This action might not be possible to undo. Are you sure you want to continue?

**Uni ver si t y I nf or mat i on Technol ogy Ser vi ces
**

Confirmatory Factor Analysis using Amos, LISREL, Mplus,

SAS/STAT CALIS*

Jeremy J. Albright and Hun Myoung Park

© 2006-2009

Last modified on April 2009

University Information Technology Services

Center for Statistical and Mathematical Computing

Indiana University

410 North Park Avenue Bloomington, IN 47408

(812) 855-4724 (317) 278-4740

http://www.indiana.edu/~statmath

*

The citation of this document should read: “Albright, Jeremy J., and Hun Myoung Park. 2009. Confirmatory

Factor Analysis Using Amos, LISREL, Mplus, and SAS/STAT CALIS. Working Paper. The University Information

Technology Services (UITS) Center for Statistical and Mathematical Computing, Indiana University.”

http://www.indiana.edu/~statmath/stat/all/cfa/index.html

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

2

This document summarizes confirmatory factor analysis and illustrates how to estimate

individual models using Amos 16.0, LISREL 8.8, Mplus 5.1, and SAS/STAT 9.1.

**

1. Introduction

2. Exploratory Factor Analysis

3. Confirmatory Factor Analysis

4. Confirmatory Factor Analysis with Missing Data

5. Confirmatory Factor Analysis with Categorical Data

6. Conclusion

1. Introduction

Factor analysis is a statistical method used to find a small set of unobserved variables (also

called latent variables, or factors) which can account for the covariance among a larger set of

observed variables (also called manifest variables). A factor is an unobservable variable that is

assumed to influence observed variables. Scores on multiple tests may be indicators of

intelligence (Spearman, 1904); political liberties and popular sovereignty may measure the

quality of a country’s democracy (Bollen, 1980); or issue emphases in election manifestos may

signify a political party’s underlying ideology (Gabel & Huber, 2000). Factor analysis is also

used to assess the reliability and validity of measurement scales (Carmines & Zeller, 1979).

Principle component analysis also reduces the number of variables, but it differs from principle

factor analysis (Brown, 2006: 22). A factor (unobserved latent variable) is assumed to exert

causal influence on observed variables, while the underlying causal relationship is reversed in

principle component analysis; observed variables are linear combinations of latent variables in

factor analysis, while principle components are (weighted) linear combinations of observed

variables (Hatcher, 1994: 9-10, 69). Principle components account for total variance, while

factors account for the common variance (as opposed to unique variance) of a total variance

(Brown, 2006: 22; Hatcher, 1994: 69).

1.1. Exploratory versus Confirmatory Factor Analysis

Investigators wish to explore patterns in the data or to test explicitly stated hypotheses.

Exploratory factor analysis (EFA), corresponding to the former task, imposes no substantive

constraints on the data; there is no restrictions on the pattern of relationships between observed

and latent variables. EFA is data driven (Brown 2006: 14). Each common factor is assumed to

affect every observed variable and that the common factors are either all correlated or

uncorrelated. Once model is estimated, factor scores, proxies of latent variables, are calculated

and used for follow-up analysis.

1

General purpose statistical software packages such as SPSS,

SAS, and Stata can perform EFA.

**

Jeremy alone wrote the first edition in 2006 and then Hun revised introduction and added exploratory factor

analysis and SAS/STAT CALIS sections to later editions.

1

A factor score is a linear composite of the optimally-weighted observed variables, while a factor-based score is

merely a linear composite (e.g., mean or sum) of the variables that demonstrated meaningful factor loadings

(Hatcher, 1994: 31).

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

3

Confirmatory factor analysis (CFA), on the other hand, is theory- or hypothesis driven. With

CFA it is possible to place substantively meaningful constraints on the factor model.

Researchers can specify the number of factors or set the effect of one latent variable on

observed variables to particular values. CFA allows researchers to test hypotheses about a

particular factor structure (e.g., factor loading between the first factor and first observed

variable is zero). Unlike EFA, CFA produces many goodness-of-fit measures to evaluate the

model but do not calculate factor scores. CFA requires special purpose software packages such

as Mplus, LISREL, Amos, EQS, and SAS/STAT CALIS.

Table 1. Explanatory and Confirmatory Factor Analysis

EFA (Data-driven) CFA (Theory-driven)

Constraint N/A Yes

Unstandardized solution N/A Yes

Standardized solution Yes Yes

Factor rotation Yes N/A

Factor scores Yes N/A

Hypothesis test N/A Yes

Goodness-of-fit N/A Yes

Software package General purpose software Mplus, LISREL, Amos, EQS, SAS CALIS

In fact, CFA is a special case of the structural equation model (SEM), also known as the

covariance structure (McDonald, 1978) or the linear structural relationship (LISREL) model

(Jöreskog & Sörbom, 2004). SEM consists of two components: a measurement model linking a

set of observed variables to a usually smaller set of latent variables and a structural model

linking the latent variables through a series of recursive and non-recursive relationships. CFA

corresponds to the measurement model of SEM. Table 1 summarizes differences and

similarities of EFA and CFA.

1.2. Model Specification and Path Diagram

It is common to display confirmatory factor models as path diagrams in which squares

represent observed variables and circles represent the latent variables. Figure 1 has two latent

variables ξ

1

and ξ

2

in circles that are manifested by six observed variables x

1

through x

6

in

squares. Single-headed arrows are used to imply a direction of assumed causal influence, and

double-headed arrows represent covariance between two latent variables. Latent variables

“cause” the observed variables, as shown by the single-headed arrows pointing away from the

circles and towards the manifest variables.

Table 2. Notation for Confirmatory Factor Analysis

Name Symbol Matrix Form Description

Ksi

ξ

Latent variable

x x X Observed variable

Lambda

λ

Λ Factor loading

Phi

ϕ

Φ Factor variance and covariance

Theta delta

δ

Θ

δ

Error variance and covariance

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

4

The circles labeled ξ (ksi) represent latent variables or (common) factors. A factor can point to

more than one observed variable; in Figure 1, ξ

1

causes three observed variables x

1

through x

3

and ξ

2

influences x

3

through x

6

. The two ξ

i

are expected to covary, as represented by ϕ

21

on the

two-headed arrow. Factor loadings are represented by λ

ij

; λ

31

is, for example, the effect

(regression slope) of ξ

1

on x

3

. The squared factor loading

2

ij

ì is referred to as a communality

representing the proportion of variance in the ith observed variable that is explained by the jth

latent variable (Brown, 2006: 61). The circles labeled δ

i

(delta) represent unique factors

because they affect only a single observed variable. The δ

i

incorporate all the variance in each

x

i

, such as measurement error, which is not captured by the common factors. Finally, error in

the measurement of x

3

is expected to correlate to some extent with measurement error of x

6

, as

represented by δ

63

.

2

Table 2 summarizes CFA notation discussed so far.

Figure 1: Path Diagram of a Confirmatory Factor Model

When observed and latent variables are mean centered to have deviations from their means, the

confirmatory factor model can be summarized by the equation

X = Ac + o

in which X is the vector of observed variables, Λ (lambda) is the matrix of factor loadings

connecting the ξ

i

to the x

i

, ξ is the vector of common factors, and δ is the vector of unique

factors. It is assumed that the error terms have a mean of zero, E(δ) = 0, and that the common

and unique factors are uncorrelated, E(ξδ’)=0. Equation 1 can be rewritten for Figure 1 as:

x

1

= ì

11

c

1

+ o

1

x

2

= ì

21

c

1

+ o

2

x

3

= ì

31

c

1

+ o

3

x

4

= ì

42

c

2

+ o

4

x

5

= ì

52

c

2

+ o

5

x

6

= ì

62

c

2

+ o

6

Here the similarities with regression analysis are evident. Each x

i

is a linear function of one or

more common factors plus an error term (there is no intercept since the variables are mean

centered). The primary difference between these factor equations and regression analysis is that

2

This may occur, for example, with panel data in which ξ

1

and ξ

2

represent the same concept measured at

different points in time; if there is measurement error at t

1

it is likely that there will be measurement error at t

2

.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

5

the ξ

i

are unobserved in CFA. Consequently, estimation proceeds in a manner distinct from the

conventional approach of regressing each x on the ξ

i

.

1.3. Identification

One essential step in CFA is determining whether the specified model is identified. If the

number of the unknown parameters to be estimated is smaller than the number of pieces of

information provided, the model is underidentified. For example, the equation 10 = 2x + 3y is

not identified because it has two unknowns but only one piece of information (one equation).

That is, an infinite number of values for x and y could make the equation true; the equation is

not solvable. To make it just-identified, another independent equation should be provided; for

example, adding 3 = x + y ends up with x=-1 and y=4. Provision of more than one independent

equation will make it overidentified.

In CFA, a model is identified if all of the unknown parameters can be rewritten in terms of the

variances and covariances of the x variables.

3

Unknown parameters of the CFA in Figure 1 are

ϕ

21

, six λ

ij

, six δ

i

, and δ

63

. Information provided is variances and covariances of observed

variables including σ

11

, σ

21

, σ

22

, σ

31

… σ

66

.

66 65 64 63 62 61

55 54 53 52 51

44 43 42 41

33 32 31

22 21

11

o o o o o o

o o o o o

o o o o

o o o

o o

o

The number of input information is 21=6(6+1)/2=p(p+1)/2, where p is the number of observed

variables. Degrees of freedom are 7=21 (knowns) -14 (unknowns); this CFA is overidentified.

Without introducing some constraints any confirmatory factor model is not identified. The

problem lies in the fact that the latent variables are unobserved and hence their scales are

unknown. To identify the model, it therefore becomes necessary to set the metric of the latent

variables in some manner. The two most common constraints are to set either the variance of

the latent variable or one of its factor loadings to one.

1.4. Estimation

When the x variables are measured as deviations from their means it is easy to show that the

sample covariance matrix for x, represented by S, can be decomposed as follows:

Θ Λ ΛΦ Σ + ' =

3

A full discussion of the topic in the context of CFA is available in Bollen (1989, chapter 7), including some

necessary and sufficient conditions for identification.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

6

where Φ (phi) represents the covariance matrix of the ξ factors and Θ (theta) represents the

covariance matrix of the unique factors δ (Bollen, 1989: 236). Estimation proceeds by finding

the parameters Λ

ˆ

, Φ

ˆ

, and Θ

ˆ

so that predicted x covariance matrix Σ (sigma) is as close to

the sample covariance matrix S as possible. Several different fitting functions exist for

determining the closeness of the implied covariance matrix to the sample covariance matrix, of

which maximum likelihood is the most common.

This document includes examples using maximum likelihood estimation (MLE), including Full

Information Maximum Likelihood (FIML) for situations in which there are missing values in

the raw data file. However, MLE assumes multivariate normality among the observed variables,

and preliminary diagnostics of sample data show strong deviations from normality for several

of the variables. Alternative estimators exist for cases of non-normal data but for the most part

lie outside the limited scope of this document. This document will also describe a weighted

least squares (WLS) approach suitable for situations in which the x variables are categorical.

1.5. Goodness of Fit

A large class of omnibus tests exists for assessing how well the model matches the observed

data. χ

2

is a classic goodness-of-fit measure to determine overall model fit. The null hypothesis

is that the implied or predicted covariance matrix Σ is equivalent to the observed sample

covariance matrix S, Σ=S. A large χ

2

and rejection of the null hypothesis means that model

estimates do not sufficiently reproduce sample covariance; the model does not fit the data well.

By contrast, a small χ

2

and failure to reject the null hypothesis is a sign of a good model fit.

However, the χ

2

test is widely recognized to be problematic (Jöreskog, 1969). It is sensitive to

sample size, and it becomes more and more difficult to retain the null hypothesis as the number

of cases increases. The χ

2

test may also be invalid when distributional assumptions are violated,

leading to the rejection of good models or the retention of bad ones. χ

2

is based on a very

stringent hypothesis of Σ=S (Brown 2006: 81).

Due to these drawbacks of χ

2

test many alternative fit statistics have been developed, though

each has its own advantages and disadvantages.

4

Another commonly reported statistic is the

Root Mean Square Error of Approximation (RMSEA), a measure of fit introduced by Steiger

and Lind (1980). RMSEA “incorporates a penalty function for poor model parsimony” and thus

becomes sensitive to the number of parameters estimated and relatively insensitive to sample

size (Brown 2006: 83-84). The Amos User’s Guide suggests that “a value of the RMSEA of

about 0.05 or less would indicate a close fit of the model in relation to the degrees of freedom,”

although “this figure is based on subjective judgment” and “cannot be regarded as infallible”

(Arbuckle, 2005: 496). The Akaike Information Criterion (Akaike, 1987) and Schwarz’s

Bayesian Information Criterion (Schwartz, 1978) can be also used to compare models with

respect to model parsimony.

Comparative fit index (CFI) evaluates “the fit of a user-specified solution in relation to a more

restricted, nested baseline model,” in which the “covariances among all input indicators are

fixed to zero” or no relationship among variables is posited (Brown 2006: 84). CFI ranges from

4

Appendix C of the Amos User’s Guide provides summaries of many different fit measures (Arbuckle, 2005).

For a thorough discussion of different tests see Bollen and Long’s (1993) edited volume.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

7

0 for a poor fit to 1 for a good fit. Tucker-Lewis index (TLI) is another index for comparative fit

that “includes a penalty function for adding freely estimated parameters” (Brown 2006: 85).

TLI can be interpreted in a similar fashion as CFI, but it can have a value outside of the range

of 0 to 1 (p. 86).

There is no single evaluation rule on which everyone agrees. Hu and Bentler (1999) provide

rules of thumb for deciding which statistics to report and choosing cut-off values for declaring

significance. When RMSEA values are close to .06 or below and CFI and TLI are close to .95

or greater, for example, the model may have a reasonably good fit. Therefore, it is

recommended to report not only χ

2

but RMSEA and CFI/TLI.

1.6. Software Issues

This document considers estimating confirmatory factor models using Amos 7.0 (Arbuckle,

2005); LISREL 8.8 (Jöreskog & Sörbom, 2004), and Mplus 5.1 (Muthén & Muthén, 2006).

CFA and SEM can also be estimated using the CALIS procedure in SAS. All four programs are

supported by the Stat/Math Center at Indiana University, while EQS, another popular SEM

program, is currently not supported.

Mplus provides a variety of useful information in a concise manner. Mplus, LISREL, and SAS

CALIS need a program describing a model to be estimated, while Amos supports both the point

and click method and the program approach. Mplus and SAS CALIS use simple syntax

structure, while LISREL (PRELIS, SIMPLIS, and LISREL) and Amos have a relatively

abstruse grammar. Amos and LISREL produce a path diagram but Mplus and SAS CALIS do

not. LISREL is able to go back and forth between a program and a path diagram, but Amos is

not. To sum, Mplus and LISREL are generally recommended for confirmatory factor analysis.

Table 3. Comparison of Amos, LISREL, Mplus, and SAS/STAT CALIS

Amos LISREL Mplus SAS CALIS

Estimation Amos Graphics,

Program Editor

SIMPLIS, LISREL,

path diagram

Program Program

Path diagram Yes Yes No No

Data format supported SPSS Many formats ASCII text Many formats

Syntax (language) Visual Basic, C# LISREL, PRELIS,

SIMPLIS

Mplus SAS CALIS

Output Messy Normal Concise Messy

Platform supported Windows Windows, UNIX Windows Windows, UNIX

Unstandardized estimates Yes Yes Yes Yes

Standardized estimates

and R

2

Yes w/o s.e. Yes w/o s.e. on the

path diagram only

Yes Yes w/o s.e.

Covariances of factors Yes w/o s.e. Yes w/o s.e. Yes Yes w/o s.e.

Correlations of factors Yes w/o s.e. Yes w/o s.e. Yes Yes w/o s.e.

Goodness-of-fit Many Many Several Many

Residual (error) variances Yes Yes Yes Yes

Modification indices Yes only in Amos

Graphics

Yes Yes No

This document provides step-by-step examples for conducting a CFA with commonly used

statistical software packages: Amos, LISREL, Mplus, and SAS/STAT CALIS. The next section

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

8

provides an example of EFA with six observed variables. Section 3 begins with two-factor

CFA with six observed indicators. Section 4 extends Section 3 to cover cases involving missing

data. Section 5 discusses the commonly encountered situation in which the observed variables

are categorical rather than continuous. Section 6 provides a brief summary.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

9

2. Exploratory Factor Analysis

Before moving on to CFA, let us discuss sample data used in this document and briefly review

exploratory factor analysis to contrast it with confirmatory factor analysis.

2.1 Sample Data

In politics commentators often use the terms left and right to describe the ideological positions

of politicians and voters, but it is not always clear what exactly these terms mean. In the United

States the political left is generally associated with favoring greater government involvement in

the economy while the right is understood to favor market autonomy. Yet on moral issues such

as abortion, assisted suicide, and gay marriage it is often the political right that favors a

stronger regulatory role for government. Does a single dimension of values underlie

Americans’ views on both economic and moral issues? Or are there in fact two distinct value

dimensions that underlie citizen attitudes?

This example uses data from the American sample of the European Values Survey (European

Values Group and World Values Survey Association, 2005) to determine whether a model with

one or two common latent factors adequately describes attitudes on economic and moral issues.

The survey queried a random sample of 1,200 respondents about their economic, political, and

moral values. Three questions summarizing economic attitudes and three questions

summarizing moral attitudes, all measured on 10point scales, will be analyzed.

The economic items asked respondents if they felt private ownership of industry should be

increased (PRIVTOWN), if the government should take more responsibility to see that all

people are provided for (GOVTRESP), and whether competition brings out the best or worst in

people (COMPETE). The moral items asked respondents how they felt about homosexuality

(HOMOSEX), legalized abortion (ABORTION), and assisted suicide (EUTHANAS).

For this section missing data is handled by listwise deletion (all cases with missing

observations on any indicator are removed). Listwise deletion resulted in dropping 40 of the

original 1,200 observations, leaving a sample size of 1,160. The data is saved as the SPSS file

values.sav located in the folder C:\temp\CFA. The data set has six variables and looks like the

following:

+---------------------------------------------------------------+

| privtown govtresp compete homosex abortion euthanas |

|---------------------------------------------------------------|

1. | 1 3 2 2 2 2 |

2. | 2 1 2 1 1 1 |

3. | 7 3 3 4 3 3 |

4. | 4 3 1 6 7 1 |

5. | 8 1 1 1 5 1 |

6. | 8 3 4 3 4 4 |

7. | 6 6 6 1 1 1 |

2.2 An Explorative Factor Analysis

EFA seeks a smaller number of latent variables to explain variance and covariance among

manifest variables. There are many methods to extract factors, such as principal factor (PF),

maximum likelihood (ML), weighted least squares (WLS), generalized least squares (GLS),

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

10

etc., of which PF and ML are most commonly used. Researchers may need to determine the

number of factors extracted using eigenvalues calculated from input correlation matrix. As a

rule of thumb, count the number of eigenvalues greater than 1 and use it as the number of

factors (Brown, 2006: 26; Hatcher, 1994: 22-23). Alternatively, researchers may use scree test

and parallel analysis (Brown 2006: 26-30). In the following output, there are two positive

eigenvalues only one of which is greater than 1.

--------------------------------------------------------------------------

Factor | Eigenvalue Difference Proportion Cumulative

-------------+------------------------------------------------------------

Factor1 | 1.18970 0.55561 0.9686 0.9686

Factor2 | 0.63409 0.65191 0.5163 1.4849

Factor3 | -0.01782 0.11383 -0.0145 1.4704

Factor4 | -0.13165 0.07293 -0.1072 1.3632

Factor5 | -0.20458 0.03691 -0.1666 1.1966

Factor6 | -0.24149 . -0.1966 1.0000

--------------------------------------------------------------------------

Once factors are extracted, researchers may rotate them to “foster interpretability by

maximizing factor loadings close to 1.0 and minimizing factor loadings close to 0” (Brown

2006: 31). Factor rotation is either orthogonal or oblique. Factors are constrained to be

uncorrelated in orthogonal rotation but not necessarily in oblique rotation. Varimax

(orthogonal) and promax (oblique) rotations are commonly used. Keep in mind that any factor

rotation does not alter factor loadings but change views of pattern matrix. Finally, researchers

may calculate factor scores for future analysis.

-------------------------------------------------

Variable | Factor1 Factor2 | Uniqueness

------------+--------------------+---------------

privtown | -0.0203 0.5494 | 0.6977

govtresp | 0.2070 0.1619 | 0.9309

compete | 0.0231 0.5547 | 0.6918

homosex | 0.6356 -0.0145 | 0.5958

abortion | 0.6572 -0.0059 | 0.5681

euthanas | 0.5514 0.0642 | 0.6919

-------------------------------------------------

The output above suggests two latent variables underlying six manifest variables. The numbers

under first and second columns are factor loadings. Factor loadings of factor 1 and 2 on the

perception on competition (COMPETE) are .0231 and .5547. The squared factor loadings,

communalities, of .0005=.0231

2

and .3077=.5547

2

are respectively the proportions of variance

in COMPETE that is explained by factor 1 and 2. That is, 31 percent of variance in COMPETE

is explained by factor 2 and almost zero percent by factor 1. Unique variance is the proportion

of variance that is not explained by any factor. For example, .6918 is calculated as 1-

(.0005+.3077); the 69 percent of variance in COMPETE is not explained by two factors.

PRIVTOWN (private ownership) and COMPETE are largely explained by factor 2, while

HOMOSEX (homosexuality), ABORTION, and EUTHANAS (assisted suicide) by factor 1.

However, neither factor 1 nor 2 can explain GOVTRESP (government responsibility)

sufficiently; it has the largest unique variance of 93 percent. Therefore, there appear to be two

moral dimensions underlying citizen attitudes.

Two factor scores predicted here have zero mean and standard deviations of .7871 and .6616.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

11

3. Confirmatory Factor Analysis

This section begins with one factor model and then moves forward to the two factor model.

3.1 One Factor Model

The following path diagram with standardized solutions illustrates the one factor model, where

a latent variable values is manifested by six observed variables.

5

An oval and a rectangle

represent a latent variable and a manifest variable, respectively. The numbers on arrows from

the latent variable to observed variables are standardized factor loadings (regression weights).

COMPETE and PRIVTOWN have large factor loadings of .72 and .58; they appear to be the

best indicators of values. R

2

is a standardized factor loading squared that means the extent that

a factor can explain the variance in a manifest variable. For example, the latent variable values

explains about 51 percent (=.72

2

) of variance in COMPETE. HOMOSEX, ABORTION, and

EUTHANAS have poor factor loadings, suggesting that they appear to indicate other factors.

The following is the LISREL output of this confirmation factor model. The coefficients of

values listed under the Measurement Equations heading are unstandardized factor loadings.

The numbers in parentheses are standard errors followed by test statistics. For instance, the

factor loading on COMPETE is 1.31; its standard error is.32; and the test statistic is 4.08

(=1.31/.32). R

2

is listed at the end of each equation.

DATE: 11/18/2008

TIME: 13:00

L I S R E L 8.80

BY

Karl G. Jöreskog & Dag Sörbom

This program is published exclusively by

Scientific Software International, Inc.

5

See the LISREL section for details about fitting a confirmatory factor model and drawing a path diagram.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

12

7383 N. Lincoln Avenue, Suite 100

Lincolnwood, IL 60712, U.S.A.

Phone: (800)247-6113, (847)675-0720, Fax: (847)675-2140

Copyright by Scientific Software International, Inc., 1981-2006

Use of this program is subject to the terms specified in the

Universal Copyright Convention.

Website: www.ssicentral.com

The following lines were read from file c:\Temp\cfa\values.spj:

One factor model

Raw Data from file 'c:\Temp\cfa\values.psf'

Sample Size = 1160

Latent Variables Values

Relationships

privtown = 1.00*Values

govtresp = Values

compete = Values

homosex = Values

abortion = Values

euthanas = Values

Path Diagram

End of Problem

Sample Size = 1160

Second Two factor model

Covariance Matrix

privtown govtresp compete homosex abortion euthanas

-------- -------- -------- -------- -------- --------

privtown 5.05

govtresp 0.51 7.20

compete 2.27 0.67 5.71

homosex -0.22 1.39 0.01 10.57

abortion -0.11 1.01 0.06 4.81 8.85

euthanas 0.14 0.72 0.38 3.00 3.05 6.03

Second Two factor model

Number of Iterations = 8

LISREL Estimates (Maximum Likelihood)

Measurement Equations

privtown = 1.00*Values, Errorvar.= 3.35 , R² = 0.34

(0.44)

7.67

govtresp = 0.33*Values, Errorvar.= 7.01 , R² = 0.026

(0.083) (0.30)

4.00 23.64

compete = 1.31*Values, Errorvar.= 2.78 , R² = 0.51

(0.32) (0.72)

4.08 3.85

homosex = 0.11*Values, Errorvar.= 10.55, R² = 0.0020

(0.094) (0.44)

1.18 24.04

abortion = 0.12*Values, Errorvar.= 8.82 , R² = 0.0030

(0.086) (0.37)

1.44 24.03

euthanas = 0.21*Values, Errorvar.= 5.96 , R² = 0.012

(0.073) (0.25)

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

13

2.81 23.89

Variances of Independent Variables

Values

--------

1.70

(0.44)

3.84

Goodness of Fit Statistics

Degrees of Freedom = 9

Minimum Fit Function Chi-Square = 641.04 (P = 0.0)

Normal Theory Weighted Least Squares Chi-Square = 718.06 (P = 0.0)

Estimated Non-centrality Parameter (NCP) = 709.06

90 Percent Confidence Interval for NCP = (624.84 ; 800.67)

Minimum Fit Function Value = 0.55

Population Discrepancy Function Value (F0) = 0.61

90 Percent Confidence Interval for F0 = (0.54 ; 0.69)

Root Mean Square Error of Approximation (RMSEA) = 0.26

90 Percent Confidence Interval for RMSEA = (0.24 ; 0.28)

P-Value for Test of Close Fit (RMSEA < 0.05) = 0.00

Expected Cross-Validation Index (ECVI) = 0.64

90 Percent Confidence Interval for ECVI = (0.57 ; 0.72)

ECVI for Saturated Model = 0.036

ECVI for Independence Model = 0.83

Chi-Square for Independence Model with 15 Degrees of Freedom = 949.26

Independence AIC = 961.26

Model AIC = 742.06

Saturated AIC = 42.00

Independence CAIC = 997.59

Model CAIC = 814.73

Saturated CAIC = 169.18

Normed Fit Index (NFI) = 0.32

Non-Normed Fit Index (NNFI) = -0.13

Parsimony Normed Fit Index (PNFI) = 0.19

Comparative Fit Index (CFI) = 0.32

Incremental Fit Index (IFI) = 0.33

Relative Fit Index (RFI) = -0.13

Critical N (CN) = 40.17

Root Mean Square Residual (RMR) = 1.45

Standardized RMR = 0.17

Goodness of Fit Index (GFI) = 0.83

Adjusted Goodness of Fit Index (AGFI) = 0.60

Parsimony Goodness of Fit Index (PGFI) = 0.36

The Modification Indices Suggest to Add an Error Covariance

Between and Decrease in Chi-Square New Estimate

compete privtown 204.2 42.00

homosex privtown 8.2 -0.62

homosex govtresp 28.1 1.35

abortion govtresp 17.0 0.96

abortion homosex 285.9 4.80

euthanas govtresp 10.3 0.62

euthanas homosex 162.9 2.99

euthanas abortion 201.6 3.04

Time used: 0.000 Seconds

A various goodness-of-fit statistics are listed under Goodness of Fit Statistics. χ

2

is 718.06,

which is so large that the null hypothesis of a good fit is rejected at the .05 level (p<.000). The

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

14

degrees of freedom is 9 =21 - 12; there are 21 sample variance and covariance elements and 12

unknown parameters including six λ

ij

and six δ

i

. Root Mean Square Error of Approximation

(RMSEA) 0.26 is also large enough to reject the null hypothesis (p<.000). Comparative Fit

Index (CFI) 0.32 is small. Therefore, this one factor model shows a poor fit and needs to be

modified somehow.

Modification indices at the bottom suggest that HOMOSEX, ABORTION, and EUTHANAS

are closely related (moral values), while PRIVTOWN and COMPETE are grouped together

(economic values). For instance, if you add covariance between HOMOSEX and ABORTION

χ

2

will decrease by 285.9; two variables appear to manifest the same latent variable. Small R

2

of HOMOSEL (.002), ABORTION (.003), and EUTHANAS (.012) support this conclusion

since the one factor explains practically no variance in these observed variables. It is likely that

a two factor model is more appropriate to describe the economic and moral values of

Americans It is not clear, however, whether GOVTRESP falls into economic values or moral

values; let us first connect to economic values though.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

15

3.2 Two Factor Model using LISREL

This section demonstrates how to estimate a two factor confirmatory factor model using

LISREL. Note that the previous subsection revealed that the single common factor model was a

poor fit to the data.

LISREL can be launched from any computer running Windows in the UITS Student

Technology Centers by going to Start → All Programs → Departmentally Sponsored →

Statistics-Math → LISREL 8.80 → LISREL 8.80.

First, you need to import the data

file into LISREL and save it as

a .psf (PRELIS system) file.

PRELIS, the pre-processor to

LISREL, can read data files from

a number of statistical programs,

including SAS, Stata, and SPSS.

To open the SPSS file values.sav

saved in the C:\temp\CFA folder,

go to File → Import Data. The

Open dialog box opens. Change

Files of Type to SPSS Data

File(*.sav), navigate to the

correct folder, and click on

values.sav.

Click Open. You will then be prompted to save the data as a .psf file. Name the file values and

click Save to store it in the working directory. A spreadsheet with the raw data will display.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

16

LISREL allows you to construct a path diagram of the model and then generate a PRELIS or

LISREL program based on the path diagram. Go to File → New and choose Path Diagram.

You will be immediately prompted to save the path diagram. Name the file values and click

Save (the .pth extension will differentiate this file from the other files named values in the

working directory). An empty window opens where you will eventually draw the diagram.

The next step is to name the variables that will be in the model. Go to Setup → Title and

Comments to open the Title and Comments dialog box. Enter First Two Factor Model in the

Title field and click Next.

The Group Names box opens, which is used to label different groups when comparing models

for multiple independent samples. Because we are interested only in the single sample of

American respondents we can skip this box by clicking Next.

The Labels dialog box then opens, which is used to identify the latent and observed variables

to be analyzed. Currently no variables have been selected. To choose variable names click on

Add/Read Variables.

This opens a new dialog box used to locate the PRELIS system file. Verify that the Read from

file radio button is chosen and pick PRELIS System File from the drop-down menu. Then

click Browse to choose the PRELIS system file created earlier. Click OK.

The names of the observed variables are now listed in the Labels box. Note that the default

variable CONST appears on the list of observed variables. To add the names of the latent

variables click Add Latent Variables. Enter ECONOMIC in the box that opens. Repeat to

enter the name of the second common factor MORALS. Click OK.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

17

Click Next, and a final dialog box opens.

Raw data from a

PRELIS system file will

be analyzed. If desired,

the data can be viewed

and edited by clicking

on the Edit button.

Because this system file

already contains

information about the

sample size it is not

necessary to make

further changes. Click

OK.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

18

It is now possible to begin drawing the path diagram. The names of the observed and latent

variables appear on the left side of the screen. Drag all of the observed variables to the drawing

pad along with the latent variables ECONOMIC and MORALS.

Next click on the single-headed arrow on the tool bar and connect the ECONOMIC factor to

PRIVTOWN, GOVTRESP, and COMPETE. Also draw arrows from MORALS to HOMOSEX,

ABORTION, and EUTHANAS. Because the usual assumption is that the latent variables

“cause” the observed variables, the arrows should point towards the six indicators. Finally,

draw a two-headed arrow connecting each latent variable.

In LISREL, it is not necessary to draw the unique factors representing measurement error for

each of the observed variables. LISREL includes these by default and automatically sets their

scales by constraining the loadings to one. To set the scale of ECONOMIC, constrain the

regression weight of the PRIVTOWN variable to one. Double-click on the line at the point

where 0.00 appears and change the loading to 1.00. LISREL will not recognize this constraint,

however, unless you then right-click on the loading and choose Fix. Do the same for the path

connecting MORALS to HOMOSEX to set the metric for the second common factor.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

19

The final step before estimation is to build from the path diagram the corresponding syntax

LISREL uses for estimation. There are actually two languages that LISREL understands:

LISREL syntax and SIMPLIS syntax. As its name suggests, SIMPLIS is more straightforward

and easy to read than LISREL syntax. A SIMPLIS syntax file can be built from the path

diagram by choosing Setup → Build SIMPLIS syntax. This opens an editor displaying the

SIMPLIS commands needed to estimate the model.

Click the Run LISREL button to begin estimation. Each time the Run LISREL button is

clicked, a text output file is written to the working directory (extension .out) and a path diagram

is presented. The text output contains unstandardized estimates, test statistics, goodness-of-fit

statistics, modification indices, and other additional information; standardized estimates do not

appear on the text output. It is always a good idea to inspect the output file for any error

messages and, in some cases, warnings that a model may not be identified. For this model the

output file is the following:

DATE: 11/18/2008

TIME: 1:33

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

20

L I S R E L 8.80

BY

Karl G. Jöreskog & Dag Sörbom

This program is published exclusively by

Scientific Software International, Inc.

7383 N. Lincoln Avenue, Suite 100

Lincolnwood, IL 60712, U.S.A.

Phone: (800)247-6113, (847)675-0720, Fax: (847)675-2140

Copyright by Scientific Software International, Inc., 1981-2006

Use of this program is subject to the terms specified in the

Universal Copyright Convention.

Website: www.ssicentral.com

The following lines were read from file c:\temp\cfa\values.spj

First Two factor model

Raw Data from file 'c:\temp\cfa\values.psf'

Sample Size = 1160

Latent Variables Economic Morals

Relationships

privtown = 1.00*Economic

govtresp = Economic

compete = Economic

homosex = 1.00*Morals

abortion = Morals

euthanas = Morals

Path Diagram

End of Problem

Sample Size = 1160

First Two factor model

Covariance Matrix

privtown govtresp compete homosex abortion euthanas

-------- -------- -------- -------- -------- --------

privtown 5.05

govtresp 0.51 7.20

compete 2.27 0.67 5.71

homosex -0.22 1.39 0.01 10.57

abortion -0.11 1.01 0.06 4.81 8.85

euthanas 0.14 0.72 0.38 3.00 3.05 6.03

First Two factor model

Number of Iterations = 5

LISREL Estimates (Maximum Likelihood)

Measurement Equations

privtown = 1.00*Economic, Errorvar.= 3.37 , R² = 0.33

(0.57)

5.88

govtresp = 0.30*Economic, Errorvar.= 7.04 , R² = 0.022

(0.084) (0.30)

3.63 23.70

compete = 1.35*Economic, Errorvar.= 2.65 , R² = 0.54

(0.45) (1.02)

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

21

3.01 2.60

homosex = 1.00*Morals, Errorvar.= 5.86 , R² = 0.45

(0.40)

14.49

abortion = 1.02*Morals, Errorvar.= 3.96 , R² = 0.55

(0.076) (0.37)

13.39 10.63

euthanas = 0.64*Morals, Errorvar.= 4.12 , R² = 0.32

(0.046) (0.22)

13.73 19.13

Covariance Matrix of Independent Variables

Economic Morals

-------- --------

Economic 1.68

(0.58)

2.91

Morals 0.10 4.72

(0.13) (0.49)

0.78 9.67

Goodness of Fit Statistics

Degrees of Freedom = 8

Minimum Fit Function Chi-Square = 42.11 (P = 0.00)

Normal Theory Weighted Least Squares Chi-Square = 41.53 (P = 0.00)

Estimated Non-centrality Parameter (NCP) = 33.53

90 Percent Confidence Interval for NCP = (17.01 ; 57.57)

Minimum Fit Function Value = 0.036

Population Discrepancy Function Value (F0) = 0.029

90 Percent Confidence Interval for F0 = (0.015 ; 0.050)

Root Mean Square Error of Approximation (RMSEA) = 0.060

90 Percent Confidence Interval for RMSEA = (0.043 ; 0.079)

P-Value for Test of Close Fit (RMSEA < 0.05) = 0.16

Expected Cross-Validation Index (ECVI) = 0.058

90 Percent Confidence Interval for ECVI = (0.044 ; 0.079)

ECVI for Saturated Model = 0.036

ECVI for Independence Model = 0.83

Chi-Square for Independence Model with 15 Degrees of Freedom = 949.26

Independence AIC = 961.26

Model AIC = 67.53

Saturated AIC = 42.00

Independence CAIC = 997.59

Model CAIC = 146.26

Saturated CAIC = 169.18

Normed Fit Index (NFI) = 0.96

Non-Normed Fit Index (NNFI) = 0.93

Parsimony Normed Fit Index (PNFI) = 0.51

Comparative Fit Index (CFI) = 0.96

Incremental Fit Index (IFI) = 0.96

Relative Fit Index (RFI) = 0.92

Critical N (CN) = 553.99

Root Mean Square Residual (RMR) = 0.41

Standardized RMR = 0.052

Goodness of Fit Index (GFI) = 0.99

Adjusted Goodness of Fit Index (AGFI) = 0.97

Parsimony Goodness of Fit Index (PGFI) = 0.38

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

22

The Modification Indices Suggest to Add the

Path to from Decrease in Chi-Square New Estimate

govtresp Morals 31.6 0.24

The Modification Indices Suggest to Add an Error Covariance

Between and Decrease in Chi-Square New Estimate

compete privtown 31.6 86.63

homosex govtresp 12.4 0.76

Time used: 0.016 Seconds

The χ

2

statistic for model fit is 42.11, which is large enough to reject the null that the model is a

good fit to the data. We report the Minimum Fit Function χ

2

in order to be consistent with the

output from other software packages. The degrees of freedom is 8 = 21 -13; there are 13

unknown parameters including six λ

ij

, six δ

i

, and ϕ

21

. The Root Mean Square Error of

Approximation (RMSEA) is .060. Using a cut-off rule of .05, the RMSEA is too high to

indicate a good fit. CFI of .96 indicates a moderate fit.

The text output presents unstandardized estimates and their standard errors. It is possible to

ascertain the statistical significance of the estimates by comparing the unstandardized loadings

displayed in the equations under the Measurement Equations heading in the output file with

their standard errors displayed in parentheses. When the unstandardized loadings are at least

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

23

twice the size of the standard errors the estimates are significant at the .05 level. In this case

each of the unconstrained estimates is significant.

The unstandardized estimates also appear in the path diagram by default (see left diagram

below). To view the standardized estimates choose Standardized Solution from the Estimates

drop-down menu (see above screenshot). Note that factor loadings of PRIVTOWN and

HOMOSEX are set 1 in the above diagram for unstandardized solutions, while variances of two

factors are set 1 in the following diagram for standardized solutions.

The standardized loadings represent the correlation between each observed variable and the

corresponding factor. Considering first the indicators of ECONOMIC, they are .58 for

PRIVOWN, .15 for GOVTRESP, and .73 for COMPETE. Considering the indicators of

MORALS, the standardized loadings are .67 for HOMOSEX, .74 for ABORTION, and .56 for

EUTHANAS.

A good deal of the variance in each observed variable, with the exception of GOVTRESP, is

accounted for. The R

2

for PRIVTOWN is .33 (=.58

2

); for COMPETE it is .54 (=.73

2

); for

HOMOSEX it is .45; for ABORTION it is .65; and for EUTHANAS it is .32. Only

GOVTRESP, with its R

2

of .022, does not fit in well with the model. It may be the case that this

survey question taps some kind of value dimension distinct from the economic dimension

measured by the PRIVTOWN and COMPETE variables.

