Stata 15 help for whatsnew14

Additions made to Stata during version 14

This file records the additions and fixes made to Stata during the 14.0, 14.1, and 14.2 releases:

+---------------------------------------------------------------+ | help file contents years | |---------------------------------------------------------------| | whatsnew Stata 15.0 and 15.1 2017 to present | | whatsnew14to15 Stata 15.0 new release 2017 | | this file Stata 14.0, 14.1, and 14.2 2015 to 2017 | | whatsnew13to14 Stata 14.0 new release 2015 | | whatsnew13 Stata 13.0 and 13.1 2013 to 2015 | | whatsnew12to13 Stata 13.0 new release 2013 | | whatsnew12 Stata 12.o and 12.1 2011 to 2013 | | whatsnew11to12 Stata 12.0 new release 2011 | | whatsnew11 Stata 11.0, 11.1, and 11.2 2009 to 2011 | | whatsnew10to11 Stata 11.0 new release 2009 | | whatsnew10 Stata 10.0 and 10.1 2007 to 2009 | | whatsnew9to10 Stata 10.0 new release 2007 | | whatsnew9 Stata 9.0, 9.1, and 9.2 2005 to 2007 | | whatsnew8to9 Stata 9.0 new release 2005 | | whatsnew8 Stata 8.0, 8.1, and 8.2 2003 to 2005 | | whatsnew7to8 Stata 8.0 new release 2003 | | whatsnew7 Stata 7.0 2001 to 2002 | | whatsnew6to7 Stata 7.0 new release 2000 | | whatsnew6 Stata 6.0 1999 to 2000 | +---------------------------------------------------------------+

Most recent changes are listed first.

-------- more recent updates --------------------------------------------------

See whatsnew14to15.

-------- update 29jan2018 -----------------------------------------------------

1. New command lmbuild makes creating Mata function libraries easier than does mata mlib.

2. estat mindices, when specified after sem with a large number of observed endogenous variables, sometimes reported a missing value for all the modification indices representing an omitted path or covariance. This has been fixed.

Our technical services team previously recommended using the undocumented option slow when estat mindices reported missing values. This option is no longer necessary.

3. ml, when specified with option technique(bhhh), sampling weights, and option group(), exited with an uninformative error message even when the weights were constant within group. This has been fixed.

-------- update 19dec2017 -----------------------------------------------------

1. Community-contributed command dataex, which generates properly formatted data examples for Statalist, is now being distributed with official Stata for the convenience of anyone who wishes to post data on Statalist.

2. areg, when fitting a model with the absorb() variable specified among the indepvars and with option vce(cluster clustvar) where clustvar is a variable other than the absorb() variable, incorrectly exited with error message "not sorted". This has been fixed.

3. predict with option ystar, ystar(a,b), or e(a,b) after churdle exponential produced incorrect results. This has been fixed.

The difference between the old and the new results is negligible when the values of the dependent variable are large, such as 100 or larger in our experiments. Large values of the dependent variable are typical with exponential mean models. However, in the rare case where the dependent variable is small, differences between old and new predictions may be larger; in our experiments, the change was at most 6%.

4. esize twosample, when a string variable was specified in option by(), incorrectly exited with an uninformative error message. This has been fixed.

5. gsem, when fitting a model with both endogenous and exogenous latent variables and with an interaction term containing a latent variable, exited with an uninformative error message. gsem now fits the model as specified.

6. gsem, when option capslatent was specified or implied, ignored variables in the model if their first letter was a capital letter. Now gsem recognizes those variables as latent variables.

7. margins has the following improvement and fixes:

a. margins has improved numerical derivatives logic for computing the Jacobian matrix.

b. margins with option asbalanced, when working with estimation results from ordinal outcome models like ologit and oprobit and when a factor variable was not among the marginsvars, incorrectly exited with error message "not conformable". This has been fixed.

c. margins with option pwcompare, when used to compute margins for two or more interaction terms that involved the same number of factor variables, incorrectly computed the values of the margins. This has been fixed.

8. marginsplot, when plotting marginal effects of multiple factor variables in options dydx() and eydx(), has improved spacing of the x-axis scale.

9. melogit and mepoisson with option intmethod(laplace}, when used to fit a two-level model, may not have converged when the model could have been identified. This has been fixed.

10. Prefix command mfp used with intreg incorrectly dropped all left- or right-censored observations from the estimation sample. mfp now keeps these observations.

11. ml with method d1debug or method d2debug, when the specified model contained factor variables, exited with error message "could not calculate numerical derivatives". This has been fixed.

12. predict after logit and probit, when the fitted model omitted the base level of a factor variable because it was a perfect predictor, did not properly assign missing values to the predictions according to the information in the matrix e(rules). This has been fixed.

13. putexcel ul_cell=picture(filename) gave an error if you used compound quotes around filename. This has been fixed.

14. pwcorr has the following fixes:

a. pwcorr with no observations and only two variables specified, incorrectly stored r(rho) as 0. It now exits with error message "no observations".

b. pwcorr with no observations for the first two variables and more than two variables specified, incorrectly stored r(rho) as 0. It now correctly stores r(rho) as missing.

15. stcox, when specified with option shared() and variable names in indepvars that exceeded the default 12-character width for the table, produced a coefficient table with misaligned elements in the row for the theta parameter estimate. This has been fixed.

16. stepwise with options lockterm1 and lr and forward selection or forward stepwise selection computed likelihood ratios at the first addition step by comparing each model with an additional term with an empty model rather than with the model that includes the forced first term. Thus, terms could be added to the model that did not satisfy the specified significance level for inclusion. This has been fixed.

17. xtile, altdef failed to use the alternative formula for calculating percentiles and instead calculated them using the default formula. This has been fixed.

18. Edit > Copy Table could omit the minus sign from negative numbers if they were in the first cell of a row being copied. This has been fixed.

19. (Mac) macOS 10.13 introduced a change that caused Stata's Data Editor to crash when used with an if expression. This has been fixed.

-------- update 10oct2017 -----------------------------------------------------

1. putexcel now supports existing option overwritefmt when using the syntax for formatting cells. This lets you write Excel cell formats more efficiently.

2. asclogit, asmprobit, asroprobit, and nlogit, when option case() specified a variable with values exceeding c(maxlong) = 2,147,483,620, grouped all the associated observations into a single case. This has been fixed.

3. graph ignored legend option ring() if the scheme placed the legend within the plot region. This has been fixed.

4. graph export, when exporting to PDF a Greek letter that had the italic font applied, did not display the Greek letter as italicized. This has been fixed.

5. import excel, when the Excel file was specified as a path and filename without an extension, exited with error message "file not found" even when the file existed. This has been fixed.

6. margins with weighted estimation results, with the first observation in the dataset containing a missing weight value, and with all the independent variables fixed to a constant value incorrectly estimated all margins as zero valued with the "(omitted)" note. This has been fixed.

7. marginsplot, after margins where a factor-variable name could also have served as an abbreviation for a continuous variable and both variables were specified in option at(), exited with error message "_marg_save has a problem. Margins not uniquely identified." This has been fixed.

8. tebalance box, when used after teffects with a treatment variable using value labels with spaces, incorrectly exited with error message "invalid syntax". This has been fixed.

9. zip and zinb no longer support option vuong, which specifies that the Vuong test for nonnested models be reported. This test was used to compare zero-inflated and noninflated models. However, recent work has shown that testing for zero inflation using the Vuong test is inappropriate. If you wish to proceed with the test anyway, use the zero-inflated commands with undocumented option forcevuong.

10. (Windows) window fopen always displayed all files instead of the filtered list of files. This has been fixed.

11. (Unix GUI) The "Save preferences" dialog box did not open to the correct directory. This has been fixed.

-------- update 26sep2017 -----------------------------------------------------

1. (Mac) A bug in macOS High Sierra (10.13) could cause Stata to crash when dialogs that were opened from the Data, Statistics, and Graphics menus were closed. This has been fixed.

-------- update 13jul2017 -----------------------------------------------------

1. mixed with option dfmethod() is now faster when computing degrees of freedom using Satterthwaite and Kenward-Roger methods. The increase in speed is substantial with clusters containing thousands of observations.

2. estat eform is now supported after svy: gsem.

3. bayesmh, when Gibbs sampling was used for a covariance matrix parameter with an inverse-Wishart prior and when the degrees of freedom was specified as a noninteger value, returned an uninformative error message. bayesmh now exits with error message "the degrees of freedom must be an integer number".

4. clogit with option vce(bootstrap) did not allow factor-variables notation. This has been fixed.

5. Expressions that included a numeric literal longer than 129 characters could cause Stata to crash. This has been fixed so that Stata can handle numeric literals of arbitrary length.

6. fvrevar now attaches the default numeric format to all indicator variables generated from a factor variable. Previously, the numeric format of the original factor variable was copied to the new variables. The old behavior is not preserved under version control.

7. graph export has the following fixes:

a. graph export, when exporting to PDF file format a pie graph where the angle of a slice was close to 0 degrees, incorrectly displayed the pie graph as a whole circle. This has been fixed.

b. graph export, when exporting to EPS file format a graph where a truetype font was specified with option fontface(), did not render text as boldface or italic. This has been fixed.

8. gsem, with multilevel latent variables, multiple outcome variables, and missing values in some of the observed variables, incorrectly exited with error message "missing values not allowed in S". This has been fixed.

9. predict with option scores, when used after gmm with option onestep and when onestep was abbreviated, exited with an uninformative error message. predict now computes the scores as specified.

10. import delimited, when used with option stringcols() or option numericcols(), could incorrectly import the first row as variable names. This has been fixed.

11. Mata functions lnwishartden(df, V, X) and lniwishartden(df, V, X) caused Stata to use more memory than was required. The practical consequence was that Stata ran slower when these functions were called repeatedly, such as by bayesmh with models using Wishart and inverse-Wishart priors. This has been fixed.

12. mfp with the default closed-test algorithm for choosing a multivariable fractional polynomial (MFP) model reported p-values that were slightly smaller than they should have been. In most cases, the difference occurred beyond the third decimal place. This has been fixed.

Note that the differences in p-values rarely affected the variables selected for the final MFP model and that the p-values reported in the regression table for the final model were correct.

13. Prefix mi estimate, when option cmdok was specified to allow estimation with unsupported command gsem, incorrectly exited with a Mata trace log and error message "name conflict" when random effects were included with gsem. mi estimate now fits the specified generalized SEM model.

14. mi impute chained with option savetrace() did not respect spaces or quotes in folder names or filenames. This has been fixed.

15. ml, when option technique() specified a switch between nr and bfgs or between nr and dfp and when constraints were specified, incorrectly exited with an error message. This has been fixed.

16. pctile with option genp(newvarp) and more than 1,000 quantiles specified in option nquantiles() did not store the percentages in the observations that held the corresponding percentiles. This has been fixed.

17. predict with option latent(), when used after gsem and when the fitted model had paths between endogenous latent variables, incorrectly exited with an error message. This has been fixed.

18. saveold with option all, when the dataset in memory contained the e(sample) variable, incorrectly exited with error message "variable name ... does not meet the restrictions of old .dta format". This has been fixed.

19. sem with a symbolic constraint name defined using @, such as @a, when the constraint was also identified with an automatically determined anchor, exited with error message "'a' not found" if the symbolic constraint name was used in any other constraints. sem now retains the symbolic constraint name.

20. stcrreg with prefix stepwise and parentheses binding the first predictor incorrectly exited with error message "option compete() required" even when option compete() was specified. This has been fixed.

21. Function tden(df,t), in extreme cases in the tails of the distribution where the density is close to 0, returned a value that was larger than the correct value. This has been fixed.

22. tebalance box and tebalance density after teffects, when the working directory contained many .gph files, incorrectly exited with error message "invalid syntax". This has been fixed.

23. unicode convertfile with option dstcallback() applied the default method stop regardless of the method specified. This has been fixed.

24. Function ustrregexs() in Stata and Mata without first performing a valid regular expression match using ustrregexm() caused Stata to crash. This has been fixed.

25. (Unix GUI) When using a MobaXterm Windows client, the Find dialog in the Do-file Editor froze if no matches were found for the text specified in the "Find what:" field. This has been fixed.

-------- update 04may2017 -----------------------------------------------------

1. arfima, when the time-series data were nonstationary and feasible initial values could not be obtained, exited with an uninformative error message. The command now exits with error message "initial values not feasible".

2. factor, when any variable in the analysis is constant, exited with error message "conformability error". Now, factor drops the constant variable and uses the remaining variables for the analysis.

3. graph box and graph hbox, when the upper adjacent value equals the 75th percentile plus 1.5 times the interquartile range (IQR) or the lower adjacent value equals the 25th percentile minus 1.5 times the IQR, plotted the data whose value equals the adjacent values as an outlier. Because an outlier is defined as a value greater than the upper or less than the lower adjacent value, this value is not an outlier. This has been fixed.

4. icd10 lookup, when the requested code range contained a formatted code (one that includes a "."), incorrectly reported "(no matches found)". This has been fixed.

5. irt rsm now imposes constraints on the threshold structure of the items by using the underlying IRT parameterization to fit the model described by Andrich (1978a, Applied Psychological Measurement 2: 581-594; 1978b, Psychometrika 43: 561-573). Previously, irt rsm imposed constraints by using the slope-intercept parameterization. The old behavior is not preserved under version control.

