Outline - Boise State University

Modern Methods to Estimate Propensity Score Weights
Dan McCaffrey & Matt Cefalu
The estimation of causal effects is the goal of many research studies. For example, analysts might want
to understand whether a new treatment or intervention leads to better outcomes for patients or
whether receipt of certain health service improves long term health. Controlled experiments are held as
the gold standard for estimating causal effects. However, experiments can be infeasible for many
reasons forcing analysts to rely on observational data in which treatment assignments are out of the
control of the researchers. Over the last 30 years, statisticians have developed new models for defining
causal effects and new methods for estimating them from observational data that complement
traditional approaches that are still commonly used in many applications such as linear and logistic
regression. Recently RAND statisticians and their colleagues have developed tools to implement these
new causal effect estimation methods.
This short course will provide an introduction to causal modeling using the potential outcomes frame
work and use of propensity score weights in the estimation of causal effects from observational data.
The course will also provide step-by-step guidelines on how to estimate and perform diagnostic checks
of propensity score weights for evaluations examining the relative effectiveness of two or more
interventions and the cumulative effects of time-varying interventions. Attendees will gain hands on
experience estimating propensity score weights for these different settings, evaluating the quality of
those weights, and utilize the weights for estimating intervention effects.
At the end of the course, participants should be familiar with how propensity score weights can be
utilized to estimate intervention effects and how to evaluate balance before and after propensity score
weighting in their own studies. The course will provide demonstrations of available software in R and
SAS for fitting the models and opportunities for conducting analyses. The primary goals of the course
are for attendees to have an understanding of how to implement propensity score weighting using
state-of-art-methods and insights into some of the practical issues that evaluating the quality of
propensity score weights involve.
Course Outline
8:30 – 9:30 Introduction to potential outcomes framework for causal modeling and the role of the
propensity score in causal effect estimation
9:30 - 9:45 Break
9:45 – 10:30 Inverse probability of treatment weighting and causal effect estimators
10:30 – 12:00 Propensity score estimation via logistic regression and GBM
a. What is GBM and how does it compare with logistic regression
b. The role of balance in propensity score estimation
c. Metrics for assessing balance
12:00 – 12:30 Lunch
12:30 – 1:30 Propensity score weighting example with two treatment conditions
a. Use the TWANG package to fit a GBM model
b. Use of TWANG is R, SAS, and Stata
c. Use of TWANG to assess balance
d. Estimation of treatment effects using estimated weights
1:30 – 2:00 Alternatives to GBM and logistic regression for propensity score estimation (CBPS or MDIA)
2:00 – 2:45 Causal effects with 3+ treatments, definitions and estimators
2:45-3:00 Break
3:00 – 4:30 Propensity score weighting example with four treatment conditions
a. Use the TWANG package to fit a GBM model
b. Use of TWANG is R, SAS, and Stata
c. Use of TWANG to assess balance
d. Estimation of treatment effects using estimated weights
4:30-5:00 Final topics
a. Marginal structural models & inverse probability of treatment weighting
b. Longitudinal analyses -- GLS vs. Independence