LISREL reports modification indices, both in the path diagram (by choosing Modification

Indices from the Estimation menu) and in the output. These indices make suggestions about

loosening certain model parameters in order to improve the overall model fit. As long as any

decisions made on the basis of modification indices are theoretically meaningful and do not

result in an unidentified model they can be helpful in improving model specification. Three

suggestions are given in the output: add a path from GOVTRESP to MORALS; add error

covariances between HOMOSEX and GOVTRESP, and between PRIVTOWN and COMPETE.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

24

GOVTRESP appears to have something in common with the morality dimension, either by

sharing measurement error with HOMOSEX and COMPETE or as a direct indicator of the

latent morality dimension. The modification index 31.60 says that adding an arrow from

MORALS to GOVTRESP will reduce χ

2

by 31.60. Because the standardized loading of

GOVTRESP on ECONOMIC was so low, it is possible that the item is actually tapping a

different values dimension.

The final model therefore adds a path from MORALS to GOVTRESP. This modification

results in the following standardized solution:

Second Two factor model

Raw Data from file 'c:\Temp\cfa\values.psf'

Sample Size = 1160

Latent Variables Economic Morals

Relationships

privtown = 1.00*Economic

govtresp = Economic Morals

compete = Economic

homosex = 1.00*Morals

abortion = Morals

euthanas = Morals

Path Diagram

End of Problem

The following output is selective, ignoring less informative parts.

Measurement Equations

privtown = 1.00*Economic, Errorvar.= 3.24 , R² = 0.36

(0.62)

5.23

govtresp = 0.28*Economic + 0.24*Morals, Errorvar.= 6.77 , R² = 0.060

(0.082) (0.043) (0.29)

3.46 5.56 23.33

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

25

compete = 1.26*Economic, Errorvar.= 2.86 , R² = 0.50

(0.42) (0.96)

2.98 2.97

homosex = 1.00*Morals, Errorvar.= 5.72 , R² = 0.46

(0.40)

14.33

abortion = 0.99*Morals, Errorvar.= 4.08 , R² = 0.54

(0.072) (0.36)

13.83 11.46

euthanas = 0.63*Morals, Errorvar.= 4.13 , R² = 0.32

(0.045) (0.21)

13.80 19.28

Covariance Matrix of Independent Variables

Economic Morals

-------- --------

Economic 1.81

(0.63)

2.89

Morals 0.04 4.86

(0.13) (0.49)

0.32 9.93

Goodness of Fit Statistics

Degrees of Freedom = 7

Minimum Fit Function Chi-Square = 9.88 (P = 0.20)

Normal Theory Weighted Least Squares Chi-Square = 9.89 (P = 0.19)

Estimated Non-centrality Parameter (NCP) = 2.89

90 Percent Confidence Interval for NCP = (0.0 ; 15.41)

Minimum Fit Function Value = 0.0085

Population Discrepancy Function Value (F0) = 0.0025

90 Percent Confidence Interval for F0 = (0.0 ; 0.013)

Root Mean Square Error of Approximation (RMSEA) = 0.019

90 Percent Confidence Interval for RMSEA = (0.0 ; 0.044)

P-Value for Test of Close Fit (RMSEA < 0.05) = 0.98

Expected Cross-Validation Index (ECVI) = 0.033

90 Percent Confidence Interval for ECVI = (0.030 ; 0.043)

ECVI for Saturated Model = 0.036

ECVI for Independence Model = 0.83

Chi-Square for Independence Model with 15 Degrees of Freedom = 949.26

Independence AIC = 961.26

Model AIC = 37.89

Saturated AIC = 42.00

Independence CAIC = 997.59

Model CAIC = 122.68

Saturated CAIC = 169.18

Normed Fit Index (NFI) = 0.99

Non-Normed Fit Index (NNFI) = 0.99

Parsimony Normed Fit Index (PNFI) = 0.46

Comparative Fit Index (CFI) = 1.00

Incremental Fit Index (IFI) = 1.00

Relative Fit Index (RFI) = 0.98

Critical N (CN) = 2167.38

Root Mean Square Residual (RMR) = 0.12

Standardized RMR = 0.018

Goodness of Fit Index (GFI) = 1.00

Adjusted Goodness of Fit Index (AGFI) = 0.99

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

26

Parsimony Goodness of Fit Index (PGFI) = 0.33

Time used: 0.000 Seconds

This model fits the data well. The χ

2

measure of model fit is 9.88, which is too small to reject

the null hypothesis of a good fit (p<.20). Adding λ

22

reduces degrees of freedom by 1 (df=7).

The RMSEA has declined to .019, which is small enough to indicate a good fit, while CFI is

almost (as oppose to numerically) 1.

The unconstrained loadings are all statistically significant at the .05 level, having estimates that

are more than twice the size of their standard errors (see the output above and left diagram

below). GOVTRESP continues to have a low correlation with the ECONOMIC factor (.14) and

has a similarly low correlation with MORALS (.20) (see right diagram below). However, the

remaining standardized loadings range from .56 (EUTHANAS) to .73 (ABORTION). In

between are PRIVTOWN (.60), HOMOSEX (.68), and COMPETE (.71).

Despite receiving a path from both common factors GOVTRESP continues to have by far the

smallest R

2

(.060). The remaining statistics are moderately well accounted for by the

corresponding factors. The R

2

values are, in order of increasing magnitude, .32 for

EUTHANAS, .36 for PRIVTOWN, .46 for HOMOSEX, .50 for COMPETE, and .54 for

ABORTION. Finally, the correlation between ECONOMIC and MORALS is a negligible -.01

(right path diagram above) and their covariance is .04 (left diagram above).

The conclusion from this analysis is that two nearly orthogonal dimensions underlie the

economic and moral values of American citizens. Additionally it is unclear whether the

GOVTRESP item is tapping either dimension. Future surveys should incorporate more reliable

measures of economic values.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

27

3.3 Two Factor Model using Mplus

This section demonstrates how to estimate a confirmatory factor model using Mplus 5.2.

Because the subsection covering Amos revealed that the single common factor model was a

poor fit to the data, this section will also begin with the two factor model.

To launch Mplus from any Windows machine in the UITS Student Technology Centers go to

Start → All Programs → Departmentally Sponsored → Statistics-Math → Mplus5.2 →

Mplus Editor. This will open Mplus and display the program’s built-in syntax editor.

Unlike Amos and LISREL, Mplus does not allow you to simply draw a path diagram and

estimate the model; you must write the syntax yourself. The Language Generator under the

Mplus menu, however, can make this task a little easier. Mplus reads only ASCII text files with

free or fixed formatted; Mplus cannot directly read a SPSS data set. Nonetheless, Mplus is an

extremely powerful program for estimating a much wider range of models than is possible with

Amos and LISREL, and this example will only scratch the surface of what Mplus can do.

The first model to be examined consists of two common factors and the six observed indicators

of economic and moral values. The Mplus syntax for estimating this model is the following:

TITLE: First Two Factor Model;

DATA: FILE IS c:\temp\cfa\values.dat;

VARIABLE: NAMES ARE privtown govtresp compete

homosex abortion euthanas;

MODEL: economic BY privtown govtresp compete;

morals BY homosex abortion euthanas;

OUTPUT: STANDARDIZED;

MODINDICES;

Save as an input file under the name values1.inp in the same folder as the values.dat file.

The TITLE statement provides a label for the particular analysis that will be run. The DATA

statement specifies where the data file is located at. Absolute path names are only necessary if

the syntax file is located in a directory different from where the data is saved. The

VARIABLE statement provides names for the six observed variables in the raw data file in the

order in which they appear. The MODEL statement specifies the particular model to be

estimated. In this case ECONOMIC is assumed to cause the three observed variables

PRIVTOWN, GOVTRESP, and COMPETE; and MORALS is assumed to cause HOMOSEX,

ABORTION, and EUTHANAS. The OUTPUT statement requests that standardized parameter

estimates and modification indices be included in the output file.

There are a few things to keep in mind when creating Mplus syntax. First, all commands end

with a semicolon; omitting the semicolon will lead to error messages. Second, Mplus cannot

read more than 80 characters in a line. One way to limit this problem is to use very short names

for variables, such as x

1

, x

2

, y

1

, y

2

(longer names are used here to be consistent with the Amos

and LISREL examples later). Commands can take up more than one line, as the semicolon

marks the command end. Finally, Mplus is not case sensitive; capital and lowercase letters can

be used interchangeably.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

28

It is also important to know that the default behavior for setting the scale of the common latent

variable is to constrain the loading for the first variable (in this case PRIVTOWN and

HOMOSEX) to one. This option can be overridden but will not be altered here to keep the

example consistent with the Amos and LISREL examples above.

After entering the syntax and saving it as an Mplus input (.inp) file, estimate the model by

clicking the Run button . This produces a text output (.out) file stored in the working

directory with the results. For this model the output file looks like the following:

Mplus VERSION 5.1

MUTHEN & MUTHEN

11/17/2008 9:58 PM

INPUT INSTRUCTIONS

TITLE: First Two Factor Model;

DATA: FILE IS C:\Temp\CFA\values.dat;

VARIABLE: NAMES ARE privtown govtresp compete

homosex abortion euthanas;

MODEL: economic BY privtown govtresp compete;

morals BY homosex abortion euthanas;

OUTPUT: STANDARDIZED;

MODINDICES;

INPUT READING TERMINATED NORMALLY

Two Factor Model 1;

SUMMARY OF ANALYSIS

Number of groups 1

Number of observations 1160

Number of dependent variables 6

Number of independent variables 0

Number of continuous latent variables 2

Observed dependent variables

Continuous

PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS

Continuous latent variables

ECONOMIC MORALS

Estimator ML

Information matrix OBSERVED

Maximum number of iterations 1000

Convergence criterion 0.500D-04

Maximum number of steepest descent iterations 20

Input data file(s)

C:\Temp\CFA\values.dat

Input data format FREE

THE MODEL ESTIMATION TERMINATED NORMALLY

TESTS OF MODEL FIT

Chi-Square Test of Model Fit

Value 42.147

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

29

Degrees of Freedom 8

P-Value 0.0000

Chi-Square Test of Model Fit for the Baseline Model

Value 891.990

Degrees of Freedom 15

P-Value 0.0000

CFI/TLI

CFI 0.961

TLI 0.927

Loglikelihood

H0 Value -16217.034

H1 Value -16195.961

Information Criteria

Number of Free Parameters 19

Akaike (AIC) 32472.068

Bayesian (BIC) 32568.135

Sample-Size Adjusted BIC 32507.785

(n* = (n + 2) / 24)

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.061

90 Percent C.I. 0.043 0.079

Probability RMSEA <= .05 0.147

SRMR (Standardized Root Mean Square Residual)

Value 0.046

MODEL RESULTS

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 1.000 0.000 999.000 999.000

GOVTRESP 0.304 0.083 3.675 0.000

COMPETE 1.350 0.433 3.120 0.002

MORALS BY

HOMOSEX 1.000 0.000 999.000 999.000

ABORTION 1.018 0.076 13.432 0.000

EUTHANAS 0.637 0.046 13.725 0.000

MORALS WITH

ECONOMIC 0.098 0.124 0.784 0.433

Intercepts

PRIVTOWN 3.553 0.066 53.876 0.000

GOVTRESP 4.312 0.079 54.775 0.000

COMPETE 3.440 0.070 49.048 0.000

HOMOSEX 4.781 0.095 50.095 0.000

ABORTION 4.352 0.087 49.845 0.000

EUTHANAS 2.638 0.072 36.590 0.000

Variances

ECONOMIC 1.677 0.557 3.011 0.003

MORALS 4.715 0.486 9.692 0.000

Residual Variances

PRIVTOWN 3.366 0.553 6.090 0.000

GOVTRESP 7.034 0.297 23.685 0.000

COMPETE 2.645 0.983 2.692 0.007

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

30

HOMOSEX 5.851 0.403 14.527 0.000

ABORTION 3.957 0.372 10.648 0.000

EUTHANAS 4.119 0.215 19.116 0.000

STANDARDIZED MODEL RESULTS

STDYX Standardization

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 0.577 0.094 6.136 0.000

GOVTRESP 0.147 0.038 3.825 0.000

COMPETE 0.732 0.118 6.223 0.000

MORALS BY

HOMOSEX 0.668 0.028 23.883 0.000

ABORTION 0.743 0.029 26.028 0.000

EUTHANAS 0.563 0.028 20.215 0.000

MORALS WITH

ECONOMIC 0.035 0.044 0.782 0.434

Intercepts

PRIVTOWN 1.582 0.044 35.908 0.000

GOVTRESP 1.608 0.044 36.171 0.000

COMPETE 1.440 0.042 34.366 0.000

HOMOSEX 1.471 0.042 34.720 0.000

ABORTION 1.463 0.042 34.637 0.000

EUTHANAS 1.074 0.037 29.136 0.000

Variances

ECONOMIC 1.000 0.000 999.000 999.000

MORALS 1.000 0.000 999.000 999.000

Residual Variances

PRIVTOWN 0.667 0.108 6.157 0.000

GOVTRESP 0.978 0.011 86.785 0.000

COMPETE 0.464 0.172 2.690 0.007

HOMOSEX 0.554 0.037 14.820 0.000

ABORTION 0.448 0.042 10.541 0.000

EUTHANAS 0.683 0.031 21.796 0.000

STDY Standardization

(skip output)

STD Standardization

(skip output)

R-SQUARE

Observed Two-Tailed

Variable Estimate S.E. Est./S.E. P-Value

PRIVTOWN 0.333 0.108 3.068 0.002

GOVTRESP 0.022 0.011 1.913 0.056

COMPETE 0.536 0.172 3.111 0.002

HOMOSEX 0.446 0.037 11.941 0.000

ABORTION 0.552 0.042 13.014 0.000

EUTHANAS 0.317 0.031 10.107 0.000

QUALITY OF NUMERICAL RESULTS

Condition Number for the Information Matrix 0.720E-03

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

31

(ratio of smallest to largest eigenvalue)

MODEL MODIFICATION INDICES

Minimum M.I. value for printing the modification index 10.000

M.I. E.P.C. Std E.P.C. StdYX E.P.C.

BY Statements

MORALS BY GOVTRESP 31.631 0.240 0.522 0.195

WITH Statements

COMPETE WITH PRIVTOWN 31.719 86.626 86.626 29.028

HOMOSEX WITH GOVTRESP 12.426 0.760 0.760 0.118

Beginning Time: 21:58:43

Ending Time: 21:58:43

Elapsed Time: 00:00:00

MUTHEN & MUTHEN

3463 Stoner Ave.

Los Angeles, CA 90066

Tel: (310) 391-9971

Fax: (310) 391-8971

Web: www.StatModel.com

Support: Support@StatModel.com

Copyright (c) 1998-2008 Muthen & Muthen

The overall model fit is not great, with a χ

2

statistic of 42.147 (df=8) large enough to reject the

null of a good fit. The RMSEA (Root Mean Square Error of Approximation) is .061, which is

higher than the cutoff value of .05 (significance level) chosen to indicate a good fit. CFI

(Comparative Fit Index) is .961.

Under the MODEL RESULTS heading the unstandardized loadings appear along with standard

errors, ratios of the estimates to their standard errors, and p-values. The ratios under Est./S.E.

column are test statistics that are equivalent to z scores. If the absolute value of the number in

this column is greater than 1.96, the estimate can be interpreted as significant at the .05 level.

Alternatively, if a p-value is smaller than .05, you may reject the null hypothesis at the same

significance level that the parameter is zero. In this case all of the unconstrained loading

estimates are significant.

Standardized factor loadings are presented under the STANDIARDIZED MODEL RESULTS heading.

6

ABORTION has the highest standardized factor loading .743; ABORTION appears to be a

reliable indicator of moral values. By contrast, GOVTRESP has the lowest .147, suggesting

that it is an unreliable indicator of economic values. Unstandardized and standardized delta

values appear under Residual Variances.

The squared multiple correlations under R-SQUARE provide information on how much variance

the common factors account for in the observed variables. R

2

is a standardized factor loading

6

The previous version presents standardized factor loadings under the StdYX column of the MODEL RESULTS.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

32

squared; for example, .535824=.732

2

for COMPETE. Economic values, for example, explains

53.6 percent of the total variation in COMPETE and moral values account for 55.2 percent of

variance of ABORTION. The low R

2

of .022 suggests that economic values can explain only a

small portion of variation of GOVTRESP.

The correlation between the two common factors (economic and moral values) is a very

small .035 (p<.434), and the covariance estimate of .098 is not statistically discernable from

zero (p<.433). You may find these statistics appear respectively on standardized and

unstandardized results. Two factors do not appear to be closely related each other. The

variances of two common factors on unstandardized results are 1.677 and 4.714, respectively.

Note that standardized variances of common factors are set 1.

The MODINDICES of OUTPUT command produces model modification indices and gives you

some hints about model specification. Under the MODEL MODIFICATION INDICES heading Mplus

makes three suggestions: 1) adding covariance between COMPETE and PRIVTOWN will

reduce χ

2

by 31.719, 2) adding a path from the MORALS latent variable to GOVTRESP by

31.631, and 3) adding a covariance between HOMOSEX and GOVTRESP by 12.426. The

GOVTRESP item has something in common with the morality dimension, either by sharing

measurement error with the HOMOSEX variable or as a direct indicator of the latent morality

dimension. Because the standardized loading of GOVTRESP on ECONOMIC was so low, it is

possible that the item is actually tapping a different values dimension. Thus the second

suggestion makes theoretical sense and will be estimated.

Add GOVTRESP to moral values in the MODEL command and then save the input file.

TITLE: Second Two Factor Model;

DATA: FILE IS values.dat;

VARIABLE: NAMES ARE privtown govtresp compete

homosex abortion euthanas;

MODEL: economic BY privtown govtresp compete;

morals BY homosex abortion euthanas govtresp;

OUTPUT: STANDARDIZED;

MODINDICES;

Mplus returns the following output. Pay attention to the parts in red. Note that the degrees of

freedom decrease from 8 to 7.

Mplus VERSION 5.1

MUTHEN & MUTHEN

11/17/2008 11:38 PM

INPUT INSTRUCTIONS

(skip output)

INPUT READING TERMINATED NORMALLY

(skip output)

THE MODEL ESTIMATION TERMINATED NORMALLY

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

33

TESTS OF MODEL FIT

Chi-Square Test of Model Fit

Value 9.893

Degrees of Freedom 7

P-Value 0.1947

Chi-Square Test of Model Fit for the Baseline Model

Value 891.990

Degrees of Freedom 15

P-Value 0.0000

CFI/TLI

CFI 0.997

TLI 0.993

Loglikelihood

H0 Value -16200.907

H1 Value -16195.961

Information Criteria

Number of Free Parameters 20

Akaike (AIC) 32441.814

Bayesian (BIC) 32542.938

Sample-Size Adjusted BIC 32479.411

(n* = (n + 2) / 24)

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.019

90 Percent C.I. 0.000 0.044

Probability RMSEA <= .05 0.985

SRMR (Standardized Root Mean Square Residual)

Value 0.016

MODEL RESULTS

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 1.000 0.000 999.000 999.000

GOVTRESP 0.285 0.082 3.488 0.000

COMPETE 1.256 0.442 2.839 0.005

MORALS BY

HOMOSEX 1.000 0.000 999.000 999.000

ABORTION 0.991 0.072 13.787 0.000

EUTHANAS 0.627 0.046 13.662 0.000

GOVTRESP 0.239 0.043 5.610 0.000

MORALS WITH

ECONOMIC 0.043 0.136 0.317 0.752

Intercepts

PRIVTOWN 3.553 0.066 53.876 0.000

GOVTRESP 4.312 0.079 54.775 0.000

COMPETE 3.440 0.070 49.048 0.000

HOMOSEX 4.781 0.095 50.096 0.000

ABORTION 4.352 0.087 49.845 0.000

EUTHANAS 2.638 0.072 36.590 0.000

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

34

Variances

ECONOMIC 1.806 0.654 2.762 0.006

MORALS 4.855 0.490 9.906 0.000

Residual Variances

PRIVTOWN 3.238 0.648 4.998 0.000

GOVTRESP 6.760 0.290 23.312 0.000

COMPETE 2.856 1.007 2.836 0.005

HOMOSEX 5.711 0.400 14.269 0.000

ABORTION 4.078 0.356 11.465 0.000

EUTHANAS 4.123 0.213 19.312 0.000

STANDARDIZED MODEL RESULTS

STDYX Standardization

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 0.598 0.107 5.608 0.000

GOVTRESP 0.143 0.038 3.785 0.000

COMPETE 0.707 0.125 5.658 0.000

MORALS BY

HOMOSEX 0.678 0.028 24.613 0.000

ABORTION 0.734 0.028 26.548 0.000

EUTHANAS 0.562 0.028 20.437 0.000

GOVTRESP 0.196 0.034 5.810 0.000

MORALS WITH

ECONOMIC 0.015 0.047 0.312 0.755

Intercepts

PRIVTOWN 1.582 0.044 35.908 0.000

GOVTRESP 1.608 0.044 36.171 0.000

COMPETE 1.440 0.042 34.366 0.000

HOMOSEX 1.471 0.042 34.721 0.000

ABORTION 1.463 0.042 34.637 0.000

EUTHANAS 1.074 0.037 29.136 0.000

Variances

ECONOMIC 1.000 0.000 999.000 999.000

MORALS 1.000 0.000 999.000 999.000

Residual Variances

PRIVTOWN 0.642 0.128 5.028 0.000

GOVTRESP 0.940 0.017 55.747 0.000

COMPETE 0.501 0.177 2.836 0.005

HOMOSEX 0.540 0.037 14.476 0.000

ABORTION 0.461 0.041 11.365 0.000

EUTHANAS 0.684 0.031 22.099 0.000

STDY Standardization

(skip output)

STD Standardization

(skip output)

R-SQUARE

Observed Two-Tailed

Variable Estimate S.E. Est./S.E. P-Value

PRIVTOWN 0.358 0.128 2.804 0.005

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

35

GOVTRESP 0.060 0.017 3.538 0.000

COMPETE 0.499 0.177 2.829 0.005

HOMOSEX 0.460 0.037 12.307 0.000

ABORTION 0.539 0.041 13.274 0.000

EUTHANAS 0.316 0.031 10.218 0.000

QUALITY OF NUMERICAL RESULTS

Condition Number for the Information Matrix 0.646E-03

(ratio of smallest to largest eigenvalue)

MODEL MODIFICATION INDICES

Minimum M.I. value for printing the modification index 10.000

M.I. E.P.C. Std E.P.C. StdYX E.P.C.

No modification indices above the minimum value.

(skip output)

Skipped are some parts of the output, which are redundant or less informative.

The χ

2

test yields a value of 9.893 (df=7), which does not reject the null hypothesis of an

overall good fit (p<.1947). The RMSEA .019 (p<.985) and CFI .997 indicate that this model

fits the data well.

All unstandardized and standardized factor loadings are statistically significant (p<.000).

Individual standardized loading values remain almost unchanged. The factor loading .196

suggests that GOVTRESP manifests both economic and moral values significantly. The

covariance and correlation of two factors are respectively .043 and .015, which are smaller than

those of model 1. Adding a relation between GOVTRESP and moral values appears to make

the model fit better. Model modification indices do not suggest any change in this model.

The conclusion from this analysis is that two nearly orthogonal dimensions underlie the

economic and moral values of American citizens. It is not clear, however, whether the

GOVTRESP item is tapping either dimension. Future surveys should incorporate more reliable

measures of economic values.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

36

3.4 Two Factor Model using SAS CALIS

This section demonstrates how to estimate a confirmatory factor model using the SAS CALIS

procedure. This procedure supports various linear models including structural equation model,

explanatory and confirmatory factor analysis, multivariate linear regression, path analysis,

simultaneous equation model, and canonical correlation. The following CALIS procedure fits

the first two factor model discussed above.

LIBNAME cfa 'c: \temp\cfa';

DATA cfa.values;

INFILE 'c: \temp\cfa\values.dat';

INPUT privtown govtresp compete homosex abortion euthanas;

RUN;

The LIBNAME statement above defines a library cfa that refers a physical space c:\temp\cfa;

SAS looks for data sets in the directory once the library cfa is specified. The DATA step above

reads six variables from an ASCII text file values.dat and save them into a SAS data set

values.sas7bdat in the library cfa.

ODS HTML;

PROC CALIS DATA=cfa.values METHOD=ML PALL;

VAR privtown govtresp compete homosex abortion euthanas;

LINEQS privtown = 1.0 f1 + e1,

govtresp = lambda2 f1 + e2,

compete = lambda3 f1 + e3,

homosex = 1.0 f2 + e4,

abortion = lambda5 f2 + e5,

euthanas = lambda6 f2 + e6;

STD f1-f2 = phi1-phi2,

e1-e6 = theta1-theta6;

COV f1-f2 = phi3;

RUN;

ODS HTML CLOSE;

PROC CALIS reads a data set values.sas7bdat from a library sas (c:\temp\cfa). The

METHOD=ML uses the maximum likelihood method and the PALL option reports all possible

output. The VAR statement lists the variables used. The LINEQS statement specifies

relationships between latent and manifest variables in equation form. Note that lambda1 and

lambda4 are set 1 for model identification. STD and COV statements are place where standard

deviations and covariances are defined. Finally, ODS HTML redirects the SAS output into the

HTML format.

The following is selected from the SAS output.

The CALIS Procedure

Covariance Structure Analysis: Maximum Likelihood Estimation

Fit Function 0.0363

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

37

Goodness of Fit Index (GFI) 0.9882

GFI Adjusted for Degrees of Freedom (AGFI) 0.9690

Root Mean Square Residual (RMR) 0.0519

Parsimonious GFI (Mulaik, 1989) 0.5270

Chi-Square 42.1104

Chi-Square DF 8

Pr > Chi-Square <.0001

Independence Model Chi-Square 891.22

Independence Model Chi-Square DF 15

RMSEA Estimate 0.0607

RMSEA 90% Lower Confidence Limit 0.0434

RMSEA 90% Upper Confidence Limit 0.0793

ECVI Estimate 0.0589

ECVI 90% Lower Confidence Limit 0.0445

ECVI 90% Upper Confidence Limit 0.0798

Probability of Close Fit 0.1472

Bentler's Comparative Fit Index 0.9611

Elliptic Corrected Chi-Square 38.0261

Pr > Elliptic Corrected Chi-Square <.0001

Normal Theory Reweighted LS Chi-Square 41.5308

Akaike's Information Criterion 26.1104

Bozdogan's (1987) CAIC -22.3390

Schwarz's Bayesian Criterion -14.3390

McDonald's (1989) Centrality 0.9854

Bentler & Bonett's (1980) Non-normed Index 0.9270

Bentler & Bonett's (1980) NFI 0.9527

James, Mulaik, & Brett (1982) Parsimonious NFI0.5081

Z-Test of Wilson & Hilferty (1931) 4.6039

Bollen (1986) Normed Index Rho1 0.9114

Bollen (1988) Non-normed Index Delta2 0.9614

Hoelter's (1983) Critical N 428

Manifest Variable Equations with Estimates

privtown = 1.0000 f1 + 1.0000 e1

govtresp = 0.2546 * f1 + 1.0000 e2

Std Err 0.0701 lambda2

t Value 3.6310

compete = 1.2699 * f1 + 1.0000 e3

Std Err 0.4224 lambda3

t Value 3.0064

homosex = 1.0000 f2 + 1.0000 e4

abortion = 1.1127 * f2 + 1.0000 e5

Std Err 0.0831 lambda5

t Value 13.3876

euthanas = 0.8426 * f2 + 1.0000 e6

Std Err 0.0614 lambda6

t Value 13.7307

Variances of Exogenous Variables

Variabl

e

Parameter Estimate Standard

Error

t Value

f1 phi1 0.33257 0.11439 2.91

f2 phi2 0.44621 0.04613 9.67

e1 theta1 0.66743 0.11357 5.88

e2 theta2 0.97844 0.04129 23.70

e3 theta3 0.46369 0.17864 2.60

e4 theta4 0.55379 0.03822 14.49

e5 theta5 0.44750 0.04211 10.63

e6 theta6 0.68319 0.03571 19.13

Covariances Among Exogenous Variables

Var1 Var2 Parameter Estimate Standard

Error

t Value

f1 f2 phi3 0.01336 0.01721 0.78

Manifest Variable Equations with Standardized Estimates

privtown = 0.5767 f1 + 0.8170 e1

govtresp = 0.1468 * f1 + 0.9892 e2

lambda2

compete = 0.7323 * f1 + 0.6809 e3

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

38

lambda3

homosex = 0.6680 f2 + 0.7442 e4

abortion = 0.7433 * f2 + 0.6690 e5

lambda5

euthanas = 0.5629 * f2 + 0.8266 e6

lambda6

Squared Multiple Correlations

Variable Error Variance Total Variance R-Square

1 privtown 0.66743 1.00000 0.3326

2 govtresp 0.97844 1.00000 0.0216

3 compete 0.46369 1.00000 0.5363

4 homosex 0.55379 1.00000 0.4462

5 abortion 0.44750 1.00000 0.5525

6 euthanas 0.68319 1.00000 0.3168

Correlations Among Exogenous Variables

Var1 Var2 Parameter Estimate

f1 f2 phi3 0.03469

Predicted Moments of

Latent Variables

f1 f2

f1 0.33257 0.01336

f2 0.01336 0.44621

This model has a large χ

2

42.1104 (p<.0001), RMSEA .0607 (p<.1472), and CFI .9611, which

indicates a poor fit. This χ

2

is slightly smaller than 42.147 that Mplus produced.

Unstandardized factor loadings of SAS CALIS are different from those Mplus returned,

although z scores are very similar. For example, the factor loading of GOVTRESP is .2546 and

its test statistic is 3.6320 (=.2546/.0701) in SAS CALIS. Corresponding statistics were

respectively .304 and 3.675 (=.304/.083) in Mplus. Similarly, the covariance of two factors

is .01336 and its test statistic is .78 (=.01336/.01721); Mplus reported .098 and .784,

respectively.

However, standardized factor loadings, R

2

, and correlation of two factors are almost the same

as corresponding statistics that Mplus computed. For instance, SAS CALIS and Mplus

respectively report the factor loading .7433 and .743 for ABORTION. R

2

are .5525 (=.7433

2

)

and .552, respectively. The correlation of two factors is .03469 in SAS CALIS and .035 in

Mplus.

The second two factor model is estimated by the following SAS codes. Note that lambda7 f2

was added to the GOVTRESP equation.

ODS HTML;

PROC CALIS DATA=sas.values METHOD=ML PALL;

VAR privtown govtresp compete homosex abortion euthanas;

LINEQS privtown = 1.0 f1 + e1,

govtresp = lambda2 f1 + lambda7 f2 + e2,

compete = lambda3 f1 + e3,

homosex = 1.0 f2 + e4,

abortion = lambda5 f2 + e5,

euthanas = lambda6 f2 + e6;

STD f1-f2 = phi1-phi2,

e1-e6 = theta1-theta6;

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

39

COV f1-f2 = phi3;

RUN;

ODS HTML CLOSE;

The output is similar to that of the first two factor model but you need to pay attention to

lambda7 in the GOVTRESP equation, which represents the factor loading of ECONOMIC.

The CALIS Procedure

Covariance Structure Analysis: Maximum Likelihood Estimation

Fit Function 0.0085

Goodness of Fit Index (GFI) 0.9972

GFI Adjusted for Degrees of Freedom (AGFI) 0.9915

Root Mean Square Residual (RMR) 0.0178

Parsimonious GFI (Mulaik, 1989) 0.4653

Chi-Square 9.8844

Chi-Square DF 7

Pr > Chi-Square 0.1952

Independence Model Chi-Square 891.22

Independence Model Chi-Square DF 15

RMSEA Estimate 0.0189

RMSEA 90% Lower Confidence Limit .

RMSEA 90% Upper Confidence Limit 0.0436

ECVI Estimate 0.0328

ECVI 90% Lower Confidence Limit .

ECVI 90% Upper Confidence Limit 0.0437

Probability of Close Fit 0.9847

Bentler's Comparative Fit Index 0.9967

Elliptic Corrected Chi-Square 8.9257

Pr > Elliptic Corrected Chi-Square 0.2580

Normal Theory Reweighted LS Chi-Square 9.8904

Akaike's Information Criterion -4.1156

Bozdogan's (1987) CAIC -46.5088

Schwarz's Bayesian Criterion -39.5088

McDonald's (1989) Centrality 0.9988

Bentler & Bonett's (1980) Non-normed Index 0.9929

Bentler & Bonett's (1980) NFI 0.9889

James, Mulaik, & Brett (1982) Parsimonious NFI 0.4615

Z-Test of Wilson & Hilferty (1931) 0.8623

Bollen (1986) Normed Index Rho1 0.9762

Bollen (1988) Non-normed Index Delta2 0.9967

Hoelter's (1983) Critical N 1651

Manifest Variable Equations with Estimates

privtown = 1.0000 f1 + 1.0000 e1

govtresp = 0.2387 * f1 + 0.2893 * f2 + 1.0000 e2

Std Err 0.0691 lambda2 0.0520 lambda7

t Value 3.4558 5.5599

compete = 1.1809 * f1 + 1.0000 e3

Std Err 0.3967 lambda3

t Value 2.9770

homosex = 1.0000 f2 + 1.0000 e4

abortion = 1.0828 * f2 + 1.0000 e5

Std Err 0.0783 lambda5

t Value 13.8252

euthanas = 0.8295 * f2 + 1.0000 e6

Std Err 0.0601 lambda6

t Value 13.8008

Variances of Exogenous Variables

Variable Parameter Estimate Standard

Error

t Value

f1 phi1 0.35809 0.12399 2.89

f2 phi2 0.45951 0.04626 9.93

e1 theta1 0.64191 0.12276 5.23

e2 theta2 0.94033 0.04030 23.33

e3 theta3 0.50067 0.16838 2.97

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

40

Variances of Exogenous Variables

Variable Parameter Estimate Standard

Error

t Value

e4 theta4 0.54049 0.03772 14.33

e5 theta5 0.46127 0.04025 11.46

e6 theta6 0.68382 0.03548 19.28

Covariances Among Exogenous Variables

Var1 Var2 Parameter Estimate Standard

Error

t Value

f1 f2 phi3 0.00589 0.01814 0.32

Manifest Variable Equations with Standardized Estimates

privtown = 0.5984 f1 + 0.8012 e1

govtresp = 0.1428 * f1 + 0.1961 * f2 + 0.9697 e2

lambda2 lambda7

compete = 0.7066 * f1 + 0.7076 e3

lambda3

homosex = 0.6779 f2 + 0.7352 e4

abortion = 0.7340 * f2 + 0.6792 e5

lambda5

euthanas = 0.5623 * f2 + 0.8269 e6

lambda6

Squared Multiple Correlations

Variable Error Variance Total Variance R-Square

1 privtown 0.64191 1.00000 0.3581

2 govtresp 0.94033 1.00000 0.0597

3 compete 0.50067 1.00000 0.4993

4 homosex 0.54049 1.00000 0.4595

5 abortion 0.46127 1.00000 0.5387

6 euthanas 0.68382 1.00000 0.3162

Correlations Among Exogenous Variables

Var1 Var2 Parameter Estimate

f1 f2 phi3 0.01453

Predicted Moments of

Latent Variables

f1 f2

f1 0.35809 0.00589

f2 0.00589 0.45951

This second model has a smaller χ

2

9.8844 (p<.1952) and RMSEA .0189 (p<.0328), which do

not reject the null hypothesis at the .05 level. CFI .9967 also indicates a good fit. Government

responsibility (GOVTRESP) is significantly explained by both economic and moral values.

As discussed above, unstandardized factor loadings of SAS CALIS are different from those of

Mplus, but their test statistics are similar. The covariance of two factors is .00589, slightly

larger than .043 in Mplus, but its test statistic is .32 (=.00589/.01814) that is very close to .317

Mplus returned. Standardardized factor loadings and correlation of two factors reported by both

software packages are virtually same. SAS CALIS reports .1961 as a standardized factor

loading of moral values on GOVTRESP, which is the same as .196 that Mplus returned.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

41

3.5 Two Factor Model using Amos

Amos consists of several applications including Amos Graphics, Program Editor, File manager,

and Seed Manager. Most people begin with Amos Graphics to fit a model because it provides a

point-and-click GUI environment, which makes it easy for beginners to specify models. This

section, however, uses Program Editor to take advantage of its simplicity at the expense of

burdensome programming. See Appendix if you want to estimate a model using Amos

Graphics.

Amos Program Editor can be launched from any computer running Windows in the UITS

Student Technology Centers by going to Start → All Programs → Departmentally

Sponsored → Statistics-Math → Amos 16 → Program Editor. At Program Editor, click File

→ Open to open a dialog box, and then select a sample program Ex08.vb listed in the default

directory Examples. The following screen will display.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

42

We are going to copy its syntax structure, which is the same as that of two factor models that

we want to fit. Amos uses Visual Basic (VB.NET) or C# language to specify a model. The

beginning of the program describes a program title and backgrounds. Main program begins

with Sub Main() and ends with End Sub. Pay attention to Sem.BeginGroup() and

Sem.AStructure() functions. The former tells where a data file is located, while the latter

specifies model to estimate. Sem.TextOutput() creates an output file and

Sem.Standardized() computes standardized estimates.

Let us rename the program file first. Click File → Save As… to open Save As dialog box.

Provide a name values and then click Save. Now, the file name became values.vb.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

43

Look at the comments in green. Replace Example 8: with First Two Factor Model and then

remove other comments (see the screenshot below). Change Sem.BeginGroup() to look like

Sem.BeginGroup("c:\temp\cfa\values.sav"). Amos reads data set values.sav from

c:\temp\cfa. In the six SEM.AStructure() functions, replace spatial with economic and

verbal with morals. Replace visperc with privtown, cubes with govtresp, and so on.

Finally, change error terms from e1 through e6 as shown in the above screenshot.

The first SEM.AStructure() tells that the observed variable privtown is explained by a latent

variable economic and error e1. Note that the estimate of factor loading is set 1 for the purpose

of identification.

Click File → Run or click Run button on the right bottom to fit this model. Amos Output pops

up if no error is found. You may see the list of results. Click Notes for Model on the left pane

to get the chi-square test of model fit (see following screenshot). Chi-square χ

2

of 42.110 (df=8)

is so large as to reject the null hypothesis of a good fit (p<.000).

Now click Model Fit on the left pane to see the other goodness-of-fit statistics. RMSEA .061

and CFI .961, which are the same as what Mplus and LISREL returned, suggests that the fit of

the model is questionable.

Model Fit Summary

CMIN

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

44

Model NPAR CMIN DF P CMIN/DF

Default model 13 42.110 8 .000 5.264

Saturated model 21 .000 0

Independence model 6 891.221 15 .000 59.415

RMR, GFI

Model RMR GFI AGFI PGFI

Default model .411 .988 .969 .376

Saturated model .000 1.000

Independence model 1.557 .786 .700 .561

Baseline Comparisons

Model

NFI

Delta1

RFI

rho1

IFI

Delta2

TLI

rho2

CFI

Default model .953 .911 .961 .927 .961

Saturated model 1.000 1.000 1.000

Independence model .000 .000 .000 .000 .000

Parsimony-Adjusted Measures

Model PRATIO PNFI PCFI

Default model .533 .508 .513

Saturated model .000 .000 .000

Independence model 1.000 .000 .000

NCP

Model NCP LO 90 HI 90

Default model 34.110 17.433 58.302

Saturated model .000 .000 .000

Independence model 876.221 782.083 977.751

FMIN

Model FMIN F0 LO 90 HI 90

Default model .036 .029 .015 .050

Saturated model .000 .000 .000 .000

Independence model .769 .756 .675 .844

RMSEA

Model RMSEA LO 90 HI 90 PCLOSE

Default model .061 .043 .079 .147

Independence model .225 .212 .237 .000

AIC

Model AIC BCC BIC CAIC

Default model 68.110 68.268 133.841 146.841

Saturated model 42.000 42.255 148.180 169.180

Independence model 903.221 903.294 933.558 939.558

ECVI

Model ECVI LO 90 HI 90 MECVI

Default model .059 .044 .080 .059

Saturated model .036 .036 .036 .036

Independence model .779 .698 .867 .779

HOELTER

Model

HOELTER

.05

HOELTER

.01

Default model 427 553

Independence model 33 40

Click Estimates on the pane to check parameter estimates. Under the Regression Weights

heading the unstandardized loadings appear along with standard errors, a critical ratios, and

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

45

p-values. The standardized estimates under Standardized Regression Weights can be

interpreted as the correlation between the observed variable and the corresponding common

factor. These unstandardized and standardized estimates are the same as what Mplus and

LISREL produced. However, standard errors are slightly different from Mplus and LISREL

counterparts.

