CHAPTER 5. METHOD
5.4 SEM Specification and Identification
Model Specification and Illustration
After the latent variables were stabilized in terms of their indicators, reliability and validity, variables were created from averaged scores of the indicators retained in the SEM analysis. Although detailed hypotheses and models were already explained in Chapter III and Chapter IV, it is necessary to explain why structural equation modelling was chosen as the modeling strategy instead of other similar types of variance- or covariance-based multivariate analysis methods.
First of all, an important feature of the current dissertation is that it is a structural- functionalist reorganization of existing theories in public diplomacy and international political communication. This means that it is concerned with the relationship of the actors and their influences on each other. Thus, a multiple regression, as Zhu and He (2002) did in their study, would be insufficient to explore the relationship between
of perception of governmental control on information resource, the first set of models in Chapter III hopes to study governmental influence on media credibility perception. Thus, the direction here is from government to citizens, not the other way around.
Since variables such as perception of media credibility require model comparisons to test some of the hypotheses in Chapter III and Chapter IV, the analytical strategy is: the dissertation will first test model in figure 5.1 (the same as figure 3.1) proposed in Chapter III. And once the initial model and hypotheses testing for Chapter III is done and the competitive relationship between foreign media and domestic media is confirmed, a post hoc analysis will be done on the model in figure 5.1 to determine if there are any missing links that were ignored.
If there is theoretical support to supplement the original model with extra links, these links will be added to the model and the model will be reevaluated in the post hoc analysis. This process will yield a final model for hypotheses in chapter III, which will then be used in the model testing for chapter IV hypotheses.
The next step is to combine the models in chapter III and chapter IV using media dependency variables (Grant, 1996) (dependency on foreign media and dependency on domestic media) as mediators to behavioral antecedents, namely problem recognition,
involvement recognition and constraint recognition (Kim and Grunig, 2011). For the initial test, the possible loop between the two types of dependencies (on either foreign or domestic media) is not allowed to maintain the simplicity of the model. After all,
theoretical discussion about competition between different types of system dependencies goes beyond the scope of this dissertation. In sum, the model shown in the following figure will be tested as the final model, as a combination of the theoretical arguments and hypotheses in chapter III and chapter IV.
Covarying Error Terms of Latent Variables
Covariance of error terms of endogenous variables in SEM has been explained by statisticians. In the models proposed by the current chapter, it is important to mention the necessity to covary error terms of some of endogenous variables and justify why they should be covaried.
Figure 5.2 Final Model
theoretically, variables in the same block tends to be spurred by a common latent variable. For example, numerous studies found that dependencies on different types of media are positively correlated. This shows the possibility that either foreign media dependency or domestic media dependency is a part of people’s general dependency on all media.
This means that although the model expects them to be different constructs, their error terms (the variance that cannot be explained by distinguishing foreign from
domestic) are very likely to be spurred by a common latent variable of general media dependency, which is not observed by the indicators in our questionnaire. This same rationale applies for variables in the media credibility block, however, both model 1 and model 2 already hypothesized direct interactions between foreign credibility and
domestic credibility, thus their shared variance is already considered in the model. Therefore, in these two models, the error terms between foreign media dependency and domestic media dependency will be proposed as covarying.
As for the variables in the communicative behavior antecedent block (see figure 5.2), previous study on the STOPS model have predominantly treated these as exogenous variables, and thus they are theoretically hypothesized as covarying. Considering that they usually covary and constitutes to a common latent factor called situational
motivation in previous models, in the current study their error terms will be proposed as covarying.
That said, it is important to remember that this study does NOT covary any error terms across blocks: so far, theoretical support is evident only for errors covarying within
a block, and the ultimate purpose of this dissertation is to illustrate 1) the interactions between different blocks of variables and 2) the comparison of psychometric effects of foreign v.s. domestic media, when controlling for their common variance.
Model Identification
Now the next step is to make sure model 1 (figure 5.1) and model 2 (figure 5.2) are identified models that can be analyzed using SEM strategy. For model 1, because there is a non-recursive block (i.e. loop between credibility of foreign media and credibility of domestic media), it is necessary to evaluate whether the model is rank- identified. Kline (2016) proposed a simple way to do this by constructing a system
(p = 𝑣(𝑣+1)
2 , v = number of latent variables in the model) and q is the number of estimated
parameters (including latent variable variances, covariances, and the variances and covariances of their residuals). We can calculate that dfModel1 = 6*7/2 –14 = 7. Obviously, dfModel1 is larger than zero, making model 1 identified. For model 2, the dfModel2 is
calculated to be 20. dfModel1 and dfModel2 will later be used to determine sample sizes given
desired power (refer to the next two sections).