Usage
Workflow, options, margins after suest2, and the rules for weights, survey data, and MI
suest2 model1name model2name [, options]model#name are stored estimates. Store each model with estimates store before running suest2.
The workflow
Fit each model, store it, combine, then do the post-estimation you want on the combined system. We recommend fitting the models with vce(robust) or vce(cluster varname) so that the suest2 results match the individual model results exactly.
sysuse nlsw88, clear(NLSW, 1988 extract)drop if missing(union, wage, ttl_exp, collgrad, age, south)(368 observations deleted)
. quietly logit union c.age##i.collgrad i.south, vce(robust)
estimates store m1
quietly logit union c.age##i.collgrad i.south c.wage c.ttl_exp, vce(robust)
estimates store m2. suest2 m1 m2Simultaneous results for m1, m2 Number of obs = 1,878
--------------------------------------------------------------------------------
| Robust
| Coefficient std. err. z P>|z| [95% conf. interval]
---------------+----------------------------------------------------------------
m1_union |
age | -0.002 0.021 -0.092 0.927 -0.043 0.039
|
collgrad |
College grad | -1.007 1.564 -0.644 0.520 -4.072 2.058
|
collgrad#c.age |
College grad | 0.039 0.040 0.973 0.331 -0.039 0.116
|
south |
South | -0.745 0.116 -6.408 0.000 -0.972 -0.517
_cons | -0.901 0.828 -1.088 0.277 -2.525 0.722
---------------+----------------------------------------------------------------
m2_union |
age | -0.004 0.021 -0.187 0.852 -0.046 0.038
|
collgrad |
College grad | -1.284 1.601 -0.802 0.423 -4.422 1.855
|
collgrad#c.age |
College grad | 0.041 0.040 1.004 0.315 -0.039 0.120
|
south |
South | -0.675 0.119 -5.690 0.000 -0.907 -0.442
wage | 0.054 0.014 3.843 0.000 0.026 0.081
ttl_exp | 0.006 0.013 0.484 0.629 -0.019 0.032
_cons | -1.302 0.838 -1.554 0.120 -2.945 0.340
--------------------------------------------------------------------------------What the combined results contain
e(b) holds every model’s coefficients, one equation per model (named model_depvar), and e(V) their joint covariance matrix. Any command that works on e(b) and e(V) works here.
test [m1_union]1.collgrad = [m2_union]1.collgrad ( 1) [m1_union]1.collgrad - [m2_union]1.collgrad = 0
chi2( 1) = 2.06
Prob > chi2 = 0.1516nlcom ([m1_union]1.collgrad - [m2_union]1.collgrad) / [m1_union]1.collgrad _nl_1: ([m1_union]1.collgrad - [m2_union]1.collgrad) / [m1_union]1.collgrad
------------------------------------------------------------------------------
| Coefficient Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
_nl_1 | -0.274 0.443 -0.620 0.536 -1.142 0.594
------------------------------------------------------------------------------The other stored results are listed in the help file.
margins after suest2
A bare margins gives predictions from every model. To use one model, name it in predict() with model():
margins, dydx(age) predict(model(m1) pr)------------------------------------------------------------------------------
| Delta-method
| Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
m1 |
age | 0.002 0.003 0.520 0.603 -0.005 0.008
------------------------------------------------------------------------------margins, dydx(age) predict(model(m2) pr)------------------------------------------------------------------------------
| Delta-method
| Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
m2 |
age | 0.001 0.003 0.428 0.669 -0.005 0.008
------------------------------------------------------------------------------The other margins options, e.g. at(), over(), and post, work as usual. For multi-category models add the outcome, e.g. predict(model(o1) pr outcome(3)); for multilevel models, e.g. predict(model(me1) mu fixedonly). The Predictions and graphs page shows marginsplot and coefplot after suest2.
margins, post followed by lincom or metest tests a marginal effect across the two models, which is what mecompare automates:
mecompare agePredicting: Pr(union)
Marginal effects and cross-model differences (N_m1=1878) (N_m2=1878)
| ME # Estimate Robust SE P>|z|
---------------------------------+---------------------------------------
age + 1 (centered) |
m1 | 1 0.002 0.003 0.603
m2 | 2 0.001 0.003 0.669
Difference | 3 0.000 0.001 0.579Options
| Option | Purpose |
|---|---|
cluster(varname) |
Cluster the joint covariance on varname. |
vce(robust), robust |
Request the route’s robust covariance. |
vce(cluster varname) |
Same as cluster(varname). |
level(#) |
Confidence level; the default is 95. |
dir |
Pass through the official suest display option. |
eform(string) |
Exponentiated coefficients where applicable, with string as the column label, e.g. eform("Odds ratio"). The argument is required. |
minus(string) |
The official suest minus convention when supported; passed verbatim. |
regressml |
The official suest ML-regression convention when supported. |
svy |
Pass through the official suest survey option when supported. Does not replace fitting each survey model with the svy: prefix. |
nowarn |
Suppress the note printed when a model was not fit with vce(robust). |
Specify only one of cluster(), vce(), and robust. Panel models cluster at the panel identifier by default and multilevel models at the common highest-level grouping variable; a larger cluster may be requested only when the panels or groups are nested within it.
