e1

e2

e3

e4

CONFIRMATORY FACTOR ANALYSIS: ONE FACTOR MODELS

Conceptual Nature of Latent Variables

Latent variables correspond to some type of hypothetical construct

measured directly

Require a specific operational definition Indicators of the construct need to be selected Data from the indicators must be consistent with certain predictions (e.g., moderately correlated with one another)

Why use latent variables?

Measurement error is inevitable and we must deal with it.

Advantageous because latent variables are less influenced by measurement error.

Latent variables are the real variables, not measured variables. Cannot use factor scores because they still have measurement error. Ability to use latent variables is the primary strength of SEM

Multi-Indicator Approach

Reduces the overall effect of measurement error of any individual observed variable on the accuracy of the results We distinguish between observed variables (indicators) and underlying latent variables or factors (constructs) measurement model: observed variables and the latent variables

Error Var1
1

Error Var2
1

Error Var3
1

Error VT
1

Error AT
1

Error ST
1

Observed Observed Observed Variable 1 Variable 2 Variable 3

Vocabulary Analogies Test Test

Writing Sample

Latent Variable

Verbal Abilities

Construct of interest

Exploratory Factor Analysis (EFA)

EFA also has to do with the latent structure of a set of variables

In EFA, no a priori specification of how many latent factors or about how measures relate to these factors

Exploratory Factor Analysis (EFA)

Problems
Communality:

must know the communality before estimation, but communality is a function of the loadings Number of factors Rotation: when there are two or more factors, the solution is not unique

Exploratory Factor Analysis (EFA)

Principal Components
Communality

is set to 1 Factor is defined as the sum of the variables Loadings chose to maximize the explanation of the variances of the measures Loadings are usually too high in that the predicted correlations are larger than the observed correlations

Exploratory Factor Analysis (EFA)

Principal Factors
Communality

iteratively estimated Factor is a latent variable Loadings chose to:

maximize the explanation of the correlations between the measures. minimizes the sum of squared residuals residual = observed correlation minus predicted correlation

Exploratory Factor Analysis (EFA)

Maximum Likelihood
Solution

is iteratively estimated Factor is a latent variable Loadings chosen to maximize the explanation of the correlations between the measures tries harder to explain the larger correlations Statistical tests available

Exploratory Factor Analysis (EFA)

Example: [Link]
Principal Axis Maximum

Variable Principal Components Likelihood

Overall

.905

.877

.816

Clear
Color Odor

.865
.921 .805

.811
.918 .710

.862
.938 .684

Note that Principal Components loadings generally larger than the other methods. PA and ML fairly similar.

EFA vs. CFA

EFA is useful when the researcher does not know how many factors there are or when it is uncertain what measures load on what factors EFA typically used as a data reduction strategy Both EFA and CFA reduce a larger number of observed variables into a smaller number of latent factors However, EFA is done with little a priori hypothesis; CFA requires a priori specification based on hypothesis

Assumptions of CFA

Multivariate normality Sufficient sample size Correct model specification Sampling Assumptions- Simple random sample

Confirmatory Factor Analysis

Representation in SEM

Latent variable represented by a circle Measured variables (indicators) represented by a square Each indicator variable has an error term

Error Var1
1

Error Var2
1

Error Var3
1

CFA Initial Specification

Observed Observed Observed Variable 1 Variable 2 Variable 3

Each variable loads on one and only one factor Factors can (and typically are) correlated Errors across indicator variables are independent
The factors are uncorrelated with the measurement errors Most (if not all) of the errors of different indicators are uncorrelated with each other

Assumptions

Latent Variable

Residuals In CFA

Item level residuals are represented as latent variables. They are not called disturbances. They represent measurement error in EFA/CTT sense. This is a tremendous advantage in hybrid models, which combine CFA and path models, because it separates measurement error from error in the model.

Errors vs. Disturbances

Are both residuals. Both necessary in their respective roles. Errors always represent measurement error. Disturbances
Represent

omitted variables (in hybrid model). If no error terms, measurement error will be in disturbance (in path model).

Standard Specification of CFA

One indicator treated as a marker or reference variable (Brown p. 61, Kline p. 170): Its loading is fixed to one Which variable should you choose?
closest in meaning to the factor most interpretable units of measurement empirical: strongest correlations with other indicators No test of statistical significance

Factor variance is freely estimated Error paths are set to one Error variances are freely estimated

Standard Specification
e1 e2 e3 e4

Identification

Identification in CFA is largely determined by the number of indicator variables used in the model (more later). Number of indicators

2 is the minimum 3 is safer, especially if factor correlations are weak 4 provides safety 5 or more is more than enough (If too many indicators then combine indicators into sets or parcels.)

