Stata 15 help for whatsnew6

Additions made to Stata during version 6

This file records the additions and fixes made to Stata during the life of Stata version 6:

+---------------------------------------------------------------+ | help file contents years | |---------------------------------------------------------------| | whatsnew Stata 15.0 and 15.1 2017 to present | | whatsnew14to15 Stata 15.0 new release 2017 | | whatsnew14 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 12.0 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 | | this file Stata 6.0 1999 to 2000 | +---------------------------------------------------------------+

Most recent changes are listed first. Note: Starred (*) items mean the update was made to the executable.

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

See whatsnew6to7.

----- update 20nov2000 --------------------------------------------------------

On-line help and search index brought up to date for STB-58

----- update 03nov2000 --------------------------------------------------------

biprobit had difficulty parsing the second equation when it was longer than 80 characters; in that case, if the first and second equations were identical for the first 80 characters, biprobit would use the first equation as the second equation, and so estimate a model which was not the one requested. This has been fixed.

glm with family(binomial) would report an error if the dependent variable did not vary even when the binomial denominator did. This has been fixed.

xtpcse with the options correlation(ar1) or correlation(psar1) was reporting an uncentered R-squared -- an R-squared with a base of 0 rather than the mean of the dependent variable -- and associated model sum of squares e(mss) and model degrees of freedom e(df_m). It now computes a centered R-squared unless option noconstant is specified.

----- update 25oct2000 --------------------------------------------------------

stsplit, as updated 19sep2000, when used with datasets containing more than 32,766 observations, would fail to split the data correctly. This has been fixed.

----- update 12oct2000 --------------------------------------------------------

stcurv, with multiple failure data and used after streg, dist(lognormal), produced a graph with a jagged line; this has been fixed. In addition, stcurv used after streg, dist(exponential) had a spelling error on the default labeling of the axis.

vif on rare occasions would fail to compute variance inflation factors for all the rhs variables in the model.

----- update 19sep2000 --------------------------------------------------------

On-line help and search index brought up to date for STB-57

bstat has a new option title() allowing the title on the output to be changed.

stcox did not save in e() the names of the variables created when the basehazard(), basechazard(), or basesurv() options were specified and fully spelled out rather than entered as abbreviations, such as baseh(), basec(), and bases(). This caused no problems whatsoever, but could cause problems if you wanted to use the saved results.

stcurv's new outfile(filename) option will save the data generated to produce the graphed curves in a Stata .dta dataset for later use.

stsplit has been improved. It runs faster, it has additional options, and it has a new syntax that facilitates splitting at failure times. See help stsplit. In addition, the new help file stsplit2 is very much like a FAQ in that it explains the usefulness of splitting at failure times and the relationship between stcox's exactp option and conditional logistic regression estimates produced by clogit; see help stsplit2. There were no bugs in the previous stsplit.

----- update 11sep2000 --------------------------------------------------------

alpha with option generate() would fail with an "invalid syntax" error message if the option was completely spelled out; this is fixed.

avplot after regress with the cluster() option would title the resulting plot using the non-robust standard error.

xtgee on rare occasions would fail with error message r(505) "matrix not symmetric". This has been fixed.

----- update 18aug2000 --------------------------------------------------------

tabstat has a new statistics(sdmean) option to produce standard deviation of the mean, the statistics() option is now case insensitive, and the default output format has been changed if only one variable is specified.

----- update 15aug2000 --------------------------------------------------------

Stata executable(*), all platforms.

rotate with the varimax option, used after factor with two factors retained, produced incorrect answers; results are now correct.

----- update 14aug2000 --------------------------------------------------------

dprobit would give an error message if any of the independent variables contained the string "_cons" in the variable name.

xtpcse with the option correlation(ar1) or correlation(psar1) now constrains estimates of panel-level autocorrelation to be in the range [-1,1]. This bounding imposes the theoretical limits on the correlations and allows the Prais-Winsten computation to always be performed. The change only affects estimation results for unusual datasets where any estimated panel-level autocorrelation fell outside the range [-1,1]. Even in these cases, the new results are asymptotically equivalent to the prior results.

----- update 02aug2000 --------------------------------------------------------

avplot after regress with the robust option would title the resulting plot using the non-robust standard error.

dfuller reported an incorrect 1% critical value when the number of observations was between 101 and 500 and options trend and noconstant were not specified.

pperron reported incorrect Z(t) and Z(rho) statistics. In addition, as with dfuller, pperron reported incorrect 1% critical values when the number of observations was between 101 and 500 and options trend and noconstant were not specified.

----- update 17jul2000 --------------------------------------------------------

On-line help and search index brought up to date for STB-56

----- update 07jul2000 --------------------------------------------------------

xtpcse when specified with the pairwise option would exit with a "matrix not positive definite" message in the unusual case when the estimated covariance matrix of the disturbances was not positive definite. The error message has been improved to state that the pairwise option may not be used. As noted in help xtpcse, this can only occur when the pairwise option is specified.

strate with option smr and option output() would fail with an error message. It now works.

stsum with option by(), and when two variables were specified in the by(), would sometimes display the stub of the table with quote characters in odd positions. This is fixed.

----- update 03jul2000 --------------------------------------------------------

stset, due to a recent update, when exit(eventvar==numlist time varname) was specified, could ignore the events in favor of the time, treating the option as if exit(time varname) had been specified. The same could happen with the enter() and origin() options. The problem could only occur when both an event list and a time were specified; there was no problem with statements such as exit(eventvar==numlist) or exit(time varname).

----- update 29jun2000 --------------------------------------------------------

ml has improved convergence properties for pathological likelihoods where one or more parameters are being driven to positive or negative infinity.

reshape, in a certain case, gave an unhelpful error message. The error message has been improved.

sts generate, in rare circumstances, would give an error message rather than producing the requested result. That is fixed.

tabstat is a new command to display a table of summary statistics; see help tabstat.

xi would give an error message if one or more of the variables in the model contained a constant; xi should have not complained and simply allowed the model to be estimated. This arose as a side effect of the 12jun2000 update and is now fixed.

