Known Bugs in RATS Version 4
Software bugs are a fact of life, although we make great efforts to minimize them. By making a list of known bugs available, we hope to limit the inconvenience these problems may cause. Please contact us if you have any questions.
This page provides a bug-fix history for Version 4 of RATS. For bugs and bug-fix information for the current release, please see the Known Bugs page.
For software like RATS, these can be grouped in three basic categories:
- Serious: Bugs that produce incorrect results that are not obvious to the
user, and/or that impose major limitations on functionality.
- Annoying: Bugs that are obvious but annoying, or impose moderate
limitations.
- Minor: Minor or very rare inconveniences, presenting little or no
limitations on functionality (i.e. interface glitches, problems with easy
workarounds, etc.).
This "short version" page includes only the "serious"
bugs—the ones we particularly want you to be aware of. For a more complete
listing, see the Known Bugs (Long Version).
For the most part, this only includes bugs that were present in release
4.10c or later (4.10c was a bug-fix update to Version 4.10). If you are using
something older than that, we would recommend updating to get the bug fixes
mentioned below (in addition to all the spiffy new features, of course). At some
point we will extend this list to include older problems, particularly those
affecting Mac users stuck at Version 4.02. Stay tuned.
Note: As of May, 1998, the current version of RATS for all platforms except
Mac is 4.31, which is identical to 4.30 except for the bug fixes noted below.
Bugs Not Fixed as of 4.31
Bugs Fixed in Version 4.31
- Equality constraints frequently would not hold in LQPROG problems (RATS
would compute a solution treating the equality constraints as <= constraints,
rather than equality constraints). Appeared in 4.30. Fixed in 4.31.
- In NNLEARN, changed the way node values were scaled to eliminate some
convergence problems. Appeared in 4.30. Fixed in 4.31.
- The formula we used for KAPPA with the cosine taper was wrong—Pi should
not appear in either the numerator or the denominator. Present since 4.0. Fixed
in 4.31.
Bugs Fixed in Version 4.21 or 4.30 (4.30 for UNIX)
Versions 4.21 (PC platforms only) and 4.30 were released at approximately
the same time. 4.21 added only bug fixes and was available for shipping/handling
costs, while 4.30 include the same bug fixes plus new functionality and was
slightly more expensive ($25 from 4.2 to 4.3).
- Original release of WinRATS-32 4.20 would not start up normally under Win95
or NT 4.0 (which shipped after RATS 4.20). Workaround was to run as minimized
icon (i.e. start in batch mode, then switch back to interactive using File-Batch
Mode). Fixed in August, 1995 build of 4.20.
- Due to some algorithm changes, the NOSQUARE option on MCOV has no effect in
Version 4.21 (i.e. produced same results as if you had used default SQUARE
setting). This affects, for example, our ROBUST.PRG example (page 5-11). As long
as the resids series is all positive, you can work around this by just taking
the square root of the series before applying MCOV. Fixed in 4.30.
- If you use AR1 with panel data, the residuals are not set correctly. Rather
than being "NA", the first obs. of each indiv. will be set to the
previous residual value. Fixed in 4.21.
- %MQFORM(A,B) produced wrong answers when the A matrix was declared as
RECTANGULAR. Worked properly when A was SYMMETRIC (workaround is to just use the
standard matrix functions for which %MQFORM is a shortcut—see Appendix B).
Fixed in 4.21.
- The DAMP option had no effect on NLSYSTEM (so standard errors were, if you
will, the "undamped" standard errors). Fixed in 4.21.
- The TEST instruction returned a Chi-squared statistic when used after a
BOXJENK—should have returned an F-statistic. Fixed in 4.21.
- Through 4.20, MVFRACTILE used the wrong weights in most cases when
interpolating between two numbers (i.e. if the fractile doesn't hit an exact
data point). Numbers were close (between the correct two numbers, but weighted
incorrectly). MEDIAN option worked OK, but most other factors would give the
wrong numbers. Fixed in 4.21.
