movestay. endogenous switching regression, 2+ outcomes selection equation

[natalia malancu]

Hi guys!

I was wondering whether there is a way to run an endogenous switching regression when the selection equation involves more than 2 outcomes. movestay does not seem to support that.

thanks,
Natalia

[FernandoRios]

Hi Natalia.
I do think you might be looking into the user written command Oheckman. The caveat, perhaps is that the first stage is an order probit, instead of a more general multinomial probit.
Perhaps that is what you need
Best
Fernando

[natalia malancu]

hi fernando!

in the first stage i'm looking at obesity levels so an ordered probit makes sense. however, isn't this ado designed for heckman selection models and not endogeneous switch ones?

natalia

[natalia malancu]

i guess my confusion is the result of the fact that as opposed to a heckman situation, in my case i observe the outcome for all regimes...so i know say, what is the wage of a regular weight, over-weight and obese person...

[FernandoRios]

Hi Natalia,
Intrinsically, the Movestay command is an adjusted version of the old -tretreg- (stata 12).
On the first stage it estimates a heckman model, and i the second, it introduces the correction terms (accordingly) into the output equations. The difference with tretreg is that instead of estimating two different models, treatreg uses only one equation model (think of it as the difference between using a dummy, or two separate equations to in a oaxaca-blinder decomposition type of analysis.
The same strategy is used in the oheckman model. For confirmation, why dont you use the oheckman (using only two categories) and you will see they give you the same results (with rounding errors).
Hope this helps
Fernando

[Maxwell Agisu]

Hi guys I would like to use a multinomial endogenous switching regression but I don't know how to do it in STATA. Please if you have a sample do file kindly share it with me. Thanks!

[Bijoyata Yonzon]

Hello Fernando,
I am using Stata 14. I have installed my oheckman command from (SJ7-2: st0123) and don't remember where did i installed movestay from (How do we check the source after installing?). I have been using movestay for my endogenous switching regression (labour market). I wanted to run oaxaca for decompostion, but could not find anything that could help me with oaxaca after movestay estimation. I found in the above comment of yours about oheckman and movstay giving same result. I thought may be i could use oheckman with oaxaca, but there are some error for oheckman command.

Command Movestay:

movestay (LgWg Female1 Rural1 NeverMarried1 TechEdu1 YearsOfEducation Age AgeSq SocialGrp1 SocialGrp2 SocialGrp3 Relign1 Relign2), select(Highwg Female1 Rural1 NeverMarried1 TechEdu1 YearsOfEducation Age AgeSq SocialGrp1 SocialGrp2 SocialGrp3 Relign1 Relign2 HoHEdu_Level ValueOfConsumption30Days)

Fitting initial values .....
Iteration 0: log likelihood = -58226.585
Iteration 1: log likelihood = -58157.096
Iteration 2: log likelihood = -58156.484
Iteration 3: log likelihood = -58156.484

Endogenous switching regression model Number of obs = 59143
Wald chi2(12) = 2794.94
Log likelihood = -58156.484 Prob > chi2 = 0.0000