----- update 26jun2000 --------------------------------------------------------

wntestb had an unnecessary global macro that was never explicitly used but could on occasion be outputted on the graph's title. This has been fixed.

xthaus when used with no covariates which vary within panel would produce a syntax error. It now gives a more informative error message.

xtpcse specified with options correlation(ar1) and pairwise, and when there were no observations on which to compute the autocorrelation parameter, would exit with the error message "matrix has missing values". It now exits with the appropriate error message "insufficient observations".

----- update 12jun2000 --------------------------------------------------------

collapse would not allow a mix of its two syntaxes, one of which specified the name of the new variable and the other of which did not, such as "collapse newvar1=oldvar1 oldvar2 newvar3=oldvar3". "collapse newvar1=oldvar1 newvar3=oldvar3" was allowed as was "collapse oldvar2", but you could not mix the two together into one statement. Now you can.

xi would give a syntax error if what followed it contained a reference to a Stata function. This was fixed on 31may2000, but not completely.

----- update 31may2000 --------------------------------------------------------

dstdize now saves numbers of observations, unadjusted, and adjusted rates for use by other programs; see help dstdize for details.

xi would give a syntax error if what followed it contained a reference to a Stata function. This has been fixed.

----- update 25may2000 --------------------------------------------------------

Stata executable(*), all platforms.

xtclog produced estimates of the parameters and standard errors that were only accurate to a few digits or, in a few cases, failed to converge. This was traced to an error in the calculation of the matrix of 2nd derivatives and is fixed.

----- update 24may2000 --------------------------------------------------------

ci with the options poisson and by() along with an if restriction would fail to produce results when one of the by groups had no observations meeting the if condition.

hausman output makes it clearer that the more efficient model in the test must be fully efficient.

----- update 08may2000 --------------------------------------------------------

On-line help and search index brought up to date for STB-55

----- update 19apr2000 --------------------------------------------------------

codebook would produce an error when sent certain variable names such as _d. This has been fixed.

fracpred with the dresid option and fracplot gave uninformative error messages after "fracpoly clogit ..." or "fracpoly probit ..." They now produce more informative errors. (The manual does not list fracpoly as working with clogit and probit, but it does. Neither clogit nor probit, however, can produce deviance residuals and so fracplot and fracpred, dresid are not allowed after fracpoly with either of these commands.)

gladder now works with variables with nonpositive values.

nl now includes an if exp on an "nl ?" call. See "Advanced programming of nlfcns" in help nl. On a query call, the syntax is now

nlfcn ? [varlist] if exp [, options]

This allows the nlfcn to perform initialization computations restricted to the relevant subsample within the portion of the code that sets the initial parameters into global macros. In addition, nl will no longer remove existing variables with the same name (or similar name based on abbreviation rules) as the parameter names. The leave option is still supported, but a check is first made that no existing variables by those names exist before creating the variables.

predict after stcox and cox reported hazard ratios of 1 rather than exp(x*b) whenever x*b>18. This has been fixed.

qladder is a new command that displays the quantiles of transforms of varname according to the ladder of powers against the quantiles of a normal distribution.

snapspan with the generate(newt0var) option now formats newt0var to have the same format as the timevar.

svyolog, svyoprob, and svymlog produced uninformative error messages when there were more than 50 outcomes (a Stata limit). It now produces the appropriate error message and return code r(149).

xtpcse now allows the time variable to be used as a regressor with any set of options. Previously, xtpcse refused to estimate models with the time variable as a regressor when options correlation(ar1) or correlation(psar1) were specified.

----- update 17apr2000 --------------------------------------------------------

Stata executable(*), all platforms.

display's _request() option now allows more than 80 characters of input.

format _all caused a crash if there were no data in memory; that is fixed.

matrix accum now produces an error when _N==1.

matrix glsaccum now produces correct results when used with time-series operators, and when the rows() option is used to repeat rows of the weighting matrix. In addition, matrix glsaccum now produces an error message when the data are not sorted by group and when the weighting matrix in nonsymmetric.

matrix vecaccum with time-series operators now correctly labels the rows.

Power calculations a^b are now calculated by repeated multiplication for integer values of b<=32; previously b<=4 was used. This makes results more accurate.

In matrix expressions, the nullmat() function worked only on the left hand side of the comma and backslash operators; it now works on both the left and right hand sides.

Stata executable(*), Unix.

Under Solaris, the message "operating system refuses to spawn new process" should no longer appear when you attempt an ls or shell command.

Under Red Hat 6.2, some people reported that Stata crashed when invoked, before the initial prompt. That is now fixed.

----- update 05apr2000 --------------------------------------------------------

xtgls with options corr(ar1) or corr(psar1) did not fully honor the force option and this could lead to the error r(504) "matrix has missing values". xtgls now always honors the force option.

xtpcse now allows estimation when the panel-level autocorrelation parameter, rho_i, cannot be computed for one or more panels and the option correlation(psar1) has been specified. Previously, xtpcse would return the error message r(504) "matrix has missing values" and refuse to estimate the model; it now assumes that these correlations are 0 and proceeds with the estimation.

----- update 27mar2000 --------------------------------------------------------

sts list with the by() option used an incorrect display format when the by-variable was numeric. Although the output was correct, the most appropriate display format was not always used.

svytab produced an "operator invalid"; r(198) error message when there were value labels containing a period. This has been fixed.

----- update 24mar2000 --------------------------------------------------------

intreg now handles aweights differently; see comment concerning 20mar2000 update to tobit and cnreg below.

kwallis was updated 09mar2000 to report an adjusted-for-ties chi-squared statistic. This new statistic was calculated incorrectly when there were missing values.

xtclog, xtlogit, xtnbreg, xtpois, xtprobit, and xtreg, all with the pa option, no longer allow aweights. With pa, however, they all now allow fweights. All now work with the noconstant option.

xtgee has had numerous changes made to it. These include

1. The default divisor for computing correlations and standard errors has been changed from 1/(N-p) to 1/N. New option nmp will produce the previous results. (Which divisor is used is of little importance since both are asymptotically correct. 1/N has the slight advantage in that it scales; see help xtgee.