6. predict with option cm after tnbreg and tpoisson, when the lower truncation point specified in option ll() of the estimation command was not 0, did not account for truncation correctly when computing the conditional mean. As a result, the conditional mean was overestimated. This has been fixed.

7. svy: tabulate, when tabulating numeric variables with value labels or string variables that contain periods (".") or colons (":"), now preserves these characters in the row and column titles of the reported table. Previously, "." and ":" were translated to "," and ";". The old behavior is not preserved under version control.

-------- update 16mar2017 -----------------------------------------------------

1. Online help and the search index have been brought up to date for Stata Journal 17(1).

-------- update 07mar2017 -----------------------------------------------------

1. estat eqgof, after fitting a model with sem that included latent endogenous variables and did not include any exogenous variables, incorrectly reported zero values for the R-squared statistics. This has been fixed.

2. import excel has the following fixes:

a. import excel with option sheet(sheetname), when the sheetname started with $, incorrectly named the sheet Sheet1. This has been fixed.

b. import excel incorrectly used Stata format %tchh:MM:SS instead of %tcHH:MM:SS when importing Excel custom time format h:mm:ss data. This has been fixed.

3. margins has the following fixes:

a. margins with options asbalanced and emptycells(reweight) treated interaction terms without factor variables as zero. This has been fixed.

b. margins with option predict() and suboption psel or xbsel after heckman, heckprobit, or heckoprobit incorrectly excluded nonselected observations with a missing value for a covariate not specified in select() at estimation. This has been fixed.

4. power oneproportion with option continuity overestimated the power for the two-sided test. This has been fixed.

5. predict with option te after xtfrontier, when estimation had been performed for a sample restricted to a certain range by using qualifiers if or in, returned incorrect results. This has been fixed.

6. Do-files and ado-files that contain nested open comment delimiters with no spaces between them (for example, /*/*) could cause Stata to crash. This has been fixed.

7. The Java Runtime Environment that is redistributed with Stata is now updated to version 8 update 121.

8. (Mac) In the Data Editor, when the Option key+e combination was used to enter an accented character into a cell and when the inline edit field for the cell had not been initiated prior to entering text, the first typed character was not entered. This has been fixed.

9. (Mac) In the Data Editor, when an international keyboard (such as a Japanese keyboard) was used to initiate the inline edit field of a cell, the first typed character was not entered. This has been fixed.

10. (Mac) Stata could crash when a document that was printed from the Do-file Editor took a long time to print. This has been fixed.

11. (Mac) When opening a file in the Do-file Editor that is not encoded in UTF-8, Stata attempted to convert the file to UTF-8 using the default character encoding for your locale. For some locales, this conversion could not be done using the default encoding, and nothing appeared in the Do-file Editor (for example, opening a do-file encoded in Windows Latin 1 from a Japanese locale). Stata now prompts you to specify the character encoding of the file before converting it to UTF-8.

12. (Mac) Programmable dialogs with combo boxes that have editing disabled did not register a selection from the combo box's drop-down list. This has been fixed.

-------- update 09jan2017 -----------------------------------------------------

1. Online help and the search index have been brought up to date for Stata Journal 16(4).

2. bayestest model, when all marginal likelihoods were evaluated to be zero because of numerical overflow, did not evaluate the posterior probabilities and reported missing values. The calculation of the posterior probabilities is now more precise, so missing values are far less likely.

3. suest, when used to compute the sandwich estimator of the asymptotic variance for the fitted coefficients from nonsvy estimation results, now uses e(V_modelbased) if it is present instead of e(V). The old behavior is not preserved under version control.

-------- update 19dec2016 -----------------------------------------------------

1. estat gof, after fitting a model with an overidentified mean structure using sem, overestimated the value of SRMR. estat gof no longer includes the modeled means when calculating SRMR. The old behavior is not preserved under version control.

2. margins has the following fixes:

a. margins after stcrreg and with option vce(unconditional) incorrectly exited with an error. This has been fixed.

b. margins after svy with multilevel weights exited with error message "svyset characteristics disagree with svy estimation results" even when the svyset characteristics were not changed. This has been fixed.

3. Mata function _docx_new(), when called after _docx_new_table(), cleared previously returned table ID codes. This prevented tables that were created before _docx_new() was called from being edited or added to the document. This has been fixed.

4. Mata function fillStataMatrix(name, colnames, rownames) incorrectly filled the first row of the table with the matrix rownames of the matrix when the colnames argument was not 0. This has been fixed.

5. Mata function lnmvnormalden(M,V,X) exited with error message "conformability error" when M and X were void vectors and V was a 0 x 0 matrix. lnmvnormalden() now returns a 0 x 0 matrix in this case.

6. Mata functions lnwishartden(df,V,X) and lniwishartden(df,V,X) exited with error message "conformability error" when V and X were 0 x 0 matrices. lnwishartden() and lniwishartden() now return a 0 x 0 matrix in this case.

7. power twoproportions has the following change in behavior and fix:

a. power twoproportions with option test(fisher), when either the control-group or the experimental-group sample size is larger than 1,000, reported a missing value. It now exits with error message "failure to compute power".

b. power twoproportions with option test(fisher), when the total sample size is larger than 1,028 and both the control-group and experimental-group sample sizes are smaller than or equal to 1,000, underestimated the power. It now exits with error message "failure to compute power".

8. (Mac) When using Stata in full-screen mode on macOS Sierra, some graphs appeared in a low resolution or did not appear at all after their window finished transitioning into a full-screen window. This has been fixed.

9. (Mac) macOS Sierra 10.12.2 introduced a change that caused Stata's Do-file Editor to crash. This has been fixed.

10. (Mac and Unix GUI) When saving an SEM path diagram, Stata prompted the user for a filename even if the diagram had been saved before. This has been fixed.

-------- update 16nov2016 -----------------------------------------------------

1. You can now comment or uncomment a selection of lines in the Do-file Editor by selecting the "Edit > Advanced > Toggle comment" menu item.

2. Mata function _docx*() has the following improvement and fix:

a. New Mata function _docx_cell_set_border(dh, tid, i, j, name, val, color) allows you to format the borders of the cell on the ith row and jth column, in which name, val, and color specify the border name, border style, and border color.

b. _docx_table_set_width(dh, tid, type, w), when argument type was auto, did not make the columns in a table automatically fit the contents. This has been fixed.

3. _diparm after meglm, melogit, meprobit, mecloglog, meologit, meoprobit, mepoisson, menbreg, and mestreg, when eqname referred to a variance component, incorrectly exited with error message "equation ___ not found". _diparm now works as documented.

4. margins has the following fixes:

a. margins with option asbalanced exited with an uninformative error message when factor variables were included in a model fit using gsem, meglm, melogit, meprobit, mecloglog, mepoisson, menbreg, meologit, meoprobit, mestreg, xtologit, or xtoprobit. margins now computes the requested marginal means, predictive margins, or marginal effects.

b. margins, when used after ologit or oprobit with no indepvars or a model Wald test with zero degrees of freedom, exited with error message "0 invalid name" instead of computing the requested margins. This has been fixed.

c. margins after xtreg, be with option wls incorrectly exited with an uninformative error message. margins now works as documented.

5. Mata function setCellTopBorderWidth(i, j, sz) incorrectly removed the right border instead of the top border of the cell on the ith row and jth column when the specified size sz was 0. This has been fixed.

6. ml plot did not update starting values for use with ml. This has been fixed.

7. predict with option reffects or mu after fitting a model with meglm, melogit, meprobit, mecloglog, meologit, meoprobit, mepoisson, menbreg, or mestreg could return zero values instead of the empirical Bayes mean estimates for groups with a large number of observations. This has been fixed.

8. predict with option latent or mu after fitting a model with gsem could return zero values instead of the empirical Bayes mean estimates for groups with a large number of observations. This has been fixed.

9. qnorm, when used with qualifier if or in, drew a y-axis scale that accounted for the observations not selected by the if or in condition. This has been fixed.

10. Stata function regexs(), when used in an expression, could crash or produce wrong results in Stata/MP. The incorrect results were most often scrambled text. This has been fixed.

11. sem with option technique(bhhh), when the model was not identified, sometimes exited with an uninformative error message before reaching the maximum number of iterations. This has been fixed.

12. stcrreg, when the data were stset using option exit(), reported incorrect numbers of competing risks and censored observations. The estimated coefficients were not affected. This has been fixed.

13. uselabel, when used to create a dataset from labels that contained numeric values requiring double precision, incorrectly stored the values in float precision. This has been fixed.

14. xtlogit, xtprobit, and xtcloglog with option vce(robust) or option vce(cluster panelvar) and prefix by incorrectly included the whole sample in the calculation of the standard errors instead of limiting the sample to each by-group. This has been fixed.

15. translate has the following improvement and fix:

a. (Mac) Translators smcl2prn, txt2prn, smcl2pdf, and txt2pdf of translate now support translator settings pagesize, pagewidth, and pageheight. These settings let you, for example, specify landscape printing for a standard letter page (11 x 8.5 inches). For more information, see [R] translate.

b. (Mac) On some Macs, using translate to export the Results or Viewer window to a PDF file could print the contents of the window instead. This has been fixed.

-------- update 29sep2016 -----------------------------------------------------

1. Online help and the search index have been brought up to date for Stata Journal 16(3).

2. destring, when used to convert a number stored as text that included leading or trailing Unicode whitespace characters, converted the number to missing instead of its numeric value. This has been fixed.

3. After the 30mar2016 update, estat concordance, when the number of comparison pairs, the number of orderings as expected, or the number of tied predictions was greater than 10 million, incorrectly reported rounded values for the statistic that was greater than 10 million. This has been fixed.

4. gmm, when used with option xtinst() and when the data contained panels that were completely dropped from the estimation sample, incorrectly exited with error message "esample() invalid". This has been fixed.

5. For graphics commands that support adding additional plots using option addplot(), suboption below was respected on the initial drawing, and the added plot was drawn below the command's own plot or plots. However, if the resulting graph was saved with graph save or combined with graph combine, then the added plot was brought back above the command's own plot or plots. This has been fixed.

-------- update 14sep2016 -----------------------------------------------------

1. (Windows) After the 06sep2016 update, Stata displayed all matching variable names as a single menu item rather than separate menu items when the Tab key was used to autocomplete a partially typed variable name in the Command window. This has been fixed.

-------- update 06sep2016 (Stata version 14.2) --------------------------------

1. The icd10 suite of commands has the following improvements and changes:

a. New command icd10 search allows you to search the text descriptions of ICD-10 codes for specified keywords. With option version(year), you can restrict the search to only codes valid in the specified year.

b. icd10 check has new option summary. summary is a reporting option that provides a frequency of each invalid or undefined code.

c. icd10 clean now requires either option generate() or replace. generate() creates a new variable that contains the cleaned ICD-10 codes. replace replaces the existing ICD-10 codes in the specified variable.

d. icd10 generate with option description can now be combined with new option addcode(begin|end). addcode(begin) adds the ICD-10 code to the beginning of the description and replaces existing option long. addcode(end) adds the code to the end of the description. long continues to work but is undocumented.

e. icd10 generate with options description and addcode() can now be combined with new options nodots and pad to control the formatting of the code added to the description.

f. icd10 lookup now accepts a range of ICD-10 codes instead of a single code.

g. icd10 clean and icd10 generate have new option check to check the formatting of the ICD-10 codes before either cleaning or creating a new variable. This is a convenience feature to avoid running icd10 check first.

h. icd10 check, icd10 clean, icd10 generate, and icd10 lookup have new option version(year), which replaces existing option year(). year() continues to work but is undocumented.

i. icd10 check has new option fmtonly, which replaces existing option any. any continues to work but is undocumented.

2. gmm has the following improvements for postestimation after generalized method of moments (GMM) estimation:

a. margins is now available after gmm. This lets you estimate marginal and conditional effects such as means, predictive margins, and marginal effects after GMM estimation.

b. predict after gmm now allows option xb. This lets you obtain the linear prediction after GMM estimation.

c. predict after gmm now takes option scores. This lets you obtain the values of the moment conditions after GMM estimation.

3. predict after irt, gsem, meglm, melogit, meprobit, mecloglog, meologit, meoprobit, mepoisson, menbreg, and mestreg has the following improvements:

a. predict now supports out-of-sample predictions of empirical Bayes means, empirical Bayes modes, and other predictions that are conditional on them.

b. predict now computes predictions of empirical Bayes means and empirical Bayes modes even when the estimation sample has changed.

The old behavior is preserved under version control.

4. sem has improved starting values logic for models with latent and observed exogenous variables when option method(mlmv) is specified or option noxconditional is specified or implied by the model specification. The result of this is faster convergence of the fitted target model.

5. bayesmh, when used with factor variables as predictors and when showbaselevels was set, incorrectly included the base-level parameters in the calculation of the maximum and average sampling efficiencies, which led to increased estimates of these summaries. The base-level parameters are omitted from the model and as such are not sampled. This has been fixed.

6. fp, when a search was performed on a model that had been fit by clogit and that model included only the fractional polynomial powers, exited with the error message "redundant or inconsistent constraints". fp now fits the model as specified.

7. Function lnigammaden(a,b,x), when x was less than or equal to 0, returned a value of 0. lnigammaden() now returns a missing value if x is less than or equal to 0.