ABORTION (.743) and COMPETE (.732) have highest standardized factor loadings,

GOVTRESP has a lowest factor loading of .147. Squared multiple correlations (R

2

)

corresponding to the six observed variables, which are arranged in an arbitrary order, indicate

that the respective factor explains a respectable portion of the variance. Economic values

(ECONOMIC) can explain only 2.2 percent (=.147

2

) of variation in GOVTRESP although

regression weights are all significant though. GOVTRESP does not seem to tap the same values

dimension as the other two economics questions.

Estimates (Group number 1 - Model 1)

Scalar Estimates (Group number 1 - Model 1)

Maximum Likelihood Estimates

Regression Weights: (Group number 1 - Model 1)

Estimate S.E. C.R. P Label

privtown <--- economic 1.000

govtresp <--- economic .304 .084 3.631 ***

compete <--- economic 1.351 .449 3.006 .003

homosex <--- morals 1.000

abortion <--- morals 1.018 .076 13.388 ***

euthanas <--- morals .637 .046 13.731 ***

Standardized Regression Weights: (Group number 1 - Model 1)

Estimate

privtown <--- economic .577

govtresp <--- economic .147

compete <--- economic .732

homosex <--- morals .668

abortion <--- morals .743

euthanas <--- morals .563

Covariances: (Group number 1 - Model 1)

Estimate S.E. C.R. P Label

economic <--> morals .098 .126 .777 .437

Correlations: (Group number 1 - Model 1)

Estimate

economic <--> morals .035

Variances: (Group number 1 - Model 1)

Estimate S.E. C.R. P Label

economic 1.677 .577 2.907 .004

morals 4.715 .487 9.674 ***

e1 3.366 .573 5.877 ***

e2 7.034 .297 23.698 ***

e3 2.645 1.019 2.596 .009

e4 5.851 .404 14.491 ***

e5 3.957 .372 10.627 ***

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

46

Estimate S.E. C.R. P Label

e6 4.119 .215 19.132 ***

Squared Multiple Correlations: (Group number 1 - Model 1)

Estimate

euthanas .317

abortion .553

homosex .446

compete .536

govtresp .022

privtown .333

Unlike LISREL, an Amos program written in Program Editor is not able to produce a path

diagram and Amos graphics does not generate a VB or C# program on the basis of a path

diagram drawn.

GOVTRESP was only weakly accounted for by the ECONOMIC variable, hinting that the

survey item was not tapping the same values dimension as the other two economic values

indicators. An alternative possibility is that GOVTRESP is also tied to the morality dimension.

Let us set a relationship between MORALS to GOVTRESP by adding + morals to the second

Sem.AStructure() below. It is like adding an arrow from MORALS to GOVTRESP on a path

diagram. Look at the right part in red.

Header

Module MainModule

' Second Two Factor Model

Sub Main()

Dim Sem As New AmosEngine

Try

Sem.TextOutput()

Sem.Standardized()

Sem.Smc()

Sem.BeginGroup( "c:\temp\cfa\values.sav")

Sem.AStructure("privtown = (1) economic + (1) e1")

Sem.AStructure("govtresp = economic + morals + (1) e2")

Sem.AStructure("compete = economic + (1) e3")

Sem.AStructure("homosex = (1) morals + (1) e4")

Sem.AStructure("abortion = morals + (1) e5")

Sem.AStructure("euthanas = morals + (1) e6")

Sem.FitModel()

Finally

Sem.Dispose()

End Try

End Sub

End Module

Run this program by clicking File → Run and then click Notes for Model in the Amos Output.

The overall model fit appears quite good. χ

2

(df=7) decreases down to 9.884, which is too small

to reject the null hypothesis of a good fit (p<.195). A small RMSEA of .019 and a large CFI

of .997 indicate a good fit of this model.

Notes for Model (Model 1)

Computation of degrees of freedom (Model 1)

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

47

Number of distinct sample moments: 21

Number of distinct parameters to be estimated: 14

Degrees of freedom (21 - 14): 7

Result (Model 1)

Minimum was achieved

Chi-square = 9.884

Degrees of freedom = 7

Probability level = .195

Model Fit Summary

CMIN

Model NPAR CMIN DF P CMIN/DF

Default model 14 9.884 7 .195 1.412

Saturated model 21 .000 0

Independence model 6 891.221 15 .000 59.415

RMR, GFI

Model RMR GFI AGFI PGFI

Default model .119 .997 .991 .332

Saturated model .000 1.000

Independence model 1.557 .786 .700 .561

Baseline Comparisons

Model

NFI

Delta1

RFI

rho1

IFI

Delta2

TLI

rho2

CFI

Default model .989 .976 .997 .993 .997

Saturated model 1.000 1.000 1.000

Independence model .000 .000 .000 .000 .000

Parsimony-Adjusted Measures

Model PRATIO PNFI PCFI

Default model .467 .461 .465

Saturated model .000 .000 .000

Independence model 1.000 .000 .000

NCP

Model NCP LO 90 HI 90

Default model 2.884 .000 15.404

Saturated model .000 .000 .000

Independence model 876.221 782.083 977.751

FMIN

Model FMIN F0 LO 90 HI 90

Default model .009 .002 .000 .013

Saturated model .000 .000 .000 .000

Independence model .769 .756 .675 .844

RMSEA

Model RMSEA LO 90 HI 90 PCLOSE

Default model .019 .000 .044 .985

Independence model .225 .212 .237 .000

AIC

Model AIC BCC BIC CAIC

Default model 37.884 38.055 108.671 122.671

Saturated model 42.000 42.255 148.180 169.180

Independence model 903.221 903.294 933.558 939.558

ECVI

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

48

Model ECVI LO 90 HI 90 MECVI

Default model .033 .030 .043 .033

Saturated model .036 .036 .036 .036

Independence model .779 .698 .867 .779

HOELTER

Model

HOELTER

.05

HOELTER

.01

Default model 1650 2167

Independence model 33 40

Amos reports the factor loading of MORALS on GOVTRESP on the third row. Unstandardized

and standardized estimates are respectively .239 and .196, which are smaller than those of other

estimates. GOVTRESP appears an unreliable indicator of both economic and moral values.

Other standardized factor loadings range from .562 (EUTHANAS) to .734 (ABORTION).

Corresponding squared multiple correlations (R

2

) range from .060 (GOVTRESP) to .499

(COMPETE) and .539 (ABORTION). Latent variables ECONOMIC and MORALS account

for 6 percent of the total variation in GOVTRESP. Covariance and correlation of two factors

are .043 and .015 but they are not statistically discernable (p<.7). Finally, the correlation

between the two common factors is a very small -.011, and the covariance estimate of -.030 is

not statistically discernable from zero (p<.808).

Estimates (Group number 1 - Model 1)

Scalar Estimates (Group number 1 - Model 1)

Maximum Likelihood Estimates

Regression Weights: (Group number 1 - Model 1)

Estimate S.E. C.R. P Label

privtown <--- economic 1.000

govtresp <--- economic .285 .082 3.456 ***

govtresp <--- morals .239 .043 5.560 ***

compete <--- economic 1.256 .422 2.977 .003

homosex <--- morals 1.000

abortion <--- morals .991 .072 13.825 ***

euthanas <--- morals .627 .045 13.801 ***

Standardized Regression Weights: (Group number 1 - Model 1)

Estimate

privtown <--- economic .598

govtresp <--- economic .143

govtresp <--- morals .196

compete <--- economic .707

homosex <--- morals .678

abortion <--- morals .734

euthanas <--- morals .562

Covariances: (Group number 1 - Model 1)

Estimate S.E. C.R. P Label

economic <--> morals .043 .132 .325 .745

Correlations: (Group number 1 - Model 1)

Estimate

economic <--> morals .015

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

49

Variances: (Group number 1 - Model 1)

Estimate S.E. C.R. P Label

economic 1.806 .625 2.888 .004

morals 4.855 .489 9.933 ***

e1 3.237 .619 5.228 ***

e2 6.760 .290 23.330 ***

e3 2.857 .960 2.974 .003

e4 5.711 .399 14.329 ***

e5 4.078 .356 11.459 ***

e6 4.123 .214 19.275 ***

Squared Multiple Correlations: (Group number 1 - Model 1)

Estimate

euthanas .316

abortion .539

homosex .460

compete .499

govtresp .060

privtown .358

In summary, there appear to be two orthogonal dimensions which underlie American attitudes

on a number of different issues: one representing economic values and the other representing

moral values. It is unclear which dimension the GOVTRESP item was tapping, however, and

future surveys should employ a more reliable measure.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

50

3.6 Summary

Tables 4 and 5 compare the unstandardized and standardized factor loadings from each

software package for both two-factor models. The tables also present standard errors in

parentheses, χ

2

, Root Mean Square Error of Approximation (RMSEA), and Comparative Fit

Index (CFI).

Table 4 Comparison of Estimates: First Two Factor Model (N=1,160, DF=8)

Amos LISREL Mplus SAS CALIS

Unstd. Standard Unstd. Standard Unstd. Standard Unstd. Standard

F1 PRIVTOWN

1.00 .577

1.00 .58 1.000 .577

(.094)

1.0000 .5767

F1 GOVTRESP

.304

(.084)

.147 .30

(.084)

.15 .304

(.083)

.147

(.038)

.2546

(.0701)

.1468

F1 COMPETE

1.351

(.449)

.732 1.35

(.45)

.73 1.350

(.433)

.732

(.118)

1.2699

(.4224)

.7323

F2 HOMOSEX

1.00 .668 1.00 .67 1.000 .668

(.028)

1.0000 .6680

F2 ABORTION

1.018

(.076)

.743 1.02

(.076)

.74 1.018

(.076)

.743

.029)

1.1127

(.0831)

.7433

F2 EUTHANAS

.637

(.046)

.563 .64

(.046)

.56 .637

(.046)

.563

(.028)

.8426

(.0614)

.5629

F1 F2

.098

a

(.126)

.035

b

.10

a

(.13)

.03

b

.098

a

(.124)

.035

b

(.044)

.0134

a

(.0172)

.0347

b

;

2

, RMSEA, CFI

42.110

.061

.961

42.11

.060

.96

42.147

.061

.961

42.1104

.0607

.9611

Standard errors appear in parentheses

a

covariance;

b

correlation between F1 (economic values) and F2 (moral values)

Table 5 : Comparison of Estimates: Second Two Factor Model (N=1,160, DF=7)

Amos LISREL Mplus SAS CALIS

Unstd. Standard Unstd. Standard Unstd. Standard Unstd. Standard

F1 PRIVTOWN

1.00 .598

1.00 .60 1.000 .598

(.107)

1.0000 .5984

F1 GOVTRESP

.285

(.082)

.143 .28

(.082)

.14 .285

(.082)

.143

(.038)

.2387

(.0691)

.1428

F1 COMPETE

1.256

(.422)

.707 1.26

(.42)

.71 1.256

(.442)

.707

(.125)

1.1809

(.3967)

.7066

F2 HOMOSEX

1.00 .678 1.00 .68 1.000 .678

(.028)

1.0000 .6779

F2 ABORTION

.991

(.072)

.734 .99

(.072)

.73 .991

(.072)

.734

.028)

1.0828

(.0783)

.7340

F2 EUTHANAS

.627

(.045)

.562 .63

(.045)

.56 .627

(.046)

.562

(.028)

.8295

(.0601)

.5623

F2 GOVTRESP

.239

(.043)

.196 .24

(.043)

.20 .239

(.043)

.196

(.034)

.2893

(.0520)

.1961

F1 F2

.043

a

(.132)

.015

b

.04

a

(.13)

.01

b

.043

a

(.136)

.015

b

(.047)

.0059

a

(.0181)

.0145

b

;

2

, RMSEA, CFI

9.884

.019

.997

9.88

.019

1.00

9.893

.019

.997

9.8844

.0189

.9967

Standard errors appear in parentheses

a

covariance;

b

correlation between F1 (economic values) and F2 (moral values)

Mplus, LISREL, and Amos produce almost same statistics. SAS/STAT CALIS reports different

unstandidized factor loadings but same standardized statistics. Mplus reports key

goodness-of-fit statistics, while other software packages provide various fit statistics including

χ

2

, RMSEA, and CFI.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

51

4. Confirmatory Factor Analysis with Missing Data

This section shows how to estimate the two-factor model when the raw data matrix includes

missing observations. The data to be analyzed has been saved as an SPSS file named

values_full.sav in the C:\temp\CFA folder. All missing observations have been coded as system

missing (.) in SPSS.

4.1 Missing Data Issue

Missing data is a pervasive problem in the social sciences. A subject may fail to complete a test

in an experimental setting, refuse to give an answer to a particular survey item, or drop out of a

panel. In many cases, researchers choose to drop all observations from subjects that have

missing observations on any of the items included in the model. This approach to handling

missing data is referred to as listwise deletion and is the default in programs such as SPSS and

Stata. Unfortunately dropping incomplete cases results in sacrificing information from the

sample and can lead to biased estimates when the data is not missing completely at random.

Another approach is pairwise deletion that removes observations with missing data in any one

of two variables when computing their covariance. This method of dealing with missing data

can use all available data but each element of covariance matrix may be based on different

observations.

Over the last 30 years more sophisticated means have emerged for dealing with missing data,

many of which have been incorporated into structural equation modeling software. Because it is

available in Amos, LISREL, and Mplus, this document will consider Full Information

Maximum Likelihood (FIML) estimation which makes maximal use of all data available from

every subject in the sample. Other approaches to dealing with missing data, such as multiple

imputation via Bayesian simulation, may also be available depending on the specific software

packages. A non-technical overview of different methods for handling missing data in the

context of structural equation models is available in Enders (2001), though the description of

capabilities of specific computer packages is already dated.

4.2 CFA with Missing Data using Amos

This section fits the model using Amos Graphics instead of Program Editor. Amos can be

launched from any computer running Windows in the UITS Student Technology Centers by

going to Start → All Programs → Departmentally Sponsored → Statistics-Math →Amos

16 → Amos Graphics. The following screen will display:

On the far left pane appear the different tools that can be used to create path diagrams. Just to

the right of the toolbar buttons is a column that will display information about the model after

estimates have been calculated. The remainder of the screen contains the area where the path

diagram will be drawn.

Click File →Data Files to load data. After the Data Files dialog box opens, click on File

Name. Navigate to the C:\temp\CFA folder and choose values_full.sav containing missing

values. Click Open, Then Okay.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

52

On the left pane, click on the Draw unobserved variable button in the tool box

(alternatively click on Diagram → Draw Unobserved) and click some place on the right

workspace to draw an oval representing a latent variable. Move the mouse pointer just below

the oval and click once to create a second oval. Add a covariance between the two latent

variables (common factors) by choosing the Draw Covariances button and clicking and

dragging a two-headed arrow from one factor to the other (see the screenshot on the next page).

Click on the Draw a latent variable button or add an indicator to a latent variable button

. Click three times inside each oval to add a total of six indicators and their respective

error terms (see the following screenshot).

7

By default Amos sets the metric of each error term

7

Alternatively, you may draw ovals and boxes and then link them with appropriate arrows manually. Click on the

blue rectangle button of Draw Observed Variable (alternatively click on Diagram → Draw Observed).

Then click six times to create a total of six boxes for observed variables. Select the blue oval and click six

times on the left of six boxes to represent measurement error specific to each of the observed indicators.

Select the Draw Paths button for drawing single headed arrows. Click and drag from the common and

unique factors to the appropriate observed variable.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

53

by constraining the path parameters to one. The factor loading of the first indicator for each

latent variable is also set to one. Rotate each latent variable by choosing the Rotate the

indicators of a latent variable button and clicking each factor three times until you are

satisfied with the appearance so that the path diagram should now look like the following.

To label the latent variables, right-click in the first oval and choose Object Properties. When

the Object Properties dialog box opens, choose the Text tab. Name the variable ECONOMIC

and then click the second oval to name it MORALS. Click the first error oval and name it d

1

and adjust the font size to 12 point. Do the same to name the error terms d

2

through d

6

. Click

X in the upper right hand corner to close the Object Properties dialog box.

To name the observed variables, choose View → Variables in Dataset. Click and drag the

names of each variable to the appropriate box in the path diagram. If the names do not fit, you

can change the position of parts of the diagram after choosing the Move objects button

or resize the box after clicking on the Change the shape of objects button . Finally, add

an arrow from the MORALS latent variable to the GOVTRESP indicator as suggested in the

earlier analysis. The path diagram now should look something like the following:

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

54

For comparison with LISREL and Mplus, let us set factor loading on PRIVTOWN and

HOMOSEX to one and factor loadings on COMPETE and EUTHANAS free. Right-click on

the arrow from ECONOMIC to COMPETE to open Object Properties dialog box and then

click on the Paramaters tab and delete 1 in the field labeled Regression weight. Click on the

arrow from ECONOMIC to PRIVTOWN and enter 1 in Object Properties dialog box (see

screenshot above). Repeat the same to set the factor loading on HOMOSEX to 1 after set factor

loadings on EUTHANAS free. Click X to close the Object Properties dialog box. The final

path diagram should look like this:

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

55

When missing values are present, it is necessary to request that Amos estimate means and

intercepts (required for FIML estimation), which is not the default. Choose View →Analysis

Properties, click the Estimation tab in the Analysis Properties dialog box, and select

Estimate means and intercepts. Next click on the Output tab. Minimization History is

checked by default. Also place checks next to Standardized Estimates and Squared Multiple

Correlations.

Now, we are ready to fit the model. Go to Analyze → Calculate Estimates. To see the results

in the path diagram click on the View the output path diagram button on the second

pane.

The unstandardized estimates are displayed by default. The path diagram for

unstandardized estimates is often hard to read especially when the diagram does not have

enough space for parameter estimates. To bring up the standardized estimates, click on the

Standardized estimates option in the column between the tools and the workspace.

Amos now displays the standardized factor loadings on arrows and the squared multiple

correlation coefficient for each observed variable. All factor loadings except for GOVTRESP

are large, while the correlation between two latent variables is negligible (-.01). Note that for

some models with many parameters and missing data, Amos (and all SEM software) may

require a large number of iterations to estimate a χ

2

statistic.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

56

To see more detail about the results, go to View → Text Output. A selected portion of the

output is the following:

Notes for Model (Default model)

Computation of degrees of freedom (Default model)

Number of distinct sample moments: 27

Number of distinct parameters to be estimated: 20

Degrees of freedom (27 - 20): 7

Result (Default model)

Minimum was achieved

Chi-square = 9.911

Degrees of freedom = 7

Probability level = .194

Estimates (Group number 1 - Default model)

Scalar Estimates (Group number 1 - Default model)

Maximum Likelihood Estimates

Regression Weights: (Group number 1 - Default model)

Estimate S.E. C.R. P Label

compete <--- Economic 1.200 .369 3.255 .001

govtresp <--- Economic .302 .082 3.695 ***

privtown <--- Economic 1.000

euthanas <--- Morals .633 .045 14.009 ***

abortion <--- Morals .992 .071 13.937 ***

homosex <--- Morals 1.000

govtresp <--- Morals .215 .042 5.070 ***

Standardized Regression Weights: (Group number 1 - Default model)

Estimate

compete <--- Economic .691

govtresp <--- Economic .154

privtown <--- Economic .615

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

57

Estimate

euthanas <--- Morals .566

abortion <--- Morals .737

homosex <--- Morals .677

govtresp <--- Morals .176

Intercepts: (Group number 1 - Default model)

Estimate S.E. C.R. P Label

compete 3.442 .069 49.743 ***

govtresp 4.312 .078 55.366 ***

privtown 3.541 .065 54.587 ***

euthanas 2.645 .071 37.076 ***

abortion 4.360 .086 50.760 ***

homosex 4.774 .095 50.472 ***

Covariances: (Group number 1 - Default model)

Estimate S.E. C.R. P Label

Morals <--> Economic -.016 .134 -.118 .906

Correlations: (Group number 1 - Default model)

Estimate

Morals <--> Economic -.005

Variances: (Group number 1 - Default model)

Estimate S.E. C.R. P Label

Economic 1.902 .604 3.150 .002

Morals 4.850 .485 10.000 ***

d3 2.998 .847 3.542 ***

d2 6.872 .289 23.745 ***

d6 4.115 .213 19.361 ***

d5 4.021 .353 11.407 ***

d4 5.732 .396 14.481 ***

d1 3.118 .596 5.235 ***

Squared Multiple Correlations: (Group number 1 - Default model)

Estimate

homosex .458

abortion .543

euthanas .321

privtown .379

govtresp .054

compete .477

Model Fit Summary

CMIN

Model NPAR CMIN DF P CMIN/DF

Default model 20 9.911 7 .194 1.416

Saturated model 27 .000 0

Independence model 6 910.164 21 .000 43.341

Baseline Comparisons

Model

NFI

Delta1

RFI

rho1

IFI

Delta2

TLI

rho2

CFI

Default model .989 .967 .997 .990 .997

Saturated model 1.000 1.000 1.000

Independence model .000 .000 .000 .000 .000

Parsimony-Adjusted Measures

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

58

Model PRATIO PNFI PCFI

Default model .333 .330 .332

Saturated model .000 .000 .000

Independence model 1.000 .000 .000

NCP

Model NCP LO 90 HI 90

Default model 2.911 .000 15.448

Saturated model .000 .000 .000

Independence model 889.164 794.150 991.574

FMIN

Model FMIN F0 LO 90 HI 90

Default model .008 .002 .000 .013

Saturated model .000 .000 .000 .000

Independence model .759 .742 .662 .827

RMSEA

Model RMSEA LO 90 HI 90 PCLOSE

Default model .019 .000 .043 .987

Independence model .188 .178 .198 .000

AIC

Model AIC BCC BIC CAIC

Default model 49.911 50.146

Saturated model 54.000 54.317

Independence model 922.164 922.235

ECVI

Model ECVI LO 90 HI 90 MECVI

Default model .042 .039 .052 .042

Saturated model .045 .045 .045 .045

Independence model .769 .690 .855 .769

HOELTER

Model

HOELTER

.05

HOELTER

.01

Default model 1702 2235

Independence model 44 52

The overall model fit appears quite good. The χ

2

test yields a statistic of 9.911 (df=7), which

has a corresponding p-value of .194. This p-value is too high to reject the null hypothesis of a

good fit. The RMSEA of .019 and CFI of .997 also suggest that the model fits the data well.

Under the Regression Weights heading, the unstandardized loadings appear along with

standard errors, critical ratios (test statistics), and p-values. A critical ratio greater than 1.96 or

a p-value smaller than .05 signifies the parameter is statistically discernable from zero at the .05

significance level. Three asterisks (

***

) indicate that the p-value is smaller than .001. In this

case all of the unconstrained estimates are statistically significant. Notice that the order of

variables listed in the table is not the same as what is specified in the path diagram; it is a bit

confusing and inconvenient to read results.

8

8

Variables are recognized from the left when created by clicking Draw a latent variable button or add an

indicator to a latent variable button to right, but we rotated the path diagram so that manifest variables are

vertically arranged.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

59

The unstandardized weights are highly sensitive to model constraints, whereas the standardized

regression weights provide more intuitive information about the strength of factor loadings.

The GOVTRESP has low standardized loadings on both factors (.154 for ECONOMIC

and .176 for MORALS), suggesting that it is an unreliable indicator of both economic and

moral values. However, the other indicators have moderate to strong standardized loadings. For

EUTHANAS the factor loading is .566, for PRIVTOWN it is .615, for HOMOSEX it is .677,

for COMPETE it is .691, and for ABORTION it is .737.

The squared multiple correlations provide information about how much variance of an

observed variable the factors can account for. Despite receiving a path from both latent

variables, GOVTRESP has a very low R

2

of only .054. The remaining R

2

statistics are, in order

of increasing magnitude, .321 (EUTHANAS), .379 (PRIVTOWN), .458 (HOMOSEX), .477

(COMPETE), and .543 (ABORTION). Finally, the covariance between the two common

factors is -.012 and their correlation is -.005, which is not statistically distinguishable from

zero.

4.3 CFA with Missing Data using LISREL

Launch LISREL and open the file values_full.sav by choosing File → Import Data. When

Open dialog box opens, change Files of type to SPSS Data File(*.sav), navigate to the folder

C:\temp\CFA, and choose values_full.sav. Click Open to import the data set. A prompt appears

immediately to save the file as a PRELIS system file (.psf). Enter values_full and click Save. A

spreadsheet will open displaying the data. Notice that missing observations are coded

-999999.0.

9

9

To make sure that PRELIS understands these are missing values, it is necessary to declare them as such. Go to

Data → Define Variables to open the Define Variables dialog box. Highlight each variable name by clicking on

PRIVTOWN, holding down the shift key, and clicking on EUTHANAS and others. Click on Missing Values to

bring up the Missing Values box. Click on the Missing Values radio button, enter -999999.0 in the first empty

field, and check the Apply to all option. Click OK, then OK again.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

60

In 3.2, we drew a path diagram for two factor CFA and then generated a corresponding

SIMPLIS or LISREL syntax. This approach is intuitive but wordy to explain each step all the

way to the final. Here let us directly write a SIMPLIS program in the text editor.

Go to File → New, choose SIMPLIS Project in the New dialog box, and click OK. When the

Save As dialog box opens, type in values_full, and then click Save. An empty text window

opens for you to type in the commands shown in the screenshot in the previous page. Notice

that values_full.psf is used as input data and the metric of the two common factors is set by

constraining these factor loadings on PRIVTOWN and HOMOSEX to equal one.

Click on the Run LISREL button to fit the model. The unstandardized estimates are

immediately displayed in the path diagram along with two measures of overall fit: ;

2

and

RMSEA. To view the standardized results, choose Standardized Solution from the Estimates

pull-down menu as we did in 3.2. The path diagrams for unstandardized (left diagram below)

and standardized (right diagram) estimates will look like these:

More detailed information can be obtained by looking at the output text file generated after

estimation. This file is given the same name as the path diagram plus an .out extension and

stored in the working directory. The file values_full.out looks like the following:

DATE: 1/17/2009

TIME: 19:09

L I S R E L 8.80

BY

Karl G. Jöreskog & Dag Sörbom

This program is published exclusively by

Scientific Software International, Inc.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

61

7383 N. Lincoln Avenue, Suite 100

Lincolnwood, IL 60712, U.S.A.

Phone: (800)247-6113, (847)675-0720, Fax: (847)675-2140

Copyright by Scientific Software International, Inc., 1981-2006

Use of this program is subject to the terms specified in the

Universal Copyright Convention.

Website: www.ssicentral.com

The following lines were read from file c: \Temp\CFA\values_full.spj:

Two factor model with missing data

Raw Data from file 'c: \temp\cfa\values_full.psf'

--------------------------------

EM Algorithm for missing Data:

--------------------------------

Number of different missing-value patterns= 12

Convergence of EM-algorithm in 3 iterations

-2 Ln(L) = 33282.55078

Percentage missing values= 0.71

Note:

The Covariances and/or Means to be analyzed are estimated

by the EM procedure and are only used to obtain starting

values for the FIML procedure

Latent Variables Economic Morals

Relationships

privtown = 1.00*Economic

govtresp = Economic Morals

compete = Economic

homosex = 1.00*Morals

abortion = Morals

euthanas = Morals

Path Diagram

End of Problem

Sample Size = 1200

Two factor model with missing values

Covariance Matrix

privtown govtresp compete homosex abortion euthanas

-------- -------- -------- -------- -------- --------

privtown 5.02

govtresp 0.55 7.27

compete 2.28 0.70 5.74

homosex -0.27 1.26 -0.08 10.59

abortion -0.15 0.89 0.00 4.81 8.81

euthanas 0.11 0.64 0.33 3.03 3.08 6.06

Two factor model with missing values

Number of Iterations = 5

LISREL Estimates (Maximum Likelihood)

Measurement Equations

privtown = 1.00*Economic, Errorvar.= 3.12 , R² = 0.38

(0.60)

5.23

govtresp = 0.30*Economic + 0.21*Morals, Errorvar.= 6.87 , R² = 0.054

(0.082) (0.042) (0.29)

3.70 5.07 23.74

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

62

compete = 1.20*Economic, Errorvar.= 3.00 , R² = 0.48

(0.37) (0.85)

3.26 3.54

homosex = 1.00*Morals, Errorvar.= 5.73 , R² = 0.46

(0.40)

14.48

abortion = 0.99*Morals, Errorvar.= 4.02 , R² = 0.54

(0.071) (0.35)

13.94 11.41

euthanas = 0.63*Morals, Errorvar.= 4.11 , R² = 0.32

(0.045) (0.21)

14.01 19.36

Covariance Matrix of Independent Variables

Economic Morals

-------- --------

Economic 1.90

(0.60)

3.15

Morals -0.02 4.85

(0.13) (0.48)

-0.12 10.00

Global Goodness of Fit Statistics, Missing Data Case

-2ln(L) for the saturated model = 33282.551

-2ln(L) for the fitted model = 33292.471

Degrees of Freedom = 7

Full Information ML Chi-Square = 9.92 (P = 0.19)

Root Mean Square Error of Approximation (RMSEA) = 0.019

90 Percent Confidence Interval for RMSEA = (0.0 ; 0.043)

P-Value for Test of Close Fit (RMSEA < 0.05) = 0.99

The result is quite similar to that of Amos. The χ

2

test yields a value of 9.92 (df=7) and

RMSEA is only .019, indicating a good fit. CFI is not, however, reported here. Unconstrained

estimates are shown under the Measurement Equations heading and standardized solutions on

the path diagram. The GOVTRESP has low standardized loadings from both factors (.15 for

ECONOMIC and .18 for MORALS) and a very low R

2

of .054, suggesting that it is a weak

indicator of both economic and moral values. However, the other indicators have moderate to

strong standardized loadings with higher squared multiple correlations ranging from .32 for

EUTHANAS to .54 for ABORTION. Finally, the covariance and correlation between the two

common factors are -.02 and -.01, respectively, which are not statistically distinguishable from

zero.

4.4 CFA with Missing Data using Mplus

Unlike Amos and LISREL, Mplus cannot directly read an SPSS data file. Instead raw data must

be saved as an ASCII file in free or fixed format. Since Mplus does not treat blanks as missing

values, missing values must be coded as a number in the raw data and explicitly specified in the

VARIABLE command of the Mplus syntax.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

63

In order to recode missing values, open SPSS, choose File →New → Syntax to open a SPSS

Syntax Editor window, and then enter the following commands. These two commands recode

missing values as -1 and write a tab-delimited text file values_full.dat in the C:\temp\CFA

directory.

RECODE privtown govtresp compete homosex abortion euthanas (SYSMIS=-1).

SAVE TRANSLATE OUTFILE='c:\temp\CFA\values_full.dat' /TYPE=TAB/MAP.

Now launch Mplus and type in the following syntax for the two factor model with missing

data.

10

TITLE: Two Factor Model with Missing Data;

DATA: FILE IS c:\temp\cfa\values_full.dat;

VARIABLE: NAMES ARE privtown govtresp compete

homosex abortion euthanas;

MISSING ARE ALL(-1);

MODEL: economic BY privtown govtresp compete;

morals BY homosex abortion euthanas govtresp;

OUTPUT: STANDARDIZED;

MODINDICES;

The TITLE line provides a short description for the analysis. The DATA statement specifies

the path name for the tab-delimited raw data file to be analyzed. The VARIABLE statement

lists the names of the variables in the order they appear in the data file. The MISSING ARE

option tells Mplus to interpret the numeric value -1 as missing for all variables. The MODEL

statement tells Mplus that there are two latent variables and six outcome (manifest) variables.

The OUTPUT statement here requests that standardized estimates and modification indices

appear in the output file.

To begin the estimation, click on Run . A text output file appears and is saved in the

working directory. The output file looks like the following:

INPUT INSTRUCTIONS

TITLE: Two Factor Model with Missing Data;

DATA: FILE IS c:\temp\cfa\values_full.dat;

VARIABLE: NAMES ARE privtown govtresp compete

homosex abortion euthanas;

MISSING ARE ALL(-1);

MODEL: economic BY privtown govtresp compete;

morals BY homosex abortion euthanas govtresp;

OUTPUT: STANDARDIZED;

MODINDICES;

INPUT READING TERMINATED NORMALLY

Two Factor Model with Missing Values;

SUMMARY OF ANALYSIS

Number of groups 1

Number of observations 1200

10

In previous edition, you need to add “ANALYSIS: TYPE=MISSING H1;” which becomes a default setting in

version 5. The TYPE = MISSING option in the ANALYSIS statement tells Mplus to use an estimator appropriate

for the presence of missing data, and the h1 requests a chi-square statistic for model fit.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

64

Number of dependent variables 6

Number of independent variables 0

Number of continuous latent variables 2

Observed dependent variables

Continuous

PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS

Continuous latent variables

ECONOMIC MORALS

Estimator ML

Information matrix OBSERVED

Maximum number of iterations 1000

Convergence criterion 0.500D-04

Maximum number of steepest descent iterations 20

Maximum number of iterations for H1 2000

Convergence criterion for H1 0.100D-03

Input data file(s)

c:\temp\cfa\values_full.dat

Input data format FREE

SUMMARY OF DATA

Number of missing data patterns 12

COVARIANCE COVERAGE OF DATA

Minimum covariance coverage value 0.100

PROPORTION OF DATA PRESENT

Covariance Coverage

PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION

________ ________ ________ ________ ________

PRIVTOWN 0.994

GOVTRESP 0.994 0.999

COMPETE 0.994 0.999 0.999

HOMOSEX 0.977 0.981 0.981 0.982

ABORTION 0.988 0.992 0.992 0.976 0.993

EUTHANAS 0.985 0.990 0.990 0.975 0.985

Covariance Coverage

EUTHANAS

________

EUTHANAS 0.991

THE MODEL ESTIMATION TERMINATED NORMALLY

TESTS OF MODEL FIT

Chi-Square Test of Model Fit

Value 9.920

Degrees of Freedom 7

P-Value 0.1932

Chi-Square Test of Model Fit for the Baseline Model

Value 910.923

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

65

Degrees of Freedom 15

P-Value 0.0000

CFI/TLI

CFI 0.997

TLI 0.993

Loglikelihood

H0 Value -16646.235

H1 Value -16641.275

Information Criteria

Number of Free Parameters 20

Akaike (AIC) 33332.470

Bayesian (BIC) 33434.272

Sample-Size Adjusted BIC 33370.744

(n* = (n + 2) / 24)

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.019

90 Percent C.I. 0.000 0.043

Probability RMSEA <= .05 0.987

SRMR (Standardized Root Mean Square Residual)

Value 0.016

MODEL RESULTS

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 1.000 0.000 999.000 999.000

GOVTRESP 0.302 0.082 3.683 0.000

COMPETE 1.201 0.384 3.124 0.002

MORALS BY

HOMOSEX 1.000 0.000 999.000 999.000

ABORTION 0.992 0.072 13.864 0.000

EUTHANAS 0.633 0.046 13.902 0.000

GOVTRESP 0.215 0.042 5.105 0.000

MORALS WITH

ECONOMIC -0.016 0.140 -0.112 0.911

Intercepts

PRIVTOWN 3.541 0.065 54.608 0.000

GOVTRESP 4.312 0.078 55.390 0.000

COMPETE 3.442 0.069 49.763 0.000

HOMOSEX 4.774 0.095 50.494 0.000

ABORTION 4.360 0.086 50.780 0.000

EUTHANAS 2.645 0.071 37.091 0.000

Variances

ECONOMIC 1.900 0.627 3.029 0.002

MORALS 4.850 0.486 9.974 0.000

Residual Variances

PRIVTOWN 3.120 0.619 5.041 0.000

GOVTRESP 6.872 0.289 23.738 0.000

COMPETE 2.995 0.882 3.395 0.001

HOMOSEX 5.731 0.398 14.409 0.000

ABORTION 4.022 0.352 11.416 0.000

EUTHANAS 4.115 0.212 19.419 0.000

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

66

STANDARDIZED MODEL RESULTS

STDYX Standardization

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 0.615 0.100 6.167 0.000

GOVTRESP 0.154 0.037 4.202 0.000

COMPETE 0.691 0.111 6.217 0.000

MORALS BY

HOMOSEX 0.677 0.027 24.755 0.000

ABORTION 0.737 0.027 26.842 0.000

EUTHANAS 0.566 0.027 20.849 0.000

GOVTRESP 0.176 0.034 5.241 0.000

MORALS WITH

ECONOMIC -0.005 0.046 -0.112 0.911

Intercepts

PRIVTOWN 1.580 0.043 36.432 0.000

GOVTRESP 1.600 0.044 36.690 0.000

COMPETE 1.437 0.041 34.904 0.000

HOMOSEX 1.468 0.042 35.047 0.000

ABORTION 1.470 0.042 35.166 0.000

EUTHANAS 1.075 0.036 29.531 0.000

Variances

ECONOMIC 1.000 0.000 999.000 999.000

MORALS 1.000 0.000 999.000 999.000

Residual Variances

PRIVTOWN 0.622 0.123 5.065 0.000

GOVTRESP 0.946 0.016 58.909 0.000

COMPETE 0.522 0.154 3.397 0.001

HOMOSEX 0.542 0.037 14.626 0.000

ABORTION 0.457 0.040 11.307 0.000

EUTHANAS 0.679 0.031 22.080 0.000

STDY Standardization

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 0.615 0.100 6.167 0.000

GOVTRESP 0.154 0.037 4.202 0.000

COMPETE 0.691 0.111 6.217 0.000

MORALS BY

HOMOSEX 0.677 0.027 24.755 0.000

ABORTION 0.737 0.027 26.842 0.000

EUTHANAS 0.566 0.027 20.849 0.000

GOVTRESP 0.176 0.034 5.241 0.000

MORALS WITH

ECONOMIC -0.005 0.046 -0.112 0.911

Intercepts

PRIVTOWN 1.580 0.043 36.432 0.000

GOVTRESP 1.600 0.044 36.690 0.000

COMPETE 1.437 0.041 34.904 0.000

HOMOSEX 1.468 0.042 35.047 0.000

ABORTION 1.470 0.042 35.166 0.000

EUTHANAS 1.075 0.036 29.531 0.000

Variances

ECONOMIC 1.000 0.000 999.000 999.000

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

67

MORALS 1.000 0.000 999.000 999.000

Residual Variances

PRIVTOWN 0.622 0.123 5.065 0.000

GOVTRESP 0.946 0.016 58.909 0.000

COMPETE 0.522 0.154 3.397 0.001

HOMOSEX 0.542 0.037 14.626 0.000

ABORTION 0.457 0.040 11.307 0.000

EUTHANAS 0.679 0.031 22.080 0.000

STD Standardization

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 1.378 0.228 6.057 0.000

GOVTRESP 0.416 0.100 4.159 0.000

COMPETE 1.656 0.271 6.117 0.000

MORALS BY

HOMOSEX 2.202 0.110 19.949 0.000

ABORTION 2.185 0.102 21.355 0.000

EUTHANAS 1.394 0.080 17.423 0.000

GOVTRESP 0.474 0.092 5.158 0.000

MORALS WITH

ECONOMIC -0.005 0.046 -0.112 0.911

Intercepts

PRIVTOWN 3.541 0.065 54.608 0.000

GOVTRESP 4.312 0.078 55.390 0.000

COMPETE 3.442 0.069 49.763 0.000

HOMOSEX 4.774 0.095 50.494 0.000

ABORTION 4.360 0.086 50.780 0.000

EUTHANAS 2.645 0.071 37.091 0.000

Variances

ECONOMIC 1.000 0.000 999.000 999.000

MORALS 1.000 0.000 999.000 999.000

Residual Variances

PRIVTOWN 3.120 0.619 5.041 0.000

GOVTRESP 6.872 0.289 23.738 0.000

COMPETE 2.995 0.882 3.395 0.001

HOMOSEX 5.731 0.398 14.409 0.000

ABORTION 4.022 0.352 11.416 0.000

EUTHANAS 4.115 0.212 19.419 0.000

R-SQUARE

Observed Two-Tailed

Variable Estimate S.E. Est./S.E. P-Value

PRIVTOWN 0.378 0.123 3.083 0.002

GOVTRESP 0.054 0.016 3.388 0.001

COMPETE 0.478 0.154 3.108 0.002

HOMOSEX 0.458 0.037 12.378 0.000

ABORTION 0.543 0.040 13.421 0.000

EUTHANAS 0.321 0.031 10.425 0.000

QUALITY OF NUMERICAL RESULTS

Condition Number for the Information Matrix 0.778E-03

(ratio of smallest to largest eigenvalue)

MODEL MODIFICATION INDICES

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

68

Minimum M.I. value for printing the modification index 10.000

M.I. E.P.C. Std E.P.C. StdYX E.P.C.

