|14 Oct 1997
- xtgee, xtpois, and xtprobit have been updated to
deal with groups containing one observation (singletons) correctly.
This bug was was discovered by Donna Speigelman who reported it on the
S-News listserver concerning S-plus. She also wrote,
"This problem appears to be present in some other versions of GEE
solvers (PROC GENMOD, stata xtgee). The problem is NOT present in the
SAS IML macro that has been maintained for some time by Dr. Ulrike
We have verified the the presence of the bug in Stata. Estimation of
the exchangeable working correlation parameter was incorrect whenever
the outcome variable included singletons (clusters of size one). The
effect of the bug was to attenuate the estimated correlation in the
exchangeable structure. If the proportion of singletons was small, the
effect was negligible.
To address this problem, your Stata executable must be dated at least
14 Oct 1997 (type query born in Stata) and you must be using at
least version 3.3.0 of xtgee (type which xtgee in Stata).
If your xtgee command is older,
download the updated ado-files as well.
- xtgee has also been updated to change the calculation of the
scale parameter. Formerly N was used as the divisor and now N-k is
used. Either will work in that any consistent estimate of phi is
adequate, but most other implementations use N-k.
- “tabdisp ... if exp ...” would say “no
observations” and then proceed to produce output if no observations
- tabulate would fail to present the variable label on some
computers when the variable(s) being tabulated were strings.
- tabulate would sometimes produce a Cramér's V with the sign
reversed in 2x2 tables.
- tabulate, gen() missing would label the variable corresponding
to the missing category oddly.
- cox and stcox could go into an endless loop —
which could be stopped by pressing Break — if basehazard()
or basesurv() was specified with certain malformed models having
coefficients equal to infinity.
- Constructs such as x1a1-x1a4 for a list of new variables
produced an invalid syntax error, r(198). They are now understood.
- insheet previously required that, if there were K variables in
the dataset, each and every record have K-1 delimiters. It no longer
requires this. (Some programs, such as Excel, will produce short
records when the cells at the end are empty.)
- Attempts to define value labels with negative values, such as label
define yesno -9 yes, produced the error message “invalid syntax”
rather than the more descriptive error message “label values must be
between 0 and 32767”.
- The programming command parse, parse() now clears all numbered
macros. Previously it cleared only the one following the last to be
defined meaning that, if the user defined the missing macro, `*' could
take on a surprising value.
- infix using dictionaryname, when the dictionary itself
contained a using filename, gave an error message when
filename contained an explicit path.
|19 Mar 1997
- infile with a dictionary, when the dictionary specified that
lines were to be read out of order, could the read the data incorrectly.
- infile with a dictionary treated extended ASCII characters as
- Very long, very narrow datasets could cause Stata to crash with a
divide-by-zero error if there was insufficient memory to hold the data.
|06 Jan 1997
- cox, logit, probit, and regress, when
used with a by prefix and pweights, no longer switch to
aweights after the first group.
- A oneway tabulate with an if expression and the
generate() option no longer fills in random values for
observations not selected by the if expression.
- set maxvar/maxobs no longer gives an error since
it plays no role in Stata 5.
- Many --more-- conditions are now properly suppressed in a
graphics (gph) program.
- graph using filename no longer closes the file being
saved by a graphics (gph) program.
- table and tabdisp now stop with a
no-observations error when they should.
|17 Oct 1996
- outsheet with a variable list saved the correct values,
but used the first k variable names in the dataset
as column headers rather than the k specified variable names.
|11 Oct 1996
- insheet no longer requires the last line of a file to end with
a hard return.
- Stata's automatic repartitioning has been improved; it now avoids
|04 Oct 1996
- Depending on your simultaneous-use license, Stata for Sun and
Stata for Linux would mistakenly determine 30 minutes into each
session that the license-tracking file was corrupt when it was not.