Online Resources for: Bowen, N. K. & Guo, S. (2011). Structural Equation Modeling. Oxford University Press.

Online resources accompanying book. Please see disclaimer, and cite appropriately. Using Mplus for SEM Disclaimer: Mplus is a powerful SEM program. Many options available in Mplus are not covered here. Information on the options that are covered is based on our experiences with recent versions of the program. These guidelines are not meant to be comprehensive or exhaustive. They may not reflect upcoming versions of the program or recommendations of the Mplus program developers. See also the Mplus website: http://www.statmodel.com/index.shtml for Mplus manual and program updates, and consult the in-program help feature. C.1. Preparing and Saving Data Files for Use with Mplus This section provides information on preparing data files for Mplus. Mplus has many options related to data and variables. The presentation is not exhaustive; we cover just some of the most common situations social work researchers may encounter. C.1.a Missing Values As part of the preparation of data for SEM analysis in Mplus, users must designate which symbols or numbers in their datasets represent missing values. Options for missing values include: period (.), asterisk (*), blank (), and numeric values that are not among valid options for a variable. Because of restrictions on the use of non-numeric flags (e.g., only one per dataset), we recommend using numeric flags. Either positive or negative numeric values can be used as missing value flags; just be sure that the value used for any one variable does not overlap with its potential valid values. In Mplus, more than one missing flag may apply to one variable, one missing value flag can be used for all variables, or different flags can be used to designate 1

missing values in different variables. Therefore, users do not have to alter existing data files to make all missing values the same. Different treatments of variable response options can be specified in Mplus without making changes to the original data file. C.1.b Saving Files for Use in Mplus Before saving an input data file, all data cleaning should be completed, any necessary data transformations and recodes should be completed, and missing values should be recoded to one of the options Mplus reads. Files may contain variables that will not be included in Mplus analyses, but all variables will have to be named and listed in the Mplus syntax in order for the file to be read correctly by Mplus. Raw data files with only numeric variables should be saved as free or fixed ASCII files. They may be saved them as tab, space, or comma delimited text files. If files are saved in free format, it is not necessary to specify the file format in Mplus syntax. Those saved in fixed format need to be described on the FORMAT line in Fortran notation (e.g., F8.2). There can be no blanks in files in free format (therefore, missing values cannot be designated with blanks in free format text files). Files saved from SPSS as tab delimited .dat files are among those read in Mplus. From Excel, files saved as text MS_DOS files (with .txt extension), or formatted text, space delimited files (with .prn extension) are among those that can be read. To use a lower triangle or full covariance matrix as the input matrix for Mplus, create a free format text file from a spreadsheet or database program with the matrix beginning on the first row of the file. To use a correlation matrix, create a free format text file that includes the means of the variables (in the same order as they occur in the correlation matrix) on the first line of the file, the standard deviations of the variables on the second line, and the first line of the full or lower triangle correlation matrix on the third line. 2

Example: 1. How to format a correlation matrix of 4 variables to be used in Mplus analyses. .46 .14 1 .13 .23 .48 1 .81 .56 1 .44 1 Correlation matrix of the four variables .34 .22 .75 .49 .52 .38 Means of the four variables Standard deviations of the four variables

C.2. Specifying Data Files and Variables in Mplus Syntax with the DATA and VARIABLE Commands and Options C.2.a Specifying Data Files Users specify where to find data files and what kinds of data are in data files with the DATA command and options. Note, the indentations of options under commands are for ease of illustration; they are not required in Mplus. Mplus code lines can be continued from one line to the next if necessary and end with a semi colon. DATA: FILE IS TYPE IS If a data file is not located in the same directory as the syntax file, path information is specified in the DATA: FILE IS line. Note that the full directory information has to be contained in quotes if the path or the file names contains any blanks. Mplus will give a message that it cannot find a data file if a path with blank spaces is not put inside quotation marks! See Example 1 below.

