelstring="
  model {
    for (j in 1:p) {
      theta[j]~dnorm(0,0.0001)
      for (i in 1:n) {
        X[i,j]~dnorm(mu+theta[j],tau)
      }
    }
    mu~dnorm(0,0.0001)
    tau~dgamma(1,0.0001)
  }
"
model=jags.model(textConnection(modelstring),data=data,inits=init)
update(model,n.iter=1000)
output=coda.samples(model=model,variable.names=c("mu","tau","theta"),n.iter=100000,thin=10)
print(summary(output))
plot(output)
autocorr.plot(output)
pairs(as.matrix(output))
crosscorr.plot(output)

On running the code we can clearly see that this naive approach leads to high posterior correlation between the mean and the effects, due to the fundamental lack of identifiability of the model. This also leads to MCMC mixing problems, but it is important to understand that this computational issue is conceptually entirely separate from the fundamental statisticial identifiability issue. Even if we could avoid MCMC entirely, the identifiability issue would remain.

A quick fix for the identifiability issue is to use “treatment contrasts”, just as for the frequentist model. We can implement that as follows.

data=list(X=X,n=n,p=p)
init=list(mu=2,tau=1)
modelstring="
  model {
    for (j in 1:p) {
      for (i in 1:n) {
        X[i,j]~dnorm(mu+theta[j],tau)
      }
    }
    theta[1]<-0
    for (j in 2:p) {
      theta[j]~dnorm(0,0.0001)
    }
    mu~dnorm(0,0.0001)
    tau~dgamma(1,0.0001)
  }
"
model=jags.model(textConnection(modelstring),data=data,inits=init)
update(model,n.iter=1000)
output=coda.samples(model=model,variable.names=c("mu","tau","theta"),n.iter=100000,thin=10)
print(summary(output))
plot(output)
autocorr.plot(output)
pairs(as.matrix(output))
crosscorr.plot(output)

Running this we see that the model now works perfectly well, mixes nicely, and gives sensible inferences for the treatment effects.

Another source of confusion for models of this type is data formating and indexing in JAGS models. For our balanced data there was not problem passing in data to JAGS as a matrix and specifying the model using nested loops. However, for unbalanced designs this is not necessarily so convenient, and so then it can be helpful to specify the model based on two-column data, as we would use for fitting using lm(). This is illustrated with the following model specification, which is exactly equivalent to the previous model, and should give identical (up to Monte Carlo error) results.

N=n*p
data=list(y=Data$values,g=Data$ind,N=N,p=p)
init=list(mu=2,tau=1)
modelstring="
  model {
    for (i in 1:N) {
      y[i]~dnorm(mu+theta[g[i]],tau)
    }
    theta[1]<-0
    for (j in 2:p) {
      theta[j]~dnorm(0,0.0001)
    }
    mu~dnorm(0,0.0001)
    tau~dgamma(1,0.0001)
  }
"
model=jags.model(textConnection(modelstring),data=data,inits=init)
update(model,n.iter=1000)
output=coda.samples(model=model,variable.names=c("mu","tau","theta"),n.iter=100000,thin=10)
print(summary(output))
plot(output)

As suggested above, this indexing scheme is much more convenient for unbalanced data, and hence widely used. However, since our data is balanced here, we will revert to the matrix approach for the remainder of the post.

One final thing to consider before moving on to random effects is the sum-contrast model. We can implement this in various ways, but I’ve tried to encode it for maximum clarity below, imposing the sum-to-zero constraint via the final effect.

data=list(X=X,n=n,p=p)
init=list(mu=2,tau=1)
modelstring="
  model {
    for (j in 1:p) {
      for (i in 1:n) {
        X[i,j]~dnorm(mu+theta[j],tau)
      }
    }
    for (j in 1:(p-1)) {
      theta[j]~dnorm(0,0.0001)
    }
    theta[p] <- -sum(theta[1:(p-1)])
    mu~dnorm(0,0.0001)
    tau~dgamma(1,0.0001)
  }
"
model=jags.model(textConnection(modelstring),data=data,inits=init)
update(model,n.iter=1000)
output=coda.samples(model=model,variable.names=c("mu","tau","theta"),n.iter=100000,thin=10)
print(summary(output))
plot(output)

Again, this works perfectly well and gives similar results to the frequentist analysis.

Random effects

The key difference between fixed and random effects in a Bayesian framework is that random effects are not independent, being drawn from a distribution with parameters which are not fixed. Essentially, there is another level of hierarchy involved in the specification of the random effects. This is best illustrated by example. A random effects model for this problem is given below.

data=list(X=X,n=n,p=p)
init=list(mu=2,tau=1)
modelstring="
  model {
    for (j in 1:p) {
      theta[j]~dnorm(0,taut)
      for (i in 1:n) {
        X[i,j]~dnorm(mu+theta[j],tau)
      }
    }
    mu~dnorm(0,0.0001)
    tau~dgamma(1,0.0001)
    taut~dgamma(1,0.0001)
  }
"
model=jags.model(textConnection(modelstring),data=data,inits=init)
update(model,n.iter=1000)
output=coda.samples(model=model,variable.names=c("mu","tau","taut","theta"),n.iter=100000,thin=10)
print(summary(output))
plot(output)

The only difference between this and our first naive attempt at a Bayesian fixed effects model is that we have put a gamma prior on the precision of the effect. Note that this model now runs and fits perfectly well, with reasonable mixing, and gives sensible parameter inferences. Although the effects here are not constrained to sum-to-zero, like in the case of sum contrasts for a fixed effects model, the prior encourages shrinkage towards zero, and so the random effect distribution can be thought of as a kind of soft version of a hard sum-to-zero constraint. From a predictive perspective, this model is much more powerful. In particular, using a random effects model, we can make strong predictions for unobserved groups (eg. a ninth University), with sensible prediction intervals based on our inferred understanding of how similar different universities are. Using a fixed effects model this isn’t really possible. Even for a Bayesian version of a fixed effects model using proper (but vague) priors, prediction intervals for unobserved groups are not really sensible.

Since we have used simulated data here, we can compare the estimated random effects with the true effects generated during the simulation.

> apply(as.matrix(output),2,mean)
           mu           tau          taut      theta[1]      theta[2] 
 3.098813e+00  9.627110e+01  7.015976e+03  2.086581e-03 -3.935511e-03 
     theta[3]      theta[4]      theta[5]      theta[6]      theta[7] 
-1.389099e-02  1.881528e-02 -1.921854e-02  5.640306e-04  9.529532e-03 
     theta[8] 
 5.227518e-03 
> RE
[1]  0.002637034 -0.008294518 -0.014616348  0.016839902 -0.015443243
[6] -0.001908871  0.010162117  0.005471262

We see that the Bayesian random effects model has done an excellent job of estimation. If we wished, we could relax the assumption of common variance across the groups by making tau a vector indexed by j, though there is not much point in persuing this here, since we know that the groups do all have the same variance.

Strong subjective priors

The above is the usual story regarding fixed and random effects in Bayesian inference. I hope this is reasonably clear, so really I should quit while I’m ahead… However, the issues are really a bit more subtle than I’ve suggested. The inferred precision of the random effects was around 7,000, so now lets re-run the original, naive, “fixed effects” model with a strong subjective Bayesian prior on the distribution of the effects.

data=list(X=X,n=n,p=p)
init=list(mu=2,tau=1)
modelstring="
  model {
    for (j in 1:p) {
      theta[j]~dnorm(0,7000)
      for (i in 1:n) {
        X[i,j]~dnorm(mu+theta[j],tau)
      }
    }
    mu~dnorm(0,0.0001)
    tau~dgamma(1,0.0001)
  }
"
model=jags.model(textConnection(modelstring),data=data,inits=init)
update(model,n.iter=1000)
output=coda.samples(model=model,variable.names=c("mu","tau","theta"),n.iter=100000,thin=10)
print(summary(output))
plot(output)

This model also runs perfectly well and gives sensible inferences, despite the fact that the effects are iid from a fixed distribution and there is no hard constraint on the effects. Similarly, we can make sensible predictions, together with appropriate prediction intervals, for an unobserved group. So it isn’t so much the fact that the effects are coupled via an extra level of hierarchy that makes things work. It’s really the fact that the effects are sensibly distributed and not just sampled directly from a vague prior. So for “real” subjective Bayesians the line between fixed and random effects is actually very blurred indeed…

Marginal likelihood from tempered Bayesian posteriors

Introduction

In the previous post I showed that it is possible to couple parallel tempered MCMC chains in order to improve mixing. Such methods can be used when the target of interest is a Bayesian posterior distribution that is difficult to sample. There are (at least) a couple of obvious ways that one can temper a Bayesian posterior distribution. Perhaps the most obvious way is a simple flattening, so that if

is the posterior distribution, then for we define

This corresponds with the tempering that is often used in statistical physics applications. We recover the posterior of interest for and tend to a flat distribution as . However, for Bayesian posterior distributions, there is a different way of tempering that is often more natural and useful, and that is to temper using the power posterior, defined by

Here we again recover the posterior for , but get the prior for . Thus, the family of distributions forms a natural bridge or path from the prior to the posterior distributions. The power posterior is a special case of the more general concept of a geometric path from distribution (at ) to (at ) defined by

where, in our case, is the prior and is the posterior.

So, given a posterior distribution that is difficult to sample, choose a temperature schedule

and run a parallel tempering scheme as outlined in the previous post. The idea is that for small values of mixing will be good, as prior-like distributions are usually well-behaved, and the mixing of these "high temperature" chains can help to improve the mixing of the "low temperature" chains that are more like the posterior (note that is really an inverse temperature parameter the way I’ve defined it here…).

