bayesian-statistics-guide

Bayesian Statistics Guide

A skill for applying Bayesian statistical methods to research data analysis. Covers prior specification, Markov chain Monte Carlo (MCMC) sampling, posterior interpretation, model comparison, and reporting standards.

Bayesian Framework Overview

Bayes' Theorem in Practice

Posterior = (Likelihood x Prior) / Evidence

P(theta | data) = P(data | theta) * P(theta) / P(data)

In practice:
  P(theta | data) is proportional to P(data | theta) * P(theta)
  (the denominator is a normalizing constant)

When to Use Bayesian Methods

Scenario	Bayesian Advantage
Small sample sizes	Priors regularize estimates
Complex hierarchical models	Natural framework for multilevel data
Sequential data collection	Update beliefs as data arrives
Prior knowledge available	Formally incorporate existing evidence
Model comparison	Bayes factors and posterior model probabilities
Prediction	Full posterior predictive distributions

Prior Specification

Types of Priors

import numpy as np
from scipy import stats
import matplotlib.pyplot as plt

def visualize_priors(parameter_name: str, prior_type: str = 'weakly_informative'):
    """
    Visualize common prior choices for a parameter.
    """
    x = np.linspace(-10, 10, 1000)

    priors = {
        'flat': {
            'dist': stats.uniform(loc=-100, scale=200),
            'description': 'Flat/Uniform: minimal prior info (often improper)',
            'recommendation': 'Avoid -- can lead to improper posteriors'
        },
        'weakly_informative': {
            'dist': stats.norm(loc=0, scale=2.5),
            'description': 'Weakly informative: Normal(0, 2.5)',
            'recommendation': 'Good default for regression coefficients'
        },
        'informative': {
            'dist': stats.norm(loc=0.5, scale=0.2),
            'description': 'Informative: based on previous studies',
            'recommendation': 'Use when strong prior evidence exists'
        },
        'horseshoe': {
            'dist': stats.cauchy(loc=0, scale=1),
            'description': 'Horseshoe-like (Cauchy): sparsity-inducing',
            'recommendation': 'Good for variable selection problems'
        }
    }

    prior = priors.get(prior_type, priors['weakly_informative'])
    return prior

# Recommended default priors (Gelman et al., 2008):
# Intercept: Normal(0, 10)
# Coefficients: Normal(0, 2.5) on standardized predictors
# Standard deviation: Half-Cauchy(0, 2.5) or Exponential(1)
# Correlation: LKJ(2) for correlation matrices