2. In the particular case of family(gaussian), the default normalization for calculating the robust VCE is now n/(n-1) instead of [n/(n-1)]*[(N-1)/(N-p)], where n = # of panels and N = # of observations, and this change was made for the same reasons as (1). New option rgf will reproduce the previous results.

3. The calculation of the working correlation matrix, and thus the resulting estimates, in the case of unbalanced panels with corr(nonstationary) or corr(unstructured), has been changed.

4. xtgee no longer allows aweights.

5. xtgee now allows fweights.

6. The way xtgee calculates weighted results has changed.

See for a technical FAQ on this.

For these changes to take effect, you must also install the 20mar2000 executable update.

----- update 20mar2000 --------------------------------------------------------

Stata executable(*), all platforms.

clogit with no right-hand-side variables reported an obviously incorrect log-likelihood value when there was more than one failure per group. In addition, clogit with the offset() option reported an incorrect model chi-squared statistic.

Matrix function det() was reported to return a 0 for the determinant when the matrix had zeros on the diagonal; that is fixed.

mlogit's convergence has been improved; it now converges for some models where previously it did not. (This resulted from improving the way stepping is handled in the internal optimizer and should improve convergence of all models).

tobit and cnreg now handle aweights differently. Previously, they were just scaled and applied to the likelihood function in the "ordinary" way, but that solves no real-world problem. Now they are applied in the correct way to deal with cell-mean data with these two likelihood functions.

(aweights have been going through a change for many releases now. They started as being defined as rescaled fweights, perhaps useful, and perhaps not, and most estimators allowed them. Later, aweights, still defined in this mechanical way, were deleted from estimators in which they solved no real-world problem. By that logic, aweights should have been deleted from tobit and cnreg. aweights are now in the process of being given a different definition: that of producing estimates when the estimator is applied to cell-mean data, even if this requires changing the mechanical definition. The mechanical definition arose because it was the solution to handling cell-mean data in linear regression.)

tobit and cnreg now produce better starting values and so should converge more quickly. There was previously one report of a model in which tobit refused to converge; the new starting values fix that.

Internal changes were made to support the 24mar2000 update to xtgee, listed above.

Stata executable(*), Unix.

Stata would refuse to execute from a Gnome terminal under Red Hat Linux 6.1. That is fixed.

----- update 13mar2000 --------------------------------------------------------

egen's robs() function with the strok option (introduced 16Aug1999) counted the number of missing instead of the number of non-missing observations. This has been fixed.

On-line help and search index brought up to date for STB-54

----- update 09mar2000 --------------------------------------------------------

biprobit has been improved in three ways. (1) biprobit did not respect the ml options; this has been fixed. (2) predict after biprobit defaulted to pmarg1 when, per the documentation, it should have defaulted to p11. (3) biprobit with partial option would report incorrect values for the comparison test of rho=0.

cttost would fail if an enter variable was specified to indicate delayed entry.

heckman has four new options for use when option twostep is specified: rhosigma, rhotrunc, rholimited, and rhoforce. These options affect the handling of estimates of rho outside the bounds [-1,1] and are relevant only with the two-step method because obtaining values outside those bounds is not possible using maximum likelihood. In addition to help heckman, see for technical details.

kwallis now reports both the unadjusted and the adjusted-for-ties chi-squared statistics and associated p-values.

strate when used with the output() option would save the person-time values in a string rather than numeric variable. This has been fixed.

stset with the time0(varname) option correctly did not use observations with missing values of varname, but incorrectly did not use such observations even when processing the entry(eventvar==numlist) option. A missing value of time0() for the first observation on a subject is reasonable and even so, that record should be able to trigger entry-into-the analysis because it is the data on the subsequent record (for which time0() is not missing) that is relevant.

----- update 21feb2000 --------------------------------------------------------

heckman with the twostep option, and only with the twostep option, produced incorrect standard errors (they were overly conservative) and incorrect estimate of rho (it was attenuated toward zero) when all (emphasis on all) the independent variables in the regression equation were nonmissing in any observation flagged as not observed.

table with the replace option would give an error message if a variable named table1 was already defined in the dataset.

webseek under Windows and Mac did not honor the results option.

----- update 09feb2000 --------------------------------------------------------

gamma, gompertz, llogist and lnormal parametric survival regression models, when used outside the st system, would fail if the failure-event variable was not coded 0/1. This has been fixed. These programs assume that the failure-event variable is 0 if the observation is censored and a value other than 0 -- typically 1 -- if the observation represents a failure.

kap now reports and saves in r(se) the standard error for the estimated kappa statistic. Also, the output table now uses a better display format for the outcome value when kap is used with more than two raters and a non-integer scale.

stset without arguments and streset failed to reset the data properly if failure(varname==numlist) was originally specified, for example failure(event==12 13 14). In addition, the text of an observation-exclusion message was changed from "obs begin on or after enter" to "obs begin on or after exit" or "obs begin on or after first failure", as appropriate. Only the text of the message was changed.

xtgls now returns the more appropriate error code 459 rather than 498 when the data do not meet the requirements of the estimator.

xthaus now refuses to perform the Hausman test and suggests use of the command hausman when the estimate of sigma_u (the estimate of the variance of the random effect) is 0. The random-effects estimator in this case has degenerated to pooled OLS and the difference of the variance-covariance matrices computed by xthaus may no longer be positive definite. The test statistics reported by the more general hausman command are preferred when the difference matrix is not positive definite.

----- update 28jan2000 --------------------------------------------------------

webseek is a new command to perform a keyword search of user-written additions to Stata. webseek searches not only the STB but the entire web, including user sites.