8. gmm has the following improvement and fixes:

a. gmm now reports the number of panels used to compute standard errors for models with panel-style instruments.

b. gmm, when option onestep was combined with option winitial(identity), reported incorrect standard errors. This has been fixed.

c. gmm with option nocommonesample, when specified with instruments that were not shared between moment equations, either did not converge or produced incorrect parameter estimates. This has been fixed.

d. gmm, when used with a moment-evaluator program that included a time-series operated instrumental variable, incorrectly excluded observations in a panel from the estimation sample when the same time-series operator was applied to a variable used as an instrument and a variable defined in the moment-evaluator program. For example, specifying L#.z in the instruments() option and L#.u in the evaluator program would result in extra observations being dropped. This has been fixed.

9. import delimited with option varnames(nonames), when the first line of the text file to be imported contained a missing value indicated by ".", incorrectly imported the column as a string, even if all subsequent values in the columns were numeric or missing. This has been fixed.

10. After the 19may2016 update, irt commands, when used to fit a model where some observations had missing values for all items or when specified with option listwise, incorrectly included the observations with missing values in e(sample). Estimation results were not affected, but subsequent commands such as predict that rely on e(sample) incorrectly included these observations. This has been fixed.

11. margins has the following fixes:

a. margins, when used after xtlogit, xtprobit, and xtcloglog had been used to fit the default random-effects model where indepvars included the same continuous variable multiple times and when that variable was not specified in the margins command, returned estimated margins that set the effect of the variable to 0 in the calculations. This has been fixed.

b. margins, when used after xtlogit, xtprobit, and xtcloglog had been used to fit the default random-effects model where indepvars included the same continuous variable multiple times and when that variable was specified in the margins command, exited with an uninformative error message. margins now computes the requested marginal effects.

c. margins with svy estimation results and the default variance estimation method (vce(delta)) ignored the design degrees of freedom. This means that confidence intervals were narrower than they should have been and p-values were smaller than they should have been. This has been fixed.

12. Mata function fputmatrix(), when used to write a Mata struct that had been previously read by fgetmatrix() and when the Mata struct contained a member of type Mata struct, did not write the struct to the file. This has been fixed.

13. predict with option het, after fitting a model with arch, ignored qualifiers if and in. This has been fixed.

14. predict with option distance, after fitting a model with teffects nnmatch that included factor variables in the outcome model, exited with an uninformative error message. The nearest-neighbor distances are now calculated as requested.

15. putexcel has the following fixes:

a. putexcel with option sheet(), when the specified Excel worksheet name included leading or trailing whitespace characters, did not write the whitespace characters with the text. This has been fixed.

b. putexcel with option sheet() did not allow commas in an Excel worksheet name. This has been fixed.

c. putexcel, when writing Stata missing values, incorrectly overwrote a cell's formatting. This has been fixed.

16. svy, when used as a prefix to etregress and fracreg, gave overly conservative standard errors. This has been fixed.

17. svy now defaults to using e(V_modelbased) instead of e(V) when e(V_modelbased) is available. This change is being made because using e(V) for commands that do not assume that the default variance-covariance matrix is independent and identically distributed results in overly conservative estimates of the standard errors. The old behavior is not preserved under version control.

18. tsreport has improved output when the time gaps in the data occur in periods where the combined number of digits in the observation numbers that identify the gap exceed 11 characters.

19. twoway fpfit with missing values in yvar did not always produce the same graph when the same command was submitted multiple times. This has been fixed.

20. xtologit and xtoprobit with singleton groups and pweights incorrectly exited with an uninformative error message. The commands now fit the model as specified.

21. (Mac) Selecting "Replace all in selection" from the Find and Replace bar in the Do-file Editor replaced all matching text in the document instead of just the matching text in the current selection. This has been fixed.

22. (Unix) After the 20jul2016 update, opening the Data Editor twice in one session on a computer with the Ubuntu 16.04.1 LTS distribution caused Stata to crash. This has been fixed.

-------- update 20jul2016 -----------------------------------------------------

1. Stata has been updated to account for the leap second that will be added by the International Earth Rotation and Reference Systems Service on December 31, 2016. Stata's datetime/C UTC date format now recognizes 31dec2016 23:59:60 as a valid time.

2. estat teffects did not correctly compute the standard errors for total and indirect effects between endogenous variables. This has been fixed.

3. menbreg with option dispersion() incorrectly exited with error message "option dispersion() not allowed". This has been fixed.

4. mixed with robust standard errors and prefix by incorrectly exited with an error message. This has been fixed.

5. predict with option yhat, option residuals, or option scores incorrectly exited with an error message when used after nl with option lnsql(). This has been fixed.

6. statsby with options by() and total, when specified as a prefix for mixed with robust standard errors, returned correct values of the requested statistics for each by-group but returned only missing values when computed using all observations in the dataset. This has been fixed.

7. xtreg with a string variable specified in option vce(cluster clustvar), when used to fit a fixed-effects or random-effects model, incorrectly exited with an error message. This has been fixed.

8. After the 06jul2016 update, all user-written commands accessed from the User menu incorrectly exited with error message "too many menus (limit of 1250)". This has been fixed.

-------- update 06jul2016 -----------------------------------------------------

1. odbc load has new option bigintasdouble, which specifies storing database BIGINT columns as Stata type double when the data are loaded instead of using the default type (string). If any integer value is larger than 9,007,199,254,740,965 or less than -9,007,199,254,740,992, this conversion is not possible, and odbc load will issue an error message.

2. Online help and the search index have been brought up to date for Stata Journal 16(2).

3. estat gof, after fitting a saturated model with poisson, could report a negative deviance statistic. This has been fixed.

4. irf create after svar displayed a dot instead of a red X when a bootstrap replication failed. It now displays a red X for each failed replication and exits with an error message if there are insufficient observations to compute bootstrap standard errors.

5. ivprobit with options first and twostep did not display the linear regressions for the endogenous covariates. This has been fixed.

6. margins with results from svy: meologit or svy: meoprobit incorrectly exited with error message "conformability error". This has been fixed.

7. Mata function fgetmatrix() did not read a Mata class or struct that contained a member of transmorphic type and nonreal contents. This has been fixed.

8. ml with option group() and the default variance estimation method calculated the correct standard errors but incorrectly reported that the standard errors were adjusted for clustering. This has been fixed.

9. nlsur with option vce(cluster clustvar), when clustvar did not vary within the estimation sample, incorrectly exited with error message "estimates repost: matrix has missing values". This has been fixed.

10. predict, after factor variables with operator bn. (no base) were specified in gsem, meglm, melogit, meprobit, mecloglog, meologit, meoprobit, mepoisson, menbreg, mestreg, xtologit, or xtoprobit, computed predictions as if the coefficients for the first levels of those factor variables were zero instead of the estimated values. This has been fixed.

11. regress with a constant factor variable and pweights or aweights could fail to omit the constant factor variable. This has been fixed.

12. tabstat with option varwidth() now displays longer variable names in the stub of the table when statistics are displayed in the table columns. The maximum length is now the maximum name length in Stata, 32, instead of 16 characters.

13. unicode translate and unicode analyze, when the specified Stata dataset contained value-label contents that were blank, did not process the dataset. This has been fixed.

14. xtivreg has the following fixes:

a. xtivreg, when used to fit the default random-effects model and with options first and small, reported normal (Z) and chi-squared statistics instead of small-sample t and F statistics for the first-stage regression. This has been fixed.

b. xtivreg, when used to fit a fixed-effects model and with option first and either option vce(robust) or option vce(cluster clustvar), reported standard errors from areg in the first stage. xtivreg, fe now reports standard errors from xtreg, fe in the first stage.

15. xtreg with option vce(cluster clustvar) and either default option re or option fe, when a string variable was specified as clustvar, incorrectly exited with error message "type mismatch". This has been fixed.

16. After the 28 Jan 2016 update, when a custom font was embedded in a graph exported to a Postscript file, text was displayed incorrectly in Postscript renderers. This has been fixed.

17. (Unix GUI) For Ubuntu 14.04 users, the xstata Data menu was disabled after opening the Data Editor. This has been fixed.

-------- update 19may2016 -----------------------------------------------------

1. A new family of functions computes probabilities and other quantities of the inverse Gaussian distribution.

igaussian(m,a,x) computes the cumulative distribution function (CDF) of the inverse Gaussian distribution with mean m and shape a.

igaussianden(m,a,x) computes the density of the inverse Gaussian distribution with mean m and shape a.

igaussiantail(m,a,x) computes the reverse CDF of the inverse Gaussian distribution with mean m and shape a.

invigaussian(m,a,p) computes the inverse CDF of the inverse Gaussian distribution: if igaussian(m,a,x) = p, then invigaussian(m,a,p) = x.

invigaussiantail(m,a,p) computes the inverse reverse CDF of the inverse Gaussian distribution: if igaussiantail(m,a,x) = p, then invigaussiantail(m,a,p) = x.

lnigaussianden(m,a,x) computes the natural logarithm of the density of the inverse Gaussian distribution with mean m and shape a.

rigaussian(m,a) computes inverse Gaussian variates with mean m and shape a.

2. gmm and ivregress now allow an optional tuning parameter that affects the number of lags selected by Newey and West's (1994) optimal lag-selection algorithm. To specify the tuning parameter, use option wmatrix(hac kernel opt [#]) or option vce(hac kernel opt [#]).

3. bayesmh has new option initrandom to randomly initialize model parameters. This option is useful for initializing multiple Markov chain Monte Carlo chains with different starting values. When parameters cannot be randomly initialized, such as with user-defined prior distributions, initrandom can be combined with option initial(), which assigns fixed starting values to selected model parameters.

4. irt commands are now faster for models that have many observations and few items.

5. roctab is now faster.

6. tab1 now allows aweights and iweights.

7. New Mata class AssociativeArray() provides an interface into asarray().

8. Mata structure scalars may now be used in an interactive Mata session.

9. New Mata function classname() returns the class name of a Mata class scalar.

10. New Mata function structname() returns the structure name of a Mata structure scalar.

11. Mata functions fputmatrix() and fgetmatrix() now allow you to save and read Mata class matrices in addition to string, numeric, and pointer matrices.

12. include has new option adopath, which searches Stata's system directories for the specified file if the file is not found at the default location.

13. areg, when the model included an interaction between a single specified level of a factor variable and a continuous variable, omitted the main effect for the factor variable. This has been fixed.

14. asclogit with alternative-specific covariates used better starting values in Stata 13, meaning that the model converged quicker (required fewer iterations). Stata 14 now uses the Stata 13 starting values.

15. churdle and fracreg, when used to fit a model that included factor variables, reported the correct model chi-squared statistic but with incorrect degrees of freedom. The degrees of freedom was too large because of the inclusion of the base levels of the factor variables, resulting in a conservative model test. This has been fixed.

Note: It is unlikely that inference about the overall model significance test will change except in cases with small values of the model chi-squared statistic.

16. cii means with option poisson, in the unusual case that the confidence level was set to 99.06 or higher, and with exactly one event, did not return results or exit until the Break key was pressed. This has been fixed.

17. estimates table, when used after bayesmh to display stored results that were not obtained using bayesmh, incorrectly exited with error message "estimates table not allowed after bayesmh". This has been fixed.

18. glm and poisson with iweights, when the weights contained negative values, incorrectly exited with error message "negative weights encountered". This has been fixed.

19. infile, when used with a dictionary file that specified more variables than the current maxvar setting allowed, caused Stata to crash. This has been fixed.

20. ivregress 2sls with options first and vce(cluster clustvar) did not scale the first-stage standard errors properly. This was a reporting error, and calculations for the full model and related inference were unaffected. This has been fixed.

21. margins, when used with estimation results that have a single predictor variable xvar and with option at(xvar=generate(exp)), used only the first observation instead of the entire estimation sample. This has been fixed.

22. Mata function asarray_remove() could fail when the key dimension for the associative array was greater than 1. This has been fixed.

23. Mata function st_subview(), when the first argument was the same as the second argument, could cause Stata to crash. This has been fixed.

24. mi impute with a user-defined imputation method and option replace, when used with data declared in flong or mlong style, exited with an error. In style flong, it exited with error message "no observations". In style mlong, it exited with error message "missing imputed values produced". This has been fixed.

25. mi reshape with option string did not properly register imputed, passive, and regular variables. In style mlong, when the reshaped variables included imputed variables, and in style wide, when the reshaped variables included imputed or passive variables, imputed data were lost for the reshaped variables. This has been fixed.

26. mixed, when the model contained a lagged variable and when vce(robust), vce(cluster clustvar), or probability weights were specified, incorrectly exited with error message "error obtaining scores for robust variance". This has been fixed.

27. nlogit ignored option collinear when checking collinearity among variables specified at different levels of the hierarchy. This has been fixed.

28. reg3 is documented to support time-series operators in options exog(), endog(), and inst(). However, the command exited with error message "time-series operators not allowed" when a time-series operator was used. This has been fixed.

29. strmatch(s1,s2) treated "*" as a literal instead of as a wildcard character if it was in the same position in string s1 and pattern s2. For example, strmatch("ab*c", "ab*") returned 0 (not matched) instead of 1 (matched). This has been fixed.

30. The stteffects commands, when used with data that were previously stset with options enter() and exit(), which are disallowed by stteffects, exited with an error message after stset was resubmitted with an allowed specification. This has been fixed.

31. svyset, when the same sampling weight variable specified as a pweight or an iweight is included in the varlist specified with option brrweight(), jkrweight(), bsrweight(), or sdrweight(), now exits with an error message.