Marginal likelihood and normalising constants

The marginal likelihood of a Bayesian model is

This quantity is of interest for many reasons, including calculation of the Bayes factor between two competing models. Note that this quantity has several different names in different fields. In particular, it is often known as the evidence, due to its role in Bayes factors. It is also worth noting that it is the normalising constant of the Bayesian posterior distribution. Although it is very easy to describe and define, it is notoriously difficult to compute reliably for complex models.

The normalising constant is conceptually very easy to estimate. From the above integral representation, it is clear that

where the expectation is taken with respect to the prior. So, given samples from the prior, , we can construct the Monte Carlo estimate

and this will be a consistent estimator of the true evidence under fairly mild regularity conditions. Unfortunately, in practice it is likely to be a very poor estimator if the posterior and prior are not very similar. Now, we could also use Bayes theorem to re-write the integral as an expectation with respect to the posterior, so we could then use samples from the posterior to estimate the evidence. This leads to the harmonic mean estimator of the evidence, which has been described as the worst Monte Carlo method ever! Now it turns out that there are many different ways one can construct estimators of the evidence using samples from the prior and the posterior, some of which are considerably better than the two I’ve outlined. This is the subject of the bridge sampling paper of Meng and Wong. However, the reality is that no method will work well if the prior and posterior are very different.

If we have tempered chains, then we have a sequence of chains targeting distributions which, by construction, are not too different, and so we can use the output from tempered chains in order to construct estimators of the evidence that are more numerically stable. If we call the evidence of the th chain , so that and , then we can write the evidence in telescoping fashion as

Now the th term in this product is , which can be estimated using the output from the th and/or th chain(s). Again, this can be done in a variety of ways, using your favourite bridge sampling estimator, but the point is that the estimator should be reasonably good due to the fact that the th and th targets are very similar. For the power posterior, the simplest method is to write

where the expectation is with respect to the th target, and hence can be estimated in the usual way using samples from the th chain.

For numerical stability, in practice we compute the log of the evidence as

The above expression is exact, and is the obvious formula to use for computation. However, it is clear that if and are sufficiently close, it will be approximately OK to switch the expectation and exponential, giving

In the continuous limit, this gives rise to the well-known path sampling identity,

So, an alternative approach to computing the evidence is to use the samples to approximately numerically integrate the above integral, say, using the trapezium rule. However, it isn’t completely clear (to me) that this is better than using directly, since there there is no numerical integration error to worry about.

Numerical illustration

We can illustrate these ideas using the simple double potential well example from the previous post. Now that example doesn’t really correspond to a Bayesian posterior, and is tempered directly, rather than as a power posterior, but essentially the same ideas follow for general parallel tempered distributions. In general, we can use the sample to estimate the ratio of the last and first normalising constants, . Here it isn’t obvious why we’d want to know that, but we’ll compute it anyway to illustrate the method. As before, we expand as a telescopic product, where the th term is now

A Monte Carlo estimate of each of these terms is formed using the samples from the th chain, and the logs of these are then summed to give . A complete R script to run the Metropolis coupled sampler and compute the evidence is given below.

U=function(gam,x)
{
  gam*(x*x-1)*(x*x-1)
}

temps=2^(0:3)
iters=1e5

chains=function(pot=U, tune=0.1, init=1)
{
  x=rep(init,length(temps))
  xmat=matrix(0,iters,length(temps))
  for (i in 1:iters) {
    can=x+rnorm(length(temps),0,tune)
    logA=unlist(Map(pot,temps,x))-unlist(Map(pot,temps,can))
    accept=(log(runif(length(temps)))<logA)
    x[accept]=can[accept]
    swap=sample(1:length(temps),2)
    logA=pot(temps[swap[1]],x[swap[1]])+pot(temps[swap[2]],x[swap[2]])-
            pot(temps[swap[1]],x[swap[2]])-pot(temps[swap[2]],x[swap[1]])
    if (log(runif(1))<logA)
      x[swap]=rev(x[swap])
    xmat[i,]=x
  }
  colnames(xmat)=paste("gamma=",temps,sep="")
  xmat
}

mat=chains()
mat=mat[,1:(length(temps)-1)]
diffs=diff(temps)
mat=(mat*mat-1)^2
mat=-t(diffs*t(mat))
mat=exp(mat)
logEvidence=sum(log(colMeans(mat)))
message(paste("The log of the ratio of the last and first normalising constants is",logEvidence))

It turns out that these double well potential densities are tractable, and so the normalising constants can be computed exactly. So, with a little help from Wolfram Alpha, I compute log of the ratio of the last and first normalising constants to be approximately -1.12. Hopefully the above script will output something a bit like that…

References

• Meng, Xiao-Li, and Wing Hung Wong. “Simulating ratios of normalizing constants via a simple identity: a theoretical exploration.” Statistica Sinica 6.4 (1996): 831-860.
• Gelman, Andrew, and Xiao-Li Meng. “Simulating normalizing constants: From importance sampling to bridge sampling to path sampling.” Statistical Science (1998): 163-185.
• Friel, Nial, and Anthony N. Pettitt. “Marginal likelihood estimation via power posteriors.” Journal of the Royal Statistical Society: Series B (Statistical Methodology) 70.3 (2008): 589-607.
• Friel, Nial, and Jason Wyse. “Estimating the evidence–a review.” Statistica Neerlandica 66.3 (2012): 288-308.

Summary stats for ABC

Introduction

In the previous post I gave a very brief introduction to ABC, including a simple example for inferring the parameters of a Markov process given some time series observations. Towards the end of the post I observed that there were (at least!) two potential problems with scaling up the simple approach described, one relating to the dimension of the data and the other relating to the dimension of the parameter space. Before moving on to the (to me, more interesting) problem of the dimension of the parameter space, I will briefly discuss the data dimension problem in this post, and provide a couple of references for further reading.

Summary stats

Recall that the simple rejection sampling approach to ABC involves first sampling a candidate parameter from the prior and then sampling a corresponding data set from the model. This simulated data set is compared with the true data using some (pseudo-)norm, , and accepting if the simulated data set is sufficiently close to the true data, . It should be clear that if we are using a proper norm then as tends to zero the distribution of the accepted values tends to the desired posterior distribution of the parameters given the data.

However, smaller choices of will lead to higher rejection rates. This will be a particular problem in the context of high-dimensional , where it is often unrealistic to expect a close match between all components of and the simulated data , even for a good choice of . In this case, it makes more sense to look for good agreement between particular aspects of , such as the mean, or variance, or auto-correlation, depending on the exact problem and context. If we can find a finite set of sufficient statistics, for , then it should be clear that replacing the acceptance criterion with will also lead to a scheme tending to the true posterior as tends to zero (assuming a proper norm on the space of sufficient statistics), and will typically be better than the naive method, since the sufficient statistics will be of lower dimension and less “noisy” that the raw data, leading to higher acceptance rates with no loss of information.

Unfortunately for most problems of practical interest it is not possible to find low-dimensional sufficient statistics, and so people in practice use domain knowledge and heuristics to come up with a set of summary statistics, which they hope will closely approximate sufficient statistics. There is still a question as to how these statistics should be weighted or transformed to give a particular norm. This can be done using theory or heuristics, and some relevant references for this problem are given at the end of the post.

Implementation in R

Let’s now look at the problem from the previous post. Here, instead of directly computing the Euclidean distance between the real and simulated data, we will look at the Euclidean distance between some (normalised) summary statistics. First we will load some packages and set some parameters.

require(smfsb)
require(parallel)
options(mc.cores=4)
data(LVdata)
 
N=1e7
bs=1e5
batches=N/bs
message(paste("N =",N," | bs =",bs," | batches =",batches))

Next we will define some summary stats for a univariate time series – the mean, the (log) variance, and the first two auto-correlations.

ssinit <- function(vec)
{
  ac23=as.vector(acf(vec,lag.max=2,plot=FALSE)$acf)[2:3]
  c(mean(vec),log(var(vec)+1),ac23)
}

Once we have this, we can define some stats for a bivariate time series by combining the stats for the two component series, along with the cross-correlation between them.

ssi <- function(ts)
{
  c(ssinit(ts[,1]),ssinit(ts[,2]),cor(ts[,1],ts[,2]))
}

This gives a set of summary stats, but these individual statistics are potentially on very different scales. They can be transformed and re-weighted in a variety of ways, usually on the basis of a pilot run which gives some information about the distribution of the summary stats. Here we will do the simplest possible thing, which is to normalise the variance of the stats on the basis of a pilot run. This is not at all optimal – see the references at the end of the post for a description of better methods.

message("Batch 0: Pilot run batch")
prior=cbind(th1=exp(runif(bs,-6,2)),th2=exp(runif(bs,-6,2)),th3=exp(runif(bs,-6,2)))
rows=lapply(1:bs,function(i){prior[i,]})
samples=mclapply(rows,function(th){simTs(c(50,100),0,30,2,stepLVc,th)})
sumstats=mclapply(samples,ssi)
sds=apply(sapply(sumstats,c),1,sd)
print(sds)

# now define a standardised distance
ss<-function(ts)
{
  ssi(ts)/sds
}

ss0=ss(LVperfect)

distance <- function(ts)
{
  diff=ss(ts)-ss0
  sum(diff*diff)
}

Now we have a normalised distance function defined, we can proceed exactly as before to obtain an ABC posterior via rejection sampling.

post=NULL
for (i in 1:batches) {
  message(paste("batch",i,"of",batches))
  prior=cbind(th1=exp(runif(bs,-6,2)),th2=exp(runif(bs,-6,2)),th3=exp(runif(bs,-6,2)))
  rows=lapply(1:bs,function(i){prior[i,]})
  samples=mclapply(rows,function(th){simTs(c(50,100),0,30,2,stepLVc,th)})
  dist=mclapply(samples,distance)
  dist=sapply(dist,c)
  cutoff=quantile(dist,1000/N,na.rm=TRUE)
  post=rbind(post,prior[dist<cutoff,])
}
message(paste("Finished. Kept",dim(post)[1],"simulations"))

Having obtained the posterior, we can use the following code to plot the results.

th=c(th1 = 1, th2 = 0.005, th3 = 0.6)
op=par(mfrow=c(2,3))
for (i in 1:3) {
  hist(post[,i],30,col=5,main=paste("Posterior for theta[",i,"]",sep=""))
  abline(v=th[i],lwd=2,col=2)
}
for (i in 1:3) {
  hist(log(post[,i]),30,col=5,main=paste("Posterior for log(theta[",i,"])",sep=""))
  abline(v=log(th[i]),lwd=2,col=2)
}
par(op)

This gives the plot shown below. From this we can see that the ABC posterior obtained here is very similar to that obtained in the previous post using the full data. Here the dimension reduction is not that great – reducing from 32 data points to 9 summary statistics – and so the improvement in performance is not that noticable. But in higher dimensional problems reducing the dimension of the data is practically essential.

