Past Gaming Experience and Cognition as Selective Predictors of Novel Game Learning Across Different Gaming Genres

One hundred and seven adults (40 females; M_Age = 34.79, SD_Age = 20.59; aged 18–77) participated in this study. Fifty-three participants were recruited from the University of Texas at Dallas and received class credit for their participation. The remaining 54 participants were recruited from Dallas and its neighboring communities and were paid $10/h for their participation. Inclusion criteria were high school education (or more), normal or corrected-to-normal vision (visual acuity >20/30), and no reported history of major medical or psychological illnesses (e.g., visual neglect, epilepsy, heart attacks, etc.). Participants aged 55 and older were required to score 26 or above in the Mini-Mental State Examination (MMSE 2nd Edition; Folstein et al., 2010), an indicator of normal cognitive health.

The study was approved by the Institutional Review Board of the University of Texas at Dallas and performed in accordance with the ethical standards for research outlined by that review board. Written and informed consent was collected from all participants included in this study prior to any testing procedures.

In this 2 h study, after informed consent, participants underwent cognitive assessments of processing speed (a measure of attention), short-term memory, and working memory. Then the participants played two novel online video games for which they had no reported prior experience. These games were played in succession for 40 min each (80 min total). The order of the games played was counterbalanced between the participants. Finally, participants answered a survey of their past gaming experience and their preferences on the two games played.

Cognitive tests administered included the following: forward digit span (FSpan), a measure of short-term memory that requires temporary storage of digits for a very short period of time (WAIS-R: Wechsler Adult Intelligence Scale—Revised; Wechsler, 1981); backward digit span (BSpan), a measure of working memory that requires, in addition to storage, manipulation of the digits in mental workspace (WAIS-R; Wechsler, 1981); and the Symbol–Digit Substitution Test (SDST, taken from the MMSE 2nd Edition), a measure of processing speed (Folstein et al., 2010).

The two video games played were Tank Attack 3D, an action game, and Sushi-Go-Round, a strategy game, on miniclip.com↗, a website which hosts free casual games spanning multiple genres (Baniqued et al., 2013). Games from minilip.com↗ were utilized due to their ease of access and short individual play duration, which allowed for multiple iterations of the game to be played in a relatively short time period and facilitated the assessment of game learning quickly. Additionally, both games feature adaptive difficulty, introducing new and more challenging mechanics in response to high performance by the player, which allowed us to assess the degree to which our participants learned these mechanics by proxy of progression through each game. Lastly, both of these selected games conform to the “action” and “strategy” genres as defined by the video game cognition literature, thereby avoiding issues of interpretation presented by more complex, longer-playing video games that involve multiple genre features (Dale and Green, 2017).

Tank Attack 3D (Figure 1) was selected as an exemplar action game due to its strong perceptual and attentional demands, coupled with relatively minimal resource management and decision-making requirements (Baniqued et al., 2013; Ray et al., 2017). In Tank Attack 3D, players are required to operate a tank through a war area to destroy enemy radars, tanks, and bases. To complete the mission successfully, players were required to arrive at their home base with remaining energy before the timer ran out. In addition to the time awareness required for successful play, players had to react to fast-moving stimuli and keep track of multiple items, such as the number of radars destroyed and life energy left. Additionally, Tank Attack 3D periodically introduces new types of enemies in response to player victories, forcing the player to adapt to those new challenges to succeed in later missions.

Sushi-Go-Round (Figure 1) was selected as an exemplar strategy game due to its strong emphasis on resource management, coordination between multiple information units and tasks, rapid reasoning, and decision making (Baniqued et al., 2013; Ray et al., 2017). In Sushi-Go-Round, players are required to serve a variety of sushi preparations to a continuous stream of customers while simultaneously monitoring customer happiness, order times, remaining ingredients, and monetary incomes and expenses. To complete a level successfully, players were required to accumulate a certain monetary value by serving customers the requested dishes, while avoiding making the customers wait too long, serving the customer an incorrect dish, or making errors in dish preparation. Additionally, Sushi-Go-Round periodically introduces new and more complex recipes to the menu in response to player success, forcing players to continuously learn new recipes to succeed in later levels.

