What’s new in general statistics

• Contrasts, which is to say, tests of linear hypotheses involving factor variables and their interactions from the most recently fit model, and that model can be virtually any model that Stata can fit. Tests include ANOVA-style tests of main effects, simple effects, interactions, and nested effects. Effects can be decomposed into comparisons with reference categories, comparisons of adjacent levels, comparisons with the grand mean, and more.
• Pairwise comparisons of means, estimated cell means, estimated marginal means, predictive margins of linear and nonlinear responses, intercepts, and slopes. In addition to ANOVA-style comparisons, comparisons can be made of population averages.
• Graphs of margins, marginal effects, contrasts, and pairwise comparisons. Margins and effects can be obtained from linear or nonlinear (for example, probability) responses.
• ROC adjusted for covariates, which is to say, you can model the ROC curve and obtain coefficients, standard errors, and graphs. Nonparametric and parametric estimation is supported.
• Estimation output improved:
  
  
  • Baseline odds now shown, which is to say, the exponentiated intercept is displayed by logistic and by logit with option or. In fact, all estimation commands show exponentiated intercepts when option eform() or its equivalent is specified. For example, poisson shows the baseline incidence rate when option irr is specified.
  • Implied zero coefficients now shown. When a coefficient is omitted, it is now shown as being zero, and the reason it was omitted—collinearity, base, empty—is shown in the standard-error column. (The word “omitted” is shown if the coefficient was omitted because of collinearity.)
  • You can set displayed precision for all values in coefficient tables using set cformat, set pformat, and set sformat. Or you may use options cformat(), pformat(), and sformat() now allowed on all estimation commands.
  • Estimation commands now respect the width of the Results window. This feature may be turned off by new display option nolstretch.
  • You can now set whether base levels, empty cells, and omitted are shown using set showbaselevels, set showemptycells, and set showomitted.
• test with coefficient names not using _b[ ] notation is now allowed, even when the specified variables no longer exist in the current dataset.
• areg now faster. areg is orders of magnitude faster when there are hundreds of absorption groups, even if you are not running Stata/MP.
• misstable summarize will now create a summary variable recording the missing-values pattern.
• margins command supports contrasts.
• sfrancia uses a better algorithm. sfrancia now uses an algorithm based on the log transformation for approximating the sampling distribution of the W′ statistic for testing normality. The old algorithm, using the Box–Cox transformation, is available under version control or via the new boxcox option. Based on simulation, the new algorithm is more powerful for sample sizes greater than 1,000 and is comparable to the old algorithm for sample sizes less than 1,000. Also, similarly to swilk, sfrancia now allows you to suppress the treatment of ties when option noties is used.
• logistic now allows option noconstant.
• Probability predictions now available. predict after count-data models, such as poisson and nbreg, can now predict the probability of any count or count range.
• Truncated count-data models now available. New estimation commands tpoisson and tnbreg fit models of count-data outcomes with any form of left truncation, including truncation that varies observation by observation. These new commands supersede ztp and ztnb.
• cnsreg checks for collinear variables prior to estimation and has new option collinear, which keeps the collinear variables instead of omitting them. The old behavior of always keeping collinear variables is preserved under version control.
• ml improved,
  
  
  • ml now distinguishes the Hessian matrix produced by technique(nr) from the other techniques that compute a substitute for the Hessian matrix. This means that ml will compute the real Hessian matrix of second derivatives to determine convergence when all other convergence tolerances are satisfied and technique(bfgs), technique(bhhh), or technique(dfp) is in effect. The old behavior was to use the nrtolerance() value with the H matrix associated with the technique() currently in effect to determine convergence; this behavior is preserved under version control.
  • ml has new option qtolerance() that distinguishes itself from ntrol() when technique(bfgs), technique(bhhh), or technique(dfp) is specified. Option qtolerance() replaces nrtolerance() when technique(bfgs), technique(bhhh), or technique(dfp) is in effect.
• margins has new option estimtolerance() for setting tolerance used to determine estimable functions.

Back to highlights

See New in Stata 17 to learn about what was added in Stata 17.