I am running the following code to predict gender X education differences in work using multiply imputed data with importance weights. I am working in a restricted VDE and cannot include the output or exact code, but this is close.
mi estimate, saving(miestfile) esample(esample0): mlogit work i.female##i.education [iweight=weight], base(2) vce(cluster pid)
est store edfem
mimrgns using miestfile, esample(esample0) at(education(0(1)3)) over(female) predict(outcome(1)) cmdmargins
marginsplot
*this produces results indicating no gender gap in the role of education for outcome 1 among those with a college degree. IE, the confidence intervals from the predicted probabilities overlap across gender when predicting outcome 1.
Following Mize, I then use spost13 (specifically, mlincom) to calculate the AMEs and contrasts for my interaction.
est restore edfem
mimrgns using miestfile, esample(esample0), at(education(0(1)3)) over(female) predict(outcome1)) post
mlincom, clear
qui mlincom 1, stat(est se p) add rowname (label)
... continues through to row 8
qui mlincom 2-1 stat(est se p) add rowname (label)
... continues through to 8-7
mlincom, title(gender gap in education, outcome 1)
qui mlincom (4-3) - (2-1), stat(est se p) addrowname(label)
mlincom, title (gender gaps and contrasts, outcome 1)
.. continues through (6-5) -(8-7)
this produces predicted probabilities that match the mimrgns results. Specifically, men and women have overlapping CIs in their likelihood to report outcome 1 if they have a college degree:
men: .26 +/- .033
women: .21 +/- .035
The issue is that the gender gap (AME) reported in mlincom is statistically significant (gender gap = .05, se = .025, p=.03). Neither the predicted probabilities + CIs in mlincom nor the predicted probabilities + CIs from mimrgns support this. I even hand calculated the CIs and they overlap. I have also tried solutions from other threads (transform_margins does not work with mlogit, and this code from a previous thread aligns with the CIs I am already getting:
matrix rtable = r(table)
matrix rtable = (rtable[1..2, 1...]\ rtable[5..6, 1...])'
_marg_save , saving(mimrgns_results, double)
use mimrgns_results, clear
list _margin _se _ci_lb _ci_ub , noobs separator(0)
Am I missing something? Is the standard error for the *gender gap/AME* calculated in a way that the confidence intervals for each gender are not necessarily aligned with the significance value of the AME? Any advice is appreciated and I apologize that I cannot include output.
mi estimate, saving(miestfile) esample(esample0): mlogit work i.female##i.education [iweight=weight], base(2) vce(cluster pid)
est store edfem
mimrgns using miestfile, esample(esample0) at(education(0(1)3)) over(female) predict(outcome(1)) cmdmargins
marginsplot
*this produces results indicating no gender gap in the role of education for outcome 1 among those with a college degree. IE, the confidence intervals from the predicted probabilities overlap across gender when predicting outcome 1.
Following Mize, I then use spost13 (specifically, mlincom) to calculate the AMEs and contrasts for my interaction.
est restore edfem
mimrgns using miestfile, esample(esample0), at(education(0(1)3)) over(female) predict(outcome1)) post
mlincom, clear
qui mlincom 1, stat(est se p) add rowname (label)
... continues through to row 8
qui mlincom 2-1 stat(est se p) add rowname (label)
... continues through to 8-7
mlincom, title(gender gap in education, outcome 1)
qui mlincom (4-3) - (2-1), stat(est se p) addrowname(label)
mlincom, title (gender gaps and contrasts, outcome 1)
.. continues through (6-5) -(8-7)
this produces predicted probabilities that match the mimrgns results. Specifically, men and women have overlapping CIs in their likelihood to report outcome 1 if they have a college degree:
men: .26 +/- .033
women: .21 +/- .035
The issue is that the gender gap (AME) reported in mlincom is statistically significant (gender gap = .05, se = .025, p=.03). Neither the predicted probabilities + CIs in mlincom nor the predicted probabilities + CIs from mimrgns support this. I even hand calculated the CIs and they overlap. I have also tried solutions from other threads (transform_margins does not work with mlogit, and this code from a previous thread aligns with the CIs I am already getting:
matrix rtable = r(table)
matrix rtable = (rtable[1..2, 1...]\ rtable[5..6, 1...])'
_marg_save , saving(mimrgns_results, double)
use mimrgns_results, clear
list _margin _se _ci_lb _ci_ub , noobs separator(0)
Am I missing something? Is the standard error for the *gender gap/AME* calculated in a way that the confidence intervals for each gender are not necessarily aligned with the significance value of the AME? Any advice is appreciated and I apologize that I cannot include output.
Comment