Video gaming and working memory: a large-scale cross-sectional correlative study

Studies have indicated that video gaming is positively associated with cognitive performance in select cognitive domains, but the magnitudes of these associations have been called into question, as they have frequently been based on extreme groups analyses that have compared video gamers with non-gamers. When including the whole range of participants, and not just extreme cases, these effects were observed to reduce markedly (Unsworth et al., 2015). To further study this issue, we compared the associations between video gaming and aspects of working memory (WM) performance in an extreme groups design to those of a design that includes the full range of participants in a large adult sample (n = 503). WM was measured with three composite scores (verbal WM, visuospatial WM, n-back). The extreme groups analyses showed that video gamers performed better than non-gamers on all three WM measures, while the whole sample analyses indicated weak positive associations between the time spent playing video games and visuospatial WM and n-back performance. Thus, study design modulated the effects, but two of the three associations between WM and video gaming were consistent across both analysis techniques. A separate study confirmed that our questionnaire-based estimate of gaming hours was reliable when compared with one-week diaries of videogame playing. While the present cross-sectional results preclude causal inferences, possible mechanisms of WM - videogame playing associations and future research directions are discussed. Overall, our results indicate that cognition - videogame playing relationships, albeit weak, are not solely due to recently discussed methodological artefacts concerning the particular analytical approach and survey reliability.

Introduction

With the growing popularity of video games, there has been an increasing interest in investigating whether and how video gaming (the habit of playing video or computer games) is related to cognitive abilities. One particularly interesting cognitive domain in this respect is working memory (WM), a short-term memory system that is involved in the maintenance and processing of currently active mental information (e.g., Cowan, 2014). Its importance lies in its key position in human cognition: WM is considered to be engaged in every conscious thought process, and it is associated with many important skills and outcomes such as mathematical achievement (Bull, Andrews Espy, & Wiebe, 2008) and fluid intelligence (Conway, MacNamara, & Engel de Abreu, 2013). Moreover, the possibility to use video gaming as a form of cognitive training has also garnished significant interest (Bediou et al., 2018; Sala, Tatlidil, & Gobet, 2018). A potential mechanism for cognitive enhancement through video gaming is provided by the so-called core training hypothesis, according to which repeated strain on a cognitive system induces plastic changes in its neural substrates and thereby leads to performance improvements (Anguera et al., 2013). Another proposed underlying mechanism is learning to learn, according to which video gaming (especially of the action game genre, see below) improves skills such as rule learning, cognitive resource allocation, and probabilistic inference that are used in many different situations (Bavelier, Green, Pouget, & Schrater, 2012).

Several previous cross-sectional studies that have investigated associations between cognition and video gaming suggest that video gamers perform better in various cognitive domains as compared to non-video gamers. Examples of such performance advantages include WM updating as measured by the n-back task (Colzato, van den Wildenberg, Zmigrod, & Hommel, 2013; Moisala et al., 2017), action cascading (i.e., goal-directed multi-component behavior) (Steenbergen, Sellaro, Stock, Beste, & Colzato, 2015), encoding speed of visual information into short-term memory (Wilms, Petersen, & Vangkilde, 2013), visual change detection (Clark, Fleck, & Mitroff, 2011), and multisensory perception and integration (Donohue, Woldorff, & Mitroff, 2010). Narrative reviews indicate that gamers show advantages especially in visuospatial aspects of cognition (Hubert-Wallander, Green, & Bavelier, 2010; Oei & Patterson, 2014), and quantitative meta-analyses have supported these conclusions to some extent (Powers, Brooks, Aldrich, Palladino, & Alfieri, 2013; Bediou, Adams, Mayer, Tipton, Green, & Bavelier, 2018; Sala et al., 2018). It is important to note that the different video games and game genres are not seen as equals in their cognitive demands and in their expected skill-transfer outcomes. Many previous studies have compared action video gamers with non-gamers due to the perceived demanding nature of action video games. Action video games have been described to be fast-paced, to set high perceptual, motor, and cognitive demands, to emphasize peripheral vision and divided attention, and to require constant predictions of future game events (Green & Bavelier, 2012). Action games are typically exemplified by the first-person shooter genre that includes games such as Halo, Doom, and Call of Duty.

Recently, the study designs used to quantify the associations between video gaming and cognition have raised discussion. Unsworth, Redick, McMillan, Hambrick, Kane, and Engle (2015) reported two experiments where they examined the cognitive advantages associated with video gaming with an extreme groups design (group comparison) vs. a whole-group design (regression analysis). When the latter method was used, many of the advantages seen in the extreme groups design were only weak or disappeared completely. This led Unsworth et al. (2015) to argue that the effects in previous meta-analyses have been overestimated because they have included a significant amount of studies comparing extreme groups with small samples and increased likelihood of Type 1 errors (see also Boot, Blakely, & Simons, 2011). Comparing two extreme groups can involve problems such as magnifying minor results and losing important information from the middle of the distribution (Preacher, Rucker, MacCallum, & Nicewander, 2005). However, Green et al., (2017) criticized the videogaming questionnaires used by Unsworth et al. (2015), in which the participants were to estimate the number of playing hours per week per game type. According to Green et al. (2017) this leads to unreliable estimates (especially if participants play multiple game types) that are suboptimal at measuring finer gradations of behavior that are relevant for whole-group analyses (for a response, see Redick, Unsworth, Kane, & Hambrick, 2017).

