# Computer Gaming and the Gender Math Gap: Cross-Country Evidence among Teenagers ☆

ISBN: 978-1-78756-462-6, eISBN: 978-1-78756-461-9

ISSN: 0147-9121

Publication date: 6 August 2018

## Abstract

Using the Program for International Student Assessment (PISA) surveys (2003–2015), this chapter explores the relationship between the gender gap in math test scores and computer (digital devices) gaming, as a potential “swimming upstream” factor in the quest to close that gap. Using a decomposition based on a pooled hybrid specification, we attribute two to three points (from 13% to 29%) of the gender math gap to gender differences in the incidence and returns to intense gaming. The comparison of the negative versus positive girl-specific effects found for collaborative games versus single-player games suggest a potential role for gaming network effects.

## Keywords

#### Citation

Algan, Y. and Fortin, N.M. (2018), "Computer Gaming and the Gender Math Gap: Cross-Country Evidence among Teenagers ", Transitions through the Labor Market (Research in Labor Economics, Vol. 46), Emerald Publishing Limited, pp. 183-228. https://doi.org/10.1108/S0147-912120180000046006

### Publisher

:

Emerald Publishing Limited

## 1. Introduction

Increasing the use of information and communication technologies (ICT) has long been part of public policies aimed at fostering economic growth, especially in middle-income (MI) countries. 1 In higher-income (HI) countries, breaking the digital divide is seen as essential to increasing human capital acquisition among low-income groups (Goolsbee & Guryan, 2006; Machin, McNally, & Silva, 2007). But it also increases other computer uses, such as computer gaming with possible unintended consequences. The latter might explain the mixed results found in the substantial economic literature on the impact of general computer use on human capital enhancement (Angrist & Lavy, 2002; Beuermann, Cristia, Cruz-Aguayo, Cueto, & Malamud, 2013; Fairlie & London, 2012; Fairlie & Robinson, 2013; Malamud & Pop-Eleches, 2011; Vigdor, Ladd, & Martinez, 2014).

Fuchs and Wößmann (2004) who study the use of computers at home and at school on student learning using Program for International Student Assessment (PISA) 2000 find that the conditional relationship between student achievement, and computer and Internet use at school has an inverted U-shape. They conjecture that the availability of computers at home seems to distract students from effective learning. However, Spiezia (2010) who uses PISA 2006 finds more positive effects for computer use at home than at school. Biagi and Loi (2013) who use PISA 2009 find that students’ test scores increase with the intensity of computer use for gaming activities while they decrease for activities more related to school curricula. To capture the increasing trends in computer (or digital devices) gaming, this chapter exploits up to six waves of the PISA surveys from 2000 to 2015. 2

### Fig. 1.

Percentage of PISA Students by Gender Who Use a Computer Outside of School for Computer Gaming (A and B) or Music Downloading (C and D).

In addition to the substantial gender divide in intense computer gaming, there is a persistent gender gap in math test scores that ranges from 8 to 17 points on the normalized (in each wave) average of 500 (2–3% discrepancy). Fig. 2 displays the male advantage in math test scores averaged across all countries participating in the ICT supplement, also separating MI and HI countries. 5 Because of potential positive effects of gaming on visual-spatial and attention skills (Green & Bavelier, 2003, 2012), problem-solving and strategic thinking (Adachi & Willoughby, 2013), we focus mainly on the relationship between everyday computer gaming and math test scores. 6

### Fig. 2.

Average Male Advantage in PISA Mathematics Test Scores by Country Income Status.

Many of the above papers on the impact of ICT use on academic performance exploit quasi-natural experiments or randomized control trials (RCT). 7 In Algan and Fortin (2016), we attempted to use an instrumental variable strategy based on Internet diffusion, seen an exogenous cause of Internet-enabled gaming among teenagers, to estimate the impact of everyday computer gaming on math test scores by gender. 8 This estimation strategy, however, yielded coefficients of everyday gaming on math scores that were imprecisely estimated. Although the significant point estimates, in line with those estimated on country averages, showed consistent negative effects for girls, but not so much for boys. At a minimum, the male minus female differences in the estimates were always positive, indicating that everyday gaming might provide a relative advantage to boys.

In the context of the gender math scores gap, computer gaming can be seen as a “swimming upstream” factor, that is a phenomenon which is increasing in prominence over time but where girls are at a disadvantage. Here, we show that this disadvantage arises not only from girls’ lower participation, but also from the negative girl-specific correlation between high-intensity gaming and math test scores. We provide estimates of the interaction between indicators for female and everyday computer gaming, along with base effects, on test scores to illustrate this finding. As explained below in terms of a pooled hybrid Oaxaca–Blinder (OB) decomposition, this implies that the role of gaming in accounting for the gender gap in math test scores arises from both composition and structural components. In this context, self-selection into intense gaming is a concern if the correlation between intense gaming and unobservables conditional on observed characteristics of male and female gamers is different. We discuss this concern using an argument à la Altonji, Elder, and Taber (2005) given our rich set of personal characteristics.

This “swimming upstream” factor is different from a typical explanatory factor. For example, the gender pay gap was anticipated to close as gender differences in labor market experience (an explanatory factor) narrowed over time (Mincer & Polachek, 1974), a pathway confirmed by Blau and Kahn (2017). A “swimming upstream” factor would be akin to increasing competitiveness in the labor market over time. 9 Because job competitiveness is associated with higher pay (Cortes & Pan, 2016), and because women on average display lower level of competitiveness than men (Croson & Gneezy, 2009) (or even when women act competitively, they are seen negatively), 10 then increasing competitiveness in the labor market will arguably impede improvements in female representation in top jobs, which slows the convergence between male and female average pay (Fortin, Bell, & Böhm, 2017).

These types of factors (competitiveness and computer gaming) are more specific but yet related to broader “cultural” factors such as gender norms (Fortin, 2005). Guiso, Monte, Sapienza, and Zingales (2008) and Fryer and Levitt (2010) used the PISA 2003 to show that, across countries, that the gender gap in math and reading test scores is positively correlated with indicators of gender equality in the society. De San Román and de la Rica Goiricelaya (2016) consider the impact of gender role attitudes, and its intergenerational transmission as measured by the mother’s labor force participation, on gender differences in PISA 2009 math scores. Nollenberger, Rodríguez-Planas, and Sevilla (2016) further investigate the role of culture across several PISA waves focusing on the gender gap in math scores among immigrants and showing that it can be linked to gender norms from their country of origin. Given that intense computer gaming is a highly gendered and popular activity, our study investigates a specific channel through which a gendered cultural trend can influence educational outcomes.