Participants were asked to play each game continuously for 40 min, and the play order was counterbalanced across participants. A round of either game lasted for about 7 min. The games also become adaptively more difficult, indexed by the game level, based on the participant’s performance. In cases where a round was in progress after 40 min had elapsed, participants were asked to finish that round and then cease play.

A learning composite (LC) measure was calculated for each game for each participant based on two outcome measures: highest level reached in each game at the end of 40 min and the learning rate. Learning rate was calculated as the slope of a logarithmic function fitted to each participant’s scores (higher scores indicate higher success in game play). A steeper slope implies faster learning and greater improvement (Basak et al., 2011; Ray et al., 2017). Highest level reached reflected the difficulty tier reached in each game and was reported as “day” for Sushi-Go-Round and “mission” for Tank Attack 3D. These two measures (learning rate and highest level reached) were first standardized (z-scored) and then averaged to create the LC. We utilized a composite measure to assess learning rates as a method of insulating against fatigue effects: the score output of both Sushi-Go Round and Tank Attack 3D fluctuates with the participant’s performance and can drop over time if the participant becomes fatigued or less engaged with the task. Highest level, conversely, can never fall as a result of poor performance in these games and, as discussed above, is a proxy of the participant’s learning of the mechanics of each game. Constructing these measures allowed us to examine two different metrics of learning on each game, while lessening the overall measure’s susceptibility to fatigue effects, and is consistent with our approach in previous published research which assessed learning rates on these two games or similar games (Boot et al., 2010; Basak et al., 2011; Ray et al., 2017).

Participants completed a survey of gaming habits and preferences adapted from that used by Basak et al. (2008, 2011) and Boot et al. (2008) to assess the level of video game skill (i.e., to identify expert or novice players). This adapted questionnaire addressed the participants’ typical video game habits on a weekly basis, as is the standard methodology used in the game cognition field to differentiate between non-players and players of various skill levels (Green and Bavelier, 2006; Feng et al., 2007; Basak et al., 2008, 2011; Boot et al., 2008; McDermott et al., 2014).

Unless otherwise noted, responses were recorded on a five-point Likert scale for the questions in the survey. Individuals rated how frequently they played different genres and formats of video games from “never” to “often”; ratings were collapsed across five common game formats (personal computer game, game arcade, home gaming console, internet gaming website, phone/tablet game) to determine an individual’s gaming Frequency. Experience with specific action game subgenres (FPS, racing, and simulation); specific strategy game subgenres [RTS, role-playing games (RPGs), simulation, and puzzle]; and casual games was assessed on the same scale. Identification as a gamer was rated on a five-point Likert scale of “do not identify at all” to “strongly identifies.”

Gaming duration was measured via a numeric report of estimated weekly hours spent gaming over the past month. These variables (Frequency, Duration, Identification) constitute an individual’s game habits for the purpose of this study.

In addition to these questions, participants also rated their level of enjoyment of playing the two games used in this study (1 = did not enjoy, 5 = highly enjoyable). The enjoyment rating for Tank Attack 3D was subtracted from that of Sushi-Go-Round to yield the game Preference score; a positive score indicates greater preference for Sushi-Go-Round, whereas negative score indicates greater preference for Tank Attack 3D.

We have sufficient power to not only examine correlations between two variables but also conduct multiple regressions required to evaluate the cognition–learning, experience–learning, and cognition–experience–learning relationships. All estimations for power analyses were conducted using G^∗power3.1 (Faul et al., 2009). Our sample size of 107 participants provides us with a power greater than 0.95 to detect an effect size of 0.5 in a regression analysis with only one predictor, and a power greater than 0.92 to detect the same effect size in a multiple regression analysis with three predictors (e.g., three cognitive variables as predictors of game learning). We also have a power greater than 0.8 to detect an effect size of 0.5 in a multiple regression analysis with all seven regressors of interest in this study, that is, three cognitive variables (FSpan, BSpan, SDST), three game habit variables (Identification, Duration, Frequency), and one game preference variable.

