|
New
in LISREL 8.54
Upgrade to LISREL 8.54
LISREL 8.54 for Windows
(May 2003) is the final version of LISREL 8.5 for Windows.
It includes the following new features and corrections that
were introduced since the release of LISREL 8.50 for Windows
in June 2001. Click here
for the upgrading instructions.
The following features,
which were not available in LISREL 8.50 for Windows, are available
in LISREL 8.54 for Windows.
1. NP Keyword (Optional)
One can put NP=<number
of decimal places> on the OU PRELIS command line. The value
of NP controls the number of printed decimals in the output
file. Default: NP=3.
2. CRAND(n) variable
(Optional)
In addition to NRAND and
URAND which generate random standard normal and uniform variables,
respectively (see PRELIS 2: User's Reference Guide, p. 191),
PRELIS 2.54 has CRAND(n) which generates a random chi-square
with n degrees of freedom. This is useful for generating non-normal
variables with different degrees of skewness. Like NRAND and
URAND, CRAND(n) can be used in expressions on NE lines (see
PRELIS 2: User's Reference Guide, pp. 67, 160).
3. Changes in the MA
keyword
A listing of the changes
that has been made in PRELIS since the release of PRELIS 2.50
is given below. These changes has to do with the specification
of MA on the OU line. For further explanation see Karl's
paper: Analysis of Ordinal Variables 2: Cross-Sectional Data
on our website.
The meaning of MA has been
changed as follows:
MA not specified implies
data screening but the meaning of data screening has been
extended.
MA = CM, MM, AM implies Alternative Parameterization (ordinal
variables)
MA = Anything else remains the same.
The output of PRELIS 2.54
is in several ways different from the output of PRELIS 2.50.
For example, "Univariate Marginal Parameters" does
not belong in the data screening and will therefore not be
printed if MA is not specified.
Also, if MA is specified the MA matrix (and means and standard
deviations) will in some cases be different, see point (vi)
below.
(i) There was a bug when
some variables are ordinal if the raw data is saved to a file
and MA is not specified. Then the ordinal variables were wrong
on the saved file. This bug has no effect when all variables
are continuous or when MA is specified.
(ii) Data screening (MA
not specified) is now done pairwise and a missing data map
(number of variables < 32) is given in the output if there
are missing values. The idea is to provide as much information
as possible about the data. But if raw data is saved to a
file, the saved file will be produced under the type of deletion
method specified even though the output file contains the
pairwise information.
(iii) If all variables
are ordinal, PRELIS gives an additional output file called
<inputfilename>.freq listing all response patterns occuring
in the sample and the frequency of occurrence of each pattern.
The 20 most common response patterns will also be listed in
the ordinary output file. The *.freq file should be regarded
as a data file. It gives the data in the most concise form.
The *.freq file may be read by PRELIS by specifying the frequency
variable as a weight variable. The reason for the *.freq file
having the name of the input file rather than the name of
the data file is that the it depends on the commands in the
input file.
(iv) If MA = PM and some
variables are ordinal, the bivariate tables will not be given
in the output file but have to be requested by BT on the OU
line. Previously these tables appeared automatically in the
output but could be excluded by XB on the OU line. Technically
this means that XB is now the default.
(v) If MA = PM and some
variables are ordinal, the threshold value for RMSEA has been
changed from 0.05 to 0.1. This affects only the P-value for
Close Fit used for testing approximate underlying bivariate
normality (see Analysis of Ordinal Variables 1: Preliminary
Analysis and Analysis of Ordinal Variables 2: Cross-Sectional
Data). Furthermore, if this P-value is less than 0.05 an additional
output file called <inputfilename>.bts is produced containing
information about the lack of underlying bivariate normality.
(vi) If MA=CM (or MM or
AM) and there are ordinal variables included, the estimation
of the MA matrix is done under the alternative parameterization.
This holds generally, i.e., also with ET, FT and FI (covariates).
These cases will be explained carefully in the next three
contributions posted on Karl's Corner.
4. XA Option (Optional)
One can put XA on the OU
line in LISREL syntax or on an Options (or LISREL Output)
line in SIMPLIS syntax. This will have the effect that only
C1 (Minimum Fit Function Chi-Square) will be computed. This
works for all methods of estimation (ULS, GLS, ML, DWLS, and
WLS) regardless of whether an AC matrix is read or not. Standard
errors are not affected. They will be the same whether XA
is used or not.