32. test and testparm, used to test hypotheses for margins estimation results computed following a survey estimation command with bootstrap or successive difference replicate standard errors, reported a missing F statistic instead of an unadjusted chi-squared statistic. This has been fixed.

33. xtgee with option nmp, when one or more indepvars were omitted because of collinearity, overestimated the standard errors, leading to conservative estimates of the reported z statistics and p-values. This has been fixed.

34. xtlogit and xtprobit with option vce(robust) or option vce(cluster panelvar), when any of the variables in the model were specified with a time-series operator, incorrectly exited with error message "calculation of robust standard errors failed". This has been fixed.

35. zip and zinb with option inflate(_cons) and either vce(bootstrap) or vce(jackknife) exited with error message "variable _cons not found". This has been fixed.

36. In the Results window, copying a table with spaces before the horizontal separators resulted in the table being copied as is instead of in a delimited format that could be pasted into another application. This has been fixed.

37. (Windows) In the Command window, selecting "Undo" from the contextual menu did not undo the last edit. This has been fixed.

-------- update 30mar2016 -----------------------------------------------------

1. pcorr now supports factor variables.

2. roccomp is now faster.

3. estat concordance after stcox is now faster.

4. Online help and the search index have been brought up to date for Stata Journal 16(1).

5. betareg with option vce(cluster clustvar), when a string variable was specified as clustvar, incorrectly exited with error message "no observations". This has been fixed.

6. bayesmh has the following fixes:

a. Suboption noglmtransform of bayesmh's option likelihood(), which was used to fit the exponential, binomial, and Poisson distributions to the outcome variable, is now undocumented. Instead, use probability distributions dexponential(), dbernoulli(), dbinomial(), and dpoisson() within option likelihood() to fit the corresponding distribution to the data.

b. bayesmh, when used on replay, did not recalculate the deviation information criterion (DIC) statistics and stored missing values in e(dic). As a result, bayesstats ic reported missing DIC values. This has been fixed.

7. dstdize with option using(), when the using dataset did not explicitly contain the .dta extension and using() was specified as a relative path that contained a dot, exited with error message "file ... not found". This has been fixed.

8. duplicates drop, when specified with a varlist that included a variable that did not exist in the dataset, incorrectly exited with error message "varlist not allowed". duplicates drop now exits with error message "variable ... not found".

9. eteffects with pweights incorrectly exited with error message "pweights not allowed". This has been fixed.

10. fracreg with option vce(cluster clustvar), when clustvar was included in an if or in condition, reported the wrong number of clusters and incorrect standard errors. This has been fixed.

11. heckprobit with option vce(opg) used the default observed information matrix (OIM) standard errors instead of the requested outer product of the gradient vectors (OPG) standard errors. This has been fixed.

12. icd9 search and icd9p search, when set varabbrev had been set to off, incorrectly exited with error message "__desc not found". This has been fixed.

13. irt grm, irt nrm, irt pcm, irt gpcm, and irt rsm incorrectly accepted items with noninteger values. The estimates from these irt commands were still correct. However, the inclusion of noninteger values caused irtgraph commands to exit with an error. These irt commands now exit with error message "item ... contains noninteger values".

14. table, when option contents() contained a string variable, exited with an uninformative error message. table now exits with error message "'...' found where numeric variable expected".

15. tssmooth dexponential, despite reporting the correct statistics, did not save the correct root mean squared error in r(rmse) or the correct sum of squared residuals in r(rss). This has been fixed.

16. xtdpd with option vce(robust), in the unusual case in which the model included no lags of the dependent variable and the dependent variable had a missing value, incorrectly exited with error message "estimates post: matrix has missing values". This has been fixed.

17. xtlogit, xtprobit, xtcloglog, and xtpoisson, normal with default option re have the following fixes:

a. xtlogit, xtprobit, xtcloglog, and xtpoisson, when option vce(oim) was specified instead of being left as the unspecified default, incorrectly reported robust standard errors instead of the requested OIM standard errors. This has been fixed.

b. xtlogit, xtprobit, xtcloglog, and xtpoisson, when constraints were specified with option noconstant and either option vce(robust) or option vce(cluster clustvar), did not apply the constraints when calculating robust standard errors. This has been fixed.

18. (Windows) estimates save did not open files with UNC paths (network paths beginning with \\) and returned error message "file filename could not be opened". This has been fixed.

19. (Windows) estimates use did not open files with UNC paths (network paths beginning with \\) and returned error message "file filename not found". This has been fixed.

-------- update 03mar2016 -----------------------------------------------------

1. icd10 check, icd10 lookup, and icd10 generate with option description have a new default year of 2016, which corresponds to the World Health Organization's Version 5. Results using previous ICD-10 codes may be obtained by specifying option year(#). Any year value from 2003 to 2016 is allowed.

2. Stata/MP performance has improved, often dramatically, for the following commands: gsem, irt, meglm, melogit, meprobit, mecloglog, meologit, meoprobit, mepoisson, menbreg, mestreg, xtologit, xtoprobit, and xtstreg.

3. cc, cs, and ir now store the reported crude and pooled estimates from a stratified analysis, along with the upper and lower bounds from the confidence intervals. The new stored results are r(crude), r(lb_crude), r(ub_crude), r(pooled), r(lb_pooled), and r(ub_pooled).

4. churdle with a string variable specified in option vce(cluster clustvar) reported missing standard errors. This has been fixed.

5. correlate, when varlist contained a string variable, incorrectly exited with error message "no observations". This has been fixed.

6. egen with function count(exp) incorrectly exited with error message "type mismatch" when exp contained a string variable. This has been fixed.

7. fracreg with a string variable specified in option vce(cluster clustvar) reported missing standard errors. This has been fixed.

8. estat overid, used after fitting a model with ivregress that contained factor variables in the set of endogenous variables and with option vce(robust) specified, incorrectly exited with error message "depvar may not be a factor variable". This has been fixed.

9. graph export, when exporting as a PDF file a pie graph that had the angle of the first slice set to a value other than the default 90 degrees, displayed the slice in wrong position in the PDF file. This has been fixed.

10. import delimited incorrectly imported string values of the form #f or #F as numeric values. For example, 12345F was imported as 12345 instead of as a string containing 12345F. This has been fixed.

11. include has new option system, which searches Stata's system directories for the specified file if the file is not found at the default location.

12. loneway with aweights, when the response variable or group variable contained missing values or when qualifier if or in was specified, could report incorrect estimates of the intraclass correlation and the standard deviation of the group effect. If an entire group comprised observations to be omitted, the results were affected. In other cases, whether incorrect results were reported depended on an internal sort. This has been fixed.

13. mgarch ccc, mgarch dcc, and mgarch vcc have the following fixes:

a. mgarch ccc, mgarch dcc, and mgarch vcc, when lagged variables were specified as indepvars or in option het(varlist), stored an incorrect minimum time in sample in e(tmin) and an incorrect formatted minimum time in e(tmins). This has been fixed.

b. mgarch ccc, mgarch dcc, and mgarch vcc, when used after tsappend, stored an incorrect maximum time in sample in e(tmax) and an incorrect formatted maximum time in e(tmaxs). This has been fixed.

14. plugin function SF_mat_store(), when changing the value of element [i,j] of a symmetric matrix, also changed the value of element [j,i]. This has been fixed for plugins written using version 3.0 of the plugin interface. Plugins written using version 2.0 of the plugin interface will continue to work as they did before.

15. post no longer requires the maximum number of bytes to hold an observation to be less than 32,767.

16. predict with option dynamic(), when the data were tsset using a business calendar and used to make predictions after dfactor, mgarch, mswitch, and sspace, incorrectly exited with error message "invalid syntax". This has been fixed.

17. predict after pca is documented to accept stub* notation when no statistic is specified. Instead, predict exited with error message "invalid name". The command now works as documented.

18. tsset and xtset, in the unusual case when there was only one observation in the dataset, incorrectly exited with error message "observation numbers out of range". This has been fixed.

19. unicode translate could corrupt datasets that could not be opened in the current session. For example, a dataset that could not be opened because it had more variables than the current maxvar setting may have become corrupted. unicode translate now ignores files with the .dta extension that cannot be opened in the current session.

20. xtreg, xtivreg, xthtaylor, xtlogit, xtprobit, xtcloglog, xtologit, xtoprobit, xtpoisson, xtstreg, and gmm, when specified with qualifier if or in and with option vce(cluster clustvar) and when the panels were not nested in clusters in the full data but were nested in clusters after applying the qualifier, incorrectly exited with error message "panels are not nested within clusters". This has been fixed.

21. xttobit, when specified without an independent variable, incorrectly exited with error message "invalid syntax". This has been fixed.

22. zinb, with a factor-variables operator on the first variable in option inflate() and option vce(bootstrap) or vce(jackknife), incorrectly exited with error message "depvar may not be a factor variable". This has been fixed.

23. The Data Editor did not allow variables that were formatted using a business calendar format (%tbcalname) to be edited with values that also had a business calendar format. This has been fixed.

24. (Windows only) graph export, when used to export a graph to EPS or PS format, could cause Stata to crash if the fontface setting for EPS or PS was changed from the default value of Helvetica. This has been fixed.

25. (Mac and Unix) Windows-only setting set autotabgraphs no longer returns error message "unrecognized command" when used on Mac and Unix platforms. Instead, a note is displayed. This prevents do-files written by users with Stata for Windows from exiting with an unnecessary error message when shared with users with Stata for Mac or Stata for Unix.

-------- update 16feb2016 -----------------------------------------------------

1. bayesmh has new likelihood() specifications dexponential(), dbernoulli(), dbinomial(), and dpoisson() that allow you to more conveniently fit exponential, Bernoulli, binomial, and Poisson distributions to data. Previously, these distributions could be fit using likelihood() specifications exponential, binlogit, and poisson with suboption noglmtransform.

2. bayesmh has new likelihood() specification binomial() that is a synonym for the existing likelihood() specification binlogit() for fitting a binomial regression with a logit link.

3. bayestest interval, when used with a probability label longer than 31 characters, incorrectly displayed the label in the legend and exited with an uninformative error message. This has been fixed.

4. cii means with option poisson did not allow noninteger exposure values. This has been fixed.

5. gsem, meglm, melogit, meprobit, meologit, meoprobit, mepoisson, menbreg, and mestreg, when used to fit a model with three or more levels and with option vce(cluster clustvar), did not exit with error message "highest-level groups are not nested with clustvar" when the groups defined by the highest level of the model specification were not nested within the clusters defined by clustvar. This has been fixed.

6. gsem, meglm, melogit, meprobit, meologit, meoprobit, mepoisson, menbreg, and mestreg, when used to fit a model with three or more levels and with prefix svy, did not exit with error message "hierarchical groups are not nested within ..." when the groups defined by the multilevel model specification were not nested within the svyset sampling units. This has been fixed.

7. mprobit, when specified without independent variables, incorrectly exited with error message "varlist required". This has been fixed.

8. predict, median, after mestreg was used to fit a model with family weibull, exponential, lognormal, or loglogistic, ignored the random-effects portion of the model when generating the predictions. This has been fixed.

-------- update 01feb2016 -----------------------------------------------------

1. (Mac) After the 28jan2016 update, the Data Editor's toolbar and menu were disabled. This has been fixed.

-------- update 28jan2016 -----------------------------------------------------

1. bayesmh, when initial values for scalar model parameters were specified in options likelihood() and prior() during parameter declaration, incorrectly exited with a syntax error. This has been fixed.

2. estat residuals and estat teffects, when used after sem with option technique(bhhh), incorrectly exited with an uninformative error message. This has been fixed.

3. estat sbsingle, for certain combinations of the number of observations in the dataset and the number of covariates in the model, incorrectly exited with error message "insufficient observations at the specified trim level" when there were enough observations to compute the test statistic. This has been fixed.

4. estat teffects after sem did not show indirect effects when they were composed of more than six paths between the variables. This has been fixed.

5. Factor variables with an fvset base of first, frequent, or last incorrectly resulted in all single-level specifications being treated as a base level. For example, for the 0/1-valued variable foreign, typing

. sysuse auto . fvset base last foreign . regress mpg 1.foreign

resulted in 1.foreign being treated as the base level and omitted from the model. Similarly, single-level specifications with the base-level operators b(first), b(frequent), or b(last) were also treated as a base level. The correct behavior is for single-level specifications to ignore the base-level setting or specification. This has been fixed.

6. Stata reported an uninformative error message when factor-variable interaction notation (# or ##) was used with two different time-series operators applied to the same factor variable within the interaction term. This has been fixed.

7. graph export has the following fixes:

a. graph export, when exporting a graph that contained text with a font size of 0 to EPS format, produced an EPS file that was unreadable by other applications. When graph export encounters text of size 0, it no longer exports that text when writing to an EPS file.

b. graph export, when exporting an SEM or a GSEM path diagram that included a B├ęzier Curve to a PDF file, incorrectly exited with error message "unable to save PDF file". This has been fixed.

8. icc now allows the target and rater variables to have a string storage type.

9. After the 29oct2015 update, ivprobit with option vce(cluster clustvar) and pweights incorrectly exited with error message "option vce(cluster) not allowed with probability weights". This has been fixed.

10. ivregress 2sls with options first and vce(cluster clustvar), when the data were not sorted by clustvar, reported incorrect standard errors and an incorrect number of clusters for the first-stage model results only. This was a reporting error, and calculations for the full model were unaffected. This has been fixed.