Table 1 reports means and standard deviations of responses to the game habit variables (Frequency, Duration, Identification) as well as means and standard deviations of the cognitive assessments administered (FSpan, BSpan, SDST) and of game Preference. Notably, the primary Preference measure (relative preference of Sushi-Go-Round over Tank Attack, S-A) indicated that participants preferred Tank Attack 3D (action game) over Sushi-Go-Round (strategy game), mean S-A = −0.29, SD = 1.72. A paired-samples t-test comparing reported preference for each of these games found this difference to be significant, t(106) = 2.26, p = 0.03.

We ran a series of bivariate correlations, using Pearson’s method, between the variables of game habits, cognition, and novel game learning, as well as Preference, to establish baseline relationships between these measures. The results of these correlations are presented in Figure 2. The game habit variables (Frequency, Duration, and Identification) were significantly inter-correlated: Frequency–Duration, r(106) = 0.7, p < 0.01; Frequency–Identification, r(106) = 0.68, p < 0.01; Duration–Identification, r(106) = 0.74, p < 0.01. Similarly, both Action LC and Strategy LC were significantly correlated, r(106) = 0.78, p < 0.01. These bivariate correlations suggest that the multiple measures of the two constructs, game habit and novel game learning, were strongly inter-correlated and were reliable measures of these constructs. For cognition, FSpan and BSpan were significantly correlated, r(106) = 0.47, p < 0.01, but neither was significantly correlated with SDST, [FSpan r(106) = −0.06, p = 0.6; BSpan r(106) = 0.13, p = 0.21].

Regarding the habit–learning relationship, all measures of game experience (Frequency, Duration, and Identification) were significantly and positively correlated with both Action LC [Frequency, r(106) = 0.67, p < 0.01; Duration, r(106) = 0.59, p < 0.01; Identification, r(106) = 0.62, p < 0.01] and Strategy LC [Frequency, r(106) = 0.53, p < 0.01; Duration, r(106) = 0.5, p < 0.01; Identification, r(106) = 0.46, p < 0.01]. Regarding habit–cognition relationships, Duration was significantly correlated with FSpan, r(108) = 0.23, p = 0.03, as was Identification, r(106) = 0.21, p = 0.04. Identification was additionally significantly correlated with b BSpan, r(106) = 0.22, p = 0.03. These results suggest that greater gaming experience was related to better learning of novel games, better working memory, and to a lesser degree, better short-term memory.

Preference was significantly and positively related to both Duration, r(106) = 0.21, p = 0.03, and Identification, r(106) = 0.25, p = 0.01, but not with Frequency, r(106) = 0.17, p = 0.08, indicating that participants who reported a higher preference for the action game than the strategy game tended to also report, on average, more hours of gameplay per week and more strongly identified as “gamers.” Preference was not related to either of the two memory span measures [FSpan, r(106) = 0.08, p = 0.48; BSpan, r(106) = 0.08, p = 0.43], nor to either LC [Action LC, r(106) = 0.1, p = 0.29; Strategy LC, r(106) = −0.11, p = 0.26]. SDST was, however, significantly but negatively correlated with preference of the action game over the strategy game, r(106) = −0.27, p = 0.01. This result suggests that individual differences in greater preference for the strategy game over the action game was related to better performance on SDST (a measure of processing speed), although preference for one type of game over another was not impacted by individual differences in memory spans, either FSpan or BSpan. The results from the effects of Preference for a specific game on learning of the two games are reported in detail in section “Additional Analyses: Effect of Game Preference on Novel Game Learning.”

In order to investigate the potential influence of demographic variables on novel game learning, we first analyzed the relationships of participants’ age and gender with the other variables (game habits, preference, cognition, and learning).

