ddml uses stacking regression as the default machine learner but may be used in combination with other methods implemented in Stata.

While formal testing of overidentifying restrictions is possible, its interpretation still hinges on the validity of an initial set of untestable just-identifying orthogonality conditions. We present the kinkyreg Stata program for kinky least-squares (KLS) inference, which adopts an alternative approach to identification. By exploiting nonorthogonality conditions in the form of bounds on the admissible degree of endogeneity, feasible test procedures can be constructed that do not require instrumental variables. The KLS confidence bands can be more informative than confidence intervals obtained from instrumental-variable estimation, in particular when the instruments are weak. Moreover, the approach facilitates a sensitivity analysis for the standard instrumental-variable inference. In particular, it allows one to assess the validity of previously untestable just-identification exclusion restrictions. Further KLS-based tests include heteroskedasticity, function form, and serial correlation tests.

The two-step approach is sometimes seen as superior to the more standard one-step approach if the numbers of observations on the cluster level become small. Additionally, two-step mulitlevel analysis may be used as a companion to the one-step approach, for instance, to check model or linearity assumptions. twostep is created specifically with this second use in mind.

The estimation and the tests follow the methodologies developed in Andrews (1993), Bai and Perron (1998), and Ditzen, Karavias, and Vesterlund (2021). For both time-series and panel-data regressions, five tools are provided: (i) a test of no structural change against the alternative of a specific number of changes; (ii) a test of the null hypothesis of no structural change against the alternative of an unknown number of structural changes; (iii) a test of the null of s changes against the alternative of s + 1 changes; (iv) consistent break date estimators; and (v) asymptotically valid confidence intervals for the break dates.

References:

Andrews, D. V. K. 1993. Tests for parameter instability and structural change with unknown change point. Econometrica 61: 821–856.

Bai, B. Y. J., and P. Perron. 1998. Estimating and testing linear models with multiple structural changes. Econometrica 66: 47–78.

Ditzen, J., Y. Karavias, and J. Vesterlund. 2021. Testing for Multiple Structural Breaks in Panel Data.

This presentation will consider three specific variants—namely, external Stata macros, INI, and MS Excel—and outline some general principles to facilitate discussion on good practices within the Stata community.

In Stata 17, we introduced a new command, bayes:var, for fitting Bayesian VAR models. Bayesian VAR models apply priors on the regression parameters and variance-covariance of the errors for a fine control over the posterior time-series process. By default, the prior on regression coefficients shrinks them toward a random-walk process that assumes no relationship between time-series variables. This assumption helps avoid overfitting the data. The Bayesian approach also provides a systematic and unambiguous way of determining the number of lags.

In this presentation, I illustrate Bayesian VAR models on some real data and show model interpretations based on their impulse–response functions. I also compute Bayesian forecasts and compare them with classical forecasts.

Particular features of the command are that it provides consistent standard errors supporting complex sample designs for all covered statistics and that the simultaneous estimation of multiple statistics across multiple variables and multiple subpopulations is possible. Furthermore, the command supports covariate balancing based on reweighting techniques (inverse probability weighting and entropy balancing), including appropriate correction of standard errors. Standard-error estimation is implemented in terms of influence functions, which can be stored for further analysis, for example, in RIF regressions.

telasso estimates the average treatment effects with high-dimensional controls while using lasso for model selection. This estimator is Neyman orthogonal because it is robust to the model-selection mistakes. It is also doubly robust, so only one of the models needs to be correctly specified.