- A perfect-fit regression, such as a constant on CONSTANT will crash
WinRATS-16 if Q-stats are calculated (i.e. if a CAL is in effect). WinRATS-32
returns NA's for Q-stat and significance level. Fixed in 4.21.
- When saving undated panel data to a RATS format file, there was a problem
with handling missing values at the beginning of the data set (the NA's weren't
saved, so the data was shifted "up"). Fixed in 4.20.
Bugs Fixed in Version 4.20
- Standard errors reported by MAXIMIZE were incorrect when using the
ROBUSTERRORS option. Note that BHHH results should be unaffected by ROBUST, but,
due to this bug, ROBUST produced different std. errors than NOROBUST. Std.
errors also wrong for BFGS with ROBUST. Fixed in 4.20.
- Leading NA's and trailing NA's in a panel data set get chopped off when
doing the PANEL instruction if you use a / for default range (i.e. the data
would be shifted "up", or earlier in time). Workaround is to use
explicit date ranges for start and end. Fixed in 4.20.
- Through 4.10, EIGEN instruction did not sort results correctly if the input
matrix was type RECTANGULAR (non-symmetric) matrix. Problem fixed in 4.20.
- For univariate models, %LOGDENSITY would return the determinant, not the
LOG(DET), of the first term, and would this give incorrect results. Works OK for
multivariate, not univariate. Workaround is to write them the old fashioned
way. Fixed in 4.20.
- Through version 4.10, ORDER(ALL) didn't sort the last (highest numbered)
series. All series are sorted if you list them explicitly (e.g. "ORDER X /
Y Z" sorts X, Y, and Z, while "ORDER(ALL) X" sorts only X and Y).
Fixed in version 4.20.
- Through version 4.10, STWISE with the METHOD=BACKWARD option would always
use the P-values from the initial (full) regression, except for slight
differences due to changing degrees of freedom. Thus STWISE could very well
report the wrong final model. Fixed with version 4.20.
- Problem with converting data from RATS format to RATS format using
STORE(CONVERT=RATS) (some data would be slightly incorrect due to bogus
frequency conversion attempt). Fixed in 4.20.
- When doing COPY(FOR=WKS,DATES), RATS 4.10c sometimes writes out the dates
incorrectly for monthly, quarterly, and annual data (daily, sevenday, weekly are
all OK). Fixed in 4.20.
- RATS ignored * symbols on "PATHS" supplementary cards for
impulse, so you may get shocks applied to the wrong equation if skipping some
equations. Workaround is to use 0 (the number zero) instead of * to skip
equations. Fixed in 4.20.
- If an individual (or time period) in a panel set is all NA (missing
values), PANEL(COMPRESS) sets the corresponding entry in the output series to 0,
rather than NA. Fixed in 4.20.
- TYPE option choices on TAPER were not recognized. Fixed in 4.20.
Bugs in Older Versions
Most bugs fixed by Version 4.10c or earlier are not listed on this page.
However, a few important ones are shown below:
- Through version 4.02, hypothesis tests done after an NLSYSTEM or MAXIMIZE
estimation were done as F-tests, when they should have been chi-square tests.
Fixed with Version 4.10.
- When doing EXCLUDE after doing AR1, the statistic used for the test is
wrong by a factor of %SEESQ. Workaround is to take chi-squared statistic from
output and divide this by %SEESQ to get the proper statistic, than apply CDF:
COMPUTE STAT = # from Exclude output/%SEESQ
CDF CHISQ STAT 1
Fixed with 4.10
- Weighted Least Squares Problem:
INST Z1 Z2
LINREG(INST,SPREAD=S) Y
# X1 X2
yields the wrong results. Instead use
INST Z1 Z2
CMOM(SPREAD=S)
# Z1 Z2
COMPUTE %CMOM = INV(%CMOM)
LINREG(INST,SPREAD=S,WMATRIX=%CMOM) Y
# X1 X2
Fixed in RATS 4.10
