Deciding on best matching algorithm with kmatch

[Linda WUR]

Dear statalisters,

Based on the paper of King and Nielson (2016) on the comparative analysis of matching methods, I decided to not 'blindly' decide on using propensity scores for matching but also applied other algorithms (Mahalanobis and Exact) to analyse the most efficient one. For this aim, I am working with the kmatch command and have tried a few things:

Kernel matching:
 kmatch md round age_1 schoolatt childrenattendschool partner howmanyHHM_1   ( pcaenvironmental1 pcaenvironmental2 ), att nate
kmatch em round age_1 schoolatt childrenattendschool partner howmanyHHM_1   ( pcaenvironmental1 pcaenvironmental2 ), att nate
kmatch ps round age_1 schoolatt childrenattendschool partner howmanyHHM_1   ( pcaenvironmental1 pcaenvironmental2 ), att nn

Nearest neighbour: 1:1 and 1:5
kmatch ps round age_1 schoolatt childrenattendschool partner howmanyHHM_1   ( pcaenvironmental1 pcaenvironmental2 ), att nn
kmatch md round age_1 schoolatt childrenattendschool partner howmanyHHM_1   ( pcaenvironmental1 pcaenvironmental2 ), att nn
kmatch md round age_1 schoolatt childrenattendschool partner howmanyHHM_1   ( pcaenvironmental1 pcaenvironmental2 ), att nn(5)
kmatch ps round age_1 schoolatt childrenattendschool partner howmanyHHM_1   ( pcaenvironmental1 pcaenvironmental2 ), att nn(5)

However, whereas for propensity score matching under psmatch2 it is quite clear how to analyse the common support (in a nice, and visual manner) I have not really found out how to clearly observe the best method. The outputs I get are e.g. for Kernel matching under kmatch:

Mahalanobis

Propensity Score

Exact matching

Based on this, I would assume that the latter (exact matching) is the best option as the standard deviation is the lowest as well as the variance ratio closest to 1? However, this method only used 41 out of the 150 treated cases..

I am a bit confused, would be great if someone would have some advise!

Linda

[Felix Bittmann]

Well, what you have to decide on is a compromise between bias and precision. Fewer cases with a better match might reduce bias but standard errors become larger. This is a common problem in statistics. Maybe you can see some similar studies for an orientation. Also make sure to both use analytical and bootstrapped standard errors after deciding for an algorithm as an additional robustness check. In my opinion, NN matching is the only algorithm I would consider to be rather outdated. All others might produce similar and fine results.