The tenets of quantile-based inference in Bayesian models
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (PREreview)
Abstract
Bayesian inference can be extended to probability distributions defined in terms of their inverse distribution function, i.e. their quantile function. This applies to both prior and likelihood. *Quantile-based likelihood* is useful in models with sampling distributions which lack an explicit probability density function. *Quantile-based prior* allows for flexible distributions to express expert knowledge. The principle of *quantile-based* Bayesian inference is demonstrated in the univariate setting with a Govindarajulu likelihood, as well as in a *parametric quantile regression*, where the error term is described by a quantile function of a Flattened Skew-Logistic distribution.
Article activity feed
-
This Zenodo record is a permanently preserved version of a PREreview. You can view the complete PREreview at https://prereview.org/reviews/5554646.
This paper gives a good overview on Bayesian inference for quantile distributions, covering the theoretical basis and practical methods. There are also plenty of useful examples. The paper also introduces novel methodology for some stages of the process. Overall I recommend the paper as an good description of the area.
My most major issue with the paper is that I would not find it easy to reproduce the results from the code supplied. Other comments are that I think the paper could be improved by adding more discussion on some topics and fixing some minor mistakes. Full details of all these issues are below.
Main comments
A lot of code is included as supplementary material, but …
This Zenodo record is a permanently preserved version of a PREreview. You can view the complete PREreview at https://prereview.org/reviews/5554646.
This paper gives a good overview on Bayesian inference for quantile distributions, covering the theoretical basis and practical methods. There are also plenty of useful examples. The paper also introduces novel methodology for some stages of the process. Overall I recommend the paper as an good description of the area.
My most major issue with the paper is that I would not find it easy to reproduce the results from the code supplied. Other comments are that I think the paper could be improved by adding more discussion on some topics and fixing some minor mistakes. Full details of all these issues are below.
Main comments
A lot of code is included as supplementary material, but it's not entirely clear how to run it to reproduce the analyses. Adding a README file summarising this would be helpful. Also, looking at the code files I can only see Stan code and an R file defining data + target distribution. Should there also be code for the fmcmc analyses? Finally, I wonder if it would help to include scripts to run the Stan analyses.
Some of the paper reviews existing methods, so it would be good to list the contributions of the paper in the introduction, in particular which parts are novel. (Some contributions are described in various parts of Section 7, but it would be helpful to have an overview in one place.)
"Indirect inference" is the name of a popular method for likelihood-free inference. See for example "Indirect inference" by Gourieroux et al and "Estimating functions in indirect inference" by Heggland and Frigessi. To avoid confusion you could use a different name (e.g. "indirect quantile inference" or "inference for indirect likelihoods"), or add a discussion to note that there is no direct connection to this literature.
I suggest adding a summary of how inference on indirect likelihoods is carried out. Section 5 discusses numerical inversion of the quantile function but it would be good to add (1) a brief discussion of how this used within Stan (and other MCMC algorithms) and (2) a summary of which inversion methods you recommend and use in your examples. (There are some comment on this in Section 5, but it would be good to have a short summary somewhere.) You could then refer to this summary in Sections 3 and 4 when you perform inference. One particular issue is that HMC algorithms require the gradient of the log likelihood. Can you comment on how this is calculated for indirect likelihoods?
It would be interesting to comment on the run time of inference algorithms using indirect likelihoods. Does the need for repeated numerical inversion slow these down noticably?
Optionally, it would be interesting to comment on possibilities for using quantile distributions beyond the case of univariate IID data.
Minor comments
There are some minor grammatical errors in the paper. The writing is still easy to understand, but it would be ideal to fix these. One particularly noticable error is missing articles (e.g. "the" or "a"). For instance I suggest adding "The" to the start of (1) the first sentence on page 10 (2) the first sentence in Section 5.3 (3) the start of most paragraphs in Appendix A.
Page 4 Figure 1: It would be good to replace "Some distributions" with at least 1 specific example e.g. uniform or exponential?
Page 4 Figure 1: There are some unusual distributions mentioned in Figure 1, and it would be good to explain to the reader where they can find more details about them. Most are described in Section 2.3, so perhaps the caption to Figure 1 could mention this. However there are some other distributions which I couldn't find described anywhere in the paper (Q-Normal and Myserson), and it would be good to add a brief comment with citations for these.
Page 5 When you introduce the definitions (1) and (2) it would be worth noting that you will often omit "|θ" and "_X" below to simplify notation
Page 9 Can you clarify the application area for the claims example. Is this about insurance claims, or something else?
Page 10 "prior beliefs about the parameter(s) of the distribution of θ can be expressed indirectly using the distribution of the quantile values corresponding to the depth v, given hyperparameter(s) of the prior distribution (9)" I think you could simplify this to just say "prior beliefs about θ can be expressed indirectly using
a distribution for the depth v".
Page 10/11 Currently the bottom equations in (9) and (10) contradict each other! To avoid this, a little more notational detail is needed in applying the change of variables formula e.g. subscripts on L to denote the random variable whose density is given.
Page 13 "we defined a shifted exponential prior". For clarity it would be helpful to state this prior mathematically e.g. using its pdf.
Page 16 In grid-matching, it's possible that x is outside the grid of Q values. What should be done in this case?
Page 16 The discussion of non-bracketing methods only describes Newton-Raphson. It might be worth briefly mentioning how other non-bracketing methods differ from this, or saying that you will only consider Newton-Raphson.
Page 20 "In order for the quantile function to be valid, it needs to be... non-decreasing. Note that it is possible for a non-decreasing quantile function to... remain valid." These two sentences contradict each other! A little more care is needed in this explanation. (I think the point is that a non-decreasing function is a valid transformation even though it's not a valid quantile function.)
Page 24 Section 7.2 might be better as an appendix as it involves some technical details of coding in Stan which some readers won't be familiar with.
Page 26 "We tried to cue initial values closer to the mode of the prior to facilitate better mixing of MCMC chains". Can you say more about how you picked suitable initial values in practice?
Technicalities and possible typos
Page 5 Equation (2) I think this should say "inf{x: ...}" rather than "inf{u : ...}".
Page 5 "If $F_X(x)$ is continuous and non-decreasing" The cdf is non-decreasing by defintion! Perhaps this should say strictly increasing.
Page 5 "simply an inverse" -> "simply the inverse"?
Page 8 It would be worth explaining the notation [1..N] (and similar notation elsewhere)
Page 8 I think "computing of" should be "computing"
Page 9 It would be more accurate to replace $\sum \underline{x}$ with $\sum_{i=1}^N x_i$
Page 10/11 Ideally the equation blocks (9) and (10) should have implication signs ($\Rightarrow$ in latex) on their 2nd lines
Page 11 The first sentence after (10) mainly repeats remarks from the previous page.
Page 15 "u|θ is the depth u, such that Ω(u; x, θ) = 0" seems slightly convoluted - could you just say "u is the depth"?
Page 15 "iteratively adjusting u|θ" Would it be more accurate to say "iteratively adjusting u"? (similarly for other uses of u|θ later)
Page 16 (and elsewhere) Typo "Newton-Rhapson"
Page 16 Typo in (12) ";;"
Page 17 Section 5.3 uses p for depth instead of u as in earlier sections. I suggest using u for consistency, or discussing why this change of notation is used.
Page 23 "the algorithm start" -> "the algorithm starts"
Page 31 Gilchrist 2000a and 2000b are the same reference.
-