No modification indices above the minimum value.

The results are quite similar to those of Amos and LISREL. Under the TESTS OF MODEL FIT

heading, the χ

2

test yields a large value of 9.920 with 7 degrees of freedom. The p-value

of .1932 is too high to reject the null hypothesis of a good fit. The RMSEA and CFI are .019

and .997, indicating that the model fits the data well.

Unstandardized estimates are listed under the MODEL RESULTS heading and standardized

estimates under STDYX Standardization of the STANDARDIZED MODEL RESULTS heading. All

parameters are statistically discernable from zero; test statistics under the Est./S.E. column

are large and corresponding p-values are small enough to reject the null hypothesis at the .05

level. Standardized estimates can be interpreted as the correlation between the latent and

observed variables. GOVTRESP has low standardized loadings of .154 and .176 on

ECONOMIC and MORALS, respectively; only 5.4 percent of variance in GOVTRESP is

accounted for by the two latent variables. Other factor loadings range from .566 for

EUTHANAS (R

2

= .321) and .737 for ABORTION (R

2

= .543). The covariance and correlation

between the two common factors are negligible -.016 and -.005, respectively.

4.5 Summary

Table 6 compares the unstandardized and standardized estimates that each statistical software

package produced. Standard errors appear in parentheses. The final row lists goodness of fit

measures of χ

2

, RMSEA, and CFI. The results are essentially identical across programs.

Table 6: Two Factor Model with Missing Data (N=1,200, DF=7)

Amos LISREL Mplus

Unstd. Standard Unstd. Standard Unstd. Standard

F1 PRIVTOWN

1.00 .615

1.00 .62 1.000 .615

(.100)

F1 GOVTRESP

.302

(.082)

.154 .30

(.082)

.15 .302

(.082)

.154

(.037)

F1 COMPETE

1.200

(.369)

.691 1.20

(.37)

.69 1.201

(.384)

.691

(.111)

F2 HOMOSEX

1.00 .677 1.00 .68 1.000 .677

(.027)

F2 ABORTION

.992

(.071)

.737 .99

(.071)

.74 .992

(.072)

.737

.027)

F2 EUTHANAS

.633

(.045)

.566 .63

(.045)

.57 .633

(.046)

.566

(.027)

F2 GOVTRESP

.215

(.042)

.176 .21

(.042)

.18 .215

(.042)

.176

(.034)

F1 F2

-.012

a

(.102)

-.005

b

-.02

a

(.13)

-.01

b

-.016

a

(.140)

-.005

b

(.046)

;

2

, RMSEA, CFI

9.911

.019

.997

9.92

.019

9.920

.019

.997

Standard errors appear in parentheses

a

covariance;

b

correlation between F1 (economic values) and F2 (moral values)

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

69

5. CFA with Categorical indicators

This section estimates a confirmatory factor model using the polychoric correlation matrix in

the presence of categorical manifest variables. Latent variables are assumed to represent

continuous (not categorical) constructs. For illustration, original responses ranging from 1 to 3

were recoded as 1; those ranging from 4 to 7 were recoded as 2; and those ranging from 8 to 10

were recoded as 3. Forty observations with missing data are dropped in order to focus only on

the problem of categorical outcome variables in the confirmatory factor model. The recoded

data are stored into the SPSS file values_ord.sav (N=1,160) in the C:\temp\CFA directory.

Values_full.sav (N=1,200) is also used to show how pairwise deletion usesas much information

in the raw data file as possible.

5.1 Background

The maximum likelihood estimation (MLE) approach used in the previous sections relied on

the strong assumption of multivariate normality. In practice, a substantial amount of social

science data is non-normal. Survey responses are often coded as yes/no or as scores on an

ordered scale (e.g. strongly disagree, disagree, neutral, agree, strongly agree). In the presence of

categorical or ordinal data, MLE may not work properly, calling for alternative estimation

methods.

Mplus, LISREL, and SAS/STAT CALIS employ a multi-step method for ordinal outcome

variables that analyzes a matrix of polychoric correlations rather than covariances. This

approach works as follows: 1) thresholds are estimated by maximum likelihood, 2) these

estimates are used to estimate a polychoric correlation matrix, which in turn is used to 3)

estimate parameters through (diagonally) weighted least squares using the inverse of the

asymptotic covariance matrix as the weight matrix (Muthén, 1984; Jöreskog, 1990).

In LISREL, the diagonally weighted least squares (DWLS) method needs to be specified.

Alternatively, the polychoric correlation matrix and asymptotic covariance matrix is estimated

and saved into a LISREL system file (.dsf) using PRELIS before fitting the model. Mplus

automatically follows above steps when the syntax includes a line identifying observed

variables as categorical. In SAS/STAT CALIS, the METHOD=WLS or METHOD=DWLS option of the

PROC CALIS statement tells SAS to fit the model using the WLS or DWLS method instead of

MLE.

5.2 CFA with Categorical Indicators Using LISREL

Launch LISREL and load the data by choosing File → Import Data. Change the file type to

SPSS Data File (*.sav) in the Open dialog box, navigate to the C:\temp\CFA folder, choose the

file values_ord.sav, and click Open. When prompted from the Save As dialog box, enter the

name values_ord in the File Name field and click Save. The data will then be displayed in

spreadsheet form.

11

11

You may explicitly define the variables as ordinal by going to Data → Define Variables, highlighting all the

variables, and opening the Variable Type menu.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

70

To open an empty text window, go to File → New, choose SIMPLIS Project, and click OK.

When prompted in the Save As dialog box, type in values_ord, and then click Save. Enter the

following commands and save the file. Notice that Method of Estimation: Diagonally

Weighted Least Squares tells LISREL to fit WLS for categorical observed variables.

12

Two factor model CFA for Ordinal Indicators

SYSTEM FILE from file 'c:\Temp\CFA\values_ord.dsf'

Latent Variables Economic Morals

Relationships

privtown = 1.00*Economic

govtresp = Economic Morals

compete = Economic

homosex = 1.00*Morals

abortion = Morals

euthanas = Morals

Path Diagram

Method of Estimation: Diagonally Weighted Least Squares

End of Problem

Click F5 or the Run LISREL button to fit the model (alternatively click on File → Run).

The unstandardized solution appears in the path diagram along with ;

2

and RMSEA statistics.

To view the standardized estimates, choose Standardized Solution from the Estimates

drop-down menu. The following display path diagrams for unstandardized (left) and

standardized (right) estimates.

Detailed information can be read from the text output file values_ordinal.out. The following is

selected from the output file.

Sample Size = 1160

Two factor model CFA for Ordinal Indicators

12

To request the weighted least squares estimator on menu, go to Output → SIMPLIS Outputs. Choose

Diagonally Weighted Least Squares.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

71

Correlation Matrix

privtown govtresp compete homosex abortion euthanas

-------- -------- -------- -------- -------- --------

privtown 1.00

govtresp 0.15 1.00

compete 0.52 0.20 1.00

homosex -0.05 0.18 -0.03 1.00

abortion -0.02 0.16 -0.01 0.58 1.00

euthanas 0.08 0.15 0.11 0.47 0.51 1.00

Two factor model CFA for Ordinal Indicators

Number of Iterations = 6

LISREL Estimates (Robust Diagonally Weighted Least Squares)

Measurement Equations

privtown = 1.00*Economic, Errorvar.= 0.61 , R² = 0.39

(0.13)

4.61

govtresp = 0.38*Economic + 0.30*Morals, Errorvar.= 0.89 , R² = 0.11

(0.079) (0.055) (0.065)

4.77 5.51 13.78

compete = 1.32*Economic, Errorvar.= 0.32 , R² = 0.68

(0.37) (0.20)

3.56 1.60

homosex = 1.00*Morals, Errorvar.= 0.46 , R² = 0.54

(0.078)

5.87

abortion = 1.06*Morals, Errorvar.= 0.39 , R² = 0.61

(0.078) (0.080)

13.56 4.91

euthanas = 0.88*Morals, Errorvar.= 0.58 , R² = 0.42

(0.068) (0.077)

12.93 7.56

Covariance Matrix of Independent Variables

Economic Morals

-------- --------

Economic 0.39

(0.12)

3.34

Morals 0.01 0.54

(0.02) (0.05)

0.30 10.68

Goodness of Fit Statistics

Degrees of Freedom = 7

Normal Theory Weighted Least Squares Chi-Square = 33.75 (P = 0.00)

Satorra-Bentler Scaled Chi-Square = 13.43 (P = 0.062)

Chi-Square Corrected for Non-Normality = 13.93 (P = 0.052)

Estimated Non-centrality Parameter (NCP) = 6.43

90 Percent Confidence Interval for NCP = (0.0 ; 20.89)

Minimum Fit Function Value = 0.010

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

72

Population Discrepancy Function Value (F0) = 0.0055

90 Percent Confidence Interval for F0 = (0.0 ; 0.018)

Root Mean Square Error of Approximation (RMSEA) = 0.028

90 Percent Confidence Interval for RMSEA = (0.0 ; 0.051)

P-Value for Test of Close Fit (RMSEA < 0.05) = 0.94

Expected Cross-Validation Index (ECVI) = 0.036

90 Percent Confidence Interval for ECVI = (0.030 ; 0.048)

ECVI for Saturated Model = 0.036

ECVI for Independence Model = 1.25

Chi-Square for Independence Model with 15 Degrees of Freedom = 1440.45

Independence AIC = 1452.45

Model AIC = 41.43

Saturated AIC = 42.00

Independence CAIC = 1488.79

Model CAIC = 126.22

Saturated CAIC = 169.18

Normed Fit Index (NFI) = 0.99

Non-Normed Fit Index (NNFI) = 0.99

Parsimony Normed Fit Index (PNFI) = 0.46

Comparative Fit Index (CFI) = 1.00

Incremental Fit Index (IFI) = 1.00

Relative Fit Index (RFI) = 0.98

Critical N (CN) = 1595.43

Root Mean Square Residual (RMR) = 0.032

Standardized RMR = 0.032

Goodness of Fit Index (GFI) = 1.00

Adjusted Goodness of Fit Index (AGFI) = 0.99

Parsimony Goodness of Fit Index (PGFI) = 0.33

The Modification Indices Suggest to Add the

Path to from Decrease in Chi-Square New Estimate

homosex Economic 81.6 -3.40

euthanas Economic 21.3 0.51

The Modification Indices Suggest to Add an Error Covariance

Between and Decrease in Chi-Square New Estimate

euthanas compete 8.2 0.16

This model appears to fit the data moderately. The Satorra-Bentler Scaled χ

2

statistic has a large

value of 13.43 with seven degrees of freedom. But its corresponding p-value is .062; we do not

reject the null hypothesis of a good fit marginally at the .05 level. However, RMSEA and CFI

are respectively .028 and 1.00, indicating a very good model fit.

Under the Measurement Equations heading appear the unstandardized estimates, standard

errors, t-values, and R

2

statistics. All of the unconstrained path coefficients are large enough to

be statistically significant. Despite receiving a path from both ECONOMIC and MORALS,

GOVTRESP has the smallest R

2

(.11). The other observed variables have moderate to high R

2

statistics, ranging from .39 (PRIVTOWN) to .68 (COMPETE).

The standardized estimates of the loadings displayed in the path diagram range from .22

(MORALS) and .24 (ECONOMIC) for GOVTRESP to .63 for PRIVTOWN and .83 for

COMPETE. This result a bit differs from one in previous models. Finally, the covariance and

correlation between the two latent variables are .01, which is not statistically discernable from

zero.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

73

5.3 CFA with Categorical Indicators using Mplus

Mplus cannot directly read data from an SPSS system file. Raw data must come from a free or

fixed format text file. See sections 4.4 for syntax to translate an SPSS file to an ASCII file. For

this example the data is saved as the tab-delimited file values_ord.dat in the C:\temp\CFA

folder.

After launching Mplus, the syntax editor appears. The following commands are used to

estimate the confirmatory factor model with ordinal observed variables.

TITLE: Two factor model CFA with ordinal indicators (listwise);

DATA: FILE IS c:\Temp\CFA\values_ord.dat;

VARIABLE: NAMES ARE privtown govtresp compete

homosex abortion euthanas;

CATEGORICAL ARE privtown govtresp compete

homosex abortion euthanas;

MODEL: economic BY privtown govtresp compete;

morals BY homosex abortion euthanas govtresp;

OUTPUT: Standardized;

The syntax is similar to previous sections except that CATEGORICAL ARE is added to the

VARIABLE statement to define the observed variables as categorical. Mplus by default

assumes all observed variables are continuous unless otherwise specified. When categorical

indicators are declared, Mplus employs by default a robust weighted least squares estimator

similar to the Diagonally Weighted Least Squares estimator in LISREL. Click on Run to

carry out the estimation. A selection of the output file is the following:

SUMMARY OF ANALYSIS

Number of groups 1

Number of observations 1160

Number of dependent variables 6

Number of independent variables 0

Number of continuous latent variables 2

Observed dependent variables

Binary and ordered categorical (ordinal)

PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS

Continuous latent variables

ECONOMIC MORALS

Estimator WLSMV

Maximum number of iterations 1000

Convergence criterion 0.500D-04

Maximum number of steepest descent iterations 20

Parameterization DELTA

Input data file(s)

c:\Temp\CFA\values_ord.dat

Input data format FREE

SUMMARY OF CATEGORICAL DATA PROPORTIONS

…

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

74

THE MODEL ESTIMATION TERMINATED NORMALLY

TESTS OF MODEL FIT

Chi-Square Test of Model Fit

Value 15.516*

Degrees of Freedom 6**

P-Value 0.0166

* The chi-square value for MLM, MLMV, MLR, ULSMV, WLSM and WLSMV cannot be used

for chi-square difference tests. MLM, MLR and WLSM chi-square difference

testing is described in the Mplus Technical Appendices at www.statmodel.com.

See chi-square difference testing in the index of the Mplus User's Guide.

** The degrees of freedom for MLMV, ULSMV and WLSMV are estimated according to

a formula given in the Mplus Technical Appendices at www.statmodel.com.

See degrees of freedom in the index of the Mplus User's Guide.

Chi-Square Test of Model Fit for the Baseline Model

Value 1054.754

Degrees of Freedom 11

P-Value 0.0000

CFI/TLI

CFI 0.991

TLI 0.983

Number of Free Parameters 20

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.037

WRMR (Weighted Root Mean Square Residual)

Value 0.671

MODEL RESULTS

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 1.000 0.000 999.000 999.000

GOVTRESP 0.375 0.073 5.166 0.000

COMPETE 1.313 0.332 3.949 0.000

MORALS BY

HOMOSEX 1.000 0.000 999.000 999.000

ABORTION 1.059 0.071 14.866 0.000

EUTHANAS 0.878 0.058 15.040 0.000

GOVTRESP 0.304 0.055 5.534 0.000

MORALS WITH

ECONOMIC 0.006 0.022 0.254 0.800

Thresholds

…

Variances

ECONOMIC 0.394 0.105 3.765 0.000

MORALS 0.543 0.045 12.113 0.000

STANDARDIZED MODEL RESULTS

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

75

STDYX Standardization

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 0.628 0.083 7.530 0.000

GOVTRESP 0.236 0.041 5.788 0.000

COMPETE 0.825 0.104 7.936 0.000

MORALS BY

HOMOSEX 0.737 0.030 24.226 0.000

ABORTION 0.780 0.030 25.647 0.000

EUTHANAS 0.647 0.035 18.679 0.000

GOVTRESP 0.224 0.040 5.654 0.000

MORALS WITH

ECONOMIC 0.012 0.048 0.254 0.799

Thresholds

…

Variances

ECONOMIC 1.000 0.000 999.000 999.000

MORALS 1.000 0.000 999.000 999.000

STDY Standardization

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 0.628 0.083 7.530 0.000

GOVTRESP 0.236 0.041 5.788 0.000

COMPETE 0.825 0.104 7.936 0.000

MORALS BY

HOMOSEX 0.737 0.030 24.226 0.000

ABORTION 0.780 0.030 25.647 0.000

EUTHANAS 0.647 0.035 18.679 0.000

GOVTRESP 0.224 0.040 5.654 0.000

MORALS WITH

ECONOMIC 0.012 0.048 0.254 0.799

Thresholds

…

Variances

ECONOMIC 1.000 0.000 999.000 999.000

MORALS 1.000 0.000 999.000 999.000

STD Standardization

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 0.628 0.083 7.530 0.000

GOVTRESP 0.236 0.041 5.788 0.000

COMPETE 0.825 0.104 7.936 0.000

MORALS BY

HOMOSEX 0.737 0.030 24.226 0.000

ABORTION 0.780 0.030 25.647 0.000

EUTHANAS 0.647 0.035 18.679 0.000

GOVTRESP 0.224 0.040 5.654 0.000

MORALS WITH

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

76

ECONOMIC 0.012 0.048 0.254 0.799

Thresholds

…

Variances

ECONOMIC 1.000 0.000 999.000 999.000

MORALS 1.000 0.000 999.000 999.000

R-SQUARE

Observed Two-Tailed Residual

Variable Estimate S.E. Est./S.E. P-Value Variance

PRIVTOWN 0.394 0.105 3.765 0.000 0.606

GOVTRESP 0.107 0.025 4.200 0.000 0.893

COMPETE 0.680 0.171 3.968 0.000 0.320

HOMOSEX 0.543 0.045 12.113 0.000 0.457

ABORTION 0.608 0.047 12.824 0.000 0.392

EUTHANAS 0.418 0.045 9.339 0.000 0.582

QUALITY OF NUMERICAL RESULTS

Condition Number for the Information Matrix 0.103E-02

(ratio of smallest to largest eigenvalue)

Both LISREL and Mplus (listwise deletion) produce virtually identical results. The χ

2

test

yields a value of 15.516, but a corresponding p-value is .0166, which is small enough to reject

the null hypothesis of a good fit at the .05 significance level. As explained in the output, Mplus

estimates 6 degrees of freedom, which is one smaller than those reported by LISREL. However,

a small RMSEA of .037 and a large CFI of .991 suggest that the model fits the data well as a

whole.

Under the MODEL RESULTS heading, the unstandardized loadings appear along with standard

errors, the ratio of the estimates to their standard errors, and p-values. Since all absolute values

of the numbers in the Est./S.E. column are greater than 1.96, all of the unconstrained

loadings estimates are significant at the .05 level.

Standardized estimates are provided under the STDYX Standardization heading. GOVTRESP

has relatively low standardized loadings on both factors (.236 for ECONOMIC and .224 for

MORALS). For PRIVTOWN the loading is .628, for COMPETE it is .825, for HOMOSEX it

is .737, for ABORTION it is .780, and for EUTHANAS it is .647. GOVTRESP has a low

squared multiple correlation of .107. The remaining R

2

statistics are, in order of increasing

magnitude, .394 (PRIVTOWN), .418 (EUTHANAS), .543 (HOMOSEX), .608 (ABORTION),

and .680 (COMPETE). Finally, the covariance and correlation between the two common

factors are .006 and .012, respectively.

5.4 CFA with categorical indicators and missing data

The previous two subsections explained how to estimate the confirmatory factor model when

the observed variables represent ordered categories. A total of 40 observations with missing

values on at least one indicator were dropped in the listwise deletion, reducing the original

sample of 1,200 to 1,160. It is possible to maximize the information available in the raw data

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

77

file using pairwise rather than listwise deletion for missing data. In pairwise deletion,

correlations will be estimated using all observations with complete data available on both

variables. The following Mplus syntax use values_full.dat instead of values_ord.dat.

TITLE: Two factor model CFA with ordinal indicators (pairwise);

DATA: FILE IS c:\Temp\CFA\values_full.dat;

VARIABLE: NAMES ARE privtown govtresp compete

homosex abortion euthanas;

CATEGORICAL ARE privtown govtresp compete

homosex abortion euthanas;

MISSING ARE all (-1);

MODEL: economic BY privtown govtresp compete;

morals BY homosex abortion euthanas govtresp;

OUTPUT: Standardized;

Notice that MISSING ARE all (-1) is added to the VARIABLE command in order to use pairwise

deletion. The following is a selection of Mplus output.

SUMMARY OF ANALYSIS

Number of groups 1

Number of observations 1200

Number of dependent variables 6

Number of independent variables 0

Number of continuous latent variables 2

Observed dependent variables

Binary and ordered categorical (ordinal)

PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS

Continuous latent variables

ECONOMIC MORALS

Estimator WLSMV

Maximum number of iterations 1000

Convergence criterion 0.500D-04

Maximum number of steepest descent iterations 20

Maximum number of iterations for H1 2000

Convergence criterion for H1 0.100D-03

Parameterization DELTA

Input data file(s)

c:\Temp\CFA\values_full.dat

Input data format FREE

SUMMARY OF DATA

Number of missing data patterns 12

COVARIANCE COVERAGE OF DATA

Minimum covariance coverage value 0.100

PROPORTION OF DATA PRESENT

Covariance Coverage

PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

78

________ ________ ________ ________ ________

PRIVTOWN 0.994

GOVTRESP 0.994 0.999

COMPETE 0.994 0.999 0.999

HOMOSEX 0.977 0.981 0.981 0.982

ABORTION 0.988 0.992 0.992 0.976 0.993

EUTHANAS 0.985 0.990 0.990 0.975 0.985

Covariance Coverage

EUTHANAS

________

EUTHANAS 0.991

THE MODEL ESTIMATION TERMINATED NORMALLY

TESTS OF MODEL FIT

Chi-Square Test of Model Fit

Value 10.417*

Degrees of Freedom 6**

P-Value 0.1081

* The chi-square value for MLM, MLMV, MLR, ULSMV, WLSM and WLSMV cannot be used

for chi-square difference tests. MLM, MLR and WLSM chi-square difference

testing is described in the Mplus Technical Appendices at www.statmodel.com.

See chi-square difference testing in the index of the Mplus User's Guide.

** The degrees of freedom for MLMV, ULSMV and WLSMV are estimated according to

a formula given in the Mplus Technical Appendices at www.statmodel.com.

See degrees of freedom in the index of the Mplus User's Guide.

Chi-Square Test of Model Fit for the Baseline Model

Value 1528.615

Degrees of Freedom 10

P-Value 0.0000

CFI/TLI

CFI 0.997

TLI 0.995

Number of Free Parameters 62

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.025

WRMR (Weighted Root Mean Square Residual)

Value 0.336

MODEL RESULTS

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 1.000 0.000 999.000 999.000

GOVTRESP 0.380 0.051 7.416 0.000

COMPETE 1.352 0.258 5.245 0.000

MORALS BY

HOMOSEX 1.000 0.000 999.000 999.000

ABORTION 1.066 0.053 20.239 0.000

EUTHANAS 0.930 0.045 20.684 0.000

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

79

GOVTRESP 0.272 0.044 6.207 0.000

MORALS WITH

ECONOMIC 0.014 0.018 0.801 0.423

Thresholds

…

Variances

ECONOMIC 0.354 0.070 5.060 0.000

MORALS 0.520 0.033 15.803 0.000

STANDARDIZED MODEL RESULTS

STDYX Standardization

Two-Tailed

Estimate S.E. Est./S.E. P-Value

ECONOMIC BY

PRIVTOWN 0.595 0.059 10.120 0.000

GOVTRESP 0.226 0.029 7.818 0.000

COMPETE 0.805 0.077 10.394 0.000

MORALS BY

HOMOSEX 0.721 0.023 31.607 0.000

ABORTION 0.769 0.023 33.214 0.000

EUTHANAS 0.670 0.027 25.218 0.000

GOVTRESP 0.196 0.031 6.236 0.000

MORALS WITH

ECONOMIC 0.033 0.040 0.811 0.417

Thresholds

…

Variances

ECONOMIC 1.000 0.000 999.000 999.000

MORALS 1.000 0.000 999.000 999.000

R-SQUARE

Observed Two-Tailed Residual

Variable Estimate S.E. Est./S.E. P-Value Variance

PRIVTOWN 0.354 0.070 5.060 0.000 0.646

GOVTRESP 0.093 0.018 5.241 0.000 0.907

COMPETE 0.648 0.125 5.197 0.000 0.352

HOMOSEX 0.520 0.033 15.803 0.000 0.480

ABORTION 0.591 0.036 16.607 0.000 0.409

EUTHANAS 0.449 0.036 12.609 0.000 0.551

This CFA model fits the data quite well. The χ

2

of 10.417 with 6 degrees of freedom and its

p-value of .1081 do not reject the null hypothesis of a good fit. A small RMSEA of .025 and a

large CFI of .997 also suggest that the model fits the data well.

All unstandardized and standardized estimates have large test statistics and appear to be

statistically significant. GOVTRESP has lowest factor loadings of .226 from economic values

and .196 for moral values. Other factor loadings range from .595 (PRIVTOWN) to .805

(COMPETE). Accordingly, R

2

ranges from .093 for GOVTRESP to .354 for PRIVTOWN

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

80

and .648 for COMPETE. The covariance and correlation between two latent variables are .014

and .033, respectively.

5.5 CFA with categorical indicators Using SAS/STAT CALIS

SAS/STAT CALIS can also fit a confirmatory factor model with categorical indicators using

(diagonally) weighted least squares estimation. In SPSS, you may convert the data set

values_ord.sav into the SAS data set by choosing Save As and select SAS v7+Windows long

extension (*.sas7bdat) as a file type.

The following is the SAS script for this two factor CFA model with ordinal data. Notice that

METHOD=WLS in the PROC CALIS statement tells SAS to employ the WLS method instead of

the default MLE.

13

Other statements for specification remain unchanged.

ODS HTML;

PROC CALIS DATA=sas.values_ord METHOD=WLS PALL;

VAR privtown govtresp compete homosex abortion euthanas;

LINEQS privtown = 1.0 f1 + e1,

govtresp = lambda2 f1 + lambda7 f2 + e2,

compete = lambda3 f1 + e3,

homosex = 1.0 f2 + e4,

abortion = lambda5 f2 + e5,

euthanas = lambda6 f2 + e6;

STD f1-f2 = phi1-phi2,

e1-e6 = theta1-theta6;

COV f1-f2 = phi3;

RUN;

ODS HTML CLOSE;

The following is a selection of the SAS output. SAS reports 21 knowns provided by the input

covariance matrix and 14 unknown parameters to be estimated; degrees of freedom are 7 = 21 -

14. Basic statistics and correlation matrix display first.

The CALIS Procedure

Covariance Structure Analysis: Weighted Least-Squares Estimation

Observations 1160 Model Terms 1

Variables 6 Model Matrices 4

Informations 21 Parameters 14

Variable Mean Std Dev Skewness Kurtosis

privtown 1.52672 0.60852 0.70316 -0.47100

govtresp 1.68103 0.71347 0.55257 -0.89216

13

METHOD=DWLS uses diagonally weighted least squares method but, despite successful convergence, failed to

report goodness of fit measures in this example.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

81

Variable Mean Std Dev Skewness Kurtosis

compete 1.49138 0.63267 0.92321 -0.21119

homosex 1.85345 0.78108 0.26214 -1.31825

abortion 1.74052 0.74246 0.45744 -1.07136

euthanas 1.33276 0.60023 1.62442 1.49187

Correlations

privtown govtresp compete homosex abortion euthanas

privtown 1.0000 0.1051 0.3917 -0.0390 -0.0123 0.0489

govtresp 0.1051 1.0000 0.1315 0.1421 0.1205 0.0909

compete 0.3917 0.1315 1.0000 -0.0218 -0.0057 0.0689

homosex -0.0390 0.1421 -0.0218 1.0000 0.4670 0.3305

abortion -0.0123 0.1205 -0.0057 0.4670 1.0000 0.3604

euthanas 0.0489 0.0909 0.0689 0.3305 0.3604 1.0000

Fit Function 0.0114

Goodness of Fit Index (GFI) 1.0000

GFI Adjusted for Degrees of Freedom (AGFI) 0.9999

Root Mean Square Residual (RMR) 0.0220

Parsimonious GFI (Mulaik, 1989) 0.4667

Chi-Square 13.2625

Chi-Square DF 7

Pr > Chi-Square 0.0660

Independence Model Chi-Square 674.00

Independence Model Chi-Square DF 15

RMSEA Estimate 0.0278

RMSEA 90% Lower Confidence Limit .

RMSEA 90% Upper Confidence Limit 0.0504

ECVI Estimate 0.0356

ECVI 90% Lower Confidence Limit .

ECVI 90% Upper Confidence Limit 0.0480

Probability of Close Fit 0.9462

Bentler's Comparative Fit Index 0.9905

Akaike's Information Criterion -0.7375

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

82

Bozdogan's (1987) CAIC -43.1308

Schwarz's Bayesian Criterion -36.1308

McDonald's (1989) Centrality 0.9973

Bentler & Bonett's (1980) Non-normed Index 0.9796

Bentler & Bonett's (1980) NFI 0.9803

James, Mulaik, & Brett (1982) Parsimonious NFI 0.4575

Z-Test of Wilson & Hilferty (1931) 1.5106

Bollen (1986) Normed Index Rho1 0.9578

Bollen (1988) Non-normed Index Delta2 0.9906

Hoelter's (1983) Critical N 1231

Manifest Variable Equations with Estimates

privtown = 1.0000 f1 + 1.0000 e1

govtresp = 0.3609 * f1 + 0.2768 * f2 + 1.0000 e2

Std Err 0.0831 lambda2 0.0546 lambda7

t Value 4.3441 5.0658

compete = 1.2040 * f1 + 1.0000 e3

Std Err 0.3369 lambda3

t Value 3.5742

homosex = 1.0000 f2 + 1.0000 e4

abortion = 1.0367 * f2 + 1.0000 e5

Std Err 0.0811 lambda5

t Value 12.7824

euthanas = 0.7616 * f2 + 1.0000 e6

Std Err 0.0633 lambda6

t Value 12.0318

Variances of Exogenous Variables

Variable Parameter Estimate Standard

Error

t Value

f1 phi1 0.31984 0.09545 3.35

f2 phi2 0.45389 0.04364 10.40

e1 theta1 0.68016 0.09550 7.12

e2 theta2 0.92369 0.02047 45.12

e3 theta3 0.53632 0.13081 4.10

e4 theta4 0.54611 0.04374 12.48

e5 theta5 0.51219 0.04580 11.18

e6 theta6 0.73673 0.03423 21.52

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

83

Covariances Among Exogenous Variables

Var1 Var2 Parameter Estimate Standard

Error

t Value

f1 f2 phi3 -0.0006067 0.01846 -0.03

Manifest Variable Equations with Standardized Estimates

privtown = 0.5655 f1 + 0.8247 e1

govtresp = 0.2041 * f1 + 0.1865 * f2 + 0.9611 e2

lambda2 lambda7

compete = 0.6809 * f1 + 0.7323 e3

lambda3

homosex = 0.6737 f2 + 0.7390 e4

abortion = 0.6984 * f2 + 0.7157 e5

lambda5

euthanas = 0.5131 * f2 + 0.8583 e6

lambda6

Squared Multiple Correlations

Variable Error Variance Total Variance R-Square

1 privtown 0.68016 1.00000 0.3198

2 govtresp 0.92369 1.00000 0.0763

3 compete 0.53632 1.00000 0.4637

4 homosex 0.54611 1.00000 0.4539

5 abortion 0.51219 1.00000 0.4878

6 euthanas 0.73673 1.00000 0.2633

Correlations Among Exogenous Variables

Var1 Var2 Parameter Estimate

f1 f2 phi3 -0.00159

The results are a bit different from those of LISREL and Mplus. The χ

2

is 13.2625 with 7

degrees of freedom and a corresponding p-value is .0660 that indicates a moderate fit. The

RMSEA of .0278 and CFI of .9905, however, suggest that this confirmatory factor model fits

the date well. This result is similar to what LISREL produced in 5.2.

Unstandardized estimates under the Manifest Variable Equations with Estimates heading

have large t statistics and appear to be statistically discernable from zero. Standardized

estimates display under Manifest Variable Equations with Standardize d Estimates Factor

loadings for GOVTRESP are .2041 for economic values and .1865 for moral values. Other

factor loadings range from .5131 for EUTHANAS to .6984 for ABORTION. Squared multiple

correlations range from .0763 for GOVTRESP to .2633 for EUTHANAS and .4878 for

ABORTION. The covariance and correlation between two latent variables are -.0006 and

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

84

-.0016, respectively. Unlike LISREL and Mplus, SAS/STAT CALIS reports negative

covariance and correlation although their magnitudes are virtually zero.

5.6 Summary

Table 7 summarizes results from the LISREL, Mplus, and SAS/STAT CALIS. LISREL and

Mplus with listwise deletion produces equal parameter estimates but Mplus reports smaller

standard errors. Mplus also supports pairwise deletion analyses using listwise deletion plus the

Mplus results using pairwise deletion that produces slightly different parameter estimates and

standard errors. The pairwise column should be considered more accurate because it is able to

incorporate the most information from the raw data; notice that pairwise deletion uses 1,200

observations without dropping 40 incomplete observations. SAS/STAT CALIS reports

goodness of fit measures similar to those LISREL produces, but unstandardized and

standardized estimates of two software packages are slightly different.

LISREL and SAS/STAT CALIS reports a χ

2

of about 13, which is different from 10 and 16 in

Mplus. LISREL and SAS/STAT CALIS have a marginal p-value of about .06 and thus do not

reject the null hypothesis of a good fit at the .05 significance level. In Mplus, CFA model with

listwise deletion rejects the null hypothesis at the .05 level, while CFA with pairwise deletion

does not. RMSEA and CFI, however, are relatively consistent regardless of models estimated

by three software packages, suggesting that all models fit the data well. Notice that Mplus uses

a different formula to calculate the degrees of freedom; LISREL and SAS/STAT CALIS report

7 degrees of freedom whereas Mplus returns 6.

Table 7: Two Factor Model with Ordinal Indicators

LISREL (listwise) Mplus (listwise) Mplus (pairwise) SAS CALIS

Unstd. Standard Unstd. Standard Unstd. Standard Unstd. Standard

F1 PRIVTOWN

1.00 .63

1.00 .628

(.083)

1.000 .595

(.059)

1.0000 .5655

F1 GOVTRESP

.38

(.079)

.24 .375

(.073)

.236

(.041)

.380

(.051)

.226

(.029)

.3609

(.0831)

.2041

F1 COMPETE

1.32

(.37)

.83 1.313

(.332)

.825

(.104)

1.352

(.258)

.805

(.077)

1.2040

(.3369)

.6809

F2 HOMOSEX

1.00 .74 1.00 .737

(.030)

1.000 .721

(.023)

1.0000 .6737

F2 ABORTION

1.06

(.078)

.78 1.059

(.071)

.780

(.030)

1.066

(.053)

.769

(.023)

1.0367

(.0811)

.6984

F2 EUTHANAS

.88

(.068)

.65 .878

(.058)

.647

(.035)

.930

(.045)

.670

(.027)

.7616

(.0633)

.5131

F2 GOVTRESP

.30

(.055)

.22 .304

(.055)

.224

(.040)

.272

(.044)

.196

(.031)

.2768

(.0546)

.1865

F1 F2

.01

a

(.02)

.01

b

.006

a

(.022)

.012

b

(.048)

.014

a

(.018)

.033

b

(.040)

-.0006

a

(.0185)

-.0016

b

Degrees of freedom

(N=1,160) 7 (N=1,160) 6 (N=1,200) 6 (N=1,160) 7

;

2

, RMSEA, CFI

13.93

.028

1.00

15.516

.037

.991

10.417

.025

.997

13.2625

.0278

.9905

Standard errors appear in parentheses

a

covariance;

b

correlation between F1 (economic values) and F2 (moral values)

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

85

6. Conclusion

Factor analysis is a widely used method for situations in which a small set of unobserved

(latent) variables is believed to underlie a larger set of observed (manifest) variables.

Exploratory factor analysis (EFA), available in most general statistics packages, is a technique

to identify structure in data and generating hypotheses without imposing any restrictions. EFA

differ from confirmatory factor analysis (CFA) in that CFA is much more theory driven (as

opposed to data driven) and is generally used to test explicit hypotheses.

CFA is the basis of the measurement model in full structural equation modeling (SEM) and can

be estimated using SEM software. Amos and LISREL are the most user-friendly, although

Mplus syntax is not at all difficult to learn. Amos and LISREL can read data files generated by

a variety of different software packages and can estimate models by simply drawing a path

diagram. SAS/STAT CALIS has the systematic grammar structure in the integrated

environment, but it has messy output and does not support advanced models with missing

values. LISREL, Mplus, and SAS/STAT CALIS can handle ordinal observed variables using

(diagonally) weighted least squares estimation methods. Amos cannot accurately estimate

models when the observed variables are categorical. All four software packages handle models

assuming the latent variable to be continuous, although Mplus can also estimate models in

which the latent variables are assumed to be categorical.

Mplus is recommended for its simplicity of syntax, informative output, and support for various

advanced models although it has poor data management capability and does not support a path

diagram. LISREL is generally suggested to take advantage of switching back and forth from a

path diagram to SIMPLIS and LISREL. By contrast, Amos Graphics and Program Editor work

independently and the output, although rich enough, is not easy to navigate.

Consult the documentation for the respective package for additional information on Amos,

LISREL, and Mplus. Additionally IU students, staff, and faculty may schedule an appointment

with a consultant at the UITS Stat/Math Center by calling 5-4724 or emailing

statmath@indiana.edu.

Acknowledgements

Amy Drayton and Takuya Noguchi offered helpful suggestions to make this document more

informative and readable, though they are not responsible for any errors that remain.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis

http://www.indiana.edu/~statmath

86

References

Akaike, H. (1987), “Factor analysis and AIC,” Psychometrika, 52, 317-332.

Bollen, K.A. (1980), “Issues in the comparative measurement of political democracy”

American Sociological Review, 45, 370-390.

Bollen, K.A. (1989), Structural Equation Models with Latent Variables, New York: Wiley &

Sons.

Bollen, K.A. and Long, J.S., eds. (1993), Testing Structural Equation Models, Newbury Park,

CA: Sage.

Brown, T. A. (2006), Confirmatory Factor Analysis for Applied Research. New York: Guilford

Press.

Carmines, E.G. and Zeller, R.A. (1979), Reliability and Validity Assessment, Beverly Hills,

CA: Sage.

Enders, C.K. (2001), “A primer on maximum likelihood algorithms for use with missing data.

Structural Equation Modeling, 8, 128-141.

European Values Study Group and World Values Survey Association. (2005), European and

World Values Surveys Integrated Data File, 1999-2002 Release I, ICPSR3975

[computer file], Ann Arbor, MI: Inter-university Consortium for Political and Social

Research [distributor], 2005-06-15.

Gabel, M.J. and Huber, J.D. (2000), “Putting parties in their place: Inferring party left- right

ideological positions from party manifestos data,” American Journal of Political

Science, 44, 94-103.