11. margins has the following fixes:

a. margins with option predict(fixedonly ...), in the unusual case when predictions were made after one of the multilevel commands, gsem, meglm, melogit, meprobit, meologit, meoprobit, mepoisson, menbreg, or mestreg, when the fitted model had a very small estimated value for a variance component (less than 1e-30), incorrectly exited with error message "could not calculate numerical derivatives -- discontinuous region with missing values encountered" even when the prediction was not a function of the variance component. This has been fixed.

b. margins, when set varabbrev off was in effect, incorrectly allowed variable abbreviations within option at(). This has been fixed.

c. margins after svy: logit, svy: logistic, or svy: probit, when a subpopulation had been specified in option subpop() and with perfect predictors and when using bootstrap or jackknife variance estimates, incorrectly reported some marginal predictions as not estimable. This has been fixed.

12. margins after xtprobit, xtlogit, and xtcloglog has the following fixes:

a. margins with option at(atspec) incorrectly computed margins at mean values even when other values were requested for atspec. This has been fixed.

b. margins with option atmeans incorrectly exited with error message "insufficient observations". This has been fixed.

13. Mata functions _docx_add_data(), _docx_add_mata(), and _docx_add_matrix() inserted an empty line before the content in each cell when creating a table in a Microsoft .docx file. This line is no longer inserted. The old behavior is not preserved under version control.

14. Mata function _docx_table_mod_cell_table(), when argument append was specified as a value other than 0, incorrectly replaced the current contents of the cell instead of appending to the contents as requested. This has been fixed.

15. The multilevel mixed-effects estimators gsem, meglm, melogit, meprobit, meologit, meoprobit, mepoisson, menbreg, and mestreg have the following improvement and fix:

a. gsem, meglm, melogit, meprobit, meologit, meoprobit, mepoisson, menbreg, and mestreg are now more likely to converge for models fit to data with large group sizes.

b. gsem, meglm, melogit, meprobit, meologit, meoprobit, mepoisson, menbreg, and mestreg, when used with datasets that had many observations per group, could indicate that the model converged and report results that did not include the large groups in the computations. To determine whether your model was affected, you can use predict to obtain empirical Bayes mean estimates for the random effects and check for zero values in large groups. This has been fixed.

16. Plugin functions SF_var_is_string() and SF_var_is_strl() returned results for the ith variable of the dataset rather than the ith variable of the varlist. This has been fixed.

17. predict, when used to compute predictions that require empirical Bayes estimates after gsem, could exit with an uninformative error message if an interaction term in the call to gsem was specified with a single level (or strict subset of the levels) of a factor variable. This has been fixed.

18. predict after regress, tobit, intreg, ivregress, ivtobit, xttobit, xtintreg, truncreg, heckman, and nl has the following improvement and fix:

a. predict, pr(a,b), predict, e(a,b), and predict, ystar(a,b) now allow a and b to be expressions. This is useful when working with a transformed dependent variable because you can specify the same transform in the prediction. For example, to predict the probability that a log-transformed dependent variable is less than 90, you can now specify pr(.,log(90)) rather than pr(.,4.4998).

b. predict, pr(a,b), predict, e(a,b), and predict, ystar(a,b) incorrectly allowed b to be less than a. This has been fixed.

19. predict with option reffects after mixed, meqrlogit, and meqrpoisson may now be specified with new option reses(). This option allows you to request predictions of the standard errors of the random effects and predictions of the random effects themselves at the same time. The old syntax (predict, reses) that required a separate command continues to work but is no longer documented.

20. sem with option ginvariant(all) and user-specified zero-valued constraints on error variances incorrectly exited with error message "inconsistent covariance setting for ...; one of these variable's underlying variance is constrained to zero". This has been fixed.

21. sts list with option strata() or option by() and with options failure and at() produced a value of 1 instead of 0 for the failure function when the time values specified in option at() preceded the earliest failure time in the data. This has been fixed.

22. var with option constraints(), when set showbaselevels on was specified, incorrectly exited with error message "option baselevels not found". This has been fixed.

23. After the 29oct2015 update, xtlogit and xtprobit with option vce(robust) or option vce(cluster panelvar), when perfect predictors caused a reduction in the sample size, reported incorrect standard errors. This has been fixed.

24. xtreg with option vce(cluster clustvar), in the unusual case when clustvar was also specified as a continuous covariate, could report an incorrect number of clusters and could produce incorrect standard errors. When the correct number of clusters was reported, the standard errors were correct even if clustvar was included as a continuous covariate. This has been fixed.

25. (Windows) export excel did not save Excel files with the .xls extension when the filename contained a Unicode character beyond the plain ASCII range. This has been fixed.

26. (Windows) Extended macro function dir files with a pattern specified, when the pattern contained an uppercase letter and when option respectcase was not specified, did not store the names of matched filenames in the requested macro. This has been fixed.

27. (Mac) The import delimited dialog box, when the name of the file to be imported contained a Unicode character, submitted an incorrect filename in the command it issued to Stata. This has been fixed.

28. (Mac) In the update dialog box, clicking on the link for manual updates did not open the "Keeping Stata 14 up to date" webpage. This has been fixed.

-------- update 21dec2015 -----------------------------------------------------

1. Online help and the search index have been brought up to date for Stata Journal 15(4).

2. bayesmh, when weights were specified using an expression, incorrectly exited with error message "variable not found". This has been fixed.

3. estat icc now allows robust or cluster-robust standard errors to be specified with the initial mixed or melogit command.

4. fp generate, when qualifier if was specified with an expression containing quotation marks (""), incorrectly exited with error message "invalid expression". This has been fixed.

5. putexcel set has new option modify, which changes the contents of an existing workbook.

After the 29oct2015 update, the default behavior of putexcel set was changed to modify an existing workbook or to create the workbook if it does not exist. Option modify is intended to prevent accidental overwriting of existing workbooks.

6. After the 29oct2015 update, putexcel, when outputting string data enclosed with compound double quotes (`"` and '"'), incorrectly exited with error message "` invalid name". This has been fixed.

7. sem, when used to fit a model with more than 256 observed variables, exited with an uninformative error. If the number of model parameters (coefficients, means, variances, and covariances) is less than or equal to 32740, then sem now fits the model as specified. If the number of parameters exceeds 32740, then sem exits with error message "matsize too small".

8. streg has the following fixes:

a. streg, in the rare case when the full model converged but the constant-only model did not, incorrectly reported a likelihood-ratio model test instead of a Wald model test. This has been fixed.

b. streg, when used to fit a Weibull model with a constraint on the constant and when option ancillary(), strata(), offset(), frailty(), or shared() was not specified, applied the specified constraint incorrectly. This has been fixed.

c. streg has improved starting values for Weibull models with a constraint on the shape parameter when option ancillary(), strata(), offset(), frailty(), or shared() is not specified.

d. streg, when used to fit a Weibull model with only a constant term in the main and ancillary equations, when a constraint was applied to the constant in the ancillary equation, and when option ancillary(), strata(), offset(), frailty(), or shared() was not specified, applied the specified constraint incorrectly. This has been fixed.

e. streg, when used to fit a Weibull model in the accelerated failure-time (AFT) metric with constraints and when option ancillary(), strata(), offset(), frailty(), or shared() was not specified, applied the constraints in the proportional-hazards metric instead of the AFT metric. This has been fixed.

f. streg, when used to fit an exponential model in the accelerated failure-time (AFT) metric with constraints, applied the constraints in the proportional-hazards metric instead of the AFT metric. This has been fixed.

g. streg with option offset(), when used to fit an exponential model in the accelerated failure-time (AFT) metric, constrained the parameter of the offset variable to one in the proportional-hazards metric instead of the AFT metric. This has been fixed.

h. streg with option frailty(), option shared(), or both, when used to fit an exponential model in the accelerated failure-time (AFT) metric, produced covariances between the main equation parameters and the frailty parameter that had the wrong sign. This has been fixed.

9. predict, with option scores after using streg to fit an exponential model in the accelerated failure-time (AFT) metric, and when streg was specified with option frailty(), option shared(), or both, produced frailty parameter scores that had the wrong sign. This has been fixed.

-------- update 01dec2015 -----------------------------------------------------

1. putexcel has two new options to write Stata dates (%td-formatted values) or datetimes (%tc-formatted values) to Excel as unformatted numbers.

a. Option asdatenum converts the specified Stata date to an Excel date and writes the result as an unformatted number. asdatenum is useful if you are adding data to cells that already have a date format and you do not want Stata to apply the default Excel date format, m/d/yyyy.

b. Option asdatetimenum converts the specified Stata datetime to an Excel datetime and writes the result as an unformatted number. asdatetime is useful if you are adding data to cells that already have a datetime format and you do not want Stata to apply the default Excel datetime format, m/d/yyyy h:mm.

2. bayesmh has the following fixes:

a. bayesmh, when option noshow() was specified for one or more model parameters, incorrectly treated the parameters as excluded in the calculation of the Laplace-Metropolis estimator of the log marginal likelihood instead of using the formula documented in the manual. This has been fixed.

b. bayesmh, when option exclude() was specified for one or more model parameters, incorrectly attempted to calculate the Laplace-Metropolis estimator of the log marginal likelihood. Because the Laplace-Metropolis estimator requires simulation results for all parameters in the model, bayesmh now reports a missing value for the log marginal likelihood if exclude() is specified.

3. bayesstats ic has new option diconly to report only the deviance information criterion (DIC).

4. estat report, irtgraph icc, irtgraph tcc, irtgraph iif, and irtgraph tif, after an IRT model was fit with prefix svy, incorrectly exited with error message "invalid subcommand report". This has been fixed.

5. forecast solve with option simulate(residuals, ...) now displays a progress log when computing the forecast residuals. The log can be suppressed with option log(off).

6. gsem and sem, when set varabbrev had been set to off and when _OEx, _LEx, _Ex, e._OEn, e._LEn, or e._En was specified in option covariance() or option covstructure(), incorrectly exited with error message "latent variable ... not found". This has been fixed.

7. memory, when used to check memory allocation in 64-bit Stata/MP, always reported that 4 GB of memory was used for overhead, even when the actual amount was far less. This has been fixed.

8. nestreg, with prefix svy and missing values in any of the survey design variables, incorrectly exited with error message "# obs. dropped because of estimability, this violates the nested model assumption". This has been fixed.

9. predict with option latent, after using gsem to fit a model with two or more latent variables that were each predicted by observed variables, incorrectly exited with error message "variable _cons not found". This has been fixed.

10. predict, when used to compute marginal predictions after fitting a weighted multilevel model with gsem, meglm, melogit, meprobit, meologit, meoprobit, mepoisson, or menbreg, and in rare cases when group variables were changed to break the original nesting structure, incorrectly exited with error message "weights not allowed with crossed models". This has been fixed.

Estimation results saved to disk prior to this fix will continue to exhibit this unwanted behavior.

11. putexcel, after the 29oct2015 update and with option asdate or option asdatetime, wrote the specified number as a string instead of the correct Excel date value. This has been fixed.

12. putexcel set, after the 29oct2015 update and with option sheet(sheetname, replace), did not clear the specified worksheet. This has been fixed.

13. sem has the following fixes:

a. sem, when specified with an empty weight expression, incorrectly exited with error message "'[]' found where '(' expected". This has been fixed.

b. sem has improved starting values for fitting models with unstructured error covariances when the data include few observations relative to the number of freely estimated parameters.

14. svy: tabulate, when value labels for a tabulated variable contained quotation marks, exited with error message "conformability error". This has been fixed.

15. unzipfile and zipfile, when the name of the file contained the word "if", "in", or "using", exited with error message "... not allowed". This has been fixed.

16. (Windows) Extended macro function dir is documented to be case-insensitive by default. The function was instead case-sensitive. It now works as documented.

17. (Windows) The Command window, when high-contrast colors are enabled in Windows, now uses the colors set by the operating system.

18. (Windows) Toolbars, when high-contrast colors are enabled in Windows, now use an appropriate background color.

19. (Windows) The Do-file Editor has the following improvements:

a. The Do-file Editor now allows you to set the cursor color. To set your cursor color preferences, choose Edit > Preferences... from the menu, and then select "Cursor color" on the Editor font tab.

b. Line numbers in the Do-file Editor now use the default text color from the operating system.

20. (Unix) graph export, when exporting to PDF and when the graph included an extended-ASCII mathematical character, could crash Stata. This has been fixed.

-------- update 29oct2015 (Stata version 14.1) --------------------------------

1. Bayesian analysis (bayesmh) has many improvements for multilevel models, panel-data models, nonlinear models, and models with latent variables:

a. bayesmh has new option reffects() for fitting univariate two-level models. Specifying reffects() implements a more computationally efficient Metropolis-Hastings algorithm for sampling random-effects parameters.

b. bayesmh with option block() has new suboption reffects. Parameters grouped in a block with suboption reffects are treated as random-effects parameters and are simulated using the new optimized version of the Metropolis-Hastings algorithm.

c. bayesmh has new option xbdefine() for defining linear combinations. This provides a more convenient and memory-efficient way of using linear combinations in substitutable expressions. This option is used when fitting nonlinear models.

d. bayesmh has new option redefine() for defining linear forms associated with random-effects variables. This option may be used to write expressions in likelihood specifications and speed up calculations, especially for nonlinear models that include random effects.

e. Option prior()'s suboptions mvnormal() and mvnormal0() provide a convenient way of modeling correlation between random-effects parameters.

mvnormal() and mvnormal0() prior distributions of dimension d can now be used with lists of m x d parameters for any positive integer m. In such cases, the list of parameters is reshaped into a matrix with d columns, and its rows are assumed to be independent samples from the specified multivariate normal distribution. See example 25 in [BAYES] bayesmh for application of this feature.

