Chapter 10 PSA with multilevel data

Question 1

1. What is the main advantage of pooling estimated treatment effects across clusters over marginal estimation of treatment effects across clusters?

Answer 1

It removes the confounding effects of both observed and unobserved cluster-level confounders.

Question 2

2. What is the main limitation of pooling estimated treatment effects across clusters?

Answer 2

The cluster sizes needs to be large enough to allow adequate common support of propensity score distributions within clusters.

Question 3

3. When is marginal estimation of treatment effects across clusters recommended instead of pooling estimated treatment effects across clusters?

Answer 3

However, if cluster sizes are not large for adequate common support, marginal estimation of the treatment effect across clusters should be used.

Question 4

4. What is the difference between the propensity score model specification for an individual-level treatment and the propensity score model specification for a cluster-level treatment?

Answer 4

?For propensity score analysis of binary treatments with multilevel data, the strong ignorability of treatment assignment assumption is needed, which states that the potential outcome distributions are independent of treatment assignment given covariates.

For analysis of treatment assignment of individuals within clusters, the treatment assignment should also be independent of cluster membership. In addition, no participant should have a probability of treatment assignment of zero or one, and adequate common support of propensity score distributions is required.

Question 5

5. Why is the stable unit treatment value assumption particularly vulnerable to violation when individuals are nested within clusters?

Answer 5

because of close proximity, regular communication, and shared resources

Question 6

6. What are two types of models that can be used for propensity score estimation with multilevel data?

Answer 6

• the multilevel logistic regression model

- the logistic regression model with fixed cluster effects

Question 7

7. How can variation in individual-level covariate effects on treatment assignment across clusters be accounted for in propensity score models?

Answer 7

1) If multilevel logistic regression models are used to estimate propensity scores, include random slopes of covariates in the model;
2) If logistic regression with fixed cluster effects is used, include covariate by cluster interactions in the model;
3) Ignore the variation of covariate effects across clusters.

Question 8

8. What is one advantage of a logistic regression model with fixed cluster effects over a multilevel logistic regression model for propensity score estimation?

Answer 8

1) It removes all confounding due to cluster-level covariates without requiring the inclusion of any cluster-level covariates in the model (Arpino & Mealli, 2011), while the multilevel logistic regression model requires that the research identifies and measures the true confounders at the cluster level;
2) It allows any correlations between individual-level predictors and the fixed cluster effects, while the multilevel logistic regression model assumes that and are uncorrelated.

Question 9

9. What is one advantage of a multilevel logistic regression model over a logistic regression model with fixed cluster effects for propensity score estimation?

Answer 9

Disadvantages of logistic regression with fixed cluster effects include: 1) Convergence difficulty if the number of clusters is large;

2) If the dataset has many small clusters, propensity score estimates may be unstable and/or predicted probabilities equal to zero or one may occur (Li, et al., 2013); 3) The requirement that there are at least two observations per cluster.

Question 10

10. What approaches can be used for marginal estimation of treatment effects across clusters that account for cluster effects on the outcome?

Answer 10

multilevel model with random intercept and random slope of treatment