If no TYPE IS line is included in the syntax, Mplus assumes the input file is in the common data format with one row for each subject and one column for each variable. If another format is used, TYPE IS must be included in the syntax. Other TYPE IS options include COVARIANCE and CORRELATION for lower triangle matrices, FULLCOV and FULLCORR for full covariance or correlation matrices, and IMPUTATION followed by a list of file names of imputed dataset. Examples: 1. The FILE IS line below indicates that the free format text file called worksatis.dat can be found on the D drive in the folder called Mplus analyses. Because there is a blank space in the folder name, the path name is in quotes. No TYPE is specified, so it is assumed that the data file has rows for records (subjects) and columns for variables. DATA: FILE IS D:\Mplus analyses\worksatis.dat; 2. The FILE IS line below indicates that the free format text file called Mplustest.dat is in the current directory. The text file contains a lower triangle correlation matrix, preceded by a line (row) of means and a line of standard deviations. The third line indicates the number of observations (subjects) represented in the correlation matrix. If the user wants to use only the first 300 observations, the NOBSERVATION line could specify 300 instead of 390. DATA: FILE IS agencyinfo.dat; TYPE IS CORRELATION MEANS STDEVIATION; NOBSERVATIONS ARE 390;

C.2.b Specifying Variables Users indicate characteristics of their variables in the VARIABLE command section of an Mplus input file. 4

VARIABLE: NAMES ARE USEVARIABLES ARE MISSING ARE CATEGORICAL ARE The first option under the VARIABLE command is the required option of specifying the names of the variables included in the input dataset. The NAMES ARE line lists in order all the names of the variables in the data file to be analyzed. Mplus reads text files without names, so the names entered here do not need to match any variable names used previously or displayed in other programs. Names should be 8 or fewer characters long and start with letters, but can contain numbers and underscores. We have found it convenient to generate a list of all variable names in a file by pasting the frequency command in SPSS (for example) into an SPSS syntax file, then copying and pasting the list into the Mplus VARIABLE: NAMES ARE line. Example: 1. The NAMES ARE line in the example below indicates that the data file contains five variables. The USEVARIABLES ARE line indicates that only four of the five variables will be used in the current analysis. VARIABLE: NAMES ARE gender age sowoexp typagen jobsatis; USEVARIABLES ARE age sowoexp typagen jobsatis; 2. The code below specifies that all of the variables in the input data file will be included in the subsequent modeling information. It is possible to list additional variables after ALL on the USEVARIABLE line if new variables have been created within the Mplus program (and therefore and not in the input file). VARIABLE: NAMES ARE gender age sowoexp typagen jobsatis; USEVARIABLES ARE ALL; 5

The NAMES ARE and USEVARIABLES lines are important. If the user attempts to model a variable later that was not included in the USEVARIABLES line, even if it appears in the NAMES ARE line, Mplus will return an error message. If a variable is listed in the USEVARIABLES line and not included in the later modeling statement, Mplus will inform the user that a variable it expected to use has not been included. Missing values are described in the VARIABLE command based on how they were treated before the input data file was saved from another program. The examples below illustrate how one numeric missing value flag can be applied to all variables in a data file (Example 1), different variables can be assigned different missing value flags (Example 2), and each variable can have more than one missing value flag (Example 2). Examples: 1. The MISSING line here indicates that for all variables, 99s should be read as missing values. VARIABLE: NAMES ARE gender age sowoexp typagen jobsatis; USEVARIABLES ARE age sowoexp typagen jobsatis; MISSING ARE ALL (99); 2. The MISSING line here indicates that missing values for variable 1 (v1) are indicated by 99, and that missing values for variable 2 (v2) are indicated by 0, 00, and 000. VARIABLE: NAMES ARE gender age sowoexp typagen jobsatis; USEVARIABLES ARE ALL; MISSING ARE v1 (99) v2 (0 00 000); One of Mplus strengths is its ability to appropriately analyze variables with different distributional and measurement qualities. The default assumption in the program is that variables are continuous. Users can specify non-continuous variables as CENSORED, CATEGORICAL, NOMINAL, OR COUNT variables if appropriate. We focus on the second and third of these 6

