COVID-19 Model Iteration 3-a-1 Structural Model Evaluation

Model plots

Structural Model Only

This is a path model showing only the structural model components.

Collinearity

Collinearity is assessed using the variance inflation factor (VIF). VIF should be < 5, ideally \(\leq\) 3.

VIF is above 3 for some constructs, specifically lower-order constructs. But it is not above 5 for any constructs.

In-sample predictive power

In-sample predictive power is assessed using variance explained R². R² \(\geq\) 0.75 indicates substantial in-sample predictive power, R² \(\geq\) 0.5 moderate and R² \(\geq\) 0.25 weak in-sample predictive power. R² \(\leq\) 0.10 indicates a lack of model predictiveness.

All R² values are above 0.25, indicating weak in-sample predictive power at a minimum. For Response Beliefs and Behavioral Intention, R² is above 0.5, indicating moderate in-sample predictive power.

Effect size

Effect size f2² measures the impact of a predictor construct on an endogenous construct. f² \(\geq\) 0.35 indicates a large effect, f² \(\geq\) 0.15 a medium and f² \(\geq\) 0.02 a small effect.

The effect sizes of the relationships including many lower-order constructs are below 0.2. Benevolence has only an effect on Perceived Response Costs. Integrity has only an effect on Perceived Response-Efficacy. Knowledge has only a negligible effect on Response Beliefs and one of its lower-order constructs, Perceived Response Costs. Other than that, all effects are small at a minimum. The effects of Perceived Response Efficacy and Response Beliefs on Behavioral Intention are medium. The effect of Distrusting Beliefs on Response Beliefs is large.

Out-of-sample predictive power

Out-of-sample predictive performance has to be evaluated using a simplified model without HOCs as an approximation as seminr as of now does not support plspredict for HOC. If the root mean square error (RMSE) or the mean absolute deviation (MAD) of the naive LM model is below the ones for the PLS model, this indicates a lack of predictive performance. However, a lack of predictive performance does not imply a lack of explanatory power.

No indicator is predicted better by the model than the LM in terms of both RMSE and MAD. The difference is mostly very small.

Significance and relevance of path coefficients

Significance is denoted by t-test and p values. A t of \(\geq\) 1.65 signifies significance at the 10 % level, t \(\geq\) 1.96 at the 5 % level and t \(\geq\) 2.57 at the 1 % level.

The only insignificant paths are those from Knowledge and Integrity to Perceived Response Costs and from Benevolence to Perceived Response Efficacy.

Summary table

Results summary

All construct and higher-order construct paths are significant, altough the effect size for the path from Knowledge to Response Beliefs is negligible. The model can be compared with other models.