Quitting the use of a habituated hedonic information system: a theoretical model and empirical examination of Facebook users

Abstract

Much research has been devoted to the early adoption and the continued and habituated use of information systems (IS). Nevertheless, less is known about quitting the use of IS by individuals, especially in habituated hedonic settings, that is, IS discontinuance. This study focuses on this phenomenon, and argues that in hedonic IS use contexts (1) IS continuance and discontinuance can be considered simultaneously yet independently by current users, and that (2) IS continuance and discontinuance drivers can have differential effects on the respective behavioral intentions. Specifically, social cognitive theory is used to point to key unique drivers of website discontinuance intentions: guilt feelings regarding the use of the website and website-specific discontinuance self-efficacy, which counterbalance the effects of continuance drivers: habit and satisfaction. The distinctiveness of continuance and discontinuance intentions and their respective nomological networks, as well as the proposed research model, were then empirically validated in a study of 510 Facebook users. The findings indicate that satisfaction reduces discontinuance intentions directly and indirectly through habit formation. However, habit can also facilitate the development of ‘addiction’ to the use of the website, which produces guilt feelings and reduces one’s self-efficacy to quit using the website. These factors, in turn, drive discontinuance intentions and possibly the quitting of the use of the website.

Appendices

Appendix A

Common method bias assessment

While the risk of CMV is mitigated by the use of multiple version of the survey, it may still exist given that data were self-reported (Sharma et al, 2009). It was examined by multiple techniques, because all CMV assessment techniques have drawbacks (Podsakoff et al, 2003, 2012; MacKenzie & Podsakoff, 2012). First, the correlation matrix was examined based on Pavlou et al’s (2007) suggestions. Very high correlations (in excess of 0.9) can be indicative of CMV, but the correlations in this study were substantially lower, and quite a few were very low and non-significant (r=0.00 ns to 0.57, P<0.01). Second, Harman’s single factor test was performed. The data for the multidimensional constructs of the model were subjected to principal component analysis with no rotation. This procedure produced seven components that explained 77% of the variance, and the largest one captured only 28% of the variation. Third, a modified Lindell & Whitney (2001) test was performed. An unrelated marker construct, namely affective state of surprise (three items from the PANAS-X inventory by Watson and Clark (1994) – amazed, surprised, astonished) was captured in the survey. These items were included in a factor analysis procedure with all items pertaining to the model’s constructs. The items loaded highly on the affective state of surprise factor (loadings of 0.79–0.81), and had low loadings on all other factors (below 0.30). Forth, a common latent factor (MacKenzie & Podsakoff, 2012; Podsakoff et al, 2012) that captures the communal variance of all model indicators was added to the CFA model. The CFA model was then estimated with and without this common latent factor, and the differences in loadings were found to be negligible (0.01–0.12) and below the 0.2 cutoff. Taken together, all of these tests indicate that common method bias is unlikely to be pertinent in the data. However, these techniques are imperfect (MacKenzie & Podsakoff, 2012; Podsakoff et al, 2012), and even their combined conclusion may be imprecise. Thus, future research can collect data such that there is temporal, proximal, or psychological separation between exogenous and endogenous variables. Future research can also go beyond the elimination of common scale properties that was utilized in this study, that is, the shuffling of the survey pages, and try to use other procedural remedies such as more balanced negative and positive items as well as statistical remedies such as measuring one’s response style or controlling for directly measured source of bias (MacKenzie & Podsakoff, 2012; Podsakoff et al, 2012).

Appendix B

Examining potential self-generated validity effects

Self-generated validity describes the potential reactive effects of measurement on outcomes (Chandon et al, 2005). It is possible that asking respondents to report on certain intentions or perceptions makes these intentions or perceptions more accessible in the memory and consequently influence resultant assessments and actions. This can lead to subsequent beliefs, attitudes, intentions, and behaviors which are more consistent with the reported intentions and perceptions (Feldman & Lynch, 1988; Chandon et al, 2005)

In the current study, it is possible that asking individuals to report say their levels of addiction may make the addiction symptoms more accessible in their memory, and as such, prime individuals to report on stronger guilt feelings and/or stronger discontinuance intentions toward the hedonic IS. Therefore, the study was designed in a way that allows the mitigation and examination of such potential self-generated validity effects. Specifically, the measures used in this study were spread over four webpages. The first page ((A) demographics and website use and behaviors) was constant, and the remaining three pages ((C) continuance factors, (D) discontinuance factors, and (O) other factors) were rotated. This process has yielded six versions of the survey. Table B1 reports sample sizes and construct means for each version. The order of the pages in each version is reflected in its name. For example, Version 2 had the following page order: demographics and use, continuance, discontinuance, and other factors.