Hatcher, Larry. (1994), Step-by-Step Approach to Using the SAS System for Factor Analysis

and Structural Equation Modeling. Cary, NC: SAS Institute.

Hu, L. and Bentler, P.M. (1999), “Cutoff criteria for fit indexes in covariance structure

analysis: conventional criteria versus new alternatives,” Structural Equation Modeling,

6, 1-55.

Jöreskog, K.G. (1969), “A general approach to confirmatory factor analysis,” Psychometrika,

34, 183-202.

Jöreskog, K.G. (1990), “New developments in LISREL: Analysis of ordinal variables using

polychoric correlations and weighted least squares,” Quality and Quantity 24, 387-404.

McDonald, R.P. (1978), “A simple comprehensive model for the analysis of covariance

structures,” British Journal of Mathematical and Statistical Psychology, 37, 234-251.

Muthén, B.O. (1984), “A general structural equation model with dichotomous, ordered,

categorical, and continuous latent variable indicators,” Pychometrika, 49, 115132.

Muthén, L.K. and Muthén, B.O. (2006), Mplus, Los Angeles: Muthén and Muthén.

Schwarz, G. (1978), “Estimating the dimension of a model,” The Annals of Statistics, 6,

461-464.

Spearman, C. (1904). “General intelligence, objectively determined and measured,” American

Journal of Psychology, 15, 201-293.

Steiger, J. H. and Lind, J. (1980), “Statistically-based tests for the number of common factors,”

Paper presented at the Annual Spring Meeting of the Psychometric Society, Iowa City.

Jöreskog, K.G. and Sörbom, D. (2004), LISREL 8.7. Scientific Software International, Inc.

Arbuckle, J.L. (2007), Amos16.0 User’s Guide, Chicago, IL: SPSS Inc.

© 2006-2009 The Trustees of Indiana University

Confirmatory Factor Analysis

2

This document summarizes confirmatory factor analysis and illustrates how to estimate individual models using Amos 16.0, LISREL 8.8, Mplus 5.1, and SAS/STAT 9.1. ** 1. 2. 3. 4. 5. 6. Introduction Exploratory Factor Analysis Confirmatory Factor Analysis Confirmatory Factor Analysis with Missing Data Confirmatory Factor Analysis with Categorical Data Conclusion

1. Introduction

Factor analysis is a statistical method used to find a small set of unobserved variables (also called latent variables, or factors) which can account for the covariance among a larger set of observed variables (also called manifest variables). A factor is an unobservable variable that is assumed to influence observed variables. Scores on multiple tests may be indicators of intelligence (Spearman, 1904); political liberties and popular sovereignty may measure the quality of a country’s democracy (Bollen, 1980); or issue emphases in election manifestos may signify a political party’s underlying ideology (Gabel & Huber, 2000). Factor analysis is also used to assess the reliability and validity of measurement scales (Carmines & Zeller, 1979). Principle component analysis also reduces the number of variables, but it differs from principle factor analysis (Brown, 2006: 22). A factor (unobserved latent variable) is assumed to exert causal influence on observed variables, while the underlying causal relationship is reversed in principle component analysis; observed variables are linear combinations of latent variables in factor analysis, while principle components are (weighted) linear combinations of observed variables (Hatcher, 1994: 9-10, 69). Principle components account for total variance, while factors account for the common variance (as opposed to unique variance) of a total variance (Brown, 2006: 22; Hatcher, 1994: 69). 1.1. Exploratory versus Confirmatory Factor Analysis Investigators wish to explore patterns in the data or to test explicitly stated hypotheses. Exploratory factor analysis (EFA), corresponding to the former task, imposes no substantive constraints on the data; there is no restrictions on the pattern of relationships between observed and latent variables. EFA is data driven (Brown 2006: 14). Each common factor is assumed to affect every observed variable and that the common factors are either all correlated or uncorrelated. Once model is estimated, factor scores, proxies of latent variables, are calculated and used for follow-up analysis.1 General purpose statistical software packages such as SPSS, SAS, and Stata can perform EFA.

Jeremy alone wrote the first edition in 2006 and then Hun revised introduction and added exploratory factor analysis and SAS/STAT CALIS sections to later editions. 1 A factor score is a linear composite of the optimally-weighted observed variables, while a factor-based score is merely a linear composite (e.g., mean or sum) of the variables that demonstrated meaningful factor loadings (Hatcher, 1994: 31).

**

http://www.indiana.edu/~statmath

© 2006-2009 The Trustees of Indiana University

Confirmatory Factor Analysis

3

Confirmatory factor analysis (CFA), on the other hand, is theory- or hypothesis driven. With CFA it is possible to place substantively meaningful constraints on the factor model. Researchers can specify the number of factors or set the effect of one latent variable on observed variables to particular values. CFA allows researchers to test hypotheses about a particular factor structure (e.g., factor loading between the first factor and first observed variable is zero). Unlike EFA, CFA produces many goodness-of-fit measures to evaluate the model but do not calculate factor scores. CFA requires special purpose software packages such as Mplus, LISREL, Amos, EQS, and SAS/STAT CALIS. Table 1. Explanatory and Confirmatory Factor Analysis EFA (Data-driven)

Constraint Unstandardized solution Standardized solution Factor rotation Factor scores Hypothesis test Goodness-of-fit Software package N/A N/A Yes Yes Yes N/A N/A General purpose software

**CFA (Theory-driven)
**

Yes Yes Yes N/A N/A Yes Yes Mplus, LISREL, Amos, EQS, SAS CALIS

In fact, CFA is a special case of the structural equation model (SEM), also known as the covariance structure (McDonald, 1978) or the linear structural relationship (LISREL) model (Jöreskog & Sörbom, 2004). SEM consists of two components: a measurement model linking a set of observed variables to a usually smaller set of latent variables and a structural model linking the latent variables through a series of recursive and non-recursive relationships. CFA corresponds to the measurement model of SEM. Table 1 summarizes differences and similarities of EFA and CFA. 1.2. Model Specification and Path Diagram It is common to display confirmatory factor models as path diagrams in which squares represent observed variables and circles represent the latent variables. Figure 1 has two latent variables ξ1 and ξ2 in circles that are manifested by six observed variables x1 through x6 in squares. Single-headed arrows are used to imply a direction of assumed causal influence, and double-headed arrows represent covariance between two latent variables. Latent variables “cause” the observed variables, as shown by the single-headed arrows pointing away from the circles and towards the manifest variables. Table 2. Notation for Confirmatory Factor Analysis Name Symbol Matrix Form Ksi ξ

x Lambda Phi Theta delta x

Description

Latent variable Observed variable Factor loading Factor variance and covariance Error variance and covariance

λ ϕ δ

X Λ Φ Θδ

http://www.indiana.edu/~statmath

© 2006-2009 The Trustees of Indiana University

Confirmatory Factor Analysis

4

The circles labeled ξ (ksi) represent latent variables or (common) factors. A factor can point to more than one observed variable; in Figure 1, ξ1 causes three observed variables x1 through x3 and ξ2 influences x3 through x6. The two ξi are expected to covary, as represented by ϕ21 on the two-headed arrow. Factor loadings are represented by λij; λ31 is, for example, the effect 2 (regression slope) of ξ1 on x3. The squared factor loading ij is referred to as a communality representing the proportion of variance in the ith observed variable that is explained by the jth latent variable (Brown, 2006: 61). The circles labeled δi (delta) represent unique factors because they affect only a single observed variable. The δi incorporate all the variance in each xi, such as measurement error, which is not captured by the common factors. Finally, error in the measurement of x3 is expected to correlate to some extent with measurement error of x6, as represented by δ63.2 Table 2 summarizes CFA notation discussed so far. Figure 1: Path Diagram of a Confirmatory Factor Model

When observed and latent variables are mean centered to have deviations from their means, the confirmatory factor model can be summarized by the equation X

in which X is the vector of observed variables, Λ (lambda) is the matrix of factor loadings connecting the ξi to the xi, ξ is the vector of common factors, and δ is the vector of unique factors. It is assumed that the error terms have a mean of zero, E(δ) = 0, and that the common and unique factors are uncorrelated, E(ξδ’)=0. Equation 1 can be rewritten for Figure 1 as: x1 = 1111 x4 = 4224 x2 = 2112 x5 = 5225 x3 = 3113 x6 = 6226

Here the similarities with regression analysis are evident. Each xi is a linear function of one or more common factors plus an error term (there is no intercept since the variables are mean centered). The primary difference between these factor equations and regression analysis is that

This may occur, for example, with panel data in which ξ1 and ξ2 represent the same concept measured at different points in time; if there is measurement error at t1 it is likely that there will be measurement error at t2.

2

http://www.indiana.edu/~statmath

for example. 3 http://www. Estimation When the x variables are measured as deviations from their means it is easy to show that the sample covariance matrix for x. The problem lies in the fact that the latent variables are unobserved and hence their scales are unknown. the equation 10 = 2x + 3y is not identified because it has two unknowns but only one piece of information (one equation). estimation proceeds in a manner distinct from the conventional approach of regressing each x on the ξi. Provision of more than one independent equation will make it overidentified. 1. an infinite number of values for x and y could make the equation true. the model is underidentified. represented by S. a model is identified if all of the unknown parameters can be rewritten in terms of the variances and covariances of the x variables. Information provided is variances and covariances of observed variables including σ11.edu/~statmath . σ31 … σ66.3. it therefore becomes necessary to set the metric of the latent variables in some manner. σ21. Without introducing some constraints any confirmatory factor model is not identified. In CFA. 1. If the number of the unknown parameters to be estimated is smaller than the number of pieces of information provided. including some necessary and sufficient conditions for identification. To identify the model. Consequently. another independent equation should be provided. The two most common constraints are to set either the variance of the latent variable or one of its factor loadings to one. That is. Degrees of freedom are 7=21 (knowns) -14 (unknowns).4. six λij. chapter 7). σ22.indiana. this CFA is overidentified. 11 21 31 41 51 61 22 32 42 52 62 33 43 44 53 54 55 63 64 65 66 The number of input information is 21=6(6+1)/2=p(p+1)/2. and δ63. To make it just-identified. can be decomposed as follows: Σ ΛΦΛ Θ A full discussion of the topic in the context of CFA is available in Bollen (1989. Identification One essential step in CFA is determining whether the specified model is identified. adding 3 = x + y ends up with x=-1 and y=4. the equation is not solvable.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 5 the ξi are unobserved in CFA.3 Unknown parameters of the CFA in Figure 1 are ϕ21. six δi. where p is the number of observed variables. For example.

and preliminary diagnostics of sample data show strong deviations from normality for several of the variables. Several different fitting functions exist for determining the closeness of the implied covariance matrix to the sample covariance matrix. MLE assumes multivariate normality among the observed variables. 1.5. Comparative fit index (CFI) evaluates “the fit of a user-specified solution in relation to a more restricted. Goodness of Fit A large class of omnibus tests exists for assessing how well the model matches the observed data. a small χ2 and failure to reject the null hypothesis is a sign of a good model fit. Due to these drawbacks of χ2 test many alternative fit statistics have been developed. a measure of fit introduced by Steiger and Lind (1980). It is sensitive to sample size. The null hypothesis is that the implied or predicted covariance matrix Σ is equivalent to the observed sample covariance matrix S. Σ=S. of which maximum likelihood is the most common. A large χ2 and rejection of the null hypothesis means that model estimates do not sufficiently reproduce sample covariance. 1978) can be also used to compare models with respect to model parsimony. and it becomes more and more difficult to retain the null hypothesis as the number of cases increases.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 6 where Φ (phi) represents the covariance matrix of the ξ factors and Θ (theta) represents the covariance matrix of the unique factors δ (Bollen.4 Another commonly reported statistic is the Root Mean Square Error of Approximation (RMSEA). though each has its own advantages and disadvantages. 1987) and Schwarz’s Bayesian Information Criterion (Schwartz. However. 1989: 236). χ2 is based on a very stringent hypothesis of Σ=S (Brown 2006: 81). Φ .indiana. This document will also describe a weighted least squares (WLS) approach suitable for situations in which the x variables are categorical. and Θ so that predicted x covariance matrix Σ (sigma) is as close to the sample covariance matrix S as possible. RMSEA “incorporates a penalty function for poor model parsimony” and thus becomes sensitive to the number of parameters estimated and relatively insensitive to sample size (Brown 2006: 83-84). Alternative estimators exist for cases of non-normal data but for the most part lie outside the limited scope of this document. 2005). The Akaike Information Criterion (Akaike. However. http://www. including Full Information Maximum Likelihood (FIML) for situations in which there are missing values in the raw data file. nested baseline model.05 or less would indicate a close fit of the model in relation to the degrees of freedom. 1969). The Amos User’s Guide suggests that “a value of the RMSEA of about 0. 2005: 496). CFI ranges from 4 Appendix C of the Amos User’s Guide provides summaries of many different fit measures (Arbuckle. the model does not fit the data well. This document includes examples using maximum likelihood estimation (MLE). leading to the rejection of good models or the retention of bad ones. For a thorough discussion of different tests see Bollen and Long’s (1993) edited volume.” although “this figure is based on subjective judgment” and “cannot be regarded as infallible” (Arbuckle. χ2 is a classic goodness-of-fit measure to determine overall model fit. the χ2 test is widely recognized to be problematic (Jöreskog. Estimation proceeds by finding ˆ ˆ ˆ the parameters Λ .” in which the “covariances among all input indicators are fixed to zero” or no relationship among variables is posited (Brown 2006: 84). The χ2 test may also be invalid when distributional assumptions are violated.edu/~statmath . By contrast.

and SAS CALIS need a program describing a model to be estimated. and SAS/STAT CALIS Amos LISREL Mplus Estimation Path diagram Data format supported Syntax (language) Output Platform supported Unstandardized estimates Standardized estimates and R2 Covariances of factors Correlations of factors Goodness-of-fit Residual (error) variances Modification indices Amos Graphics.e. while LISREL (PRELIS. 2006).e.1 (Muthén & Muthén. 2004). Yes w/o s. LISREL.95 or greater. UNIX Yes Yes w/o s. Yes w/o s. for example.e. while Amos supports both the point and click method and the program approach. is currently not supported. it is recommended to report not only χ2 but RMSEA and CFI/TLI. Many Yes No This document provides step-by-step examples for conducting a CFA with commonly used statistical software packages: Amos. Tucker-Lewis index (TLI) is another index for comparative fit that “includes a penalty function for adding freely estimated parameters” (Brown 2006: 85). the model may have a reasonably good fit.e. but Amos is not. C# Messy Windows Yes Yes w/o s. LISREL is able to go back and forth between a program and a path diagram. CFA and SEM can also be estimated using the CALIS procedure in SAS. Mplus and LISREL are generally recommended for confirmatory factor analysis.6. Program Editor Yes SPSS Visual Basic. but it can have a value outside of the range of 0 to 1 (p.e. 86). Comparison of Amos.06 or below and CFI and TLI are close to . Mplus. Mplus and SAS CALIS use simple syntax structure. LISREL. on the path diagram only Yes w/o s. another popular SEM program. 1. Mplus. Amos and LISREL produce a path diagram but Mplus and SAS CALIS do not. and LISREL) and Amos have a relatively abstruse grammar. Therefore. To sum. LISREL.0 (Arbuckle.e.edu/~statmath . All four programs are supported by the Stat/Math Center at Indiana University. LISREL. 2005). There is no single evaluation rule on which everyone agrees. SIMPLIS.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 7 0 for a poor fit to 1 for a good fit. TLI can be interpreted in a similar fashion as CFI.8 (Jöreskog & Sörbom. Yes w/o s. PRELIS. while EQS. The next section http://www. Software Issues This document considers estimating confirmatory factor models using Amos 7.e. Many Yes Yes only in Amos Graphics SIMPLIS.e. When RMSEA values are close to . LISREL 8. UNIX Yes Yes w/o s. and SAS/STAT CALIS.e. Yes w/o s. Mplus. Many Yes Yes Program No ASCII text Mplus Concise Windows Yes Yes Yes Yes Several Yes Yes SAS CALIS Program No Many formats SAS CALIS Messy Windows. SIMPLIS Normal Windows. and Mplus 5.indiana. Table 3. Hu and Bentler (1999) provide rules of thumb for deciding which statistics to report and choosing cut-off values for declaring significance. Mplus provides a variety of useful information in a concise manner. Yes w/o s. path diagram Yes Many formats LISREL.

http://www. Section 5 discusses the commonly encountered situation in which the observed variables are categorical rather than continuous. Section 3 begins with two-factor CFA with six observed indicators. Section 6 provides a brief summary.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 8 provides an example of EFA with six observed variables.indiana. Section 4 extends Section 3 to cover cases involving missing data.edu/~statmath .

generalized least squares (GLS). The survey queried a random sample of 1. and moral values.1 Sample Data In politics commentators often use the terms left and right to describe the ideological positions of politicians and voters.indiana. 4. Does a single dimension of values underlie Americans’ views on both economic and moral issues? Or are there in fact two distinct value dimensions that underlie citizen attitudes? This example uses data from the American sample of the European Values Survey (European Values Group and World Values Survey Association. weighted least squares (WLS). http://www. such as principal factor (PF).edu/~statmath . There are many methods to extract factors. 2. will be analyzed. and assisted suicide (EUTHANAS). 6. maximum likelihood (ML). assisted suicide.200 observations. but it is not always clear what exactly these terms mean. Exploratory Factor Analysis Before moving on to CFA. leaving a sample size of 1. political. The economic items asked respondents if they felt private ownership of industry should be increased (PRIVTOWN). and gay marriage it is often the political right that favors a stronger regulatory role for government.200 respondents about their economic. Yet on moral issues such as abortion. and whether competition brings out the best or worst in people (COMPETE). let us discuss sample data used in this document and briefly review exploratory factor analysis to contrast it with confirmatory factor analysis. For this section missing data is handled by listwise deletion (all cases with missing observations on any indicator are removed).sav located in the folder C:\temp\CFA. if the government should take more responsibility to see that all people are provided for (GOVTRESP). 2. legalized abortion (ABORTION). 2005) to determine whether a model with one or two common latent factors adequately describes attitudes on economic and moral issues. The data set has six variables and looks like the following: +---------------------------------------------------------------+ | privtown govtresp compete homosex abortion euthanas | |---------------------------------------------------------------| | 1 3 2 2 2 2 | | 2 1 2 1 1 1 | | 7 3 3 4 3 3 | | 4 3 1 6 7 1 | | 8 1 1 1 5 1 | | 8 3 4 3 4 4 | | 6 6 6 1 1 1 | 1. 5.2 An Explorative Factor Analysis EFA seeks a smaller number of latent variables to explain variance and covariance among manifest variables. 2. Listwise deletion resulted in dropping 40 of the original 1. In the United States the political left is generally associated with favoring greater government involvement in the economy while the right is understood to favor market autonomy. 7.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 9 2. The data is saved as the SPSS file values.160. all measured on 10point scales. Three questions summarizing economic attitudes and three questions summarizing moral attitudes. The moral items asked respondents how they felt about homosexuality (HOMOSEX). 3.

http://www.1966 Factor6 | -0.0145 | 0. Alternatively. Factor loadings of factor 1 and 2 on the perception on competition (COMPETE) are .07293 -0.0231 0. it has the largest unique variance of 93 percent.55561 0. In the following output. researchers may rotate them to “foster interpretability by maximizing factor loadings close to 1. while HOMOSEX (homosexuality).5494 | 0.0005=. ------------------------------------------------Variable | Factor1 Factor2 | Uniqueness ------------+--------------------+--------------privtown | -0. .13165 0.3077). 31 percent of variance in COMPETE is explained by factor 2 and almost zero percent by factor 1. Factor rotation is either orthogonal or oblique. Researchers may need to determine the number of factors extracted using eigenvalues calculated from input correlation matrix.0 and minimizing factor loadings close to 0” (Brown 2006: 31).edu/~statmath .9686 Factor2 | 0.4704 Factor4 | -0. and EUTHANAS (assisted suicide) by factor 1. For example.5547.02312 and .9309 compete | 0.1619 | 0.2070 0.0059 | 0.6616. researchers may use scree test and parallel analysis (Brown 2006: 26-30). Unique variance is the proportion of variance that is not explained by any factor.0203 0. 2006: 26. Factors are constrained to be uncorrelated in orthogonal rotation but not necessarily in oblique rotation. ABORTION. Hatcher. researchers may calculate factor scores for future analysis.0145 1. Varimax (orthogonal) and promax (oblique) rotations are commonly used.55472 are respectively the proportions of variance in COMPETE that is explained by factor 1 and 2. However.03691 -0. Therefore.6918 is calculated as 1(.65191 0.. The squared factor loadings. Finally.5163 1.1666 1.6977 govtresp | 0.6918 homosex | 0.5681 euthanas | 0. Two factor scores predicted here have zero mean and standard deviations of .63409 0.01782 0.24149 . -------------------------------------------------------------------------Factor | Eigenvalue Difference Proportion Cumulative -------------+-----------------------------------------------------------Factor1 | 1.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 10 etc.6572 -0.3077=. Keep in mind that any factor rotation does not alter factor loadings but change views of pattern matrix.6919 ------------------------------------------------- The output above suggests two latent variables underlying six manifest variables. of which PF and ML are most commonly used.0000 -------------------------------------------------------------------------- Once factors are extracted.5958 abortion | 0.6356 -0. of .0231 and .indiana. -0.5514 0. there appear to be two moral dimensions underlying citizen attitudes.9686 0.7871 and .20458 0. there are two positive eigenvalues only one of which is greater than 1.1966 1. neither factor 1 nor 2 can explain GOVTRESP (government responsibility) sufficiently.4849 Factor3 | -0. That is. count the number of eigenvalues greater than 1 and use it as the number of factors (Brown.1072 1. 1994: 22-23).3632 Factor5 | -0.0642 | 0. communalities.5547 | 0. PRIVTOWN (private ownership) and COMPETE are largely explained by factor 2.0005+. The numbers under first and second columns are factor loadings.11383 -0. the 69 percent of variance in COMPETE is not explained by two factors. As a rule of thumb.18970 0.

58. http://www. Confirmatory Factor Analysis This section begins with one factor model and then moves forward to the two factor model.08 (=1. respectively. COMPETE and PRIVTOWN have large factor loadings of .32). The numbers in parentheses are standard errors followed by test statistics.722) of variance in COMPETE. For example. R2 is a standardized factor loading squared that means the extent that a factor can explain the variance in a manifest variable.80 This program is published exclusively by Scientific Software International.5 An oval and a rectangle represent a latent variable and a manifest variable. HOMOSEX. The coefficients of values listed under the Measurement Equations heading are unstandardized factor loadings.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 11 3. For instance. 3. 5 See the LISREL section for details about fitting a confirmatory factor model and drawing a path diagram. R2 is listed at the end of each equation. Inc. The numbers on arrows from the latent variable to observed variables are standardized factor loadings (regression weights).edu/~statmath .31/.72 and . The following is the LISREL output of this confirmation factor model. DATE: 11/18/2008 TIME: 13:00 L I S R E L BY Karl G. and EUTHANAS have poor factor loadings.32. they appear to be the best indicators of values. its standard error is. the factor loading on COMPETE is 1.indiana. ABORTION. where a latent variable values is manifested by six observed variables. suggesting that they appear to indicate other factors. the latent variable values explains about 51 percent (=.31. Jöreskog & Dag Sörbom 8. and the test statistic is 4.1 One Factor Model The following path diagram with standardized solutions illustrates the one factor model.

85 homosex = 0.51 2.= 2.= 7.ssicentral.indiana.30) 4.01 0.72) 4. Phone: (800)247-6113.96 . Errorvar.11 0.00*Values. U.01 0. Website: www.psf' Sample Size = 1160 Latent Variables Values Relationships privtown = 1.11*Values.31*Values. Fax: (847)675-2140 Copyright by Scientific Software International.27 -0.094) (0.72 compete -------homosex -------abortion -------euthanas -------- privtown govtresp compete homosex abortion euthanas 5.21*Values.39 1.04 abortion = 0. Errorvar.= 8. R² = 0. Errorvar.33*Values.12*Values.073) (0. R² = 0.spj: One factor model Raw Data from file 'c:\Temp\cfa\values.44 24.81 3.34 (0.03 euthanas = 0.com The following lines were read from file c:\Temp\cfa\values.S. Errorvar. Suite 100 Lincolnwood.0020 (0.= 5.67 govtresp = 0.edu/~statmath .083) (0. IL 60712.A.82 . R² = 0.78 .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 12 7383 N.20 0.44) 7.026 (0.00*Values govtresp = Values compete = Values homosex = Values abortion = Values euthanas = Values Path Diagram End of Problem Sample Size = 1160 Second Two factor model Covariance Matrix privtown -------5.67 1.85 3.00 23. Errorvar.14 govtresp -------7. R² = 0.012 (0.18 24.71 0.086) (0. 1981-2006 Use of this program is subject to the terms specified in the Universal Copyright Convention.57 4.32) (0.51 (0.55.03 Second Two factor model Number of Iterations = 8 LISREL Estimates (Maximum Likelihood) Measurement Equations privtown = 1.00 8.64 compete = 1. Lincoln Avenue. Errorvar.05 6.= 10..01 .0030 (0.08 3. R² = 0. (847)675-0720.= 3.38 10.37) 1.35 . Inc.25) http://www.06 0.22 -0.44) 1.05 0. R² = 0.

33 Relative Fit Index (RFI) = -0.73 Saturated CAIC = 169.44) 3.000).000 Seconds A various goodness-of-fit statistics are listed under Goodness of Fit Statistics.72) ECVI for Saturated Model = 0.32 Non-Normed Fit Index (NNFI) = -0.indiana.55 Population Discrepancy Function Value (F0) = 0.0) Normal Theory Weighted Least Squares Chi-Square = 718.0 0.81 23.036 ECVI for Independence Model = 0.89 Variances of Independent Variables Values -------1.60 Parsimony Goodness of Fit Index (PGFI) = 0.3 0.70 (0.84 Goodness of Fit Statistics Degrees of Freedom = 9 Minimum Fit Function Chi-Square = 641.45 Standardized RMR = 0.0) Estimated Non-centrality Parameter (NCP) = 709.54 .59 Model CAIC = 814.00 homosex privtown 8.13 Parsimony Normed Fit Index (PNFI) = 0. The http://www.96 abortion homosex 285.2 42. 0.26 Independence AIC = 961.57 .62 euthanas homosex 162.26 Model AIC = 742.99 euthanas abortion 201.05 level (p<. 0.83 Chi-Square for Independence Model with 15 Degrees of Freedom = 949.36 The Modification Indices Suggest to Add an Error Covariance Between and Decrease in Chi-Square New Estimate compete privtown 204.05) = 0.62 homosex govtresp 28.6 3.00 Expected Cross-Validation Index (ECVI) = 0.19 Comparative Fit Index (CFI) = 0.2 -0.35 abortion govtresp 17. χ2 is 718.67) Minimum Fit Function Value = 0.04 Time used: 0.28) P-Value for Test of Close Fit (RMSEA < 0.24 .00 Independence CAIC = 997. 800. 0.04 (P = 0.17 Goodness of Fit Index (GFI) = 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 13 2.26 90 Percent Confidence Interval for RMSEA = (0.06.18 Normed Fit Index (NFI) = 0. which is so large that the null hypothesis of a good fit is rejected at the .69) Root Mean Square Error of Approximation (RMSEA) = 0.61 90 Percent Confidence Interval for F0 = (0.06 90 Percent Confidence Interval for NCP = (624.06 Saturated AIC = 42.17 Root Mean Square Residual (RMR) = 1.64 90 Percent Confidence Interval for ECVI = (0.80 euthanas govtresp 10.9 2.9 4.84 .13 Critical N (CN) = 40.83 Adjusted Goodness of Fit Index (AGFI) = 0.1 1.06 (P = 0.32 Incremental Fit Index (IFI) = 0.edu/~statmath .

12.9. however. if you add covariance between HOMOSEX and ABORTION χ2 will decrease by 285.002).003).000). while PRIVTOWN and COMPETE are grouped together (economic values). Modification indices at the bottom suggest that HOMOSEX. It is likely that a two factor model is more appropriate to describe the economic and moral values of Americans It is not clear. there are 21 sample variance and covariance elements and 12 unknown parameters including six λij and six δi. http://www.26 is also large enough to reject the null hypothesis (p<. Root Mean Square Error of Approximation (RMSEA) 0. let us first connect to economic values though. whether GOVTRESP falls into economic values or moral values. and EUTHANAS are closely related (moral values). and EUTHANAS (.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 14 degrees of freedom is 9 =21 . ABORTION. Small R2 of HOMOSEL (.indiana.edu/~statmath . Therefore. For instance. Comparative Fit Index (CFI) 0.32 is small. this one factor model shows a poor fit and needs to be modified somehow. two variables appear to manifest the same latent variable.012) support this conclusion since the one factor explains practically no variance in these observed variables. ABORTION (.

80. Change Files of Type to SPSS Data File(*. and SPSS.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 15 3. http://www. Click Open.psf (PRELIS system) file. The Open dialog box opens.sav saved in the C:\temp\CFA folder. go to File → Import Data. PRELIS. To open the SPSS file values. including SAS. Name the file values and click Save to store it in the working directory. You will then be prompted to save the data as a .sav.80 → LISREL 8. A spreadsheet with the raw data will display. and click on values.indiana. you need to import the data file into LISREL and save it as a .psf file. First. the pre-processor to LISREL. can read data files from a number of statistical programs.edu/~statmath .2 Two Factor Model using LISREL This section demonstrates how to estimate a two factor confirmatory factor model using LISREL. Note that the previous subsection revealed that the single common factor model was a poor fit to the data. LISREL can be launched from any computer running Windows in the UITS Student Technology Centers by going to Start → All Programs → Departmentally Sponsored → Statistics-Math → LISREL 8. Stata.sav). navigate to the correct folder.

Name the file values and click Save (the . Enter ECONOMIC in the box that opens. Go to Setup → Title and Comments to open the Title and Comments dialog box. Repeat to enter the name of the second common factor MORALS. which is used to label different groups when comparing models for multiple independent samples. The Labels dialog box then opens. An empty window opens where you will eventually draw the diagram. The names of the observed variables are now listed in the Labels box. which is used to identify the latent and observed variables to be analyzed.pth extension will differentiate this file from the other files named values in the working directory).indiana. Because we are interested only in the single sample of American respondents we can skip this box by clicking Next. Go to File → New and choose Path Diagram. The Group Names box opens.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 16 LISREL allows you to construct a path diagram of the model and then generate a PRELIS or LISREL program based on the path diagram. Verify that the Read from file radio button is chosen and pick PRELIS System File from the drop-down menu.edu/~statmath . To choose variable names click on Add/Read Variables. You will be immediately prompted to save the path diagram. http://www. Click OK. Note that the default variable CONST appears on the list of observed variables. This opens a new dialog box used to locate the PRELIS system file. Enter First Two Factor Model in the Title field and click Next. Currently no variables have been selected. Click OK. The next step is to name the variables that will be in the model. Then click Browse to choose the PRELIS system file created earlier. To add the names of the latent variables click Add Latent Variables.

Click OK.indiana. http://www. Because this system file already contains information about the sample size it is not necessary to make further changes. and a final dialog box opens. Raw data from a PRELIS system file will be analyzed. the data can be viewed and edited by clicking on the Edit button.edu/~statmath . If desired.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 17 Click Next.

Do the same for the path connecting MORALS to HOMOSEX to set the metric for the second common factor.00 appears and change the loading to 1. Also draw arrows from MORALS to HOMOSEX. the arrows should point towards the six indicators. however.indiana.00. Finally.edu/~statmath . Drag all of the observed variables to the drawing pad along with the latent variables ECONOMIC and MORALS. draw a two-headed arrow connecting each latent variable. and COMPETE. Next click on the single-headed arrow on the tool bar and connect the ECONOMIC factor to PRIVTOWN. constrain the regression weight of the PRIVTOWN variable to one. unless you then right-click on the loading and choose Fix. LISREL includes these by default and automatically sets their scales by constraining the loadings to one. GOVTRESP. http://www. Because the usual assumption is that the latent variables “cause” the observed variables. it is not necessary to draw the unique factors representing measurement error for each of the observed variables. and EUTHANAS. The names of the observed and latent variables appear on the left side of the screen. ABORTION.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 18 It is now possible to begin drawing the path diagram. To set the scale of ECONOMIC. In LISREL. Double-click on the line at the point where 0. LISREL will not recognize this constraint.

As its name suggests. This opens an editor displaying the SIMPLIS commands needed to estimate the model. SIMPLIS is more straightforward and easy to read than LISREL syntax. in some cases. and other additional information. For this model the output file is the following: DATE: 11/18/2008 TIME: 1:33 http://www.out) and a path diagram is presented. goodness-of-fit statistics.edu/~statmath . to begin estimation. warnings that a model may not be identified. It is always a good idea to inspect the output file for any error messages and. A SIMPLIS syntax file can be built from the path diagram by choosing Setup → Build SIMPLIS syntax. modification indices. Each time the Run LISREL button is Click the Run LISREL button clicked. a text output file is written to the working directory (extension . standardized estimates do not appear on the text output.indiana.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 19 The final step before estimation is to build from the path diagram the corresponding syntax LISREL uses for estimation. The text output contains unstandardized estimates. test statistics. There are actually two languages that LISREL understands: LISREL syntax and SIMPLIS syntax.

01 0.= 2.37 .05 0.ssicentral. Phone: (800)247-6113.edu/~statmath .05 6.00 8. R² = 0.com The following lines were read from file c:\temp\cfa\values. IL 60712.35*Economic.57) 5.01 0.30) 3..= 3.spj First Two factor model Raw Data from file 'c:\temp\cfa\values.85 3.65 .80 Karl G.14 govtresp -------7.70 compete = 1.11 0.51 2. Errorvar.00*Morals abortion = Morals euthanas = Morals Path Diagram End of Problem Sample Size = 1160 First Two factor model Covariance Matrix privtown -------5. Lincoln Avenue.72 compete -------homosex -------abortion -------euthanas -------- privtown govtresp compete homosex abortion euthanas 5.54 (0.71 0.33 (0. 7383 N.084) (0. (847)675-0720.22 -0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 20 L I S R E L BY 8. 1981-2006 Use of this program is subject to the terms specified in the Universal Copyright Convention. Inc. Fax: (847)675-2140 Copyright by Scientific Software International.67 1. Website: www.A.= 7.81 3. Jöreskog & Dag Sörbom This program is published exclusively by Scientific Software International.022 (0.04 . Errorvar.27 -0.00*Economic govtresp = Economic compete = Economic homosex = 1. Suite 100 Lincolnwood.psf' Sample Size = 1160 Latent Variables Economic Morals Relationships privtown = 1.38 10. Errorvar. Inc.57 4. R² = 0.03 First Two factor model Number of Iterations = 5 LISREL Estimates (Maximum Likelihood) Measurement Equations privtown = 1.indiana.20 0.30*Economic.63 23. U. R² = 0.06 0.00*Economic.45) (1.02) http://www.88 govtresp = 0.39 1.S.

57) Minimum Fit Function Value = 0.01 2. 57.53 Saturated AIC = 42.92 Critical N (CN) = 553.86 .91 Morals 0.13 Covariance Matrix of Independent Variables Economic -------Economic 1.076) (0.60 homosex = 1. 0. R² = 0.41 Standardized RMR = 0.53 (P = 0.96 Non-Normed Fit Index (NNFI) = 0.99 Root Mean Square Residual (RMR) = 0.00*Morals.01 .00) Estimated Non-centrality Parameter (NCP) = 33.060 90 Percent Confidence Interval for RMSEA = (0.00) Normal Theory Weighted Least Squares Chi-Square = 41. 0.96 Relative Fit Index (RFI) = 0.= 5.079) P-Value for Test of Close Fit (RMSEA < 0.39 10.12 .45 (0.046) (0.93 Parsimony Normed Fit Index (PNFI) = 0.59 Model CAIC = 146.97 Parsimony Goodness of Fit Index (PGFI) = 0.26 Model AIC = 67. Errorvar.036 ECVI for Independence Model = 0.38 http://www.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 21 3.64*Morals.18 Normed Fit Index (NFI) = 0.96 .015 .052 Goodness of Fit Index (GFI) = 0.32 (0. R² = 0.63 euthanas = 0.26 Independence AIC = 961.51 Comparative Fit Index (CFI) = 0.050) Root Mean Square Error of Approximation (RMSEA) = 0.00 Independence CAIC = 997. Errorvar.11 (P = 0.99 Adjusted Goodness of Fit Index (AGFI) = 0.indiana.02*Morals.10 (0.40) 14.96 Incremental Fit Index (IFI) = 0.036 Population Discrepancy Function Value (F0) = 0.49 abortion = 1. R² = 0.53 90 Percent Confidence Interval for NCP = (17.16 Expected Cross-Validation Index (ECVI) = 0. Errorvar.55 (0.73 19.058 90 Percent Confidence Interval for ECVI = (0.043 .83 Chi-Square for Independence Model with 15 Degrees of Freedom = 949.58) 2.26 Saturated CAIC = 169.78 Morals -------- 4. 0.72 (0.49) 9.edu/~statmath .044 .079) ECVI for Saturated Model = 0.67 Goodness of Fit Statistics Degrees of Freedom = 8 Minimum Fit Function Chi-Square = 42.37) 13.13) 0.22) 13.029 90 Percent Confidence Interval for F0 = (0.= 4.05) = 0.= 3.68 (0.

CFI of . It is possible to ascertain the statistical significance of the estimates by comparing the unstandardized loadings displayed in the equations under the Measurement Equations heading in the output file with their standard errors displayed in parentheses. We report the Minimum Fit Function χ2 in order to be consistent with the output from other software packages.4 0. The text output presents unstandardized estimates and their standard errors. The Root Mean Square Error of Approximation (RMSEA) is . When the unstandardized loadings are at least http://www.6 86. and ϕ21. there are 13 unknown parameters including six λij.edu/~statmath .63 homosex govtresp 12. six δi . the RMSEA is too high to indicate a good fit.indiana.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 22 The Modification Indices Suggest to Add the Path to from Decrease in Chi-Square New Estimate govtresp Morals 31. which is large enough to reject the null that the model is a good fit to the data.060.05.6 0. The degrees of freedom is 8 = 21 -13.016 Seconds The χ2 statistic for model fit is 42.24 The Modification Indices Suggest to Add an Error Covariance Between and Decrease in Chi-Square New Estimate compete privtown 31.76 Time used: 0.11. Using a cut-off rule of .96 indicates a moderate fit.

Only GOVTRESP. does not fit in well with the model.33 (=.15 for GOVTRESP. and . and .05 level. is accounted for. The unstandardized estimates also appear in the path diagram by default (see left diagram below). the standardized loadings are . Considering the indicators of MORALS. These indices make suggestions about loosening certain model parameters in order to improve the overall model fit. for ABORTION it is . both in the path diagram (by choosing Modification Indices from the Estimation menu) and in the output.582). http://www.45. The R2 for PRIVTOWN is . while variances of two factors are set 1 in the following diagram for standardized solutions.indiana.732). for COMPETE it is . As long as any decisions made on the basis of modification indices are theoretically meaningful and do not result in an unidentified model they can be helpful in improving model specification.73 for COMPETE. and for EUTHANAS it is . with its R2 of . The standardized loadings represent the correlation between each observed variable and the corresponding factor. In this case each of the unconstrained estimates is significant. and between PRIVTOWN and COMPETE.74 for ABORTION.54 (=. To view the standardized estimates choose Standardized Solution from the Estimates drop-down menu (see above screenshot). they are .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 23 twice the size of the standard errors the estimates are significant at the . Considering first the indicators of ECONOMIC. It may be the case that this survey question taps some kind of value dimension distinct from the economic dimension measured by the PRIVTOWN and COMPETE variables. A good deal of the variance in each observed variable. add error covariances between HOMOSEX and GOVTRESP. with the exception of GOVTRESP.022. .65.58 for PRIVOWN.edu/~statmath . . Note that factor loadings of PRIVTOWN and HOMOSEX are set 1 in the above diagram for unstandardized solutions. for HOMOSEX it is .32. LISREL reports modification indices.67 for HOMOSEX.56 for EUTHANAS. Three suggestions are given in the output: add a path from GOVTRESP to MORALS.