types because they have been referenced throughout the book. The CATEGORICAL specification is for variables with between 2 and 10 ordered response options. CATEGORICAL is the appropriate designation for observed latent variable indicators that are measured with Likesupt scale or other ordinal measures. The NOMINAL designation is used for variables with between 2 and 10 unordered options. Variables that are specified as either CATEGORICAL or NOMINAL are recoded by Mplus such that the response option with the lowest level becomes 0, with other options assigned higher values in increment of 1. Example: 1. In the example below, the CATEGORICAL and NOMINAL lines indicate that variables ordv1 through ordv5 are what we have referred to throughout this book as ordinal variables; gender is nominal. Age is being specified by default as a continuous variable. VARIABLE: NAMES ARE gender age ordv1 ordv2 ordv3 ordv4 ordv5; USEVARIABLES ARE gender age ordv1 ordv2 ordv3 ordv4 ordv5; MISSING ARE ALL (.); CATEGORICAL ARE ordv1 ordv2 ordv3 ordv4 ordv5; NOMINAL ARE gender; C.2.c Mplus Language Generator Mplus has a language generator that facilitates the specification of data, variables, and estimation options. The language generator allows the user to generate part of the code for an analysis quickly and accurately. The language generator does not create code specifying the model; that code is added later by the user to the generated syntax. The choices that appear throughout the language generator sequence vary slightly depending on your preliminary choice of SEM, SEM with missing data, and SEM with clustered data. For example, if you select SEM, you will see the option of indicating that you have individual (raw) data or summary data (a 7

covariance or correlation matrix). This option does not appear if you choose SEM with missing or clustered data because summary data cannot be used for those types of analyses. Here we provide a quick tour of the language generator as it would apply to an analysis of clustered data with missing values. The text created from the language generated can be edited and amended after it is pasted. Some users may prefer to start their syntax files with the generator; others may prefer writing their own code from scratch. Go to the Mplus pulldown menu at the top of the screen you see when Mplus opens. Choose language generator. Choose SEM with clustered data. (Even if you plan to do an SEM with clustered data AND missing data, select SEM with clustered data.) Enter a title for the analysis you are about to do if you would like to have a title on your input code. For this example, we enter: Example of Language Created with Language Generator: SEM with Clustered Data. Browse for the input data file and select it. Click on Next. Indicate if the data file is in free or fixed format. (The examples we gave above assumed free format, and correlation and covariance matrix data must be in free format.) Indicate if there are missing values in the data, and whether they are identified with numeric or non-numeric values. Click Next. In the next screen you are asked to enter variable names and add them to the list of variables in the data file one by one. If you have a dataset with many variables, we suggest entering just one or two variable names and cutting and pasting the rest into the code after the generated language is pasted into the Mplus editor. Click Next. In the next screen, the list of variables you have entered will appear on the left. Select the one variable (or more) you entered and move it into the Usevariable box to the right. This is where the Usevariable line is created. Add only those variables you plan to use in the current analysis. If you entered only one or two 8

variable names at the previous screen, the rest can be pasted in later. Another option on this screen is to restrict the observations used to those with specific values on one or more variables. For example, if the proposed analysis is of individuals over 50 as indicated in a variable called age, specify in the box that Age GE 50. More than one restriction can be specified using AND, OR, NOT, and the following EQ (equal), NE (not equal), GE (greater than or equal to), LE (less than or equal to), LT (less than), GT (greater than). We do not specify and subset of observations for this example. We specify that we will be using the five ordinal variables in the analysis. We do not list cluster in the USEVARIABLE line; it will be given a special role as the clustering variable in the next screen. In the next screen, you indicate what values represent missing values in your dataset. We specify that all variables use 99 as a missing variable flag. Click Next. In the subsequent screen, variables can be assigned special rolesthe grouping variable in a multiple group analysis, the cluster variable in a multilevel model, variables measured at level 2 in a two level model (between variable), and weighting variable. Because this example assumes an analysis with clustered data, we indicate that the variable called cluster in our dataset indicates how our level 1 subjects are grouped (e.g., within classrooms, agencies, states, etc.). The following screen has analysis options. We choose complex because we want to correct for standard error biases created by the grouped nature of our data. (If you planned to model level 2 predictors, you would choose twolevel.) The default estimator for a complex analysis is MLM, which uses the Satorra Bentler chi square statistic corrected for clustering of cases. We leave the next two options related to optimization at their default levels. They can be altered if we have convergence problems later.

