[MODELING WEBINAR] -- Bayesian Causal Inference & Propensity Scores, with Nathaniel Forde

My Intuitive Bayes Online Courses: www.intuitivebayes.com/
1:1 Mentorship with me: topmate.io/alex_andorra
📚 In this webinar, we will explore the world of causal inference and how propensity scores can be a powerful tool. Learn how to estimate propensity scores and use them to tackle selection bias in your analysis.
🔍 Agenda
- Propensity Score Estimation: Discover how to calculate propensity scores and their significance in causal analysis. 📊
- Bayesian Analyst's Perspective: See how propensity score weighting can enrich your Bayesian models with valuable information. 🧮
- Machine Learning & Causal Inference: Explore the application of propensity scores in debiasing machine learning for causal inference. 🤖
- Contrast: We'll highlight the differences between non-parametric BART models and simpler regression models in estimating propensity scores and causal effects.
📈 Takeaways
- Propensity scores are useful in observational data to evaluate the effect of a treatment.
- Different weighting schemes, such as raw inverse weighting and doubly robust methods, can be used with propensity scores.
- Bayesian additive regression trees (BART) models offer flexibility in modeling propensity scores.
- Extreme propensity scores can be dealt with by removing individuals or using matching algorithms.
- Propensity scores can be used in regression modeling to estimate treatment effects. Propensity score adjustment, doubly robust estimation, and mediation analysis are methods used in causal inference.
- Beliefs and assumptions about the data generating process underlie these methodologies.
- It is important to be Bayesian in our approach to causal inference.
- Non-parametric estimation can help address miscalibration and overfitting risks.
- Understanding the causal structure and considering mediation effects are crucial in causal inference.
📜 Webinar Structure
- Non-parametric Approaches: We'll showcase various non-parametric methods for estimating causal effects. Some are spot-on, while others can mislead you. ⚙️
- Propensity Scores in Selection Effect Bias: Learn how to apply propensity scores to tackle selection bias head-on. 🛡️
- Debiased/Double ML for ATE Estimation: Discover how to use Debiased/Double Machine Learning to estimate the Average Treatment Effect (ATE). 📈
- Mediation Analysis: Dive into mediation analysis and estimate Direct and Indirect Effects. 🔄
🎙️ Our guest speaker, Nathaniel Forde, is a data scientist specializing in probabilistic modeling for the study of risk and causal inference. He has experience in model development, deployment, multivariate testing and monitoring, and his academic background is in mathematical logic and philosophy.
🎁 If you're a Patron of the Learning Bayesian Statistics podcast, you can submit questions in advance, enjoy early access to all webinar recordings, and get at least a 50% discount on future webinars ( / learnbayesstats ).
References:
Notebook: www.pymc.io/projects/examples...
Slides: nathanielf.github.io/talks/lb...
Previous webinars: • Modeling Webinars
LBS #34, Multilevel Regression, Post-stratification & Missing Data, with Lauren Kennedy: • #34 Multilevel Regress...
LBS #83 Multilevel Regression, Post-Stratification & Electoral Dynamics, with Tarmo Jüristo: • #83 Multilevel Regress...
Bayesian mediation analysis: www.pymc.io/projects/examples...
Interventional distributions and graph mutation with the do-operator: www.pymc.io/projects/examples...
Complementary references: topmate.io/alex_andorra/775666
Chapters:
00:00 Introduction and Welcome
03:10 Introduction to Causal Inference and Propensity Scores
05:07 Propensity Score Analysis
09:03 Nonparametric Causal Inference
13:29 Dealing with Extreme Propensity Scores
21:22 Doubly Robust Methods
27:01 Balance of Covariate Distributions
38:59 Inverse Weighting Schemes
43:02 Doubly Robust Estimator
45:27 Comparison of Logistic Regression and BART Models
51:42 Flexibility of BART Models
54:37 Using Propensity Scores in Regression Modeling
58:31 Miscalibrated Propensity Scores and Overfitting Risks
59:57 Conditional Average Treatment Effect
01:02:18 Imbalanced Treatment and Control Groups
01:03:42 Fitting a BART Model
01:05:05 Addressing Miscalibrated Propensity Scores
01:06:02 Contrasting Raw and Reweighted Outcome Variables
01:06:31 Robust and Doubly Robust Average Treatment Effects
01:09:21 Regression Model for Adjusting Propensity Scores
01:11:39 Debiased Machine Learning and Frisch-Waugh-Lovall Theorem
01:12:36 Non-Parametric Estimation of Conditional Average Treatment Effect

[@ABG1788]

amazing talk

[@cerioscha]

Is there a typo in the slide @ 5:58 with reads: "P( theta | D ) proportional to P( D | theta ) P(D)". Should this not be: " "P( theta | D ) proportional to P( D | theta ) P(theta)" ?

[@nathanielforde8952]

😅 good catch thanks

[@cerioscha]

@@nathanielforde8952 Not al all, thanks for a great talk. It's a great time to be a Bayesian, we've come a long from "I think therefore i am, to i act because I infer'".