# Reddit reviews Statistical Rethinking: A Bayesian Course with Examples in R and Stan (Chapman & Hall/CRC Texts in Statistical Science)

We found 13 Reddit comments about Statistical Rethinking: A Bayesian Course with Examples in R and Stan (Chapman & Hall/CRC Texts in Statistical Science). Here are the top ones, ranked by their Reddit score.

## 13 Reddit comments about Statistical Rethinking: A Bayesian Course with Examples in R and Stan (Chapman & Hall/CRC Texts in Statistical Science):

Statistical Rethinking: https://www.youtube.com/playlist?list=PLDcUM9US4XdM9_N6XUUFrhghGJ4K25bFc

Also has the book: https://www.amazon.com/Statistical-Rethinking-Bayesian-Examples-Chapman/dp/1482253445

So off the top off my head, I can’t think of any courses. Here are three books that include exercieses that are relatively comprehensive and explain their material well. They all touch upon basic methods that are good to know but also how to do analyses with them.

But there are many many others.

Then there are plenty of tutorials to python, R or how to handle databases (probably the core programming languages, unless you want to go the GUI route).

One of the post-docs in my department did his dissertation with Bayesian stats and he essentially had to teach himself! He strongly recommended this as a place to start if you are interested in that topic -- https://www.amazon.com/gp/product/1482253445/ref=oh_aui_detailpage_o09_s00?ie=UTF8&amp;psc=1 (I have not read it yet.)

One of our computer science profs teaches Bayes for the CS folks and said he would be willing to put together a class for psych folks in conjunction with some other people, so that's a place where I am hoping to develop some competency at some point. I strongly recommend reaching outside of your department, especially if you are at a larger university!

I can offer my two cents. I’m a Googler who uses machine learning to detect abuse, where my work is somewhere between analyst and software engineer. I’m also 50% done through the OMSCS program. Here’s what I’ve observed:

Yes, Reinforcement Learning, Computer Vision, and Machine Learning are 100% relevant for a career in data science. But data science is vague; it means different things depending on the company and role. There are three types of data science tasks and each specific job may be weighted more heavily in one of these three directions: (1) data analytics, reporting, and business intelligence focused, (2) statistical theory and model prototyping focused and (3) software engineering focused by launching models into production, but with less empathsis on statistical theory.

I've had to do a bit of all three types of work. The two most important aspects are (1) defining your problem as a data science/machine learning problem, and (2) launching the thing in a distributed production environment.

If you already have features and labeled data, you should be able to get a sense of what model you want to use within 24 hours on your laptop based on a sample of the data (

this can be much much harder when you can't actually sample the data before you build the prod job because the data is already distributed and hard to wrangle). Getting the data, ensuring it represents your problem, and ensuring you have processes in place to monitor, re-train, evaluate, and manage FPs/FNs will take a vast majority of your time. Read this paper too: https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdfAcademic classes will not teach you how to do this in a work environment. Instead, expect them to give you a toolbox of ideas to use, and it’s up to you to match the tool with the problem. Remember that the algorithm will just spit out numbers. You'll need to really understand what's going on, and what assumptions you are making before you use each model (e.g. in real life few random variables are nicely gaussian).

I do use a good amount of deep learning at work. But try not to - if a logistic regression or gradient boosted tree works, then use it. Else, you will need to fiddle with hyper parameters, try multiple different neural architectures (e.g. with time series prediction, do you start with a CNN with attention? CNN for preprocessing then DNN? LSTM-Autoencoder? Or LSTM-AE + Deep Regressor, or classical VAR or SARIMAX models...what about missing values?), and rapidly evaluate performance before moving forward. You can also pick up a deep learning book or watch Stanford lectures on the side; first have the fundamentals down. There are many, many ways you can re-frame and tackle the same problem. The biggest risk is going down a rabbit hole before you can validate that your approach will work, and wasting a lot of time and resources. ML/Data Science project outcomes are very binary: it will work well or it won’t be prod ready and you have zero impact.