Bivariate correlations using Pearson’s method indicated that age was inversely correlated with Frequency, r(106) = −0.48, p < 0.01, Duration, r(106) = −0.3, p < 0.01, and Identification, r(106) = −0.32, p < 0.01, as well as LCs of both action, r(106) = 0.75, p < 0.01, and strategy, r(106) = −0.73, p < 0.01, games. Age was not significantly correlated with Preference, r(106) = 0.09, p = 0.36. For cognition, age was inversely correlated with working memory capacity (BSpan), r(106) = −0.2, p = 0.06, and processing speed (SDST), r(106) = −0.65, p < 0.01, but not with STM (FSpan), r(106) = −0.14, p = 0.19 (Figure 1A).

Considering the asymmetry of the age distribution of our sample, we next fitted linear models to each of the age-to-habit, age-to-cognition, and age-to-learning relationships to establish if these relationships are linear and, therefore, can be subjected to linear regressions in subsequent analyses. Regarding habit variables, the relationship between age and Frequency, F(1/105) = 31.61, p < 0.01, R² = 0.23, age and Duration, F(1/105) = 13.8, p < 0.01, R² = 0.11, and age and Identification, F(1/105) = 14.14, p < 0.01, R² = 0.12, was significantly linear. Age and Preference demonstrated no such relationship, linear or nonlinear, F(1/105) = 0.87, p = 0.35, R² = 0.01. Regarding Cognition, age and BSpan, F(1/105) = 7.03, p = 0.01, R² = 0.06, and age and SDST, F(1/105) = 117.67, p < 0.01, R² = 0.52, demonstrated significant linearity (Figures 3A,C), but age and FSpan displayed no significant relationship (Figure 3A), F(1/105) = 0.71, p = 0.4, R² = 0.01. Age also demonstrated a linear relationship with both the Action LC, F(1/105) = 138.17, p < 0.01, R² = 0.57, and the Strategy LC, F(1/105) = 124.11, p < 0.01, R² = 0.54 (Figure 3B).

Additionally, we ran a series of independent sample T-tests comparing male and female participants on habits, preference, learning, and cognitive variables. No significant gender differences were observed in these comparisons (Table 2).

The primary intent of this study is to examine the relationship between past gaming habits, cognitive abilities, game preference, and the learning of novel games, across life span, irrespective of age and gender differences. Although these variables are correlated (see section “Relationships Between Cognition, Game Habits, Game Preference, and Novel Game Learning”), it not clear to what extent these relationships are driven by individual differences in age. Therefore, we conducted correlations between these variables, after controlling for age and gender. Action LC was still significantly correlated with Frequency, r(104) = 0.49, p < 0.01, Duration, r(104) = 0.52, p < 0.01, and Identification, r(104) = 0.58, p < 0.01. Likewise, Strategy LC was correlated with Frequency, r(106) = 0.28, p = 0.04, Duration, r(104) = 0.41, p < 0.01, and Identification, r(104) = 0.33, p < 0.02 (Figure 1B). Preference was now marginally correlated with Action LC, r(104) = 0.25, p = 0.07, but not with Strategy LC, r(104) = −0.08, p = 0.59, after controlling for age and gender. None of the cognitive variables examined were found to be correlated with either Action LC [FSpan: r(104) = 0.09, p = 0.61; BSpan: r(104) = 0.18, p = 0.29; SDST: r(104) = 0.03, p = 0.87] or Strategy LC [FSpan: r(104) = 0.21, p = 0.22; BSpan: r(104) = 0.23, p = 0.16; SDST: r(104) = 0.15, p = 0.35], after controlling for age and gender.

