Structural Equation Modeling of Spatial Presence: The Influence of Cognitive Processes and Traits

Abstract

This article examines the formation process of spatial presence, which is conceived as a two-step process involving the construction of a mental model of the mediated environment, followed by the emergence of spatial presence. During both stages, cognitive processes and user traits are in effect. We present data derived from a pooled set of data of three studies using the same virtual environment. Structural equation modeling is used to confirm the proposed theoretical model. The results show that attention and the trait of visual spatial imagery are positive predictors of the mental model of the mediated environment. In the second step, the formation of spatial presence is governed by involvement, the suspension of disbelief, and the domain-specific interest, together with the mental model.

Notes

*Significant at Level p < .05 for two tailed t-test.

**Significant at Level p < .01 for two tailed t-test.

***Significant at Level p < .001 for two tailed t-test.

1. For a discussion on the (un-)realness of virtual environments (cf. Lee, 2004).

2. Note that there are two alternative conceptualizations. In addition to the dispositional view, Krapp, Hidi, and Renninger (1992) identified interest as a characteristic of a learning environment or as a psychological state.

3. We conceive involvement to be the active and intense processing of a stimulus (cf. Rothschild, 1983; Salmon, 1986; Wirth, 2006).

4. A sample size greater than 150 is deemed appropriate for the use of a structural equation model. Such a sample size yields sufficiently small standard errors of the estimates and is adequate to obtain converged and proper solutions for models with three or more indicators per factor (Anderson & Gerbing, 1988).

5. The sample used for the analysis is a pooled data set. Both of these method factors did not significantly influence the experience of spatial presence.

6. Because our data are derived from experimental studies, the dataset contains very few missing values (i.e., four data points). Nevertheless, we estimated the model using the full information maximum likelihood method, which shows the best performance in handling missing data compared with procedures such as listwise or pairwise deletion (cf. Enders & Bandalos, 2001).

7. The chi-square statistic (its p value, respectively) seemingly indicates that the theoretical model does not fit the data. However, the usefulness of the chi-square statistic as an indicator of model fit has been challenged for several reasons (Bentler, 1990): a) The statistic is sensitive to sample size and likely to be significant for large sample sizes (Hooper, Coughlan, & Mullen, 2008). In our case, the sample size is moderate and it is not clear whether the sample size is a problem. In addition, Marsh and Balla (1994) have argued that the chi-square statistic is “generally inappropriate for evaluating the size of the empirical discrepancy” (p. 186), which is defined as the difference between the fitted covariance matrix (E) and the sample covariance matrix (S). Accordingly, the chi-square statistic is likely to be biased. In order to minimize the impact of the sample size on the model chi-square, one should use the normed χ/df statistic. A good model fit is indicated if the χ/df value is between 0 and 2. Thus, in our case χ/df = 1.298), the model has a good fit (cf. Bollen, 1989; Schermelleh-Engel, Moosburgger, & Müller, 2003). b) To avoid chi-square's general inappropriateness, other fit indices ought to be used to assess model fit. We used CFI, RMSEA, PCLOSE, and SRMR. In our case, all of these indices indicate a good model fit. For example, our CFI—a measure that is least affected by sample size (Fan, Thompson, & Wang, 1999)—has a value of .954, which indicates a good fit (Schermelleh-Engel, et al., 2003).