Summary and References

As before, I recommend the wikipedia article on approximate Bayesian computation for further information and a comprehensive set of references for further reading. Here I just want to highlight two references particularly relevant to the issue of summary statistics. It is quite difficult to give much practical advice on how to construct good summary statistics, but how to transform a set of summary stats in a “good” way is a problem that is reasonably well understood. In this post I did something rather naive (normalising the variance), but the following two papers describe much better approaches.

• Blum, M. G. B., et al. “A comparative review of dimension reduction methods in approximate Bayesian computation.” Statistical Science 28.2 (2013): 189-208.
• Fearnhead, Paul, and Dennis Prangle. “Constructing summary statistics for approximate Bayesian computation: semi‐automatic approximate Bayesian computation.” Journal of the Royal Statistical Society: Series B (Statistical Methodology) 74.3 (2012): 419-474.

I still haven’t addressed the issue of a high-dimensional parameter space – that will be the topic of a subsequent post.

The complete R script

require(smfsb)
require(parallel)
options(mc.cores=4)
data(LVdata)
 
N=1e6
bs=1e5
batches=N/bs
message(paste("N =",N," | bs =",bs," | batches =",batches))
 
ssinit <- function(vec)
{
  ac23=as.vector(acf(vec,lag.max=2,plot=FALSE)$acf)[2:3]
  c(mean(vec),log(var(vec)+1),ac23)
}

ssi <- function(ts)
{
  c(ssinit(ts[,1]),ssinit(ts[,2]),cor(ts[,1],ts[,2]))
}

message("Batch 0: Pilot run batch")
prior=cbind(th1=exp(runif(bs,-6,2)),th2=exp(runif(bs,-6,2)),th3=exp(runif(bs,-6,2)))
rows=lapply(1:bs,function(i){prior[i,]})
samples=mclapply(rows,function(th){simTs(c(50,100),0,30,2,stepLVc,th)})
sumstats=mclapply(samples,ssi)
sds=apply(sapply(sumstats,c),1,sd)
print(sds)

# now define a standardised distance
ss<-function(ts)
{
  ssi(ts)/sds
}

ss0=ss(LVperfect)

distance <- function(ts)
{
  diff=ss(ts)-ss0
  sum(diff*diff)
}

post=NULL
for (i in 1:batches) {
  message(paste("batch",i,"of",batches))
  prior=cbind(th1=exp(runif(bs,-6,2)),th2=exp(runif(bs,-6,2)),th3=exp(runif(bs,-6,2)))
  rows=lapply(1:bs,function(i){prior[i,]})
  samples=mclapply(rows,function(th){simTs(c(50,100),0,30,2,stepLVc,th)})
  dist=mclapply(samples,distance)
  dist=sapply(dist,c)
  cutoff=quantile(dist,1000/N,na.rm=TRUE)
  post=rbind(post,prior[dist<cutoff,])
}
message(paste("Finished. Kept",dim(post)[1],"simulations"))

# plot the results
th=c(th1 = 1, th2 = 0.005, th3 = 0.6)
op=par(mfrow=c(2,3))
for (i in 1:3) {
  hist(post[,i],30,col=5,main=paste("Posterior for theta[",i,"]",sep=""))
  abline(v=th[i],lwd=2,col=2)
}
for (i in 1:3) {
  hist(log(post[,i]),30,col=5,main=paste("Posterior for log(theta[",i,"])",sep=""))
  abline(v=log(th[i]),lwd=2,col=2)
}
par(op)

Introduction to Approximate Bayesian Computation (ABC)

Many of the posts in this blog have been concerned with using MCMC based methods for Bayesian inference. These methods are typically “exact” in the sense that they have the exact posterior distribution of interest as their target equilibrium distribution, but are obviously “approximate”, in that for any finite amount of computing time, we can only generate a finite sample of correlated realisations from a Markov chain that we hope is close to equilibrium.

Approximate Bayesian Computation (ABC) methods go a step further, and generate samples from a distribution which is not the true posterior distribution of interest, but a distribution which is hoped to be close to the real posterior distribution of interest. There are many variants on ABC, and I won’t get around to explaining all of them in this blog. The wikipedia page on ABC is a good starting point for further reading. In this post I’ll explain the most basic rejection sampling version of ABC, and in a subsequent post, I’ll explain a sequential refinement, often referred to as ABC-SMC. As usual, I’ll use R code to illustrate the ideas.

Basic idea

There is a close connection between “likelihood free” MCMC methods and those of approximate Bayesian computation (ABC). To keep things simple, consider the case of a perfectly observed system, so that there is no latent variable layer. Then there are model parameters described by a prior , and a forwards-simulation model for the data , defined by . It is clear that a simple algorithm for simulating from the desired posterior can be obtained as follows. First simulate from the joint distribution by simulating and then . This gives a sample from the joint distribution. A simple rejection algorithm which rejects the proposed pair unless matches the true data clearly gives a sample from the required posterior distribution.

Exact rejection sampling

• 1. Sample
• 2. Sample
• 3. If , keep as a sample from , otherwise reject.
• 4. Return to step 1.

This algorithm is exact, and for discrete will have a non-zero acceptance rate. However, in most interesting problems the rejection rate will be intolerably high. In particular, the acceptance rate will typically be zero for continuous valued .

ABC rejection sampling

The ABC “approximation” is to accept values provided that is “sufficiently close” to . In the first instance, we can formulate this as follows.

• 1. Sample
• 2. Sample
• 3. If , keep as a sample from , otherwise reject.
• 4. Return to step 1.

Euclidean distance is usually chosen as the norm, though any norm can be used. This procedure is “honest”, in the sense that it produces exact realisations from

For suitable small choice of , this will closely approximate the true posterior. However, smaller choices of will lead to higher rejection rates. This will be a particular problem in the context of high-dimensional , where it is often unrealistic to expect a close match between all components of and the simulated data , even for a good choice of . In this case, it makes more sense to look for good agreement between particular aspects of , such as the mean, or variance, or auto-correlation, depending on the exact problem and context.

In the simplest case, this is done by forming a (vector of) summary statistic(s), (ideally a sufficient statistic), and accepting provided that for some suitable choice of metric and . We will return to this issue in a subsequent post.

Inference for an intractable Markov process

I’ll illustrate ABC in the context of parameter inference for a Markov process with an intractable transition kernel: the discrete stochastic Lotka-Volterra model. A function for simulating exact realisations from the intractable kernel is included in the smfsb CRAN package discussed in a previous post. Using pMCMC to solve the parameter inference problem is discussed in another post. It may be helpful to skim through those posts quickly to become familiar with this problem before proceeding.

So, for a given proposed set of parameters, realisations from the process can be sampled using the functions simTs and stepLV (from the smfsb package). We will use the sample data set LVperfect (from the LVdata dataset) as our “true”, or “target” data, and try to find parameters for the process which are consistent with this data. A fairly minimal R script for this problem is given below.

require(smfsb)
data(LVdata)

N=1e5
message(paste("N =",N))
prior=cbind(th1=exp(runif(N,-6,2)),th2=exp(runif(N,-6,2)),th3=exp(runif(N,-6,2)))
rows=lapply(1:N,function(i){prior[i,]})
message("starting simulation")
samples=lapply(rows,function(th){simTs(c(50,100),0,30,2,stepLVc,th)})
message("finished simulation")

distance<-function(ts)
{
  diff=ts-LVperfect
  sum(diff*diff)
}

message("computing distances")
dist=lapply(samples,distance)
message("distances computed")

dist=sapply(dist,c)
cutoff=quantile(dist,1000/N)
post=prior[dist<cutoff,]

op=par(mfrow=c(2,3))
apply(post,2,hist,30)
apply(log(post),2,hist,30)
par(op)

This script should take 5-10 minutes to run on a decent laptop, and will result in histograms of the posterior marginals for the components of and . Note that I have deliberately adopted a functional programming style, making use of the lapply function for the most computationally intensive steps. The reason for this will soon become apparent. Note that rather than pre-specifying a cutoff , I’ve instead picked a quantile of the distance distribution. This is common practice in scenarios where the distance is difficult to have good intuition about. In fact here I’ve gone a step further and chosen a quantile to give a final sample of size 1000. Obviously then in this case I could have just selected out the top 1000 directly, but I wanted to illustrate the quantile based approach.