suest2 m1 m2, eform("Odds ratio") nowarnSimultaneous results for m1, m2 Number of obs = 1,878
--------------------------------------------------------------------------------
| Robust
| Odds ratio std. err. z P>|z| [95% conf. interval]
---------------+----------------------------------------------------------------
m1_union |
age | 0.998 0.021 -0.092 0.927 0.958 1.040
|
collgrad |
College grad | 0.365 0.571 -0.644 0.520 0.017 7.828
|
collgrad#c.age |
College grad | 1.039 0.041 0.973 0.331 0.962 1.123
|
south |
South | 0.475 0.055 -6.408 0.000 0.378 0.596
_cons | 0.406 0.336 -1.088 0.277 0.080 2.059
---------------+----------------------------------------------------------------
m2_union |
age | 0.996 0.021 -0.187 0.852 0.955 1.039
|
collgrad |
College grad | 0.277 0.444 -0.802 0.423 0.012 6.390
|
collgrad#c.age |
College grad | 1.041 0.042 1.004 0.315 0.962 1.127
|
south |
South | 0.509 0.060 -5.690 0.000 0.404 0.643
wage | 1.055 0.015 3.843 0.000 1.027 1.085
ttl_exp | 1.006 0.013 0.484 0.629 0.981 1.033
_cons | 0.272 0.228 -1.554 0.120 0.053 1.406
--------------------------------------------------------------------------------Weights, survey data, and multiple imputation
Survey data. For weighted analyses and complex samples, svyset the data and fit each model with the svy: prefix, then pass the stored names to suest2 without a prefix. The design comes from svyset; do not specify cluster(), vce(), or robust on suest2. All models must use the same subpopulation specification, and replicate-weight VCEs are not supported.
Pweights. Pweighted systems support regress, logit/logistic, probit, poisson, nbreg, ologit, oprobit, and mlogit, which may be combined. The weights must agree on observations shared by the estimation samples. Request system clustering with cluster() on suest2.
Multiple imputation. Fit and store each model with mi estimate, post:, then pass the stored names to suest2. It combines the models within each imputation and pools the results with Rubin’s rules; the MI data must not change in between. Use mimrgns or mecompare for post-estimation; ordinary predict and margins are not available after MI pooling. The supported families are listed in the help file.
Supported estimators
The lists follow the help file; the notes at the end of each group give its rules.
Ordinary single-level models
regress; logit and logistic; probit; ologit; oprobit; mlogit; poisson; nbreg; zip; and zinb with either inflation link.
glm; cloglog; tobit; intreg; maximum-likelihood heckman; and parametric streg. glm predictions support the identity, log, logit, probit, cloglog, and loglog links.
gologit2, including proportional-odds, partial proportional-odds, unrestricted, alternative-link, and autofit() results.
ivregress 2sls. fracreg with estimators logit and probit. Every model in an instrumental-variables system must use the same estimation mode: either unweighted conventional estimates, or linearized svy: ivregress 2sls estimates under one active survey design.
betareg with links logit, probit, cloglog, and loglog. truncreg. hetprobit. biprobit. ivprobit. ivtobit.
Panel models
xtreg with estimators re, fe, be, mle, and pa.
xtlogit with estimators re, fe, and pa. xtprobit with estimators re and pa.
xtologit and xtoprobit.
xtmlogit with estimators re and fe.
xtpoisson with estimators re, fe, and pa. xtnbreg with estimators re and pa.
xtcloglog with estimators re and pa.
Panel models must be unweighted and share a panel identifier. Fit them with conventional standard errors and request robust or clustered ones on suest2. Correlated random effects models are described below.
Multilevel models
mixed, mle; melogit; meprobit; mecloglog; mepoisson; menbreg; meologit; meoprobit; and mestreg.
meglm with the family-link pairs Gaussian-identity, Bernoulli-logit, Bernoulli-probit, Bernoulli-cloglog, Poisson-log, negative-binomial-log, Gamma-log, ordinal-logit, and ordinal-probit.
mestreg with the exponential, Weibull, lognormal, loglogistic, and gamma distributions, including the applicable proportional-hazards and accelerated-failure-time forms.
Multilevel models may be combined, and every model must use the same highest-level grouping variable. Multilevel and panel models may also be combined with ordinary single-level models – regress, logit, probit, cloglog, poisson, nbreg, ologit, and oprobit – e.g., a logit with a melogit or xtlogit. Fit the ordinary model unweighted, without a prefix, and with conventional standard errors; its standard errors are clustered on the highest-level group of the multilevel or panel model.
More than two models
suest2 combines any number of stored models. Here are the four nested logits of Example 6.2 of Mize, Doan, and Long (2019), in one system:
use https://tdmize.github.io/data/data/gss_cme, clear(gss_cme.dta | GSS 1972 - 2016 Weighted | 2018-07-10)drop if year < 2000(38,116 observations deleted)drop if employed != 1(9,556 observations deleted)drop if missing(vhappy, college, wages, occprest, age, married, parent, woman, conserv, reltrad)(5,578 observations deleted)
. quietly logit vhappy i.college, vce(robust)
estimates store m1
quietly logit vhappy i.college i.married i.parent i.woman i.conserv i.reltrad i.year c.age##c.age,
vce(robust)
estimates store m2
quietly logit vhappy i.college c.wages i.married i.parent i.woman i.conserv i.reltrad i.year c.age
##c.age, vce(robust)
estimates store m3
quietly logit vhappy i.college c.wages c.occprest i.married i.parent i.woman i.conserv i.reltrad i
.year c.age##c.age, vce(robust)
estimates store m4. quietly suest2 m1 m2 m3 m4
test [m1_vhappy]1.college = [m2_vhappy]1.college ( 1) [m1_vhappy]1.college - [m2_vhappy]1.college = 0
chi2( 1) = 3.84
Prob > chi2 = 0.0499The marginal-effects version is mecompare college, models(m1 m2 m3 m4) followed by metest.