----- update 27jan2000 --------------------------------------------------------

arch, heckman and intreg now check for collinearity between the dependent and independent variables. If collinearity exists, an error message explaining which variables are collinear is displayed.

stcurv now accepts multiple at(varname=# ...) options, called at1(), at2(), ..., at10(). This allows multiple curves to be plotted on the same graph.

----- update 21jan2000 --------------------------------------------------------

Stata executable(*), all platforms.

Stata can now allocate up to 16,384 gigabytes of memory if the hardware supports it. (With the exception of the 64-bit DEC Alpha, most hardware can support only 4 gigabytes.)

There is a new -m (Unix) /m (Windows) start up option that works like -k (/k) except that memory is specified in megabytes, not kilobytes.

Stata misparsed expressions with a very unusual set of characteristics, for example "(substr(`"a"bc"',1,1)=="f")|(1)" and similar expressions that had (1) open and close parentheses at the beginning and end, (2) the parenthesis did not match, (3) a function was used inside the parentheses, and (4) the expression contained both simple and compound double quotes. This has been fixed.

graph, c(L) mistakenly connected points when values of x[_n-1]>x[_n] but the values were exceedingly close. This has been fixed.

matrix and scalar will no longer create matrices and scalars with the reserved names _pred, _b, _n, _N, float, double, int, long, _uniform, _coef, _rc, _pi, and byte.

matrix name = exp could sometimes report an "expression too long" error when the expression was not too long, depending on the expressions previously executed (technically, matrix did not clear the literal stack). This has been fixed.

predict, residual after regress, robust now produces correct results when the dependent variable contains time-series operators.

Stata executable(*), Windows 3.1: dir had a y2k bug and reported two-digit years for the file date and reported "100" for files dated 2000. This has been fixed.

Stata executable(*), Windows 2000/98/95/NT: On the print log, the radio button for printing a selection is now disabled as it always should have been.

Stata executable(*), Windows 2000/98/95/NT and Unix: Do-files with Mac end-of-line characters (carriage return only) can now be executed without translation.

----- update 19jan2000 --------------------------------------------------------

areg and xtreg with the fe, be, and re options (xtreg without mle) now produce more accurate results. For most datasets, this improvement changes coefficient estimates in the 9th or 10th decimal place.

Internally, all of these estimators work by transforming the data into differences from group means, pseudo differences from group means, or the group means themselves. Previously, the transformed data was calculated in double precision but stored in float precision unless the original variable was double. Now transformed results are stored in double precision in all cases.

xtdata has new option nodouble. xtdata's default behavior has been changed so that the transformed variables are recast to double, making results slightly more precise. Option nodouble restores the previous behavior.

----- update 14jan2000 --------------------------------------------------------

arch now computes, in all cases, the model chi-square test using all coefficients entering the specification of the conditional mean. Previously, in the specific case where option archm was specified with options ar() or ma() and no independent variable list was specified, the ar() and/or ma() terms were excluded from the test.

areg now saves the adjusted R-square in e(ar2).

prtesti has new option count. count specifies that counts rather than proportions are specified. For example, "prtesti 20 .2 30 .3" tests whether an observed proportion of .2 in 20 observations could be from the same population and an observed proportion of .3 in 30 observations. "prtesti 20 4 30 9, count" performs the same test but the observed frequencies are specified as counts.

xtgls previously implemented a generalization of PCSEs that allowed users too easily to request covariance structures that are NOT what most people currently mean by PCSE. This would occur if panels(heteroskedastic), panels(independent), or no panels() option was specified.

Now xtgls, when specified with options pcse or ols, invokes the new command xtpcse to perform OLS and Prais-Winsten regressions with panel corrected standard errors (PCSE).

If options ols or pcse are specified and the requested covariance structure is not panels(correlated), xtgls will now display an error message. xtgls's previous behavior can be obtained by specifying the option forcepcse.

xtgls has also been modified to return the estimated disturbance e(Sigma) that was used in computing the coefficient and variance estimates and to use this estimate in computing the log likelihood. Previously, if the option twostep was specified, xtgls returned an estimate of Sigma based on the FGLS coefficient estimates. The two estimates of Sigma are asymptotically equivalent. The numerical accuracy of xtgls has also been improved when corr(ar1) or corr(psar1) are specified.

xtpcse is a new command to more accurately reflect the use of the term PCSE in the literature. In addition to providing the functionality of xtgls, ols and xtgls, pcse, xtpcse allows for unbalanced panels, estimation of autocorrelation parameters when there are gaps in the time series, and the use of time-series operators.

----- update 13jan2000 --------------------------------------------------------

On-line help and search index brought up to date for STB-53.

----- update 03jan2000 --------------------------------------------------------

ivreg now produces a better error message when the equal sign is not specified to separate the instrumented variables from the instrument list.

xtgls's help now clearly documents how parameter estimates are computed when options pcse or ols are combined with a specification for autocorrelation -- option corr(ar1) or corr(psar1).

predict after reg3 when the residual option is specified and more than one equation has the same dependent variable, now computes residuals properly. Previously, in this unusual event, the negative of the Xb prediction was returned instead of the residual.

----- update 22dec1999 --------------------------------------------------------

biprobit now identifies collinearity among variables and its error messages have been improved.

collapse's sum and rawsum statistics now compute all sums in double precision. Previously sums were calculated in double precision only if the original variable was of type long or double.

collapse now also allows spaces between the parentheses and the statistic name. For example, before you had to type "collapse (mean) x", now you can type "collapse ( mean ) x".

for incorrectly limited the number of concurrent lists to 8 instead of 9. This has been fixed.

hausman could not perform the test when one estimator used equation names and the other did not. New option equations() solves this problem; see help hausman.

tsfill with the full option could on occasion produce entries within a panel with incorrect time values resulting in two entries with the same time value and another time value absent. This has been corrected.