60.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 24 GOVTRESP appears to have something in common with the morality dimension.00*Economic. The final model therefore adds a path from MORALS to GOVTRESP.60 says that adding an arrow from MORALS to GOVTRESP will reduce χ2 by 31.56 23. R² = 0.29) 3.33 http://www. R² = 0.24*Morals.043) (0.62) 5.060 (0. ignoring less informative parts. Measurement Equations privtown = 1.00*Morals abortion = Morals euthanas = Morals Path Diagram End of Problem The following output is selective.082) (0.36 (0. The modification index 31.46 5. either by sharing measurement error with HOMOSEX and COMPETE or as a direct indicator of the latent morality dimension. Errorvar.28*Economic + 0.24 .77 . it is possible that the item is actually tapping a different values dimension. This modification results in the following standardized solution: Second Two factor model Raw Data from file 'c:\Temp\cfa\values.psf' Sample Size = 1160 Latent Variables Economic Morals Relationships privtown = 1.= 3.23 govtresp = 0.indiana.00*Economic govtresp = Economic Morals compete = Economic homosex = 1. Errorvar.= 6. Because the standardized loading of GOVTRESP on ECONOMIC was so low.edu/~statmath .

© 2006-2009 The Trustees of Indiana University

Confirmatory Factor Analysis 25

compete = 1.26*Economic, Errorvar.= 2.86 , R² = 0.50 (0.42) (0.96) 2.98 2.97 homosex = 1.00*Morals, Errorvar.= 5.72 , R² = 0.46 (0.40) 14.33 abortion = 0.99*Morals, Errorvar.= 4.08 , R² = 0.54 (0.072) (0.36) 13.83 11.46 euthanas = 0.63*Morals, Errorvar.= 4.13 , R² = 0.32 (0.045) (0.21) 13.80 19.28

Covariance Matrix of Independent Variables Economic -------Economic 1.81 (0.63) 2.89 Morals 0.04 (0.13) 0.32 Morals --------

4.86 (0.49) 9.93

Goodness of Fit Statistics Degrees of Freedom = 7 Minimum Fit Function Chi-Square = 9.88 (P = 0.20) Normal Theory Weighted Least Squares Chi-Square = 9.89 (P = 0.19) Estimated Non-centrality Parameter (NCP) = 2.89 90 Percent Confidence Interval for NCP = (0.0 ; 15.41) Minimum Fit Function Value = 0.0085 Population Discrepancy Function Value (F0) = 0.0025 90 Percent Confidence Interval for F0 = (0.0 ; 0.013) Root Mean Square Error of Approximation (RMSEA) = 0.019 90 Percent Confidence Interval for RMSEA = (0.0 ; 0.044) P-Value for Test of Close Fit (RMSEA < 0.05) = 0.98 Expected Cross-Validation Index (ECVI) = 0.033 90 Percent Confidence Interval for ECVI = (0.030 ; 0.043) ECVI for Saturated Model = 0.036 ECVI for Independence Model = 0.83 Chi-Square for Independence Model with 15 Degrees of Freedom = 949.26 Independence AIC = 961.26 Model AIC = 37.89 Saturated AIC = 42.00 Independence CAIC = 997.59 Model CAIC = 122.68 Saturated CAIC = 169.18 Normed Fit Index (NFI) = 0.99 Non-Normed Fit Index (NNFI) = 0.99 Parsimony Normed Fit Index (PNFI) = 0.46 Comparative Fit Index (CFI) = 1.00 Incremental Fit Index (IFI) = 1.00 Relative Fit Index (RFI) = 0.98 Critical N (CN) = 2167.38 Root Mean Square Residual (RMR) = 0.12 Standardized RMR = 0.018 Goodness of Fit Index (GFI) = 1.00 Adjusted Goodness of Fit Index (AGFI) = 0.99

http://www.indiana.edu/~statmath

© 2006-2009 The Trustees of Indiana University

Confirmatory Factor Analysis 26

Parsimony Goodness of Fit Index (PGFI) = 0.33 Time used: 0.000 Seconds

This model fits the data well. The χ2 measure of model fit is 9.88, which is too small to reject the null hypothesis of a good fit (p<.20). Adding λ22 reduces degrees of freedom by 1 (df=7). The RMSEA has declined to .019, which is small enough to indicate a good fit, while CFI is almost (as oppose to numerically) 1. The unconstrained loadings are all statistically significant at the .05 level, having estimates that are more than twice the size of their standard errors (see the output above and left diagram below). GOVTRESP continues to have a low correlation with the ECONOMIC factor (.14) and has a similarly low correlation with MORALS (.20) (see right diagram below). However, the remaining standardized loadings range from .56 (EUTHANAS) to .73 (ABORTION). In between are PRIVTOWN (.60), HOMOSEX (.68), and COMPETE (.71).

Despite receiving a path from both common factors GOVTRESP continues to have by far the smallest R2 (.060). The remaining statistics are moderately well accounted for by the corresponding factors. The R2 values are, in order of increasing magnitude, .32 for EUTHANAS, .36 for PRIVTOWN, .46 for HOMOSEX, .50 for COMPETE, and .54 for ABORTION. Finally, the correlation between ECONOMIC and MORALS is a negligible -.01 (right path diagram above) and their covariance is .04 (left diagram above). The conclusion from this analysis is that two nearly orthogonal dimensions underlie the economic and moral values of American citizens. Additionally it is unclear whether the GOVTRESP item is tapping either dimension. Future surveys should incorporate more reliable measures of economic values.

http://www.indiana.edu/~statmath

© 2006-2009 The Trustees of Indiana University

Confirmatory Factor Analysis 27

3.3 Two Factor Model using Mplus

This section demonstrates how to estimate a confirmatory factor model using Mplus 5.2. Because the subsection covering Amos revealed that the single common factor model was a poor fit to the data, this section will also begin with the two factor model. To launch Mplus from any Windows machine in the UITS Student Technology Centers go to Start → All Programs → Departmentally Sponsored → Statistics-Math → Mplus5.2 → Mplus Editor. This will open Mplus and display the program’s built-in syntax editor. Unlike Amos and LISREL, Mplus does not allow you to simply draw a path diagram and estimate the model; you must write the syntax yourself. The Language Generator under the Mplus menu, however, can make this task a little easier. Mplus reads only ASCII text files with free or fixed formatted; Mplus cannot directly read a SPSS data set. Nonetheless, Mplus is an extremely powerful program for estimating a much wider range of models than is possible with Amos and LISREL, and this example will only scratch the surface of what Mplus can do. The first model to be examined consists of two common factors and the six observed indicators of economic and moral values. The Mplus syntax for estimating this model is the following:

TITLE: DATA: VARIABLE: MODEL: OUTPUT: First Two Factor Model; FILE IS c:\temp\cfa\values.dat; NAMES ARE privtown govtresp compete homosex abortion euthanas; economic BY privtown govtresp compete; morals BY homosex abortion euthanas; STANDARDIZED; MODINDICES;

Save as an input file under the name values1.inp in the same folder as the values.dat file. The TITLE statement provides a label for the particular analysis that will be run. The DATA statement specifies where the data file is located at. Absolute path names are only necessary if the syntax file is located in a directory different from where the data is saved. The VARIABLE statement provides names for the six observed variables in the raw data file in the order in which they appear. The MODEL statement specifies the particular model to be estimated. In this case ECONOMIC is assumed to cause the three observed variables PRIVTOWN, GOVTRESP, and COMPETE; and MORALS is assumed to cause HOMOSEX, ABORTION, and EUTHANAS. The OUTPUT statement requests that standardized parameter estimates and modification indices be included in the output file. There are a few things to keep in mind when creating Mplus syntax. First, all commands end with a semicolon; omitting the semicolon will lead to error messages. Second, Mplus cannot read more than 80 characters in a line. One way to limit this problem is to use very short names for variables, such as x1, x2, y1, y2 (longer names are used here to be consistent with the Amos and LISREL examples later). Commands can take up more than one line, as the semicolon marks the command end. Finally, Mplus is not case sensitive; capital and lowercase letters can be used interchangeably.

http://www.indiana.edu/~statmath

© 2006-2009 The Trustees of Indiana University

Confirmatory Factor Analysis 28

It is also important to know that the default behavior for setting the scale of the common latent variable is to constrain the loading for the first variable (in this case PRIVTOWN and HOMOSEX) to one. This option can be overridden but will not be altered here to keep the example consistent with the Amos and LISREL examples above. After entering the syntax and saving it as an Mplus input (.inp) file, estimate the model by . This produces a text output (.out) file stored in the working clicking the Run button directory with the results. For this model the output file looks like the following:

Mplus VERSION 5.1 MUTHEN & MUTHEN 11/17/2008 9:58 PM INPUT INSTRUCTIONS TITLE: First Two Factor Model; DATA: FILE IS C:\Temp\CFA\values.dat; VARIABLE: NAMES ARE privtown govtresp compete homosex abortion euthanas; MODEL: economic BY privtown govtresp compete; morals BY homosex abortion euthanas; OUTPUT: STANDARDIZED; MODINDICES;

INPUT READING TERMINATED NORMALLY Two Factor Model 1; SUMMARY OF ANALYSIS Number of groups Number of observations Number of dependent variables Number of independent variables Number of continuous latent variables Observed dependent variables Continuous PRIVTOWN 1 1160 6 0 2

GOVTRESP

COMPETE

HOMOSEX

ABORTION

EUTHANAS

Continuous latent variables ECONOMIC MORALS Estimator Information matrix Maximum number of iterations Convergence criterion Maximum number of steepest descent iterations Input data file(s) C:\Temp\CFA\values.dat Input data format FREE ML OBSERVED 1000 0.500D-04 20

THE MODEL ESTIMATION TERMINATED NORMALLY TESTS OF MODEL FIT Chi-Square Test of Model Fit Value 42.147

http://www.indiana.edu/~statmath

048 50.003 0.090 23.876 54.I.433 999.645 0.076 0.070 0.000 0.072 53.000 0.124 0.845 36.637 0.000 0.018 0. Two-Tailed Est.775 49.147 SRMR (Standardized Root Mean Square Residual) Value 0.034 -16195.046 999.000 0.715 0.677 4.638 0.000 0.066 0.366 GOVTRESP 7.000 13.000 3.553 0.312 3.785 -16217.433 3.0000 Chi-Square Test of Model Fit for the Baseline Model Value Degrees of Freedom P-Value CFI/TLI CFI TLI Loglikelihood H0 Value H1 Value Information Criteria Number of Free Parameters Akaike (AIC) Bayesian (BIC) Sample-Size Adjusted BIC (n* = (n + 2) / 24) 19 32472.0000 RMSEA (Root Mean Square Error Of Approximation) Estimate 90 Percent C.083 0.432 13.557 0.553 4.000 0.590 0.961 0.000 0.350 0.352 2.046 MODEL RESULTS Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS MORALS WITH ECONOMIC Intercepts PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Variances ECONOMIC MORALS S.725 999.000 1.000 0.927 891. P-Value 1.043 0.000 0.440 4.007 http://www.000 1.961 0.990 15 0.781 4.675 3.692 0.E.000 0.486 3.indiana.304 1.edu/~statmath .079 0.120 999. Probability RMSEA <= .011 9.000 Residual Variances PRIVTOWN 3./S.784 0.002 1.297 0.E.000 0.068 32568.000 0.095 49.000 0.685 2.098 0.079 0.05 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 29 8 Degrees of Freedom P-Value 0.692 0.061 0.034 COMPETE 2.087 0.983 6.000 0.000 0.095 0.135 32507.

037 0.000 0.000 0.042 0.448 EUTHANAS 0.000 Residual Variances PRIVTOWN 0.668 0.743 0.000 0.157 86.000 0.000 999.000 0.577 0.471 1.119 0.171 34.014 10.E.782 0.028 23. Est.440 1.317 QUALITY OF NUMERICAL RESULTS Condition Number for the Information Matrix 0.446 0.000 999.366 34.957 4.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 30 0.172 0.000 0.637 29.042 0.608 1.554 ABORTION 0.785 2.031 3./S.172 0.000 0.108 0.044 0.042 0.372 0.000 1.463 1.000 0.464 HOMOSEX 0.136 0.000 0./S.215 0.223 0.037 35.000 0.215 14.007 0.908 36.720E-03 http://www.667 GOVTRESP 0.028 0.000 0.000 Estimate 0.913 3.000 0.136 3.044 0.333 0.403 0.108 0.683 0.147 0. P-Value 0.851 3.000 0.074 0.edu/~statmath .031 6.038 0.000 1.648 19.978 COMPETE 0.037 0.002 0.111 11.011 0.029 0.indiana.563 0.035 0.042 0.434 1.690 14.552 0.820 10.582 1. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS MORALS WITH ECONOMIC Intercepts PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Variances ECONOMIC MORALS S.000 0.107 0.E.056 0.732 0.527 10.116 STANDARDIZED MODEL RESULTS STDYX Standardization Two-Tailed Est.825 6.941 13.720 34.042 0.E.000 HOMOSEX ABORTION EUTHANAS 5.000 0.536 0.796 0. 0.028 20.002 0.000 STDY Standardization (skip output) STD Standardization (skip output) R-SQUARE Observed Variable PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Two-Tailed S.E.011 0.068 1.883 26.000 0.000 999.022 0.000 0.541 21.044 0.118 6.000 0.000 0.000 0.000 999.000 0.094 0.

CA 90066 Tel: (310) 391-9971 Fax: (310) 391-8971 Web: www. Std E. if a p-value is smaller than . http://www. If the absolute value of the number in this column is greater than 1. GOVTRESP has the lowest . suggesting that it is an unreliable indicator of economic values.E. value for printing the modification index M.719 12.com Copyright (c) 1998-2008 Muthen & Muthen The overall model fit is not great.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 31 (ratio of smallest to largest eigenvalue) MODEL MODIFICATION INDICES Minimum M.96.147. Unstandardized and standardized delta values appear under Residual Variances.631 0. which is higher than the cutoff value of .147 (df=8) large enough to reject the null of a good fit.05.05 level. WITH Statements COMPETE HOMOSEX WITH PRIVTOWN WITH GOVTRESP 31.C. In this case all of the unconstrained loading estimates are significant.760 29.061.I.StatModel.028 0.626 0.000 StdYX E.6 ABORTION has the highest standardized factor loading . column are test statistics that are equivalent to z scores. ABORTION appears to be a reliable indicator of moral values. ratios of the estimates to their standard errors. Under the MODEL RESULTS heading the unstandardized loadings appear along with standard errors. 10.743.05 (significance level) chosen to indicate a good fit. R2 is a standardized factor loading 6 The previous version presents standardized factor loadings under the StdYX column of the MODEL RESULTS.760 86.com Support: Support@StatModel.426 86. The ratios under Est. By contrast.P.indiana. Alternatively. Standardized factor loadings are presented under the STANDIARDIZED MODEL RESULTS heading.522 0. and p-values.C.C. The RMSEA (Root Mean Square Error of Approximation) is . The squared multiple correlations under R-SQUARE provide information on how much variance the common factors account for in the observed variables. with a χ2 statistic of 42.edu/~statmath .195 E.118 Beginning Time: 21:58:43 Ending Time: 21:58:43 Elapsed Time: 00:00:00 MUTHEN & MUTHEN 3463 Stoner Ave. BY Statements MORALS BY GOVTRESP 31./S.961.I. you may reject the null hypothesis at the same significance level that the parameter is zero. Los Angeles.240 0.P.P. CFI (Comparative Fit Index) is . the estimate can be interpreted as significant at the .626 0.

.7322 for COMPETE.677 and 4. and the covariance estimate of . The low R2 of . Under the MODEL MODIFICATION INDICES heading Mplus makes three suggestions: 1) adding covariance between COMPETE and PRIVTOWN will reduce χ2 by 31. for example. for example.2 percent of variance of ABORTION. You may find these statistics appear respectively on standardized and unstandardized results. Two factors do not appear to be closely related each other.433). Because the standardized loading of GOVTRESP on ECONOMIC was so low. The variances of two common factors on unstandardized results are 1. The MODINDICES of OUTPUT command produces model modification indices and gives you some hints about model specification.035 (p<. 2) adding a path from the MORALS latent variable to GOVTRESP by 31. and 3) adding a covariance between HOMOSEX and GOVTRESP by 12.714. morals BY homosex abortion euthanas govtresp.098 is not statistically discernable from zero (p<.535824=.6 percent of the total variation in COMPETE and moral values account for 55. The correlation between the two common factors (economic and moral values) is a very small .edu/~statmath .426. Pay attention to the parts in red. either by sharing measurement error with the HOMOSEX variable or as a direct indicator of the latent morality dimension. Add GOVTRESP to moral values in the MODEL command and then save the input file. NAMES ARE privtown govtresp compete homosex abortion euthanas. Note that the degrees of freedom decrease from 8 to 7. FILE IS values. respectively. economic BY privtown govtresp compete. Note that standardized variances of common factors are set 1. The GOVTRESP item has something in common with the morality dimension. Mplus returns the following output.022 suggests that economic values can explain only a small portion of variation of GOVTRESP. Mplus VERSION 5.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 32 squared. explains 53. MODINDICES. STANDARDIZED. Thus the second suggestion makes theoretical sense and will be estimated. it is possible that the item is actually tapping a different values dimension.1 MUTHEN & MUTHEN 11/17/2008 11:38 PM INPUT INSTRUCTIONS (skip output) INPUT READING TERMINATED NORMALLY (skip output) THE MODEL ESTIMATION TERMINATED NORMALLY http://www.434).dat.631. Economic values.indiana.719. TITLE: DATA: VARIABLE: MODEL: OUTPUT: Second Two Factor Model.

627 0.488 2.411 -16200.000 0.285 1.000 0.072 53.000 3.991 0.938 32479.781 4.000 0.787 13.000 0.046 0.442 999.814 32542.239 0. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH ECONOMIC Intercepts PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS S.0000 RMSEA (Root Mean Square Error Of Approximation) Estimate 90 Percent C.000 0.005 1.000 0.edu/~statmath .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 33 TESTS OF MODEL FIT Chi-Square Test of Model Fit Value Degrees of Freedom P-Value 9.096 49.070 0.752 3.000 0.072 0.590 0.876 54.043 0.997 0.095 0.553 4.05 0.000 0.638 0.312 3.990 15 0.082 0.E./S.136 0.000 13.352 2.610 999.indiana.043 999.000 0.961 0.079 0.044 0.000 0.000 0.256 0.000 0.E.440 4.1947 Chi-Square Test of Model Fit for the Baseline Model Value Degrees of Freedom P-Value CFI/TLI CFI TLI Loglikelihood H0 Value H1 Value Information Criteria Number of Free Parameters Akaike (AIC) Bayesian (BIC) Sample-Size Adjusted BIC (n* = (n + 2) / 24) 20 32441.048 50.000 0.000 http://www.839 999.993 891.066 0.893 7 0.985 SRMR (Standardized Root Mean Square Residual) Value 0.907 -16195. Probability RMSEA <= .000 0.016 MODEL RESULTS Two-Tailed Est.087 0.019 0.845 36.662 5.000 0.I.775 49.000 0.317 0. 1.

037 35.000 999.366 34.755 1.836 14.760 COMPETE 2.490 2.000 0.806 4.000 0.000 0.582 1.000 0.006 0.400 0.734 0.312 0.540 ABORTION 0.707 0.028 0.044 0.000 0.136 0.437 5.034 24.000 0.005 Estimate 0.471 1.358 http://www.312 0.028 55.indiana.078 EUTHANAS 4.000 1.E.440 1.007 0.762 9.548 20.654 0.810 0.476 11.678 0.000 999.015 0.785 5.000 0.047 0.598 0.711 ABORTION 4.906 0./S.000 STDY Standardization (skip output) STD Standardization (skip output) R-SQUARE Observed Variable PRIVTOWN Two-Tailed S.856 HOMOSEX 5.000 0.000 999.642 GOVTRESP 0.123 0.000 0.038 0.356 0.128 2.E.099 0.042 0.000 0.461 EUTHANAS 0.171 34.037 0.000 0.177 0.269 11.000 0.608 1.017 0.041 0.908 36.000 0.000 Residual Variances PRIVTOWN 3.107 0.044 0.000 0.E.042 0.684 0.290 1.465 19.000 999.562 0.213 4.196 0.501 HOMOSEX 0.000 1.608 3.804 0.000 0.005 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 34 Variances ECONOMIC MORALS 1. 0.128 0.365 22. P-Value 0.000 0.125 5.000 0.000 0.000 0.000 STANDARDIZED MODEL RESULTS STDYX Standardization Two-Tailed Est.000 0.028 0.658 0.855 0.747 2.000 0.005 0.000 0.042 0.238 GOVTRESP 6. Est.648 0.940 COMPETE 0.028 0.836 14./S.000 Residual Variances PRIVTOWN 0.637 29.031 5.E.edu/~statmath .998 23. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH ECONOMIC Intercepts PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Variances ECONOMIC MORALS S.721 34.074 0.613 26.463 1.143 0.312 2.

307 13.000 GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS 0.019 (p<.P. however. (skip output) Skipped are some parts of the output. E. The factor loading .1947).646E-03 MODEL MODIFICATION INDICES Minimum M.C. The RMSEA .I.539 0. which are redundant or less informative.316 0.893 (df=7). The conclusion from this analysis is that two nearly orthogonal dimensions underlie the economic and moral values of American citizens. 10.041 0.985) and CFI . which does not reject the null hypothesis of an overall good fit (p<. http://www.P.274 10. It is not clear. which are smaller than those of model 1.218 QUALITY OF NUMERICAL RESULTS Condition Number for the Information Matrix (ratio of smallest to largest eigenvalue) 0.829 12.538 2. Individual standardized loading values remain almost unchanged. whether the GOVTRESP item is tapping either dimension.177 0.000 StdYX E.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 35 0.000).031 3. No modification indices above the minimum value.060 0.000 0.C. Adding a relation between GOVTRESP and moral values appears to make the model fit better.460 0.edu/~statmath . Std E. All unstandardized and standardized factor loadings are statistically significant (p<.000 0.037 0.I. The χ2 test yields a value of 9.043 and .017 0. value for printing the modification index M.C.499 0. The covariance and correlation of two factors are respectively .196 suggests that GOVTRESP manifests both economic and moral values significantly.015. Future surveys should incorporate more reliable measures of economic values.indiana.997 indicate that this model fits the data well. Model modification indices do not suggest any change in this model.005 0.000 0.P.

indiana. PROC CALIS reads a data set values. The LIBNAME statement above defines a library cfa that refers a physical space c:\temp\cfa.values METHOD=ML PALL. INFILE 'c: \temp\cfa\values. RUN. INPUT privtown govtresp compete homosex abortion euthanas. govtresp = lambda2 f1 + e2. ODS HTML redirects the SAS output into the HTML format. homosex = 1. Note that lambda1 and lambda4 are set 1 for model identification. The following is selected from the SAS output. e1-e6 = theta1-theta6.0 f2 + e4. ODS HTML CLOSE. explanatory and confirmatory factor analysis. LINEQS privtown = 1. abortion = lambda5 f2 + e5. ODS HTML.values. The LINEQS statement specifies relationships between latent and manifest variables in equation form. Finally. SAS looks for data sets in the directory once the library cfa is specified. STD f1-f2 = phi1-phi2. LIBNAME cfa 'c: \temp\cfa'.0 f1 + e1. and canonical correlation. DATA cfa. The following CALIS procedure fits the first two factor model discussed above. euthanas = lambda6 f2 + e6. multivariate linear regression. STD and COV statements are place where standard deviations and covariances are defined.0363 http://www. The CALIS Procedure Covariance Structure Analysis: Maximum Likelihood Estimation Fit Function 0.dat and save them into a SAS data set values. RUN. PROC CALIS DATA=cfa. path analysis. This procedure supports various linear models including structural equation model. simultaneous equation model. COV f1-f2 = phi3.dat'. The DATA step above reads six variables from an ASCII text file values. compete = lambda3 f1 + e3. The METHOD=ML uses the maximum likelihood method and the PALL option reports all possible output.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 36 3.sas7bdat from a library sas (c:\temp\cfa).4 Two Factor Model using SAS CALIS This section demonstrates how to estimate a confirmatory factor model using the SAS CALIS procedure. VAR privtown govtresp compete homosex abortion euthanas.sas7bdat in the library cfa. The VAR statement lists the variables used.edu/~statmath .

3390 Schwarz's Bayesian Criterion -14.2699 * f1 + 1.0000 f2 + 1.68319 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 37 Goodness of Fit Index (GFI) 0.1127 * f2 + 1.7307 Variances of Exogenous Variables Variabl Parameter Estimate Standard t Value e Error f1 phi1 0.0000 e1 govtresp = 0.04211 10.0701 lambda2 t Value 3.9854 Bentler & Bonett's (1980) Non-normed Index 0.1104 Chi-Square DF 8 Pr > Chi-Square <.01336 0. & Brett (1982) Parsimonious NFI0.1468 * f1 + 0.0589 ECVI 90% Lower Confidence Limit 0.0445 ECVI 90% Upper Confidence Limit 0.0519 Parsimonious GFI (Mulaik.0434 RMSEA 90% Upper Confidence Limit 0.9270 Bentler & Bonett's (1980) NFI 0.17864 2.91 f2 phi2 0.6039 Bollen (1986) Normed Index Rho1 0.60 e4 theta4 0.0831 lambda5 t Value 13.5767 f1 + 0.0000 f1 + 1.1104 Bozdogan's (1987) CAIC -22.22 Independence Model Chi-Square DF 15 RMSEA Estimate 0.7323 * f1 + 0. 1989) 0.3390 McDonald's (1989) Centrality 0.55379 0.01721 0.9614 Hoelter's (1983) Critical N 428 Manifest Variable Equations with Estimates privtown = 1.03822 14.9611 Elliptic Corrected Chi-Square 38.97844 0.9892 e2 lambda2 compete = 0.70 e3 theta3 0.44621 0.0793 ECVI Estimate 0.04613 9.0798 Probability of Close Fit 0.1472 Bentler's Comparative Fit Index 0.9690 Root Mean Square Residual (RMR) 0.63 e6 theta6 0.5081 Z-Test of Wilson & Hilferty (1931) 4.0000 e2 Std Err 0.8170 e1 govtresp = 0.3876 euthanas = 0.11357 5.0000 e3 Std Err 0.indiana.0000 e6 Std Err 0.03571 19.0000 e5 Std Err 0.44750 0.6310 compete = 1.0000 e4 abortion = 1.49 e5 theta5 0.0607 RMSEA 90% Lower Confidence Limit 0. Mulaik.0261 Pr > Elliptic Corrected Chi-Square <.11439 2.66743 0.67 e1 theta1 0.5270 Chi-Square 42.33257 0.0001 Normal Theory Reweighted LS Chi-Square 41.9882 GFI Adjusted for Degrees of Freedom (AGFI) 0.13 Covariances Among Exogenous Variables Var2 Parameter Estimate Standard t Value Error f2 phi3 0.6809 e3 http://www.8426 * f2 + 1.2546 * f1 + 1.0614 lambda6 t Value 13.78 Var1 f1 Manifest Variable Equations with Standardized Estimates privtown = 0.5308 Akaike's Information Criterion 26.88 e2 theta2 0.4224 lambda3 t Value 3.46369 0.edu/~statmath .9527 James.0064 homosex = 1.0001 Independence Model Chi-Square 891.9114 Bollen (1988) Non-normed Index Delta2 0.04129 23.

PROC CALIS DATA=sas. the covariance of two factors is .2546 and its test statistic is 3. Corresponding statistics were respectively .675 (=. e1-e6 = theta1-theta6.0 f2 + e4.9611.indiana.74332) and .035 in Mplus.44750 1. Mplus reported . govtresp = lambda2 f1 + lambda7 f2 + e2.66743 1. respectively.00000 0. This χ2 is slightly smaller than 42.00000 0.01336 0. which indicates a poor fit. Unstandardized factor loadings of SAS CALIS are different from those Mplus returned.03469 in SAS CALIS and .01336/. compete = lambda3 f1 + e3. SAS CALIS and Mplus respectively report the factor loading .7442 abortion = 0.6690 lambda5 euthanas = 0.098 and . standardized factor loadings.55379 1. However. For instance.7433 * f2 + 0.4462 0.304 and 3.00000 0. RMSEA .0001).5363 0. LINEQS privtown = 1.46369 1.083) in Mplus.6320 (=. Similarly.01721).00000 0. ODS HTML. and correlation of two factors are almost the same as corresponding statistics that Mplus computed.552.0 f1 + e1.00000 0.6680 f2 + 0. The second two factor model is estimated by the following SAS codes.743 for ABORTION. abortion = lambda5 f2 + e5. Note that lambda7 f2 was added to the GOVTRESP equation.01336 and its test statistic is .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 38 lambda3 homosex = 0. euthanas = lambda6 f2 + e6.3326 0. For example. the factor loading of GOVTRESP is . R2.78 (=.03469 Predicted Moments of Latent Variables f1 f1 0.5525 0.values METHOD=ML PALL.0701) in SAS CALIS.1104 (p<. VAR privtown govtresp compete homosex abortion euthanas. The correlation of two factors is .7433 and .01336 f2 0.00000 R-Square 0. http://www.784. R2 are . and CFI .33257 0. respectively.44621 f2 This model has a large χ2 42. although z scores are very similar. homosex = 1.5525 (=.147 that Mplus produced.edu/~statmath .0607 (p<.5629 * f2 + 0.0216 0.3168 Correlations Among Exogenous Variables Var1 Var2 Parameter Estimate f1 f2 phi3 0.97844 1.8266 lambda6 e4 e5 e6 1 2 3 4 5 6 Variable privtown govtresp compete homosex abortion euthanas Squared Multiple Correlations Error Variance Total Variance 0.1472).304/.2546/. STD f1-f2 = phi1-phi2.68319 1.

5088 0. ODS HTML CLOSE.indiana.94033 0.45951 0.33 2.1809 * f1 + 0.9915 0.0000 e6 Variable f1 f2 e1 e2 e3 Variances of Exogenous Variables Parameter Estimate Standard Error phi1 0.9257 0.0520 lambda7 5.5599 1.0328 .8295 * f2 + 0.0000 0.0783 lambda5 13.4653 9.0000 f1 + = 0.9967 8.04626 theta1 0.12276 theta2 0.0000 f2 + = 1.3967 lambda3 2. 1989) Chi-Square Chi-Square DF Pr > Chi-Square Independence Model Chi-Square Independence Model Chi-Square DF RMSEA Estimate RMSEA 90% Lower Confidence Limit RMSEA 90% Upper Confidence Limit ECVI Estimate ECVI 90% Lower Confidence Limit ECVI 90% Upper Confidence Limit Probability of Close Fit Bentler's Comparative Fit Index Elliptic Corrected Chi-Square Pr > Elliptic Corrected Chi-Square Normal Theory Reweighted LS Chi-Square Akaike's Information Criterion Bozdogan's (1987) CAIC Schwarz's Bayesian Criterion McDonald's (1989) Centrality Bentler & Bonett's (1980) Non-normed Index Bentler & Bonett's (1980) NFI James.23 23.1156 -46.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 39 COV RUN.4615 0.2580 9.50067 0.0000 e1 0.8623 0.16838 t Value 2.0000 e4 e5 1.8252 = 0. & Brett (1982) Parsimonious NFI Z-Test of Wilson & Hilferty (1931) Bollen (1986) Normed Index Rho1 Bollen (1988) Non-normed Index Delta2 Hoelter's (1983) Critical N Manifest privtown govtresp Std Err t Value compete Std Err t Value homosex abortion Std Err t Value euthanas Std Err t Value Variable Equations with = 1.0189 .04030 theta3 0. f1-f2 = phi3.9889 0.0000 e3 e2 1.0437 0.2387 * f1 + 0.0601 lambda6 13.93 5.0085 0.5088 -39.97 http://www.edu/~statmath .0691 lambda2 3.8904 -4.9847 0.9929 0.9967 1651 Estimates 1. which represents the factor loading of ECONOMIC.9770 = 1.4558 = 1. Mulaik.35809 0.89 9.1952 891.64191 0. The output is similar to that of the first two factor model but you need to pay attention to lambda7 in the GOVTRESP equation.8844 7 0. 0.12399 phi2 0.9972 0. 0.0178 0. The CALIS Procedure Covariance Structure Analysis: Maximum Likelihood Estimation Fit Function Goodness of Fit Index (GFI) GFI Adjusted for Degrees of Freedom (AGFI) Root Mean Square Residual (RMR) Parsimonious GFI (Mulaik.0000 1.0828 * f2 + 0.9988 0.22 15 0.9762 0.8008 0.2893 * f2 + 1.0436 0.

46 theta6 0.32 Manifest Variable Equations with Standardized Estimates privtown = 0.7066 * f1 + 0.5387 0. Standardardized factor loadings and correlation of two factors reported by both software packages are virtually same. The covariance of two factors is .1961 * f2 + 0.33 theta5 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 40 Variable e4 e5 e6 Variances of Exogenous Variables Parameter Estimate Standard t Value Error theta4 0.00000 0.32 (=.00000 e2 1 2 3 4 5 6 Variable privtown govtresp compete homosex abortion euthanas R-Square 0.7340 * f2 + 0.indiana. but their test statistics are similar.8844 (p<.01814 0.5623 * f2 + 0.4993 0.8269 e6 lambda6 Squared Multiple Correlations Error Variance Total Variance 0.9697 lambda2 lambda7 compete = 0.3581 0.0328).05 level.00000 0.68382 1.00589 0.4595 0.45951 This second model has a smaller χ2 9.28 Var1 f1 Covariances Among Exogenous Variables Var2 Parameter Estimate Standard t Value Error f2 phi3 0.00589 f2 0. http://www.6779 f2 + 0. CFI .00589. Government responsibility (GOVTRESP) is significantly explained by both economic and moral values.01814) that is very close to .00589 0.35809 0.196 that Mplus returned. which do not reject the null hypothesis at the . but its test statistic is .46127 1. unstandardized factor loadings of SAS CALIS are different from those of Mplus.0189 (p<.00000 0. As discussed above.03772 14.6792 e5 lambda5 euthanas = 0.00589/.64191 1.01453 Predicted Moments of Latent Variables f1 f2 f1 0.00000 0.7352 e4 abortion = 0.edu/~statmath . SAS CALIS reports .1428 * f1 + 0.68382 0. slightly larger than .8012 e1 govtresp = 0.03548 19.50067 1.0597 0. which is the same as .54049 1.043 in Mplus.94033 1.04025 11.1961 as a standardized factor loading of moral values on GOVTRESP.54049 0.46127 0.3162 Correlations Among Exogenous Variables Var1 Var2 Parameter Estimate f1 f2 phi3 0.1952) and RMSEA .5984 f1 + 0.7076 e3 lambda3 homosex = 0.317 Mplus returned.9967 also indicates a good fit.00000 0.

vb listed in the default directory Examples. Program Editor. This section. At Program Editor. and then select a sample program Ex08. See Appendix if you want to estimate a model using Amos Graphics. Amos Program Editor can be launched from any computer running Windows in the UITS Student Technology Centers by going to Start → All Programs → Departmentally Sponsored → Statistics-Math → Amos 16 → Program Editor. The following screen will display.5 Two Factor Model using Amos Amos consists of several applications including Amos Graphics. http://www. click File → Open to open a dialog box. however. uses Program Editor to take advantage of its simplicity at the expense of burdensome programming. and Seed Manager. File manager.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 41 3.indiana. which makes it easy for beginners to specify models. Most people begin with Amos Graphics to fit a model because it provides a point-and-click GUI environment.edu/~statmath .

The beginning of the program describes a program title and backgrounds.Standardized() computes standardized estimates. which is the same as that of two factor models that we want to fit.BeginGroup() and Sem. http://www.vb. Sem. Pay attention to Sem.TextOutput() creates an output file and Sem. Click File → Save As… to open Save As dialog box. Main program begins with Sub Main() and ends with End Sub.NET) or C# language to specify a model. Provide a name values and then click Save. Amos uses Visual Basic (VB. the file name became values. Now.edu/~statmath .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 42 We are going to copy its syntax structure. while the latter specifies model to estimate.AStructure() functions.indiana. Let us rename the program file first. The former tells where a data file is located.

Now click Model Fit on the left pane to see the other goodness-of-fit statistics.000).sav"). cubes with govtresp. Click Notes for Model on the left pane to get the chi-square test of model fit (see following screenshot). RMSEA .061 and CFI . change error terms from e1 through e6 as shown in the above screenshot. and so on. Replace visperc with privtown.indiana. Finally. Note that the estimate of factor loading is set 1 for the purpose of identification. which are the same as what Mplus and LISREL returned.edu/~statmath .BeginGroup() to look like Sem. The first SEM.AStructure() functions. In the six SEM.sav from c:\temp\cfa. Change Sem. Amos reads data set values. Chi-square χ2 of 42.110 (df=8) is so large as to reject the null hypothesis of a good fit (p<. Click File → Run or click Run button on the right bottom to fit this model. Amos Output pops up if no error is found.961.AStructure() tells that the observed variable privtown is explained by a latent variable economic and error e1. You may see the list of results. suggests that the fit of the model is questionable. replace spatial with economic and verbal with morals. Replace Example 8: with First Two Factor Model and then remove other comments (see the screenshot below).© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 43 Look at the comments in green. Model Fit Summary CMIN http://www.BeginGroup("c:\temp\cfa\values.

225 LO 90 .000 1.000 .751 PRATIO .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 44 Model Default model Saturated model Independence model RMR.000 903.237 PCLOSE .000 .000 .000 PNFI .988 1.000 Click Estimates on the pane to check parameter estimates.557 GFI .000 PCFI .147 .508 .698 HI 90 .044 .036 .953 1.036 .000 FMIN .561 NFI Delta1 .000 977.edu/~statmath .779 LO 90 . a critical ratios.433 .180 933.029 .061 .079 .221 DF 8 0 15 P .000 782.059 .000 .221 LO 90 17. Under the Regression Weights heading the unstandardized loadings appear along with standard errors.969 .294 BIC 133.841 169.043 .015 .059 .769 F0 .675 HI 90 .961 1.indiana.000 .268 42.867 MECVI .000 .000 CMIN/DF 5.000 .000 IFI Delta2 .779 AIC 68.000 .080 .000 .05 427 33 HOELTER .558 CAIC 146.180 939.000 RFI rho1 .533 .000 891.01 553 40 ECVI .000 TLI rho2 .415 RMR .302 .000 1.110 .000 .411 .110 42.221 BCC 68.841 148.036 .083 HI 90 58.911 .036 .558 RMSEA .050 .756 LO 90 .036 . GFI Model Default model Saturated model Independence model Baseline Comparisons Model Default model Saturated model Independence model NPAR 13 21 6 CMIN 42.264 59.212 HI 90 .927 .786 AGFI .000 Parsimony-Adjusted Measures Model Default model Saturated model Independence model NCP Model Default model Saturated model Independence model FMIN Model Default model Saturated model Independence model RMSEA Model Default model Independence model AIC Model Default model Saturated model Independence model ECVI Model Default model Saturated model Independence model HOELTER Model Default model Independence model HOELTER . and http://www.961 1.844 NCP 34.700 PGFI .513 .110 .000 .000 CFI .255 903.376 .000 876.

