Hey, I don’t think I ever posted a link to this. It’s a discussion in the journal Statistics in Medicine of an article by David Lunn, David Spielhalter, Andrew Thomas ,and Nicky Best. (Sorry but I can’t find the Lunn et al. article online, or I’d link to it.) Anyway, here’s my discussion. Once upon a time . . .

I first saw BUGS in a demonstration version at a conference in 1991, but I didn’t take it seriously until over a decade later, when I found that some of my Ph.D. students in political science were using Bugs to fit their models. It turned out that Bugs’s modeling language was ideal for students who wanted to fit complex models but didn’t have a full grasp of the mathematics of likelihood functions, let alone Bayesian inference and integration. I also learned that the modular structure of BUGS was a great way for students, and researchers in general, to think more about modeling and less about deciding which conventional structure should be fit to data.

Since then, my enthusiasm for BUGS has waxed and waned, depending on what sorts of problems I was working on. For example, in our study of income and voting in U.S. states [1], my colleagues fit all our models in BUGS. Meanwhile we kept running into difficulty when we tried to expand our model in different ways, most notably when going from varying-intercept multilevel regressions, to varying-intercept, varying-slope regressions, to models with more than two varying coefficients per group. Around this time I discovered lmer [2], a function in R which fits multilevel linear and generalized linear models allowing for varying intercepts and slopes. The lmer function can have convergence problems and does not account for uncertainty in the variance parameters, but it is faster than Bugs and in many cases more reliable-so much so that Jennifer Hill and I retooled our book on multilevel models to foreground lmer and de-emphasize Bugs, using the latter more as a way of illustrating models than as a practical tool.

What does BUGS do best and what does it do worst?

In short, BUGS excels with complicated models for small datasets. For example, we fit a fairly elaborate discrete choice model to data from a laboratory economics experiment with 30 trials on each of about 100 participants [3]. The sample size was small enough that we wanted to use a hierarchical model to obtain stable inference for all the people, so we programmed something up in BUGS and it worked right away.

Well, not right away. We first had to devise a work-around because BUGS was crashing on a logistic model. We added a couple lines in the model to bound the probabilities between 0.01 and 0.99. As with many situations in which a specification is set up for computational reasons [4], it turned out that the model made statistical sense as well: bounding the probabilities allowed for the occasional outlier, an issue we had never previously thought about in the context of binary data. (see [5] for a cleaner solution to this problem using the t distribution in place of the logistic.) Anyway, the point is that in this example, BUGS worked well, it was the direct solution to a good answer, and even its problems led to a new solution which would actually have been difficult to implement in non-BUGS software.

When does BUGS not work so well? With large datasets, multivariate structures, and regression coefficients. For one thing, it can be difficult to keep track of regression coefficients in the BUGS modeling language, which allows no subroutines or macros.

For example, instead of:

for (i in 1:n){ y[i] ~ dnorm (y.hat[i], tau.y) y.hat[i] <- a[county[i]] + b[county[i]]*x[i] e.y[i] <- y[i] - y.hat[i] } tau.y <- pow(sigma.y, -2) sigma.y ~ dunif (0, 1000)we want something like:

y ~ norm (a[county] + b[county]*x, sigma.y)This is all algorithmic and could be programmed, but Bugs is not set up so as to allow this.

Convergence can be slow, and, in fact, each iteration can take a long time to run. When generalizing our model of income and voting to allow four coefficients for each state (corresponding to individual income, religious attendance, their interaction, and a constant term), BUGS basically ground to a halt, to the extent that I wouldn't have trusted its results, even had I decided to be patient and run it all night.

Winston Churchill said that sometimes the truth is so precious, it must be attended by a bodyguard of lies. Similarly, for a model to be believed, it must, except in the simplest of cases, be accompanied by similar models that either give similar results or, if they differ, do so in a way that can be understood. In the example of voting by income and religion, we simply took a step back and fit the model using lmer. This didn't always work either, but when lmer had problems we stripped down the model in various ways until it worked. It ran fast enough that we were able to experiment.

What are the great things about BUGS?

1. It really works! I can use it in my applied research. (For software to be useful to me in this way, it doesn't have to work all the time, it only has to solve some problems that can't easily be solved in other ways, and to break down in recognizable ways, so that it doesn't pretend to work when it doesn't.)

2. BUGS is easy to use and to teach, with intuitive syntax.

3. It is free.

4. It can be called directly from R (see [6] for many illustrations of the advantages of running BUGS in this way, most notably for preprocessing of data and postprocessing of inferences.)What are some problems with BUGS?

1. It often needs lots of hand-holding adjustment of the model to work.

2. Efficiently-programmed models can get really long, ugly, and bug-prone, as I have learned in implementing redundant parameterizations in hierarchical models [7].

3. You can't debug it by running interactively, as you can in R, Matlab, C, Fortran, etc.

4. There's no easy way to go inside and improve the sampling, either by writing shortcuts in the BUGS language or by inserting your own jumping rules inside the inference engine. Attempts to "trick" BUGS into using efficient updates can sometimes backfire.

5. As far as I can tell, BUGS does not understand hierarchical models but rather thinks of scalar parameters as individual entities. This may be partly a historical issue-before the mid-1990s, hierarchical modeling was generally considered a special topic rather than central to concepts of Bayesian inference and prior information-but, whatever the reason, the current implementation of BUGS does not allow easy setup or fast computation with hierarchical models.There is also a problem-not really the fault of BUGS itself-that its users typically focus on inference as the only goal, ignoring the other two stages of Bayesian data analysis: model building and model checking. Often it can take awhile to get the model to converge, and once that happens there's a tendency to just stop, relax, and present the results. I believe that an important future development should be automatic implementations of model checking using posterior simulations. This can be done-for example, by creating a "shadow" replicated variable for each variable in the model, and then allowing user-specified or default graphical comparisons-but, like everything else, it requires work.

