On The Use And Reporting Of Covariance-Based Structural Equations Models In Assessing Survey Research.

Wynne W. Chin
University of Houston

Feel free to send comments to: wchin@uh.edu

originally created: November 28, 1995
last update: November 29, 1995 

Accesses since last random clear out by provider. 

1. Introduction

1.1. Surveys are frequently used to validate models in MIS research.

A critical aspect is to assess whether the survey items represent the underlying constructs they were developed to measure. Ideally, the test of the reliability and validity of these items should be made in the context of the theoretical model being studied. Covariance Based Structural Equation Modeling (SEM) represents a technique which allows the researcher to perform such an assessment in a holistic fashion. Software currently available to perform such analysis are LISREL, EQS, AMOS, EZPath, SEPATH, CALIS, MX, and RAMONA. Alternatively, a variance based approach (not discussed here) known as Partial Least Squares can also be used in assessing survey instruments.

1.2. How does it work?

A covariance matrix can be created from the survey items that a researcher plans to use - this represents the empirical data. An underlying model is hypothesized to exist and in turn explains the sample item covariances. Parameter estimates are chosen for this model in such a fashion that the covariance matrix implied by the model is as close as possible to the sample covariance matrix. Thus, SEM software attempts to minimize the differences between the sample covariances and those predicted by the theoretical model.

2. Five Generic Steps In Structural Equations Modeling

2.1. Model Specification

2.1.1. Graphical representation

As an example, we begin with the standard convention (see Figure 1) of using circles (or ovals) to represent latent variables (i.e., factors or constructs). Squares are used for actual measures such as questionnaire items (often referred as manifest variables or indicators). For the models examined, the parameters estimated are numbered and assigned an asterisk next to it. Those paths or loadings that are fixed are labeled according to these numeric values. Finally, the variances for exogenous factors (i.e., independent factors which are not affecting by other factors) are assumed to be set at 1 and not labeled (otherwise an asterisk would be next to the factor label).

Figure 1. Standard conventions for representing variables.

In Figure 2, a single factor model is presented. The latent variable F1 is assumed to have a variance of one since no asterisk is assigned to it. Three manifest variables (V1 through V3) are presented and assumed to reflect the underlying factor. The extent to which these measured items actually tap into the underlying factor are determined by estimating their respective path loadings (P1* through P3*). Finally, the model indicates that each manifest variable may also measure other factors beyond the particular one of interest. Alternatively, there may be a certain amount of measurement error reflected in each manifest variable. This assumption is modeled by having each manifest variable affected by a unique factor (U1* through U3*). In the figure, the path loadings from each unique factor are fixed at 1 with the variance of each unique factor needing to be estimated (as indicated by an asterisk). An alternative approach would be to fix the variances for each unique factor to 1 and estimate the corresponding paths.

Figure 2: A single factor model.


Note that in the model presented, it is not enough to diagram the paths among factors and manifest variables. Such path diagrams are often presented in IS papers without indicating which parameters are being estimated and which are set at particular values. While the path diagrams among latent and manifest variables help convey the general ideas behind causal models, there is still enough information left out to make it difficult for other researchers to reproduce the analysis.

2.1.2. Data (i.e., manifest variables) description Context of data acquisition and sampling procedure Population demographics Statistical description (skewness, kurtosis - univariate and multivariate; outliers) Number of subjects Sample covariance matrix

2.2. Identification

2.2.1. Is there one unique set of values that exist for the model parameters being estimated?

2.2.2. For all correlated first order factors - two indicators per factor.

2.2.3. For orthogonal factors model - each factor must have three indicators (Anderson and Rubin, 1956)

2.2.4. If no rules are available - determine identification via algebraic manipulation.

2.2.5. Scale Indeterminacy (failure to set the metric associated with each factor)

2.3. Estimation

2.3.1. Computer program and version number

2.3.2. Starting values used for estimation

2.3.3. Computational procedure employed (i.e., ML, ULS, GLS, WLS) consistent with the data?

2.3.4. Anomalies encountered during estimation Were there improper solutions such as negative variances, standardized paths greater than +1 or -1? Did the program set offending (i.e., negative) variances to zero?

2.3.5. Are the results presented confirmatory or exploratory?

2.4. Testing Fit

2.4.1. Absolute fit indices: extent to which the overall model (structural and measurement) predicts the observed covariance (Chi-Square, GFI, Residuals)

2.4.2. Incremental fit indices: extent to which the specified model performs better than a baseline model (e.g., no underlying factors) (Bentler-Bonnet Normed Fit, Tucker-Lewis) .

2.4.3. Parsimonious fit: accounts for the degrees of freedom necessary to achieve the fit (Normed Chi-Square (Chi-Square/df), AGF, Akaike information criterion

2.4.4. Resampling Procedures Bootstrapping the Data Set Cross-validation Permutation of the indicators

2.4.5. Statistical Power of the Test Importance of power analysis: A) If the model is correct, the chi-squared test statistic should be non-significant B) If the model is false, the chi-squared test statistic should be significant Need appropriate power in order to ensure that statement B occurs. Power is influenced by alpha level the effect size of the path(s) for the correct model number of indicators used the alternative model used as a basis of comparison Standard Procedure for testing statistical power: 1) Using the original model (Ho) to be tested, specify an alternative model (Ha) in terms of additional parameter(s) (e.g., a path loading or correlation between factors) and their specific effect size values. 3) Take the resulting covariance matrix and analyze it using the original model (stipulating the sample size to be the same as the original model). 4) The Likelihood ratio value (l) resulting from this analysis (i.e., the chi-square statistic from the program) should approximate a non-central chi-square distribution with degrees of freedom equal to the difference between the two models (Ho-Ha).

2.5. Respecification



ISWorld Net Navigation Map

The ISWorld Net Home Page
The ISWorld Research Home Page
MIS Surveys Research
On The Use And Reporting Of Covariance-Based Structural Eations Models In Asessing Research. Where you are right now!
Return to the Table of Contents