----- update 17dec1999 --------------------------------------------------------

prtesti reset the default confidence level for confidence intervals if you specified its level() option, meaning all subsequent commands showed confidence intervals of that level until you "set level 95" again. prtesti no longer does that.

stset's origin(varname=#s), enter(varname=#s), and exit(varname=#s) options, did not work unless varname was the same as the failure variable. E.g., "failure(outcome==5) exit(outcome==3)" did work but "failure(outcome==5) exit(diag2==32)" did not. This is fixed.

----- update 06dec1999 --------------------------------------------------------

heckman and heckprob, if the dependent variable was abbreviated, used the abbreviated name on the output (as well as to label the saved results e(b) and e(V)).

kdensity, when aweights were specified, would mistakenly report "... real changes made".

strate would distort the table alignment when reporting large values.

tabodds produced an ambiguous error message when case_var was not numeric. The error message has been changed.

ttest and sdtest, when used with by(), and in particular when the by-variable had a variable label containing parentheses, quotation marks, etc., would fail with an uninformative error message.

----- update 30nov1999 --------------------------------------------------------

ivreg now saves in e(insts) the full list of exogenous variables rather than an abbreviated list containing the first few exogenous variables followed by an ellipse (...).

prtesti with level(#) option would incorrectly produce an error message; this has been fixed.

----- update 22nov1999 --------------------------------------------------------

arch's predict will now produce predictions for ARCH models with ARCH-family terms and with multiplicative heteroskedasticity. Previously, predict refused to produce predictions for these models. predict will also now handle lags and other operated forms of the dependent variable in the multiplicative heteroskedastic component, that is, when the original model was estimated with arch's het() option.

dprobit now names the columns of the saved matrices e(dfdx) and e(se_dfdx) with the variable names from the estimated model.

heckprob with nolog option would incorrectly produce an "unrecognized command: nolog" error message; this has been fixed.

logistic has new option coef which causes logistic to report the estimated coefficients rather than the odds ratios (exponentiated coefficients).

sts graph with option enter and not lost, would not report the number who enter and the number censored on the graph; you had to specify both the enter and lost options. Now enter implies lost.

xtgls now uses a default maximum iterations of 16,000 when option igls is specified.

----- update 17nov1999 --------------------------------------------------------

cs's by() option has been enhanced to take more than one stratification variable.

stdes with the weight option would incorrectly produce an "invalid syntax" error message; this has been fixed.

svytab with only one row or column would produce an error "name conflict"; r(507). This has been fixed.

svytotal, svymean, and svyratio produced an "operator invalid"; r(198) error message when the by() option variable had value labels that contained a period. This has been fixed.

xtgee will now estimate models with options binomial and link(logit) or link(cloglog) when the model contains a perfect predictor for some of the observations. Previously, xtgee would refuse to estimate such models and report "Unable to identify sample".

xtgls now estimates models if option corr(ar1) is specified and the autocorrelation is indeterminate for one or more panels. Prior to this update xtgls refused to estimate such models.

----- update 12nov1999 --------------------------------------------------------

On-line help and search index brought up to date for STB-52.

----- update 05nov1999 --------------------------------------------------------

whatsnew (this file) improperly described the 4nov1999 update made to table. The wording below is fixed.

----- update 04nov1999 --------------------------------------------------------

bsample with the cluster() option did not produce a bootstrapped sample if the data were not sorted by the cluster variable. This is now corrected. The problem did not affect bs or bstrap because they called bsample after sorting on the cluster variable.

egen's count() function now allows string as well as numeric expressions. In the case of count(exp) where exp is a string, observations in which exp evaluates to "" are considered to be missing.

ksm with the weight option would incorrectly produce an "invalid syntax" error message; this has been fixed.

separate objected to the use of quotes in its by() option, such as "separate make, by(make=="Subaru")". This is fixed.

stset: the text of an informatory message was changed.

table displayed strange (temporary) variable names in the heading if five statistics were requested and the variables had long names. Variable names are now shortened (truncated) in the heading.

xtcorr has new option compact that displays only the estimated parameters (alpha) of the matrix of within-group correlations rather than the entire matrix.

----- update 15oct1999 --------------------------------------------------------

kdensity now correctly reports the number of observations used when the number specified in n() is larger than the total sample size. Also, when specifying at() with tied values, kdensity now retains the original order of the dataset. Additionally, kdensity now runs faster.

qnorm no longer produces an error when the connect() or symbol() graph options are specified.

_rmdcoll is a new programmer's command that identifies and removes collinearity from variable lists. See help _rmdcoll.

svytab with the subpop() and obs options now displays subpopulation counts in each cell rather than the full population counts previously displayed.

----- update 12oct1999 --------------------------------------------------------

Stata executable(*), all platforms.

Stata has had a long history (since Stata 1.0) of confusing expressions such as e+3 with numbers such as 1e+3, leading to syntax errors and users avoiding the name e for variables or scalars. That has been fixed.

anova with aweights now works. The precision improvements made to anova on 2aug1999 resulted in aweighted models being reported with all zero coefficients and missing test statistics. That has been fixed.

arima, or rather, "predict, dynamic()" used after "arima, ar(1)" now produces dynamic predictions. Previously "predict, dynamic()" after "arima, ar(1)" produced structural predictions. The problem existed only with pure AR(1) models -- "predict, dynamic()" used after "arima, ar(1) ma(1)", for instance, always produced dynamic predictions. Now all produce dynamic predictions.

cnreg now saves e(censored) containing the censoring variable name along with the other saved results.

estimates matrix and return matrix will no longer create multiple matrices with the same name. Instead, the previous matrix is erased and the new matrix replaces it. This problem only caused a waste of memory.

infile and infix no longer issue an error message when the using() option is specified and the (irrelevant) data file mentioned in the dictionary (if any) does not exist.

insheet will no longer allow creation of str81 variables.

ivreg now produces more digits of precision when used with ill-conditioned data.

macro drop now allows multiple wildcard references such as "macro drop A* B*".

table now honors specified formats wider than 9 characters.

tobit now saves e(llopt) and e(ulopt) -- the lower and upper bounds -- along with the other saved results.