With XA, C1 is still an
asymptotically correct chi-square for GLS, ML, and WLS but
not for ULS and DWLS (without XA, C1 is not given for ULS
and DWLS in the output for this reason). The XA option is
only intended for those who have very large models and cannot
afford (or do not want) to let the computer run for an hour
or so. An alternative is to use ML without an AC matrix. Because
of its scale dependence, ULS is best applied to a correlation
matrix. To get a correct asymptotic chi-square for ULS and
DWLS one must use C2 under normality and C3 or C4 under non-normality.
In general, it is a good idea to use ML when an AC matrix
is available and use C3 for decision making (particularly
if the sample size is not very large). But for large models
this will take time (unless XA is used).
5. Multilevel modeling
module
SSI has also made enhancements
to the multilevel modeling module. A brief description of
the changes are:
1) The options paragraph may contain all or any of the additional
keywords:
SUMMARY=NONE, EFFECTS=YES, NFREE=, and DEVIANCE=<-2log-likelihood
of previous model> SUMMARY=NONE requests that the detailed
data summary, usually produced as the first part of the output
should be suppressed.
EFFECTS=YES requests the print-out of effects sizes for the
regression coefficients.
Specification of the NFREE and DEVIANCE values of a previously
fitted model results in the printout of a Chi-Square test
statistic for testing the current model versus the previous
(nested) model. If the previous model is the saturated model,
RMSEA statistics are additionally produced.
2) When dummy variables are created via the Response and Fixed
variables dialog box, these new variables are added to the
current PSF file.
3) Dialog boxes contain addtional instructions to assist the
user in building non-standard hierachical linear models.
4) The output produced when testing contrasts has been substantially
improved. Output also contains the number of free parameters
(NFREE) and -2log(L) is additionally labeled as DEVIANCE.
6. Independence Chi-square
With maximum likelihood
(ML) estimation of the model there are two different chi-square
measures of fit given in the output file. One is the Minimum
Fit Function Chi-Square C1 evaluated according to equation
(A28) in the "LISREL 8: New Statistical Features"
book. The other is the Normal Theory Weighted Least Squares
Chi-Square C2. This is N-1 times the minimum value of (A22),
where N is the sample size. Alternatively, it can be computed
by equation (A29). Under multivariate normality C1 and C2
are asymptotically equivalent but in most samples they are
different. Many of the other fit measures given in the output
are functions of chi-square, and LISREL then uses C2 rather
than C1. One change made in LISREL 8.54 concerns the chi-square
for the independence model which is also used in some of the
other fit measures. Previously we used C1 for the independence
model, which is inconsistent with the use of C2 for the estimated
model. In LISREL 8.54 we have changed that so that C2 is used
for both the independence model and the estimated model. This
change affects all fit measures which depend on the chi-square
for the independence model.
7. Using weighted least
squares (WLS)
Another change made concerns
the chi-square for the independence model when the model is
estimated by weighted least squares (WLS), also called asymptotically
distribution free method (ADF). The independence model specifies
that the population covariance matrix is diagonal and it would
seem that the most reasonable estimates of the population
variances are the sample variances. These are in fact ML,
ULS, and DWLS estimates but they are not GLS and WLS estimates.
Previously, we have fixed it so that the GLS method gives
the "correct" chi-square for the independence model
but previous versions of LISREL 8.5 the sample variances were
used to compute the independence chi-square in the WLS case.
In LISREL 8.54 we have changed that so that the WLS estimates
of the variances are used in computing the chi-square for
the independence model. This change affects also all fit measures
which depend on the chi-square for the independence model.
8. Censored Regression
A censored variable has
a large fraction of observations at the minimum or maximum.
Because the censored variable is not observed over its entire
range, ordinary estimates of the mean and variance of a censored
variable will be biased. Ordinary least squares (OLS) estimates
of its regression on a set of explanatory variables will also
be biased. These estimates are not consistent, i.e., the bias
does not get smaller when the sample size increases. Examples
of censored variables are:
Number of extramarital
affairs
Vacation expenses
Censored variables are
also common in biomedical, epidemiological, survival, and
duration studies. For more examples in other fields of application,
the reader is referred to the Karl Jöreskog (2002) contribution
on censored regression which can be downloaded from www.ssicentral.com/lisrel/column12.htm.
Note that you need version 8.54 of LISREL to run the examples.
When upgrading to LISREL 8.54 a folder called censor containing
the examples and datasets will be automatically created.