I shouldn't really complain-after all, BUGS is free and, as noted in the article under discussion, has had an incredibly beneficial influence, especially considering that it was put together by just a few dedicated people. I hope that the next twenty years of BUGS are as successful as the first.

References

1. Gelman, A., Shor, B., Bafumi, J., and Park, D. Rich state, poor state, red state, blue state: What's the matter with Connecticut? Quarterly Journal of Political Science 2007; 2, 345-367.

2. Bates, D. Fitting linear models in R using the lme4 package. R News 2005; 5 (1), 27-30.

3. Casella, A., Gelman, A., and Palfrey, T. R. An experimental study of storable votes. Games and Economic Behavior 2005; 57, 123-154.

4. Gelman, A. Parameterization and Bayesian modeling. Journal of the American Statistical Association 2004; 99, 537-545.

5. Liu, C. Robit regression: a simple robust alternative to logistic and probit regression. In Applied Bayesian Modeling and Causal Inference from Incomplete-Data Perspectives, ed. A. Gelman and X. L. Meng, 2004; 227-238. London: Wiley.

6. Gelman, A., and Hill, J. Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press, 2007.

7. Gelman, A., Huang, Z., van Dyk, D., and Boscardin, W. J. Using redundant parameters to fit hierarchical models. Journal of Computational and Graphical Statistics 2007; 17, 95-122.

I had two questions on reading this:

1. Would all of these comments/criticisms apply also to Jags?

2. What's the alternative? Rolling your own in Mathematica?

What has happened to further BUGS development? It seems to have ground to a halt with WinBUGS 1.4.3 and OpenBUGS 3.0.3 (released in 2007). I was especially hopeful of further development following the move to OpenBUGS, but this doesn't seem to have occurred.

http://www.ncbi.nlm.nih.gov/pubmed/19630097

Lunn D, Spiegelhalter D, Thomas A, Best N. The BUGS project: Evolution, critique and future directions. Stat Med. 2009 Nov 10;28(25):3049-67.

Chris:

1. Yes. But I want to emphasize that most of my comments are positive, not critical.

2. For now, yes. Or maybe Matlab is better. Or using some approximation such as lmer(). Or maybe AD Model Builder or some other program that I've never tried.

Dana:

Good question. Jags has been moving forward a bit. But they're all on miniature budgets. I'd say this is a problem with open-source software, but seeing as there's no payware equivalent, I don't know.

Jimmy:

Yup, that's the article that the above entry is a discussion of.

a naive question, for the multilevel linear models, what is your comment on SAS proc mixed?

Say I have 80,000 level one units and 120 level two units. I take it this would be too much for BUGS? What about LMER?

Can anyone point me to a discussion of the advantages and disadvantages of various packages for multilevel modeling with large samples?

I like and have used BUGs a lot (WinBUGs, OpenBUGs, JAGs).

But I've recently started using pymc to the same end which has some handy features (like writing any samples/simulated data directly to a choice of "databases" – ranging from a text file to MySQL as well as some computational nicenesses I don't understand yet but look useful). On some relatively standard models (hierarchical logistic regression) it seems quite quick with large datasets.

http://code.google.com/p/pymc/

I doubt it has anything like the same userbase as BUGs, but it would be interesting to see a more formal head to head comparison.

I'm particularly interested in a comparison of HLM, STATA, SAS, and R for large datasets. Anything more obscure is probably less than ideal for a social scientist with limited programming background.

I'm not an expert, but from what I've read, SAS and Stata handle large datasets well. I've not heard of HLM. I use R and while there are packages to attempt to handle datasets larger than will fit in memory, it's not the same as if R could do so natively.

You need to define "large", though. With the current 64-bit R and a boatload of RAM, it would also probably handle things well. (Personally, I find SAS's 1960's punch card syntax repulsive, but it is the Name Brand in many circles and some people simply cannot get over R's

Dana:

OpenBUGS development is continuing at http://www.openbugs.info

The latest version is 3.0.6

Chris:

1) Yes, everything applies to JAGS. I want JAGS to be a tool for statistical modelling and not just a computational engine so I am paying close attention to Andrew's ideas on model criticism.

2) Most novel sampling methods seem to be prototyped in Matlab. Ideally, I want to see reference implementations of new methods in BUGS. Right now it is up to Andrew, or Dave or me to do the necessary programming, which is OK but is obviously a rate-limiting step.

Martyn,

Thanks so much for the OpenBUGS link. I, and a bunch of other people, were unaware development was continuing, and this is a great holiday gift!

Its not clear at all to me that OpenBUGS development is continuing. Look at the dates on those releases — nothing newer than 2007. Perhaps that is due to the choice of coding BUGS in Pascal? In any case, I think the future of MCMC software is elsewhere. JAGS, perhaps.

Hi all,

Just to keep you posted on what's been going on with OpenBUGS… There's been lots of work going on behind the scenes in preparation for a major new release. In particular, lots of people have been working on documentation, testing and reliability. The idea has been to produce something at least as reliable as WinBUGS — no jokes please ;-) — in the hope that we can convince most WinBUGS users to move over to OpenBUGS, since WinBUGS will not be further developed… We've also been working on a "Foundation" for BUGS, through which we can solicit, receive and disburse funds for development work. In short, we have been laying the foundations for seriously taking the project forwards, with a single source of code that can (potentially) be developed by many… Sorry for the recent lack of communication.

All the best,

Dave Lunn