Treatment Effects Analysis
An Online Seminar Taught by Stephen Vaisey, Ph.D.
To see a sample of the course materials, click here.
To watch a sample video, click here.
This course offers an in-depth survey of a family of techniques known as treatment-effects estimators. Treatment-effects analysis is a quasi-experimental technique for estimating causal effects from observational data using the potential outcomes or counterfactual framework. These techniques — which include propensity-score matching, inverse probability weighting, and “doubly-robust” estimators — are now widely used in the social sciences, health sciences, and public policy.
The goal of treatment-effects analysis is to identify the causal effect of a treatment on an outcome, such as the effect of a college education on earnings, the effect of divorce on child outcomes, or the effect of a training program on employee productivity. A major advantage of treatment-effects techniques over standard regression methods is that they can produce different estimates of causal effects for subjects who are likely to receive the treatment and for those who are unlikely to receive it, an important distinction for policy work.
This seminar will take participants from simple exact matching to recent developments like coarsened exact matching, doubly-robust estimators and entropy balancing. Participants will get extensive practical experience by working through case studies from economics, sociology, medicine, and public health.
Though the seminar will focus on hands-on understanding, we will also use causal graphs to look more deeply into the assumptions required to achieve unbiased estimates. Participants will learn to see how these techniques can be used in their own research.
We will cover a variety of topics including exact matching, propensity score matching and weighting, other forms of non-parametric matching and weighting, regression adjustment, and various forms of doubly-robust estimators. We will also consider tests for violations of assumptions and ways to test the sensitivity of results to violations of untestable assumptions. Although we will focus primarily on binary treatments, we will briefly explore how these techniques can be applied to multivalued treatments as well.
The course will take place over a four-week period. Each week you will be asked to watch approximately three hours of video recordings of Professor Vaiesey’s lectures. There will also be assigned readings, a short multiple choice quiz after each of 16 modules, and a weekly computer exercise. A discussion board is available for you to ask questions or post comments on an any aspect of the course. Professor Vaisey will respond to all questions within 24 hours of posting.The materials can be accessed with any recent web browser on almost any platform, including iPhone, iPad, and Android devices.
The instructor will use Stata (and some R) to demonstrate the techniques. To do the hands-on exercises, you will need to have Stata installed on your computer. Stata 14 or 15 (any flavor) are preferred, though Stata 13 can do at least 95% of what we will cover in this course. Students do NOT need any prior knowledge of Stata to be able to complete the exercises.
WHO SHOULD sign up?
This course is for any who want to learn to apply this family of techniques to observational data. Participants should have a basic foundation in linear and logistic regression.
1. The potential outcomes framework
a. The experimental ideal
b. Quasi-experiments and self-selection
c. Directed acyclic graphs (DAGs)
d. ATT: average treatment effect on the treated
e. ATU: average treatment effect on the untreated
f. ATE: average treatment effect
2. Exact matching
c. Requirements for estimating ATT/ATU/ATE
d. Other assumptions
3. Propensity score matching
a. P-scores as a solution to sparseness
b. Estimating the selection model
c. Matching on propensity scores
e. Testing overlap assumption
4. Propensity score weighting
a. “Missing data” motivation
b. Comparing weighting and matching
5. Non-parametric matching
a. Nearest-neighbor (Mahalanobis distance) matching
b. Coarsened exact matching
6. Regression adjustment
a. Balancing vs. conditioning
b. DAG motivation
7. Doubly-robust techniques
a. DAG motivation for DR techniques
b. Combining weighting and regression adjustment
8. Overview of advanced topics
a. Sensitivity analyses
b. Extensions to multivalued treatments
“This course did an excellent job covering both the theory and practice for methods related to propensity score models and other methods for dealing with selection bias in observational data. It actually could serve as a good primer for the course specifically focused on propensity score, but in fact goes beyond to cover other potentially even more meaningful methods. The instructor did an excellent job incorporating a review of methods related to survival data in response to student interest.”
Colleen Jay, UT Health Science Center
“Great practical examples and hands-on experience conducting treatment effects analysis during the course. I would highly recommend this course to colleagues.”
Nikki Wooten, University of South Carolina
“Steve was very fluid in his delivery. He has excellent mastery of the material. Very useful course.”
Alex Appiah, Bank of America
“The treatment effects analysis course was extremely useful. I’m looking forward to applying these methods correctly in my work. The instructor was very enthusiastic about the materials, as well as being extremely knowledgeable. Even with a very theoretical subject, the course was taught on a practical level which was greatly appreciated.”
Heather Litman, Corrona, LLC