715 .006 . C.732) have highest standardized factor loadings.000 .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 45 p-values.019 5.147 .668 .076 13.388 *** .877 *** 23.034 .R.126 .674 *** 5.R.957 .907 .627 *** Label http://www.Model 1) Maximum Likelihood Estimates Regression Weights: (Group number 1 . indicate that the respective factor explains a respectable portion of the variance.573 7.372 C.437 Label Correlations: (Group number 1 .631 *** 1.404 3.304 .018 . ABORTION (.economic homosex <--.E.Model 1) privtown <--.morals Estimate S. Squared multiple correlations (R2) corresponding to the six observed variables.E.366 .economic govtresp <--.035 Variances: (Group number 1 .edu/~statmath . However.743 .449 3.009 14.046 13.Model 1) economic morals e1 e2 e3 e4 e5 Estimate S.487 3.098 . GOVTRESP has a lowest factor loading of . GOVTRESP does not seem to tap the same values dimension as the other two economics questions.637 . C.596 .698 *** 2.731 *** Label Standardized Regression Weights: (Group number 1 .732 .563 Covariances: (Group number 1 .economic compete <--.1472) of variation in GOVTRESP although regression weights are all significant though. The standardized estimates under Standardized Regression Weights can be interpreted as the correlation between the observed variable and the corresponding common factor. 1.147. P .morals abortion <--. standard errors are slightly different from Mplus and LISREL counterparts.Model 1) economic <--> morals Estimate S. which are arranged in an arbitrary order.743) and COMPETE (.851 . These unstandardized and standardized estimates are the same as what Mplus and LISREL produced. P 1.003 1.645 1.indiana.004 9.577 4. P 2.Model 1) privtown <--govtresp <--compete <--homosex <--abortion <--euthanas <--economic economic economic morals morals morals Estimate .R.777 .E.morals euthanas <--.084 3.491 *** 10.Model 1) economic <--> morals Estimate . Estimates (Group number 1 .351 . Economic values (ECONOMIC) can explain only 2.577 .2 percent (=.000 1.677 .297 2.Model 1) Scalar Estimates (Group number 1 .

edu/~statmath . Notes for Model (Model 1) Computation of degrees of freedom (Model 1) http://www. It is like adding an arrow from MORALS to GOVTRESP on a path diagram.317 . A small RMSEA of .195). Header Module MainModule ' Second Two Factor Model Sub Main() Dim Sem As New AmosEngine Try Sem.119 .Smc() Sem.AStructure("abortion = morals + (1) e5") Sem.Model 1) euthanas abortion homosex compete govtresp privtown Estimate .BeginGroup( "c:\temp\cfa\values.TextOutput() Sem.AStructure("compete = economic + (1) e3") Sem.Standardized() Sem. Let us set a relationship between MORALS to GOVTRESP by adding + morals to the second Sem.536 .sav") Sem.FitModel() Finally Sem.AStructure("govtresp = economic + morals + (1) e2") Sem. An alternative possibility is that GOVTRESP is also tied to the morality dimension.R. which is too small to reject the null hypothesis of a good fit (p<. an Amos program written in Program Editor is not able to produce a path diagram and Amos graphics does not generate a VB or C# program on the basis of a path diagram drawn.884. GOVTRESP was only weakly accounted for by the ECONOMIC variable.E.AStructure() below.AStructure("privtown = (1) economic + (1) e1") Sem.Dispose() End Try End Sub End Module Run this program by clicking File → Run and then click Notes for Model in the Amos Output.022 .indiana.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 46 e6 Estimate S. Look at the right part in red.215 19. C.AStructure("homosex = (1) morals + (1) e4") Sem.132 P *** Label Squared Multiple Correlations: (Group number 1 .553 .997 indicate a good fit of this model. hinting that the survey item was not tapping the same values dimension as the other two economic values indicators. χ2 (df=7) decreases down to 9.333 Unlike LISREL.019 and a large CFI of . 4.446 . The overall model fit appears quite good.AStructure("euthanas = morals + (1) e6") Sem.

000 .180 933.558 CAIC 122.221 LO 90 .985 .019 . GFI Model Default model Saturated model Independence model Baseline Comparisons Model Default model Saturated model Independence model NFI Delta1 .000 .195 .558 RMSEA .000 .221 DF 7 0 15 P .991 .997 1.000 876.000 .055 42.000 .044 .465 .671 169.000 977.000 CFI .000 .009 .700 PGFI .993 .332 .edu/~statmath .671 148.786 AGFI .14): Result (Model 1) Minimum was achieved Chi-square = 9.000 .237 PCLOSE .884 .083 HI 90 15.561 NPAR 14 21 6 CMIN 9.000 PCFI .013 .844 NCP 2.000 782.884 Degrees of freedom = 7 Probability level = .000 RFI rho1 .769 F0 .756 LO 90 .000 1.000 IFI Delta2 .000 .989 1.884 42.000 .000 CMIN/DF 1.255 903.000 TLI rho2 .000 903.467 .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 47 21 14 7 Number of distinct sample moments: Number of distinct parameters to be estimated: Degrees of freedom (21 .119 .221 BCC 38.461 .indiana.000 .997 1.225 LO 90 .675 HI 90 .000 PNFI .976 .997 1.557 GFI .415 Parsimony-Adjusted Measures Model Default model Saturated model Independence model NCP Model Default model Saturated model Independence model FMIN Model Default model Saturated model Independence model RMSEA Model Default model Independence model AIC Model Default model Saturated model Independence model ECVI AIC 37.412 59.751 PRATIO .000 http://www.000 RMR .000 .000 FMIN .884 .195 Model Fit Summary CMIN Model Default model Saturated model Independence model RMR.294 BIC 108.000 .000 891.000 1.212 HI 90 .404 .000 .180 939.002 .

030 is not statistically discernable from zero (p<.598 .072 13.R.043 and .239 .977 .morals abortion <--.143 . Estimates (Group number 1 .E.285 .043 5.256 . the correlation between the two common factors is a very small -.000 .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 48 Model Default model Saturated model Independence model HOELTER Model Default model Independence model ECVI . Latent variables ECONOMIC and MORALS account for 6 percent of the total variation in GOVTRESP.05 1650 33 HOELTER . P 1.Model 1) Scalar Estimates (Group number 1 .562 Covariances: (Group number 1 .036 . GOVTRESP appears an unreliable indicator of both economic and moral values.745 Label Correlations: (Group number 1 .325 .560 *** 1.808).043 .003 1.economic homosex <--. Unstandardized and standardized estimates are respectively .082 3.011.779 LO 90 .033 . Corresponding squared multiple correlations (R2) range from .Model 1) Maximum Likelihood Estimates Regression Weights: (Group number 1 .030 .991 . C.707 . and the covariance estimate of -.Model 1) privtown <--govtresp <--govtresp <--compete <--homosex <--abortion <--euthanas <--economic economic morals economic morals morals morals Estimate .036 . Covariance and correlation of two factors are .015 http://www. P .867 MECVI .825 *** .539 (ABORTION).Model 1) privtown <--.196 .627 .499 (COMPETE) and .R.033 .456 *** .422 2.678 .196.indiana.801 *** Label Standardized Regression Weights: (Group number 1 .015 but they are not statistically discernable (p<.7).036 .779 HOELTER .Model 1) economic <--> morals Estimate S.239 and . Other standardized factor loadings range from .000 .E. Finally.economic govtresp <--.045 13.043 .Model 1) economic <--> morals Estimate .edu/~statmath .562 (EUTHANAS) to .morals compete <--. C. which are smaller than those of other estimates.economic govtresp <--.132 .morals Estimate S.036 .698 HI 90 .734 (ABORTION).morals euthanas <--.734 .060 (GOVTRESP) to .01 2167 40 Amos reports the factor loading of MORALS on GOVTRESP on the third row.

078 .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 49 Variances: (Group number 1 .399 4.R.619 6.Model 1) economic morals e1 e2 e3 e4 e5 e6 Estimate S.888 .indiana.625 4. there appear to be two orthogonal dimensions which underlie American attitudes on a number of different issues: one representing economic values and the other representing moral values.Model 1) euthanas abortion homosex compete govtresp privtown Estimate . 1.214 C.358 In summary.460 . P 2.edu/~statmath .974 .060 .857 .489 3.E.760 .330 *** 2.004 9.237 .228 *** 23.003 14.459 *** 19.806 . however. and future surveys should employ a more reliable measure.539 .356 4.855 . It is unclear which dimension the GOVTRESP item was tapping.933 *** 5.711 .123 .499 .290 2.275 *** Label Squared Multiple Correlations: (Group number 1 .329 *** 11.316 . http://www.960 5.

018 (.00 .00 1.094) .028) .6 Summary Tables 4 and 5 compare the unstandardized and standardized factor loadings from each software package for both two-factor models.098a (.084) 1. DF=8) LISREL Mplus SAS CALIS Amos Unstd. b correlation between F1 (economic values) and F2 (moral values) Mplus.000 .061 .01b 9.04a (.147 .893 .082) 1.0701) 1.637 (.0691) 1.1961 .668 .678 (.0000 .019 .60 . The tables also present standard errors in parentheses.563 .68 . http://www.076) .0172) Standard .02 (.14 . 1.035b (.627 (.2387 (. Table 4 Comparison of Estimates: First Two Factor Model (N=1.6680 .038) .0614) .637 (.1809 (. RMSEA.884 .2546 (.015b 9.732 (.110 .160. while other software packages provide various fit statistics including χ2.707 (.143 (.edu/~statmath .627 (.072) .5623 .30 (.5767 .9611 2.7433 . F1 PRIVTOWN F1 GOVTRESP F1 COMPETE F2 HOMOSEX F2 ABORTION F2 EUTHANAS F1 F2 1.0607 .00 .00 .147 . and Amos produce almost same statistics.26 (.046) .256 (.00 1. and Comparative Fit Index (CFI).74 .126) 42.0189 .73 .24 (.71 .7340 .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 50 3.9967 2.28 (.029) .577 (.045) . F1 PRIVTOWN F1 GOVTRESP F1 COMPETE F2 HOMOSEX F2 ABORTION F2 EUTHANAS F2 GOVTRESP F1 F2 1.072) .15 .351 (.562 (.1468 .3967) 1.044) 42.0000 1.0000 1.8844 .598 (.8295 (.028) .00 .56 .8426 (.10a (.0059a (. 1.56 .083) 1.076) .99 (.00 .018 (.285 (.442) 1.88 .991 (.63 (.000 1.0347b 42.098a (.45) 1.0145b 9.13) Standard .997 Unstd. LISREL.239 (.743 .732 .67 .42) 1.6779 .019 1. DF=7) LISREL Mplus SAS CALIS Amos Unstd.00 Unstd.046) . χ2.433) 1.147 (.997 Unstd.0783) .0000 .1104 . and CFI.285 (.043) .043a (.58 .060 .598 .046) .076) .0828 (.35 (.734 . 1.136) Standard .046) .422) 1. CFI Standard errors appear in parentheses a covariance.678 . 1.028) .304 (.038) .043) .0601) .0831) .0181) Standard .03b 42.015b (.143 .196 (.0134a (. 1.256 (.304 (.64 (.061 .045) .5629 . Mplus reports key goodness-of-fit statistics. CFI Standard errors appear in parentheses a covariance.13) Standard .107) .562 .082) 1.196 .1428 .7066 .961 Standard .707 .indiana.449) 1. SAS/STAT CALIS reports different unstandidized factor loadings but same standardized statistics.743 .2893 (.028) . 1. Root Mean Square Error of Approximation (RMSEA).577 .035b Unstd.043) .1127 (.0520) .043a (.7323 .961 Unstd.000 .734 .00 .118) .047) 9.124) Standard .20 .072) .132) Standard .239 (.4224) 1.350 (.2699 (.96 Unstd.000 .028) . RMSEA.019 .084) 1. RMSEA.160.991 (.11 .034) .668 (.082) 1.5984 . b correlation between F1 (economic values) and F2 (moral values) Table 5 : Comparison of Estimates: Second Two Factor Model (N=1.73 .563 (.125) .

) in SPSS. Just to the right of the toolbar buttons is a column that will display information about the model after estimates have been calculated. or drop out of a panel. A subject may fail to complete a test in an experimental setting. http://www. A non-technical overview of different methods for handling missing data in the context of structural equation models is available in Enders (2001). After the Data Files dialog box opens.sav containing missing values.edu/~statmath . Another approach is pairwise deletion that removes observations with missing data in any one of two variables when computing their covariance.1 Missing Data Issue Missing data is a pervasive problem in the social sciences. 4. Other approaches to dealing with missing data. Over the last 30 years more sophisticated means have emerged for dealing with missing data. and Mplus. Navigate to the C:\temp\CFA folder and choose values_full. LISREL. researchers choose to drop all observations from subjects that have missing observations on any of the items included in the model. refuse to give an answer to a particular survey item.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 51 4. Then Okay. The data to be analyzed has been saved as an SPSS file named values_full. The following screen will display: On the far left pane appear the different tools that can be used to create path diagrams. may also be available depending on the specific software packages. All missing observations have been coded as system missing (. click on File Name. though the description of capabilities of specific computer packages is already dated. The remainder of the screen contains the area where the path diagram will be drawn. this document will consider Full Information Maximum Likelihood (FIML) estimation which makes maximal use of all data available from every subject in the sample. Because it is available in Amos. Confirmatory Factor Analysis with Missing Data This section shows how to estimate the two-factor model when the raw data matrix includes missing observations. Click Open. 4. In many cases. This method of dealing with missing data can use all available data but each element of covariance matrix may be based on different observations. such as multiple imputation via Bayesian simulation. Amos can be launched from any computer running Windows in the UITS Student Technology Centers by going to Start → All Programs → Departmentally Sponsored → Statistics-Math →Amos 16 → Amos Graphics. many of which have been incorporated into structural equation modeling software. Unfortunately dropping incomplete cases results in sacrificing information from the sample and can lead to biased estimates when the data is not missing completely at random.sav in the C:\temp\CFA folder. Click File →Data Files to load data. This approach to handling missing data is referred to as listwise deletion and is the default in programs such as SPSS and Stata.indiana.2 CFA with Missing Data using Amos This section fits the model using Amos Graphics instead of Program Editor.

indiana. Add a covariance between the two latent and clicking and variables (common factors) by choosing the Draw Covariances button dragging a two-headed arrow from one factor to the other (see the screenshot on the next page). Move the mouse pointer just below the oval and click once to create a second oval. http://www. Select the blue oval and click six times on the left of six boxes to represent measurement error specific to each of the observed indicators. Click on the of Draw Observed Variable (alternatively click on Diagram → Draw Observed). blue rectangle button Then click six times to create a total of six boxes for observed variables.7 By default Amos sets the metric of each error term 7 Alternatively. click on the Draw unobserved variable button (alternatively click on Diagram → Draw Unobserved) and click some place on the right workspace to draw an oval representing a latent variable. Click three times inside each oval to add a total of six indicators and their respective error terms (see the following screenshot).© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 52 in the tool box On the left pane. Click on the Draw a latent variable button or add an indicator to a latent variable button . you may draw ovals and boxes and then link them with appropriate arrows manually. Select the Draw Paths button for drawing single headed arrows. Click and drag from the common and unique factors to the appropriate observed variable.edu/~statmath .

The factor loading of the first indicator for each latent variable is also set to one. Name the variable ECONOMIC and then click the second oval to name it MORALS. To name the observed variables. Click the first error oval and name it d1 and adjust the font size to 12 point. add an arrow from the MORALS latent variable to the GOVTRESP indicator as suggested in the earlier analysis. choose the Text tab. When the Object Properties dialog box opens. Finally. right-click in the first oval and choose Object Properties. choose View → Variables in Dataset. Rotate each latent variable by choosing the Rotate the and clicking each factor three times until you are indicators of a latent variable button satisfied with the appearance so that the path diagram should now look like the following. you can change the position of parts of the diagram after choosing the Move objects button or resize the box after clicking on the Change the shape of objects button . Click and drag the names of each variable to the appropriate box in the path diagram. Do the same to name the error terms d2 through d6. To label the latent variables.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 53 by constraining the path parameters to one.edu/~statmath .indiana. If the names do not fit. Click X in the upper right hand corner to close the Object Properties dialog box. The path diagram now should look something like the following: http://www.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 54 For comparison with LISREL and Mplus. let us set factor loading on PRIVTOWN and HOMOSEX to one and factor loadings on COMPETE and EUTHANAS free. Click on the arrow from ECONOMIC to PRIVTOWN and enter 1 in Object Properties dialog box (see screenshot above). Repeat the same to set the factor loading on HOMOSEX to 1 after set factor loadings on EUTHANAS free. The final path diagram should look like this: http://www. Right-click on the arrow from ECONOMIC to COMPETE to open Object Properties dialog box and then click on the Paramaters tab and delete 1 in the field labeled Regression weight.edu/~statmath .indiana. Click X to close the Object Properties dialog box.

Also place checks next to Standardized Estimates and Squared Multiple Correlations. Now. it is necessary to request that Amos estimate means and intercepts (required for FIML estimation). click on the Standardized estimates option in the column between the tools and the workspace. we are ready to fit the model. All factor loadings except for GOVTRESP are large. Amos now displays the standardized factor loadings on arrows and the squared multiple correlation coefficient for each observed variable.01). The unstandardized estimates are displayed by default.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 55 When missing values are present. Next click on the Output tab. To bring up the standardized estimates. which is not the default. while the correlation between two latent variables is negligible (-. and select Estimate means and intercepts. To see the results in the path diagram click on the View the output path diagram button on the second pane.edu/~statmath .indiana. Note that for some models with many parameters and missing data. Amos (and all SEM software) may require a large number of iterations to estimate a χ2 statistic. Choose View →Analysis Properties. Go to Analyze → Calculate Estimates. The path diagram for unstandardized estimates is often hard to read especially when the diagram does not have enough space for parameter estimates. click the Estimation tab in the Analysis Properties dialog box. Minimization History is checked by default. http://www.

indiana.20): Result (Default model) Minimum was achieved Chi-square = 9. P 1. C.Morals govtresp <--.Default model) Estimate S.009 *** .911 Degrees of freedom = 7 Probability level = .633 . go to View → Text Output.042 5.302 .edu/~statmath .154 .200 .Economic euthanas <--.369 3.Default model) Maximum Likelihood Estimates Regression Weights: (Group number 1 .© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 56 To see more detail about the results.Morals homosex <--.691 .045 14.000 .Economic govtresp <--.615 compete <--.Morals abortion <--.255 .071 13.Economic http://www.070 *** Label compete <--.992 .R.000 . A selected portion of the output is the following: Notes for Model (Default model) Computation of degrees of freedom (Default model) Number of distinct sample moments: Number of distinct parameters to be estimated: Degrees of freedom (27 .695 *** 1.Morals Standardized Regression Weights: (Group number 1 .194 27 20 7 Estimates (Group number 1 .Default model) Estimate .001 .E.Economic privtown <--.Economic govtresp <--.215 .Economic privtown <--.937 *** 1.Default model) Scalar Estimates (Group number 1 .082 3.

760 50. P 3.312 .Default model) Estimate -.472 P *** *** *** *** *** *** Label compete govtresp privtown euthanas abortion homosex Covariances: (Group number 1 .542 *** 23.016 .911 .115 .065 2.774 .R.596 C.000 NPAR 20 27 6 CMIN 9.997 1.998 .477 homosex abortion euthanas privtown govtresp compete Model Fit Summary CMIN Model Default model Saturated model Independence model Baseline Comparisons Model Default model Saturated model Independence model NFI Delta1 .745 *** 19.194 . P -.407 *** 14.078 3.847 6.902 .E.000 CMIN/DF 1.076 50.000 910.164 DF 7 0 21 P .737 .990 .002 10.118 .134 -.361 *** 11.366 54.indiana.E.071 4. 3.000 .Default model) Estimate S.150 .Default model) Estimate S.118 .000 CFI .379 .967 .095 C.587 37.543 .396 3.997 1.850 .000 . 1.235 *** Label Economic Morals d3 d2 d6 d5 d4 d1 Squared Multiple Correlations: (Group number 1 .321 .566 .R.289 4.732 .000 RFI rho1 .E.353 5.069 4.989 1. 49.645 .Default model) Estimate S.677 .Default model) Estimate .360 .743 55.000 TLI rho2 .541 .000 .000 *** 3. C.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 57 euthanas <--abortion <--homosex <--govtresp <--- Morals Morals Morals Morals Estimate .000 IFI Delta2 .604 4.edu/~statmath .458 .176 Intercepts: (Group number 1 .213 4.481 *** 5.872 .005 Morals <--> Economic Variances: (Group number 1 .054 .906 Label Morals <--> Economic Correlations: (Group number 1 .R.021 .416 43.341 Parsimony-Adjusted Measures http://www.485 2.086 4.442 .

194.164 BCC 50. 8 http://www.013 .001.05 significance level.002 .150 HI 90 15.574 FMIN . Three asterisks (***) indicate that the p-value is smaller than .042 .317 922.235 BIC CAIC ECVI .000 PCFI .000 1.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 58 Model Default model Saturated model Independence model NCP Model Default model Saturated model Independence model FMIN Model Default model Saturated model Independence model RMSEA Model Default model Independence model AIC Model Default model Saturated model Independence model ECVI Model Default model Saturated model Independence model HOELTER Model Default model Independence model PRATIO .759 F0 .000 PNFI .000 .742 LO 90 . but we rotated the path diagram so that manifest variables are vertically arranged.05 signifies the parameter is statistically discernable from zero at the .448 .146 54. The RMSEA of .000 AIC 49. Under the Regression Weights heading.000 922.045 .045 .000 .330 .188 LO 90 .000 .043 .911 (df=7).769 HOELTER .000 .690 HI 90 .045 .000 .827 RMSEA .019 .045 .000 .008 . In this case all of the unconstrained estimates are statistically significant.019 and CFI of .000 .178 HI 90 . This p-value is too high to reject the null hypothesis of a good fit.855 MECVI .769 LO 90 . A critical ratio greater than 1.052 . it is a bit confusing and inconvenient to read results. the unstandardized loadings appear along with standard errors.000 889. The χ2 test yields a statistic of 9. and p-values.8 Variables are recognized from the left when created by clicking Draw a latent variable button or add an indicator to a latent variable button to right.198 PCLOSE . Notice that the order of variables listed in the table is not the same as what is specified in the path diagram.911 .333 .000 .911 54.000 794. critical ratios (test statistics).662 HI 90 .indiana.000 991.01 2235 52 The overall model fit appears quite good.05 1702 44 HOELTER .000 . which has a corresponding p-value of .96 or a p-value smaller than .000 NCP 2.987 .164 LO 90 .042 .997 also suggest that the model fits the data well.332 .039 .edu/~statmath .

GOVTRESP has a very low R2 of only . and clicking on EUTHANAS and others. in order of increasing magnitude. A prompt appears immediately to save the file as a PRELIS system file (.3 CFA with Missing Data using LISREL Launch LISREL and open the file values_full.691.054. Finally.psf). whereas the standardized regression weights provide more intuitive information about the strength of factor loadings. and for ABORTION it is . .154 for ECONOMIC and . The squared multiple correlations provide information about how much variance of an observed variable the factors can account for. Click Open to import the data set.indiana.379 (PRIVTOWN). enter -999999. which is not statistically distinguishable from zero. .sav).176 for MORALS). Enter values_full and click Save.0 in the first empty field. The remaining R2 statistics are. and .677. For EUTHANAS the factor loading is . Highlight each variable name by clicking on PRIVTOWN.edu/~statmath . Despite receiving a path from both latent variables.sav by choosing File → Import Data.sav. Click on the Missing Values radio button. for HOMOSEX it is . holding down the shift key.458 (HOMOSEX). 9 http://www. Click on Missing Values to bring up the Missing Values box. it is necessary to declare them as such. the other indicators have moderate to strong standardized loadings.543 (ABORTION).321 (EUTHANAS). Click OK. Go to Data → Define Variables to open the Define Variables dialog box. and choose values_full. then OK again.9 To make sure that PRELIS understands these are missing values.005.615.566. Notice that missing observations are coded -999999. 4. the covariance between the two common factors is -.737. However. When Open dialog box opens. and check the Apply to all option.012 and their correlation is -. change Files of type to SPSS Data File(*. navigate to the folder C:\temp\CFA. A spreadsheet will open displaying the data. for COMPETE it is .477 (COMPETE). suggesting that it is an unreliable indicator of both economic and moral values. . The GOVTRESP has low standardized loadings on both factors (.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 59 The unstandardized weights are highly sensitive to model constraints. . for PRIVTOWN it is .0.

we drew a path diagram for two factor CFA and then generated a corresponding SIMPLIS or LISREL syntax.out extension and stored in the working directory. Go to File → New.indiana. type in values_full.2. This approach is intuitive but wordy to explain each step all the way to the final.2. to fit the model. choose SIMPLIS Project in the New dialog box. The unstandardized estimates are Click on the Run LISREL button immediately displayed in the path diagram along with two measures of overall fit: 2 and RMSEA. To view the standardized results. When the Save As dialog box opens.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 60 In 3. The file values_full. and click OK.out looks like the following: DATE: 1/17/2009 TIME: 19:09 L I S R E L BY 8.80 Karl G. Here let us directly write a SIMPLIS program in the text editor. choose Standardized Solution from the Estimates pull-down menu as we did in 3. Notice that values_full. Inc. An empty text window opens for you to type in the commands shown in the screenshot in the previous page. http://www. Jöreskog & Dag Sörbom This program is published exclusively by Scientific Software International.edu/~statmath . The path diagrams for unstandardized (left diagram below) and standardized (right diagram) estimates will look like these: More detailed information can be obtained by looking at the output text file generated after estimation.psf is used as input data and the metric of the two common factors is set by constraining these factor loadings on PRIVTOWN and HOMOSEX to equal one. This file is given the same name as the path diagram plus an . and then click Save.

12 ..64 compete -------homosex -------abortion -------euthanas -------- privtown govtresp compete homosex abortion euthanas 5.74 http://www.29) 3.33 10.07 23. Website: www.21*Morals.15 0. (847)675-0720.08 6.042) (0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 61 7383 N.70 1.spj: Two factor model with missing data Raw Data from file 'c: \temp\cfa\values_full. Phone: (800)247-6113.03 8. Suite 100 Lincolnwood.08 0.psf' -------------------------------EM Algorithm for missing Data: -------------------------------Number of different missing-value patterns= Convergence of EM-algorithm in 3 iterations -2 Ln(L) = 33282.38 (0. IL 60712.81 3.71 12 Note: The Covariances and/or Means to be analyzed are estimated by the EM procedure and are only used to obtain starting values for the FIML procedure Latent Variables Economic Morals Relationships privtown = 1.edu/~statmath .00*Economic govtresp = Economic Morals compete = Economic homosex = 1. Errorvar.11 govtresp -------7.89 0.55 2.28 -0.70 5.23 govtresp = 0.55078 Percentage missing values= 0.S.= 6.A.054 (0.ssicentral.59 4.indiana.02 0.26 0.00*Economic. R² = 0.74 -0. Inc. R² = 0.27 0. Errorvar.082) (0.com The following lines were read from file c: \Temp\CFA\values_full.87 .06 Two factor model with missing values Number of Iterations = 5 LISREL Estimates (Maximum Likelihood) Measurement Equations privtown = 1. Lincoln Avenue.00 0.= 3.00*Morals abortion = Morals euthanas = Morals Path Diagram End of Problem Sample Size = 1200 Two factor model with missing values Covariance Matrix privtown -------5. Fax: (847)675-2140 Copyright by Scientific Software International. U.60) 5.81 3.30*Economic + 0. 1981-2006 Use of this program is subject to the terms specified in the Universal Copyright Convention.27 -0.

which are not statistically distinguishable from zero.indiana.01 19.48) 10.02 and -.35) 13.85 (0.00 Global Goodness of Fit Statistics.edu/~statmath .19) Root Mean Square Error of Approximation (RMSEA) = 0.= 4. the covariance and correlation between the two common factors are -. Since Mplus does not treat blanks as missing values.99*Morals.12 Morals -------- 4.48 abortion = 0. Errorvar.00 .41 euthanas = 0. Errorvar. Finally.= 5.11 .019 90 Percent Confidence Interval for RMSEA = (0. R² = 0. http://www. 0. missing values must be coded as a number in the raw data and explicitly specified in the VARIABLE command of the Mplus syntax. 4.02 . respectively.37) (0.0 .90 (0. indicating a good fit.40) 14.54 for ABORTION. Errorvar. However.15 for ECONOMIC and .= 3.21) 14. however.01.73 .46 (0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 62 compete = 1.13) -0. reported here.48 (0.26 3.043) P-Value for Test of Close Fit (RMSEA < 0.32 (0. R² = 0.92 (df=7) and RMSEA is only . suggesting that it is a weak indicator of both economic and moral values.36 Covariance Matrix of Independent Variables Economic -------Economic 1.32 for EUTHANAS to .63*Morals. Missing Data Case -2ln(L) for the saturated model = -2ln(L) for the fitted model = 33282.20*Economic.54 (0. Errorvar. The χ2 test yields a value of 9. the other indicators have moderate to strong standardized loadings with higher squared multiple correlations ranging from .02 (0.071) (0.= 4.94 11. CFI is not. Mplus cannot directly read an SPSS data file.045) (0.054.54 homosex = 1.4 CFA with Missing Data using Mplus Unlike Amos and LISREL.60) 3.551 33292. R² = 0.471 Degrees of Freedom = 7 Full Information ML Chi-Square = 9.85) 3.99 The result is quite similar to that of Amos. Instead raw data must be saved as an ASCII file in free or fixed format.019.92 (P = 0. The GOVTRESP has low standardized loadings from both factors (.15 Morals -0. R² = 0. Unconstrained estimates are shown under the Measurement Equations heading and standardized solutions on the path diagram.18 for MORALS) and a very low R2 of .00*Morals.05) = 0.

The VARIABLE statement lists the names of the variables in the order they appear in the data file. The TITLE line provides a short description for the analysis. These two commands recode missing values as -1 and write a tab-delimited text file values_full. MISSING ARE ALL(-1). choose File →New → Syntax to open a SPSS Syntax Editor window. . The MODEL statement tells Mplus that there are two latent variables and six outcome (manifest) variables. The OUTPUT statement here requests that standardized estimates and modification indices appear in the output file. VARIABLE: NAMES ARE privtown govtresp compete homosex abortion euthanas.edu/~statmath . DATA: FILE IS c:\temp\cfa\values_full. A text output file appears and is saved in the To begin the estimation. The DATA statement specifies the path name for the tab-delimited raw data file to be analyzed. MODEL: economic BY privtown govtresp compete. and the h1 requests a chi-square statistic for model fit. SAVE TRANSLATE OUTFILE='c:\temp\CFA\values_full. MISSING ARE ALL(-1). and then enter the following commands.indiana. MODINDICES. morals BY homosex abortion euthanas govtresp. Now launch Mplus and type in the following syntax for the two factor model with missing data. The TYPE = MISSING option in the ANALYSIS statement tells Mplus to use an estimator appropriate for the presence of missing data. morals BY homosex abortion euthanas govtresp. RECODE privtown govtresp compete homosex abortion euthanas (SYSMIS=-1).dat. MODEL: economic BY privtown govtresp compete. SUMMARY OF ANALYSIS Number of groups Number of observations 10 1 1200 In previous edition.” which becomes a default setting in version 5. OUTPUT: STANDARDIZED. DATA: FILE IS c:\temp\cfa\values_full. The MISSING ARE option tells Mplus to interpret the numeric value -1 as missing for all variables. The output file looks like the following: INPUT INSTRUCTIONS TITLE: Two Factor Model with Missing Data. http://www.10 TITLE: Two Factor Model with Missing Data.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 63 In order to recode missing values.dat in the C:\temp\CFA directory. open SPSS. MODINDICES. VARIABLE: NAMES ARE privtown govtresp compete homosex abortion euthanas.dat' /TYPE=TAB/MAP. you need to add “ANALYSIS: TYPE=MISSING H1.dat. click on Run working directory. INPUT READING TERMINATED NORMALLY Two Factor Model with Missing Values. OUTPUT: STANDARDIZED.

999 COMPETE 0.994 0.999 0.edu/~statmath .100D-03 SUMMARY OF DATA Number of missing data patterns 12 COVARIANCE COVERAGE OF DATA Minimum covariance coverage value 0.992 0.100 PROPORTION OF DATA PRESENT Covariance Coverage PRIVTOWN GOVTRESP ________ ________ PRIVTOWN 0.988 0.990 COMPETE ________ HOMOSEX ________ ABORTION ________ 0.977 0.975 0.981 0.976 0.500D-04 20 2000 0.991 THE MODEL ESTIMATION TERMINATED NORMALLY TESTS OF MODEL FIT Chi-Square Test of Model Fit Value Degrees of Freedom P-Value 9.994 GOVTRESP 0.1932 Chi-Square Test of Model Fit for the Baseline Model Value 910.992 EUTHANAS 0.981 ABORTION 0.990 0.985 Covariance Coverage EUTHANAS ________ EUTHANAS 0.994 0.920 7 0.indiana.985 0.923 http://www.993 0.999 HOMOSEX 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 64 Number of dependent variables Number of independent variables Number of continuous latent variables Observed dependent variables Continuous PRIVTOWN 6 0 2 GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Continuous latent variables ECONOMIC MORALS Estimator Information matrix Maximum number of iterations Convergence criterion Maximum number of steepest descent iterations Maximum number of iterations for H1 Convergence criterion for H1 Input data file(s) c:\temp\cfa\values_full.982 0.dat Input data format FREE ML OBSERVED 1000 0.

© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 65 Degrees of Freedom P-Value CFI/TLI CFI TLI Loglikelihood H0 Value H1 Value Information Criteria Number of Free Parameters Akaike (AIC) Bayesian (BIC) Sample-Size Adjusted BIC (n* = (n + 2) / 24) 15 0.086 0.272 33370.000 0./S.071 54.046 0.124 999.872 COMPETE 2.000 0.002 0. P-Value 1.275 20 33332.633 0.911 3.019 0.312 3.850 0.619 0.000 0.000 0.398 0.992 0.384 999.072 0.215 0.995 HOMOSEX 5.409 11.486 3.987 SRMR (Standardized Root Mean Square Residual) Value 0.042 999.022 EUTHANAS 4.000 0.235 -16641. Probability RMSEA <= .289 0.095 0.000 0.000 Residual Variances PRIVTOWN 3.indiana.I.395 14.002 1.000 0.016 0.302 1.000 0.738 3.000 3.000 0.900 4.731 ABORTION 4.000 0.000 0.780 37.352 0.091 0.608 55.082 0.000 0.000 -0.774 4.000 0.419 0.442 4.360 2.744 RMSEA (Root Mean Square Error Of Approximation) Estimate 90 Percent C.494 50.470 33434.016 MODEL RESULTS Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH ECONOMIC Intercepts PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Variances ECONOMIC MORALS S.043 0.000 13.390 49.140 -0.212 5.541 4.041 23.000 0.029 9.645 0.120 GOVTRESP 6.902 5.001 0. Two-Tailed Est.0000 0.201 0.edu/~statmath .069 0.993 -16646.05 0.000 0.763 50.E.997 0.683 3.627 0.882 0.112 0.416 19.000 0.065 0.974 0.E.000 http://www.864 13.000 0.000 1.105 999.078 0.115 0.000 0.000 0.

044 0.000 0.580 1.080 0.000 999.166 29.042 0.154 0.000 -0.000 0.001 0.202 6.000 999.000 0.E.217 0.842 20.000 0.217 0.690 34.000 0.047 35.531 0.909 3.000 0.123 0.000 999.167 4.031 5.E.042 0.468 1.112 0.202 6.000 0.166 29.677 0.000 0.679 0.307 22.737 0.755 26.100 0.904 35.036 36.000 0.432 36.000 1.176 0.042 0.075 0.041 0.946 COMPETE 0.000 0./S.531 0.027 0. 0.E.397 14.000 0.043 0.037 0.566 0.000 http://www.437 1.000 Residual Variances PRIVTOWN 0.000 0.622 GOVTRESP 0.027 0.111 6.E.027 0.154 0.000 0.041 0.000 0.100 0.000 1.615 0.600 1.037 0.432 36.691 0.000 0.111 6.046 -0.470 1.065 58.580 1.000 0. 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 66 STANDARDIZED MODEL RESULTS STDYX Standardization Two-Tailed Est.000 999.904 35.075 0.000 0.755 26.522 HOMOSEX 0.000 0.241 0.000 0.566 0.027 0.044 0.040 0.034 24.000 999.849 5.690 34.737 0.437 1.000 0.042 0.154 0.027 0.037 0.911 1.027 0.034 24.indiana.000 0.615 0.000 0.000 0.000 -0.000 0.000 STDY Standardization Two-Tailed Est.468 1.000 0.600 1.167 4.000 0.626 11./S.000 0.849 5.241 0.691 0.043 0.000 1.457 EUTHANAS 0.911 1.000 0.046 -0.005 0.677 0.470 1.005 0.000 999. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH ECONOMIC Intercepts PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Variances ECONOMIC MORALS S.edu/~statmath .112 0. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH ECONOMIC Intercepts PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Variances ECONOMIC S.000 0.176 0.016 0.036 36.842 20.542 ABORTION 0.047 35.

355 17.001 0.522 HOMOSEX 0.542 ABORTION 0.494 50.419 0./S.031 3.390 49.000 R-SQUARE Observed Variable PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Two-Tailed S.040 0.395 14. 1. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH ECONOMIC Intercepts PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Variances ECONOMIC MORALS S.123 0.000 999.378 13.000 999.656 0.065 0.626 11.005 0.543 0.619 0.046 -0.001 0.086 0.000 0.289 0.360 2.040 0.378 0.154 0.083 3./S.E.442 4.000 999.000 0.000 0.000 0.425 0.307 22.911 3.398 0.000 0.037 0.000 0.000 999.000 0.080 0.778E-03 MODEL MODIFICATION INDICES http://www.321 QUALITY OF NUMERICAL RESULTS Condition Number for the Information Matrix (ratio of smallest to largest eigenvalue) 0.645 0.001 0.117 0.edu/~statmath .000 -0.000 0.000 STD Standardization Two-Tailed Est.608 55.112 0.000 0.E.457 EUTHANAS 0.000 Residual Variances PRIVTOWN 3.949 21. P-Value 0.780 37.000 999.774 4.000 0.000 2.002 0.000 0.078 0.000 0.882 0.E.071 54.731 ABORTION 4.000 0.057 4.000 Residual Variances PRIVTOWN 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 67 999.041 23.069 0.909 3.474 0.397 14.108 12.352 0.000 0.000 0.120 GOVTRESP 6.indiana.378 0.416 1.271 6.154 0.312 3.000 1. Est.110 0.416 19.000 0.054 0.000 MORALS 1.185 1.202 2.115 0.031 5.000 0.421 10.102 0.092 19.080 0.065 58.478 0.423 5.002 0.000 0.E.541 4.763 50.000 1.622 GOVTRESP 0.679 0.000 0.212 5.000 0.022 EUTHANAS 4.458 0.946 COMPETE 0.000 0.388 3.037 0.738 3.995 HOMOSEX 5.095 0.228 0.000 0.409 11.016 0.123 0.872 COMPETE 2.394 0.159 6.100 0.000 0.000 Estimate 0.158 0.091 0.016 0.

