Analizing direct and indirect effect in a panel data using SEM command

[Jose Ferreras]

Good afternoon,

I have a panel data that looks like the following:

ID	year	Bth	Dth	int_y	ptn_y	o_in	m_in	patnum	pat_e	sales	size	exportn	industry	gextid	country	age
20	2008	8	2	1	3	1	1	0	0	39389620	63	25.1	15	7.4	1	61
20	2011	10	2	4	5	3	4	1	0	52623691	152	40.7	15	0	1	64
20	2014	10	3	1	6	3	4	1	0	110100000	101	36.6	15	42.4	1	67
22	2008	10	1	2	3	2	0	0	0	496600000	1489	7	22	0	1	66
22	2011	10	4	1	6	3	2	8	4	511000000	1784	40.6	22	0	1	69
22	2014	10	4	2	5	2	0	7	2	384200000	1569	72.7	22	0	1	72
25	2008	5	0	0	0	0	0	0	0	21834200	154	5.4	15	16.4	1	60
25	2011	8	0	0	0	0	0	0	0	21796024	134	4.9	15	2	1	63
25	2014	8	0	0	1	0	1	0	0	23578268	124	6.4	15	0	1	66
27	2008	9	3	1	3	2	1	4	3	52746431	462	1	12	3.3	1	44
27	2011	9	3	1	3	1	0	0	0	42999533	330	0	12	0	1	47
27	2014	8	0	1	2	1	1	0	0	44002188	357	51.9	12	0	1	50
30	2008	7	3	2	3	3	1	0	0	574600000	126	0	13	31.3	1	43
30	2011	10	5	0	0	2	2	0	0	381700000	134	4.2	13	12	1	46
30	2014	10	4	1	2	0	0	1	0	354700000	121	2.2	13	5.5	1	49
32	2008	7	2	0	2	2	1	0	0	37786689	88	8.9	23	0	1	39
32	2011	7	2	0	0	0	2	0	0	17159611	67	0.9	23	0	1	42
32	2014	8	4	0	1	0	0	0	0	12281532	83	1	23	0	1	45
33	2008	2	0	0	0	2	3	0	0	62449255	378	15	14	0	1	34
33	2011	2	0	0	0	2	3	0	0	31658394	240	0	14	0	1	37
33	2014	1	0	0	0	0	0	0	0	26131847	200	0	14	0	1	40
36	2008	6	2	0	0	0	0	2	2	4349675	56	0	18	0	1	52
36	2011	5	2	0	0	3	4	5	5	8511439	92	0	18	0	1	55
36	2014	6	2	2	1	2	3	0	0	7360923	82	50.3	18	0	1	58
40	2008	10	2	1	4	0	0	0	0	25614595	44	0	10	7.4	0	33
40	2011	10	3	1	2	0	0	0	0	25171278	45	0.8	10	6.3	0	36
40	2014	10	4	0	0	0	0	0	0	24140456	44	2.5	10	0	0	39
42	2008	6	0	1	2	1	0	0	0	372300000	238	3.5	14	0	1	33
42	2011	9	0	1	3	3	0	0	0	262400000	225	5.2	14	20.3	1	36
42	2014	9	0	1	5	3	0	0	0	284300000	228	0.5	14	16.4	1	39
43	2008	8	1	0	0	2	1	0	0	18024408	80	0	19	20.9	1	35
43	2011	7	0	0	0	2	1	0	0	19699301	90	0	19	16.1	1	38
43	2014	7	0	0	0	2	0	0	0	20152976	99	0	19	0	1	41
46	2008	8	3	1	3	1	1	0	0	17146389	75	0.5	17	2.4	1	30
46	2011	9	2	0	2	1	1	0	0	20721491	70	0.3	17	2.7	1	33
46	2014	7	2	0	0	2	0	0	0	27607344	75	0.3	17	19.3	1	36
57	2008	6	2	0	0	0	0	1	0	22878934	21	1.9	17	99.5	1	20
57	2011	10	4	0	0	0	0	1	0	14147998	26	1.9	17	100	1	23
57	2014	10	4	0	0	0	1	1	0	10609470	26	2.8	17	100	1	26
62	2008	10	2	0	0	0	0	0	0	2290688	17	34.5	16	0	1	15
62	2011	10	2	0	0	0	0	0	0	2346570	15	15.2	16	0	1	18
62	2014	10	0	0	0	0	0	0	0	1719451	14	8.9	16	0	1	21
63	2008	9	1	0	0	0	0	0	0	21529736	108	42.9	15	0	1	18
64	2008	10	3	3	4	3	4	3	2	10300167	52	51.6	15	0	1	18
64	2011	10	1	0	1	3	4	0	0	9212141	35	44	15	0	1	21
64	2014	9	5	0	5	3	3	1	0	10609470	40	46.4	15	0	1	24
65	2011	5	2	0	0	2	0	0	0	3239057	30	30.9	12	0	1	21
66	2008	9	0	0	1	0	0	4	1	22136526	100	21.8	17	6.7	0	17
66	2011	10	0	0	0	0	0	2	1	12129950	77	25	17	0	1	20
66	2014	10	0	0	0	0	2	0	0	11536365	62	51.4	17	21.2	1	23
68	2008	6	2	0	0	0	0	0	0	19244274	107	0.4	17	13.6	1
72	2008	7	1	2	4	1	0	0	0	375100000	833	47.7	14	7.5	1	15
72	2011	7	1	1	4	3	0	0	0	231000000	749	38.1	14	25.2	1	18
72	2014	7	2	1	5	1	0	0	0	324900000	798	53.1	14	21.2	1	21
74	2008	1	1	0	0	0	0	0	0	17487783	307	0	21	0	1	15
74	2014	4	0	0	0	0	0	0	0	49747173	2591	4.4	21	100	1	21
75	2008	1	0	0	0	0	0	0	0	182500000	248	0.9	14	0	1	14
75	2014	5	2	0	1	0	0	0	0	69436368	190	2.5	14	0	1	20
77	2008	10	2	1	2	2	0	0	0	6023716	50	7	15	10	1	14
77	2011	10	0	0	2	1	0	0	0	7237005	58	0	15	0	1	17
82	2008	10	2	0	0	0	0	0	0	42801299	100	0	18	0	0	12
82	2011	10	2	0	0	0	0	0	0	26827256	52	0	18	0	0	15
82	2014	10	2	0	4	0	0	13	0	11847118	55	0	18	0	0	18
88	2008	5	0	0	0	0	0	0	0	61785588	365	67.8	15	0	1	11
88	2011	5	0	0	0	0	0	0	0	70115789	332	34.5	15	0	1	14
88	2014	4	0	0	0	0	0	1	0	77315502	372	41	15	0	1	17

