The UPEQ questionnaire showed reliability (between-item reliability α = 0.930, main factors α = 0.930 and between subfactors α = 0.915) and all its main and subfactors significantly correlated with the rating of the game. Although rating itself showed a negative correlation (τ = − 0.033) with ‘playtime hrs’ and ‘daysplayed’ (τ = − 0.043), it is assumed that higher rating as well as higher scores on factors and subfactors of UPEQ indicate a higher satisfaction with the game in general.
Over 200 pairs of Kruskal–Wallis H tests were conducted; therefore, we only report the differences that are shown to reject the null hypothesis—0.05 significance level, the distributions of the testing variable’s two samples (generations in this case) are the same.
Among multiple factors of motivation, there were significant differences regarding Playstyle between Millennials (M = 3.82, SD = 1.03) and both Generation Z (M = 3.91, SD = 1.02) and Baby Boomers (M = 4.01, SD = 0.91).
Agency displayed the greatest number of significant pairwise comparisons across the generations. Although the difference between Generation Z (M = 3.19, SD = 1.13) and Millennials (M = 3.14, SD = 1.0) did not reject the null hypothesis, all other comparisons, including with Baby Boomers (M = 3.63, SD = 0.98) and Generation X (M = 3.35, SD = 1.01), were significant.
For Growth, only Baby Boomers (M = 4.33, SD = 0.86) showed a significant difference across the four generations. While Generation X showed (M = 4.18, SD = 0.85) similar means to Millennials (M = 4.12, SD = 0.91) and Generation Z, (M = 4.17, SD = 0.91).
Mastery saw significant differences between Generation X (M = 3.97, SD = 0.8), and both Millennials (M = 4.03, SD = 0.85) and Generation Z (M = 4.08, SD = 0.84).
Closeness displayed one significant difference, namely between Baby Boomers (M = 3.62, SD = 1.2) and Generation X (M = 3.48, SD = 1.17).
The distribution of ranks in Interdependence was the same across the different generations. While Closeness to non-playable characters (NPC Closeness) was not significantly different between Baby Boomers (M = 3.15, SD = 1.11),Generation X (M = 3.07, SD = 1.07) or between Millennials (M = 2.83, SD = 1.17) and Generation Z (M = 2.85, SD = 1.22), it was significant for all other comparisons.
For measures of Presence, only narrative presence showed a significant difference in three degrees. Baby Boomers (M = 3.76, SD = 0.85) had significant mean differences with both Millennials (M = 3.55, SD = 1.00) and Generation Z (M = 3.54, SD = 0.99). The third significant mean difference was between Generation X (M = 3.65, SD = 0.9) and Millennials (M = 3.55, SD = 1.0).
A variety of the Behavioral measures were more prominent, such as ‘daysplayed’, ‘playtime hrs’,’prelvl30 playtime’, rejecting the null hypothesis for all six pairwise comparisons and only Baby Boomers and Generation X being the same for ‘daysingroup’, ‘playtime group hrs’, ‘prelvl30 playtimegroup’ and ‘max gearscore’. ‘Max level’ was the only variable that showed significant differences between Generation Z and other groups as well as between Millennials and Generation X. Means and standard deviations for these measures are presented in Table 1.
Based on the self-reported measures, Rating of the game only differed between Baby Boomers (M = 8.16, SD = 1.81) and Millennials (M = 7.91, SD = 1.79); Players self-reported the number of hours per week which were spent on playing video games, and the sole difference identified for this measure was for Generation Z (Median = 21–30 h per week) compared to the other generations (Median = 11–20 h per week).
Additionally, analyses of self-reporting based on the Money spent on video games per month resulted in a decreasing trend of reporting with age, displaying the lowest amount for Baby Boomers (Median = 11–30 $) in comparison with Generation Z (Median = 31–60 $).
Kendall’s tau for non-parametric correlations with age and generation index (categorical number associated with each generation) were particularly high. Table 2 presents Kendall’s tau values of Age and Generation index correlates with measured variables. Among those, strongest significant correlation was observed between generation index and ‘prelvl30 playtime hrs’ at τ = 0.326, which is significant at the 0.01 level (2-tailed).