To determine if the relationships between novel game learning and game habits, preference, and cognition are linear, we fit linear models to these relationships. The linear relationships between Action LC and Frequency, F(1/105) = 70.09, p < 0.01, R² = 0.4, Duration, F(1/105) = 63.92, p < 0.01, R² = 0.38, Identification, F(1/105) = 63.59, p < 0.01. R² = 0.37, BSpan, F(1/105) = 11, p < 0.01, R² = 0.09, and SDST, F(1/105) = 47.7, p < 0.01, R² = 0.31, were significant. The linear relationships between Action LC and Preference, F(1/105), p = 0.33, R² = 0.01, and between Action LC and FSpan, F(1/105) = 2.33, p = 0.12, R² = 0.02, were not significant. The relationships between the Sushi LC and variables of game habit and cognition had similar patterns [Frequency: F(1/105) = 41.71, p < 0.01, R² = 0.28; Duration: F(1/105) = 40.91, p < 0.01, R² = 0.28; Identification: F(1/105) = 30.24, p < 0.01, R² = 0.22; BSpan: F(1/105) = 10.8, p < 0.01, R² = 0.09; SDST: F(1/105) = 62.36, p < 0.01, R² = 0.37; Preference: F(1/105) = 1.32, p = 0.25, R² = 0.01; FSpan: F(1/105) = 3.18, p = 0.08, R² = 0.03].

These results above demonstrate that measures of novel game learning in both genres are related individually to measures of game experience, over and beyond the effects of age and gender. However, they do not reflect the potentially interrelated effects of our variables of interest on novel game learning. To investigate this possibility, we next performed a series of multiple regressions to examine the combined effect of the collected gaming experience and cognition on the game LCs. Two sets of stepwise multiple regressions were conducted, one set using Action LC and another set using Strategy LC as the dependent variable. In both stepwise multiple regressions, age and gender were entered as control variables in the first step. In the second, game experience variables (Frequency, Duration, Identification), and cognitive variables (FSpan, BSpan, SDST) were entered, which resulted in a handful of significant predictors. The results of these analyses are presented in Table 3. The final regression model for Action LC, R² = 0.67, F(4,102) = 41.9, p < 0.01, included Identification, β = 0.15, t(102) = 2.78, p < 0.01, and Duration, β = 0.11, t(102) = 2.46, p = 0.02, as significant predictors. The final regression model for Strategy LC, R² = 0.57 F(4,102) = 27.67, p < 0.01, included Duration, β = 0.21, t(102) = 4.79, p < 0.01, and BSpan, β = 0.15 t(102) = 2.58, p = 0.01, as significant predictors.

The above multiple regressions indicate that gaming duration is a common predictor of both action and strategy learning, but this result can be driven by the inter-correlation between the Action and Strategy LCs. A correlation analysis indeed demonstrated that Action LC and Strategy LC were significantly related, r(106) = 0.78, p < 0.01. To account for this inter-correlation, two sets of stepwise multiple regression analyses were conducted, such that the effects of gaming experience and cognition on either game learning could be determined, over and beyond the effects of the other game learning.

As in previous regressions, the first step included age and gender, and the second step included the other game LC (that is, Strategy LC when Action LC was the dependent variable, and Action LC when Strategy LC was the dependent variable). For the third step, only the significant predictors of earlier regression analyses (see Table 3) for a specific game learning were used to determine their specificity on that game learning. For example, for Action LC, Identification and Duration were entered in the third step. For Strategy LC, BSpan and Duration were entered in the third step. The results of these analyses are reported in Table 4.

Identification remained a significant predictor of Action LC even after accounting for Strategy LC, β = 0.16, t(101) = 3.16, p < 0.01, but Duration was no more a significant predictor, β = 0.05, t(101) = 1.13, p = 0.26. However, Duration remained a significant predictor of Strategy LC even after accounting for Action LC, β = 0.14, t(101) = 2.83, p < 0.01, with BSpan still being a marginally significant predictor of Strategy LC, β = 0.11, t(101) = 1.94, p = 0.056.