In fact, we focus on three potential mechanisms by which intense gaming (by comparison with lower intensities) could affect test scores: (1) a negative distracting effect linked to time spent gaming and rewards earned, (2) a positive effect on cognitive skills, and (3) social networking effects, likely positive for boys and negative for girls, as discussed below. 11 These mechanisms are explored using within-school-year models where we control for school climate factors, such as teachers and students hindrance to learning and sense of belonging. We can include a much richer set of contextual school variables than studies based on administrative data (e.g., Vigdor et al., 2014). These models are thought to capture the type of school demands that could prevent students from gaming everyday. In all specifications, we control for home computer and Internet access, as well as a host of individual, family, school, country, and ICT variables.

We discuss our findings on the differential effects by gender of the three mechanisms of interest below. For example, we find a positive interaction between single-player games (SPGs) and the female indicator, but a negative interaction with MMOs suggesting a role for gaming network effects, either directly or through negative selection. We view these suggestive findings on the social network effects of gaming as a novel contribution of the chapter. Our decompositions indicate that we can attribute two to three points (from 13% to 29%) of the gender gap in math test scores to gender differences in the incidence and the returns to intense computer gaming. Given the increasing trends in this highly gendered activity, we argue that it can be seen as a “swimming upstream” factor in the quest to close that gap.

The chapter is organized as follows. Section 2 discusses the sources of gender differences in computer gaming as well as their anticipated effects on test scores. Section 3 presents the data and some descriptive results. Section 4 introduces our pooled hybrid decomposition, discusses our empirical strategies, and reports the main results. Finally, Section 5 concludes.