RMSEA.046) 9.920 .176 on ECONOMIC and MORALS.154 .68 .indiana. 4.37) 1.633 (.737 . value for printing the modification index M. Std E.019 Unstd.62 .215 (.15 . 1.633 (.997.I. respectively.P.016 and -. 10.302 (. Standard errors appear in parentheses. GOVTRESP has low standardized loadings of .042) -.566 for EUTHANAS (R2 = . b correlation between F1 (economic values) and F2 (moral values) http://www.176 (.1932 is too high to reject the null hypothesis of a good fit.027) .997 2.201 (.072) .005 b Unstd.027) .911 . The covariance and correlation between the two common factors are negligible -.042) -. indicating that the model fits the data well.C. RMSEA.154 and .302 (.677 .200 (.082) 1.200. column are large and corresponding p-values are small enough to reject the null hypothesis at the .176 -. Table 6: Two Factor Model with Missing Data (N=1. The RMSEA and CFI are .071) .154 (. only 5.30 (. F1 PRIVTOWN F1 GOVTRESP F1 COMPETE F2 HOMOSEX F2 ABORTION F2 EUTHANAS F2 GOVTRESP F1 F2 1.01b 9.005b (.082) 1. CFI Standard errors appear in parentheses a covariance.543).992 (.019 .111) .18 -. Unstandardized estimates are listed under the MODEL RESULTS heading and standardized estimates under STDYX Standardization of the STANDARDIZED MODEL RESULTS heading.046) .992 (.E. All parameters are statistically discernable from zero.00 .C.000 . test statistics under the Est.20 (.019 .369) 1.05 level. The p-value of ./S. and CFI.74 .4 percent of variance in GOVTRESP is accounted for by the two latent variables.63 (.691 .045) . The final row lists goodness of fit measures of χ2.566 (.P.034) -. No modification indices above the minimum value. Under the TESTS OF MODEL FIT heading.140) Standard .082) 1. Other factor loadings range from .042) -.615 . 1.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 68 Minimum M.02a (.027) .00 .045) .037) . DF=7) LISREL Mplus Amos Unstd.691 (.I.321) and .69 .920 with 7 degrees of freedom.99 (.000 StdYX E.21 (. Standardized estimates can be interpreted as the correlation between the latent and observed variables.997 Standard .102) 9.100) .57 .615 (. E.016a (.012a (.P. respectively.000 .C. the χ2 test yields a large value of 9.737 for ABORTION (R2 = .92 . The results are quite similar to those of Amos and LISREL.677 (.384) 1.5 Summary Table 6 compares the unstandardized and standardized estimates that each statistical software package produced.00 .00 .edu/~statmath .019 and .737 . The results are essentially identical across programs.566 .005.071) .215 (.13) Standard .

disagree. Forty observations with missing data are dropped in order to focus only on the problem of categorical outcome variables in the confirmatory factor model. CFA with Categorical indicators This section estimates a confirmatory factor model using the polychoric correlation matrix in the presence of categorical manifest variables. Latent variables are assumed to represent continuous (not categorical) constructs. Survey responses are often coded as yes/no or as scores on an ordered scale (e. Mplus.sav. Change the file type to SPSS Data File (*. In practice. agree.indiana. 5. The recoded data are stored into the SPSS file values_ord. In SAS/STAT CALIS. LISREL.200) is also used to show how pairwise deletion usesas much information in the raw data file as possible.g. choose the file values_ord. This approach works as follows: 1) thresholds are estimated by maximum likelihood. 5. which in turn is used to 3) estimate parameters through (diagonally) weighted least squares using the inverse of the asymptotic covariance matrix as the weight matrix (Muthén. In LISREL. For illustration. In the presence of categorical or ordinal data. The data will then be displayed in spreadsheet form. Mplus automatically follows above steps when the syntax includes a line identifying observed variables as categorical. http://www. MLE may not work properly. the polychoric correlation matrix and asymptotic covariance matrix is estimated and saved into a LISREL system file (. and SAS/STAT CALIS employ a multi-step method for ordinal outcome variables that analyzes a matrix of polychoric correlations rather than covariances.dsf) using PRELIS before fitting the model. those ranging from 4 to 7 were recoded as 2. When prompted from the Save As dialog box.edu/~statmath .sav (N=1. Jöreskog. Alternatively. calling for alternative estimation methods.2 CFA with Categorical Indicators Using LISREL Launch LISREL and load the data by choosing File → Import Data. navigate to the C:\temp\CFA folder. Values_full. strongly agree).1 Background The maximum likelihood estimation (MLE) approach used in the previous sections relied on the strong assumption of multivariate normality. a substantial amount of social science data is non-normal. 1990).11 11 You may explicitly define the variables as ordinal by going to Data → Define Variables. the METHOD=WLS or METHOD=DWLS option of the PROC CALIS statement tells SAS to fit the model using the WLS or DWLS method instead of MLE. enter the name values_ord in the File Name field and click Save. strongly disagree. neutral.sav) in the Open dialog box. and click Open. 2) these estimates are used to estimate a polychoric correlation matrix. 1984. and those ranging from 8 to 10 were recoded as 3. highlighting all the variables. and opening the Variable Type menu.sav (N=1.160) in the C:\temp\CFA directory.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 69 5. original responses ranging from 1 to 3 were recoded as 1. the diagonally weighted least squares (DWLS) method needs to be specified.

and click OK.dsf' Latent Variables Economic Morals Relationships privtown = 1. go to File → New. When prompted in the Save As dialog box. type in values_ord. Choose Diagonally Weighted Least Squares.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 70 To open an empty text window. Notice that Method of Estimation: Diagonally 12 Weighted Least Squares tells LISREL to fit WLS for categorical observed variables. The following display path diagrams for unstandardized (left) and standardized (right) estimates. Two factor model CFA for Ordinal Indicators SYSTEM FILE from file 'c:\Temp\CFA\values_ord. Sample Size = 1160 Two factor model CFA for Ordinal Indicators To request the weighted least squares estimator on menu. Enter the following commands and save the file. choose Standardized Solution from the Estimates drop-down menu. To view the standardized estimates. and then click Save.00*Morals abortion = Morals euthanas = Morals Path Diagram Method of Estimation: Diagonally Weighted Least Squares End of Problem Click F5 or the Run LISREL button to fit the model (alternatively click on File → Run). 12 http://www. choose SIMPLIS Project. The unstandardized solution appears in the path diagram along with 2 and RMSEA statistics.out.edu/~statmath . go to Output → SIMPLIS Outputs. Detailed information can be read from the text output file values_ordinal.00*Economic govtresp = Economic Morals compete = Economic homosex = 1. The following is selected from the output file.indiana.

062) Chi-Square Corrected for Non-Normality = 13.00 0.58 0.68 Goodness of Fit Statistics Degrees of Freedom = 7 Normal Theory Weighted Least Squares Chi-Square = 33.= 0.= 0.05 -0.61 .34 Morals 0.= 0. 20.75 (P = 0.18 0.52 -0.indiana.080) 13.93 (P = 0.39 (0.= 0.01 0.39 (0.00 0. Errorvar.32*Economic. R² = 0.56 .30*Morals.37) (0. R² = 0.078) (0.12) 3. R² = 0. R² = 0.= 0.0 .54 (0.15 compete -------homosex -------abortion -------euthanas -------- privtown govtresp compete homosex abortion euthanas 1.00 -0.11 .43 90 Percent Confidence Interval for NCP = (0.68 (0.00*Morals.02) 0.068) (0.01 (0.03 -0. Errorvar.11 1.61 govtresp = 0.055) (0. Errorvar.16 0.079) (0.13) 4.06*Morals.61 . Errorvar.02 0.00) Satorra-Bentler Scaled Chi-Square = 13.065) 4.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 71 Correlation Matrix privtown -------1. R² = 0.078) 5.60 homosex = 1.78 compete = 1.77 5.077) 12.51 1.87 abortion = 1.00 0.20) 3.42 Covariance Matrix of Independent Variables Economic -------Economic 0. Errorvar.05) 10. Errorvar.00*Economic.= 0.32 .30 Morals -------- 0.56 4.edu/~statmath .20 0.052) Estimated Non-centrality Parameter (NCP) = 6.15 0.43 (P = 0.08 govtresp -------1.54 .51 13.00 Two factor model CFA for Ordinal Indicators Number of Iterations = 6 LISREL Estimates (Robust Diagonally Weighted Least Squares) Measurement Equations privtown = 1.010 http://www.89 (0.89) Minimum Fit Function Value = 0.58 (0. R² = 0.39 (0.46 (0.88*Morals.47 1.38*Economic + 0.91 euthanas = 0.93 7.56 1.00 0.

which is not statistically discernable from zero.00 Adjusted Goodness of Fit Index (AGFI) = 0.99 Parsimony Goodness of Fit Index (PGFI) = 0.018) Root Mean Square Error of Approximation (RMSEA) = 0.028 and 1.18 Normed Fit Index (NFI) = 0.032 Standardized RMR = 0.25 Chi-Square for Independence Model with 15 Degrees of Freedom = 1440.45 Model AIC = 41.01. Under the Measurement Equations heading appear the unstandardized estimates. The standardized estimates of the loadings displayed in the path diagram range from .036 90 Percent Confidence Interval for ECVI = (0.062.048) ECVI for Saturated Model = 0. The Satorra-Bentler Scaled χ2 statistic has a large value of 13.22 (MORALS) and .16 This model appears to fit the data moderately. All of the unconstrained path coefficients are large enough to be statistically significant.00. http://www.036 ECVI for Independence Model = 1. RMSEA and CFI are respectively .68 (COMPETE).051) P-Value for Test of Close Fit (RMSEA < 0.51 The Modification Indices Suggest to Add an Error Covariance Between and Decrease in Chi-Square New Estimate euthanas compete 8.11).6 -3.00 Independence CAIC = 1488. and R2 statistics.33 The Path to homosex euthanas Modification Indices Suggest to Add the from Decrease in Chi-Square New Estimate Economic 81.05) = 0.indiana.0055 90 Percent Confidence Interval for F0 = (0.43 Saturated AIC = 42. However.22 Saturated CAIC = 169. indicating a very good model fit.63 for PRIVTOWN and .98 Critical N (CN) = 1595.99 Non-Normed Fit Index (NNFI) = 0. we do not reject the null hypothesis of a good fit marginally at the .43 with seven degrees of freedom.79 Model CAIC = 126.032 Goodness of Fit Index (GFI) = 1.00 Relative Fit Index (RFI) = 0.028 90 Percent Confidence Interval for RMSEA = (0. But its corresponding p-value is .39 (PRIVTOWN) to .94 Expected Cross-Validation Index (ECVI) = 0.00 Incremental Fit Index (IFI) = 1.030 .edu/~statmath .43 Root Mean Square Residual (RMR) = 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 72 Population Discrepancy Function Value (F0) = 0.24 (ECONOMIC) for GOVTRESP to . the covariance and correlation between the two latent variables are .40 Economic 21.83 for COMPETE. 0.0 .45 Independence AIC = 1452.0 . The other observed variables have moderate to high R2 statistics. This result a bit differs from one in previous models. standard errors. Finally. GOVTRESP has the smallest R2 (.46 Comparative Fit Index (CFI) = 1.2 0. 0.05 level. 0. ranging from .3 0. t-values.99 Parsimony Normed Fit Index (PNFI) = 0. Despite receiving a path from both ECONOMIC and MORALS.

CATEGORICAL ARE privtown govtresp compete homosex abortion euthanas.4 for syntax to translate an SPSS file to an ASCII file. TITLE: Two factor model CFA with ordinal indicators (listwise). When categorical indicators are declared. See sections 4. VARIABLE: NAMES ARE privtown govtresp compete homosex abortion euthanas. The syntax is similar to previous sections except that CATEGORICAL ARE is added to the VARIABLE statement to define the observed variables as categorical.3 CFA with Categorical Indicators using Mplus Mplus cannot directly read data from an SPSS system file. The following commands are used to estimate the confirmatory factor model with ordinal observed variables. Mplus employs by default a robust weighted least squares estimator to similar to the Diagonally Weighted Least Squares estimator in LISREL. the syntax editor appears. Raw data must come from a free or fixed format text file.dat in the C:\temp\CFA folder. A selection of the output file is the following: SUMMARY OF ANALYSIS Number of groups Number of observations Number of dependent variables Number of independent variables Number of continuous latent variables Observed dependent variables Binary and ordered categorical (ordinal) PRIVTOWN GOVTRESP COMPETE HOMOSEX Continuous latent variables ECONOMIC MORALS 1 1160 6 0 2 ABORTION EUTHANAS Estimator Maximum number of iterations Convergence criterion Maximum number of steepest descent iterations Parameterization Input data file(s) c:\Temp\CFA\values_ord.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 73 5.indiana.edu/~statmath . Click on Run carry out the estimation. OUTPUT: Standardized. MODEL: economic BY privtown govtresp compete. DATA: FILE IS c:\Temp\CFA\values_ord.dat Input data format FREE WLSMV 1000 0. Mplus by default assumes all observed variables are continuous unless otherwise specified.500D-04 20 DELTA SUMMARY OF CATEGORICAL DATA PROPORTIONS … http://www. morals BY homosex abortion euthanas govtresp. For this example the data is saved as the tab-delimited file values_ord.dat. After launching Mplus.

edu/~statmath .304 0.000 0.166 3.statmodel. MLR. ** Chi-Square Test of Model Fit for the Baseline Model Value Degrees of Freedom P-Value CFI/TLI CFI TLI Number of Free Parameters 0.671 MODEL RESULTS Two-Tailed Est.045 3.000 5.394 0.0000 RMSEA (Root Mean Square Error Of Approximation) Estimate 0.055 999.000 0.000 0.113 0.040 5.375 1.0166 The chi-square value for MLM.000 0.949 999.543 0. ULSMV and WLSMV are estimated according to a formula given in the Mplus Technical Appendices at www.058 0.071 0.000 0.000 14.000 0.000 STANDARDIZED MODEL RESULTS http://www. 1.991 0.E.878 0.254 0.105 0.indiana. MLR and WLSM chi-square difference testing is described in the Mplus Technical Appendices at www. MLM.037 WRMR (Weighted Root Mean Square Residual) Value 0.059 0. ULSMV.com./S.332 999.313 0.000 0.E.000 0.516* 6** 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 74 THE MODEL ESTIMATION TERMINATED NORMALLY TESTS OF MODEL FIT Chi-Square Test of Model Fit Value Degrees of Freedom P-Value * 15. See chi-square difference testing in the index of the Mplus User's Guide. MLMV.754 11 0.com.000 0.006 0.000 1.022 0.983 20 1054.073 0.statmodel.765 12. The degrees of freedom for MLMV.800 0.534 999.866 15.000 1. See degrees of freedom in the index of the Mplus User's Guide. WLSM and WLSMV cannot be used for chi-square difference tests. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH ECONOMIC Thresholds … Variances ECONOMIC MORALS S.000 0.

254 0./S.000 0.000 0.030 0.647 0.E.628 0.000 999.048 0.254 0.083 0.E.799 1.041 0.654 0.226 25.edu/~statmath .224 0.030 0.000 999.825 0.679 5./S.000 0.647 0.000 0.000 0. 0. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH ECONOMIC Thresholds … Variances ECONOMIC MORALS S.788 7.825 0.000 999.000 0.000 1.647 18.000 http://www.737 0.030 0./S.048 0.679 5.083 0.236 0.030 0.000 999.000 0.000 0.000 999.040 24.000 STDY Standardization Two-Tailed Est.224 0. 0.936 0.654 0.000 0.737 0.000 1.628 0.040 24.647 0.530 5.indiana.737 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 75 STDYX Standardization Two-Tailed Est.000 0.788 7.000 0.825 0.236 0.788 7.000 0.799 1. 0.936 0.000 0.E.224 0.679 5.000 999.000 999.780 0.012 0.000 0.000 0.104 7.035 0.236 0.000 0.104 7.647 18.E.000 0.030 0.226 25.083 0.000 999.000 STD Standardization Two-Tailed Est.030 0.E.012 0.654 0.780 0.041 0.780 0.104 7.628 0.035 0.000 0.000 0.000 0.000 0.936 0.000 0.530 5.041 0. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH ECONOMIC Thresholds … Variances ECONOMIC MORALS S. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH S.E.226 25.040 24.647 18.000 0.000 0.530 5.035 0.

/S.765 4. for COMPETE it is .05 level. GOVTRESP has a low squared multiple correlation of .000 R-SQUARE Observed Variable PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Two-Tailed Residual S. and . and p-values.647. and for EUTHANAS it is .392 0.045 0. for HOMOSEX it is . .893 0.0166.418 QUALITY OF NUMERICAL RESULTS Condition Number for the Information Matrix (ratio of smallest to largest eigenvalue) 0.006 and . column are greater than 1.680 0.indiana.000 0. for ABORTION it is .824 9.825.160.418 (EUTHANAS).224 for MORALS).105 0.047 0.05 significance level.113 12.000 1.608 (ABORTION).000 999. which is small enough to reject the null hypothesis of a good fit at the .107. Standardized estimates are provided under the STDYX Standardization heading.628. reducing the original sample of 1.000 0. which is one smaller than those reported by LISREL. P-Value Variance 0.000 999.606 0.608 0. a small RMSEA of .680 (COMPETE).991 suggest that the model fits the data well as a whole.394 0. Mplus estimates 6 degrees of freedom.582 Estimate 0.254 1. .037 and a large CFI of . the unstandardized loadings appear along with standard errors.4 CFA with categorical indicators and missing data The previous two subsections explained how to estimate the confirmatory factor model when the observed variables represent ordered categories.780. . the ratio of the estimates to their standard errors.339 0. GOVTRESP has relatively low standardized loadings on both factors (. but a corresponding p-value is .236 for ECONOMIC and ./S. Under the MODEL RESULTS heading. However.107 0.012.000 0.012 0.E.200 3.543 (HOMOSEX).000 999. in order of increasing magnitude.457 0. It is possible to maximize the information available in the raw data http://www. A total of 40 observations with missing values on at least one indicator were dropped in the listwise deletion.000 0.045 3.E.543 0. As explained in the output. .737.96.394 (PRIVTOWN).516.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 76 0.320 0.edu/~statmath .025 0. The remaining R2 statistics are.200 to 1.103E-02 Both LISREL and Mplus (listwise deletion) produce virtually identical results.000 0. the covariance and correlation between the two common factors are . respectively.000 0.000 0.000 999.048 0. 5. Finally.171 0. For PRIVTOWN the loading is .799 ECONOMIC Thresholds … Variances ECONOMIC MORALS 0.E. Since all absolute values of the numbers in the Est.000 0. all of the unconstrained loadings estimates are significant at the . The χ2 test yields a value of 15. Est.968 12.

TITLE: DATA: VARIABLE: Two factor model CFA with ordinal indicators (pairwise). In pairwise deletion. FILE IS c:\Temp\CFA\values_full. morals BY homosex abortion euthanas govtresp.dat.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 77 file using pairwise rather than listwise deletion for missing data. economic BY privtown govtresp compete. The following Mplus syntax use values_full. SUMMARY OF ANALYSIS Number of groups Number of observations Number of dependent variables Number of independent variables Number of continuous latent variables Observed dependent variables Binary and ordered categorical (ordinal) PRIVTOWN GOVTRESP COMPETE HOMOSEX Continuous latent variables ECONOMIC MORALS 1 1200 6 0 2 ABORTION EUTHANAS Estimator Maximum number of iterations Convergence criterion Maximum number of steepest descent iterations Maximum number of iterations for H1 Convergence criterion for H1 Parameterization Input data file(s) c:\Temp\CFA\values_full. correlations will be estimated using all observations with complete data available on both variables. MISSING ARE all (-1).indiana. NAMES ARE privtown govtresp compete homosex abortion euthanas.100 COMPETE HOMOSEX ABORTION http://www.dat.100D-03 DELTA SUMMARY OF DATA Number of missing data patterns 12 COVARIANCE COVERAGE OF DATA Minimum covariance coverage value PROPORTION OF DATA PRESENT Covariance Coverage PRIVTOWN GOVTRESP 0. CATEGORICAL ARE privtown govtresp compete homosex abortion euthanas.dat Input data format FREE WLSMV 1000 0.dat instead of values_ord.500D-04 20 2000 0. Standardized.edu/~statmath . The following is a selection of Mplus output. MODEL: OUTPUT: Notice that MISSING ARE all (-1) is added to the VARIABLE command in order to use pairwise deletion.

See degrees of freedom in the index of the Mplus User's Guide.994 0.0000 RMSEA (Root Mean Square Error Of Approximation) Estimate 0.051 0.000 0.380 1.258 999.000 1. Two-Tailed Est.993 0.E.975 0.999 0. The degrees of freedom for MLMV.245 999.684 999.992 0.994 0. ** Chi-Square Test of Model Fit for the Baseline Model Value Degrees of Freedom P-Value CFI/TLI CFI TLI Number of Free Parameters 0.995 62 1528./S.1081 The chi-square value for MLM.994 0.999 0.025 WRMR (Weighted Root Mean Square Residual) Value 0. See chi-square difference testing in the index of the Mplus User's Guide.982 0. MLR and WLSM chi-square difference testing is described in the Mplus Technical Appendices at www. ULSMV. ULSMV and WLSMV are estimated according to a formula given in the Mplus Technical Appendices at www.991 THE MODEL ESTIMATION TERMINATED NORMALLY TESTS OF MODEL FIT Chi-Square Test of Model Fit Value Degrees of Freedom P-Value * 10.045 999.336 MODEL RESULTS Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS S.indiana.000 0.000 0. MLR.997 0.981 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 78 ________ ________ PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS ________ 0. WLSM and WLSMV cannot be used for chi-square difference tests.977 0.000 7.066 0.990 ________ 0.edu/~statmath .352 0.976 0.000 http://www.statmodel.615 10 0.statmodel. MLM. MLMV.com.000 0.com.417* 6** 0.239 20.000 0.000 1.985 Covariance Coverage EUTHANAS ________ EUTHANAS 0.988 0.000 20.992 0.416 5.053 0.930 0.E.999 0.990 0.981 0.000 0. P-Value 1.000 0.985 ________ 0.

Other factor loadings range from .033 0. Est.000 0.000 999.520 0.423 0.595 (PRIVTOWN) to .E.196 0.997 also suggest that the model fits the data well. R2 ranges from .214 25.000 1.352 0.670 0.000 STANDARDIZED MODEL RESULTS STDYX Standardization Two-Tailed Est.000 0. All unstandardized and standardized estimates have large test statistics and appear to be statistically significant.125 0.077 10.036 5.000 0. A small RMSEA of .000 0./S.449 This CFA model fits the data quite well.409 0.000 0.000 0.000 0. P-Value Variance 0.218 6.000 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 79 0.591 0.394 0. 0.607 33.417 with 6 degrees of freedom and its p-value of . GOVTRESP has lowest factor loadings of .207 0.000 0.000 0.033 5.025 and a large CFI of .803 0.769 0.E.036 0.721 0.272 0.520 0.023 0.000 0.060 5.000 R-SQUARE Observed Variable PRIVTOWN GOVTRESP COMPETE HOMOSEX ABORTION EUTHANAS Two-Tailed Residual S.031 31.595 0.1081 do not reject the null hypothesis of a good fit.044 6.000 0.000 0. Accordingly.033 0.241 5.000 999.607 12.197 15.093 0.edu/~statmath .000 999./S.811 0.805 (COMPETE).551 Estimate 0.801 0.070 0.805 0.000 999.029 0.093 for GOVTRESP to .018 0. The χ2 of 10.000 0.226 0.803 16.023 0.480 0.000 0.354 0.014 0.226 from economic values and .indiana.070 0.196 for moral values.059 0.646 0.E.060 15.000 GOVTRESP MORALS WITH ECONOMIC Thresholds … Variances ECONOMIC MORALS 0.907 0.120 7.236 0.818 10.000 0.648 0.027 0.E.417 1.040 0.354 for PRIVTOWN http://www.354 0. P-Value Estimate ECONOMIC BY PRIVTOWN GOVTRESP COMPETE MORALS BY HOMOSEX ABORTION EUTHANAS GOVTRESP MORALS WITH ECONOMIC Thresholds … Variances ECONOMIC MORALS S.609 0.018 0.

compete = lambda3 f1 + e3.89216 13 METHOD=DWLS uses diagonally weighted least squares method but.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 80 and . http://www. PROC CALIS DATA=sas. SAS reports 21 knowns provided by the input covariance matrix and 14 unknown parameters to be estimated. RUN.033.sas7bdat) as a file type.648 for COMPETE.edu/~statmath .0 f2 + e4.55257 Kurtosis -0. homosex = 1. degrees of freedom are 7 = 21 14.47100 -0. The CALIS Procedure Covariance Structure Analysis: Weighted Least-Squares Estimation Observations Variables Informations 1160 6 21 Model Terms Model Matrices Parameters 1 4 14 Variable privtown govtresp Mean 1. The following is the SAS script for this two factor CFA model with ordinal data. STD f1-f2 = phi1-phi2. failed to report goodness of fit measures in this example.70316 0. In SPSS. despite successful convergence. COV f1-f2 = phi3. 5. The following is a selection of the SAS output. e1-e6 = theta1-theta6.52672 1.5 CFA with categorical indicators Using SAS/STAT CALIS SAS/STAT CALIS can also fit a confirmatory factor model with categorical indicators using (diagonally) weighted least squares estimation. Basic statistics and correlation matrix display first. euthanas = lambda6 f2 + e6.values_ord METHOD=WLS PALL. LINEQS privtown = 1.0 f1 + e1.13 Other statements for specification remain unchanged. ODS HTML CLOSE. ODS HTML.indiana.sav into the SAS data set by choosing Save As and select SAS v7+Windows long extension (*.014 and .71347 Skewness 0.60852 0. Notice that METHOD=WLS in the PROC CALIS statement tells SAS to employ the WLS method instead of the default MLE. you may convert the data set values_ord. govtresp = lambda2 f1 + lambda7 f2 + e2. VAR privtown govtresp compete homosex abortion euthanas. respectively. The covariance and correlation between two latent variables are .68103 Std Dev 0. abortion = lambda5 f2 + e5.

edu/~statmath .0489 0.0218 -0. 0.31825 -1.3917 0.0218 1.4667 13.9462 0.0000 0.0356 .2625 7 0.00 15 0.1051 1.9905 -0.0390 -0.1205 0.9999 0.0278 .0000 0.0000 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 81 Variable compete homosex abortion euthanas Mean 1.0057 0.0909 compete 0.49138 1.74246 0.3917 -0. 1989) Chi-Square Chi-Square DF Pr > Chi-Square Independence Model Chi-Square Independence Model Chi-Square DF RMSEA Estimate RMSEA 90% Lower Confidence Limit RMSEA 90% Upper Confidence Limit ECVI Estimate ECVI 90% Lower Confidence Limit ECVI 90% Upper Confidence Limit Probability of Close Fit Bentler's Comparative Fit Index Akaike's Information Criterion 0.45744 1.4670 1.0660 674.92321 0.4670 0.0057 0.74052 1.26214 0.0390 0.0489 govtresp 0.0909 0.0000 Fit Function Goodness of Fit Index (GFI) GFI Adjusted for Degrees of Freedom (AGFI) Root Mean Square Residual (RMR) Parsimonious GFI (Mulaik.78108 0.3604 1.85345 1.0000 -0.7375 http://www.0220 0.1205 -0.0114 1.60023 Skewness 0.1421 0.21119 -1.0000 0.1421 -0.0689 0. 0.1315 0.07136 1.62442 Kurtosis -0.0123 0.33276 Std Dev 0.49187 Correlations privtown privtown govtresp compete homosex abortion euthanas 1.0504 0.63267 0.indiana.3305 abortion -0.0689 homosex -0.1315 1.1051 0.3305 0.0000 0.0480 0.3604 euthanas 0.0123 0.

& Brett (1982) Parsimonious NFI Z-Test of Wilson & Hilferty (1931) Bollen (1986) Normed Index Rho1 Bollen (1988) Non-normed Index Delta2 Hoelter's (1983) Critical N Manifest Variable Equations with Estimates privtown = 1.92369 0.3369 3.7616 Std Err t Value 0.9796 0.02047 0.51219 0.09545 0.3441 = 1.9906 1231 + 0.1308 -36.9578 0.35 10.13081 0.4575 1.03423 t Value f1 f2 e1 e2 e3 e4 e5 e6 phi1 phi2 theta1 theta2 theta3 theta4 theta5 theta6 0.indiana.0367 Std Err t Value 0. Mulaik.1308 0.09550 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 82 Bozdogan's (1987) CAIC Schwarz's Bayesian Criterion McDonald's (1989) Centrality Bentler & Bonett's (1980) Non-normed Index Bentler & Bonett's (1980) NFI James.0000 e6 Variances of Exogenous Variables Variable Parameter Estimate Standard Error 0.2040 0.0000 e2 abortion = 1.9803 0.0546 5.0831 4.04374 0.18 21.10 12.45389 0.0811 12.5106 0.0658 + 1.48 11.12 4.53632 0.04364 0.0000 f2 * f2 lambda5 + 1.0000 + 1.5742 = 1.2768 * f2 0.68016 0.0000 e4 e5 * f1 lambda3 f1 * f1 lambda2 + 1.0000 govtresp = 0.40 7.9973 0.3609 Std Err t Value compete Std Err t Value homosex 0.73673 3.52 http://www.04580 0.0000 e1 -43.31984 0.0633 12.12 45.54611 0.0000 e3 lambda7 + 1.7824 euthanas = 0.edu/~statmath .0318 * f2 lambda6 + 1.

2041 for economic values and .8583 e6 Squared Multiple Correlations Variable 1 2 3 4 5 6 privtown govtresp compete homosex abortion euthanas Error Variance 0.0006 and http://www.2633 for EUTHANAS and .3198 0.7390 + 0.2633 Correlations Among Exogenous Variables Var1 f1 Var2 f2 Parameter phi3 Estimate -0.0763 for GOVTRESP to . suggest that this confirmatory factor model fits the date well.00159 The results are a bit different from those of LISREL and Mplus.edu/~statmath .4539 0.2625 with 7 degrees of freedom and a corresponding p-value is .51219 0.03 Manifest Variable Equations with Standardized Estimates privtown = 0. Standardized estimates display under Manifest Variable Equations with Standardize d Estimates Factor loadings for GOVTRESP are . This result is similar to what LISREL produced in 5.6809 * f1 lambda3 homosex = 0.0763 0.01846 t Value f1 f2 phi3 -0.00000 1. Squared multiple correlations range from .8247 e1 + 0.1865 * f2 lambda7 + 0.indiana.2041 * f1 lambda2 compete = 0.7323 e3 + 0.0006067 -0. The RMSEA of . however.9611 e2 govtresp = 0.0278 and CFI of .6984 * f2 lambda5 euthanas = 0. The covariance and correlation between two latent variables are -.00000 R-Square 0.6737 f2 + 0.73673 Total Variance 1.92369 0.00000 1.0660 that indicates a moderate fit.68016 0.6984 for ABORTION.1865 for moral values.00000 1.5131 for EUTHANAS to .5131 * f2 lambda6 + 0. Unstandardized estimates under the Manifest Variable Equations with Estimates heading have large t statistics and appear to be statistically discernable from zero.9905.00000 1.00000 1.4637 0.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 83 Covariances Among Exogenous Variables Var1 Var2 Parameter Estimate Standard Error 0. The χ2 is 13.5655 f1 + 0.54611 0.53632 0.4878 0. Other factor loadings range from .4878 for ABORTION.7157 e4 e5 abortion = 0.2.

indiana.991 Standard .037 .023) .6809 .236 (.2040 (.2768 (.012b (.88 (.160) Mplus (pairwise) Unstd.028 1. RMSEA.780 (. are relatively consistent regardless of models estimated by three software packages.22 .02) (N=1.00 1. Mplus.041) .0185) (N=1. Unlike LISREL and Mplus.048) 6 15.029) . 1.edu/~statmath .05 significance level.023) .160) Standard .24 .304 (.78 .272 (.071) .068) .37) 1.027) .32 (.380 (.059 (.670 (. CFA model with listwise deletion rejects the null hypothesis at the . LISREL and SAS/STAT CALIS report 7 degrees of freedom whereas Mplus returns 6.805 (.006a (.035) . while CFA with pairwise deletion does not.997 Standard .0016.00 .077) . RMSEA and CFI.079) 1.033b (. Notice that Mplus uses a different formula to calculate the degrees of freedom.053) .9905 Standard errors appear in parentheses a covariance.721 (.030) .258) 1.878 (. SAS/STAT CALIS reports goodness of fit measures similar to those LISREL produces.083) .022) (N=1.38 (.000 .014a (.3609 (.01b 7 13.01a (.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 84 -.769 (.352 (. and SAS/STAT CALIS.595 (.031) . LISREL and SAS/STAT CALIS reports a χ2 of about 13.040) .332) 1. 5. SAS/STAT CALIS reports negative covariance and correlation although their magnitudes are virtually zero.06 and thus do not reject the null hypothesis of a good fit at the .045) .647 (.63 .0006a (. The pairwise column should be considered more accurate because it is able to incorporate the most information from the raw data.226 (.040) 6 10.058) .5655 .65 .0367 (.74 .6737 .066 (.3369) 1.073) 1.737 (.6984 . In Mplus.83 .930 (.200) SAS CALIS Unstd. Mplus also supports pairwise deletion analyses using listwise deletion plus the Mplus results using pairwise deletion that produces slightly different parameter estimates and standard errors.2041 . suggesting that all models fit the data well.0546) -.0278 .825 (.160) Standard .200 observations without dropping 40 incomplete observations.055) .0000 1. which is different from 10 and 16 in Mplus.104) .0016b 7 13.030) .0811) .000 1.93 .5131 .00 1.1865 -. 1.6 Summary Table 7 summarizes results from the LISREL.224 (. F1 PRIVTOWN F1 GOVTRESP F1 COMPETE F2 HOMOSEX F2 ABORTION F2 EUTHANAS F2 GOVTRESP F1 F2 Degrees of freedom 2.30 (.417 . LISREL and SAS/STAT CALIS have a marginal p-value of about .313 (. LISREL and Mplus with listwise deletion produces equal parameter estimates but Mplus reports smaller standard errors.0000 .2625 .05 level. but unstandardized and standardized estimates of two software packages are slightly different.7616 (. CFI 1.196 (. b correlation between F1 (economic values) and F2 (moral values) http://www.018) (N=1.055) .0831) 1.0633) .00 . however.059) .628 (.375 (. Table 7: Two Factor Model with Ordinal Indicators LISREL (listwise) Mplus (listwise) Unstd. 1. respectively.078) .044) .06 (.516 . notice that pairwise deletion uses 1.025 .051) 1.00 Unstd.

and faculty may schedule an appointment with a consultant at the UITS Stat/Math Center by calling 5-4724 or emailing statmath@indiana. Mplus is recommended for its simplicity of syntax. LISREL. staff. CFA is the basis of the measurement model in full structural equation modeling (SEM) and can be estimated using SEM software. is a technique to identify structure in data and generating hypotheses without imposing any restrictions. LISREL. although Mplus syntax is not at all difficult to learn. although Mplus can also estimate models in which the latent variables are assumed to be categorical. Mplus. Amos and LISREL are the most user-friendly. though they are not responsible for any errors that remain. http://www. Amos and LISREL can read data files generated by a variety of different software packages and can estimate models by simply drawing a path diagram. although rich enough. All four software packages handle models assuming the latent variable to be continuous.edu/~statmath . SAS/STAT CALIS has the systematic grammar structure in the integrated environment. Consult the documentation for the respective package for additional information on Amos. is not easy to navigate.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 85 6. informative output. Amos cannot accurately estimate models when the observed variables are categorical. and SAS/STAT CALIS can handle ordinal observed variables using (diagonally) weighted least squares estimation methods. By contrast.indiana. Additionally IU students. Acknowledgements Amy Drayton and Takuya Noguchi offered helpful suggestions to make this document more informative and readable.edu. Exploratory factor analysis (EFA). Amos Graphics and Program Editor work independently and the output. and support for various advanced models although it has poor data management capability and does not support a path diagram. and Mplus. but it has messy output and does not support advanced models with missing values. Conclusion Factor analysis is a widely used method for situations in which a small set of unobserved (latent) variables is believed to underlie a larger set of observed (manifest) variables. LISREL is generally suggested to take advantage of switching back and forth from a path diagram to SIMPLIS and LISREL. EFA differ from confirmatory factor analysis (CFA) in that CFA is much more theory driven (as opposed to data driven) and is generally used to test explicit hypotheses. available in most general statistics packages.

objectively determined and measured. K. and Long. Beverly Hills. and Huber. Jöreskog. Structural Equation Models with Latent Variables. Spearman. Gabel.right ideological positions from party manifestos data.7. L. ordered. Mplus. 8. CA: Sage. Enders.G. Amos16. (1904).J. Hatcher.A. 183-202. Bollen. McDonald.© 2006-2009 The Trustees of Indiana University Confirmatory Factor Analysis 86 References Akaike. Structural Equation Modeling.” Psychometrika.M.” British Journal of Mathematical and Statistical Psychology. (2004). 461-464. B.” Pychometrika. Ann Arbor. “Cutoff criteria for fit indexes in covariance structure analysis: conventional criteria versus new alternatives. 49. 387-404. Muthén.G. B. H.A. 128-141. 2005-06-15. (1994). “A general structural equation model with dichotomous. (1979). C. 44. Chicago. R. H.O. (2006). J. (1987). Larry. 370-390. Step-by-Step Approach to Using the SAS System for Factor Analysis and Structural Equation Modeling.0 User’s Guide.D. J. K. (1980). Jöreskog.” Paper presented at the Annual Spring Meeting of the Psychometric Society. and Sörbom. http://www. (2000). Bollen.G.K. Inc. “Issues in the comparative measurement of political democracy” American Sociological Review. A. 45. 37. 52. (1993). Scientific Software International. G. New York: Wiley & Sons. Steiger. 234-251. “Statistically-based tests for the number of common factors. 115132. D. K. 317-332. (1969). Jöreskog.” Quality and Quantity 24.” American Journal of Psychology. K. C. Bollen. (1978). and Lind. eds. Cary. M. E. J. Iowa City.” The Annals of Statistics. 6. R. and continuous latent variable indicators.A. categorical. ICPSR3975 [computer file].O.” American Journal of Political Science. “A primer on maximum likelihood algorithms for use with missing data. J. T. J. (1978).” Psychometrika. (1984). CA: Sage. 94-103. Reliability and Validity Assessment. “Estimating the dimension of a model. Testing Structural Equation Models.indiana. “A general approach to confirmatory factor analysis. MI: Inter-university Consortium for Political and Social Research [distributor]. P. European and World Values Surveys Integrated Data File. Muthén.G. 1999-2002 Release I.P. and Muthén. NC: SAS Institute. K. “Factor analysis and AIC. Arbuckle. (2007). and Zeller. 201-293. European Values Study Group and World Values Survey Association. Schwarz. (2006). “New developments in LISREL: Analysis of ordinal variables using polychoric correlations and weighted least squares. “General intelligence. Carmines.. Confirmatory Factor Analysis for Applied Research. Brown.edu/~statmath . (1999). IL: SPSS Inc. 15. L.K.A. New York: Guilford Press. (1989). 1-55. 6. (1980). and Bentler. 34. (1990). Los Angeles: Muthén and Muthén.L.S. “Putting parties in their place: Inferring party left. K. (2005). (2001). “A simple comprehensive model for the analysis of covariance structures. LISREL 8. Hu.” Structural Equation Modeling. Newbury Park.

Sign up to vote on this title

UsefulNot useful- Using Amos for Structural Equation Modeling in Market Research
- Structural_Equation_Modeling_with_AMOS
- 985 Structural Equation Modeling in Educational Research
- Structural Equation Modelling (SEM) Part 2 of 3
- AMOS Tutorial
- Barbara Byrne Structural Equation Modeling With AMOS Basic Concepts, Applications, And Programming Multivariate Applications, 2nd Edition 2009
- Sem
- Factor
- Handout StructuralEquationModelling
- Mplus Users Guide v6
- Structural Equation Modelling
- Structural Equation Modeling
- LISREL for Dummies English
- Structural Equation Modelling and its Application
- A General Model for Testing Mediation and Moderation Effects
- confirmatory factor analysis
- 2011 an Introduction to Confirmatory Factor Analysis CFA
- Structural equation modeling
- Structural equation modeling using R
- [Joseph F. Hair]a Primer on Partial Least Squares Structural Equation Modeling (PLS-SEM)(PDF){Zzzzz}
- Partial Least Squares Regression (PLSR)
- Introduction to Path Analysis and SEM with AMOS
- SmartPLS Utilities
- Basics of Structural Equation Modeling
- Manual Smart Pls
- DETERMINING SAMPLE SIZE FOR RESEARCH ACTIVITIES
- IBM SPSS Amos User's Guide
- SEM Manual
- Applied Multivariate Statistical Analysis 6th Edition
- Confirmatory Factor Analysis Using Amos, LISREL, Mplus,