Coef. Std. Err. z	P>z	[95% Conf.	Interval]
LgWg_1
Female1 .0971449 .0094778 10.25	0.000	.0785687	.1157211
Rural1 .0045184 .0063062 0.72	0.474	.0078414	.0168783
NeverMarried1 -.026173 .0119653 -2.19	0.029	.0496245	-.0027215
TechEdu1 .1318623 .0101472 12.99	0.000	.1119741	.1517505
YearsOfEducation .028232 .000898 31.44	0.000	.0264719	.029992
Age -.0058623 .0026352 -2.22	0.026	.0110272	-.0006973
AgeSq .0002276 .0000317 7.18	0.000	.0001654	.0002897
SocialGrp1 .016831 .0104722 1.61	0.108	.0036943	.0373562
SocialGrp2 -.0551694 .0096879 -5.69	0.000	.0741574	-.0361814
SocialGrp3 -.0799899 .0072541 -11.03	0.000	.0942076	-.0657721
Relign1 -.0136886 .0099006 -1.38	0.167	.0330933	.0057161
Relign2 -.0339452 .0099969 -3.40	0.001	.0535388	-.0143517
_cons 5.972815 .0571467 104.52	0.000	5.860809	6.084821
LgWg_0
Female1 -.332986 .0052325 -63.64	0.000	.3432415	-.3227304
Rural1 -.0076181 .0046665 -1.63	0.103	.0167643	.0015281
NeverMarried1 -.0545071 .0076402 -7.13	0.000	.0694816	-.0395326
TechEdu1 .0618608 .0169353 3.65	0.000	.0286683	.0950534
YearsOfEducation .0030634 .0006276 4.88	0.000	.0018334	.0042934
Age .0164929 .0016059 10.27	0.000	.0133454	.0196403
AgeSq -.0002267 .000021 -10.81	0.000	.0002679	-.0001856
SocialGrp1 -.0509528 .0080047 -6.37	0.000	.0666417	-.035264
SocialGrp2 .0002649 .0067132 0.04	0.969	.0128927	.0134225
SocialGrp3 -.0074975 .005853 -1.28	0.200	.0189692	.0039742
Relign1 -.0139758 .0070668 -1.98	0.048	.0278265	-.0001251
Relign2 .1185262 .0085419 13.88	0.000	.1017843	.135268
_cons 4.572867 .030442 150.22	0.000	4.513202	4.632533
Highwg
Female1 -.6574639 .0170263 -38.61	0.000	-.690835	-.6240929
Rural1 -.2027037 .0127465 -15.90	0.000	.2276863	-.1777211
NeverMarried1 -.0537769 .0227233 -2.37	0.018	.0983138	-.00924
YearsOfEducation .0761885 .0021021 36.24	0.000	.0720684	.0803086
Age .081488 .0049085 16.60	0.000	.0718675	.0911084
AgeSq -.0006553 .0000618 -10.61	0.000	.0007763	-.0005342
SocialGrp2 -.0810457 .0190536 -4.25	0.000	.1183901	-.0437013
SocialGrp3 -.034727 .0154754 -2.24	0.025	.0650582	-.0043958
Relign1 .1032603 .0197185 5.24	0.000	.0646128	.1419078
Relign2 .2597465 .0214172 12.13	0.000	.2177696	.3017235
TechEdu1 .2575649 .0300161 8.58	0.000	.1987345	.3163954
SocialGrp1 .0798819 .0221549 3.61	0.000	.036459	.1233047
HoHEdu_Level .0496705 .002896 17.15	0.000	.0439944	.0553467
ValueOfConsumption30Days .0000673 1.26e-06 53.39	0.000	.0000649	.0000698
_cons -3.558502 .0962428 -36.97	0.000	3.747134	-3.369869
/lns1 -.7197982	.0067224 -107.07 0.000	.7329739	.7066225
/lns2 -.901243	.0041086 -219.35 0.000	.9092957	.8931903
/r1 -.755531	.0200177 -37.74 0.000	.7947649	.7162971
/r2 -.224644	.0185302 -12.12 0.000	.2609627	.1883254
sigma_1 .4868505	.0032728	.480478	.4933076
sigma_2 .4060646	.0016684	.4028078	.4093477
rho_1 -.6384371	.0118584	.6610999	.6146104
rho_2 -.2209399	.0176257	.2551957	.1861302
LR test of indep. eqns. :	chi2(1) = 11561.06	Prob > chi2	= 0.0000

Command oheckman:
oheckman LgWg Female1 Rural1 NeverMarried1 TechEdu1 YearsOfEducation Age AgeSq SocialGrp1 SocialGrp2 SocialGrp3 Relign1 Relign2, select(Highwg Female1 Rural1 NeverMarried1 TechEdu1 YearsOfEducation Age AgeSq SocialGrp1 SocialGrp2 SocialGrp3 Relign1 Relign2 HoHEdu_Level ValueOfConsumption30Days)
Result
initial values not feasible
r(1400);

Also, while using oaxaca command, I got confused about the categorical option in the command syntax. I do have some dummies