One problem with the above script is that all proposed samples are stored in memory at once. This is problematic for problems involving large numbers of samples. However, it is convenient to do simulations in large batches, both for computation of quantiles, and also for efficient parallelisation. The script below illustrates how to implement a batch parallelisation strategy for this problem. Samples are generated in batches of size 10^4, and only the best fitting samples are stored before the next batch is processed. This strategy can be used to get a good sized sample based on a more stringent acceptance criterion at the cost of addition simulation time. Note that the parallelisation code will only work with recent versions of R, and works by replacing calls to lapply with the parallel version, mclapply. You should notice an appreciable speed-up on a multicore machine.

require(smfsb)
require(parallel)
options(mc.cores=4)
data(LVdata)

N=1e5
bs=1e4
batches=N/bs
message(paste("N =",N," | bs =",bs," | batches =",batches))

distance<-function(ts)
{
  diff=ts[,1]-LVprey
  sum(diff*diff)
}

post=NULL
for (i in 1:batches) {
  message(paste("batch",i,"of",batches))
  prior=cbind(th1=exp(runif(bs,-6,2)),th2=exp(runif(bs,-6,2)),th3=exp(runif(bs,-6,2)))
  rows=lapply(1:bs,function(i){prior[i,]})
  samples=mclapply(rows,function(th){simTs(c(50,100),0,30,2,stepLVc,th)})
  dist=mclapply(samples,distance)
  dist=sapply(dist,c)
  cutoff=quantile(dist,1000/N)
  post=rbind(post,prior[dist<cutoff,])
}
message(paste("Finished. Kept",dim(post)[1],"simulations"))

op=par(mfrow=c(2,3))
apply(post,2,hist,30)
apply(log(post),2,hist,30)
par(op)

Note that there is an additional approximation here, since the top 100 samples from each of 10 batches of simulations won’t correspond exactly to the top 1000 samples overall, but given all of the other approximations going on in ABC, this one is likely to be the least of your worries.

Now, if you compare the approximate posteriors obtained here with the “true” posteriors obtained in an earlier post using pMCMC, you will see that these posteriors are really quite poor. However, this isn’t a very fair comparison, since we’ve only done 10^5 simulations. Jacking N up to 10^7 gives the ABC posterior below.

ABC posterior from 10^7 samples

This is a bit better, but really not great. There are two basic problems with the simplistic ABC strategy adopted here, one related to the dimensionality of the data and the other the dimensionality of the parameter space. The most basic problem that we have here is the dimensionality of the data. We have 16 (bivariate) observations, so we want our stochastic simulation to shoot at a point in a 16- or 32-dimensional space. That’s a tough call. The standard way to address this problem is to reduce the dimension of the data by introducing a few carefully chosen summary statistics and then just attempting to hit those. I’ll illustrate this in a subsequent post. The other problem is that often the prior and posterior over the parameters are quite different, and this problem too is exacerbated as the dimension of the parameter space increases. The standard way to deal with this is to sequentially adapt from the prior through a sequence of ABC posteriors. I’ll examine this in a future post as well, once I’ve also posted an introduction to the use of sequential Monte Carlo (SMC) samplers for static problems.

Further reading

For further reading, I suggest browsing the reference list of the Wikipedia page for ABC. Also look through the list of software on that page. In particular, note that there is a CRAN package, abc, providing R support for ABC. There is a vignette for this package which should be sufficient to get started.

Getting started with Bayesian variable selection using JAGS and rjags

Bayesian variable selection

In a previous post I gave a quick introduction to using the rjags R package to access the JAGS Bayesian inference from within R. In this post I want to give a quick guide to using rjags for Bayesian variable selection. I intend to use this post as a starting point for future posts on Bayesian model and variable selection using more sophisticated approaches.

I will use the simple example of multiple linear regression to illustrate the ideas, but it should be noted that I’m just using that as an example. It turns out that in the context of linear regression there are lots of algebraic and computational tricks which can be used to simplify the variable selection problem. The approach I give here is therefore rather inefficient for linear regression, but generalises to more complex (non-linear) problems where analytical and computational short-cuts can’t be used so easily.

Consider a linear regression problem with n observations and p covariates, which we can write in matrix form as

where is an matrix. The idea of variable selection is that probably not all of the p covariates are useful for predicting y, and therefore it would be useful to identify the variables which are, and just use those. Clearly each combination of variables corresponds to a different model, and so the variable selection amounts to choosing among the possible models. For large values of p it won’t be practical to consider each possible model separately, and so the idea of Bayesian variable selection is to consider a model containing all of the possible model combinations as sub-models, and the variable selection problem as just another aspect of the model which must be estimated from data. I’m simplifying and glossing over lots of details here, but there is a very nice review paper by O’Hara and Sillanpaa (2009) which the reader is referred to for further details.

The simplest and most natural way to tackle the variable selection problem from a Bayesian perspective is to introduce an indicator random variable for each covariate, and introduce these into the model in order to “zero out” inactive covariates. That is we write the ith regression coefficient as , so that is the regression coefficient when , and “doesn’t matter” when . There are various ways to choose the prior over and , but the simplest and most natural choice is to make them independent. This approach was used in Kuo and Mallick (1998), and hence is referred to as the Kuo and Mallick approach in O’Hara and Sillanpaa.

Simulating some data

In order to see how things work, let’s first simulate some data from a regression model with geometrically decaying regression coefficients.

n=500
p=20
X=matrix(rnorm(n*p),ncol=p)
beta=2^(0:(1-p))
print(beta)
alpha=3
tau=2
eps=rnorm(n,0,1/sqrt(tau))
y=alpha+as.vector(X%*%beta + eps)

Let’s also fit the model by least squares.

mod=lm(y~X)
print(summary(mod))

This should give output something like the following.

Call:
lm(formula = y ~ X)

Residuals:
     Min       1Q   Median       3Q      Max 
-1.62390 -0.48917 -0.02355  0.45683  2.35448 

Coefficients:
              Estimate Std. Error t value Pr(>|t|)    
(Intercept)  3.0565406  0.0332104  92.036  < 2e-16 ***
X1           0.9676415  0.0322847  29.972  < 2e-16 ***
X2           0.4840052  0.0333444  14.515  < 2e-16 ***
X3           0.2680482  0.0320577   8.361  6.8e-16 ***
X4           0.1127954  0.0314472   3.587 0.000369 ***
X5           0.0781860  0.0334818   2.335 0.019946 *  
X6           0.0136591  0.0335817   0.407 0.684379    
X7           0.0035329  0.0321935   0.110 0.912662    
X8           0.0445844  0.0329189   1.354 0.176257    
X9           0.0269504  0.0318558   0.846 0.397968    
X10          0.0114942  0.0326022   0.353 0.724575    
X11         -0.0045308  0.0330039  -0.137 0.890868    
X12          0.0111247  0.0342482   0.325 0.745455    
X13         -0.0584796  0.0317723  -1.841 0.066301 .  
X14         -0.0005005  0.0343499  -0.015 0.988381    
X15         -0.0410424  0.0334723  -1.226 0.220742    
X16          0.0084832  0.0329650   0.257 0.797026    
X17          0.0346331  0.0327433   1.058 0.290718    
X18          0.0013258  0.0328920   0.040 0.967865    
X19         -0.0086980  0.0354804  -0.245 0.806446    
X20          0.0093156  0.0342376   0.272 0.785671    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 

Residual standard error: 0.7251 on 479 degrees of freedom
Multiple R-squared: 0.7187,     Adjusted R-squared: 0.707 
F-statistic:  61.2 on 20 and 479 DF,  p-value: < 2.2e-16 

The first 4 variables are “highly significant” and the 5th is borderline.

Saturated model

We can fit the saturated model using JAGS with the following code.

require(rjags)
data=list(y=y,X=X,n=n,p=p)
init=list(tau=1,alpha=0,beta=rep(0,p))
modelstring="
  model {
    for (i in 1:n) {
      mean[i]<-alpha+inprod(X[i,],beta)
      y[i]~dnorm(mean[i],tau)
    }
    for (j in 1:p) {
      beta[j]~dnorm(0,0.001)
    }
    alpha~dnorm(0,0.0001)
    tau~dgamma(1,0.001)
  }
"
model=jags.model(textConnection(modelstring),
				data=data,inits=init)
update(model,n.iter=100)
output=coda.samples(model=model,variable.names=c("alpha","beta","tau"),
			n.iter=10000,thin=1)
print(summary(output))
plot(output)

I’ve hard-coded various hyper-parameters in the script which are vaguely reasonable for this kind of problem. I won’t include all of the output in this post, but this works fine and gives sensible results. However, it does not address the variable selection problem.

Basic variable selection

Let’s now modify the above script to do basic variable selection in the style of Kuo and Mallick.

data=list(y=y,X=X,n=n,p=p)
init=list(tau=1,alpha=0,betaT=rep(0,p),ind=rep(0,p))
modelstring="
  model {
    for (i in 1:n) {
      mean[i]<-alpha+inprod(X[i,],beta)
      y[i]~dnorm(mean[i],tau)
    }
    for (j in 1:p) {
      ind[j]~dbern(0.2)
      betaT[j]~dnorm(0,0.001)
      beta[j]<-ind[j]*betaT[j]
    }
    alpha~dnorm(0,0.0001)
    tau~dgamma(1,0.001)
  }
"
model=jags.model(textConnection(modelstring),
				data=data,inits=init)
update(model,n.iter=1000)
output=coda.samples(model=model,
		variable.names=c("alpha","beta","ind","tau"),
		n.iter=10000,thin=1)
print(summary(output))
plot(output)

Note that I’ve hard-coded an expectation that around 20% of variables should be included in the model. Again, I won’t include all of the output here, but the posterior mean of the indicator variables can be interpreted as posterior probabilities that the variables should be included in the model. Inspecting the output then reveals that the first three variables have a posterior probability of very close to one, the 4th variable has a small but non-negligible probability of inclusion, and the other variables all have very small probabilities of inclusion.