The final screen in the language generator asks what output we would like in addition to the default output. We indicate that we would like standardized parameter estimates. The code created by the Mplus language generator from the above example and pasted into the Mplus editor is as follows: Example of Language Generator Syntax before Editing: TITLE: Example of Language Created with Language Generator: SEM with Clustered Data DATA: FILE IS "F:\Mplus analyses\Langexample\Mydata.dat"; VARIABLE: NAMES ARE age gender ordv1 ordv2 ordv3 ordv4 ordv5 cluster; USEVARIABLES ARE ordv1 ordv2 ordv3 ordv4 ordv5; MISSING ARE ALL (99); CLUSTER IS cluster; ANALYSIS: TYPE IS COMPLEX; ESTIMATOR IS MLM; ITERATIONS = 1000; CONVERGENCE = 0.00005; OUTPUT: STANDARDIZED;

The code pasted into the Mplus editor by the language generator is now ready for editing and amendment. If there are more variables in the dataset than we entered in the generator, they can be added as appropriate to the NAMES ARE and USEVARIABLES ARE lines. The generator for SEM with clustered data did not give us the option of specifying categorical or nominal level variables. We add two lines under the VARIABLE command to specify the measurement level of our variables. The default estimator for SEM analyses with categorical variables is WLSMV, so we change MLM to WLSMV. (Mplus syntax will run when an inappropriate estimator is specified, but it will use the appropriate default estimator and notify 1 0

the user of the change in the analysis output.) We also add a line under the ANALYSIS command specifying that we have missing values and we want the analysis to address the missing values with FIML. Then we add a MODEL statement in between the ANALYSIS and OUTPUT commands. In this example we indicate that the five categorical variables load on one latent variable. Specification is discussed in more detail in the next section. Example of Language Generator Syntax After Editing to Make it Complete: TITLE: Example of Edited Syntax: SEM with Clustered Data DATA: FILE IS "F:\Mplus analyses\Langexample\Mydata.dat"; VARIABLE: NAMES ARE age gender ordv1 ordv2 ordv3 ordv4 ordv5 cluster; USEVARIABLES ARE ordv1 ordv2 ordv3 ordv4 ordv5; MISSING ARE ALL (99); CATEGORICAL ARE ordv1 ordv2 ordv3 ordv4 ordv5 ; 5 variables are categorical. NOMINAL IS gender; 1 variables is nominal. ANALYSIS: TYPE IS COMPLEX; TYPE IS MISSING H1; ESTIMATOR IS WLSMV; ITERATIONS = 1000; CONVERGENCE = 0.00005;

We want correction of standard errors appropriate for our multilevel data Missing tells Mplus to use FIML; H1 requests a chi square statistic For categorical data; uses a polychoric correlation matrix Tells Mplus to try up to 1000 times to achieve a converged solution (default) Defines convergence as one minimization value being this close to a prior iterations minimization value

MODEL: Factor1 by ordv1 ordv2 ordv3 ordv4 ordv5;

Specifies the model to be tested: as our 5 observed variables loading on one new latent variable (Factor 1)

OUTPUT: STANDARDIZED;

Requests standardized parameter estimates in addition to default output

1 1

C.3 Model Specification Model specification occurs in the MODEL command. Here we provide annotated examples of a number of common types of CFA and general SEM models. Variations in the formatting of the code are presented to illustrate their equivalence. For example, ARE, IS and = are equivalent. Commands can be spelled out in full or, in most cases, just the first four letters can be provided. If the default setting of a command is desired, the user does not have to specify the option. For example, for one-level CFA and SEM, the type of analysis is GENERAL. We do not have to include TYPE=GENERAL in the syntax, because it is the default. If there are missing data, we use the line TYPE=MISSING H1 to let the program know that we want FIML and a chi square value to be calculated even though it can increase computation time. Mplus is not case sensitive. Also, note that Mplus will save output in an output file with the same name as an input file. If you change a model and want to save a new output file, save the changed input file under a new name. Examples of Mplus Syntax for Measurement and General Structural Models Example 1. 3-factor CFA with 9 continuous, normally distributed observed variables, no missing values Measurement Model A. This is the CFA model specified in the measurement model examples below unless otherwise indicated:

Place Figure 1 about here Mplus Code TITLE: Measurement Model of Job Satisfaction and Commitment, Continuous, Normal Variables, No Missing Values DATA: FILE IS "C:\Mplus\jobdata.dat"; VARIABLE: NAMES ARE 1 2 Annotation

Mplus reads text files; the default is free format, so it is not specified here.

agency sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3 age gender; USEVARIABLES ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; ANALYSIS: ESTIMATOR IS ML ITERATIONS = 1000; CONVERGENCE = 0.00001;

The dataset contains 12 variables. Only 9 of the variables will be used in this analysis. ML is the default; appropriate for interval or higher normally-distributed variables These are the defaults. The number of iterations or the convergence criterion can be increased if the model has trouble converging. The model is a 3-factor model with 3 observed indicators loading on each factor. BY indicates that JOBSATIS is being measured by SAT1-SAT3. Latent factors are named here for the first time. The observed variable names must match the NAMES ARE and USEVARIABLE lines above. Correlations among the three factors are assumed (default). Standardized parameter estimates, modifications indices that are .5 or higher and the residual matrix are requested. Fit indices are provided by default.

