Relearning statistics with a Bayesian approach. My undergrad education was in social science research methods and I spent 4 years learning the strict frequentist approach of orthodox statistics. It made sense for highly-controlled experiments and simple dice games. But it broke down horribly when faced with any complexity and I never understood why. 25 years later, it's time to fill in the gaps. My reading list:
- E.T. Janes "Probability Theory: The Logic of Science" provides the fundamental theory.
- Robert McElreath "Rethinking Statistics" provides a practical application of the theory in R.
- Andrew Clayton's "Bernoulli's Fallacy" is a non-technical book that provides historical context to the frequentist vs bayesian debate.
I'm fairly convinced now that Bayesian approaches have more mathematical rigor than the crusty old heuristics of traditional statistics. But in terms of user-experience, doing Bayesian calculations still requires more effort on model design and more compute power. It's flexible to a fault, without a well-defined workflow. There is a strong temptation to follow the easy path - shove your data into a black box and publish if p<0.05. It's going to take a generation of (re)training and improvements to statistical software before Bayesian methods are widespread.
- E.T. Janes "Probability Theory: The Logic of Science" provides the fundamental theory.
- Robert McElreath "Rethinking Statistics" provides a practical application of the theory in R.
- Andrew Clayton's "Bernoulli's Fallacy" is a non-technical book that provides historical context to the frequentist vs bayesian debate.
I'm fairly convinced now that Bayesian approaches have more mathematical rigor than the crusty old heuristics of traditional statistics. But in terms of user-experience, doing Bayesian calculations still requires more effort on model design and more compute power. It's flexible to a fault, without a well-defined workflow. There is a strong temptation to follow the easy path - shove your data into a black box and publish if p<0.05. It's going to take a generation of (re)training and improvements to statistical software before Bayesian methods are widespread.