Given the controversy on the adequacy of the study designs and videogaming questionnaires employed in this field, further research on the relationships between videogaming and cognition is warranted. Here we attempted a systematic replication of the study by Unsworth et al. (2015; see also Redick et al., 2017) by using a somewhat different videogaming questionnaire that we also validated in a separate videogaming diary study. To avoid potential sample- or task-specific confounds, we compared the outcomes of extreme groups vs. whole-group designs with a single large sample by using WM composites derived from a factor analysis of multiple WM tasks performed by the same basic sample (Waris et al., 2017).

In Study 1, we examined cross-sectional associations between WM performance and common video gaming habits in a large adult sample (n = 503), and compared the outcomes of extreme groups analyses (video gamers vs. non-gamers 1 ) with whole-sample analyses, both conducted within the same sample. Our self-report video gaming questionnaire had participants estimate their total playing time in hours per week, as well as the percentage of playing time devoted to a set of game types. This differed from Unsworth et al. (2015) and Green et al. (2017), who asked participants to estimate how many hours they spent playing games within certain genres (i.e., estimated genre by genre). A secondary aim was to investigate whether we would observe differential associations between visuospatial vs. numerical-verbal WM and video gaming, as spatial cognition has been shown to be more strongly related to video gaming than verbal cognition (Bediou et al., 2018; Sala et al., 2018). When considering our three WM composites (numerical-verbal WM, visuospatial WM, n-back), the core training hypothesis would predict a stronger association between visuospatial WM and video gaming than for verbal WM because video games are a visually dominant media and the cognitive demands of video games are more pronounced in the visual domain (e.g., short-term maintenance of object/enemy/etc. locations that are rapidly updated, tracking multiple moving objects, predicting trajectories, navigating in a virtual world etc.). In the same vein, n-back tasks that require flexible updating of the WM contents should also show a stronger association with video gaming than verbal WM. The learning-to-learn hypothesis would further predict a more general advantage that is not related to any specific domain, but rather depends on how well the skills hypothetically boosted by video gaming can be implemented in a specific WM task. However, it is also important to point out that self-selection or task-specific learning could explain any observed cross-sectional associations.

Furthermore, in Study 2 we evaluated the reliability of our questionnaire on video gamer’s self-estimated video gaming time. As noted above, the reliability of video gamers’ selfevaluations and thereby the whole-group analysis approach on video gaming - cognition relationships has been called into question (Green et al., 2017). There is surprisingly little empirical evidence on the reliability of self-estimated video gaming time, but thus far the results have been rather disheartening. Greenberg et al. (2005) reported correlations of only .207 (offline gaming hours) and .289 (online gaming hours) between self-reported estimated playing time and diary-reports (diary-reports refers to participants keeping track of their gaming time for a specific time period). However, their estimates and diaries only encompassed single separate days, which makes the measures susceptible to significant temporary variation. Furthermore, Greenberg et al. (2005) did not evaluate the accuracy of the diaries in any way. Similarly, Kahn, Ratan, and Williams (2014) found a correlation of only .365 between estimates of weekly playing time on a single game (EverQuest II) with game log data on playing time. However, as pointed out by the authors, the game log data encompassed the entire existence of the player’s game account while the survey question was vague on the time frame (i.e., how many hours the person usually plays), which could result in higher discrepancy if the player had changed playing habits. Furthermore, the game log data, which counted every second a player was logged on to their account, did not apparently distinguish between actual playing time and time when the player was away from the keyboard. This could account for a portion of the discrepancy as players in these types of games (so-called massively multiplayer online games) can gain substantial in-game benefits without actively playing (e.g., keeping the game running over a night) by using automated commands (Ducheneaut, Moore, & Nickell, 2007). Considering the scarcity of research on the topic, we therefore correlated video gamers’ estimates of weekly video gaming time in a self-report questionnaire with their diaries of gaming time during one week. The second study also allowed us to test if the number of genres a person plays affects their estimates of time spent playing video games. This point of criticism was made by Green et al. (2017), who reported that subdividing the estimated playing time on multiple genres lead to overestimations of playing time.

Study 1

This study tested whether video gaming was associated with higher WM performance, and whether the possible WM advantages would weaken or dissipate when moving from extreme groups analysis (video gamers vs. non-gamers) to whole-sample analysis (Unsworth et al., 2015). To increase the reliability of our study, we conducted the two analyses within a single sample and employed WM composite measures (rather than single WM task scores) that were derived from a latent variable analysis (Waris et al., 2017).

