PS methods for multiple treatments Flashcards
What is weak unconfoundness?
requires that assignment to each treatment is independent of the potential outcomes of the respective treatment
What is the difference between weak unconfoundness and strong ignorability of treatment assignment?
? Weak ignorability does not require that the assignment to one treatment is independent of all potential outcomes.
What are the assumptions necessary for propensity score methods for multiple treatments?
- weak ignorability of treatment assignment
- overlap
- stable unit treatment value assumption (SUTVA).
What is the generalized propensity score?
conditional predicted probabilities of individuals receiving a particular version of treatment j given observed covariates X
formula in ppt
What are two methods that can be used to estimate generalized propensity scores?
- multinomial logistic regression
* data mining methods
How can data mining methods be used to estimate generalized propensity scores?
- Create a dummy indicator for each treatment version
- Predict each dummy indicator
- The generalized propensity score for each individual is the predicted probability of the treatment version that the individual was exposed to.
How can common support be assessed with multiple treatments?
- common support should hold for all pairs of treatment versions with respect to each of the J vectors of GPS (for multinomial logistic regression)
- by examination of minimums and maximums and with box-and-whiskers plots (for the GPS obtained with GBM)
How can weights based on the generalized propensity score be calculated?
formula in ppt
How are marginal mean weights through stratification calculated with multiple treatments?
- Define treatment levels.
- Obtain generalized propensity score
- Check for overlap.
- Divide propensity scores into strata.
- Obtain marginal mean weights through stratification
- Estimate the treatment effects across strata using the weights for units in propensity stratum s and assigned to treatment version j, (formula in ppt)
How can covariate balance be assessed with multiple treatments?
by comparison with all other versions, or by pairwise comparison of treatment versions.
What are two methods to estimate the ATE with multiple treatments?
weighted mean differences
generalized linear models
How can control for covariates be performed in the estimation of effects of multiple treatments?
Covariates and interactions between covariates and outcomes can be added to remove additional bias, increase power and achieve double-robustness.