Stata executable(*), Unix.

Stata for Unix is now 8-bit cleaner. Stata now defaults to specifying -e ("stata -e") when loading the pipeline driver (pd) automatically. This allows 8-bit streams to be sent to the terminal.

----- update 28sep1999 --------------------------------------------------------

On-line help and search index brought up to date for STB-51.

joinby has been made faster and some new options were added. There were no problems with the previous joinby.

qreg, iqreg, sqreg, and bsqreg sometimes reported that a collinear variable was being dropped whereas in actuality another of the collinear variables was dropped. The reported variable dropped now always agrees with the actual variable dropped. In rare instances, this could lead to a "conformability error" when using bsqreg; this no longer occurs.

pwcorr stopped with an error message if a pair of variables had no observations in common; now pwcorr reports missing value for the correlation.

_rmcoll is now documented on-line. This is an advanced programming command for use in writing estimation commands that are to run efficiently.

xttab produced an inappropriate error message when run on zero observations.

----- update 10sep1999 --------------------------------------------------------

cloglog with fweights reported incorrect number of zeros and ones. The estimation results were correct. This has been fixed.

dotplot now accepts a string variable in the by() option.

quadchk now works correctly after xttobit. xttobit now saves returned results e(ulopt) and e(llopt) for use by quadchk but which other programs may now use. e(ulopt) and e(llopt) are the ul() and ll() options specified by the user.

recast has a new force option that forces the change to be made even if that change would cause a loss of precision, introduction of missing values, etc.

reshape now allows variable-name abbreviation for varlist and j() option when changing data from long to wide form.

reg3 and sureg with weights specified now save the weighting expression in e(wexp) and the weight type in e(wtype).

zinb option nbreg has been replaced by option vuong. vuong requests that a Vuong test of ZINB vs. negative binomial be done.

zip option poisson has been replaced by option vuong. vuong requests that a Vuong test of ZIP vs. Poisson be done.

----- update 30aug1999 --------------------------------------------------------

xi, when producing interactions of variables taking on hundreds of different values, could fail with the error message "syserr: length |5| not right"; r(198). This has been fixed.

----- update 27aug1999 --------------------------------------------------------

cchart and pchart now correctly plot and report the number of units out of control if the lower control limit is zero or lower.

predict after streg, dist(exponential) and streg, dist(weibull) now calculates predicted time without including Euler's constant.

predict after streg, dist(gamma) now produces correct predicted time values when kappa is negative. Also the way it calculates hazard values has been made more numerically stable.

recode, when used on a variable of storage type long or double, and the variable contained lots of significant digits, could incorrectly include in recode ranges values that should have been excluded. This has been fixed.

sts generate newvar = s now produces survival probabilities equal to 1 rather than missing when there are no failures in the data.

tsset now returns values for the time unit (that is, periodicity) and the time-series format. r(unit) contains daily, weekly, quarterly, etc. r(unit1) contains d, w, q, etc. r(tsfmt) contains the time-series format.

----- update 17aug1999 --------------------------------------------------------

codebook now handles date variables more elegantly. In addition, the new option notes displays notes associated with variables and the new option header reports the dataset's name, etc., above the report on the individual variables. New option all is equivalent to specifying both header and notes.

xi now works properly with commands that use bound (parenthesized) equation or variable lists (such as ivreg) and with commands that take equations or variable lists in options (such as heckman).

----- update 16aug1999 --------------------------------------------------------

arima with the arima() option specified as arima(k,0,0) now estimates ARIMA models with k AR terms and no MA terms; previously arima(k,0,0) was incorrectly taken to mean k AR and k MA terms.

egen's robs() function has a new strok option. strok allows string variables to be specified and they are counted as containing missing when the string variables contain "".

streg, dist(exponential) and ereg, when the robust option is specified, now label the chi-squared model test "Wald chi2". They previously incorrectly labeled the test "LR chi2".

svytab can now tabulate variables whose variable labels contain quotes.

----- update 05aug1999 --------------------------------------------------------

lrecomp is a new command for making tables of number of correctly estimated digits (LREs) when comparing calculated results to known-to-be-correct results.

----- update 02aug1999 --------------------------------------------------------

Stata executable(*), all platforms.

anova now provides more digits of precision when used with ill-conditioned data.

mlogit, ologit, and oprobit now produce correct standard errors when pweights are specified with option cluster(). Previously, absurdly large standard errors were reported.

net now works with paths including ".." for all HTTP servers, including Mac servers. Previously, net relied on the server to handle relative paths.

regress now provides more digits of precision in R-squared near 1 regressions.

Typing "save, replace" now works when the remembered filename exceeds 64 characters.

The subinstr extended macro function with the option word now correctly handles words at the beginning of a string when the first character is doubled.

table now handles spaces at the beginning and end of string variables more elegantly.

tab2 now allows option exact to be abbreviated e as per the documentation.

Stata for Windows executable(*).

The do-file editor now includes .dct files in the list of editable files.

The log window now remembers the printer font as well as the screen font between sessions.

Copying and pasting into the data editor no longer creates a row of missing values below the pasted data.

Stata now reports attempts to open log files in read-only directories.

Stata for Mac executable(*).

Copying and pasting into the data editor no longer creates a row of missing values below the pasted data.

----- update 29jul1999 --------------------------------------------------------

On-line help and search index brought up to date for STB-50.

sts graph now includes option censored() which allows placing tick marks on the graph to indicate censorings.

----- update 23jul1999 --------------------------------------------------------

dprobit added two matrices to saved results -- e(dfdx) records the vector marginal effects and e(se_dfdx) records their standard errors.

cc now suppresses the test for homogeneity when any of the estimated odds ratios are zero or missing.

reshape. A message reshape sometimes displays has been reworded to correct a typographical error.