## 2. Gender Differences and Anticipated Impact of Computer Gaming

Beyond use by adolescents, the advent of Internet-enabled gaming and music downloading and streaming has transformed the entertainment industry. The entertainment software industry (computer and video games) is poised to surpass the movies industry in economic might. The US Entertainment Software Association (ESA, 2017) boasted that total consumer spending in the gaming industry was to close $23.5 billion in 2015. The Association is preoccupied with female representation among gamers and reported that women represented 41% of gamers of all ages in 2017, down from the 45% of gamers in 2012. In PISA 2015, we find that girls represent 46% of computer gamers, down from 48% in PISA 2012. These relative incidence numbers do not reflect the fact that girls play computer games much less frequently than boys, as shown in Fig. 1. The literature investigating the sources of gender differences in computer/video gaming has identified four sets of explanations for these differences: They pertain to game content, the gaming community, gender differences in the social and competitive appeal of gaming, and gender differences in physiological responses. First, most game content presents a highly stereotypical view of women. Female game characters are under-represented, significantly more helpless and sexually provocative than male characters who are likely to be strong, more aggressive, and powerful (Ogletree & Drake, 2007). Males are more likely to be main characters and heroes, while females were more often supplemental characters, more sexy and innocent, and also wear more revealing clothing which likely attracts male players (Beasley & Collins Standley, 2002; Miller & Summers, 2007). A quick look at the free online portion of the most popular computer games of 2012 – Diablo III, Guild Wars 2, and World of Warcraft: Mists of Pandaria (according to ESA, 2013) – did indeed show well-endowed scantly clothed female characters. 12 A consequence of this portrayal of female characters has been that girl gamers often experience the gaming culture as secondary gamers (Schott & Horrell, 2000). Girl gamers do not extend their playing habits to engage in game play outside the home or participate to the same extent in the broader gaming culture. Some of the parents’ positive views about games as opportunities for their teens to “connect with friends” (ESA, 2013) may not apply to girls to the same extent. In addition to sexual gender stereotyping of game characters, female gamers disliked the lack of meaningful social interaction and the violent content of games, and are less engaged by the competitive elements of games (Hartmann & Klimmt, 2006). Women who have expressed discontent at the sexism in games and demanded a more balanced portrayal of males and females in games have been harassed by the online community, including anonymous threats of violence against these women (Kendall-Morwick, 2015). 13 Harassment leads female gamers to limit their social interaction online (chat online) which puts them a disadvantage in competition, particularly in strategic games that require teamwork and may result in negative reactions from fellow teammates (Richard, 2013). Both the social and competitive aspect of computer gaming deter female gamers. Also, differential physiological responses to reward intensity help explain the gender differences in engagement with gaming. Hoeft, Watson, Kesler, Bettinger, and Reiss (2008) find that compared to females, males showed greater activation and functional connectivity in their mesocorticolimbic system when playing computer/video games. Substantial gender differences emerge in reward prediction, learning reward values, and cognitive state during computer/video gaming. It is interesting to note that the prospect of monetary rewards in males also engages a broader network of mesolimbic brain regions compared to only limited activation for social rewards, while for females both types of rewards lead to similar activations (Spreckelmeyer et al., 2009). Hamlen (2010) finds that for boys, the increased play time leads to increased feelings of success and achievement, which then prompts more time playing; girls actually feel just as competent in their gaming ability, but lack the initial motivation for the rewards offered. The literature on gaming is also useful to help understand the potential link to math test scores. Although parents in the United States (ESA, 2013) impose time usage limits on computer/video games more than any other form of entertainment: 83% of parents place time limits on video game playing versus 76% of parents on television viewing. A vast majority of parents (71%) believe that game play provides mental stimulation and education to their children. In psychology, the vast majority of research on the effects of gaming has focused on potential negative impacts: the potential harm related to violence, addiction, and depression. Papers (Rehbein, Psych, Kleimann, Mediasci, & Möße, 2010; Skoric, Teo, & Neo, 2009) attempting to link these potential effects to academic achievement found that addiction (or clinical dependency) is indeed associated with lower academic achievement. In Rehbein et al. (2010), the percentage of German teenagers showing a dependency to gaming is relatively low (3% among boys and 0.3% among girls). 14 There are also many psychological studies focusing on the positive cognitive aspects of gaming, such improvements in visual-spatial and attention skills and in problem-solving skills (e.g., Green & Bavelier, 2003, 2012). Adachi and Willoughby (2013) in a longitudinal analysis contrasting the use of more or less strategic video game play among teenagers find that higher frequency of strategic video game play is predictive of greater self-reported problem-solving skills. Jackson, Von Eye, Witt, Zhao, and Fitzgerald (2011) report more nuanced correlational findings: for youth with low initial levels of visual-spatial skills, playing video games facilitates the development of visual-spatial skills. Increases in visual-spatial skills are found to be correlated with mathematical skills, but not with average grades in school. In summary, the existing literature on the effects of gaming confirms that our emphasis on intense gaming and on gender differences in gaming is well-placed. Gender differences arise along three of the five dimensions outlined by Gentile (2011) to evaluate the impacts of games on youth: amount of play, game content, and context of play. Previous research also provides clues on the mechanisms to explore. The negative effects of intense gaming would arise from distraction/displacement effects associated with a lot of time spent playing and with reward systems that may be more immediate than incremental improvements in grades. Some positive effects of gaming would likely arise through the enhancement of visual-spatial and problem-solving skills which, in turn, could improve math scores, and through increased social interactions, more so for boys than for girls. For girls, additional negative effects may arise from the sexualized portrayal of female characters which would undermine their confidence in mathematics in ways similar to those found in the stereotype treat literature (e.g., Spencer, Steele, & Quinn, 1999). ## 3. Data and Basic Facts In this section, we first present the student, school, country, and ICT data used. We include some data visualizations that illustrate differences in trends between HI and MI countries. Next, we turn to student and family variables to compare how everyday gamers and other students might differ across observables between HI/MI countries. Finally, we discuss the type of school-level measures available, and the school climate factors important to educators. ### 3.1. Data Sets Few data sets contain information on both standardized tests scores and gaming habits. These include surveys from the PISA targeting 15-year-olds and conducted by the Organization for Economic Cooperation and Development (OECD), as well as the Trends in International Mathematics and Science Study (TIMMS) and Progress in International Reading Literacy Study (PIRLS) administered to 4th- and 8th-grade students by the International Association for the Evaluation of Educational Achievement (IEA). Here, we focus on the PISA surveys because of a broader coverage of countries over more years, noting that only a subset of countries (listed in Table AI) participates in the ICT supplement. Table AI also indicates whether a country is HI or MI according to the UN (2012) country classification. 