Stata release 16 - New Features



You can summarize results from multiple studies with Stata’s new meta-analysis suite. Use random-effects, fixed-effects, or common-effect meta-analysis to combine individual results and compute overall effect size. Forest plots allow you to visualize the results. With subgroup analysis or meta-regression, you can explore heterogeneity of studies. And you can evaluate publication bias using funnel plots and the trim-and-fill method.


Lasso-based machine learning

You can use lasso and elastic net for model selection and prediction. Want to estimate effects and test coefficients? With cutting-edge inferential methods, you can make inferences for variables of interest. Analyze continuous, binary, and count outcomes. You can even account for unobserved confounding.


Truly reproducible reporting

With Stata’s reporting features, you can easily incorporate Stata results and graphs with formatted text and tables in Word, PDF, HTML, and Excel formats. Take advantage of Stata’s integrated versioning to create reproducible reports. Dynamic documents can be updated as your data change. With Stata 16, you can create Word documents from Markdown; easily include headers, footers, page numbers, and large blocks of text in Word; and convert HTML to Word or Word to PDF. The new Reporting Reference Manual guides you with many examples and workflows..


New Bayesian analysis features

The most requested additions for Bayesian analysis— multiple chains and Bayesian predictions—are now available. You can use multiple chains with Bayesian estimation to evaluate MCMC convergence. And you can now evaluate convergence using the Gelman–Rubin convergence diagnostic. With Bayesian predictions, you can check model fit and predict out-of-sample observations.


Python integration

You can now embed and execute Python code within Stata. Invoke Python interactively or within do-files or ado-files. With the new Stata Function Interface (sfi) Python module, you can pass data back and forth seamlessly. This means that you can now use any Python package directly within Stata. For instance, you might use Matplotlib to draw 3-dimensional graphs, Scrapy to scrape data from the web, or TensorFlow and scikit-learn to access additional machine-learning techniques.


Multiple-dose pharmacokinetic modeling

Stata’s menl command for fitting nonlinear multilevel models now allows you to include lags, leads, and differences in your model specification. This means you can now fit more pharmacokinetic models, including multiple-dose models.

DSGE Models

Nonlinear DSGE models

Dynamic stochastic general equilibrium (DSGE) models consist of systems of equations that describe the structure of the economy. Equations in these models are almost always nonlinear. With Stata 16’s new dsgenl command, you no longer need to linearize the equations before fitting your DSGE models. And after fitting your model, you can obtain policy and transition matrices, identify the model’s steady state, estimate covariances and autocovariances, and create and graph impulse–response functions.


Choice models

Stata 16 introduces a new, unified suite of features for summarizing and modeling choice data. In addition, you can now fit mixed logit models for panel data. And here’s the best part: margins now works after fitting choice models. This means that you can now easily interpret your results. For example, estimate how much wait times at the airport affect the probability of traveling by air or even by train. And you can answer these types of questions whether you just fit a conditional logit, multinomial probit, mixed logit, rank-ordered probit, or another choice model.

Panel data

Panel-data extended regression models

Extended regression models account for common problems—endogenous covariates, sample selection, and treatment—either alone or in combination. In Stata 16, add panel data to that list. Fit random-effects extended regression models for linear, interval-censored (including tobit), binary, and ordinal outcomes.


Nonparametric series regression

Don’t know the functional form of the relationship between your outcome and covariates? Don’t worry. Nonparametric series regression can select a polynomial, B-spline, or spline function that closely approximates the mean of your outcome. And you can still make inferences. Explore the response surface, estimate population-averaged effects, and obtain tests and confidence intervals.


Multiple-group IRT

With IRT, you can explore the relationship between an unobserved latent trait, such as mathematical ability, and an instrument designed to measure that trait, such as a test. In Stata 16, you can now fit multiple-group IRT models and evaluate whether tests perform equally across different subpopulations. Do students in urban and rural schools respond in the same way to test questions, or are some questions worded unfairly for one group? With multiple-group IRT, you can perform an IRT model-based test of this hypothesis and of similar hypotheses related to differential item functioning.