1. Do I need to write down all the dummy variables used in this option, even the two category variables which has one dummy?
2. Since in the categorical option we write all the dummies even the reference category, it would mean 3 dummies for 3 category variable?, do we write these reference category dummy in the model1 and model2 option as well and in the indepvars?

syntax from Ben Jann (2004) https://core.ac.uk/download/pdf/6442665.pdf:
oaxaca depvar indepvars if in weight , by(groupvar) swap detail (dlist) adjust(varlist) threefold (reverse) weight(# # . . . ) pooled (model opts) omega (model opts) reference(name) split x1(names and values) x2(names and values) categorical(clist) svy (svyspec) vce(vcetype) cluster(varname) fixed (varlist) no suest nose model1(model opts) model2(model opts) xb noisily level(#) eform

Sorry for the long post and number of questions.
Thank you

[FernandoRios]

Hi Bijoyata
Im surprised about what you find with oheckman. As I said before, Oheckman is a generalization of the movestay command. So they should provide the same output, with rounding errors.
But perhaps they way it was programmed simply has less stable conditions for the maximization.
That being said. There is no clear suggestion, as far as I know, about how to correct for sample selection into each group when implementing oaxaca Decompositions.
There is a discussion in Neuman and Oaxaca (2004) https://link.springer.com/article/10...28395.38694.4b
The simplest approach is to elaborate the decomposition as a function of the selectivity corrected coefficients, excluding the selection term.
In that case, you could do the following:

probit group x1 x2 x3
predict imr, score
oaxaca y x1 x2 x3, by(group) adjust(imr) w(0)

This should adjust for IMR (the selection correction term, before doing the decomposition.
HTH
Fernando

[Bijoyata Yonzon]

Thank you Fernando. Will try and let you know.

[Bijoyata Yonzon]

Hi Fernando,
I did run the above command and

oaxaca LgWg Rural1 NeverMarried1 TechEdu1 YearsOfSchooling SocialGrp1 SocialGrp2 SocialGrp3 Relign1 Relign2 Age, by(Female1) adjust(imr) w(0)

adjust(): imr not found in models
r(111);

Then I used imr as one of the regressor after which I got the following result:

oaxaca LgWg Rural1 NeverMarried1 TechEdu1 YearsOfSchooling SocialGrp1 SocialGrp2 SocialGrp3 Relign1 Relign2 Age imr, by(Female1) adjust(imr) w(0)

Blinder-Oaxaca decomposition Number of obs = 84,279
Model = linear
Group 1: Female1 = 0 N of obs 1 = 63988
Group 2: Female1 = 1 N of obs 2 = 20291

LgWg Coef.	Std. Err.	z	P>z	[95% Conf.	Interval]
overall
group_1 4.526128	.003349	1351.48	0.000	4.519564	4.532692
group_2 3.882628	.0063409	612.31	0.000	3.8702	3.895056
difference .6435	.007171	89.74	0.000	.629445	.6575549
adjusted
group_1 3.65692	.0608573	60.09	0.000	3.537642	3.776198
group_2 -8.936544	.8185817	-10.92	0.000	-10.54094	-7.332154
difference 12.59346	.8208408	15.34	0.000	10.98465	14.20228
explained -1.057167	.0827838	-12.77	0.000	-1.21942	-.8949137
unexplained 13.65063	.9015892	15.14	0.000	11.88355	15.41771
explained
Rural1 .0174847	.0013835	12.64	0.000	.014773	.0201964
NeverMarried1 -.1560421	.0118881	-13.13	0.000	-.1793424	-.1327418
TechEdu1 .0374981	.0054727	6.85	0.000	.0267718	.0482244
YearsOfSchooling -.7378792	.0620924	-11.88	0.000	-.859578	-.6161804
SocialGrp1 -.0178812	.0020675	-8.65	0.000	-.0219334	-.013829
SocialGrp2 .0121051	.0015508	7.81	0.000	.0090657	.0151445
SocialGrp3 -.0046387	.0009473	-4.90	0.000	-.0064953	-.0027821
Relign1 -.2045997	.0150904	-13.56	0.000	-.2341762	-.1750231
Relign2 .0007131	.0015656	0.46	0.649	-.0023553	.0037816
Age -.0039271	.0043378	-0.91	0.365	-.0124291	.0045748
unexplained
Rural1 .072083	.0077662	9.28	0.000	.0568615	.0873046
NeverMarried1 .599324	.0345889	17.33	0.000	.5315311	.6671169
TechEdu1 -.1809348	.0105748	-17.11	0.000	-.2016612	-.1602085
YearsOfSchooling 3.050308	.1807955	16.87	0.000	2.695956	3.404661
SocialGrp1 -.0926602	.0049768	-18.62	0.000	-.1024146	-.0829059
SocialGrp2 .0611215	.005582	10.95	0.000	.050181	.0720621
SocialGrp3 -.1291391	.0088536	-14.59	0.000	-.1464918	-.1117863
Relign1 .4689542	.0275437	17.03	0.000	.4149695	.522939
Relign2 -.0432731	.0034984	-12.37	0.000	-.0501297	-.0364164
Age 2.400971	.1238874	19.38	0.000	2.158156	2.643785
_cons 7.443877	.5650106	13.17	0.000	6.336477	8.551278
(adjusted by imr)