MODEL: JOBSATIS LIKESUPR JOBCOMM

BY sat1 sat2 sat3 ; BY sup1 sup3; BY com1 - com3;

OUTPUT: STANDARDIZED MODINDICES (.5) RESIDUAL;

Note: By default, the loading of the first indicator of a factor (first variable listed after BY) is fixed at 1; covariances between pairs of exogenous variables are freely estimated, and error paths are fixed at 1. To change the default setting of the first loading of JOBSATIS to 1, you make a notation in the MODEL command to free the first loading and a notation to fix an alternative loading at 1 (here the loading of sat2 is fixed at 1). For example: JOBSATIS BY sat1* sat2@1 sat3; The * indicates that the default fixing of sat1s value to 1 is changed, and the @ 1 specifies that the loading of sat2 is fixed at 1 instead. These specifications occur in the matrix. 1 3

To set the metric of JOBSATIS by fixing its variance to 1 instead of one of its loadings, similar code is used: JOBSATIS BY sat1* sat2 sat3; JOBSATIS @ 1; To change the default setting of an inter-factor correlation to 0, a WITH statement is added to the model. For example: JOBSATIS WITH JOBCOMM @ 0; This code specifies that the correlation between the two factors is 0. Inter-factor correlations are found on the off-diagonals in the matrix.

Example 2. 3-factor CFA with 9 continuous, normally distributed observed variables, and missing values Mplus Code TITLE: Measurement Model of Job Satisfaction and Commitment with Continuous, Normal Variables, and Missing Values DATA: FILE IS "C:\Mplus\jobdata.dat"; VARIABLE: NAMES ARE agency sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3 age gender; USEVARIABLES ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; MISSING ARE ALL (99); ANALYSIS: TYPE IS MISSING H1 ; ESTIMATOR IS ML; ITERATIONS = 1000; CONVERGENCE = 0.00001; MODEL: Same as previous OUTPUT: SAMPSTATS STANDARDIZED MODINDICES (1) RESIDUAL; Annotation

The missing value flag for all variables in the current analysis is 99. The new TYPE line indicates that FIML and a chi square are requested.

Statistics on the input variables, standardized parameters, modifications indices that change by more than 1 are requested, and the residual matrix are requested as output. 1 4

Example 3. 3-factor CFA with 9 continuous, normally distributed observed variables, missing values, and two correlated measurement errors Measurement Model B. The CFA model specified in Example #3 has two correlated measurement errors as pictured here: Place Figure 2 about here

The code for this example is the same as the code for the previous example except for the addition of two lines to the MODEL statement. Only part of the code is presented.

Mplus Code TITLE: Measurement Model of Job Satisfaction and Commitment with Missing Values, Continuous Variables, Two Correlated Errors DATA: FILE IS "C:\Mplus\jobdata.dat"; MODEL: JOBSATIS BY sat1 sat2 sat3 ; LIKESUPR BY sup1 sup3; JOBCOMM BY com1 - com3; sat1 WITH sat2; com1 WITH com3;

Annotation

Each factor has three observed indicators. One pair of indicators of JOBSATIS and one pair of indicators for JOBCOMM have correlated measurement errors.

The two WITH statements (in conjunction with the BY statements that specify the correlated variables as indicators of latent variables) specify that the two pairs of measurement error variances are correlated (not the variables themselves). The estimated parameters will be in the matrix.

Example 4. 3-factor CFA with 9 continuous, non-normal observed variables, and missing values The code for this example would be the same as the code for the models above except for a change in the ESTIMATOR option. Mplus Code TITLE: Measurement Model of Job Satisfaction and Commitment with Non Normal Continuous 5 Annotation

Variables and Missing Values DATA: FILE IS "C:\Mplus\jobdata.dat"; ANALYSIS: TYPE IS MISSING H1 ; ESTIMATOR IS MLM; ITERATIONS = 1000; CONVERGENCE = 0.00001;

The MLM estimator corrects standard errors and the chi square statistic for non-normality.