2. ci and cii have the following improvements and changes:

a. With new commands ci variances and cii variances, you can now compute confidence intervals for variances or standard deviations. Classical confidence intervals for normal data and Bonett confidence intervals for nonnormal data are provided.

b. ci means and cii means are now used to compute confidence intervals for means of normally and Poisson distributed variates. The previous commands were ci and cii. The old syntax is preserved under version control.

c. ci proportions and cii proportions are now used to compute confidence intervals for proportions. The previous commands were ci with option binomial and cii with option binomial. The old syntax is preserved under version control.

3. New command diflogistic performs the logistic regression test for uniform and nonuniform differential item functioning (DIF) in the context of item response theory (IRT) analysis. Tests for DIF detect whether an item performs differently for individuals with the same ability who come from different groups.

4. difmh has the following improvements:

a. difmh has new option maxp(#) that causes difmh to display results only for items with p < #.

b. difmh has new option sformat() that changes the display format of chi-squared values.

c. difmh has new option pformat() that changes the display format of p-values.

d. difmh has new option oformat() that changes the display format of odds-ratio statistics.

e. difmh can now replay results, and on replay the user can specify reporting options maxp(), sformat(), pformat(), oformat(), and level().

5. ivprobit with two or more endogenous variables now converges more reliably and presents easier to interpret output for the ancillary parameters. Their names in e(b) are now athrho and lnsigma with index suffixes. The old behavior is preserved under version control.

6. ivprobit has two new postestimation commands:

a. estat covariance displays the covariance matrix of the errors.

b. estat correlation displays the correlation matrix of the errors.

7. predict with option pr after ivprobit has been improved to compute probability estimates that account for endogeneity. The old behavior is preserved under version control.

This means that the results of margins, estat classification, lroc, and lsens after ivprobit also have been improved to account for endogeneity.

8. predict with options pr(), e(), and ystar() after ivtobit have been improved to compute probabilities, truncated expectations, and censored expectations that account for endogeneity. The old behavior is preserved under version control.

This means that the results of margins also have been improved to account for endogeneity.

9. You can now add your own imputation methods to mi impute by writing an ado-file program that follows simple conventions. You use the program to impute your variables once, and then use it with mi impute to obtain multiple imputations. mi impute will properly create imputations according to the chosen mi style. All of mi impute's global options, except by() and noupdate, are supported. See User-defined imputation methods for details.

10. contrast after mixed ..., dfmethod() has new option small to perform small-sample inference for contrasts of fixed effects. This allows you to request that degree-of-freedom adjustments, such as Kenward-Roger and Satterthwaite, be applied to tests for individual contrasts and tests of main effects, simple effects, and interactions.

11. pwcompare after mixed ..., dfmethod() has new option small to perform small-sample inference for pairwise comparisons of fixed effects. This allows you to request that degree-of-freedom adjustments, such as Kenward-Roger and Satterthwaite, be applied to tests for pairwise comparisons.

12. mlexp has three new options:

a. Option constraints() allows linear constraints to be specified.

b. Option collinear allows mlexp to retain collinear variables.

c. Option debug may be specified with option derivative() to report the difference between the numerical derivatives and analytical derivatives. This option is useful for determining if there are problems with user-provided derivatives.

13. predict after mlexp now takes option xb. This lets you obtain the linear prediction after maximum likelihood estimation of user-specified expressions.

14. margins is now available after mlexp. This lets you estimate marginal and conditional effects such as means, predictive margins, and marginal effects after maximum likelihood estimation of user-specified expressions.

15. teffects psmatch is now much faster for large datasets.

16. xtprobit, re, xtlogit, re, and xtcloglog, re now allow predictions of marginal probabilities. The marginal probability is the probability of a positive outcome that is marginal with respect to the random effect. This is now the default statistic for margins after these commands. The old default behavior is preserved under version control.

17. xtpoisson, re with option normal now allows prediction of the number of events to be marginal with respect to the random effect. This prediction is the new default statistic for margins after xtpoisson, re. The old default behavior is preserved under version control.

18. xtologit and xtoprobit now allow predictions of marginal probabilities. The marginal probability is the probability of the specified outcome that is marginal with respect to the random effect. The marginal probability for each outcome is the new default statistic for margins after these commands. The old default behavior is preserved under version control.

19. xtstreg now allows predictions of the marginal mean survival time, marginal hazard, and marginal survivor function. The mean survival time, hazard, and survivor function are marginal with respect to the random effect. The marginal mean survival time is the new default statistic for margins after xtstreg. The old default behavior is preserved under version control.

20. stcurve now defaults to marginal predictions after xtstreg. The old default behavior is preserved under version control.

21. icd10 now works with ICD-10 codes from many more years. Any year from 2003 to 2015 is allowed. This year range covers the full history of ICD-10 back to version 2. Type icd10 query for complete information.

22. icd10 check, icd10 lookup, and icd10 generate with option description default to using the most current ICD-10 codes in Stata if year() is not specified. The current default is now year(2015).

23. putexcel has a greatly simplified syntax and a more intuitive dialog.

24. Plugins now support strLs (long strings). See the webpage Creating and using Stata plugins.

25. Stata has improved matrix stripe logic that now allows matrices to be combined with a mixture of factor-variable operators, including contrast operators and the @ operator. Because matrix stripe behavior is affected, the old behavior is preserved under version control.

26. eteffects, teffects, and stteffects no longer allow the r operator in references to coefficients in the POmean equation. The old behavior of requiring the r operator is preserved under version control.

27. bayesmh's model summary of specified priors, when the prior included substitutable expressions that were not enclosed in parentheses, did not display the argument to the parameter distribution. For example, {mpg:_cons} ~ exponential(sqrt(_pi)) was displayed as {mpg:_cons} ~ exponential. This has been fixed.

28. collapse has the following fixes:

a. collapse with option percent and missing values in a by() variable used the wrong denominator to calculate the percentages. This has been fixed.

b. collapse with option percent and iweights used the unweighted total instead of the weighted total to calculate the percentages. This has been fixed.

c. collapse with option percent and aweights ignored the specified weights. This has been fixed.

29. etregress with vce(cluster clustvar) exited with an uninformative error message when clustvar had missing values. This has been fixed.

30. gllamm in Stata/MP with link(ologit) or link(oprobit) exited with error message "(error occurred in ML computation)" even when the model was correctly specified. This has been fixed.

31. icd9 and icd9p have the following fixes:

a. icd9 check and icd9p check, for a small number of category and subcategory codes, incorrectly indicated that the codes were not defined. This has been fixed. Note that no billable ICD-9-CM codes were affected.

b. icd9 generate and icd9p generate with option description returned a blank description for a small number of codes instead of indicating that these codes were further subdivided. This has been fixed.

c. icd9 generate with option description for the codes 176, 764, 765, V29, and V69 returned unofficial descriptions created by StataCorp that included the text "(Stata)". The description for these codes is now "*", which is the documented indicator that a code has been subdivided into billable codes.

d. icd9 lookup and icd9p lookup, when looking up a small number of category codes and subcategory codes without specifying a range, would incorrectly return "(no matches found)". For example, typing icd9 lookup 040.4 would trigger this note, but icd9 lookup 040.4* would not. This has been fixed.

32. import delimited now uses a comma instead of a tab character (\t) as the default delimiter if it is unable to automatically determine the delimiter used in the file. This behavior is more consistent with import delimited's use of .csv as the default file extension. The old behavior is preserved under version control.

33. import delimited with option varnames(), with files having more than 5,000 rows and where the data type for a variable switched between numeric and string after row 5,000, resulted in variable names being imported from the wrong line of the file. This has been fixed.

34. predict with stub* and option pr, after using ologit or oprobit to fit a model without covariates, incorrectly exited with error message "varname required". This has been fixed.

35. predict after xtintreg and xttobit and with option pr0(), option e0(), or option ystar0() calculated marginal predictions, typically more useful than predictions that assume a zero-valued random effect. However, the documentation stated that predictions assumed random effects were 0. The documentation has been updated to reflect the behavior of the command and, for clarity, the options have been renamed pr(), e(), and ystar(). The old syntax is preserved under version control.

36. rnormal(), despite producing random-normal variates with all the desirable statistical properties of the documented algorithm, on rare occasions did not exactly implement the documented algorithm. This has been fixed.

The sequence of random-normal variates can differ in rare cases from those that were previously drawn when starting from the same seed. Commands dependent on random-normal variates, such as bayesmh and mi impute mvn could produce different results. The previous behavior is preserved under user-version control.

37. xtlogit and xtprobit with option vce(robust) or option vce(cluster panelvar), when perfect predictors caused a reduction in the sample size, incorrectly exited with error message "calculation of robust standard errors failed". This has been fixed.

38. Contrary to the documentation, xtpoisson, re with option normal did not accept time-series operators in depvar or indepvars. This has been fixed.

39. xtregar has the following fixes:

a. xtregar with option re, in the rare case of a negative autocorrelation coefficient, produced standard errors that were too large. This has been fixed.

b. xtregar with option fe, in the rare case of a negative autocorrelation coefficient, produced standard errors that were too large and a point estimate for the constant term that was biased toward 0. This has been fixed.

40. (Windows) Icons are now rendered more clearly when the DPI setting in Windows is above 100%.

41. (Windows) Dialog boxes for import delimited and power now use proper scaling when the DPI setting in Windows is above 100%.

42. (Windows) graph use did not open files with UNC paths (network paths beginning with \\) and returned error message "file filename not found". This has been fixed.

43. (Windows) The Graph Recorder, when a recording was selected after choosing Recorder and then Play from the Tools menu or after clicking on the Play recording button in the Standard Toolbar, did not play the selected recording and put Stata in a state where no commands could be submitted. This has been fixed.

44. (Mac) view browse did not open https URI schemes or URLs that were enclosed in quotes. This has been fixed.

45. (Mac) Selecting Paste special... from the Edit menu and then clicking on OK in the resulting dialog caused Stata to crash. This has been fixed.

46. (Mac GUI) The System Integrity Protection feature of OS X 10.11 (El Capitan) prevents users from making modifications to the /usr/bin directory. Because Stata can no longer write to the /usr/bin directory, Stata now creates a symbolic link to the console version of Stata/MP or Stata/SE in the /usr/local/bin directory if the user selects "Install Terminal Utility..." from the Apple menu. Stata will create the /usr/local/bin directory if it does not exist.

-------- update 07oct2015 -----------------------------------------------------

1. Online help and the search index have been brought up to date for Stata Journal 15(3).

2. egen cut(varname), group(#), when # > 1000 and when the data were not previously sorted by varname, did not produce # groups. This has been fixed.

3. estat phtest with option detail, when used after stcox with option vce(robust) and with multiple-record data, used robust standard errors instead of the appropriate cluster-robust standard errors when computing rho statistics and p-values for tests of the proportional-hazards assumption for individual covariates. This has been fixed.

4. eteffects has the following fixes and improvement:

a. eteffects after the 3 Aug 2015 update, when there were missing predicted values for the treatment probability, produced incorrect estimates. This has been fixed.

b. eteffects attempted to fit a model when the treatment variable had a negative value and returned an uninformative error message. The command now returns error message "treatment variable tvar must be positive".

c. eteffects now allows the treatment variable to identify the treatment levels using any two nonnegative integers. Previously, the treatment levels were required to be coded 0 and 1. By default, eteffects assumes that the lower value of the treatment variable is the control.

5. forecast solve with option simulate(residuals) returned a Mata error message when used to define an identity before estimates were defined or before a coefficient vector was added. This has been fixed.

6. ivtobit with models having more than one endogenous variable specified is now more likely to converge.

7. margins with options vce(unconditional) and expression(), if the expression was a function of more than one prediction, exited with error message "varlist required". This has been fixed.

8. pctile newvar = exp, nquantiles(#), when # > 1000 and when the data were not previously sorted by exp, correctly computed the quantile values but did not store the values in the first #-1 observations of the dataset. This has been fixed.

9. predict with option scaledsch, when used after stcox with option vce(robust) and with multiple-record data, used robust standard errors instead of the appropriate cluster-robust standard errors when computing scaled Schoenfeld residuals. This has been fixed.

10. teffects nnmatch, in the unusual case that fweights were specified with option vce(robust), produced standard errors that were too small. This has been fixed.

-------- update 22sep2015 -----------------------------------------------------

1. Stata has improved execution for operations using the https:// protocol and when using a proxy server. Stata now sends a CONNECT request to the proxy server, which is the standard Java implementation and is the preferred method for secure sockets layer (SSL) connections.

2. eteffects has improved starting values with outcome model fractional and with option atet, meaning that these models now converge faster. The estimated average treatment effect on the treated (ATET) and associated standard error are unchanged.

3. etregress with option cfunction has an improved method for computing the moment conditions and gradient. Most users will notice a change only in the values displayed in the iteration log. However, in rare cases, some complex models that could not be fit before will now converge.

4. contrast with the observation-weighted orthogonal polynomial operators pw(numlist). and qw(numlist). did not correctly interpret the numlist as the degree of the polynomial. When either pw. or qw. was specified, numlist was interpreted as the value of the factor variable. This has been fixed.