In this panel data:

Independent variables: Bth, Dth, int_y, ptn_y

Mediating variables: o_in, m_in

Depending variables: patnum, pat_epus

Control variables: sales, size, exportn, industry, gextid, country, newage

With this panel data I would like to analyze the direct and indirect effect that the independents variables have on the dependent variables. Looking for recommendation of how to perform this analysis I have found that in stata, the SEM command can be used to obtain both the direct and indirect effect when performing mediation. As the data need to be arranged in wide format to perform the analysis, a limitation that I have seen is that I am only able to perfom the analysis in a cross sectional matter. Is there a way to perform this analysis with SEM including the longitudinal characteristic of the dataset?

Also, as the dependent variables patnum and patent_epus represent the number of “pat” requested of each type for the firms in the years analyzed, is it ok to use the “sem” command or is it better to use the “gsem” command?

Thank you very much in advanced for your help.

Jose Luis

[Roman Mostazir]

While both SEM and GSEM can be applied to test mediation analysis, for nesting data structure, you need GSEM (Generalized Structural Equation Model) command. Once the model is fitted for random intercepts/coefficients, you can test the indirect effects which are tested using post-estimation command -nlcom- for path x->z->y. See the help file for -sem- and -gsem-

[Richard Williams]

You may want to look at xtdpdml, available from SSC. What you want to do is more complicated than what it does but it can give you some ideas of how to code it. The support page is at

https://www3.nd.edu/~rwilliam/dynamic/index.html

See especially sections 2.3 and 4 of

https://www3.nd.edu/~rwilliam/dynamic/SJPaper.pdf

[Jose Ferreras]

Thank you very much for your replying. Is it possible to perform the gsem command or the xtdpdml in Stata 12?

[Richard Williams]

Xtdpdml, probably. Use the v12 option. I don’t think gsem showed up til later but I could be wrong. In any event you can look at xtdpdml to see how it does things and then see if you can get ideas for your own problem.

[Stephanie Retz]

Dear all,
Dear Richard,

I have unbalanced panel data for a maximum of four waves per individual. Based on this data I estimate the following regression with individual fixed effects:

xtreg y i.x1 x2 i.x3 x4, fe vce(cluster id)

where y is a 5-category variable and x1 is also a 5-category variable.

Now, I would like to decompose the effect of x1 into a direct effect and an indirect effect via mediation analysis. The mediator that I have in mind is a continuous variable z, which has a positive effect on y, while x1 has a negative effect on y.

How can I estimate the direct and indirect effect while controlling for unobserved heterogeneity using individual fixed effects?

Does xtdpdml by Richard Williams fit my problem? And if so, how would I apply the command, i.e. how to indicate that I have panel data, that I want fixed effects and that I want the standard errors clustered at the individual level etc.?

I have posted this question separately earlier last week, but still don't know how to go ahead: https://www.statalist.org/forums/for...direct-effects I hope, someone here can help me.

Thanks a lot,
Stephanie

(I work with Stata 15 on Windows.)

[Stephanie Retz]

Dear all,

I would really apprechiate some help regarding my problem. Please let me know in case my question was not stated clearly and I will try to rephrase it.

Best wishes,
Staphanie

[Richard Williams]

Sorry, I did not see the September 2019 message.

I don't know if xtdpdml can do what you want or not. I have never tried it.

You could try specifying the staywide option in xtdpdml. The data then stay in wide format. The full range of sem post-estimation commands should then be available to you, including estat teffects.

xtdpdml also supports the vce option, although I've only tried a few options with it.

You can also output generated code into a file. If the generated code is close but not quite to the code you want, you may be able to make some tweaks by hand.

Again, I don't really know. But I have often found my programs can do things I had never thought of because they can take advantage of features built into Stata.

A few other points: if y and x1 are 5 category variables, do you think it is appropriate to treat them as continuous? sem estimates linear models.

There isn't going to be a single x1. Instead there will be an x1time1, x1time2, etc. Does that really give you want?

xtdpdml is going to include lagged y as a regressor by default. If you don't want that, you should use the ylag(0) option.

[Stephanie Retz]

Dear Richard,
Thank you very much for your response.
I will look closer into xtdpdml and get back to you if I have further questions.
All the best,
Stephanie

[John Sgr]

Dear Richard,

As you said staywide options enables to use post-estimation commands of SEM. I have a particular question on that. I have a panel data for 700 individuals throughout 12 months. Once I run xtdpml by using option staywide and esttab teffects I get direct, indirect, total effects for all lags. But this is not what I'm looking for. I am more interested in average effects as a whole rather than effects per all lags separately.

Do you have any suggestion about how to get average direct, indirect effects like a cross-sectional design?

Thanks in advance!

Best regards,
John