Hi all,
I want to compare the effects from ATE and ATET in propensity score and nearest neighbor matching estimations. I set a caliper to eliminate some observations and satisfy the common support requirement. For both commands this takes some iteration before the errors stop (see below), and at the end t he number of eliminated observations differ (which I understand).
a) teffects psmatch (y) (treatment matching_vars*, probit), caliper(0.005) vce(iid) osample(osample_ate#)
...teffects psmatch (y) (treatment matching_vars*, probit) if osample_ate# == 0, caliper(0.005) vce(iid)
b) teffects psmatch (y) (treatment matching_vars*, probit), caliper(0.005) vce(iid) osample(osample_atet#) atet
...teffects psmatch (y) (treatment matching_vars*, probit) if osample_atet# == 0, caliper(0.005) vce(iid) atet
After a) and b) I generated graphs to show the common support (see below):
teffects overlap, ptlevel(1)
(ATE)
(ATET)
Question 1: Why are observations with very low propensity scores from the control croup NOT eliminated when doing the atet estimation? The graph clearly shows that this sample is unbalanced. Can I still use it as it is?
Question 2: How do I set the caliper for teffects nnmatch? It does not use propensity scores so the rule of caliper = 0.2*sd(propensity core) does not work here. Most of my observations are eliminated at that value.
Best,
Robin
I want to compare the effects from ATE and ATET in propensity score and nearest neighbor matching estimations. I set a caliper to eliminate some observations and satisfy the common support requirement. For both commands this takes some iteration before the errors stop (see below), and at the end t he number of eliminated observations differ (which I understand).
a) teffects psmatch (y) (treatment matching_vars*, probit), caliper(0.005) vce(iid) osample(osample_ate#)
...teffects psmatch (y) (treatment matching_vars*, probit) if osample_ate# == 0, caliper(0.005) vce(iid)
b) teffects psmatch (y) (treatment matching_vars*, probit), caliper(0.005) vce(iid) osample(osample_atet#) atet
...teffects psmatch (y) (treatment matching_vars*, probit) if osample_atet# == 0, caliper(0.005) vce(iid) atet
After a) and b) I generated graphs to show the common support (see below):
teffects overlap, ptlevel(1)
(ATE)
Question 1: Why are observations with very low propensity scores from the control croup NOT eliminated when doing the atet estimation? The graph clearly shows that this sample is unbalanced. Can I still use it as it is?
Question 2: How do I set the caliper for teffects nnmatch? It does not use propensity scores so the rule of caliper = 0.2*sd(propensity core) does not work here. Most of my observations are eliminated at that value.
Best,
Robin