So far, our analyses have focused on the effects of game habits, preference, and cognition on novel game learning, irrespective of age. However, it is plausible that the pattern of these relationships varies across the adult life span. We therefore performed moderator analyses using multiple regressions to account for the effects of age on these relationships. Gender was introduced in the first step of these multiple regressions. Age and the significant predictors demonstrated previously (Identification and Duration for the Action LC, Duration and BSpan for the Strategy LC) were entered in the second step. The third step included the moderation terms that accounted for the interaction between age and the two variables, which were entered in the second step. The results of these analyses are reported in Table 5.

While age was demonstrated to be a significant predictor of both Action LC, β = −0.03, t(100) = −6.61, p < 0.01, and Strategy LC, β = −0.03, t(100) = −0.4.18, p = 0.01, none of the interaction terms with Age demonstrated significance in either analyses, and neither of the regression models in the third step demonstrably improved model fit compared to the previous step, indicating that age is not a meaningful moderator of the previously observed relationships.

In addition to the general game experience/habits data collected, we also gathered information from each participant regarding their frequency of play with eight specific genres of video games. We examined the potential relationship between experience on a specific game genre and novel game learning through correlation analyses that compared self-rated frequency of playing of these genres with Action and Strategy LCs. These correlations were controlled for individual differences in age and gender. The results are displayed in Figure 1C.

Action LC was found to be positively correlated with frequent experience with FPS games, r(103) = 0.42, p < 0.01, RTS games, r(103) = 0.29, p < 0.01, and RPGs, r(103) = 0.51, p < 0.01. Strategy LC was positively correlated with frequent experience with RPG, r(103) = 0.32, p = 0.01, and casual games, r(103) = 0.25, p = 0.01 Additionally, Strategy LC was inversely correlated with frequent experience with sports video games, r(103) = −0.2, p = 0.04.

To further examine the impact of game preference on novel game learning, we next divided our participants into subgroups based on self-reported preference for the two games utilized in this task. Subjects were divided into three groups based on their reported enjoyment of Tank Attack 3D and Sushi-Go-Round. The Prefer Tank group (n = 34, 10 females, M_Age = 29.79, SD_Age = 18.14) reported higher enjoyment of Tank Attack 3D compared to Sushi-go-Round. The Prefer Sushi (n = 46, 17 females, M_Age = 25.89, SD_Age = 11.53) group reported higher enjoyment of Sushi-go-Round than Tank Attack 3D. The No Preference (n = 27, 13 females, M_Age = 55.79, SD_Age = 20.96) group reported identical levels of enjoyment of both games.

We first assessed how these preference groups differed in regard to the game LCs. A one-way ANOVA demonstrated that preference group had a significant effect on Tank LC, F(2,104) = 29.86, p < 0.01. Post hoc comparisons using the Sidak method demonstrated that Tank LC was significantly higher for the Prefer Tank group compared to the No Preference group (p < 0.01) but did not differ significantly from the Prefer Sushi group (p = 0.36). A second one-way ANOVA demonstrated that preference group had a significant effect on Sushi LC as well, F(2,104) = 20.71, p < 0.01. Post hoc comparisons using the Sidak method demonstrated that Sushi LC was significantly higher for the Prefer Sushi group compared to the No Preference group (p < 0.01) but did not differ significantly from the Prefer Tank group (p = 0.26). These results indicate that individuals with a preference for either game demonstrated greater LCs on both the action and strategy games utilized. Therefore, greater preference for a specific game was not related to better learning of that game.

Considering the apparent disparity in age between the three preference groups as reported above, we next compared participant age across these three groups. A one-way ANOVA indeed demonstrated a significant difference in age between the three groups, F(2,104) = 31.06, p < 0.01. Post hoc comparisons using the Sidak method demonstrated that the No Preference group was significantly older than both the Prefer Tank group (p < 0.01) and the Prefer Sushi group (p < 0.01), but there was no significant age difference between the Prefer Tank and Prefer Sushi groups (p = 0.65). These results suggest that the differences in relative learning rates between the preference groups and the no-preference group are possibly driven by participant age.