Thank you

[FernandoRios]

Hi Bijoyata
I just look at your results, and It seems odd to me that you are now finding that the expected wage of women is negative! (well, almost zero).
Something I didnt ask. How are you defining your probit. Is "female=1" your dependent variable? Or is it "earn high wages"? I think that the IMR will have to be adjusted, depending on what type of selection you are trying to account for here.
Best
Fernando

[Bijoyata Yonzon]

Hi Fernando,
Thank you for noticing the odd result. Yes I have taken female=1 n my probit function.

probit Female1 Rural1 NeverMarried1 TechEdu1 YearsOfSchooling SocialGrp1 SocialGrp2 SocialGrp3 Relign1 Relign2 Age

Iteration 0: log likelihood = -46528.897
Iteration 1: log likelihood = -44490.64
Iteration 2: log likelihood = -44478.906
Iteration 3: log likelihood = -44478.903

Probit regression Number of obs = 84,293
LR chi2(10) = 4099.99
Prob > chi2 = 0.0000
Log likelihood = -44478.903 Pseudo R2 = 0.0441

----------------------------------------------------------------------------------
Female1 | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-----------------+----------------------------------------------------------------
Rural1 | -.0161325 .0103094 -1.56 0.118 -.0363387 .0040736
NeverMarried1 | -.280515 .0146186 -19.19 0.000 -.309167 -.251863
TechEdu1 | .2913284 .0219962 13.24 0.000 .2482167 .3344401
YearsOfSchooling | -.0513405 .001036 -49.56 0.000 -.0533709 -.04931
SocialGrp1 | .0710058 .0173734 4.09 0.000 .0369546 .1050571
SocialGrp2 | -.0406757 .0148374 -2.74 0.006 -.0697564 -.011595
SocialGrp3 | .0408341 .0128577 3.18 0.001 .0156334 .0660348
Relign1 | -.4472978 .0185717 -24.08 0.000 -.4836977 -.4108979
Relign2 | .0486743 .0164405 2.96 0.003 .0164515 .0808972
Age | -.0069367 .0004892 -14.18 0.000 -.0078955 -.0059778
_cons | -.088205 .0245496 -3.59 0.000 -.1363214 -.0400886
----------------------------------------------------------------------------------

[Bijoyata Yonzon]

Hi Fernando,
the probit fuction imr is giving me the same doubtful result, since I had ready ran movestay command and realised that even with movestay command I can estimate imr (mills ratio) for both regimes with mspredict.. varname, imr1 and mspredict varname, imr2 command. Would it be correct to use these mills ratio as my selectivity correction term?
When I used these imr I got an explainable result, but the results are different for each one. What difference will be in my result by using the imr for regime 1 and imr for regime ?.

[Mohammed Adem Ali]

why movestay command result is not convergent?