----- update 13jul1999 --------------------------------------------------------

aorder now handles multiple number fields within variable names.

centile now correctly handles the case of no observations (it previously went into an endless loop) and it correctly handles if exp and in range when there are multiple variables.

cf could become confused when different variables appeared in both datasets and the name of one of the variables could be interpreted as an abbreviation of the other.

dotplot now recognizes labels for the y variable.

egen's rmean() and rsum() functions now allow one or more variables rather than requiring at least two.

nl has a new delta(#) option and a default value for same that should improve the accuracy of coefficients and standard-error estimates.

predict after streg, dist(gamma) produced incorrect hazard values. This also caused stcurv to make incorrect plots.

heckman with the twostep option sometimes reported the message "matrix not symmetric".

ksmirnov, after reporting an uncorrected P-value of zero, reported the corrected P-value as zero rather than missing.

mhodds now displays better headers when option by() is specified with string variables.

streset would fail if the before option was specified.

xtnbreg now honors the nolog option.

----- update 28jun1999 --------------------------------------------------------

arima sometimes produced unclear and misleading messages during optimization. Messages have been reworded.

cf's option verbose did nothing; it is fixed.

collapse no longer retains the name of the original dataset and so typing "save, replace" after collapse will not replace the original, uncollapsed data.

compare run on some files that contained string variables presented a syntax error and stopped; that is fixed.

corrgram, ac, and pac have been given minor improvements: corrgram now stores results in r(), ac is faster, and pac handles errors better.

fracpoly. Setting the random-number seed did not re-create an analysis when the add option was specified.

lsens's replace option did not work; it now does.

stsplit and stjoin. stsplit's way of recording time could cause stjoin, in some cases, not to be able to rejoin all the records possible.

svytab with the subpop() option mistakenly printed the word "dash" in its header; it no longer does.

----- update 22jun1999 --------------------------------------------------------

Stata executable(*), all platforms. HTTP proxy authorization has been added.

In processing numlists, Stata issued an error when numlists of the form "#(#)#" or "# # to #" did not start with an integer element.

The numlist parsing command issued an out-of-range error when option range(<#) was specified even when the list did not contain out-of-range elements.

summarize now saves r(sd) = sqrt(r(Var)).

tabulate truncated a character in the labels of one-way tabulations when the labels were longer than 8 or 9 characters.

Stata for Windows executable(*). --more-- conditions now work as expected with dialog boxes.

The do-file editor now clears the edited filename when a new instance is started. Thus, restarting the do-file editor and clicking the save button now results in the do-file editor asking for the name under which the file is to be saved.

----- update 31may1999 --------------------------------------------------------

On-line help and search index brought up to date for STB-49.

cloglog with the noconstant option sometimes dropped variables when it should not have. That has been fixed and this updated version also converges more reliably.

lrtest now handles models with no RHS variables.

ml graph did not work when there were more than 10 iterations; it now does.

xtgls would, for some models, reset the iis and tis variables so that they were interchanged. This is fixed.

----- update 13may1999 --------------------------------------------------------

collapse previously produced a float result when working with long variables and now produces a double result, thus maintaining all possible precision.

heckprob now works correctly for selection dependent variables that are zero/not zero. Previously, it required that the selection dependent variable be 0/1.

reshape mislabeled the heading of its report when reshaping from long to wide. It said "wide -> long" when it should have said "long -> wide".

stset's exit(var=numlist) option would not work if numlist contained more than one variable.

----- update 05may1999 --------------------------------------------------------

prais issued an error message when used with more than 10 or 20 variables. Now it is fixed.

----- update 27apr1999 --------------------------------------------------------

arima and arch. To make the interpretation of dynamic predictions more consistent for a wider range of models, the meaning of the dynamic() option has changed. Now, all references to the dependent variable, say y, prior to the time period specified in dynamic() evaluate to the observed (not predicted) value of y; all values of y on or after dynamic() evaluate to the predicted value of y.

xi would not work with time-series operators if the dataset was tsset as panel data.

----- update 26apr1999 --------------------------------------------------------

tabi now works correctly with larger than 9 x k or k x 9 tables.

----- update 23apr1999 --------------------------------------------------------

grmeanby did not work when weights were specified.

ivreg would fail if the number of instruments exceeded about 450.

ltable did not present the log-rank test if weights were specified.

pause now runs in version-6 mode, meaning commands you type in debug mode are interpreted according to the modern rules.

sttocc did not necessarily reselect the same "random" sample when rerun with the same random-number seed; it now does. In addition, sttocc's match() option now allows a varlist rather than just a varname.

stmh did not, under certain conditions, report results when option by() was specified.

sts now uses the value labels to label graphs when you have attached value labels to the analytic time variable _t (as unlikely as that may be).

----- update 21apr1999 --------------------------------------------------------

Stata executable(*), all platforms. cox and stcox now have a new noadjust option for use with the robust option. This option prevents the estimated variance matrix from being multiplied by N/(N-1) or g/(g-1) (g the number of clusters). Most users will probably not wish to use this new option.

gettoken and tokenize now support long macros.

*.pkg files may now contain "f ../etc." references to mean parent directory.

Stata executable(*), Unix. Some users saw $<#> in help files when text was supposed to be highlighted. The text is now shown in highlighted form just as intended.

----- update 12apr1999 --------------------------------------------------------

On-line help and search index brought up to date for STB-48.