9. Import Data in Free
Format
The Import Data in Free
Format option has been changed to enable one to
Add Variable Names as the
first line(s) of the data file
Read comma separated files (.csv)
Read Tab-delimited files (.txt)
Read SPSS for Windows (.sav) files
The SPSS option is included mainly for users of the LISREL
student edition. Users of the full edition should import SPSS
data via the "Import External in Other Formats"
option.
10. Export LISREL Data
Once a PSF file is opened,
the new option "Export LISREL Data..." appears on
the file menu. If this option is selected, one can save the
contents of the PSF file to one of the following formats:
Comma separated files (.csv)
Tab-delimited files (.txt)
SPSS for Windows (.sav) files
ASCII file with no variables on top and fixed format of F15.6
The advantage of the csv file is that one can read it directly
with MS EXCEL, whereas the tab-delimited file can be read
by many software packages (e.g., SPSS).
11. Multiple Imputation
The Multiple Imputation
procedure was changed so that
All variables are carried
over to the imputed file, even if only a subset is selected
for imputation.
One can choose from one of 3 options which determine how the
imputation procedure deals with those cases where all the
variables selected for imputation are missing.
12. Factor Analysis
by MINRES
The MINRES (MINimum RESiduals)
method for exploratory factor analysis is based on the direct
minimization of least squares rather than the ULS minimization
in Jöreskog (1977), which is based on eigenvalues and
eigenvectors of the reduced correlation matrix. However, it
can be shown that, up to an orthogonal transformation, the
two methods are equivalent.
MINRES can be used with
small samples even when the number of variables is large and
when the correlation matrix is not positive definite for other
reasons (for example, this might be the case for a matrix
of tetrachoric or polychoric correlations). It is particularly
suited for exploratory factor analysis when only parameter
estimates (and not standard error estimates and chi-square
values) are of interest.
For a detailed discussion
of this topic and interpretation of the output, the reader
is referred to the contribution on MINRES by Karl Jöreskog,
which can be downloaded from www.ssicentral.com/lisrel/column13.htm.
The examples discussed
are based on data and syntax files contained in the lis850ex
and ls8ex subdirectories of LISREL 8.54 for Windows.
13. Mean Structures
in Multilevel Structural Equation Models
A multivariate level-2
model consists of a fixed part (the population means) and
a random part (the between-clusters and within-clusters covariance
matrices).
Several examples of structural
equation models imposed on the between-clusters and the within-clusters
covariance matrices are given in the MSEMEX subfolder of the
folder in which LISREL 8.5 for Windows is installed. It is
interesting to note that these SEM models are fitted by assuming
that Group1 refers to the between-clusters variation whereas
Group2 refers to the within-clusters variation. In spite of
this method for setting up the model, observations from clustered
data cannot be split into two mutually exclusive groups. Furthermore,
there is only one set of means.
In longitudinal studies,
it is often the case that the trend in means is of interest.
For example, cholesterol levels of patients at 60 hospitals
are measured on 4 occasions. Patients are randomly assigned
to a control and a treatment group. Using the treatment group
data, we may want to test the hypothesis that there is no
change in the mean cholesterol level over time. On the other
hand, it may be evident that the treatment results in a decrease
in cholesterol level over time. In this case, we may be interested
in testing the hypothesis that the cholesterol levels decrease
(or increase) linearly over time. A third type of mean structure
that may be of interest is to assume that variables such as
treatment group (in which case the control group data is also
used), gender and age influence the values of, for example,
the intercept parameter, but not the slope parameter of a
latent trait model.
For practical applications
and more detailed information, the user should consult the
LISREL 8.54 for Windows Help file. Examples illustrating the
imposition of structured means are available in the MSEMEX
folder. These examples are clearly distinguished by the addition
of "_means" or "_trend" descriptions to
the syntax filenames.
14. List of problems
corrected
Since the release of LISREL
8.50 a few problems have been brought to our attention. These
have been corrected, and include:
For a large number of variables,
the SE (select a subset of variables) command was incorrectly
generated.
A problem with the FO option, which indicates a formatted
covariance matrix (CM command), has been corrected.
A problem with stacked LISREL syntax files when performing
several exploratory factor analyses has been corrected.
A problem with the IR command (constrain parameters to a specific
interval) has been resolved.
When a model has no free parameters the minimum fit chi-square
statistic (C1) is shown in the path diagram instead of C2.
A problem with the calculation of latent variable scores for
a model with a mean structure has been corrected. These scores
are now computed correctly from a LISREL Data System (DSF)
file.
File and folder names that contain a "-" symbol
is now supported.