5. estat df with option method(repeated) reported incorrect degrees of freedom when observations in the dataset were not grouped by the subject variable. This has been fixed.

Note that results reported by mixed with option dfmethod(repeated) were not affected.

6. estimates table with multiple prais estimation results ignored all but the first coefficient. This has been fixed.

7. generate and replace, when the right-hand-side expression included references to elements of Stata matrices, did not benefit from parallelization. These commands now exhibit nearly perfect parallelization when the matrix subscripts are specified as numeric scalars.

8. graph export, when used to export to EPS format a graph that had a font with glyphs of more than 2,000 points, could cause Stata to crash. This has been fixed.

9. gsem with prefix svy and when the specified model included a multinomial outcome and did not include any latent variables incorrectly exited with an error message. This has been fixed.

10. Mata functions lnmvnormalden(), lnwishartden(), and lniwishartden() with an input argument that was created with st_view() returned incorrect results. This has been fixed.

11. me commands have the following fixes:

a. me commands used with a dependent variable that began with a capital letter and without a random-effects equation incorrectly exited with the error message "model not identified". This has been fixed.

b. me commands, in the unusual case when multilevel commands were used to fit a one-level model and when the model included an independent variable that began with a capital letter, fit the model as if the capitalized variable were a level-2 random effect and reported a note that indepvar was treated as a latent variable. This has been fixed.

12. prtest and prtesti, when used for a two-sample test of proportions, reported the two-sided p-value of the test with the label Pr(|Z| < |z|). The label should have been Pr(|Z| > |z|). This has been fixed.

13. pwcorr, when pairwise correlations were requested for more than seven variables, did not store results as documented. Matrix r(C) stored only a portion of the correlations, and scalars r(rho) and r(N) contained missing values. This has been fixed.

14. sem with constrained coefficients on paths from two or more latent variables to a single observed variable incorrectly exited with an error message when the model was identified. This has been fixed.

15. Postestimation command varlmar, when the dataset contained variables beginning with e# and # was less than or equal to the number of equations in the previous model, dropped those variables from the dataset. For instance, variables named e1, e2, and e1a were dropped. If the variable was also included in the previous var or svar command, varlmar also exited with error message "variable varname not found". This has been fixed.

16. (Mac) When a table with multiple rows was copied from the Results window as HTML, the HTML output would appear as a table with just one row. This has been fixed.

17. (Mac) Graph dialogs on Mac OS X 10.10, when combined with this specific sequence of steps -- 1) select a variable from a variable control; 2) switch to a tab that contains no active text edit fields; and 3) click on the OK button without making any other changes -- displayed a warning message and, in rare cases, caused Stata to crash. This has been fixed.

18. (Mac and Unix GUI) The Do-file Editor now lets you replace all text within a selection. To use this feature, select a block of text in which you want to make changes and then

a. (Mac) open the Replace dialog within the Do-file Editor. The "Replace all" button will display an arrow icon. Click on the button to display a menu that prompts you to either replace all text within the document or replace only text within the selection. Choose "Replace all in selection".

b. (Unix GUI) open the Replace dialog in the Do-file Editor, check the "Replace all in selection" checkbox, and click on the "Replace all" button.

19. (Unix GUI) graph export, when exporting to an .eps or a .ps file, no longer requires you to set font directories. Truetype fonts are now automatically mapped to font settings specified by graph set {ps|eps} fontface by searching system default font directories. Alternative search paths to find corresponding truetype fonts may be specified by using graph set {ps|eps} fontdir.

-------- update 03aug2015 -----------------------------------------------------

1. putexcel has new option locale(), which lets you specify the locale to write extended ASCII characters to a workbook.

2. ereturn post and ereturn repost have new options buildfvinfo and findomitted. These are new utilities for programmers interested in adding factor-variables support to their estimation commands.

3. ereturn repost has new option resize. This option allows programmers to repost estimation results of a different dimension than the original.

4. areg with option vce(bootstrap) or option vce(jackknife) resampled the absorption groups identified by the variable specified in option absorb() instead of resampling the observations. Because the model for areg is a linear regression with a large dummy-variable set, observations are the appropriate resampling unit. This has been fixed.

The old behavior is not preserved under version control but can be reproduced by using xtreg, fe instead of areg.

5. bayesmh with likelihood(probit) now uses a more precise method of calculating the log likelihood for zero outcomes. The previous method could produce numerically unstable results when an observation-level likelihood was close to zero.

6. eteffects produced incorrect standard errors. This has been fixed. Note that the point estimates of the treatment effects were correct.

7. gmm using vce(hac kernel opt), where the optimal number of kernel lags is used, would report a missing value for the number of kernel lags used. This has been fixed. The missing value reported is not a numerical problem, only a reporting error.

8. gsem, meologit, and meoprobit, in the rare case when depvar contained noninteger values and the command was specified with prefix svy, produced invalid linearized standard-error values. This has been fixed.

9. gsem with intercepts or fixed predictors on some but not all latent variables would fail to converge, even for models that were otherwise statistically identified. Instead, the command reported the message "(not concave)" or "(backed up)" with unchanging log-likelihood values in the iteration log. This has been fixed.

10. The documentation for gsem states that time-series operators are allowed. However, when a time-series-operated depvar and multilevel latent variables were specified, gsem exited with an uninformative error message. This has been fixed.

11. margins has the following fixes:

a. margins with weighted estimation results incorrectly reported some margins as "(not estimable)" in the results table if the margins were the result of fixing all the indepvars at a constant value and the first observation in the dataset had a zero-valued weight. This has been fixed.

b. margins, in the unusual case where xtlogit, re or xtlogit, pa was also specified with option noconstant and either option vce(bootstrap) or option vce(jackknife), incorrectly reported the error message "conformability error". This has been fixed.

12. The documentation for meologit and meoprobit states that they support time-series operators for depvar. However, when a time-series-operated depvar was specified, these commands exited with an uninformative error message. This has been fixed.

13. putexcel, when writing data or adding new formats, did not preserve existing column formats. This has been fixed.

14. Stata function regexm(s, re) could crash Stata when one or both arguments were binary strL. This has been fixed.

15. Stata function strproper(s) could crash Stata if the argument was binary strL. This has been fixed.

16. Stata function ustrlen(s) caused Stata to use more memory than was required. In general, this was not noticeable. However, in rare cases, this could cause Stata to slow down. This has been fixed.

17. svy when used to fit a regression model to data from a multistage survey design that 1) did not have a sampling unit or stratum variable in the final stage and 2) had FPC variables in all stages could exit with the error message "fpc for all observations within a stratum must be the same", depending on the sort order of the data. This has been fixed.

18. In the Results or Viewer window, highlighting text that included Unicode characters, right-clicking to open a contextual menu, and then selecting "Copy table" or "Copy table as HTML" produced an incorrectly formatted table when the copied text was pasted to a new location. This has been fixed.

19. (Windows) graph export when exporting to an .eps or a .ps file has the following fixes:

a. graph export with option fontface(fontname) ignored the specified fontname when rendering text in the PostScript file. This has been fixed.

b. graph export ignored the font settings for rendering text in the PostScript file that were specified through graph set. This has been fixed.

20. (Windows) The SEM Builder did not display the statistical category in the Details pane when showing associated results. This has been fixed.

21. (Mac) After the 10 Jun 2015 update, the toolbar buttons and some menu items for the Data Editor were disabled. This has been fixed.

22. (Mac) In the Command window or Do-file Editor, dragging and dropping text could cause Stata to crash. This has been fixed.

23. (Mac) In the Find dialog, selecting "Match whole word only" while searching in the Do-file Editor caused Stata to correctly find the search term in the current search but incorrectly save the setting. In subsequent searches, results only returned text that started with the requested search term rather than the whole word. This has been fixed.

24. (Mac) Stata graphs could incorrectly render to the screen the first point of polygons, such as arrowheads. The first point was shifted from where it should have been drawn by one pixel. This has been fixed.

25. (Unix GUI) On Ubuntu 12.04 LTS or later, Stata dialogs with multiple tabs would not display any dialog controls for Japanese. This has been fixed.

-------- update 23jun2015 -----------------------------------------------------

1. Online help and the search index have been brought up to date for Stata Journal 15(2).

2. estat abond now stores p-values in r(arm).

3. In the 10jun2015 update, we reported that import delimited, when importing a file with end-of-line characters from Mac OS 9 and earlier (\r), skipped portions of the data and that this was resolved. import delimited continued to have problems with \r. This has been fixed.

4. xtabond, xtdpd, and xtdpdsys with option twostep reported standard errors that provided poor coverage when the second-step GMM weighting matrix was not full rank. For example, this may have happened when the number of instruments was too large relative to the number of panels. When the weighting matrix is not full rank, these commands now issue an error message, and the user is advised to use the default one-step estimates.

-------- update 10jun2015 -----------------------------------------------------

1. difmh is a new command that performs the Mantel-Haenszel chi-squared test for uniform differential item functioning (DIF) in the context of IRT analysis. Tests for DIF are used to detect whether an item performs differently for individuals with the same ability who come from different groups. The Mantel-Haenszel test for DIF is used to determine whether the responses to an item are independent of the group to which an individual belongs.

2. teffects now allows fractional-outcome models with teffects ra, teffects ipwra, and teffects aipw. This lets you estimate the average treatment effect, the average treatment effect on the treated, and the potential-outcome means for responses such as rates and proportions.

3. pwcorr now stores the pairwise correlations in matrix r(C).

4. arima returned an uninformative error message when options mar(numlist, #s) or mma(numlist, #s) were incorrectly specified by omitting numlist or #s. This has been fixed.

5. bayesmh, when used with a program evaluator that calls an estimation command and when option saving() was not specified, exited with error r(601). This has been fixed.

6. bayesstats, bayestest, and bayesgraph, when used after bayesmh with option noshow() and when statistics, tests, or graphs were requested for the parameters specified in noshow(), exited with error message "parameter not found". This has been fixed.

7. betareg, churdle, and eteffects failed to report a footnote when convergence was not achieved. This has been fixed.

8. After the 05may2015 update, export excel with option datestring() generated strings containing incorrect dates when the specified datetime format was weekly, monthly, quarterly, half-yearly, or yearly. This has been fixed.

9. gsem has the following fixes:

a. gsem specified with more than one outcome variable, with weights, and with latent variables exited with an uninformative error message when perfect predictors caused a reduction in the sample size for a subset of binary outcome variables. This has been fixed.

b. gsem and some me commands reported an unhelpful error message when there were variables defined in the current dataset but no observations. In this case, these commands now report the error message "no observations".

10. import delimited has the following fixes:

a. import delimited, when string data were not present until row number 5,000 or higher for a variable in the imported text file, incorrectly chose a numeric data type instead of a string data type for that variable. This has been fixed.

b. import delimited, in rare cases, when importing a file with end-of-line characters from Mac OS version 9 and earlier (\r), skipped portions of the data. This has been fixed.

c. import delimited, when importing a file with end-of-line characters from Mac OS 9 and earlier (\r) and with extra blank lines at the end, displayed error message "java.lang.ArrayIndexOutOfBoundsException" and failed to import the file. This has been fixed.

11. import haver has the following fixes:

a. import haver, when aggregating a daily series to weekly, defined weeks as beginning on Mondays. Friday is now treated as the first day of the week.

b. import haver, when a series contained no data, loaded the series as a variable containing all missing values instead of dropping the series from the query. This has been fixed.

c. import haver, describe with option saving() always displayed the series meta-information in the Results window. The meta-information should be displayed only when suboption verbose is specified in saving(). This has been fixed.

12. irf graph with option ci#opts() did not change the rendition of the confidence interval for the #th statistic. This has been fixed.

13. irt with prefix svy, contrary to the documentation, reported the fitted parameters using the slope-intercept parameterization instead of the standard IRT parameterization. This has been fixed.

14. ivprobit for models with more than one endogenous variable specified is now more likely to converge.

15. margins has the following fixes:

a. margins with multiple predict() options and with option atmeans, option at(), option over(), or a factor variable in marginslist did not preserve the grouping of equations identified by the indepvars specified in the marginal effects options dydx(), dyex(), eydx(), and eyex(). Predictions were correctly calculated, but the output did not display results in the most logical way. This has been fixed.

b. margins after me commands with one or more factor variables that have their base level fvset incorrectly exited with an error. This has been fixed.

c. margins, when using estimation results that contain an offset or exposure, and either when option atmeans was specified or when all indepvars were fixed to constants, used one observation instead of the entire estimation sample to compute margins and marginal effects. This has been fixed.

16. mi impute intreg exited with Mata error message "st_store(): 3203 colvector required" when incomplete data contained only one censored observation. This has been fixed.

17. mlexp ignored reporting options associated with factor variables. This has been fixed.

18. nl now allows iweights to be specified with option vce(robust).

19. predict with option lnsigma after teffects ra, when the specified prediction created one new variable instead of a group of variables, incorrectly exited with an error message. This has been fixed.

20. putexcel has the following fixes:

a. putexcel, when writing a formula, created an invalid Excel formula if the formula to be written to the spreadsheet contained double quotes. This has been fixed.

b. putexcel reset a cell's existing font format to the default format for the workbook and applied only the font format specified in font(), bold(), italic(), strikeout(), or underline(). This has been fixed.

c. putexcel, when used with multiple cell formatting options (font(), bold(), italic(), strikeout(), or underline()), applied only the last specified formatting option. This has been fixed.