Example 5. 3-factor CFA with 9 categorical observed variables, and missing values The code for this example is the same as the code for the models above except for changes in the VARIABLE Command and ESTIMATOR option. Mplus Code TITLE: Measurement Model of Job Satisfaction and Commitment, Categorical Variables and Missing Values, DATA: FILE IS "C:\Mplus\jobdata.dat"; VARIABLE: NAMES ARE agency sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3 age gender; USEVARIABLES ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; MISSING ARE ALL (99); CATEGORICAL ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; ANALYSIS: TYPE IS MISSING H1 ; ESTIMATOR IS WLSMV; ITERATIONS = 1000; CONVERGENCE = 0.00001; MODEL: JOBSATIS BY sat1 sat2 sat3 ; LIKESUPR BY sup1 sup3; JOBCOMM BY com1 - com3; OUTPUT: SAMPSTATS STANDARDIZED MODINDICES (1) RESIDUAL; Annotation

This line indicates that the observed indicators of the latent variable are categorical (ordinal) variables. The WLSMV estimator creates and analyses a matrix of polychoric, tetrachoric, and/or polyserial correlations and an associated weight matrix.

A smaller number of fit indices is available with WLSMV and is provided by default in the output. No confidence interval is printed for RMSEA. 1 6

Note: This code illustrates one proper way to estimate a CFA with ordinal variables. If categorical variables are specified and the ML or MLM estimator is chosen, Mplus will change the estimator to WLSMV (and let you know with a warning in the output).

Example 6. CFA with 9 categorical observed variables, missing values, and clustered data TITLE: Measurement Model of Job Satisfaction and Commitment, Categorical Variables, Missing Values, and Clustered Data DATA: FILE IS "C:\Mplus\jobdata.dat"; VARIABLE: NAMES ARE agency sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3 age gender; USEVARIABLES ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; MISSING ARE ALL (99); CATEGORICAL ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; CLUSTER IS agency;

Dont include agency on the USEVARIABLES line.

ANALYSIS: TYPE IS COMPLEX; TYPE IS MISSING H1 ; ESTIMATOR IS WLSMV; ITERATIONS = 1000; CONVERGENCE = 0.00001; MODEL: JOBSATIS BY sat1 sat2 sat3 ; LIKESUPR BY sup1 sup3; JOBCOMM BY com1 - com3; OUTPUT: STANDARDIZED MODINDICES (.5) RESIDUAL;

Data on individuals are clustered within the variable agency. Agency is not included in the MODEL command; its role is specified here. With TYPE =COMPLEX, standard errors and chi square will be corrected for clustering. FIML will be used to handle missing values; Chi square will be computed. The categorical nature of the data is handled by WLSMV.

Example 7. CFA with Categorical and Clustered Data, and Code for Chi Square Difference Test--1 When users use certain estimators in Mplus, they will see the following notice under the part of the output: The chi-square value for MLM, MLMV, MLR, ULS, WLSM and WLSMV 1 7

cannot be used for chi-square difference tests. Both the and the df are calculated differently with these estimators. Users will note, for example, that the df is not the difference between the number of unique sample moments (input covariance matrix elements) and the number of parameters being estimated. To compare nested models when the above estimators are used in Mplus requires a special process. The user cannot just compare the change in per change in df. First the user runs the less restrictive model, and includes the SAVEDATA command. TITLE: Measurement Model of Job Satisfaction and Commitment, Categorical Variables, Missing Values, Clustered Data, and Difftest SAVEDATA option DATA: FILE IS C:\Mplus\jobdata.dat; VARIABLE: Same as previous ANALYSIS: TYPE IS COMPLEX; TYPE IS MISSING H1 ; ESTIMATOR IS WLSMV; ITERATIONS = 1000; CONVERGENCE = 0.00001; MODEL: Same as previous OUTPUT: STANDARDIZED MODINDICES (.5) RESIDUAL; SAVEDATA: DIFFTEST IS jobsat1;

With TYPE =COMPLEX, standard errors and chi square will be corrected for clustering. FIML will be used to handle missing values. The categorical nature of the data is handled by WLSMV.

A small text file will be saved in the input file directory for use by the program in computing the difference test when the nested model is run.

