Change Notes for Mx
Fixed some bugs with the use of the Highest command for ordinal data
Added intelligent evaluation of integrals for ordinal data when covariance matrix is of simple structure (permits efficient generalized marginal maximum likelihood via mixture distributions).
Added flushing of buffer for linux version to allow output to be useful while job still running.
Added #repeat #if #else #endif for conditional structuring of scripts.
Confidence intervals not permitted in constraint groups due to possible incorrect calculation.
Fixed RMSEA calculation not taking notice of DF= option (thx to Kris Preacher)
Fixed problem which makes samecov fail to give same answer as when samecov
is not being used. IDIFFT status was incorrectly computed for cases where
definition variables were being used. This would have led to incorrect
results when option samecov is used, and the vectors were different.
Fixed problem with reassignment of definition variables' numbers
when automatically #define'ing definition variables as -1 -2 etc.
Previous versions have a subtle bug that means if a different set
of definition variables are used in an earlier group they will
not be overwritten by the new group definitions. (Thx to me)
Also fixed problem with duplicate #NG not being understood and misread
as title. (Thanks to Andrew Heath)
Constraint problem in multiple groups: inequality constraints become
equality constraints (thanks to Caroline van Baal) - fixed.
Added ISSUB and SUB options like ISSAT and SAT= options.
Added system command to make system calls.
Fixed df when definition variables are included; Fixed some bootstrapping bugs
Added bootstrap confidence intervals; some minor bugs so use 1.54 or higher instead
Added features for mixed ordinal and continuous analysis, but still too buggy for general consumption :(
Added allint function to compute all hypervolumes of multivariate normal sliced and diced by thresholds
Improved robustness of mixture distribution modeling. New version of Genz's integration routines incorporated
Bug with printing observed means where the first one lt zero fixed
Cosmetic Bug with printing Observed means with VL/Rec files fixed.
Previously, when one or more variables contained no observations, the
observed means would be messed up. A shorter list of means for the
variables that did have data was printed, which was out of alignment after
a variable that didn't have data. The shortened list of observed means
was followed by some junk. Observed means are now correctly printed, and
0.000 is printed for variables that do not contain any observations.
Labels, when supplied, are now printed on the summary of vl data.
Bug with reading AC matrices in FU format fixed. This never worked
before, usually indicating an non-positive definite weight matrix.
VL and REC files may now be read instream with the Mx job (preferably in a
.dat file) and end with End VL or End Rectangular commands.
Option SE eliminated. Due to its susceptibility to numerical inaccuracy
and the superior properties of likelihood-based CI's, SE has been retired.
Option SA for Saturated model allows fit & df of a saturated model (or any
supermodel - except Cindy Crawford :) ) to be entered
to obtain difference chisquares. E.g., option SA=5.36,6
Option ISSAT identifies a model as saturated; subsequent models in multiple fit
mode will be compared with it.
Version 1.43   added the
following features and fixed the following problems:
Unpredictable results when a definition variable is put into more than one
Reading AV matrices, earlier versions spuriously complains that not enough
elements have been read.
Speed of multivariate normal integrator increased on average; relative accuracy
is no longer required, only absolute, which can be varied with Option
Added Option append (on all systems except IBM AIX due to the retarded compiler)
will append output matrices to files, rather than overwrite them.
Version 1.42 contains minor bugfixes for blank lines with
indented comments, and a #include feature to enable reading in parts of files
(most commonly the Data parts used in many scripts). The second line of a
script can now be NGroups=2 to keep Ngroups out of Data lines (handy with
Matrices=Group1 would erroneously read the next line looking for a number
after a space. This could leave a matrix undefined. Matrices=Groupx now
generates an error and the correct syntax Matrices = Group 1 is offered.
Blank lines with embedded comments no
longer misinterpreted as end of input file
Problem with changing missing value string to shorter string fixed Aug 6 96.
Problem with -99 not working as missing value when -99.00 does ?!
Stops optimize with correlations out of bound for mnor - now uses mode -1
Crashes internal to multivariate normal integrator too frequent for comfort
Problem with floating point overflow in numerical routine sadmvn due to r>.99
circumvented with ifail=-1
Added better debug info for cases where ; or / forgotten in a formula
Fixed memory bug that sometimes occurs after saving data to binary file
in multiple fit mode
Dropped cases where all data missing in rectangular file
Added reporting of such cases
Fixed bug where subsequent groups CI's didn't work
Option ML gives silly warning about means if rect file read. Should be
ignored with VL/Rec data. Fixed.
Get bivnor warnings in contingency table analyses - NOTED not fixed.
Problem with confidence intervals if estimates started at exactly zero -
add .01 nudge?
mxa= filename was incorrectly read as a blank filename. Leading blanks
Printout of formulae added to output.
Printing of the word Sigma & some numbers when using the momnor function
Confidence intervals first added.
No warning issued if NG missing in group 1!! Causes exception. FIXED.
Weights and NMODEL features added
New, much faster numerical integration routine added for the /muln function.
Command line options -f -k # and -h for the pc version were fixed
-k # allows larger workspace (in 1000's of words) to be allocated
All ascii characters from space downwards (to null 00) are treated as
legal delimiters. This helps with some mysterious problems with certain
editors in AIX, and with users who like tabs.
Binary file saves were improved, but rendered incompatible with earlier
versions, which can no longer be read by later versions of Mx
Simulation group idea started but aborted due to lack of time.
Various improvements to the front-end interface. Should be user-invisible.
Direct memory access for faster optimization, especially with large jobs
with many groups, matrices and/or parameters.
Notification of which raw data vectors contain uncalculable likelihoods
added to mx%p=file output files.
Workspace array A taken out of common and made allocatable for PC version.
Allows larger workspace and has smaller binary executable.
Improved trapping of zero-length formulae. Compute statement no longer
required in calculation groups.
Revised calling structure for NPSOL.
Bugs in reading full files as symmetric automatically (introduced at 1.28)
Change to Algebra and equality constraints!
Equality constraints between matrices constrained to be equal across groups
but which were defined in a Begin Algebra section were formerly problematic.
The way they used to work was that the formulae were recalcuated in the
later groups. Apart from being inefficient, this was a potential cause of
mysterious errors. The revision works the way you would expect: a matrix
constrained to equal an earlier 'Begin Algebra' matrix just receives a copy
of the result, and does not recalculate.
Second, the equate all matrices across groups has changed. The syntax
Matrices = Group 1
is still legal, but Matrices computed with Algebra statements are not copied.
Attempts to redefine matrices using matrix algebra were formerly flagged as
errors. Now it is possible to redefine matrices; a warning is issued
Warnings were standardized and the number of warnings in the run is printed
at the end.
Multiple fit option
If the multiple fit option was given anywhere other than the last group,
erratic behavior could result. Fixed.
Bugs in \momnor function fixed. Also added protection against invalid
arguments (return IFAIL=-1) so alternative points are used. Prior to
bug fix, moments of truncated distributions were correct, but not those
of correlated distributions.
For information about the
changes made in upgrades prior to 1.27,
you can browse the versions listed below.
Updates in earlier versions:
1.26 1.23 1.20 1.17
1.25 1.22 1.19 1.16
1.24 1.21 1.18 1.15