Table B1 Sample sizes and construct means for the six versions

As can be seen, the differences between construct means across versions seem small, and all appear to be in the same range across versions, and in no particular direction. To empirically test whether these small differences in means manifest from version differences, a Multivariate Analysis of Variance (MANOVA) model was specified and tested in SPSS 20. In this model, the version was specified as a fixed factor and all the measures were specified as dependent variables. Pilai’s Trace of 0.36 (F(186)=0.98) was not significant (P<0.57), indicating no omnibus version-based differences among the data sets. Thus, the order of the presentation of the scales did not significantly influence the way individuals responded to these questions. Hence, it was concluded that self-generation does not seem to be a problem in this data set.

Appendix C

Assessment of differences between continuance and discontinuance intentions

As another preliminary analysis, this study sought to empirically examine whether continuance and discontinuance intentions should be treated as separate factors and whether predictors can have differential effects on both. To this end, differences between these two concepts were established using the criteria for dual-factor phenomena (Cacioppo & Berntson, 1994; Cacioppo et al, 1999). First, the existence of cases in each quadrant in Table 1 was examined. Using median splits for continuance and discontinuance intentions 124 participants (24%) were in Quadrant 1, 160 (31%) in Quadrant 2, 188 (37%) in Quadrant 3, and 38 (7%) in Quadrant 4. Thus, there are possible cases of ‘apathy’ and ‘ambivalence’ in the sample, which indicate that people can have simultaneous low (high) continuance and discontinuance intentions.

Second, a principal component EFA with Promax rotation with the default of κ=4 was performed on continuance and discontinuance intention items. Promax is an oblique rotation which was chosen since the two constructs are expected to be at least moderately correlated. Two components emerged, with high loadings over 0.95 and low cross-loadings, below 0.45. The correlation between the components was −0.56, which demonstrated moderate to high negative correlations, but no major overlap (the constructs share about 30% of the variance).

Third, a CFA model was specified and estimated using AMOS 20 using Maximum Likelihood estimates. The fit indices were acceptable [χ²(df=8)=18.1, χ²/df=2.26, CFI=0.99, IFI=0.99, RMSEA=0.050 with P-close=0.46, and SRMR=0.012]. All loadings were significant (P<0.001), and the correlation between the two factors was −0.58 (P<0.001). A χ² test was used to contrast this model with a single factor model allowing all six items to load on a single factor [χ²(df=9)=1422.5, χ²/df=158, CFI=0.65, IFI=0.65, RMSEA=0.56 with P-close=0.00, and SRMR=0.20]. The χ² difference statistic was significant [χ²(1) =1404.4, P<0.000]. Therefore, estimating the correlation between the two factors statistically significantly reduced the χ² statistic, and the two-factor model significantly fits the data better than does the single-factor model.

Lastly, a sequence of χ² difference tests was performed on a model that included both continuance and discontinuance intention and the hypothesized predictors of discontinuance intentions. The coefficients and explained variances are reported in Table C1, and the χ² comparisons are reported in Table C2. The results show that all the predictors, except for satisfaction, separately and together, had statistically different influences on continuance and discontinuance intentions. Habit had a much stronger effect on continuance (β=0.35) than on discontinuance intentions (β=−0.16), and guilt had the opposite pattern – much stronger effect on discontinuance (β=0.38) than on continuance (β=−0.12) intentions. Similarly, self-efficacy to discontinue had a significant effect on discontinuance intentions (β=0.17), but no significant effect on continuance intention. Satisfaction was the only predictor that equally catered to both continuance and discontinuance decisions. The correlation between continuance and discontinuance intentions was −0.45, P<0.001. These findings imply that discontinuance intentions can have their own predictors (e.g., guilt and self-efficacy to discontinue) and a nomological net different from this of continuance intentions.

Table C1 Path coefficients of a model for establishing nomological differences between continuance and discontinuance intentions

Table C2 Fit indices and χ ² difference tests for establishing differences in the nomological nets of continuance and discontinuance intentions