Stata 17 released
We just announced Stata 17. Visit stata.com/new-in-stata to read all about its 29 major new features.
They are
- Tables
- Bayesian econometrics
- Faster Stata
- Difference-in-differences (DID) and DDD models
- Interval-censored Cox model
- Multivariate meta-analysis
- Bayesian VAR models
- Treatment-effects lasso estimation
- Bayesian multilevel models: Nonlinear, joint, SEM-like, and more
- Galbraith plots
- Leave-one-out meta-analysis
- Bayesian longitudinal/panel-data models
- Panel-data multinomial logit model
- Zero-inflated ordered logit model
- Nonparametric tests for trend
- Bayesian dynamic forecasting
- Bayesian IRF and FEVD analysis
- BIC for lasso penalty selection
- Lasso with clustered data
- Bayesian linear and nonlinear DSGE models
- Do-file Editor enhancements
- New functions for dates and times
- Intel Math Kernel Library (MKL)
- Stata on Apple Silicon
- JDBC
- Java integration
- H2O integration
- PyStata
- Jupyter Notebook with Stata
Looking over this list of features, someone suggested that a potential marketing tagline for Stata 17 could be “Better. Faster. Stronger.” I thought Daft Punk might not like it if we used that, so we aren’t, but really, it is a great overall description of the new version.
I’ll share my thoughts on some of the new features below.
Customizable tables
There has been a long tradition in the Stata user community of commands that build various tables. These are among some of the most-used community-contributed commands! There has been an equally long tradition of the Stata user community asking us to provide more official features to assist with flexible table creation and export. Stata’s table command has been completely revamped, and a new collect command allows you to gather and manage results from multiple commands, which can then be shown in tabular form. Excel, HTML, LaTeX, Markdown, PDF, Stata SMCL, Word, and plain text are supported as export formats. I suspect almost all users will be adding this to their Stata repertoire.
Bayesian econometrics
In Stata 17, we have added many features for Bayesian econometrics, including
- Bayesian VAR models
- Bayesian IRF and FEVD analysis
- Bayesian dynamic forecasting
- Bayesian longitudinal/panel-data models
- Bayesian linear and nonlinear DSGE models
Many of you have been asking us for Bayesian VAR models. That’s not surprising. VAR models have many parameters but often not enough data to estimate them reliably. The Bayesian approach provides a solution by incorporating specialized priors to allow you to obtain more stable parameter estimates.
As with classical VAR models, you can perform IRF analysis and obtain dynamic forecasts but now within the Bayesian paradigm.
Bayesian panel-data models are appealing when you have few panels or when you would like to study and compare panel-specific effects.
With Bayesian DSGE models, prior distributions give you a natural way to incorporate knowledge about model parameters that is motivated by the economic theory.
As with Stata’s other Bayesian features, our aim is to make specification of these models as intuitive as possible and as similar to the specifications of the frequentist counterparts as possible.
All the above is in addition to other existing Bayesian features of interest to econometricians such as Bayesian generalized linear models and Bayesian sample-selection models.
Bayesian multilevel models
Stata users span many disciplines. In addition to the new Bayesian features above that will be of most interest to econometricians, Stata 17 also adds Bayesian multilevel modeling with support for nonlinear, joint, SEM-like, and even more models. One notable feature is the ability to fit multivariate nonlinear models containing random effects such as multivariate nonlinear growth models.
Give me more speed
As computer capabilities have grown, so have dataset sizes. With larger datasets and more computationally intensive methods comes the above request, “Give me more speed.” Sometimes, our developers say, “I’m giving her all she’s got, Captain!”, but for Stata 17, they gave us all more.
We achieved speed gains in part through careful algorithm selection and implementation and in part through integration of the Intel Math Kernel Library (MKL) to underpin many of Mata’s linear algebra functions and operators.
Read the details, including tables of specific speed gains.
Difference-in-differences (DID) models
DID and difference-in-difference-in-differences (DDD) models are appealing to many disciplines, including econometrics, epidemiology, political science, public policy, and many more. If you are studying the effect of a treatment (such as a drug regimen or policy) in observational data and are concerned that the effect may be influenced by time or by some other group effects, DID and DDD models provide intuitive methods to control for such unobserved effects.