This is fine so far as it goes, but is not entirely satisfactory. One problem is that the choice of a “fixed effects” prior for the regression coefficients of the included variables is likely to lead to a Lindley’s paradox type situation, and a consequent under-selection of variables. It is arguably better to model the distribution of included variables using a “random effects” approach, leading to a more appropriate distribution for the included variables.

Variable selection with random effects

Adopting a random effects distribution for the included coefficients that is normal with mean zero and unknown variance helps to combat Lindley’s paradox, and can be implemented as follows.

data=list(y=y,X=X,n=n,p=p)
init=list(tau=1,taub=1,alpha=0,betaT=rep(0,p),ind=rep(0,p))
modelstring="
  model {
    for (i in 1:n) {
      mean[i]<-alpha+inprod(X[i,],beta)
      y[i]~dnorm(mean[i],tau)
    }
    for (j in 1:p) {
      ind[j]~dbern(0.2)
      betaT[j]~dnorm(0,taub)
      beta[j]<-ind[j]*betaT[j]
    }
    alpha~dnorm(0,0.0001)
    tau~dgamma(1,0.001)
    taub~dgamma(1,0.001)
  }
"
model=jags.model(textConnection(modelstring),
				data=data,inits=init)
update(model,n.iter=1000)
output=coda.samples(model=model,
		variable.names=c("alpha","beta","ind","tau","taub"),
		n.iter=10000,thin=1)
print(summary(output))
plot(output)

This leads to a large inclusion probability for the 4th variable, and non-negligible inclusion probabilities for the next few (it is obviously somewhat dependent on the simulated data set). This random effects variable selection modelling approach generally performs better, but it still has the potentially undesirable feature of hard-coding the probability of variable inclusion. Under the prior model, the number of variables included is binomial, and the binomial distribution is rather concentrated about its mean. Where there is a general desire to control the degree of sparsity in the model, this is a good thing, but if there is considerable uncertainty about the degree of sparsity that is anticipated, then a more flexible model may be desirable.

Variable selection with random effects and a prior on the inclusion probability

The previous model can be modified by introducing a Beta prior for the model inclusion probability. This induces a distribution for the number of included variables which has longer tails than the binomial distribution, allowing the model to learn about the degree of sparsity.

data=list(y=y,X=X,n=n,p=p)
init=list(tau=1,taub=1,pind=0.5,alpha=0,betaT=rep(0,p),ind=rep(0,p))
modelstring="
  model {
    for (i in 1:n) {
      mean[i]<-alpha+inprod(X[i,],beta)
      y[i]~dnorm(mean[i],tau)
    }
    for (j in 1:p) {
      ind[j]~dbern(pind)
      betaT[j]~dnorm(0,taub)
      beta[j]<-ind[j]*betaT[j]
    }
    alpha~dnorm(0,0.0001)
    tau~dgamma(1,0.001)
    taub~dgamma(1,0.001)
    pind~dbeta(2,8)
  }
"
model=jags.model(textConnection(modelstring),
				data=data,inits=init)
update(model,n.iter=1000)
output=coda.samples(model=model,
		variable.names=c("alpha","beta","ind","tau","taub","pind"),
		n.iter=10000,thin=1)
print(summary(output))
plot(output)

It turns out that for this particular problem the posterior distribution is not very different to the previous case, as for this problem the hard-coded choice of 20% is quite consistent with the data. However, the variable inclusion probabilities can be rather sensitive to the choice of hard-coded proportion.

Conclusion

Bayesian variable selection (and model selection more generally) is a very delicate topic, and there is much more to say about it. In this post I’ve concentrated on the practicalities of introducing variable selection into JAGS models. For further reading, I highly recommend the review of O’Hara and Sillanpaa (2009), which discusses other computational algorithms for variable selection. I intend to discuss some of the other methods in future posts.

References

O’Hara, R. and Sillanpaa, M. (2009) A review of Bayesian variable selection methods: what, how and which. Bayesian Analysis, 4(1):85-118. [DOI, PDF, Supp, BUGS Code]

Kuo, L. and Mallick, B. (1998) Variable selection for regression models. Sankhya B, 60(1):65-81.

Inlining JAGS models in R scripts for rjags

JAGS (Just Another Gibbs Sampler) is a general purpose MCMC engine similar to WinBUGS and OpenBUGS. I have a slight preference for JAGS as it is free and portable, works well on Linux, and interfaces well with R. It is tempting to write a tutorial introduction to JAGS and the corresponding R package, rjags, but there is a lot of material freely available on-line already, so it isn’t really necessary. If you are new to JAGS, I suggest starting with Getting Started with JAGS, rjags, and Bayesian Modelling. In this post I want to focus specifically on the problem of inlining JAGS models in R scripts as it can be very useful, and is usually skipped in introductory material.

JAGS and rjags on Ubuntu Linux

On recent versions of Ubuntu, assuming that R is already installed, the simplest way to install JAGS and rjags is using the command

sudo apt-get install jags r-cran-rjags

Now rjags is a CRAN package, so it can be installed in the usual way with install.packages("rjags"). However, taking JAGS and rjags direct from the Ubuntu repos should help to ensure that the versions of JAGS and rjags are in sync, which is a good thing.

Toy model

For this post, I will use a trivial toy example of inference for the mean and precision of a normal random sample. That is, we will assume data

with priors on and of the form

Separate model file

The usual way to fit this model in R using rjags is to first create a separate file containing the model

  model {
    for (i in 1:n) {
      x[i]~dnorm(mu,tau)
    }
    mu~dnorm(cc,d)
    tau~dgamma(a,b)
  }

Then, supposing that this file is called jags1.jags, an R session to fit the model could be constructed as follows:

require(rjags)
x=rnorm(15,25,2)
data=list(x=x,n=length(x))
hyper=list(a=3,b=11,cc=10,d=1/100)
init=list(mu=0,tau=1)
model=jags.model("jags1.jags",data=append(data,hyper), inits=init)
update(model,n.iter=100)
output=coda.samples(model=model,variable.names=c("mu", "tau"), n.iter=10000, thin=1)
print(summary(output))
plot(output)

This is all fine, and it can be very useful to have the model declared in a separate file, especially if the model is large and complex, and you might want to use it from outside R. However, very often for simple models it can be quite inconvenient to have the model separate from the R script which runs it. In particular, people often have issues with naming files correctly, making sure R is looking in the correct directory, moving the model with the R script, etc. So it would be nice to be able to just inline the JAGS model within an R script, to keep the model, the data, and the analysis all together in one place.

Using a temporary file

What we want to do is declare the JAGS model within a text string inside an R script and then somehow pass this into the call to jags.model(). The obvious way to do this is to write the string to a text file, and then pass the name of that text file into jags.model(). This works fine, but some care needs to be taken to make sure this works in a generic platform independent way. For example, you need to write to a file that you know doesn’t exist in a directory that is writable using a filename that is valid on the OS on which the script is being run. For this purpose R has an excellent little function called tempfile() which solves exactly this naming problem. It should always return the name of a file which does not exist in a writable directly within the standard temporary file location on the OS on which R is being run. This function is exceedingly useful for all kinds of things, but doesn’t seem to be very well known by newcomers to R. Using this we can construct a stand-alone R script to fit the model as follows:

require(rjags)
x=rnorm(15,25,2)
data=list(x=x,n=length(x))
hyper=list(a=3,b=11,cc=10,d=1/100)
init=list(mu=0,tau=1)
modelstring="
  model {
    for (i in 1:n) {
      x[i]~dnorm(mu,tau)
    }
    mu~dnorm(cc,d)
    tau~dgamma(a,b)
  }
"
tmpf=tempfile()
tmps=file(tmpf,"w")
cat(modelstring,file=tmps)
close(tmps)
model=jags.model(tmpf,data=append(data,hyper), inits=init)
update(model,n.iter=100)
output=coda.samples(model=model,variable.names=c("mu", "tau"), n.iter=10000, thin=1)
print(summary(output))
plot(output)

Now, although there is a file containing the model temporarily involved, the script is stand-alone and portable.

Using a text connection

The solution above works fine, but still involves writing a file to disk and reading it back in again, which is a bit pointless in this case. We can solve this by using another under-appreciated R function, textConnection(). Many R functions which take a file as an argument will work fine if instead passed a textConnection object, and the rjags function jags.model() is no exception. Here, instead of writing the model string to disk, we can turn it into a textConnection object and then pass that directly into jags.model() without ever actually writing the model file to disk. This is faster, neater and cleaner. An R session which takes this approach is given below.

require(rjags)
x=rnorm(15,25,2)
data=list(x=x,n=length(x))
hyper=list(a=3,b=11,cc=10,d=1/100)
init=list(mu=0,tau=1)
modelstring="
  model {
    for (i in 1:n) {
      x[i]~dnorm(mu,tau)
    }
    mu~dnorm(cc,d)
    tau~dgamma(a,b)
  }
"
model=jags.model(textConnection(modelstring), data=append(data,hyper), inits=init)
update(model,n.iter=100)
output=coda.samples(model=model,variable.names=c("mu", "tau"), n.iter=10000, thin=1)
print(summary(output))
plot(output)

This is my preferred way to use rjags. Note again that textConnection objects have many and varied uses and applications that have nothing to do with rjags.

