mimrgns and emargins - average marginal effects the same as coefficient values

[Mollie Bourne]

Dear all,

I have run a logistic regression model on mi data and now I would like to get average marginal effects.

I have tried using both the mimrgns command and the emargins program, but both of these return average marginal effects values that are the same as the original coefficient values. Please see my code and output below:

1. Example with mimrgns:

mi estimate: logit th ib2.cohort i.fert c.stat c.edu i.sex1

Output:
th	Coef.	Std. Err.	t	P	[95% Conf.	Interval]
cohort (ref 2)
1. Cohort1	-.0611989	.0289399	-2.11	0.036	-.1182399	-.004158
3. Cohort 3	.6489638	.0336769	19.27	0.000	.5825208	.7154069
fert (ref 1)
2.	-.3343231	.0612948	-5.45	0.000	-.4555199	-.2131262
3.	-.3815099	.077192	-4.94	0.000	-.5355125	-.2275072
4.	-.5661	.0702629	-8.06	0.000	-.7053152	-.4268849
5.	-.5632703	.069933	-8.05	0.000	-.7020042	-.4245364
6.	-.7277115	.0733454	-9.92	0.000	-.8734783	-.5819447
7.	-.9045366	.0735842	-12.29	0.000	-1.050667	-.7584065
stat	.8950652	.0698831	12.81	0.000	.7563983	1.033732
edu	.9673386	.0453642	21.32	0.000	.877517	1.05716
1.sex1	.1226577	.0234141	5.24	0.000	.0765511	.1687643
_cons	.4311072	.077804	5.54	0.000	.2770226	.5851919

mimrgns, dydx(cohort fert stat edu sex1)

Output:

Multiple-imputation estimates	Imputations	= 20
Average marginal effects	Number of obs	= 52,537
Average RVI	= 0.5409
Largest FMI	= 0.5388
DF adjustment: Large sample	DF: min	= 68.78
avg	= 134.24
Within VCE type: Delta-method	max	= 258.53
Expression : Linear prediction	(log odds), predict(xb)
dy/dx w.r.t. : 1.cohort 3.cohort	2.fert 3.fert 4.fert 5.fert	6.fert 7.fert	stat	edu..

dy/dx	Std. Err.	t	P>t	[95% Conf.	Interval]
cohort
1. Cohort 1	-.0611989	.0289399	-2.11	0.036	-.1182399	-.004158
3. Cohort 3	.6489638	.0336769	19.27	0.000	.5825208	.7154069
fert
2.	-.3343231	.0612948	-5.45	0.000	-.4555199	-.2131262
3.	-.3815099	.077192	-4.94	0.000	-.5355125	-.2275072
4.	-.5661	.0702629	-8.06	0.000	-.7053152	-.4268849
5.	-.5632703	.069933	-8.05	0.000	-.7020042	-.4245364
6.	-.7277115	.0733454	-9.92	0.000	-.8734783	-.5819447
7.	-.9045366	.0735842	-12.29	0.000	-1.050667	-.7584065
stat	.8950652	.0698831	12.81	0.000	.7563983	1.033732
edu	.9673386	.0453642	21.32	0.000	.877517	1.05716
1.sex1	.1226577	.0234141	5.24	0.000	.0765511	.1687643

2. Example with emargins:

program emargins, eclass properties(mi)
version 12
args outcome
logit th ib2.cohort i.fert c.stat c.edu i.sex1
margins, dydx(cohort fert stat edu sex1)
end
mi estimate, cmdok: emargins 1

Output:

Multiple-imputation estimates	Imputations	=	20
Logistic regression	Number of obs	=	51,576
Average RVI	=	0.5032
Largest FMI	=	0.5284
DF adjustment: Large sample	DF: min	=	71.54
avg	=	138.78
max	=	239.46
Model F test: Equal FMI	F( 11, 1732.8)	=	308.70
Within VCE type: OIM	Prob > F	=	0.0000

th	Coef.	Std. Err.	t	P>t	[95% Conf.	Interval]
cohort
1. Cohort1	-.0625422	.0291314	-2.15	0.033	-.119964	-.0051203
3. Cohort3	.6486531	.0335712	19.32	0.000	.5824863	.7148199
fert
2.	-.3334438	.0601941	-5.54	0.000	-.4522741	-.2146135
3.	-.3806243	.0770744	-4.94	0.000	-.5342863	-.2269623
4.	-.5691863	.0703006	-8.10	0.000	-.7084076	-.4299649
5.	-.5655533	.0698688	-8.09	0.000	-.704074	-.4270326
6.	-.7310476	.0739668	-9.88	0.000	-.8780468	-.5840485
7.	-.9087949	.0740891	-12.27	0.000	-1.055914	-.7616757
stat	.8871267	.0699085	12.69	0.000	.7484814	1.025772
edu	.9617701	.0448082	21.46	0.000	.8731478	1.050392
sex1
2. Female	.1203853	.0237677	5.07	0.000	.0735649	.1672058
_cons	.4392243	.0789807	5.56	0.000	.2827478	.5957008

Please can anyone advise on why this command does not appear to be working?

Thank you in advance.


Mollie Bourne

[daniel klein]

First, note that the two approaches yield different results! The reported sample size does also differ. You need to address these issues.

The second approach should be expected to be the same as the original coefficients, because you omit the post option in the margins call in your emargins program. This way, mi reports the combined results from logit. Adding the post option will yield (average) marginal effects in terms of probabilities - the default after logit.

The mimrgns command (from SSC) reports (average) marginal effects in terms of the linear predictor, as you can clearly see in the output, stating

Expression : Linear prediction (log odds), predict(xb)

The help file explains why this is the case and how you can get predicted probabilities. I suggest you read the help carefully, then code

mimrgns , dydx(cohort fert stat edu sex) predict(pr)

Best
Daniel

Edit:
Please use code delimiters to post results.

[Mollie Bourne]

Great - thank you very much!

[Steffen Mauch]

daniel klein Do I understand correctly that

mimrgns , dydx(cohort fert stat edu sex) predict(pr)

does the same thing in case of imputed data as

margins if e(sample), dydx(cohort fert stat edu sex) post

after a normal probit command? Or is it necessary to use the -predict-option in case of the margins command as well?

[daniel klein]

It is never a problem to explicitly specify the predict() option. It also makes the code more readable. If you omit the option, margins' default prediction depends on the last estimation command. After probit, margins default to predict(pr).