When a path diagram is displayed, LISREL crashed when the
"Select LISREL Outputs" checkbox on the SIMPLIS
Outputs menu (activated by selecting the Output option on
the main menu bar) was clicked. This problem has been corrected.
Parameter specifications
for fixed elements of the covariance LISREL parameter matrices
are correctly reported.
Modification indices for fixed parameters are computed and
reported.
SIMPLIS syntax for equal error covariances are processed correctly.
Correct standardized indirect effects are computed and reported
for FIML estimation.
The GF keyword on the OU command line for FIML estimation
has been updated.
Correct Chi-square test statistic values are produced in the
case of a nonpositive definite estimated asymptotic covariance
matrix.
The name of the file containing the estimated asymptotic covariance
matrix used within a Data System File (DSF) may include blank
spaces.
Large bootstrap samples
do not require the -SIZE parameter.
Correct Z-scores for univariate tests for skewness and kurtosis
are produced.
Correct estimated asymptotic variances are produced.
The correct bootstrap asymptotic covariance matrix and means
are produced.
The correct augmented moment matrix is produced.
Data values written to a PSF-file when recoding ordinal variables
are now correct.
Chi-square statistic when
testing a set of simultaneous contrasts for multilevel models
is now correct.
The LISREL ProJect (LPJ)
and SIMPLIS ProJect (SPJ) dialog boxes are operating correctly.
The estimated asymptotic
covariance matrix was not produced when there are ordinal
variables present and MA=CM and RP>1. This has been corrected.
With ET (equal thresholds) lines included and MA=PM, the standard
deviations at the end of the PRELIS output are incorrectly
given as 1.000 (although they are correctly given in the beginning
of the output). This has been corrected.
ET (equal thresholds) does not work with Probit and Logit
regression. An error message is now produced in this case.
The reference variable solution obtained with exploratory
factor analysis is now based on Sigma-hat instead of S. This
makes the solution closer to an ML solution.
Saved raw data in ASCII or PSF form is now correct without
MA specified.
One can put XU on the OU line to exclude univariate tables
in the output.
A bug in the standardized
solution when PH=ID on the MO line in LISREL syntax has been
corrected.
A bug that occurs when there are missing values and the LISREL
syntax contains CO lines has been corrected.
The reference variable solution obtained with exploratory
factor analysis is now based on Sigma-hat instead of S. This
makes the solution closer to an ML solution.
ML is now default even if an asymptotic covariance matrix
is read. Previously WLS was default in this case.
6. Applying the upgrade
When prompted to save the downloadable file to disk, save
it in your LISREL directory (where liswin32.exe is). To be
on the safe side, restart your computer before applying the
patch and make sure that the LISREL 8.5 application is closed.
The P-value for the null
hypothesis of a close model fit for multiple group analyses
is now based on the RMSEA recommendations in Steiger, J.H.
(1998). A Note on Multiple Sample Extensions of the RMSEA
Fit Index. Structural Equation Modeling, 5, 411-419.
The default estimation method is now always Maximum Likelihood
(ML) even if an estimated asymptotic covariance matrix is
specified. Please note that the estimation method should be
specified on the OU command line of the FIRST group in a multiple
group LISREL syntax file.
The correct imputed data set is produced if imputation by
matching for ordinal variables is used.
The correct decision table for the number of factors is reported
in the exploratory factor analysis results if the number of
factors is not specified.
SIMPLIS syntax files with equal error covariances (off-diagonal
elements of PSI) are now processed correctly.
Spaces are now permissible in the name of the file for the
estimated asymptotic covariance matrix specified in a Data
System File (DSF).
The estimated asymptotic covariance matrix of the parameter
estimators of the parameters of multilevel structural equation
models is now printed correctly in the file specified in the
EC option on the LISREL OU command line.
The RP command is now available for the Full Information Maximum
Likelihood (FIML) method for data with missing values.
The calculation of the degrees of freedom for the equal thresholds
hypothesis test has been corrected (ET command in PRELIS).
 |
The patch file upgradeto854.exe
can be used to upgrade LISREL 8.50, 8.51, 8.52 and 8.53 for
Windows to LISREL 8.54 for Windows (May 2003).
Instructions
- Download the patch file upgradeto854.exe to the folder
in which LISREL 8.5 for Windows is installed, for example,
c:\lisrel850, c:\lisrel851, c:\lisrel852 or c:\Program Files\lisrel853.
- Ensure that the downloaded file is of the same size as
shown below. If not, please download the file again.
- Use the Run option on the Windows Start menu to run the
patch file upgradeto854.exe.
|