Example 8. CFA with Categorical and Clustered Data, and Code for Chi Square Difference Test--2 Next the user runs the nested model (with fewer parameters freely estimated than in the previous model). A new line is added to the ANALYSIS options. TITLE: Measurement Model of Job Satisfaction and Commitment, Categorical Variables, Missing Values, Clustered Data, and DIFFTEST Analysis Option 1 8

DATA: FILE IS C:\Mplus\jobdata.dat; VARIABLE: Same as previous ANALYSIS: TYPE IS COMPLEX; TYPE IS MISSING H1 ; ESTIMATOR IS WLSMV; ITERATIONS = 1000; CONVERGENCE = 0.00001; DIFFTEST=jobsat1; MODEL: Any model nested in previous model OUTPUT: STANDARDIZED MODINDICES (.5) RESIDUAL;

The last option under the ANALYSIS command tells the program to use the data saved in file created when the first model was run, to test whether the fit of the more restrictive (second) model is significantly worse than the fit of the first.

The output for the second model will include results of the difference test. If the difference has a p value less than .05, the restrictions added to create the nested model make the fit significantly worse and the first model is retained. If did not become significantly worse (p > .05), the second, more parsimonious, model should be retained.

Example 9. Second Order Factor Model Measurement Model C. The diagram here shows a higher order factor Work Experience that influences the three lower order factors: Place Figure 3 about here

TITLE: Second Order Factor Model with Categorical Variables and Missing Values DATA: FILE IS "C:\Mplus\jobdata.dat"; VARIABLE: NAMES ARE agency sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3 age gender; USEVARIABLES ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; MISSING ARE ALL (99); CATEGORICAL ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; 1 9

ANALYSIS: TYPE IS MISSING H1 ; ESTIMATOR IS WLSMV; ITERATIONS = 1000; CONVERGENCE = 0.00001; MODEL: JOBSATIS BY sat1 sat2 sat3 ; LIKESUPR BY sup1 sup3; JOBCOMM BY com1 - com3; WORKEXP BY JOBSATIS LIKESUPR JOBCOM; Sat1 with sat2; Com1 with com3; OUTPUT: STANDARDIZED MODINDICES (.5) RESIDUAL;

The new line of code names a new latent factor and specifies that it is measured by the first three latent factors. The first order factors become endogenous variables.

Example 10. Multiple Group CFA A measurement model (or general SEM) can be tested to see if it is invariant across groups. Additional information is required in the VARIABLE, ANALYSIS, and MODEL commands. Multiple group model instructions vary depending on a number of variable and model issues. The example here is simple. A more detailed example is provided in the online book resources. Readers are also referred to the online Mplus Users Guide for more information. TITLE: Multiple Group CFA with Categorical Variables and Missing Values DATA: FILE IS "C:\Mplus\jobdata.dat"; VARIABLE: NAMES ARE agency sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3 age gender; USEVARIABLES ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; MISSING ARE ALL (99); CATEGORICAL ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; GROUPING IS gender (0=women 1=men) ANALYSIS: MODEL = NOMEANSTRUCTURE; TYPE IS MISSING H1 ; ESTIMATOR IS WLSMV; 2 0

ITERATIONS = 1000; CONVERGENCE = 0.00001; MODEL: JOBSATIS BY sat1 sat2 sat3 ; LIKESUPR BY sup1 sup3; JOBCOMM BY com1 - com3; Sat1 with sat2; Com1 with com3; MODEL men: sat1 with sat2 @0; com1 with com3@0;

The model statement specifies the model that applies to both groups.

The MODEL men statement indicates where the model differs across the two groups. In this example, the correlated measurement errors do not apply to the mens model.

Example 11. General SEM Specification Most aspects of model specification are the same for general SEM as for CFA. The same reasons for choosing different estimators and analysis options apply. Multiple group and multi-level modeling can be done. See the measurement model examples above for information on syntax that applies to both types of models. Differences will be in the MODEL command, where relationships between factors are specified for general SEMs. After the first general SEM example, just the TITLE and MODEL command sections are presented for different structural models. General SEM A. The model pictured below is based on the measurement model used in the examples above. Note that the structural part of the model is just-identified (df = 0). In a real study, we would have to address this issue in order to test the hypothesized relationships among latent variables. General SEM B shows one way we could make the model over-identified. Place Figure 4 about here

