Longitudinal/panel-data analysis

1. New estimation commands xtmelogit and xtmepoisson fit nested, hierarchical, and mixed models with binary and count responses; i.e., you can fit logistic and Poisson models with complex, nested error components. Syntax is the same as for Stata’s linear mixed-model estimator, xtmixed. To fit a model of graduation with a fixed coefficient on x1 and random coefficient on x2 at the school level, and with random intercepts at both the school and class-within-school level, you type

           . xtmelogit graduate x1 x2 || school: x2 || class:, model(logistic) 

   predict after xtmelogit and xtmepoisson will calculate predicted random effects. See [XT] xtmelogit, [XT] xtmelogit postestimation, [XT] xtmepoisson, and [XT] xtmepoisson postestimation.
2. New estimation commands are available for fitting dynamic panel-data models:
   
   
   1. Existing estimation command xtabond fits dynamic panel-data models by using the Arellano–Bond estimator but now reports results in levels rather than differences. Also xtabond will now compute the Windmeijer bias-corrected two-step robust VCE. See [XT] xtabond.
   2. New estimation command xtdpdsys fits dynamic panel-data models by using the Arellano–Bover/Blundell–Bond system estimator. xtdpdsys is an extension of xtabond and produces estimates with smaller bias when the AR process is too persistent. xtpdsys is also more efficient than xtabond. Whereas xtabond uses moment conditions based on the differenced errors in producing results, xtpdsys uses moment conditions based on differences and levels. See [XT] xtdpdsys.
   3. New estimation command xtdpd fits dynamic panel-data models extending the Arellano–Bond or the Arellano–Bover/Blundell–Bond system estimator, and allows a richer syntax for specifying models and so will fit a broader class of models than either xtabond or xtdpdsys. xtdpd can be used to fit models with serially correlated idiosyncratic errors, whereas xtdpdsys and xtabond assume no serial correlation. xtdpd can be used with models where the structure of the predetermined variables is more complicated than that assumed by xtdpdsys or xtabond. See [XT] xtdpd.
   4. New postestimation command estat abond tests for serial correlation in the first-differenced errors. See [XT] xtabond postestimation, [XT] xtdpdsys postestimation, and [XT] xtdpd postestimation.
   5. New postestimation command estat sargan performs the Sargan test of overidentifying restrictions. See [XT] xtabond postestimation, [XT] xtdpdsys postestimation, and [XT] xtdpd postestimation.
3. Existing estimation command xtreg, fe now accepts aweights, fweights, and pweights. Also new option dfadj specifies that the cluster-robust VCE be adjusted for the within transform. This was previously the default behavior. See [XT] xtreg.
4. New command xtset declares a dataset to be panel data and designates the variable that identifies the panels. In previous versions of Stata, you specified options i(groupvar) and sometimes t(timevar) to identify the panels. You specified the i() and t() options on the xt command you wanted to use. Now you “xtset groupvar” or “xtset groupvar timevar” first. The values you set will be remembered from one session to the next if you save your dataset.
   
   xtset also provides a new feature. xtset allows option delta() to specify the frequency of the time-series data, something you will need to do if you are using Stata’s new date/time variables.
   
   Finally, you can still specify old options i() and t(), but they are no longer documented. Similarly, old commands iis and tis continue to work but are no longer documented. See [XT] xtset.
5. Existing estimation commands xtreg, fe and xtreg, re used to be willing to produce cluster-robust VCEs when the panels were not nested within the clusters. Sometimes this VCE is consistent and other times it is not. You must now specify the new nonest option to obtain a cluster-robust VCE when the panels are not nested within the clusters.
6. The numerical method used to evaluate distributions, known as quadrature, has been improved. This method is used by the xt random-effect estimation commands xtlogit, xtprobit, xtcloglog, xtintreg, and xttobit, and xtpoisson, re normal.
   
   
   1. For the estimation commands, the default method is now intmethod(mvaghermite). The old default was intmethod(aghermite).
   2. Option intpoints(#) for the commands now allows up to 195 quadrature points. The default is 12, and the old upper limit was 30. (Models with large random effects often require more quadrature points.)
   3. The estimation commands may now be used with constraints regardless of the quadrature method chosen.
   4. Command quadchk, for use after estimation to verify that the quadrature approximation was sufficiently accurate, now produces a more informative comparison table. Before, four fewer and four more quadrature points were used, and that was reasonable if the number of quadrature points was, say, n_q = 12. Now you may specify significantly larger n_q and the ±4 is not useful. Now quadchk uses n_q − int(n_q / 3) and n_q + int(n_q / 3).
   5. quadchk has new option nofrom that forces refitted models to start from scratch rather than starting from the previous estimation results. This is important if you use the old intmethod(aghermite), which is sensitive to starting values, but not important if you are using the new default intmethod(mvaghermite).
   See [XT] quadchk.
7. All xt estimation commands now accept option vce(vcetype). As mentioned in the More new statistical features, vce(robust) and vce(cluster varname) are the right ways to specify the old robust and cluster() options, and option vce() allows other VCE calculations as well.
8. Existing estimation command xtcloglog has new option eform that requests exponentiated coefficients be reported; see [XT] xtcloglog.
9. Existing estimation command xthtaylor now allows users to specify only endogenous time-invariant variables, only endogenous time-varying variables, or both. Previously, both were required. See [XT] xthtaylor.
10. Most xt estimation commands have new option collinear, which specifies that collinear variables are not to be removed. Typically, you do not want to specify this option. It is for use when you specify constraints on the coefficients such that, even though the variables are collinear, the model is fully identified. See [XT] estimation options.
11. Existing command xtdes has been renamed to xtdescribe. xtdes continues to work as a synonym for xtdescribe. See [XT] xtdescribe.
12. The [XT] manual has an expanded glossary.

Back to highlights