1. (StataNow) Machine learning via H2O: Ensemble decision trees. Machine learning methods are often used to solve research and business problems focused on prediction when the problems require more advanced modeling than linear or generalized linear models. Ensemble decision tree methods, which combine multiple trees for better predictions, are popular for such tasks. H2O is a scalable machine learning platform that supports data analysis and machine learning, including ensemble decision tree methods such as random forest and gradient boosting.
2. (StataNow) stmgintcox now supports computation of the clustered sandwich estimator of variance using option vce(cluster clustvar). It can also be specified on replay.
3. (StataNow) list has new option relative. By default, list with qualifiers if and in displays the absolute observation numbers for the subset of observations. By adding new option relative, the relative observation numbers will be displayed instead. That is, observations will be numbered 1, 2, 3, etc.
4. Stata and H2O Integration has the following improvements:
   1. The h2o command has new subcommand h2o clear, which discards all resources within the H2O cluster.
   2. h2o init has new option maxmem() for specifying the maximum memory to allocate when starting a local H2O cluster.
   3. The _h2oframe command has new subcommand _h2oframe distinct, which lists the distinct values of a column in the current H2O frame.
   4. _h2oframe put, _h2oframe create, _h2oframe import, and _h2oframe upload have new option current, which sets the H2O frame as the current (working) H2O frame.
   5. _h2oframe put, _h2oframe import, _h2oframe upload, and _h2oframe split have new option replace, which replaces the H2O frame if it already exists.
   6. _h2oframe put has new option coltype(varlist, type), which sets the column type for the specified variable(s) in the destination H2O frame.
   7. _h2oframe split has new option split(numlist), which splits an existing H2O frame into a list of H2O frames based on the proportions or ratio specified in numlist.
5. hdidregress twfe and xthdidregress twfe are now significantly faster. Treatment effects are computed analytically rather than using margins.
6. regress with many covariates and option vce(hc2 [clustvar], dfadjust) is now faster.
7. Online help and the search index have been brought up to date for Stata Journal 24(4).
8. The Java Development Kit (OpenJDK) that is redistributed with Stata is now updated to version 17.0.14+7-LTS acquired from Azul Systems.
9. areg has the following fixes:
   1. areg with option vce(hc2) or vce(hc2 clustvar), where clustvar was not nested with the absorb variable, produced standard errors and confidence intervals that were too small. This has been fixed.
   2. areg with option vce(hc2) or vce(hc2 clustvar), where clustvar is not nested with the absorb variable, now disallows weights.
10. areg and xtreg, fe with no independent variables and with option vce(hc2) or vce(hc2 clustvar) and regress with no independent variables and with option vce(hc2 clustvar) exited with an uninformative error message. This has been fixed.
11. churdle with option vce(cluster clustvar) reported the wrong number of clusters in the coefficient table. This did not affect standard error computations. This has been fixed.

© Copyright 1996–2026 StataCorp LLC. All rights reserved.

12. collect export, when exporting to LaTeX, did not apply background color changes to table titles or notes. This has been fixed.
13. collect recode, collect remap, collect addtags, and collect stars with option fortags(taglist), where taglist was not empty but also did not match any items in the collection, incorrectly behaved as if fortags() was not specified. For example,

collect recode coleq xb = mpg, fortags(cmdset[2]#coleq[mpg])

(assuming that cmdset[2]#coleq[mpg] did not match any items in the collection) behaved like

collect recode coleq xb = mpg

meaning that all items with coleq[xb] were recoded to coleq[mpg]. The correct behavior is for no items to change, because no items are tagged with cmdset[2]#coleq[mpg]. This has been fixed.

14. estat aggregation has the following fixes:
    1. estat aggregation after hdidregress twfe and xthdidregress twfe in models with covariates could return aggregated treatment-effect estimates even when these effects were not estimable. This has been fixed.
    2. estat aggregation after hdidregress twfe and xthdidregress twfe, when repeated values were specified in option time() or option cohort(), returned an uninformative error message. This has been fixed.
15. gsem, when a latent endogenous variable (L) was specified before any of the observed endogenous variables, resulted in the last specified observed endogenous variable’s equation being split by the equation for L. This has been fixed. The estimates were correct; just the order they were displayed in the estimation table and stored in e(b) was not standard.
16. ivregress 2sls with options first and vce(unadjusted) produced error message “varname not found” if varname was dropped in the first-stage regression. This has been fixed.
17. hdidregress twfe and xthdidregress twfe in models with covariates could return treatment-effect estimates even when these effects were not estimable. Typically, this nonestimability was the result of empty cells formed by interactions with levels of factor covariates. This has been fixed.
18. regress with options noconstant and vce(hc2 clustvar) or vce(hc2 [clustvar], dfadjust) exited with an uninformative error message. This has been fixed.
19. reshape wide with option favor(speed) and option j() that identified a subset of values incorrectly exited with an uninformative error message. This has been fixed.
20. xtgls with option corr(ar1) could mark e(sample) incorrectly when some groups had only one observation. This has been fixed.
21. Pasting new data into the Data Editor, when observations were filtered and where new observations extended beyond the original filtered view, did not extend the dataset to the correct size, causing some new observations not to be included. This has been fixed.
22. (Windows) Initializing a virtual Python environment created using venv from within Stata caused Stata to crash. This has been fixed.
23. (Unix) The Python environment within Stata failed to find Python modules and packages within directories PLUS and PERSONAL. This has been fixed.