Methods

Ethical statement

The study was approved by the Joint Ethics Committee at the Departments of Psychology and Logopedics, Abo Akademi University, and by the Human Research Review Board at the University of California, Riverside. Informed consent was obtained from all participants, participation was anonymous, and all participants were informed of their right to withdraw from the study at any time.

Procedure

Every aspect of the study was completed online. Participants were recruited via the crowdsourcing site Amazon Mechanical Turk, and an extensive background questionnaire (concerning, e.g., video gaming habit, medical history, and education) and all WM tasks were administered with an in-house developed web-based test platform. Participants were paid $10 US for completing the entire study, which took 1.5h on average to complete. The background questionnaire was completed first, followed by ten WM tasks (see below). The order of the WM tasks was randomized for every participant with one exception: the forward simple span was always followed by the respective simple span backward.

Participants

711 participants completed the entire study. We excluded 38 participants as they reported using external tools (such as note-taking) to help them solve one or more of the WM tasks. Four participants were excluded for having missing values on one or more of the WM measures and one participant was excluded for taking more than one day to complete the entire study. Next, in order to minimize the effect that depression possibly plays in WM performance (Christopher & MacDonald, 2005; Harvey et al., 2004; Rose & Ebmeier, 2006), we excluded 136 participants who reported a depression score that corresponded to moderate, severe, or very severe depressive symptoms according to the QIDS (Quick Inventory of Depressive symptoms, Rush et al., 2003). Sixteen participants were additionally excluded for having missing depression scale data. Finally, 13 participants were excluded for being multivariate outliers on the WM task variables according to Mahalanobis distance. This gave us a final sample of 503 participants. The mean age of this sample was 34.2 years (SD = 10.6, range: 18-71); the gender distribution was 56.5% female, 43.3% male, and 0.2% other; 53.7% of the sample reported having a Bachelor’s or Master’s degree; and the mean estimated household wealth during childhood was 3.88 (on a scale from 1, very poor, to 7, very wealthy; see Waris, Soveri, Lukasik, Lehtonen, & Laine, 2018). To test whether possible prior experience with the WM tasks used here was not an issue in our sample, we ran independent samples t-tests where we compared the WM composite score performances (verbal, visuospatial, n-back, see below) of those who post assessment reported any prior experience with similar tasks (n = 81) with those who reported no prior experience (n = 422). All t-tests were non-significant (verbal WM, t(501)=1.27, p = .204; visuospatial WM, t(501)= 0.05, p = .962; n-back: t(501)=0.03, p = .973), which indicates that this was not an issue in our sample.

Working memory measures

The WM measures are only briefly described here, as further details are provided in Waris et al. (2017). We assessed WM with ten separate tasks that involved four different task paradigms. The task paradigms were simple span (both forward and backward), complex span, running memory span, and n-back. There was one numerical-verbal and one visuospatial variant of each task paradigm (hence 5×2 = 10 tasks). The numerical-verbal tasks used the digits 1-9 as stimuli, while the visuospatial tasks used locations in a 3×3 grid. For the complex span tasks, the distracting items that were placed between every to-be-remembered item consisted of simple arithmetic problems in the verbal task and mental combination of partially filled matrices in the visuospatial task. Accuracy rates were used as outcome measures in all tasks. For the simple, complex, and running memory spans, the number of correctly recalled individual items was used as the outcome variable. For the n-back tasks, the outcome variable was the corrected recognition score, that is, the total number of hits (correct targets) minus the total number of false alarms (no-targets that were incorrectly selected as targets).

We first Box-Cox transformed the WM measures to better approximate normal distributions (Osborne, 2010). Then, following the exploratory factor analysis results of Waris et al. (2017), three WM composite scores were created by summing and then averaging the respective z-transformed WM measures. The composites were (1) numerical-verbal WM that consisted of the verbal simple, complex, and running memory spans, (2) visuospatial WM that consisted of the visuospatial simple, complex, and running memory spans, and (3) n-back that consisted of the verbal and visuospatial n-back tasks. The first two composites allowed us to specifically assess numerical-verbal and visuospatial WM. The Box-Cox and z-transformations were done separately for the whole sample and extreme groups sample.

Video gaming

To assess video gaming habits, the participants were asked whether they regularly played computer, console, or similar games, and how many hours they played on average per week. Additionally, they were asked to evaluate percentage-wise how much they played nine separate game types and whether they typically played it alone (single player) or with others (multiplayer) (see Table 1 ). The game types were: (1) Card, (2) Mobile, (3) Action, (4) Shooter/First person shooter, (5) Exercise, music, and party, (6) Adventure, puzzle, and role-playing, (7) Simulation, (8) Strategy, and (9) Brain training and educational (see Table 2 ). We note that these categories are only a rough approximation of the activities performed across individual video game play (e.g. different playing styles, diversity of game-types in a genre, etc., see Dale & Green, 2017) and that a detailed comparison of aspects of game play and WM performance are beyond the scope of the present manuscript.

Table 1

Video gaming items and response alternatives in the background questionnaire.