15 We also prefer to focus on adolescents because they are more likely attracted to popular MMOs and more in control of their time use than children. However, the PISA does not have good time use data to illustrate the intensity of daily gaming. To discuss the issue of time substitution, we turn to the TIMMS. Eight graders in TIMMS 2008 were asked “How many hours a day” they devote to several activities using five categories. For “playing computer games” comparing the percentages of boys and girls in these categories, we find 25% of boys versus 41% of girls spend no time at all, 25% versus 18% less than 1 hour, 24% versus 8% 1 to 2 hours, 13% versus 8% more than 2 but less than 4 hours, and 13% of boys versus 6% of girls play 4 or more hours. To summarize these numbers more concisely, we assign hours corresponding to the mid-point of the first four categories, and a conservative 6 hours to the “4 or more hours” category; we find that on average boys play computer games 2.6 hours a day versus 1.5 hours for girls. We use the same averaging scheme to compare the average number of hours dedicated to several activities in Table 1. They show that everyday gamers, those spending at least one hour a day playing computer gamers, generally enjoy more screen and Internet time, and more time with friends (possibly in the course of these activities) than nongamers. They also play more sports. Computer gaming only seems to take time away from household chores and homework, but relatively little. One important expandable activity, not included in the survey, is sleep time. Rehbein et al. (2010) do indeed mention that an essential indicator of problem gaming is “losing sleep over it,” which could lead teenagers to subsequently miss their morning classes, a question we explore below. Table 1. Average Number of Daily Hours Spent on Several Activities by Gender and Gaming Style. TIMMS 2008 − 8th Graders Boys Girls Activities Everyday Gamers Others Everyday Gamers Others Play or talk with friends 2.7 1.8 2.7 1.9 Watch TV and videos 2.6 1.5 2.7 1.7 Use the Internet 2.4 0.8 2.5 1.0 Play sports 2.3 1.9 1.4 1.1 Do homework 1.5 1.6 1.9 2.0 Do jobs or chores at home 1.2 1.3 1.3 1.5 Read a book for enjoyment 0.9 0.8 1.0 0.9 Work at a paid job 0.7 0.7 0.4 0.3 Notes: Everyday gamers are those who spend at least 1 hour a day gaming. The averages are computed as the percentage of boys and girls in each of five categories multiplied by the mid-point of the hours boundaries of each category: 0 for the no time, 0.5 for the less than 1 hour, 1.5 for the 1–2 hours, 3 hours for 2 but less than 4 hours, and 6 hours for the 4 or more hours. Our study uses data from the PISA 2000, 2003, 2006, 2009, 2012, and 2015 surveys, comprising 100,000–300,000 observations with nonmissing information per wave. The panel of countries is unbalanced with many developed countries not participating in the ICT supplements. Notably, France participated only 2015, but we have several waves of data from Germany, Canada, and Japan (the United States for 2000 and 2003). Our country-level control variables include the logarithm of GDP, the female labor force participation, youth male and youth female unemployment rates, all from the World Bank Indicators. 16 Female labor force participation (LFP) has been used in previous studies to capture the impact of social norms on the math gender differentials. Youth unemployment rates are meant to capture alternative uses of time, such as the relative un/availability of part-time work. Our ICT variables are retrieved from the International Telecommunications Union (ITU). They include the percentage of individual Internet users and the percentage of homes with a computer in each country in each year; these variables are thought to capture the density of potential gamers. Most PISA countries employ a two-stage stratified sampling technique to get a representative sample of the target population in each country. The first stage draws a random sample of schools in which 15-year-old students are enrolled and the second stage randomly targets 34 students (on average) in that age range in each school. 17 We do not know whether the same schools are sampled across the different waves. Thus we conduct within-school-year analyses; for these, we exclude students whose school peers were not interviewed. In the PISA scholastic assessment, tests on reading, mathematics, and science are divided into several item clusters, with each item cluster requiring 30 minutes of test time. Each student completes a subset of the clusters randomized across several booklets, and thus undergoes two hours of total testing divided by a five-minute break. For each test and each student, PISA reports five plausible random values drawn from the posterior distributions of the students’ scores (OECD, 2014). Because only a small subset of the universe of all students in each country is observed, PISA provides alternative sets of weights to capture this sampling error. Estimations have to be carried out separately for each one of the five plausible values using a set of 80 replicate weights that account for the two-stage design. 18 We use the “repest” procedure of Avvisati and Keslair (2017) in STATA to perform this multistage estimation and for the computation of standard errors. We note the point estimates are identical to those obtained using the average of the five (or ten) plausible values, but the standard errors are about two to three times as large. The PISA mathematics reporting scale is normalized to mean 500 and standard deviation of 100 across 30 OECD countries; it is directly comparable across PISA waves from 2003 onward. The average PISA math scores for boys and girls in HI and MI countries are reported in the tables of results below. The questions on electronic gaming have become more sophisticated over time in the PISAs, separating video games, SPGs from collaborative games (MMO) in 2009, 2012, and 2015. This is potentially quite important given that Smyth (2007) finds after comparing other types of games to MMOs that the latter represent a “different gaming experience with different consequences.” He finds that the number of weekly hours played by participants (three-quarter males) randomly assigned to MMOs is twice (14.4 hours) as large as those assigned to SPGs (6.6 hours), and leads to greater interference with real-life socializing and academic work, but to higher acquisition of new “friendships.” ESA (2017) report that 53% of the most frequent gamers in the United States play multiplayer games at least once a week, spending an average of 6 hours playing with others online and 5 hours playing with others in person. In 2015, PISA added two questions asking when, in their daily routine, students played video games: the answers were “before going to school,” 45% of boys versus 16% of girls, “after leaving school,” 66% of boys versus 25% of girls. There are many sources of information confirming that gaming has become a daily activity, especially for boys. To include earlier waves, we recode the answers to lowest common denominator: any games. But, we also perform separate analyses only for 2009, 2012, and 2015 using the finer distinctions. As indicated above, these activities are listed under “How often do you use a computer (digital devices) for the following activities outside of school?,” and vary by waves; they include “playing games” and “downloading music” among others. The answers record the intensity of the activities: everyday/almost everyday, a few times a week, between once a week and once a month, less than once a month/hardly ever/never. Following the literature on anticipated effects of gaming, we focus on the high-intensity gamers (also sometimes called “hardcore” players) compared to the more casual users (omitted group). Essentially we think that playing once a month or once a week is unlikely to confer the presumed cognitive enhancement or to represent a substantial distraction. 