The simplest method to perform propensity score matching is one-to-one greedy matching. Even though more modern methods, such as genetic matching and optimal matching will perform better than one-to-one greedy matching if evaluated across a large number of studies, one-to-one greedy matching is able to obtain adequate covariate balance in many situations. In the video below, I demonstrate one-to-one greed matching.
Variable ratio greedy matching is a variation of one-to-one greedy matching where the number of matches is determined by the data. For example, some treated cases will receive one match from the control group, while others may receive more than one match. This method results in weights to be used in the analysis to account for the fact that treated cases received different numbers of matches.
In the video below I demonstrate optimal full matching, which uses network flow optimization algorithms to create strata containing at least one treated and one control observations, minimizing the distance between them with respect to propensity scores. One way to think of optimal full matching is that it is a type of stratification where the maximum number of strata is determined from the data by an optimization algorithm.
In the video below I demonstrate genetic matching, which is matching that is optimized with a genetic algorithm. Genetic algorithm is a class of metaheuristic optimization algorithms based on principles of genetics, such as crossover and mutation. One advantage of matching with a genetic algorithm over other methods is that it can matching using the covariate values themselves instead of the propensity score. Therefore, it allows the researcher to skip the propensity score estimation step.
Data for Propensity Score Matching Example
| chapter5_data_breastfeeding_example.rdata |
| chapter5_data_with_propensity_scores_and_formula.rdata |
R Notebook for Propensity Score Matching
| guide_to_propensity_score_matching.rmd |
R Code for Propensity Score Matching
| chapter5_part1__propensity_score_estimation.r |
| chapter5_part2__one_to_one_greedy_matching.r |
| chapter5_part3_variable_ratio_greedy_matching.r |
| chapter5_part4_genetic_matching_with_propensity_scores_plus_covariates.r |
| chapter5_part5_genetic_matching_with_propensity_scores_only.r |
| chapter5_part6_optimal_one_to_one_matching.r |
| chapter5_part7_optimal_full_matching.r |
Related Publications:
Leite, W. L. (2015). Latent growth modeling of longitudinal data with propensity score matched groups In Wei Pan, & Haiyan Bai. Propensity Score Analysis: Fundamentals, Developments, and Extensions, (pp. 191-216.) New York: Guilford.
Leite, W. L., Sandbach, R., Jin, R., MacInnes, J., & Jackman, G. A. (2012). An Evaluation of Latent Growth Models for Propensity Score Matched Groups. Structural Equation Modeling. 19, 437–456. DOI: 10.1080/10705511.2012.687666
Lee, Y., Leite, W. L., & Leroux, A. J. (2023). Propensity Score Matching with Cross-Classified Data Structures: A Comparison of Methods. The Journal of Experimental Education, 1-18. https://doi.org/10.1080/00220973.2023.2164843
Leite, W. L. (2015). Latent growth modeling of longitudinal data with propensity score matched groups In Wei Pan, & Haiyan Bai. Propensity Score Analysis: Fundamentals, Developments, and Extensions, (pp. 191-216.) New York: Guilford.
Leite, W. L., Sandbach, R., Jin, R., MacInnes, J., & Jackman, G. A. (2012). An Evaluation of Latent Growth Models for Propensity Score Matched Groups. Structural Equation Modeling. 19, 437–456. DOI: 10.1080/10705511.2012.687666
Lee, Y., Leite, W. L., & Leroux, A. J. (2023). Propensity Score Matching with Cross-Classified Data Structures: A Comparison of Methods. The Journal of Experimental Education, 1-18. https://doi.org/10.1080/00220973.2023.2164843