----- update 02apr1999 --------------------------------------------------------

adjust. The reported standard errors were incorrect when some variables were left unspecified. Results were correct when every variable in the model was specified either in the varlist or in the by() option.

arima and arch could not estimate models if there were insufficient contiguous observations for the starting value algorithm. In such cases, they now default to starting values of 0 for the ARMA and/or ARCH parameters.

hausman has new options prior() and current() to allow the labels for the coefficient columns to be changed.

heckman. The counts of censored and uncensored observations were reversed in the output.

heckprob. (1) The noconstant option was not honored in the selection equation; it now is. (2) When there were missing values among the RHS variables of the primary equation for observations that were not selected, these observations were incorrectly omitted from the estimation sample. (3) Better starting values are now used so models should converge more quickly.

integ incorrectly required the user to specify the generate() option.

istdize would sometimes mistakenly issue the error message that the number of cases exceeded the population size and so refused to run, even when that was not true.

kdensity's title() option now understands title(" ") to mean no title is to be displayed and the title() option will accept compound double quotes.

quadchk now works with xtintreg.

scobit's model test is deleted because the ancillary parameter alpha and the constant are not both identifiable in the constant-only model. scobit now also converges more reliably.

xtgls. (1) xtgls now reports the correct log-likelihood. (2) Previously, if option igls was specified, xtgls reported consistent, iterated results, but these were not the MLE; now MLE results are produced. (3) Previously, when panel(correlated) was specified without option igls, the GLS matrix was based on a panel-heteroskedastic model. It is now computed from OLS. Both yield consistent results but the new approach is more in line with what most people would expect.

xtnbreg, fe now explicitly omits all observations that do not contribute to the likelihood (for example, groups of size one). xtnbreg, both fe and re, also now converges more reliably for difficult problems.

xtpois, fe now explicitly omits all observations that do not contribute to the likelihood (for example, groups of size one). xtpois, re normal now converges more reliably for the constant-only model.

----- update 19mar1999 --------------------------------------------------------

snapspan refused to transform the data and instead issued a syntax error when there was more than one event variable in the dataset.

tabodds. When weights were used, the score test for trend reported by tabodds was incorrect. (All other statistics were fine.) The problem has been fixed.

----- update 15mar1999 --------------------------------------------------------

nbreg has been enhanced to optionally estimate an alternative parameterization of the negative binomial model. Choosing the option dispersion(constant) yields a model in which the overdispersion is constant from observation to observation. See help nbreg for details.

stphplot would sometimes connect the points in the resulting graph in an odd order.

xtnbreg, fe should have included a constant term, but it did not. When weights were specified in the re and fe models, it would not converge to the right answer; it now converges reliably. The comparison model for the re case is the new nbreg, dispersion(constant) alternative parameterization. The command is now much faster as well.

xtpois, re (without weights) could sometimes have a problem converging. When weights were specified in the re and fe models, it would not converge to the right answer. It now converges reliably with and without weights, and it is much faster as well.

----- update 04mar1999 --------------------------------------------------------

fracpoly. The way the adjust() option works has changed.

sw now works with ologit and oprobit.

Stata executable(*), all platforms. When version is set to 5, label save now produces files compatible with Stata 5.0.

cox and stcox no longer allow specifying option robust with mexact or pexact options.

ologit and oprobit now report the Wald model chi2 statistic if option offset() is specified.

inspect caused fracpoly to crash under certain circumstances; that is fixed.

Stata for Mac executable(*). can now be located in a folder that contains spaces in its name.

The Log button is now synchronized with the command line.

Stata for Windows executable(*). The Review window now maintains its size even when used with TrueType fonts. can now be located in a folder that contains spaces in its name.

With some printers, the do-file editor would only print on a quarter of a page. It now prints on the whole page.

----- update 22feb1999 --------------------------------------------------------

cnsreg was not saving the log-likelihood.

dstdize's saving() option now works and its base() option now works with value labels.

heckman's mills() option did not work if an optional dependent variable was specified simultaneously with the select() option. In addition, option nshazard() has been added as a synonym for mills() because it more accurately describes what is calculated (nshazard stands for nonselection hazard).

linktest would not run if you set version prior to 6.

lrtest. The reported degrees-of-freedom for the test would be incorrect if the estimator's option constraints() was used in the estimation of either model and it was the constraint being tested. (This could only affect reg3 and mlogit).

strate's jackknife option now works with pweights and iweights.

svylc, show did not work; it now does.

----- update 10feb1999 --------------------------------------------------------

areg. predict after areg did not take into account weights if they were specified.

bs has been improved so as to post a missing value when a statistic cannot be calculated, or when the bootstrap sample causes the estimation command to produce an error.

bs, bstrap, bsample gave an error message when the cluster() or size() options were specified. These options now work.

bstat now handles missing values on a variable-by-variable basis.

fracpred. Typing fracpred newvar, for(xvar) where xvar was not a predictor produced an incorrectly worded error message.

streg, dist(gamma) could produce the message "initial values invalid", and then refuse to estimate the model, if _t in the first observation was missing.

svymean, svytotal, and svyratio gave an error message when they encountered value labels longer than 8 characters.

svymlog gave an error message when it encountered value labels with embedded quotes.

xi could not be used with the version-5 (backward compatibility mode) st commands or other version-5 commands.

----- update 28jan1999 --------------------------------------------------------

Stata executable(*), all platforms. The outfile command did not respect date formats. The sysdir set command incorrectly handled trailing slashes.

Stata for Windows executable(*). If an application such as //":StataQuest changed the Help menu, later the original menu bar could not be restored.

----- update 27jan1999 --------------------------------------------------------

Stata for Mac executable(*). In some cases, Stata had difficulty finding its library that allows it to communicate over the Internet.

----- update 19jan1999 --------------------------------------------------------

Stata executable(*), all platforms. Expressions containing nested, compound double quotes (e.g, `"`"this"'"') in which the quoted string itself contained +, -, *, /, would result in a syntax error.

Stata for Windows executable(*). File--Close did nothing under the do-file editor, although the Close button did work. In addition, double clicking on very long lines in the Review window would cause the line to be truncated when executed.

Stata for Mac executable(*). Long string variables (longer than 20 characters) could cause problems in the data editor.

----- update 13jan1999 --------------------------------------------------------

gnbreg. The degrees of freedom for the likelihood ratio test of alpha=0 (comparison with Poisson model) were set to one in all cases. The degrees of freedom should be the number of parameters in the lnalpha equation. This is fixed.

heckprob did not honor if exp or in range.


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