Stratification and weighting via the propensity score in estimation of causal treatment effects: a comparative study
read more
Citations
An Introduction to Propensity Score Methods for Reducing the Effects of Confounding in Observational Studies
Matching Methods for Causal Inference: A Review and a Look Forward
Moving towards best practice when using inverse probability of treatment weighting (IPTW) using the propensity score to estimate causal treatment effects in observational studies
Doubly robust estimation in missing data and causal inference models
Counterfactuals and Causal Inference: Methods and Principles for Social Research
References
The central role of the propensity score in observational studies for causal effects
Estimating causal effects of treatments in randomized and nonrandomized studies.
Propensity score methods for bias reduction in the comparison of a treatment to a non‐randomized control group
Marginal Structural Models and Causal Inference in Epidemiology
A generalization of sampling without replacement from a finite universe.
Related Papers (5)
The central role of the propensity score in observational studies for causal effects
Marginal Structural Models and Causal Inference in Epidemiology
Frequently Asked Questions (14)
Q2. What future works have the authors mentioned in the paper "Stratification and weighting via the propensity score in estimation of causal treatment effects: a comparative study" ?
An interesting avenue for future research would be to establish guidelines for choosing the number of strata based on theoretical analysis of the rate at which the number of strata should increase with sample size to eliminate bias.
Q3. What is the common method for estimating the difference of two treatment means?
A popular method for estimating the (causal) difference of two treatment means isthat of Rosenbaum and Rubin [7], where individuals are stratified based on estimated propensityscores and the difference estimated as the average of within-stratum effects.
Q4. What is the implication of incorporating covariates in the propensity model?
Thepractical implication is that, at least in large samples, for these weighted estimators, incorporatingcovariates in the propensity model that are not related to treatment exposure but are associatedwith potential response will always lead to precision for estimating ∆ at least as great as thatattained by disregarding such covariates.
Q5. How did the authors determine the relative performance of the subjects in the simulations?
Toinvestigate relative performance in such a realistic setting, the authors carried out simulations involving anumber of continuous and discrete covariates and a continuous response such that ∆0 > 0, where larger values of the response are preferred, so that treatment is beneficial.
Q6. Why are low coverages for S due to the residual biases in Table I?
Low coverages for ∆̂S are due to the residual biases in Table I, as estimated standard errors from (29) performed well, closely tracking the MC standard deviations.
Q7. Why is DR the efficient estimator in the class?
Because ∆̂DR is the efficient estimator in the class, in large samples, it has smaller variance than ∆̂IPW1 or ∆̂IPW2 , often dramatically so.
Q8. How was the joint distribution of (X, V ) specified?
The joint distribution of (X, V ) was specified by taking X3 ∼ Bernoulli(0.2) and then generating V3 as Bernoulli with P (V3 = 1|X3) = 0.75X3 + 0.25(1−X3).
Q9. What is the effect of the scaling on the probability of a complete case?
the scaling has the effectin practice of offering stability in the case where some complete-case probabilities may be small orare highly variable.
Q10. What is the effect of a covariate profile on the propensity for treatment?
All scenarios are such that values of X associated with lower responses arealso associated with increased propensity for treatment, so that subjects with a covariate profileindicating poor response are those more likely to be treated.
Q11. What is the effect of including V in the propensity score model?
From (32) and these analogous expressions, the effect of including V in the propensity score model is to reduce the variance relative to that in the case where V is excluded.
Q12. What settings were chosen to represent the degree of association of the corresponding covariate to Z?
Settings of β and ξ that achieve the features described above were chosen to represent varyingdegrees of association of the corresponding covariate to Z or Y .
Q13. What is the heuristic account of large-sample results for S?
3.2 Stratification EstimatorsHere, the authors present a heuristic account of large-sample results for ∆̂S and ∆̂SR based on representing the stratification and within-stratum estimation schemes for each as solutions to sets ofestimating equations.
Q14. What is the difference between the two classes of estimators?
as shown in Section 3.2, for fixed K,∆̂S is not consistent and evidently neither ∆̂S nor ∆̂SR makes use of inverse weighting, so these estimators are not members of this class.