Identification
Overidentified model = knowns > unknowns
Number of knowns = Number of variances and covariances of observed variables computed by k(k+1)/2, where k is the number of observed variables Number of unknowns (free parameters) is based on the specified model. It is typically a sum of the number of: exogenous variables (one variance estimated for each) endogenous variables (one error variance each) correlations between variables (one covariance for each pairing) regression paths (arrows linking exogenous variables to endogenous variables) Latent variables indicator variables (one error variance for each) paths from latent variables to indicator variables (excluding those fixed to 1)

Parameters in a CFA Model

Factor loadings: the effect of latent variable on that observed measure

Can be unstandardized or standardized If measure loads on only one factor, standardized factor loading is the correlation between the measure and the factor (and square root of measures reliability).

Factor covariances or correlations: association between each pair of latent variables Error variance: variance in the observed measure that is not explained by latent variable

Error variance is variance not explained by the factor (but not necessarily random or meaningless variance)

Correlated errors: correlation between a pair of error terms

Parameters in a CFA Model

V3

W1 1 V1 W4 1
BSI8 Depression Factor BSI5 e5

V4 1
e8

V5 W2
BSI10

Number of knowns = 28 (7*8)/2=28

1
e10

W5 1 W6 W3 1
BSI16 e16 BSI14 e14

V6

V7

V8 1
BSI18 e18

Number of unknowns = 14
Variance of latent factor (1) Free factor loadings (6) Variances of error terms (7)

CONFIRMATORY FACTOR ANALYSIS: One Factor of BSI Depression with all parameters labeled

If a researchers measurement model is reasonably correct, then (Kline)

2.

Items (indicators) specified to measure a common underlying factor should have relatively high loadings on that factor (convergent validity) Estimated correlations between the factors should not be excessively high (>.85) (discriminant validity)

Discriminant validity refers to the distinctiveness of the factors measured by different sets of indicators.

What to examine

Parameter estimates (all should be statistically significant different from zero)

variance (one-tailed test)

error variances (one-tailed test) error correlations (two tailed) Check for Heywood cases!!!! (Negative error variances)

Problems In Estimation: Heywood Cases

Heywood Cases: Negative error variance (or a standardized loading larger than 1) Why?
Misspecification

Outliers
Small

sample+2 indicators per factor Empirical under-identification

Problems In Estimation: Heywood Cases

How to eliminate Heywood cases:

Search for specification error

Sometimes a measure has correlated error Create 2 factors

Fix error variance to zero

Creates an extra df as one parameter is not estimated Need to adjust chi-square and fit indices

Non-linear constraints that error variances cannot be negative (always in EQS) Set loadings equal (must use covariance matrix) Use an alternative estimation method beside ML Empirical underidentification: make sure correlations are not weak

AMOS Bollen Example

Respecification

Simpler Model
Set

loadings equal: use covariance matrix and variables must be in the same metric

More Complex Model

errors: a priori justification Second factor

Use Diagnostics: residuals

These are nested

If the two-headed arrow in model b is set to 1, that would be saying there is only one latent trait.
Thus model b has one more path than model a.

Proposed 1 Factor Structure

Testing: Comparison of Nested Models

Two models
Base

Model More Complex Model (e.g., base model with additional paths)

df If the base model is good fitting, then the more complex model must also be good fitting. Chi square and degrees of freedom are subtracted to test constraints made in the base model The more complex model should be a good fitting model, otherwise the conclusion is that one model is less poor than another.

fewer

Testing: Comparison of Nested Models

Chi-square difference test

Run both models Subtract Chi-square and df: Simple-Complex

Complex: more parameterized, less parsimonious Simpler: less parameterized, more parsimonious

2 diff = n.s. favor parsimonious model 2 diff = sig favor more parameterized model

Notes about Nested Models

Models must have same variables

will not be nested if your respecification involves deleting a variable(s)

Can add parameters, delete parameters, but cannot do both

Nested Models?

Nested Models?

Path Analytic Specification

All loadings are freely estimated Factor variance are set to one Error paths are freely estimated:
(Standardized)

error path equals the square root of one minus the standardized factor loading squared

Error variances are set to one

Path Analytic Specification

e1

e2

e3

e4

4 basic steps- CFA:

Define the factor model.

select the number of factors to determine the nature of the paths between the factors and the measures. Paths can be fixed at zero, fixed at another constant value, allowed to vary freely, or be allowed to vary under specified constraints (such as being equal to each other).

2.

Fit the model to the data.

4 Basic Steps for CFA

When the factor model is fit to the data, the factor loadings are chosen to minimize the discrepancy between the correlation matrix implied by the model and the actual observed matrix. The amount of discrepancy after the best parameters are chosen can be used as a measure of how consistent the model is with the data. Fit statistics

4 Basic Steps for CFA

To compare two nested models, examine the difference between their c2 statistics. Most tests of individual factor loadings can be made as comparisons of full and reduced factor models. For non-nested models, you can compare the Root mean square error of approximation (RMSEA), an estimate of discrepancy per degree of freedom in the model, other fit indices, and the AIC and BIC.