SMABS 2004 Jena University
SMABS 2004 Home Organization About Jena Sponsors Links Imprint / Contact SMABS Home

European Association of Methodology

Department of methodology and evaluation research

Jena University

Contributions: Abstract

This course will present the Rubin Causal Model perspective on understanding and teaching statistical inference for causal effects through potential outcomes. There are three parts to the course. The first part establishes the primitives that form the foundation. The second part presents inference based solely on the assignment mechanism; this perspective generalizes Fisher's (1925) and Neyman's (1923) randomization-based methods, and emphasizes the central role of the propensity score (Rosenbaum and Rubin, 1983). The third part presents inference based on predictive models for the distribution of the missing potential outcomes, formally, Bayesian posterior predictive inference (Rubin, 1978). In practice, the predictive approach is ideal for creating statistical procedures, whereas the assignment-based approach of Fisher is ideal for traditional confirmatory inference, and the assignment-based approach of Neyman is ideal for evaluating procedures. For best practice, being facile with all three approaches is important. There is essentially no prerequisite knowledge for this course, as the material is based on an introductory course taught at Harvard University and designed for students with very little quantitative background. The material is, however, conceptually demanding. Examples are presented from a variety of fields, including medicine, education, economics and other branches of social science.

Course outline:

I . Framework

1. Primitives: Units, treatments, potential outcomes
2. Learning about causal effects: Replication, stability, the assignment mechanism
3. The transition to statistical inference: Introduction to randomized experiments and the Rubin Causal Model
4. Examples of assignment mechanisms

II. Causal inference based on the assignment mechanism

5. Fisherian significance levels in CR experiment
6. Neymanian repeated sampling evaluations in CR experiment
7. Extension to studies with variable but known propensities
8. Extension to studies with unknown propensities
9. Examples of methods for estimating propensities

III. Causal inference based on predictive distributions of potential outcomes

10. Predictive inference intuition under ignorability
11. Matching to impute missing potential outcomes
12. Fitting predictive models within each treatment group
13. Formal predictive inference Bayesian [Rubin,1978]
14. Principal stratification: Dealing with post-treatment variables (noncompliance, surrogate outcomes, censoring due to death, etc.)