I am modelling consumers' intention to discontinue using mobile payments. The proposed model comprises of 3 independent variables, one moderating variable, and dependent variable. 5-point Likert scale used to collect responses (250+) through a questionnaire (seems to be ordinal data, but could be argued). I could not find a clear guidance in the literature or from my tutors regarding which type of statistical analysis should I use to test my model: SEM, ordinal regression, linear regression, or other.
You could still use path analysis in the context of SEM and assume that the variables are categorical or ordinal. This could be a better approach if you want to see direct and indirect relationships through the moderating variable. With multiple steps in ordinal regression, you may get more or less the same results as well.
Thank you Ette and Okan.
Overall descriptive statistic makes sense.
The data meets Normality, Linearity and Multicollinearity, and Auto-Correlation is OK, with just about 5 outliers.
I have tried Ordinal Regression (SPSS), just to see what it would show, and there are more questions than answers: Model Fitting seems Ok (Sig.= .000), Goodness-of-Fit is perfect (both Pearson & Deviance= 1.000), Pseudo R-Square seems Ok (Nagelkerke= .937), Test of Parallel Lines (Sig.=.641; Log-Likelihood value=.000 > may be a complete separation in the data), however, Parameter Estimates are just mess, jumping from positive to negative from item to item, with Significance level varies from .990 to .000 within many items. I have tried to do item by item, and adding Gender as factor, and Usage frequency (moderating variable) as covariant, but the results are the same.
On top of that, majority of papers from the field (IS research) use SEM in modelling, while Ordinal Regression is not used at all, from what I could find.
Maksym,
Can you tell us more again how you process the Likert-type items? Do they define a latent variable, or do you just use the items directly as dependent and independent variables in the analysis? SEM would be more flexible to handle all of these options in one place.
I coded both IVs and DV as Strongly Agree - 5 ... Neutral -3 ... Strongly Disagree - 1. Each IV comprises of 4 items, while DV of 2 items. Moderator had 4 answer options and was coded as low and high.
It seems that SEM is more viable option as more and more people suggest that.
I suggest R lavaan project or partial least squares with nonparametric estimation (i.e. SmartPLS program) which accept ordinal latent or observed variables, and with any (or very less) assumptions.
Thank you all for your suggestions.
Consequently, I have successfully used SEM as statistical analysis technique. It appeared quite easy and effective. Soon I will publish my research findings.
- Badges
- Science topic
- Similar topics
- Psychology
- Psychometrics