New in Stata 18
Stata 18 has something for everyone. Below we list the highlights of the release.
Stata 18 has something for everyone. Below we list the highlights of the release.
Bayesian model averaging (BMA) — Traditionally, you choose a model and perform analysis based on this model. The results are conditional on the chosen model. In the presence of multiple plausible models, this approach may not be reliable. Model averaging allows you to perform analysis based on multiple models and thus account for model uncertainty in the results. BMA accounts for model uncertainty according to Bayesian principles, which can be applied universally to any data analysis. In the regression setting, model uncertainty describes the uncertainty about which predictors should be included in a regression model.
The new command bmaregress performs BMA for linear regression and can be used for inference, prediction, and, if desired, even model selection. For instance,
. bmaregress y x1 x2
considers all four possible models for outcome y that include or exclude predictors x1 and x2 and combines these models according to how likely each model is based on the observed data. You can choose from a variety of prior distributions to explore the effect of assumptions about a model's and predictors' importance on the results.
Postestimation commands allow you to estimate the probability of a model, identify important predictors, explore model complexity, obtain predictive means, evaluate predictive performance, and perform inference on regression coefficients.
Causal mediation analysis — Causal inference aims to identify and quantify the causal effect of a treatment on an outcome. In causal mediation analysis, we aim to further explore how this effect arises. Maybe exercise increases the level of a hormone that, in turn, increases well-being. Maybe an import quota increases the market power of local companies which, in turn, increases the prices of goods.
With the new mediate command, we can estimate the total effect of a treatment on an outcome and decompose it into direct effects and indirect effects (via a mediator such as hormone level). In fact, multiple types of decompositions can be computed, depending on the hypothesis of interest. In addition, estat proportion reports the proportion of the total effect that occurs through the mediator.
mediate is very flexible—the outcome can be continuous, binary, or count; the mediator can be continuous, binary, or count; and the treatment can be binary, multivalued, or continuous.
Tables of descriptive statistics — The new dtable command creates a table of descriptive statistics.
dtable reports summary statistics for continuous and categorical factor variables. You can select which statistics you want to report for each variable; select from the mean, standard deviation, median, interquartile range, percentage, proportion, and many others. You can also easily compare statistics across categories of another variable.
Tables created by dtable can be customized in many ways—statistics to be reported, numeric and string formats, notes, titles, labels, and more. The table can be exported directly to Microsoft Word, Microsoft Excel, HTML, Markdown, PDF, LaTeX, SMCL, or plain text.
Heterogeneous difference in differences (DID) — DID models are used to estimate the average treatment effect on the treated (ATET) with repeated-measures data. A treatment effect can be an effect of a drug regimen on blood pressure or an effect of a training program on employment. Unlike with the standard cross-sectional analysis, provided by the existing teffects command, DID analysis controls for group and time effects when estimating the ATET, where groups identify repeated measures.
Heterogeneous DID models additionally account for variation in treatment effects arising from groups being treated at different points in time and effects varying over time within groups.
Suppose that several school districts introduce an exercise and a nutrition program to improve students' health outcomes. Different school districts introduce the program at different points in time. Is it sensible to assume the effect of the program on students’ health outcomes does not change over time and is the same regardless of when the program was adopted? Maybe not. We can use heterogeneous DID models to account for the potential differences in effects.
The new commands hdidregress and xthdidregress fit heterogeneous DID models. hdidregress works with repeated-cross-sectional data, and xthdidregress works with longitudinal/panel data.
Group sequential designs (GSDs) — GSDs are types of adaptive design that allow researchers to stop a trial early if they find compelling evidence that a treatment is effective or ineffective.
Suppose that we want to design a study to test whether a type of chemotherapy is effective for treating tumors and that we expect data to be collected over a few years’ time. Rather than performing one analysis once all the data have been collected, GSDs allow us to perform interim analyses as the data are collected. Each interim analysis provides the opportunity to stop the trial or continue collecting data. The trial can be stopped early if there is strong evidence of efficacy. The trial can also be stopped early if there is strong evidence of futility; this avoids exposing additional participants to an inadequate treatment.
Stata 18 offers a suite of commands for GSDs. The new gsbounds command calculates efficacy and futility bounds based on the number of analyses (also called looks), the desired overall Type I error, and the desired power. You can select from seven boundary-calculation methods—choose whether you want classical or error-spending methods and whether you want more conservative or less conservative bounds for early analyses. The new gsdesign command calculates efficacy and futility boundaries and provides sample sizes for the interim and final analyses for tests of means, proportions, and survivor functions.
Graphs make it easy to visualize the boundaries across all interim and final analyses.
Multilevel meta-analysis — When researchers want to analyze results from multiple studies, they use meta-analysis to combine results and estimate an overall effect size. The existing meta suite is used to perform standard and multivariate meta-analysis.
