Panayiota Touloupou (Warwick), Naif Alzahrani, Peter Neal, Simon Spencer (Warwick) and Trevelyan McKinley arXived a paper yesterday on Model comparison with missing data using MCMC and importance sampling, where they proposed an importance sampling strategy based on an early MCMC run to approximate the marginal likelihood a.k.a. the evidence. Another instance of estimating a constant. It is thus similar to our Frontier paper with Jean-Michel, as well as to the recent Pima Indian survey of James and Nicolas. The authors give the difficulty to calibrate reversible jump MCMC as the starting point to their research. The importance sampler they use is the natural choice of a Gaussian or t distribution centred at some estimate of θ and with covariance matrix associated with Fisher’s information. Or derived from the warmup MCMC run. The comparison between the different approximations to the evidence are done first over longitudinal epidemiological models. Involving 11 parameters in the example processed therein. The competitors to the 9 versions of importance samplers investigated in the paper are the raw harmonic mean [rather than our HPD truncated version], Chib’s, path sampling and RJMCMC [which does not make much sense when comparing two models]. But neither bridge sampling, nor nested sampling. Without any surprise (!) harmonic means do not converge to the right value, but more surprisingly Chib’s method happens to be less accurate than most importance solutions studied therein. It may be due to the fact that Chib’s approximation requires three MCMC runs and hence is quite costly. The fact that the mixture (or defensive) importance sampling [with 5% weight on the prior] did best begs for a comparison with bridge sampling, no? The difficulty with such study is obviously that the results only apply in the setting of the simulation, hence that e.g. another mixture importance sampler or Chib’s solution would behave differently in another model. In particular, it is hard to judge of the impact of the dimensions of the parameter and of the missing data.
Archive for nested sampling
Clément Walter from Paris just pointed me to an arXived paper he had very recently gotten accepted for publication in Statistics and Computing. (Congrats!) Because his paper relates to nested sampling. And connects it with rare event simulation via interacting particle systems. And multilevel Monte Carlo. I had missed it when it came out on arXiv last December [as the title was unrelated with nested sampling if not Monte Carlo], but the paper brings fairly interesting new results about an ideal version of nested sampling that is
- unbiased when using an infinite number of terms;
- always better than the standard Monte Carlo estimator, variance-wise;
- connected with an implicit marked Poisson process; and
- enjoying a finite variance provided the quantity of interest has an 1+ε moment.
Of course, such results only hold for an ideal version and do not address the issue of the conditional simulations required by nested sampling. (Which has an impact on the computing time as the conditional simulation becomes more and more expensive as the likelihood value increases.) The explanation therein of the approximation of tail probabilities by a Poisson estimate makes the link with deterministic nested sampling much clearer to me. Point 2 above means that the nested sampling estimate always does better than the average of the likelihood values produced by an iid
or MCMC simulation from the prior distribution. The paper also borrows from the debiasing approach of Rhee and Glynn (already used by the Russian roulette) to turn truncated versions of the nested sampling estimator into an unbiased estimator, with a limited impact on the variance of the estimator. Truncation is associated with the generation of a geometric stopping time which parameter needs to be optimised. Without a more detailed reading, I am somewhat lost as to this optimisation remains feasible in complex settings… The paper contains an illustration for a Pareto distribution where optimisation and calibration can be conducted quite far. It also re-analyses the Mexican hat example of Skilling (2006), showing that our stopping rule may induce bias.
.[Following my posting of a misfiled 2013 blog, Ewan Cameron told me of the impact of this paper in starting his own blog and I asked him for a guest post, resulting in this analysis, much deeper than mine. No warning necessary this time!]
Back in February 2013 when “Bayesian Model Averaging in Astrophysics: A Review” by Parkinson & Liddle (hereafter PL13) first appeared on the arXiv I was a keen, young(ish) postdoc eager to get stuck into debates about anything and everything ‘astro-statistical’. And with its seemingly glaring flaws, PL13 was more grist to the mill. However, despite my best efforts on various forums I couldn’t get a decent fight started over the right way to do model averaging (BMA) in astronomy, so out of sheer frustration two months later I made my own soapbox to shout from at Another Astrostatistics Blog. Having seen PL13 reviewed recently here on Xian’s Og it feels like the right time to revisit the subject and reflect on where BMA in astronomy is today.
As pointed out to me back in 2013 by Tom Loredo, the act of Bayesian model averaging has been around much longer than its name; indeed an early astronomical example appears in Gregory & Loredo (1992) in which the posterior mean representation of an unknown signal is constructed for an astronomical “light-curve”, averaging over a set of constant and periodic candidate models. Nevertheless the wider popularisation of model averaging in astronomy has only recently taken place through a variety of applications in cosmology: e.g. Liddle, Mukherjee, Parkinson & Wang (2006) and Vardanyan, Trotta & Silk (2011).