TITLE: General SEM of Job Satisfaction and Commitment with Categorical Variables, Missing Values, Clustered Data DATA: FILE IS C:\Mplus\jobdata.dat; VARIABLE: NAMES ARE 2 1

agency sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3 age gender; USEVARIABLES ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; MISSING ARE ALL (99); CATEGORICAL ARE sat1 sat2 sat3 sup1 sup2 sup3 com1 com2 com3; CLUSTER IS agency; ANALYSIS: TYPE IS COMPLEX; TYPE IS MISSING H1 ; ESTIMATOR IS WLSMV; ITERATIONS = 1000; CONVERGENCE = 0.00001 MODEL: JOBSATIS BY sat1 sat2 sat3 ; LIKESUP BY sup1 sup3; JOBCOMM BY com1 - com3; sat1 WITH sat2; com1 WITH com3; JOBCOMM ON JOBSATIS; JOBCOMM ON LIKESUPR;

The measurement model is specified as before. Here we have included the correlated errors of measurement presented earlier. The ON statements indicate regression relationships among the latent variables. The correlation between JOBSATIS and LIKESUPR is assumed by Mplus.

OUTPUT: STANDARDIZED MODINDICES (.5) RESIDUAL;

12. General SEM with Latent and Observed Predictors General SEM B. The inclusion of gender as a predictor of Job Commitment makes the structural part of the model over-identified (df = 2).

Place Figure 5 about here

TITLE: General SEM of Job Satisfaction and Commitment with Categorical Variables, Missing Values, Clustered Data DATA: FILE IS C:\Mplus\jobdata.dat; MODEL: JOBSATIS BY sat1 sat2 sat3 ; LIKESUP BY sup1 sup3; JOBCOMM BY com1 - com3; 2

The new ON statement indicates that JOBCOMM is also regressed on gender. Mplus knows gender is an observed variable because it is named 2

sat1 WITH sat2; com1 WITH com3; JOBCOMM ON JOBSATIS; JOBCOMM ON LIKESUP; JOBCOMM ON gender;

in the VARIABLE command, and because it is not measured BY any other variables.

13. General SEM with Latent and Observed Predictors and a Mediational Hypothesis General SEM C. In the following SEM, Jobsat has both direct and indirect effects on Jobcomm. Indirect effects are through LikeSupr. Place Figure 6 about here

TITLE: General SEM of Job Satisfaction and Commitment with Categorical Variables, Missing Values, Clustered Data, and Mediation DATA: FILE IS C:\Mplus\jobdata.dat; MODEL: JOBSATIS BY sat1 sat2 sat3 ; LIKESUP BY sup1 sup3; JOBCOMM BY com1 - com3; sat1 WITH sat2; com1 WITH com3; JOBCOMM ON JOBSATIS; JOBCOMM ON LIKESUP; JOBCOMM ON gender; JOBSATIS ON LIKESUPR;

The new ON statement indicates that JOBSATIS is regressed on LIKESUPR. JOBSATIS is both predicted and a predictor in the model. The implication is that the effects of LIKESUP on JOBCOMM are partially mediated by its effect on JOBSATIS.

2 3

Jobsat 1 sat1 1 sat2 1 sat3 1 1 sup1 1

LikeSupr 1 sup2 1 sup3 1 com1 1

Jobcomm

com2 1

com3 1

Figure 1. Measurement Model A

Jobsat 1 sat1 1 sat2 1 sat3 1 1 sup1 1

LikeSupr 1 sup2 1 sup3 1 com1 1

Jobcomm

com2 1

com3 1

Figure 2. Measurement Model B

2 4

WorkExp
1

z
1 sat1 1

1 Jobsat

z
1 sat2 1 sat3 1 sup1 1

1 LikeSupr 1 sup2 1 sup3 1 com1 1

Jobcomm

com2 1

com3 1

Figure 3. Measurement Model C

2 5

1 sat1 1 sat2 1 sat3 1 com3 1 Jobcomm 1 sup1 1 sup2 1 sup3 1 LikeSupr 1 com1 com2 1 1 Jobsat

Figure 4. General SEM A

2 6

1 sat1 1 sat2 1 sat3 1 com3 1 Jobcomm 1 sup1 1 sup2 1 sup3 1 LikeSupr gender 1 com1 com2 1 1 Jobsat

Figure 5. General SEM B

2 7

1 sat1 1 sat2 1 sat3 1 com3 1 Jobcomm 1 sup1 1 sup2 1 sup3 1 LikeSupr gender 1 com1 com2 1 1 Jobsat

Figure 6. General SEM C

2 8