19 ### 3.2. Cross-country Visualization In Fig. 3, we provide a visualization of the cross-country data at hand. The cross-country scatter-plots illustrate the relationship between the percentage of everyday computer gamers and average math scores by PISA wave and gender. These scatter-plots represent a tell-tale figure of the primary stylized fact uncovered in this chapter. For boys, Fig. 3A shows a positive relationship among MI countries (in orange/lighter color) and a negative or neutral relationship among HI countries (in blue/darker color). Dividing the countries by income status shows a tighter relationship between the two variables that the overall cloud leads to believe. It also illustrates the substantial differences in the average test scores between HI and MI countries, in the 75 points range for boys and the 70 points range for girls. This relationship is consistent with the view that, at the country level, higher ranges of math scores are reached only when a smaller percentage of students engage in these distracting activities. Interestingly, Japan (before 2012), Singapore, and Korea figure among the high math score countries with lower than average percentage of everyday gamers. Strikingly, however, Fig. 3B shows a uniformly negative relationship for girls between the percentage of everyday computing gamers and average math scores in both MI and HI countries, although the relationship turned neutral in HI countries in 2015. ### Fig. 3. Percentage of Everyday Gamers and Math Test Scores by Country, Year, and Gender (Panel A for Boys and Panel B for Girls). ### 3.3. Individual and Family Characteristics Our analyses include a host of student characteristics and family variables: presence of an Internet link in the home, presence of a computer in the home, the student’s age, international grade, first- or second-generation immigrant status, father and mother education (primary (omitted category), secondary or tertiary education) and occupation (white-collar high-skilled, white-collar low-skilled, blue-collar high-skilled, blue-collar low-skilled (omitted category), and the number of books in the home (5 categories – lowest omitted). In Table AII, we report the means of these variables by everyday gamers and other users, by gender and by HI/MI country category to see whether boys and girls intense gamers are different from each other, and from other users, on the basis of observables. The more salient differences between everyday gamers and others are unsurprisingly coming from Internet related variables, such as a higher percentage of music downloaders, Internet link and home computer among the everyday gamers, especially among boys. In MI countries, there are also some significant higher percentage of country-level Internet and home computer penetration among intense gamers than nonintense gamers. In terms of socio-economic status variables, there are more differences between intense gamers and others in MI countries. In HI countries, everyday gamers come from families with parents of somewhat lower education level, but the differences are not large. In MI countries, the differences between everyday gamers and others are more striking. Intense gamers come from families with more educated parents employed in occupations with higher status, with more books in the home and they are more likely to be immigrants. On the basis of individual and family characteristics, there are little statistically significant differences between male and female everyday gamers. In HI countries, everyday female gamers come from families with fewer fathers with tertiary education and high-skilled white-collar jobs. In MI countries, female gamers have less of the higher socio-economic characteristics that their male counterparts have. To further check whether female everyday gamers are a glaringly negatively selected group, we computed kernel densities of the PISA math test scores comparing everyday gamers and nonusers to the entire group. These are displayed for 2015 in Fig. A1 by HI and MI countries. 20 In Checchi and Flabbi (2013), test score densities of German and Italian students from different high school tracks show clear stochastic dominance patterns of the academic track over the vocational track, for example. No such evidence appears here in HI countries, although that might be the case in MI countries. By comparison with the entire sample of boys, everyday boy gamers are over-represented on the middle of the distribution. Although there are differences across games, which we discussed above, if the gaming industry at large is optimizing to attract the largest number of players, we would expect computer games to target the median (male) player. Among everyday female gamers, there is less mass in the lower middle and more mass in the upper middle of the test score distribution than among all girls. In MI countries, by contrast, the distribution of test scores of everyday gamers appears to stochastically dominate that of nonusers. Again, this underscores the need to study HI and MI countries separately. Also, note that our control group is made of casual gamers, those who play a few times a week or between once a week and once a month, rather than nongamers which might be differently selected groups by gender. ### 3.4. School Variables In some within-school-year models, we add the following school-level measures: proportion of girls in the school, the percentage of certified teachers, the student–teacher ratio, dummies for instructional material not lacking and strongly lacking. The relative abundance of computer resources in the school is captured by percentage of computers connected to the Internet, and the ratio of computers to school size. 21 In addition, we include some school climate factors, namely learning hindrance factors and sense of belonging in some specifications. These school climate factors, described below, are deemed very important by educators and speak to the interaction between identity and social effects (Turner, Reynolds, Lee, Subasic, & Bromhead, 2014). A seemingly important variable such as time spent on homework is not recorded consistently over the waves. This may be due to the changing nature of homework over the period, part of which increasingly takes place in the school, and the increased use of tutors. Another confounding factor for time spent on homework is the increasing role of “multitasking while doing homework” (Pabilonia, 2016). This implies that longer time spent on homework does necessarily mean a higher quantity (or higher quality) of homework done given that many students may be simultaneously texting or Facebooking. For 2009, 2012, and 2015, we use a simple measure of homework available in categories similar to gaming and other activities: “Doing homework every day.” Some gender differences in social norms about doing homework every day (Bishop, 2006) also make it difficult to anticipate the effect of homework on test scores. In some boys’ culture, it may be best to score high without spending much time on homework. Bursztyn and Jensen (2015) present experimental evidence on the role of reputation or peer effects on not expanding effort at school to be popular. Doing one’s homework at school on the morning of the due date, rather than the night before at home, would be the visible signal that some students could use to remain popular. School climate factors have been found to be among the most important determinants of student engagement (Algan, Cahuc, & Shleifer, 2013). We use two critical sets of factors: these include factors hindering learning, available at the school level, and factors capturing students’ sense of belonging to the school, available at the student level. 22 We divide the factors hindering learning into teachers’ factors, such as teachers’ not meeting individual students’ needs, and students’ factors, such as student absenteeism and student skipping classes. Following the tradition in the psychology literature, summary measures of the teachers’ behavior and of students’ behavior are constructed using the Cronbach’s alpha coefficient of internal consistency. 