The particle marginal Metropolis-Hastings (PMMH) particle MCMC algorithm

Introduction

In the previous post I explained how one can use an unbiased estimate of marginal likelihood derived from a particle filter within a Metropolis-Hastings MCMC algorithm in order to construct an exact pseudo-marginal MCMC scheme for the posterior distribution of the model parameters given some time course data. This idea is closely related to that of the particle marginal Metropolis-Hastings (PMMH) algorithm of Andreiu et al (2010), but not really exactly the same. This is because for a Bayesian model with parameters , latent variables and data , of the form

the pseudo-marginal algorithm which exploits the fact that the particle filter’s estimate of likelihood is unbiased is an MCMC algorithm which directly targets the marginal posterior distribution . On the other hand, the PMMH algorithm is an MCMC algorithm which targets the full joint posterior distribution . Now, the PMMH scheme does reduce to the pseudo-marginal scheme if samples of are not generated and stored in the state of the Markov chain, and it certainly is the case that the pseudo-marginal algorithm gives some insight into why the PMMH algorithm works. However, the PMMH algorithm is much more powerful, as it solves the “smoothing” and parameter estimation problem simultaneously and exactly, including the “initial value” problem (computing the posterior distribution of the initial state, ). Below I will describe the algorithm and explain why it works, but first it is necessary to understand the relationship between marginal, joint and “likelihood-free” MCMC updating schemes for such latent variable models.

MCMC for latent variable models

Marginal approach

If we want to target directly, we can use a Metropolis-Hastings scheme with a fairly arbitrary proposal distribution for exploring , where a new is proposed from and accepted with probability , where

As previously discussed, the problem with this scheme is that the marginal likelihood required in the acceptance ratio is often difficult to compute.

Likelihood-free MCMC

A simple “likelihood-free” scheme targets the full joint posterior distribution . It works by exploiting the fact that we can often simulate from the model for the latent variables even when we can’t evaluate it, or marginalise out of the problem. Here the Metropolis-Hastings proposal is constructed in two stages. First, a proposed new is sampled from and then a corresponding is simulated from the model . The pair is then jointly accepted with ratio

The proposal mechanism ensures that the proposed is consistent with the proposed , and so the procedure can work provided that the dimension of the data is low. However, in order to work well more generally, we would want the proposed latent variables to be consistent with the data as well as the model parameters.

Ideal joint update

Motivated by the likelihood-free scheme, we would really like to target the joint posterior by first proposing from and then a corresponding from the conditional distribution . The pair is then jointly accepted with ratio

Notice how the acceptance ratio simplifies, using the basic marginal likelihood identity (BMI) of Chib (1995), and drops out of the ratio completely in order to give exactly the ratio used for the marginal updating scheme. Thus, the “ideal” joint updating scheme reduces to the marginal updating scheme if is not sampled and stored as a component of the Markov chain.

Understanding the relationship between these schemes is useful for understanding the PMMH algorithm. Indeed, we will see that the “ideal” joint updating scheme (and the marginal scheme) corresponds to PMMH using infinitely many particles in the particle filter, and that the likelihood-free scheme corresponds to PMMH using exactly one particle in the particle filter. For an intermediate number of particles, the PMMH scheme is a compromise between the “ideal” scheme and the “blind” likelihood-free scheme, but is always likelihood-free (when used with a bootstrap particle filter) and always has an acceptance ratio leaving the exact posterior invariant.

The PMMH algorithm

The algorithm

The PMMH algorithm is an MCMC algorithm for state space models jointly updating and , as the algorithms above. First, a proposed new is generated from a proposal , and then a corresponding is generated by running a bootstrap particle filter (as described in the previous post, and below) using the proposed new model parameters, , and selecting a single trajectory by sampling once from the final set of particles using the final set of weights. This proposed pair is accepted using the Metropolis-Hastings ratio

where is the particle filter’s (unbiased) estimate of marginal likelihood, described in the previous post, and below. Note that this approach tends to the perfect joint/marginal updating scheme as the number of particles used in the filter tends to infinity. Note also that for a single particle, the particle filter just blindly forward simulates from and that the filter’s estimate of marginal likelihood is just the observed data likelihood leading precisely to the simple likelihood-free scheme. To understand for an arbitrary finite number of particles, , one needs to think carefully about the structure of the particle filter.

Why it works

To understand why PMMH works, it is necessary to think about the joint distribution of all random variables used in the bootstrap particle filter. To this end, it is helpful to re-visit the particle filter, thinking carefully about the resampling and propagation steps.

First introduce notation for the “particle cloud”: , , . Initialise the particle filter with , where and (note that is undefined). Now suppose at time we have a sample from : . First resample by sampling , . Here we use for the discrete distribution on with probability mass function . Next sample . Set and . Finally, propagate to the next step… We define the filter’s estimate of likelihood as and . See Doucet et al (2001) for further theoretical background on particle filters and SMC more generally.

Describing the filter carefully as above allows us to write down the joint density of all random variables in the filter as

For PMMH we also sample a final index from giving the joint density

We write the final selected trajectory as

where , and . If we now think about the structure of the PMMH algorithm, our proposal on the space of all random variables in the problem is in fact

and by considering the proposal and the acceptance ratio, it is clear that detailed balance for the chain is satisfied by the target with density proportional to

We want to show that this target marginalises down to the correct posterior when we consider just the parameters and the selected trajectory. But if we consider the terms in the joint distribution of the proposal corresponding to the trajectory selected by , this is given by

which, by expanding the in terms of the unnormalised weights, simplifies to

It is worth dwelling on this result, as this is the key insight required to understand why the PMMH algorithm works. The whole point is that the terms in the joint density of the proposal corresponding to the selected trajectory exactly represent the required joint distribution modulo a couple of normalising constants, one of which is the particle filter’s estimate of marginal likelihood. Thus, by including in the acceptance ratio, we knock out the normalising constant, allowing all of the other terms in the proposal to be marginalised away. In other words, the target of the chain can be written as proportional to

The other terms are all probabilities of random variables which do not occur elsewhere in the target, and hence can all be marginalised away to leave the correct posterior

Thus the PMMH algorithm targets the correct posterior for any number of particles, . Also note the implied uniform distribution on the selected indices in the target.

I will give some code examples in a future post.

Getting started with parallel MCMC

Introduction to parallel MCMC for Bayesian inference, using C, MPI, the GSL and SPRNG

Introduction

This post is aimed at people who already know how to code up Markov Chain Monte Carlo (MCMC) algorithms in C, but are interested in how to parallelise their code to run on multi-core machines and HPC clusters. I discussed different languages for coding MCMC algorithms in a previous post. The advantage of C is that it is fast, standard and has excellent scientific library support. Ultimately, people pursuing this route will be interested in running their code on large clusters of fast servers, but for the purposes of development and testing, this really isn’t necessary. A single dual-core laptop, or similar, is absolutely fine. I develop and test on a dual-core laptop running Ubuntu linux, so that is what I will assume for the rest of this post.

There are several possible environments for parallel computing, but I will focus on the Message-Passing Interface (MPI). This is a well-established standard for parallel computing, there are many implementations, and it is by far the most commonly used high performance computing (HPC) framework today. Even if you are ultimately interested in writing code for novel architectures such as GPUs, learning the basics of parallel computation using MPI will be time well spent.

MPI

The whole point of MPI is that it is a standard, so code written for one implementation should run fine with any other. There are many implementations. On Linux platforms, the most popular are OpenMPI, LAM, and MPICH. There are various pros and cons associated with each implementation, and if installing on a powerful HPC cluster, serious consideration should be given to which will be the most beneficial. For basic development and testing, however, it really doesn’t matter which is used. I use OpenMPI on my Ubuntu laptop, which can be installed with a simple:

sudo apt-get install openmpi-bin libopenmpi-dev

That’s it! You’re ready to go… You can test your installation with a simple “Hello world” program such as:

#include <stdio.h>
#include <mpi.h>

int main (int argc,char **argv)
{
  int rank, size;
  MPI_Init (&argc, &argv);
  MPI_Comm_rank (MPI_COMM_WORLD, &rank);
  MPI_Comm_size (MPI_COMM_WORLD, &size);	
  printf( "Hello world from process %d of %d\n", rank, size );
  MPI_Finalize();
  return 0;
}

You should be able to compile this with

mpicc -o helloworld helloworld.c

and run (on 2 processors) with

mpirun -np 2 helloworld

GSL

If you are writing non-trivial MCMC codes, you are almost certainly going to need to use a sophisticated math library and associated random number generation (RNG) routines. I typically use the GSL. On Ubuntu, the GSL can be installed with:

sudo apt-get install gsl-bin libgsl0-dev

A simple script to generate some non-uniform random numbers is given below.

#include <gsl/gsl_rng.h>
#include <gsl/gsl_randist.h>

int main(void)
{
  int i; double z; gsl_rng *r;
  r = gsl_rng_alloc(gsl_rng_mt19937);
  gsl_rng_set(r,0);
  for (i=0;i<10;i++) {
    z = gsl_ran_gaussian(r,1.0);
    printf("z(%d) = %f\n",i,z);
  }
  exit(EXIT_SUCCESS);
}

This can be compiled with a command like:

gcc gsl_ran_demo.c -o gsl_ran_demo -lgsl -lgslcblas

and run with

./gsl_ran_demo

SPRNG

When writing parallel Monte Carlo codes, it is important to be able to use independent streams of random numbers on each processor. Although it is tempting to “fudge” things by using a random number generator with a different seed on each processor, this does not guarantee independence of the streams, and an unfortunate choice of seeds could potentially lead to bad behaviour of your algorithm. The solution to this problem is to use a parallel random number generator (PRNG), designed specifically to give independent streams on different processors. Unfortunately the GSL does not have native support for such parallel random number generators, so an external library should be used. SPRNG 2.0 is a popular choice. SPRNG is designed so that it can be used with MPI, but also that it does not have to be. This is an issue, as the standard binary packages distributed with Ubuntu (libsprng2, libsprng2-dev) are compiled without MPI support. If you are going to be using SPRNG with MPI, things are simpler with MPI support, so it makes sense to download sprng2.0b.tar.gz from the SPRNG web site, and build it from source, carefully following the instructions for including MPI support. If you are not familiar with building libraries from source, you may need help from someone who is.

