|
Post by bayes on Apr 20, 2015 23:39:18 GMT -5
What are the best resources/textbooks for self-teaching bayesian statistics for someone who has only ever had classes in more frequentist methods? Especially geared towards the social sciences?
|
|
|
Post by initial primer on Apr 22, 2015 17:34:06 GMT -5
What are the best resources/textbooks for self-teaching bayesian statistics for someone who has only ever had classes in more frequentist methods? Especially geared towards the social sciences? Western, Bruce. 1999. “Bayesian Analysis for Sociologists: An Introduction.” Sociological Methods and Research. 28:7–34.
|
|
|
Post by 820d on Apr 23, 2015 22:06:56 GMT -5
Really trying to avoid working on my dissertation right now, so I'll give a longer answer. The best advice is to start a project using Bayesian modeling and learn enough to get started, then read to figure out the details as you go along.
The Gelman book is generally the best if you have the background for it (strong calculus and probability, decent linear algebra). The third edition is really good, and spends a decent amount of time on Hamiltonian Monte Carlo and Gelman's NUTS sampler, which are probably going to be used a lot in the future. Also worth noting that Gelman is now using STAN rather than BUGS as the default language, which isn't as widely adopted in the Bayesian community but is much faster/better than the alternatives (with a few exceptions). STAN is also relatively scalable, making working with large, but not huge, datasets feasible. Easily extendible and fun to hack if you are good at C++.
Another good source not yet mentioned is Simon Jackman's book. Same coverage, more or less, than Gill and Gelman, but worth picking up as either a primary or an additional resource. Jackman's book is also written for JAGS, which is great software that is FOSS and has easily installable 64-bit versions for Linux, OSX, and Windows. There are no existing modules in JAGS for spatial analysis though, so if that is your thing, probably pick up a different book. If you have good C++ skills you can always extend it.
Gill's book is good, though terse and marginally less helpful to get up and running models compared to Gelman or Jackman. The third edition wasn't a major improvement over the second, IMHO, but it is definitely a good place to start. He definitely goes deeper into MCMC methods than the other comparable books, which is a good thing. The code is all R and BUGS, with some coverage of JAGS.
Lynch's book will give you a good sense of the basics. You only need baby calculus and basic probability to keep on top of everything and will definitely get you started. Well worth checking out if your math is rusty and you want some clear contrasts to frequentist models. If your math skills are a bit further along, Jackman/Gill/Gelman make more sense. Code here is in R and WinBUGS.
Other than that, "The BUGS Book" is heavy on application but has some coverage of theory. It could take you a long way getting off the ground. "Bayesian Modeling using WinBUGS" is also a great way to learn how to port frequentist models you know into Bayesian methods. Even less theory, but code for tons of different models specifications. Both of those are based in WinBUGS, which despite its name, can be installed in Linux and presumably OSX using Wine. Last, Gelman and Hill's 2007 book provides basic coverage of both frequentist and Bayesian approaches to hierarchical models, and provides a good primer if that is the extent of your interests.
Stata is great, but I don't know that it would be right platform for Bayesian analysis. Nearly all people with in R/JAGS/BUGS/STAN and there is a relatively strong emphasis on FOSS in the Bayesian community. It will probably take Stata years to catch up with what is there today.
|
|
|
Post by How important on Apr 24, 2015 10:20:11 GMT -5
How important is it to know Calculus to do Bayesian statistical analysis? The more I read about Bayesian statistics the more it appeals to me, but then I look at the actual process and they're talking about stuff I've never heard of.
|
|
|
Post by very on Apr 27, 2015 0:27:10 GMT -5
I'd say it is very important to know some calculus (and linear algebra). I'd say it is important to know calculus for frequentist statistics as well, but since that is the standard you can get away with just following convention. That is, in sociology you can probably get away with just doing the standard methods, following the p-values and so on. Since Bayesian statistics isn't the standard, you have to be able not only to internalize it well in order to do it, but also know enough to explain to reviewers.
|
|
|
Post by 820d on Apr 27, 2015 9:34:53 GMT -5
Yes, calculus is very important. Probably is essential, and linear algebra is necessary to make sense of what you will read. There is really no way to get around those areas of math if you want to do Bayesian analysis. One could say the same for MLE or least squares estimation, but many sociology grad programs typically don't delve deeply into those topics.
|
|
|
Post by How important on Apr 27, 2015 10:43:16 GMT -5
Thanks for the tips. My calculus is pretty limited--I was able to understand maximum likelihood just fine, but if I'm calculating out probability distributions using calculus on a regular basis I'm pretty sure I'd be screwed.
I really wish I could go back in time and take more math. Maybe for my sabbatical...
|
|