23 In some specifications, we include the single variables that show the highest correlation with our dependent variables, that is “Student skipping classes.” Our sense of belonging variables include a measure of popularity “I make friends easily” and the opposite “I feel awkward and out of place.” If the sexualized content of some MMOs could make some players feel awkward, this would apply more to girls than boys. ## 4. Empirical Strategies and Results ### 4.1. Decomposition Method In this section, we explain our empirical strategies for the estimation of the gender gap in math scores, T i , and the role of intense gaming, G i , in accounting for the discrepancies. There are two classic approaches to the estimation of gender gaps in math scores stemming from the estimation of the gender pay gap. The first approach pools the data on both genders and includes a group indicator (female F i in this case) in a regression of explanatory variables, X i , on math scores with the goal of shrinking the coefficient of the female indicator, (1) T i = δ p F i + γ p G i + X i β p + υ i , where E [ υ i | F i , X i ] = 0 . Importantly in this approach, until all possible explanatory variables have been included, the coefficient of the female dummy δ p suffers from an omitted variables bias. As explained in Gelbach (2016), in order to assess the impact of a particular explanatory variable, one has to compare the full specification with the other that leaves out the variable of interest. Hence, the effect of intense gaming would difference out the coefficient of the female dummy in specifications with and without that variable, but including all other covariates. A second approach, the OB decomposition, avoids the issue of the “order of the decomposition matters” (Fortin, Lemieux, & Firpo, 2011) by running full linear specifications separately by gender (2) T i g = γ g G i g + X i g β g + υ i g , for g = f , m under the zero conditional mean assumption, E [ υ i g | X i ] = 0 . The approach appeals to counterfactual average scores that substitute female characteristics in the male regression (or the other way around), these are then added and subtracted from the gender differences in average scores. Letting Δ X ¯ = X ¯ m X ¯ f and Δ β ̂ = β ̂ m β ̂ f , the OB decomposition can be written using either male returns or female returns to characteristics as reference test scores structure: (3) Δ T O B = T ¯ m T ¯ f = ( G ¯ m G ¯ f ) γ ̂ m + G ¯ f ( γ ̂ m γ ̂ f ) + Δ X ¯ β ̂ m + X ¯ f Δ β ̂ or = ( G ¯ m G ¯ f ) γ ̂ f G ¯ m ( γ ̂ m γ ̂ f ) + Δ X ¯ β ̂ f X ¯ m Δ β ̂ , where the first two terms correspond the so-called explained and unexplained component of the gender gap in test scores attributable to intense gaming, and the third and fourth terms correspond to the same components for the other covariates. If the test scores returns to gaming, γ f and γ m , are vastly different by gender, for example, if they are of different signs, the part of the test score gap accounted by the gender composition of intense gaming might be positive in the first case and negative in other, an undesirable outcome. In the case of the gender pay gap, the choice of reference wage structure amounts to choosing the wage structure that would prevail in the absence of discrimination, either the male or the female wage structure, or a mix of the two. Fortin (2008) shows that using the pooled coefficients in the presence of a female dummy, as in Equation (1), in a pooled OB decomposition (4) Δ T O B P = ( G ¯ m G ¯ f ) γ ̂ p + [ G ¯ m ( γ ̂ m γ ̂ p ) G ¯ f ( γ ̂ f γ ̂ p ) ] + Δ X β ̂ p + [ X ¯ m ( β ̂ m β ̂ p ) X ¯ f ( β ̂ f β ̂ p ) ] is equivalent to running that regression in the sense that the coefficient of the female dummy corresponds the unexplained disadvantage of women: δ ̂ p = [ X ¯ f ( β ̂ f β ̂ p ) + G ¯ f ( γ ̂ f γ ̂ p ) ] . 24 In the context of test scores, there is no such presumption that the test score “returns” to individual, family or school characteristics, or more importantly country and school fixed effects, should be different by gender. However, as explained above, we anticipate the coefficient of intense gaming to be different by gender. In order to simplify the exposition and to increase the precision of our estimates, we use a pooled hybrid specification that includes an interaction between the female and everyday gaming indicators, (5) T i = δ p h F i + γ p h G i + γ f h F i G i + X i β + υ i , where γ f h = γ f γ p h is the gender-specific return (or penalty) that female gamers get in addition to γ p h . The coefficient of the female dummy, δ h , corresponds to the unexplained gender gap among nongamers and, δ h + γ h + γ f h , gives us the unexplained gender gap among gamers. By analogy with the pooled decomposition, the part of the gender gap attributable to intense gaming is: (6) Δ T P H | G = ( G ¯ m G ¯ f ) γ ̂ p h G ¯ f ( γ ̂ f γ ̂ p h ) . The first component captures the composition effects or the part explained by gender differences in the proportion of intense gamers evaluated at the returns to gaming common to all players. The second component, also referred to as the unexplained part arising from gender differences in the returns to gaming, captures the “structural” effects of the underlying test score functions. 25 The analogy with the role of de-unionization in accounting for increases in wage inequality between two time periods (Fortin et al., 2011) is a composition effect coming from the decrease in unionization rate and a wage structure effect arising from a decline in the union premium. Like union workers, hardcore gamers may not be a random sample of casual gamers; rather they self-select into gaming everyday. The concern with this “structural” interpretation is that the conditional independence assumption (also known as the ignorability assumption), G υ | X i , might not hold. This would happen if correlation between intense gaming and unobservables conditional on observed characteristics was different for male and female gamers. We alleviate these concerns below using an argument à la Altonji et al. (2005). If the female and male gaming effects show the same sensitivity to the addition of a host of covariates, in particular, personality variables linked to intense gaming, this indicates that the effect of omitted unobserved variables, is likely unimportant. In Section 4.5, we will turn to gender-specific analyses of the determinants of intense gaming to address the issue of self-selection into intense gaming in more detail. 26 ### 4.2. Results from Pooled Analyses Our first analyses begin by pooling male and female students from all countries and all comparable waves (years) of the PISA surveys. We estimate the following country and year fixed effects model, (7) T i c t = δ p h F i c t + γ p h G i c t + γ f h F i c t G i c t + X i c t β + X c t Λ + π c + θ t + ϵ i c t , where X c t represents country-level variables, and π c and θ t country and time fixed effects, respectively. The results of the country-year fixed effects models are presented for all ICT countries in Table 2 for math scores and in Table 3 for reading scores. With five waves of the PISA surveys, we have close to 1,300,000 observations from 56 countries that participated in the ICT supplements. The raw gender math scores gap is displayed in column (1) as the OLS coefficient of the female indicator: it shows a gender gap of 13.31 (0.532) out of an average of 485.85, a 2.7% difference. The following columns display specifications that gradually add more explanatory variables. Adding country-level variables in column (2) brings down the gap by a sizeable 1.63 points (although not a statistically significant difference) and has the R-squared jump to 0.20, showing that a fifth of the variance in test scores arises between countries. Adding individual and family variables in column (3) increases the gender gap, effectively dis-explaining the gap, but adds 0.15 to the R-squared. Column (4) begins to add the ICT variables, having an Internet connection or a home computer contribute a 10–20 points advantage in math test scores. Interestingly, when ICT variables are included in the regression the statistical significance of female labor force participation as a measure of gender norms disappears, consistent with our hypothesis that information technologies channel gender norms about girls and math. This point is reinforced by the fact that in the regressions on reading scores, presented in Table 3, the coefficient of female labor force participation remains highly significant across the entire table. Table 2. Country-year Fixed Effects Estimates on PISA Math Scores 