Once you have compiled SPRNG with MPI support, you can test it with the following code:

#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>
#define SIMPLE_SPRNG
#define USE_MPI
#include "sprng.h"

int main(int argc,char *argv[])
{
  double rn; int i,k;
  MPI_Init(&argc,&argv);
  MPI_Comm_rank(MPI_COMM_WORLD,&k);
  init_sprng(DEFAULT_RNG_TYPE,0,SPRNG_DEFAULT);
  for (i=0;i<10;i++)
  {
    rn = sprng();
    printf("Process %d, random number %d: %f\n", k, i+1, rn);
  }
  MPI_Finalize();
  exit(EXIT_SUCCESS);
}

which can be compiled with a command like:

mpicc -I/usr/local/src/sprng2.0/include -L/usr/local/src/sprng2.0/lib -o sprng_demo sprng_demo.c -lsprng -lgmp

You will need to edit paths here to match your installation. If if builds, it can be run on 2 processors with a command like:

mpirun -np 2 sprng_demo

If it doesn’t build, there are many possible reasons. Check the error messages carefully. However, if the compilation fails at the linking stage with obscure messages about not being able to find certain SPRNG MPI functions, one possibility is that the SPRNG library has not been compiled with MPI support.

The problem with SPRNG is that it only provides a uniform random number generator. Of course we would really like to be able to use the SPRNG generator in conjunction with all of the sophisticated GSL routines for non-uniform random number generation. Many years ago I wrote a small piece of code to accomplish this, gsl-sprng.h. Download this and put it in your include path for the following example:

#include <mpi.h>
#include <gsl/gsl_rng.h>
#include "gsl-sprng.h"
#include <gsl/gsl_randist.h>

int main(int argc,char *argv[])
{
  int i,k,po; gsl_rng *r;
  MPI_Init(&argc,&argv);
  MPI_Comm_rank(MPI_COMM_WORLD,&k);
  r=gsl_rng_alloc(gsl_rng_sprng20);
  for (i=0;i<10;i++)
  {
    po = gsl_ran_poisson(r,2.0);
    printf("Process %d, random number %d: %d\n", k, i+1, po);
  }
  MPI_Finalize();
  exit(EXIT_SUCCESS);
}

A new GSL RNG, gsl_rng_sprng20 is created, by including gsl-sprng.h immediately after gsl_rng.h. If you want to set a seed, do so in the usual GSL way, but make sure to set it to be the same on each processor. I have had several emails recently from people who claim that gsl-sprng.h “doesn’t work”. All I can say is that it still works for me! I suspect the problem is that people are attempting to use it with a version of SPRNG without MPI support. That won’t work… Check that the previous SPRNG example works, first.

I can compile and run the above code with

mpicc -I/usr/local/src/sprng2.0/include -L/usr/local/src/sprng2.0/lib -o gsl-sprng_demo gsl-sprng_demo.c -lsprng -lgmp -lgsl -lgslcblas
mpirun -np 2 gsl-sprng_demo

Parallel Monte Carlo

Now we have parallel random number streams, we can think about how to do parallel Monte Carlo simulations. Here is a simple example:

#include <math.h>
#include <mpi.h>
#include <gsl/gsl_rng.h>
#include "gsl-sprng.h"

int main(int argc,char *argv[])
{
  int i,k,N; double u,ksum,Nsum; gsl_rng *r;
  MPI_Init(&argc,&argv);
  MPI_Comm_size(MPI_COMM_WORLD,&N);
  MPI_Comm_rank(MPI_COMM_WORLD,&k);
  r=gsl_rng_alloc(gsl_rng_sprng20);
  for (i=0;i<10000;i++) {
    u = gsl_rng_uniform(r);
    ksum += exp(-u*u);
  }
  MPI_Reduce(&ksum,&Nsum,1,MPI_DOUBLE,MPI_SUM,0,MPI_COMM_WORLD);
  if (k == 0) {
    printf("Monte carlo estimate is %f\n", (Nsum/10000)/N );
  }
  MPI_Finalize();
  exit(EXIT_SUCCESS);
}

which calculates a Monte Carlo estimate of the integral

using 10k variates on each available processor. The MPI command MPI_Reduce is used to summarise the values obtained independently in each process. I compile and run with

mpicc -I/usr/local/src/sprng2.0/include -L/usr/local/src/sprng2.0/lib -o monte-carlo monte-carlo.c -lsprng -lgmp -lgsl -lgslcblas
mpirun -np 2 monte-carlo

Parallel chains MCMC

Once parallel Monte Carlo has been mastered, it is time to move on to parallel MCMC. First it makes sense to understand how to run parallel MCMC chains in an MPI environment. I will illustrate this with a simple Metropolis-Hastings algorithm to sample a standard normal using uniform proposals, as discussed in a previous post. Here an independent chain is run on each processor, and the results are written into separate files.

#include <gsl/gsl_rng.h>
#include "gsl-sprng.h"
#include <gsl/gsl_randist.h>
#include <mpi.h>

int main(int argc,char *argv[])
{
  int k,i,iters; double x,can,a,alpha; gsl_rng *r;
  FILE *s; char filename[15];
  MPI_Init(&argc,&argv);
  MPI_Comm_rank(MPI_COMM_WORLD,&k);
  if ((argc != 3)) {
    if (k == 0)
      fprintf(stderr,"Usage: %s <iters> <alpha>\n",argv[0]);
    MPI_Finalize(); return(EXIT_FAILURE);
  }
  iters=atoi(argv[1]); alpha=atof(argv[2]);
  r=gsl_rng_alloc(gsl_rng_sprng20);
  sprintf(filename,"chain-%03d.tab",k);
  s=fopen(filename,"w");
  if (s==NULL) {
    perror("Failed open");
    MPI_Finalize(); return(EXIT_FAILURE);
  }
  x = gsl_ran_flat(r,-20,20);
  fprintf(s,"Iter X\n");
  for (i=0;i<iters;i++) {
    can = x + gsl_ran_flat(r,-alpha,alpha);
    a = gsl_ran_ugaussian_pdf(can) / gsl_ran_ugaussian_pdf(x);
    if (gsl_rng_uniform(r) < a)
      x = can;
    fprintf(s,"%d %f\n",i,x);
  }
  fclose(s);
  MPI_Finalize(); return(EXIT_SUCCESS);
}

I can compile and run this with the following commands

mpicc -I/usr/local/src/sprng2.0/include -L/usr/local/src/sprng2.0/lib -o mcmc mcmc.c -lsprng -lgmp -lgsl -lgslcblas
mpirun -np 2 mcmc 100000 1

Parallelising a single MCMC chain

The parallel chains approach turns out to be surprisingly effective in practice. Obviously the disadvantage of that approach is that “burn in” has to be repeated on every processor, which limits how much efficiency gain can be acheived by running across many processors. Consequently it is often desirable to try and parallelise a single MCMC chain. As MCMC algorithms are inherently sequential, parallelisation is not completely trivial, and most (but not all) approaches to parallelising a single MCMC chain focus on the parallelisation of each iteration. In order for this to be worthwhile, it is necessary that the problem being considered is non-trivial, having a large state space. The strategy is then to divide the state space into “chunks” which can be updated in parallel. I don’t have time to go through a real example in detail in this blog post, but fortunately I wrote a book chapter on this topic almost 10 years ago which is still valid and relevant today. The citation details are:

Wilkinson, D. J. (2005) Parallel Bayesian Computation, Chapter 16 in E. J. Kontoghiorghes (ed.) Handbook of Parallel Computing and Statistics, Marcel Dekker/CRC Press, 481-512.

The book was eventually published in 2005 after a long delay. The publisher which originally commisioned the handbook (Marcel Dekker) was taken over by CRC Press before publication, and the project lay dormant for a couple of years until the new publisher picked it up again and decided to proceed with publication. I have a draft of my original submission in PDF which I recommend reading for further information. The code examples used are also available for download, including several of the examples used in this post, as well as an extended case study on parallelisation of a single chain for Bayesian inference in a stochastic volatility model. Although the chapter is nearly 10 years old, the issues discussed are all still remarkably up-to-date, and the code examples all still work. I think that is a testament to the stability of the technology adopted (C, MPI, GSL). Some of the other handbook chapters have not stood the test of time so well.

For basic information on getting started with MPI and key MPI commands for implementing parallel MCMC algorithms, the above mentioned book chapter is a reasonable place to start. Read it all through carefully, run the examples, and carefully study the code for the parallel stochastic volatility example. Once that is understood, you should find it possible to start writing your own parallel MCMC algorithms. For further information about more sophisticated MPI usage and additional commands, I find the annotated specification: MPI – The complete reference to be as good a source as any.

Metropolis-Hastings MCMC algorithms

Background

Over the next few months I’m intending to have a series of posts on recent developments in MCMC and algorithms of Metropolis-Hastings type. These posts will assume a basic familiarity with stochastic simulation and R. For reference, I have some old notes on stochastic simulation and MCMC from a course I used to teach. There were a set of computer practicals associated with the course which focused on understanding the algorithms, mainly using R. Indeed, this post is going to make use of the example from this old practical. If anything about the rest of this post doesn’t make sense, you might need to review some of this background material. My favourite introductory text book on this subject is Gamerman’s Markov Chain Monte Carlo. In this post I’m going to briefly recap the key idea behind the Metropolis-Hastings algorithm, and illustrate how these kinds of algorithms are typically implemented. In order to keep things as simple as possible, I’m going to use R for implementation, but as discussed in a previous post, that probably won’t be a good idea for non-trivial applications.