21. sem, when option covariance() specified more than one covariance term and when starting values were specified for a subset of those covariance terms with the init() option, sometimes incorrectly exited with an error message. Whether an error message was reported depended on the order of the terms within the covariance() option. For example, the option covariance(x1*x2 (x1*x3, init(.2))) would trigger this error. This has been fixed.

22. stcurve, when plotting curves for a model that included an interaction between a factor variable and a continuous variable and when option at() set the continuous variable to a value that was not observed in the data, incorrectly reported error message "# is not a valid level for factor variable varname". This has been fixed.

23. svy with multilevel models and survey weights specified using the [pw=varname] syntax ignored the weights. An error message should have been issued because survey-weighted multilevel models require each stage-level weight variable to be svyset using the stage's corresponding weight() option. This has been fixed.

24. vecrank, when run in Small Stata, always exited with error message "expression too long". This has been fixed.

25. xtreg, fe with vce(robust) or vce(cluster panelvar) and exactly two observations per panel reported missing standard errors. This has been fixed.

26. xtreg, mle now has improved starting values for the constant-only model that is used to calculate the likelihood-ratio (LR) test for model fit. In rare cases where option mle is specified with xtreg for models fit to datasets with few panels and very few observations per panel, the constant-only log likelihood and the LR statistic may change slightly.

27. In the Data Editor, when a value label had been applied to a variable, the display format was ignored when values of that variable were undefined in the value label. This has been fixed.

28. (Windows) In the Graph Editor, selecting Titles from the Graph menu caused Stata to crash. This has been fixed.

29. (Mac and Windows) Pressing the Tab key to complete a filename in the Command window now displays a list of filenames when there is more than one possible match.

30. (Mac and Windows) After the 05may2015 update, filename completion in the Command window was broken. This has been fixed.

31. (Mac) The SEM Builder did not render a preview of an object as the object was dragged in the Builder window. This has been fixed.

32. (Mac) The Properties pane in the main Stata window showed the wrong variable notes for the selected variable if the variable names in the Variables pane were sorted by clicking the "Name" column and the selected variable was not in the same position when sorted by Name and by #. This has been fixed.

-------- update 05may2015 -----------------------------------------------------

1. bayesmh has the following new features:

a. bayesmh now supports the suboption noglmtransform for binomial regression via likelihood(binlogit(), noglmtransform). This option allows you to fit a binomial distribution to the data. It requests that bayesmh not apply the inverse logit transformation to the linear predictor and is useful with constant-only models to directly model the probability of success.

b. bayesmh now supports factor-variable specifications of model parameters specified in options prior(), block(), initial(), exclude(), and noshow().

2. bayesgraph, bayesstats summary, and bayesstats ess now support factor-variable specifications when referring to model parameters.

3. bayesgraph with option wait now pauses until the --more-- condition is cleared even if set more off has been previously specified.

4. codebook, when a variable had more than 999 extended missing values, exited with error message "options not allowed". This has been fixed.

5. estat mfx, after asclogit or asmprobit with a value label assigned to the varname specified in alternatives(), exited with an error if any of the strings in the value label were longer than 26 characters. This has been fixed.

6. Any estimation command that allowed factor variables would slow down when matsize was set larger than 1,000 even for models with relatively few parameters. This has been fixed.

7. export excel, when exporting variables with any calendar date format other than %td, incorrectly exported the variables as a daily date instead of respecting the weekly, monthly, quarterly, half-yearly, or yearly format. This has been fixed.

8. For variable names of the form d#, D#, e#, and E#, Stata's expression parser misinterpreted some factor-variable specifications, such as 1.d1 and 1.e2, as numeric literals instead of the intended indicator variables. This has been fixed.

9. fracreg has the following fixes:

a. fracreg did not respect qualifiers if and in when verifying that the dependent variable was between the values zero and one. This has been fixed.

b. fracreg, when levels of a factor variable were specified in parentheses, exited with error message "variable i not found". This has been fixed.

10. graph replay with option wait now pauses until the --more-- condition is cleared even if set more off has been previously specified.

11. gsem has the following fixes:

a. gsem, with a multilevel model specification and prefix bootstrap or prefix jackknife, exited with an error message stating the prefix is not allowed with group-level models. This has been fixed.

b. After the 22apr2015 update, the estimation options dialog for gsem would not submit a command. This has been fixed.

12. Stata matrix row and column names parsing logic, when presented with interactions where the order of factor variables was not the same in each term, did not properly keep track of the levels. For example,

. matrix colnames b = 0.foreign#1.rep78 2.rep78#1.foreign

resulted in colnames

. matrix colnames b = 0.foreign#1.rep78 2.foreign#1.rep78

instead of the intended

. matrix colnames b = 0.foreign#1.rep78 1.foreign#2.rep78

This has been fixed.

13. meglm, melogit, meprobit, mecloglog, meologit, meoprobit, mepoisson, menbreg, and mestreg with option nolog failed to suppress the iteration log as requested. This has been fixed.

14. meglm with prefix svy and one of options eform, irr, or or failed to display transformed coefficients. This has been fixed.

15. melogit and meologit with prefix svy and option or failed to report odds ratios instead of coefficients. This has been fixed.

16. mestreg with prefix svy has the following fixes:

a. mestreg with prefix svy exited with error message "option distribution() required". This has been fixed.

b. mestreg with prefix svy and option nohr failed to report coefficients instead of hazard ratios. This has been fixed.

c. mestreg with prefix svy and option tratio failed to report time ratios. This has been fixed.

17. ml and Mata function moptimize() with lf1 specified as the evaluator type were passing the value 2 instead of 1 for todo while computing numerical second derivatives. This has been fixed.

18. odbc has the following new features and fixes:

a. set odbcdriver is a new command to specify the type of driver that odbc should use to connect to an ODBC data source.

b. In the odbc load dialog box, if the selected data source used Microsoft Access or Microsoft Excel ODBC drivers, the list of available tables was not displayed. This has been fixed.

19. saveold, when the first character of a variable name was a multibyte Unicode character and when the first byte of that character would have been an invalid starting character for a variable name in an older version of Stata, issued an error message to that effect but still saved the .dta file rather than exiting with that error message. The saved .dta file could not be read by Stata. saveold now exits with an error message and does not save the file.

20. Seasonal difference time-series operator S# did not appropriately translate the zero-period difference (S0). Depending on the command with which it was specified, S0 was translated to S1 or resulted in a variable containing only zero values. In all cases, typing S0.var is now equivalent to typing var; this translation is the same as that used for the other time-series operators.

21. sem, when fitting a model with no observed exogenous variables, with more than one latent variable, and with at least two coefficients constrained to one on paths from a latent variable to its observed measurements, incorrectly exited with an error message. This has been fixed.

22. The SEM Builder, when using a .stsem file created prior to Stata 14 and whose model was fit before saving, successfully opened the file but produced an error message when that model or any modified version of that model was refit from the Builder. This has been fixed.

23. xtivreg with option vce(bootstrap) or option vce(jackknife) exited with an error message when an independent variable was included after the endogenous variable and instruments were specified in parentheses. For example, the placement of variable not_smsa in

. xtivreg ln_wage c.age##c.age (tenure = union south) not_smsa, vce(bootstrap)

would trigger this error. Variables can now be specified after the expression in parentheses.

24. Executing an operation using the https:// protocol when an internal content buffer was less than 100 bytes would cause Stata to exit with return code r(1). This typically happened only when the results to be returned by the operation were very short. This has been fixed.

25. In the Data Editor, selecting one or more columns, right-clicking on the selection to open a context menu, and then choosing "Hide selected variables" or "Show only selected variables" from that menu could cause Stata/MP and Stata/SE to crash. This has been fixed.

26. (Windows and Unix) translate with translators smcl2pdf, log2pdf, txt2pdf, Viewer2pdf, and Results2pdf now provides faster exporting to PDF.

27. (Windows) graph use and graph save did not use or save, respectively, a graph when the current directory or the directory specified with the full path in the filename included non-ASCII Unicode characters. This has been fixed.

28. (Windows) In the Do-file Editor, when a subset of commands was selected, Stata failed to do or run the selected commands when the full path for temporary files included non-ASCII Unicode characters. This has been fixed.

29. (Windows) In the Review window, right-clicking and selecting "Save all..." or "Save selected..." would create or replace a do-file but would not include the commands. This has been fixed.

30. (32-bit Windows) update failed to request that Windows elevate permissions to administrator status. This could result in error r(608) if the user did not have permission to write to the installation directory. This has been fixed.

31. (Windows and Mac) Pressing the Tab key to complete a variable name in the Command window now displays a menu of variable names when there is more than one possible match.

32. (Mac) The current graph disappeared when its window entered full-screen mode. This has been fixed.

33. (Mac) The small circle symbol in graphs was shown in the upper-left corner of the graph instead of in the correct location. This has been fixed.

34. (Mac) Files .dta and .gph did not display the correct 16x16 icons for their file types. This has been fixed.

35. (Mac) The Do-file Editor has the following fixes:

a. Replacing text with Unicode characters resulted in the new text containing invalid Unicode characters instead of the Unicode characters that were typed. This has been fixed.

b. Selecting a subset of commands and clicking on "Execute selection" from the Tools menu or clicking on the Do button would execute more lines than those that were selected if the text preceding the selection or within the selection contained Unicode characters. This has been fixed.

c. In the Do-file Editor, using the Find dialog and selecting "Replace all" might not replace all the text specified in "Find what:" if 1) the text to be replaced occurs more than once; 2) the length of the text specified in "Replace what:" is longer than the text that is being replaced; and 3) the text that is being replaced comes very close to the end of the document.

36. (Mac) Using the Variables Manager to add or to edit a variable label would cause Stata to crash. This has been fixed.

37. (Unix and Mac) import excel and export excel used the system locale as the default locale for option locale() when reading or writing extended ASCII characters. Both commands now use UTF-8 as the default locale.

38. (Unix GUI) On some Linux distributions, Stata either did not launch or did not display icons in the toolbar. This has been fixed.

39. (Unix GUI) The Do-file Editor could incorrectly determine that a text file contained invalid Unicode characters, depending on where the Unicode characters existed in the file. This has been fixed.

40. (Unix console) translator set did not allow the default font directory (fontdir) to be set for translators smcl2pdf, log2pdf, txt2pdf, and gph2pdf. This affected only logs and graph files that needed TrueType fonts when option addfonts had been specified with translator set. This has been fixed.

-------- update 22apr2015 -----------------------------------------------------

1. Online help and the search index have been brought up to date for Stata Journal 15(1).

2. estat endogenous is a new command that tests the null hypothesis of no endogeneity for the treatment assignment after eteffects. If no endogeneity exists, teffects can be used to estimate the treatment effect.

3. churdle has improved starting values for the constant-only model, meaning that it can now report a likelihood-ratio test in rare cases where a missing value was previously reported.

4. eteffects has the following improvements and fixes:

a. eteffects has improved starting values for the exponential outcome model, meaning that many models now converge faster.

b. eteffects produced an error message when some treatment values were missing instead of excluding those observations and proceeding with estimation. This has been fixed.

5. mixed has the following improvements and fixes:

a. mixed's option dfmethod() now supports two new suboptions, eim and oim, with the Kenward-Roger and Satterthwaite DF methods. These suboptions allow you to use either the expected information matrix (the default) or the observed information matrix to compute degrees of freedom. The suboption oim may be useful if you wish to match results from other software packages that use the observed information matrix in their computations. estat df after mixed now also supports options eim and oim to provide the same functionality when computing Kenward-Roger and Satterthwaite degrees of freedom in postestimation.

b. mixed with option dfmethod(repeated) reported incorrect denominator degrees of freedom or a Mata error message if two conditions were true: 1) the model included one or more variables that varied within subject; and 2) the first subject in the data had no within-subject variation for one of those variables. This has been fixed.

6. bayesmh has the following improvements and fixes:

a. bayesmh, contrary to the documentation, required expressions to be enclosed in parentheses when specified in option likelihood(llf()), prior(density()), or prior(logdensity()). Expressions may now be specified as documented.

b. bayesmh displayed a nonmissing log marginal-likelihood as missing when redisplaying results after estimation. This has been fixed.

7. destring with option replace, when specified with a variable that already had a characteristic named "destring", exited with the error message "characteristic already exists". This has been fixed.

8. margins with option generate() after mlogit, ologit, and oprobit incorrectly exited with the error message "variable not found". This has been fixed.

9. pause would become stuck in an endless loop if you tried to start an interactive Mata session while Stata was already in a paused state. pause now returns an error message in this case.

10. stcurve, after mestreg with one or more factor variables, failed to set factor variables to the levels specified in the at() options when plotting the curves. This has been fixed.

11. svy with multilevel models and without stage-level weights specified in svyset failed to exit with an error when multilevel groups were not nested within the sampling stages. The affected commands are meglm, melogit, meprobit, mecloglog, meologit, meoprobit, mepoisson, menbreg, mestreg, and gsem. This has been fixed.

12. The SEM Builder, when used to fit generalized SEM with the Survey data estimation option selected on the SE/Robust tab of the GSEM estimation options dialog box, did not submit the command and prevented submission of all subsequent commands. This has been fixed.

-------- previous updates -----------------------------------------------------

See whatsnew13to14.

-------------------------------------------------------------------------------


© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index