In contrast to earlier studies like Gregory & Loredo (1992)—or the classic review on BMA by Hoeting et al. (1999)—in which the target of model averaging is typically either a utility function, a set of future observations, or a latent parameter of the observational process (e.g. the unknown “light-curve” shape) shared naturally by all competing models, the proposal of cosmological BMA studies is to produce a model-averaged version of the posterior for a given ‘shared’ parameter: a so-called “model-averaged PDF”. This proposal didn’t sit well with me back in 2013, and it still doesn’t sit well with me today. Philosophically: without a model a parameter has no meaning, so why should we want to seek meaning in the marginalised distribution of a parameter over an entire set of models? And, practically: to put it another way, without knowing the model ‘label’ to which a given mass of model-averaged parameter probability belongs there’s nothing much useful we can do with this ‘PDF’: nothing much we can say about the data we’ve just analysed and nothing much we can say about future experiments. Whereas the space of the observed data is shared automatically by all competing models, it seems to me to be somehow “un-Bayesian” to place the further restriction that the parameters of separate models share the same scale and topology. I say “un-Bayesian” since this mode of model averaging suggests a formulation of the parameter space + prior pairing stronger than the statement of one’s prior beliefs for the distribution of observable data given the model. But I would be happy to hear arguments from the other side in the comments box below … ! Continue reading
A few months ago, Nick Polson and James Scott arXived a paper on one of my favourite problems, namely the approximation of normalising constants (and it went way under my radar, as I only became aware of it quite recently!, then it remained in my travel bag for an extra few weeks…). The method for approximating the constant Z draws from an analogy with the energy level sampling methods found in physics, like the Wang-Landau algorithm. The authors rely on a one-dimensional slice sampling representation of the posterior distribution and [main innovation in the paper] add a weight function on the auxiliary uniform. The choice of the weight function links the approach with the dreaded harmonic estimator (!), but also with power-posterior and bridge sampling. The paper recommends a specific weighting function, based on a “score-function heuristic” I do not get. Further, the optimal weight depends on intractable cumulative functions as in nested sampling. It would be fantastic if one could draw directly from the prior distribution of the likelihood function—rather than draw an x [from the prior or from something better, as suggested in our 2009 Biometrika paper] and transform it into L(x)—but as in all existing alternatives this alas is not the case. (Which is why I find the recommendations in the paper for practical implementation rather impractical, since, were the prior cdf of L(X) available, direct simulation of L(X) would be feasible. Maybe not the optimal choice though.)
“What is the distribution of the likelihood ordinates calculated via nested sampling? The answer is surprising: it is essentially the same as the distribution of likelihood ordinates by recommended weight function from Section 4.”
The approach is thus very much related to nested sampling, at least in spirit. As the authors later demonstrate, nested sampling is another case of weighting, Both versions require simulations under truncated likelihood values. Albeit with a possibility of going down [in likelihood values] with the current version. Actually, more weighting could prove [more] efficient as both the original nested and vertical sampling simulate from the prior under the likelihood constraint. Getting away from the prior should help. (I am quite curious to see how the method is received and applied.)
This morning, in the train to Dauphine (train that was even more delayed than usual!), I read a recent arXival of Brendon Brewer and Courtney Donovan. Entitled Fast Bayesian inference for exoplanet discovery in radial velocity data, the paper suggests to associate Matthew Stephens’ (2000) birth-and-death MCMC approach with nested sampling to infer about the number N of exoplanets in an exoplanetary system. The paper is somewhat sparse in its description of the suggested approach, but states that the birth-date moves involves adding a planet with parameters simulated from the prior and removing a planet at random, both being accepted under a likelihood constraint associated with nested sampling. I actually wonder if this actually is the birth-date version of Peter Green’s (1995) RJMCMC rather than the continuous time birth-and-death process version of Matthew…
“The traditional approach to inferring N also contradicts fundamental ideas in Bayesian computation. Imagine we are trying to compute the posterior distribution for a parameter a in the presence of a nuisance parameter b. This is usually solved by exploring the joint posterior for a and b, and then only looking at the generated values of a. Nobody would suggest the wasteful alternative of using a discrete grid of possible a values and doing an entire Nested Sampling run for each, to get the marginal likelihood as a function of a.”
This criticism is receivable when there is a huge number of possible values of N, even though I see no fundamental contradiction with my ideas about Bayesian computation. However, it is more debatable when there are a few possible values for N, given that the exploration of the augmented space by a RJMCMC algorithm is often very inefficient, in particular when the proposed parameters are generated from the prior. The more when nested sampling is involved and simulations are run under the likelihood constraint! In the astronomy examples given in the paper, N never exceeds 15… Furthermore, by merging all N’s together, it is unclear how the evidences associated with the various values of N can be computed. At least, those are not reported in the paper.
The paper also omits to provide the likelihood function so I do not completely understand where “label switching” occurs therein. My first impression is that this is not a mixture model. However if the observed signal (from an exoplanetary system) is the sum of N signals corresponding to N planets, this makes more sense.
Today was the final session of our Reading Classics Seminar for the academic year 2014-2015. I have not reported on this seminar much so far because it has had starting problems, namely hardly any student present on the first classes and therefore several re-starts until we reached a small group of interested students. And this is truly The End for this enjoyable experiment as this is the final year for my TSI Master at Paris-Dauphine, as it will become integrated within the new MASH Master next year.
As a last presentation for the entire series, my student picked John Skilling’s Nested Sampling, not that it was in my list of “classics”, but he had worked on the paper in a summer project and was thus reasonably fluent with the topic. As he did a good enough job (!), here are his slides.
Some of the questions that came to me during the talk were on how to run nested sampling sequentially, both in the data and in the number of simulated points, and on incorporating more deterministic moves in order to remove some of the Monte Carlo variability. I was about to ask about (!) the Hamiltonian version of nested sampling but then he mentioned his last summer internship on this very topic! I also realised during that talk that the formula (for positive random variables)
does not require absolute continuity of the distribution F.
In conjunction with the recent PNAS paper on massive model choice, Rob Johnson†, Paul Kirk and Michael Stumpf published in Bioinformatics an implementation of nested sampling that is designed for biological applications, called SYSBIONS. Hence the NS for nested sampling! The C software is available on-line. (I had planned to post this news next to my earlier comments but it went under the radar…)