Metropolis-Hastings

The motivation is a need to understand a complex probability distribution, p(x). In applications to Bayesian statistics, this distribution is usually a posterior from a Bayesian analysis. A good way to understand an intractable distribution is to simulate realisations from it and study those. However, this often isn’t easy, either. The idea behind MCMC is to simulate a Markov chain whose equilibrium distribution is p(x). Metropolis-Hastings (M-H) provides a way to correct a fairly arbitrary transition kernel q(x’|x) (which typically won’t have p(x) as its equilibrium) to give a chain which does have the desired target. In M-H, the transition kernel is used to generate a proposed new value, x’ for the chain, which is then accepted as the new state at random with a particular probability a(x’|x)=min(1,A), where A = p(x’)q(x|x’)/[p(x)q(x’|x)].

If the value is accepted, then the chain moves to the proposed new state, x’. If the value is not accepted, the chain still advances to the next step, but the new state is given by the previous state of the chain, x.

It is clear that this algorithm is also a Markov chain, and that for x’ != x, the transition kernel of this chain is q(x’|x)a(x’|x). A couple of lines of algebra will then confirm that this “corrected” Markov chain satisfies detailed balance for the target p(x). Therefore, under some regularity conditions, this chain will have p(x) as its equilibrium, and this gives a convenient way of simulating values from this target.

Special case: the Metropolis method

It is clear from the form of the acceptance probability that a considerable simplification arises in the case where the proposal kernel is symmetric, having q(x’|x)=q(x|x’). This is useful, as a convenient updating strategy is often to update the current value of x by adding an innovation sampled from a distribution that is symmetric about zero. This clearly gives a symmetric kernel, which then drops out of the acceptance ratio, to give A=p(x’)/p(x).

Example

The example we will use for illustration is a Metropolis method for the standard normal distribution using innovations from a U(-eps,eps) distribution. Below is a simple R implementation of the algorithm

metrop1=function(n=1000,eps=0.5) 
{
        vec=vector("numeric", n)
        x=0
        vec[1]=x
        for (i in 2:n) {
                innov=runif(1,-eps,eps)
                can=x+innov
                aprob=min(1,dnorm(can)/dnorm(x))
                u=runif(1)
                if (u < aprob) 
                        x=can
                vec[i]=x
        }
        vec
}

First a candidate value (can) is constructed by perturbing the current state of the chain with the uniform innovation. Then the acceptance probability is computed, and the chain is updated appropriately depending on whether the proposed new value is accepted. Note the standard trick of picking an event with probability p by checking to see if u<p, where u is a draw from a U(0,1).

We can execute this code and plot the results with the following peice of code:

plot.mcmc<-function(mcmc.out)
{
	op=par(mfrow=c(2,2))
	plot(ts(mcmc.out),col=2)
	hist(mcmc.out,30,col=3)
	qqnorm(mcmc.out,col=4)
	abline(0,1,col=2)
	acf(mcmc.out,col=2,lag.max=100)
	par(op)
}

metrop.out<-metrop1(10000,1)
plot.mcmc(metrop.out)

Plots showing convergence of the Markov chain to a N(0,1) equilibrium distribution

From these plots we see that the algorithm seems to be working as expected. Before finishing this post, it is worth explaining how to improve this code to get something which looks a bit more like the kind of code that people would actually write in practice. The next version of the code (below) makes use of the fact that min(1,A) is redundant if you are just going to compare to a uniform (since values of the ratio larger than 1 will be accepted with probability 1, as they should), and also that it is unnecessary to recalculate the likelihood of the current value of the chain at every iteration – better to store and re-use the value. That obviously doesn’t make much difference here, but for real problems likelihood computations are often the main computational bottleneck.

metrop2=function(n=1000,eps=0.5) 
{
        vec=vector("numeric", n)
        x=0
        oldlik=dnorm(x)
        vec[1]=x
        for (i in 2:n) {
                innov=runif(1,-eps,eps)
                can=x+innov
                lik=dnorm(can)
                a=lik/oldlik
                u=runif(1)
                if (u < a) { 
                        x=can
                        oldlik=lik
                        }
                vec[i]=x
        }
        vec
}

However, this code still has a very serious flaw. It computes likelihoods! For this problem it isn’t a major issue, but in general likelihoods are the product of a very large number of small numbers, and numerical underflow is a constant hazard. For this reason (and others), no-one ever computes likelihoods in code if they can possibly avoid it, but instead log-likelihoods (which are the sum of a large number of reasonably-sized numbers, and therefore numerically very stable). We can use these log-likelihoods to calculate a log-acceptance ratio, which can then be compared to the log of a uniform for the accept-reject decision. We end up with the code below, which now doesn’t look too different to the kind of code one might actually write…

metrop3=function(n=1000,eps=0.5) 
{
        vec=vector("numeric", n)
        x=0
        oldll=dnorm(x,log=TRUE)
        vec[1]=x
        for (i in 2:n) {
                can=x+runif(1,-eps,eps)
                loglik=dnorm(can,log=TRUE)
                loga=loglik-oldll
                if (log(runif(1)) < loga) { 
                        x=can
                        oldll=loglik
                        }
                vec[i]=x
        }
        vec
}

It is easy to verify that all 3 of these implementations behave identically. Incidentally, most statisticians would consider all 3 of the above codes to represent exactly the same algorithm. They are just slightly different implementations of the same algorithm. I mention this because I’ve noticed that computing scientists often have a slightly lower-level view of what constitutes an algorithm, and would sometimes consider the above 3 algorithms to be different.

The last Valencia meeting on Bayesian Statistics and the future of Bayesian computation

I’ve spent the last week in Benidorm, Spain, for the 9th and final Valencia meeting on Bayesian Statistics. Nine of us travelled from Newcastle University, making us one of the best represented groups at the meeting. This was my fifth Valencia meeting – the first I attended was Valencia 5 which took place in Alicante back in 1994 when I was a PhD student at Durham University working with Michael Goldstein. Our contributed paper to the proceedings of that meeting was my first publication, and I’ve been to every meeting since. Michael is one of the few people to have attended all 9 Valencia meetings (in addition to the somewhat mythical “Valencia 0”). It was therefore very fitting that Michael opened the Valencia 9 invited programme (with a great talk on Bayesian analysis of complex computer models). Like many others, I’ll be a little sad to think that the Valencia meetings have come to an end.

The meeting itself was scientifically very interesting. I wish that I had the energy to give a summary of the scientific programme, but unfortunately I don’t! However, anyone who does want to get something of the flavour of the programme should take a look at the “Valencia Snapshots” on Christian Robert’s blog. My own talk gets a mention in Snapshot 4. I presented a paper entitled Parameter inference for stochastic kinetic models of bacterial gene regulation: a Bayesian approach to systems biology. Unfortunately my discussant, Sam Kou, was unable to travel to the meeting due to passport problems, but very kindly produced a pre-recorded video discussion to be played to me and the audience at the end of my talk. After a brief problem with the audio (a recurring theme of the meeting!), this actually worked quite well, though it felt slightly strange replying to his discussion knowing that he could not hear what I was saying!

There were several talks discussing Bayesian approaches to challenging problems in bioinformatics and molecular biology, and these were especially interesting to me. I was also particularly interested in the talks on Bayesian computation. Several talks mentioned the possibility of speeding up Bayesian computation using GPUs, and Chris Holmes gave a nice overview of the current technology and its potential, together with a link to a website providing further information. Although there is no doubt that GPU technology can provide fairly impressive speedups for certain Bayesian computations, I’m actually a little bit of a GPU-sceptic, so let me explain why. There are many reasons. First, I’m always a bit suspicious of a technology that is fairly closed and proprietary being pushed by a large powerful company – I prefer my hardware to be open, and my software to be free and open. Next, there isn’t really anything that you can do on a GPU that you can’t do on a decent multicore server or cluster using standard well established technologies such as MPI and OpenMP. Also, GPUs are relatively difficult to program, and time taken for software development is a very expensive cost which in many cases will dwarf differences in hardware costs. Also, in the days when 64 bit chips and many GB of RAM are sitting on everyone’s desktops, do I really want to go back to 1 GB of RAM, single precision arithmetic and no math libraries?! That hardly seems like the future I’m envisaging! Next, there are other related products like the Intel Knights Corner on the horizon that are likely to offer similar performance gains while being much simpler to develop for. Next, it seems likely to me that machines in the future are going to feature massively multicore CPUs, rendering GPU computing obsolete. Finally, although GPUs offer one possible approach to tackling the problem of speedup, they do little for the far more general and important problem of scalability of Bayesian computing and software. From that perspective, I really enjoyed the talk by Andrew McCallum on Probabilistic programming with imperatively-defined factor graphs. Andrew was talking about a flexible machine learning library he is developing called factorie for the interesting new language Scala. Whilst that particular library is not exactly what I need, fundamentally, his talk was all about building frameworks for Bayesian computation which really scale. I think this is the real big issue facing Bayesian computation, and modern languages and software platforms, including so-called cloud computing approaches, and technologies like Hadoop and MapReduce probably represent some of the directions we should be looking in. There is an interesting project called CRdata which is a first step in that direction. Clearly these technologies are somewhat orthogonal to the GPU/speedup issue, but I don’t think they are completely unrelated.