Sometimes the reported effect sizes are nested within higher-level groupings, such as geographical locations (states or countries) or administrative units (school districts). Effect sizes within the same groups (for example, districts) are likely to be similar and thus dependent. In this case, you can use multilevel meta-analysis. The goal of multilevel meta-analysis is to not only synthesize an overall effect size but also account for this dependence and assess the variability among the effect sizes at different hierarchical levels. The new estimation commands meta meregress and meta multilevel are used to perform multilevel meta-analysis.
Say that we have studies reporting effects (mean differences) of two teaching methods on math test scores, y, and sampling standard errors of y in se. The effect sizes are nested within schools, and schools are nested in districts. We can fit a three-level random intercepts model with
. meta meregress y || district: || school:, essevariable(se)
or
. meta multilevel y, relevels(district school) essevariable(se)
If we have covariates and want to include random slopes, we can use meta meregress:
. meta meregress y x1 x2 || district: x1 x2 || school:, essevariable(se)
After fitting the model, postestimation commands are available for computing multilevel heterogeneity statistics, displaying estimated random-effects covariance matrices, and more.
Meta-analysis for prevalence — The meta esize command performs meta-analysis of two-sample binary or continuous data. Now, it also performs meta-analysis of one-sample binary data, also known as meta-analysis of proportions or meta-analysis of prevalence.
These types of data commonly appear in meta-analysis studies when pooling results from studies that each estimate one proportion. For instance, you may have studies reporting the prevalence of a particular disease or the proportion of students who drop out of high school. In this setting, effect sizes such as Freeman–Tukey-transformed proportions or logit-transformed proportions are typically used in the meta-analysis.
After meta esize, use other commands in the meta suite for further analysis. For instance, create a forest plot with meta forestplot, perform subgroup analysis by adding the subgroup() option to meta forestplot, summarize meta-analysis data with meta summarize, or construct a funnel plot with meta funnelplot.
Robust inference for linear models — Reliable standard errors are crucial to drawing appropriate inferences in research. Stata 18 offers new ways to obtain standard errors and confidence intervals for linear models fit with regress, areg, and xtreg, fe. The new methods aim to provide better inference when large-sample approximations do not work well. Perhaps you have clustered data with only a few clusters or with an uneven number of observations per cluster. You can now add the vce(hc2 clustvar) option to obtain HC2 cluster–robust standard errors. Perhaps you have more than one variable that identifies clusters in your data. You can now add the vce(cluster clustvar1 clustvar2 …) option to obtain multiway cluster standard errors.
Wild cluster bootstrap — The wild cluster bootstrap provides another new option for robust inference when researchers have data with a few clusters, an uneven number of observations across clusters, or both.
The new wildbootstrap command computes wild cluster bootstrap p-values and confidence intervals for tests of simple and composite linear hypotheses about parameters from linear regression models. You can type
. wildbootstrap regress y x1 x2 …
or
. wildbootstrap areg y x1 x2 …, absorb(x3)
or
. xtset id
. wildbootstrap xtreg y x1 x2 …
to fit a linear regression model, a linear regression model with a large dummy-variable set, or a fixed-effects linear regression model for panel data, respectively, and to obtain the wild cluster bootstrap statistics.
Local projections for impulse–response functions (IRFs) — The new lpirf command provides local projections of IRFs. Local projections are used in time-series analysis to estimate the effect of shocks on outcome variables. For instance, we might evaluate the effect of an unexpected change in interest rates on a country’s output and inflation rate.
You can type
. lpirf y1 y2
to obtain local-projection estimates of IRFs for y1 and y2. You can add the exog() option to estimate dynamic multipliers, which are responses of endogenous variables to a shock to an exogenous variable.
The new lpirf command works seamlessly with the existing irf commands, allowing you to create graphs and tables of IRFs, orthogonalized IRFs, and dynamic multipliers.
As with linear models mentioned above, robust standard errors are often important in IRF estimation. Robust and Newey–West standard errors are available.
Flexible demand system models — Often, researchers are interested in estimating demand for a basket of goods. The new demandsys command provides extensive tools for computing demand and measuring how sensitive demand for goods is to price and expenditure changes by calculating their corresponding elasticities.
We can use demandsys to fit eight different demand system models:
With the estat elasticities command, we can estimate various elasticities—expenditure elasticities, uncompensated own-price and cross-price elasticities, and compensated own-price and cross-price elasticities—to explore how sensitive demand is to changes in prices and expenditures.
TVCs with interval-censored Cox model — In event-time data, interval-censoring occurs when the time to an event of interest, such as recurrence of cancer, is not directly observed but is known to lie within an interval. The existing stintcox command fits semiparametric interval-censored Cox proportional hazards models. In Stata 18, stintcox allows time-varying covariates.
stintcox now supports multiple-record-per-subject interval-censored data, which include a record for each examination time for each subject. This format makes it easy to accommodate time-varying covariates; the data record the values of the covariates at each examination time. Multiple-record-per-subject data also provide a convenient way to specify current status data.
stintcox also has new options tvc(varlist_t) and texp(exp) that provide a convenient way to include time-interacted covariates, which are formed by the interaction between covariates specified in tvc() and a deterministic function of time specified in texp().