New meta-analysis features
Stata 17 adds the following to the excellent (IMHO) meta-analysis suite introduced in Stata 16:
If you recognize the above, you know you want them!
And we have made these new features just as easy to use as the rest of the meta suite.
Interval-censored Cox model
The Cox proportional hazards model is used routinely by researchers from many disciplines to analyze right-censored event-time data, where time to an event of interest is observed exactly. In Stata 17, you can use it with interval-censored event-time data too!
With interval-censored event-time data, we only know that the time to an event of interest lies in an interval. For instance, think of time to cancer recurrence or to an infection or, for that matter, to any asymptomatic disease that can be detected only through periodic examinations. These are all interval-censored data, and you now have a new powerful tool to analyze them.
New lasso features
Stata 17 adds the following to the popular lasso features first introduced in Stata 16:
In fact, treatment-effects lasso combines two popular features: treatment effects and lasso. You can now incorporate many (hundreds, thousands, and more) covariates in your treatment-effects models.
You can account for clustered observations in your lasso analysis. And you can use the BIC criterion to select lasso penalty parameters.
Integration with other software and languages
Two of Stata’s great features are its extensibility and reproducibility. Stata 17 builds on that tradition by greatly enhancing its interoperability with Python and Java, adding support for Jupyter Notebook, adding JDBC support, and giving you experimental access to the H2O platform.
You could already call Python code from Stata code. Now you can call Stata from any stand-alone Python environment. Do you write Python code in Jupyter Notebook, Spyder IDE, or PyCharm IDE? Now you can call Stata directly from those environments, passing data, metadata, and results between Stata and Python seamlessly. Even your Stata graphs will show up directly in Jupyter Notebook. Our nickname for all the ways you can connect Python and Stata is ‘PyStata‘.
For Java, you could already compile Java code into .jar files and call those as plugins from Stata. Now you can embed Java code directly in your Stata do-files and ado-files, just like Mata code and Python code. Stata will compile and execute your Java code on the fly. You can interchange data, metadata, and results at will.
One of the first steps of any analysis is importing your data. Stata 17 supports JDBC for importing data from and writing data to databases that provide JDBC drivers. An important advantage JDBC has over ODBC is that JDBC drivers are platform independent, so if a database vendor provides a JDBC driver, it will work seamlessly on Windows, Mac, and Linux.
Finally, some of our developers have been experimenting with connecting Stata to H2O, a scalable and distributed open-source machine-learning and predictive analytics platform. We decided to release our experiment to you. Is this something you’d like us to do more with? We look forward to your feedback.
New date and time functions
Stata 17 adds a plethora of date and time convenience functions in three main areas:
- Datetime durations, such as ages
- Relative dates, such as the next birthday relative to a reference date
- Datetime components, functions that extract various components from datetime values
You’ll undoubtedly find these make your life easier when working with date and time values.
New Do-file Editor features
Don’t miss the enhancements in the Do-file Editor, including persistent bookmarks that are saved with your code, a Navigation Control providing quick access to those bookmarks and defined programs, syntax highlighting support for Java and XML (in addition to the existing support for Stata ado, Python, and Markdown), and autocompletion of quotes, parentheses, and brackets around a selection.
And there’s more
Health scientists who deal with ordinal outcomes with an overabundance of values in the lowest category will want to try the new ziologit command.
Those of you interested in panel data and categorical outcomes will be pleased to know that you can now analyze both together easily with the new xtmlogit command.
And for those of you interested in nonparametric tests of trend, we have added three new tests to the existing nptrend command.
Finally, Stata 17 runs fully natively on Apple’s new M1 Macs, known as Apple Silicon. Stata ships as a universal application that has everything necessary to run natively on both M1 Macs and Intel-based Macs for the best performance no matter your choice of hardware platform.
It has been a lot of fun to see this release come together at StataCorp, and it is a tremendous pleasure to be able to release it to you.