Correlations with generation index were also significant (2-tailed at the 0.01 level) for the number of days played (‘daysplayed’) and the amount of time (hrs) played (‘playtime hrs’) (both τ = 0.324), as well as ‘days in group’ (τ = 0.209), while ‘max gear score’ (τ = 0.177), ‘playtime group hrs’ (τ = 0.153), ‘prelvl30 playtime group hrs’ (τ = 0.139), ‘max level’(τ = 0.121), self-reported hours per week (τ = − 0.084), NPC Closeness (τ = 0.057), Agency (τ = 0.055), and Mastery (τ = − 0.050).
Age showed similar power in correlations with ‘daysplayed’ (τ = 0.317) ‘playtime hrs’ (τ = 0.314), ‘daysingroup’ (τ = 0.205), ‘max gear score’ (τ = 0.171), ‘playtime group hrs’ (τ = 0.149), ‘prelvl30 playtime group hrs’ (τ = 0.138), ‘max level’(τ = 0.115), self-reported hours per week (τ = − 0.079), NPCCloseness (τ = 0.060), Agency (τ = 0.053), and Mastery (τ = − 0.049).
Correlations for Age and Generation Index
We employed a series of machine learning techniques to further examine the significance of behavioral and self-reported measures in prediction of Age and generational index. As this testing conditions include 64 classes of age in 4 classes of generations, a random attempt at predicting age and generation will have baseline accuracies of 1.56% and 25% respectively.
First, we used a step forward neural network, which is a logistic regression classifier where the input is first transformed using a learnt non-linear transformation. This transformation projects the input data into a space where it becomes linearly separable. This intermediate layer is referred to as a hidden layer (Rumelhart et al. 1986).
In this experiment, a multilayer perceptron with 13 neurons in the hidden layer yielded a 30.00% accurate model in predicting age based on the input variables, including ‘playtime hrs’ (Predictor Importance = 0.11), Competence (PI = 0.10), ‘prelvl30 playtime hrs’ (PI = 0.08), Relatedness (PI = 0.07), ‘playtimegroup hrs’ (PI = 0.06), Growth and ‘daysplayed’ (both PI = 0.05), and Closeness (PI = 0.04). For the prediction of generation, a hidden layer consisting of 5 neurons reached the accuracy of 24.10% with the variables of ‘playtime hrs’ (PI = 0.13), ‘prelvl30 playtime hrs’ (PI = 0.09), ‘daysplayed’ and ‘max level’ (both PI = 0.07), as well as ‘prelvl30 playtime group hrs’ (PI = 0.05).
The support vector machine regression module, which aims to fit the best regression line within a margin of error, was fed with all aforementioned input variables, predicted the age of our participants with a 26.05% success rate, compared to prediction of generation, which was accurate at 58.31% level. This difference in accuracy gains might be due to the number of categories for prediction (4 classes of generation and 64 classes of age in this study).
The linear model with a forward stepwise selection produced a 25.40% accurate model in predicting age, with predictors such as ‘playtime hrs’ (PI = 0.28), ‘playtime group hrs’ (PI = 0.27), Agency and ‘daysplayed’ (both PI = 0.08), self-reported hours per week (PI = 0.07), Emotional Presence (PI = 0.05), as well as Growth (PI = 0.04). For generation, the accuracy was at a 30.70% range, with the predictors ‘playtime hrs’ (PI = 0.28), self-reported hours per week (PI = 0.11), ‘playtime group hrs’, Agency and ‘daysplayed’ (all PI = 0.10), Emotional Presence (PI = 0.06), as well as Growth and Competence (both PI = 0.05).
Finally, a multimodal ensemble, which uses a supervised learning algorithm that combines a set of classifiers into a meta-classifier by taking weighted voting of their prediction for the final forecast, for age and generation prediction respectively produced 55.30% and 58.70% accurate results which were the highest rates among all tested models. Please refer to “Appendix 2” for graphs showing model accuracy gain for prediction of age and generation.