I do think the triple threat of academic knowledge for success in this area would be graduate level statistics, computer science, and economics. I am weakest in theoretical statistics and really need to brush up on bayesian stats (https://www.amazon.com/Statistical-Rethinking-Bayesian-Examples-Chapman/dp/1482253445). But 9/10 times a gradient boosted tree with good features (it's all about representation) will work, and getting it in prod plus getting in buy-in from a variety of teams will be your bottleneck. In abuse and fraud; the distributions shift all the time because the nature of the problem is adversarial, so every day is interesting.

The short version is that in a bayesian model your likelihood is how you're choosing to model the data, aka P(x|\theta) encodes how you think your data was generated. If you think your data comes from a binomial, e.g. you have something representing a series of success/failure trials like coin flips, you'd model your data with a binomial likelihood. There's no right or wrong way to choose the likelihood, it's entirely based on how you, the statistician, thinks the data should be modeled. The prior, P(\theta), is just a way to specify what you think \theta might be beforehand, e.g. if you have no clue in the binomial example what your rate of success might be you put a uniform prior over the unit interval. Then, assuming you understand bayes theorem, we find that we can estimate the parameter \theta given the data by calculating P(\theta|x)=P(x|\theta)P(\theta)/P(x) . That is the entire bayesian model in a nutshell. The problem, and where mcmc comes in, is that given real data, the way to calculate P(x) is usually intractable, as it amounts to integrating or summing over P(x|\theta)P(\theta), which isn't easy when you have multiple data points (since P(x|\theta) becomes \prod_{i} P(x_i|\theta) ). You use mcmc (and other approximate inference methods) to get around calculating P(x) exactly. I'm not sure where you've learned bayesian stats from before, but I've heard good things , for gaining intuition (which it seems is what you need), about Statistical Rethinking (https://www.amazon.com/Statistical-Rethinking-Bayesian-Examples-Chapman/dp/1482253445), the authors website includes more resources including his lectures. Doing Bayesian data analysis (https://www.amazon.com/Doing-Bayesian-Data-Analysis-Second/dp/0124058884/ref=pd_lpo_sbs_14_t_1?_encoding=UTF8&amp;psc=1&amp;refRID=58357AYY9N1EZRG0WAMY) also seems to be another beginner friendly book.

There seem to be a few options. I've had this and this on my reading list for a while, but haven't got further than that.

I'm also interested in recommendations.

May I recommend this book to you then, Statistical Rethinking: A Bayesian Course with Examples in R and Stan (Chapman & Hall/CRC Texts in Statistical Science) https://www.amazon.co.uk/dp/1482253445/ref=cm_sw_r_cp_apa_Iu6.BbJE7EECQ

Not long! For this purpose I highly

highlyrecommend Richard McElreath'sStatistical Rethinking(this one here). It's SO good. The math is exceptionally straightforward for someone familiar with regression, and it's huge on developing intuition. Bonus: he sets you up with all the tools you need to do your own analyses, and there are tons of examples that he works from a lot of different angles. He even does hierarchical regression.It's an easy math book to read cover to cover by yourself, to be honest. He really holds your hand the whole way through.

Jesus, he should pay me to rep his book.

This book is excellent:

https://www.amazon.com/Statistical-Rethinking-Bayesian-Examples-Chapman/dp/1482253445

I personally think you should brush up on frequentist statistics as well as linear models before heading to Bayesian Statistics. A list of recommendations directed at your background:

They're not free, but Doing Bayesian Data Analysis and Statistical Rethinking are worth their weight in gold.

This may vary by school, but it's been my experience that there aren't a lot of classes explicitly labeled as "artificial intelligence" (especially at the undergraduate level). However, AI is a very broad and interdisciplinary field so one thing I would recommend is that you take courses from fields that form the foundation of AI. (Math, Statistics, Computer Science, Psychology, Philosophy, Neurobiology, etc.)

In addition, take advantage of the resources you can find online! Self-study the topics you're interested in and try to get some hands on experience if possible: read blogs, read papers , browse subreddits, program a game-playing AI, etc.

Given that you're specifically interested in reasoning:

*As a disclaimer: I'm fairly new to this field of study myself. What I've shared with you is my best understanding, but given my lack of experience it may not be completely accurate. (Anyone, please feel free to correct me if I'm mistaken on any of these points)

If you want an extremely practical book to complement BDA3, try Statistical Rethinking.

It's got some of the clearest writing I've seen in a stats book, and there are some good R and STAN code examples.