After fitting a model, the standard and special-interest postestimation features are available and appropriately account for the time-varying covariates. You can use the new estat gofplot command to produce a goodness-of-fit plot. You can predict the relative hazard. And you can use stcurve to plot survivor and related functions. When you have multiple record data, you can use the new stcurve option attmeans to evaluate the function at time-specific means of the covariates or the new option atframe(framename) to evaluate the function at values of variables specified in the framename frame.
Lasso for Cox model — We use lasso for prediction and model selection when we have many potential covariates. (And when we say many, we mean hundreds, thousands, or more!) We previously introduced the lasso command to perform lasso for linear, logit, probit, and Poisson models. New in Stata 18 is lasso for Cox proportional hazards models. lasso cox can be used to select covariates using lasso and fit a Cox model to survival-time data. elasticnet cox can similarly be used to select covariates using elastic net and fit a Cox model.
After lasso cox and elasticnet cox, you can use predict to predict the hazard ratio; use stcurve to plot the survivor, hazard, or cumulative hazard functions; or use any of the other postestimation tools available after lasso and elasticnet to examine the lasso results.
Relative excess risk due to interaction (RERI) — Epidemiologists often need to determine how two exposures interact to put subjects at a higher risk of experiencing an outcome of interest. For example, you might want to investigate how exposures to cigarette smoke and asbestos interact to increase the risk of lung cancer. With the new reri command, you can measure two-way interactions in an additive model of relative risk, while accounting for other risk factors.
Researchers can choose from various supported models, such as logistic, binomial generalized linear, Poisson, negative binomial, Cox, parametric survival, interval-censored parametric survival, and interval-censored Cox models. They can evaluate an additive model for the interaction of smoke and asbestos by using three related statistics: RERI, attributable proportion, and synergy index.
Instrumental-variables (IV) quantile regression — When we want to study the effects of covariates on different quantiles of the outcome, not just on the mean, we use quantile regression. For instance, we might be interested in modeling the grade distribution of students and how it is affected by changes in covariates. The existing qreg command fits quantile regression models, but what if we suspect that one of our covariates is endogenous? This endogeneity might arise for reasons such as self-selection of study participants, omission of a relevant variable from the model, or measurement error. The new ivqregress command allows us to model quantiles of the outcome and, at the same time, control for problems that arise from endogeneity using IV.
After fitting an IV quantile regression model, you can plot the coefficients across quantiles with the estat coefplot command. You can test for endogeneity using the estat endogeffects command. And you can estimate dual confidence intervals that are robust to weak instruments using estat dualci.
IV fractional probit model — Fractional outcomes are common. You might be modeling participation rates in a 401(k) pension plan, the pass rate on standardized tests, expenditure shares, or the like.
Fractional response models are a flexible and intuitive way to model outcomes that lie between 0 and 1. They do not have the problem of linear models that will yield predictions outside 0 and 1 or the problem of log-odds models that are undefined at 0 and 1. Fractional response models can be fit using the existing fracreg command.
What if you are concerned that one or more of your model covariates are endogenous? With the new ivfprobit command, you can fit a model for a fractional dependent variable and account for endogeneity in one or more of the covariates.
Alias variables across frames — Since Stata 16, Stata has supported multiple datasets in memory. Each dataset resides in a frame. When datasets are related, you can link their frames by using the frlink command and identifying the variables that match the observations in the current frame with observations in the related frame.
In Stata 18, you can use the new fralias add command to create alias variables across linked frames and easily perform analyses using variables stored in separate frames.
Alias variables behave as if you copied them from one frame into another, but because they are stored in the original frame, they take up very little memory.
To see how easy alias variables are to use, say that y is a variable in the current frame and that x is a variable from a linked frame named frame2. To create an alias to x in the current frame, you type
. fralias add x, from(frame2)
Then you can fit a regression by typing
. regress y x
just as if x were stored in the current frame.
Data Editor enhancements — The Data Editor has many enhancements in Stata 18:
Do-file Editor enhancements — The Do-file Editor also has enhancements in Stata 18:
All-new graph style — Graphs in Stata 18 have a new look.
The new default graph scheme (or, the new appearance for Stata graphs) includes the following much-requested features:
In fact, we introduced four new graph schemes: stcolor, stcolor_alt, stgcolor, and stgcolor_alt. The new default is stcolor and the other schemes are variations on stcolor that provide different widths and legend placements.
For a complete list of available features, visit the Stata website.
Buy online or contact our sales team for a customised quote based on:
Save yourself valuable time. Find out about available training courses and resources to become proficient in Stata.
Join our Stata mailing list to get notified of new releases and training.
Stata 18 is available - 25 Apr 2023
Upcoming free Stata webinars - 18 Jan 2022
Webinar: Stata 17—See the new features in action - 17 May 2021