Bayesian modeling with R and Stan (3): Simple hierarchical Bayesian model
In 2 previous posts, you learned what Bayesian modeling and Stan are and how to install them. Now you are ready to try it on some very Bayesian problems - as many people love - such as hierarchical Bayesian model.
Definition of hierarchical Bayesian models
Prior to tackling with a practical example, let's overview what and how hierarchical Bayesian model is. A famous book on Bayesian modeling with MCMC, written by Toshiro Tango and Taeko Becque and published in Japan, describes as below*1.
- 作者: 丹後俊郎,Taeko Becque
- 出版社/メーカー: 朝倉書店
- 発売日: 2011/10/14
- メディア: ?行本-精装
- クリック: 8回
- この商品を含むブログを見る
In a fixed-effects model of frequentist, each result is assumed to have a common average .
On the other hand, in a random-effects model, each result is assumed to have a distinct average and it is distributed around a global average .
Bayesian hierarchical models assume prior probability for parameters of a probability distribution of in a random-effects model, such as
It is said that such models have a hierarchical structure with two levels, that is,
- 1st level: a probability distribution is assumed for
- 2nd level: one more probability distribution is assumed for parameters of the 1st level
This is a textbook definition of hierarchical models, but I think it can be understood more intuitively; in hierarchical Bayesian models, often the models have to handle some excessive fluctuations as nonlinear effects more than expected in usual frequentist's models. Priors used in such models can be seen as an "absorber" that can absorb various kinds of fluctuations distributed around true parameters.
*1:Its original text is of course in Japanese, so this is just my own interpretation
Bayesian modeling with R and Stan (2): Installation and an easy example
The previous post overviewed what and how is Stan on R.
Are you ready now? OK, this post reviews how to install Stan. Let's start here! :) In principle this post just follows a content of "RStan Getting Started" but some tips are added in order to fix less known problems.
Warning: this post assumes you are an Windows user. If you use Mac OS or Linux, please see notification for each OS.
Read moreBayesian modeling with R and Stan (1): Overview
Although I've written a series of posts titled "Machine Learning for package uses in R", usually I don't run machine learning on daily analytic works because my current coverage is so-called an ad-hoc analysis.
Instead of machine learning, ad-hoc analysts often use statistical modeling such as linear models (called "multiple regression" in general), generalized linear models (GLM) and/or econometric time series analysis. But in some situations such linear model and its variants would not work because of nonlinear components and/or individual variance, called "random effect".
In general, random effect can be well handled by generalized linear mixed models (GLMM) and for example CRAN has some related packages. But in some cases random effects cannot be formulated concisely and explicitly... if so, we have a strong alternative method to resolve it: "Bayesian using Markov Chain Monte Carlo (MCMC) method".
As one of the strongest methods for ad-hoc analysis, a series of posts will argue about Bayesian modeling with MCMC and its apllication. For the first time, this post overviews it.
Read more