2003–2015. All ICT Countries (1) (2) (3) (4) (5) (6) (7) Average Math Score: 485.80 Female −13.307 −11.681 −12.981 −12.719 −13.934 −13.888 −10.874 (0.532) (0.520) (0.448) (0.442) (0.477) (0.476) (0.556) Everyday gamer −4.638 −0.440 2.643 (0.491) (0.501) (0.571) Female* everyday gamer −16.051 −14.849 (0.408) (0.537) Everyday music downloader −9.455 (0.825) Female* everyday music downloader −2.561 (0.682) Home link to Internet 11.335 11.609 15.226 15.183 (0.637) (0.632) (0.628) (0.627) Home computer 18.616 19.242 19.559 19.594 (0.697) (0.710) (0.706) (0.705) Female LFP 0.605 0.724 0.107 0.080 −0.037 −0.054 (0.227) (0.205) (0.198) (0.198) (0.197) (0.198) Individual controls No No Yes Yes Yes Yes Yes Family controls No No Yes Yes Yes Yes Yes ICT controls No No No Yes Yes Yes Yes Country controls No Yes Yes Yes Yes Yes Yes No. of observations 1,295,608 1,295,608 1,295,608 1,295,608 1,295,608 1,295,608 1,295,608 R-squared 0.004 0.210 0.347 0.356 0.356 0.361 0.362 Notes: Country and year fixed effects estimates (except column 1) that control for individual level, family, ICT, and country-level macro variables as indicated. Individual controls include age, grade level, immigrant status (2); family controls include father and mother education (3 categories) and occupation (4 categories), books in the home (5 categories); ICT controls include home Internet link, home computer, percent Internet users at country level; percent home with computer, both at country level; macro variables include log GDP per capita, female labor force participation rate, male and female youth unemployment. Standard errors in parentheses are computed using the “repest” STATA procedure; all coefficients are statistically significant at the p < 0.01 level, unless otherwise indicated: **p < 0.05; *p < 0.1; †not statistically significant. Table 3. Country-year Fixed Effects Estimates on PISA Reading Scores 2003–2015. All ICT Countries (1) (2) (3) (4) (5) (6) (7) Average Reading Score: 485.54 Female 29.227 30.806 29.285 29.600 28.446 28.484 31.649 (0.560) (0.554) (0.483) (0.470) (0.513) (0.513) (0.582) Everyday gamer −4.405 −0.864* 2.921 (0.493) (0.510) (0.611) Female* everyday gamer −13.539 −12.774 (0.424) (0.521) Everyday music downloader −11.338 (0.916) Female* everyday music downloader −1.687** (0.674) Home link to Internet 12.197 12.457 15.508 15.448 (0.666) (0.665) (0.677) (0.677) Home computer 19.927 20.521 20.789 20.822 (0.589) (0.599) (0.596) (0.595) Female LFP 1.880 1.984 0.974 0.949 0.850 0.832 (0.217) (0.190) (0.181) (0.181) (0.182) (0.182) Individual controls No No Yes Yes Yes Yes Yes Family controls No No Yes Yes Yes Yes Yes ICT controls No No No Yes Yes Yes Yes Country controls No Yes Yes Yes Yes Yes Yes No. of observations 1,295,608 1,295,608 1,295,608 1,295,608 1,295,608 1,295,608 1,295,608 R-squared 0.022 0.170 0.309 0.321 0.322 0.325 0.326 Notes: Country and year fixed effects estimates (except column 1) that control for individual level, family, ICT, and country-level macro variables as indicated. Individual controls include age, grade level, immigrant status (2); family controls include father and mother education (3 categories) and occupation (4 categories), books in the home (5 categories); ICT controls include home Internet link, home computer, percent Internet users at country level; percent home with computer, both at country level; macro variables include log GDP per capita, female labor force participation rate, male and female youth unemployment. Standard errors in parentheses are computed using the “repest” STATA procedure; all coefficients are statistically significant at the p < 0.01 level, unless otherwise indicated **p < 0.05; *p < 0.1; †not statistically significant. Column (5) adds a dummy for everyday gaming which increases the gender gap by 1.22 point, a very sizeable change for a single variable, and shows a negative association with test scores. So far, the point estimates of the female dummy in these specifications range from −11.84 to −13.9 with standard errors around 0.5; therefore, estimates have to differ by about two points to be statistically different. Although the estimates are thus not statistically different, there are increasing in magnitude (rather than decreasing) as more explanatory variables are included: given their individual and family characteristics girls’ math test scores should be higher, but they are not. In Table 3, these explanatory variables are similarly ineffective at accounting for the positive gender gap in reading scores. These results are line with those of Fryer and Levitt (2010), for example, who find little explanatory power of individual and family variables (actually dis-explanatory power as found here) toward the math gender gap at younger ages, but larger explanatory power of country-level variables. Column (6) adds an interaction between the female and the everyday gaming indicator. The base effect of intense gaming turns positive and significant while the interaction is highly significant and negative, of an order of magnitude comparable to the positive effect of having a home Internet link. The coefficient of the female indicator captures the gender gap among nonintense gamers and is found not to be statistically different from column (5). Column (7) adds indicators for an interaction between downloading music and the female dummy along with base effects; here, the base effect is negative and highly significant, while the coefficient of the interaction also negative and significant is of smaller magnitude. Now the coefficient of the female dummy indicates the gender gap in test scores among nonintense gamers and music downloaders and is of statistically lower magnitude than among all students (column (5)). Consistent with a cognitive enhancement channel, we find general positive effects of gaming on math scores (and reading scores) that contrast with the larger negative effects of music downloading on these scores. 27 The female interaction coefficient with computing gaming is much larger and negative than for music downloading suggesting a gender-specific effect (or selection) for girls who engage in intense gaming. ### 4.3. Results from Within-school-year Models Given that individual and family variables have little explanatory power toward the gender gap in test scores and to explore the potential mechanisms at play to account for gender differences in the effect of gaming, we turn to models where the country-level variables and country-fixed effects of Equation (7) are replaced by school-level variables and school-year fixed effects. 28 We find substantial differences across HI and MI countries in the estimated coefficients of many school variables. Our first goal is to appraise the three mechanisms of interest by which intense gaming is thought to affect gender differences in math test scores. Our second goal is to further assess issues of selection into gaming, at the individual-family level and at the school level. As argued earlier, the schooling environment introduces many students to computers which can become the gateway to gaming. Table 4 presents the results of models that include school-level variables in the odd columns, and models that add school-year fixed effects in the even columns. The results are reported for all ICT countries, and for HI and MI countries separately to alleviate concerns that computing gaming might be capturing some unobserved factor that varies by type of country. Columns (1) and (2) report estimates of the female indicator (among nongamers and nonmusic downloaders) very similar to those of Table 2. However, the girl-specific effects of intense gaming are less negative in school fixed-effects models than those estimated only with school variables, suggesting that there is a yet unobserved school factor compensating – rather than an individual factor – for the negative effects of intense gaming for girls. Table 4. School-year Fixed Effects Estimates on PISA Math Scores 2003–2015 by Country Income Status. (1) (2) (3) (4) (5) (6) Average Score All ICT Countries HI Countries MI Countries 491.87 517.03 447.23 Female −12.693 −13.201 −15.490 −15.544 −9.057 −10.559 (0.498) (0.470) (0.623) (0.567) (0.900) (0.828) Everyday gamer 2.578 4.727 3.143 5.296 2.504** 4.425 (0.614) (0.611) (0.767) (0.737) (0.989) (0.804) Female* everyday gamer −8.294 −6.473 −6.697 −4.718 −10.573 −8.941 (0.963) (0.841) (1.130) (0.940) (1.768) (1.606) Everyday music downloader −16.035 −12.835 −18.784 −14.956 −8.469 −7.635 (0.579) (0.517) (0.534) (0.538) (1.284) (1.143) Female* everyday music downloader −2.087 −1.719** −2.034 −1.733** −1.239† −0.923† (0.806) (0.745) (0.795) (0.791) (1.559) (1.419) Home link to Internet 14.402 9.765 20.589 14.605 6.782 3.071 (0.716) (0.883) (0.839) (1.168) (1.123) (1.110) Home computer 17.136 12.416 18.158 13.421 17.804 12.591 (0.801) (0.750) (0.811) (0.846) (1.284) (1.146) School computers with web access 9.587 16.193 18.370 28.545 6.409** 6.743† (2.246) (4.264) (3.694) (5.237) (3.117) (6.067) Student–teacher ratio −0.072† −0.244† 1.085 1.444 −0.566 −1.456 (0.081) (0.190) (0.151) (0.263) (0.091) (0.318) School fixed effects No Yes No Yes No Yes Country controls Yes No Yes No Yes No No. of schools 28,510 21,363 3,431 No. of observations 977,780 977,780 765,520 765,520 212,260 212,260 R-squared 0.35 0.49 0.27 0.42 0.29 0.45 Notes: Regressions without school-year fixed effects include country and year fixed effects. Other included variables are age, immigrant status, second-generation immigrant status, father and mother education (three categories) and occupation (four categories), books in the home (five categories), international grade, ratio of computers to school size, percentage of girls in the school, of certified teachers, school’s community location (six categories), dummies of instructional material not lacking and strongly lacking. Standard errors in parentheses are computed using the “repest” STATA procedure; all coefficients are statistically significant at the p < 0.01 level, unless otherwise indicated: **p < 0.05; *p < 0.1; †not statistically significant. The comparisons of HI and MI countries in columns (3) and (5), and (4) and (6) show that several ICT variables have statistically different coefficients by type of countries, notably home linked to the Internet and the percentage of school computer linked to the Internet: the correlations are larger in HI countries. However, the base effect of intense gaming appears higher in HI countries and the girl-specific effect less negative in HI countries than in MI countries, but these differences are not significant. Table 5 omits PISA 2006 for which the hindrance to learning factors are not available and reports estimates similar to those of the previous table while controlling for these factors. 29 The estimates of students hindrance to learning factors, in particular student skipping classes, are negative and generally significant. If intense gamers are also students who tend to skip classes or otherwise disturb the learning experience, these controls should account for the negative effects of that behavior on math scores. Like in the previous table, these controls appear to yield more positive coefficients of intense gaming, and less negative girl-specific coefficients, but the differences are not statistically significant. Indeed, as the data visualization exercise of Fig. 3 suggested, the relationship between everyday gaming and math test score is becoming less negative over time and in this sample weighted toward the later years, the total effects of gaming for girls are no longer negative. However, controlling for hindrance to learning factors has little impact on the math gender gap. Again, the base estimates of music downloading are large and negative; given the lack of statistical significance of the girl-specific estimates, the effects of music downloading are of the same order of magnitude for boys and girls. This underscores the importance of the distraction channel and the distinctiveness of gender differences in gaming by comparison with this other Internet-enabled activity. Table 5. School-year Fixed Effects Estimates on PISA Math Scores 2003, 2009–2015 by Country Income Status: Accounting for Classroom Behavior. (1) (2) (3) (4) (5) (6) Average Score All ICT Countries HI Countries MI Countries 491.87 517.41 447.19 Female −13.225 −13.194 −14.997 −15.012 −11.429 −11.372 (0.518) (0.523) (0.646) (0.649) (0.932) (0.941) Everyday gamer 6.262 6.265 6.622 6.598 6.177 6.187 (0.625) (0.626) (0.745) (0.747) (0.938) (0.939) Female* everyday gamer −5.603 −5.650 −3.546 −3.563 −8.464 −8.537 (0.914) (0.924) (1.044) (1.048) (1.753) (1.760) Everyday music downloader −11.556 −11.582 −13.500 −13.540 −6.470 −6.516 (0.536) (0.539) (0.550) (0.554) (1.108) (1.109) Female* everyday music downloader −1.070† −1.064† −1.639* −1.586* 0.031† 0.021† (0.737) (0.740) (0.840) (0.843) (1.364) (1.364) School computers with web access 12.546 12.086 34.469 31.583 −4.691† −3.762† (4.268) (4.331) (6.026) (6.352) (6.021) (6.294) Student–teacher ratio −0.431* −0.358† 1.448 1.298 −1.516 −1.434 (0.247) (0.241) (0.418) (0.366) (0.343) (0.350) Hindrance to learning Teacher factors 12.621 14.693 4.068 (3.201) (4.374) (5.048) Student factors −33.511 −43.595 −16.133 (3.197) (3.798) (6.314) Student skipping class −17.161 −24.027 −5.224* (1.669) (1.490) (2.744) No. of schools 26,838 26,838 20,435 20,435 6,403 6,403 No. of observations 786,600 786,600 613,298 613,298 173,302 173,302 R-squared 0.52 0.52 0.46 0.46 0.47 0.47 Notes: In addition to school-year fixed effects, the regressions include control for individual level, family, and ICT variables as indicated. Individual controls include age, grade level, immigrant status (2 home link internet, home computer); family controls include father and mother education (3 categories) and occupation (four categories), books in the home (five categories). Standard errors in parentheses are computed using the “repest” STATA procedure; all coefficients are statistically significant at the p < 0.01 level, unless otherwise indicated **p < 0.05; *p < 0.1; †not statistically significant. Beyond the effect of computer gaming on test scores, there are other interesting HI and MI countries differences that may speak to the mixed results of the experimental literature (e.g., Beuermann et al., 2013). School computers with web access have larger positive coefficients in HI countries but are generally not significant in MI countries. The coefficients of the student–teacher ratio display the counterintuitive positive sign in HI countries, where at-risk students are enrolled in smaller classes, but the coefficients are negative and highly significant in MI countries. 30 To investigate the gaming network channel, we present within-school-year analyses focusing on PISA 2009, 2012, and 2015 that differentiate SPGs from collaborative games (MMOs) and allow us to control for the frequency of homework. The differentiation between games that are played individually and over the Internet is our strongest test of network effects from the “gaming culture,” which exists mainly in the context of MMOs. In Table 6, we present school-year fixed effects models that illustrate a stark difference by gender between SPGs and MMOs. The base coefficients of SPGs and the girl-specific coefficients are positive, although these coefficients lack significance in MI countries. For MMOs, the base coefficients are also positive, but the girl-specific effects are sufficiently negative to imply overall negative effects of MMOs for girls. These results are consistent with the problem-solving and social networking enhancing aspects of MMOs, which are less likely to flourish in MI countries given the lower levels of high-speed connections. They are also consistent with the anticipated negative effects discussed in the literature about the secondary role of girl gamers in MMOs. We cannot rule out that female MMOs hardcore players are negatively selected; in this case, it would imply that the MMOs environment would detract more average girls who play SPGs from playing online. Table 6. School-year Fixed Effects Estimates on PISA Math Scores 2009–2015 by Country Income Status: Separating Single Player and Collaborative Games. (1) (2) (3) Average Score All ICT Countries HI Countries MI Countries 491.24 517.41 447.19 Female −14.898 −17.177 −13.335 (0.590) (0.663) (0.991) Everyday SPG 2.774 2.516 3.493 (0.665) (0.788) (1.251) Female* everyday SPG 4.389 6.266 1.544† (1.065) (1.240) (1.915) Everyday MMO 4.182 4.906 2.097† (0.587) (0.671) (1.296) Female* everyday MMO −9.628 −8.563 −8.710 (1.305) (1.511) (2.246) Everyday music downloader −8.124 −11.705 −1.113† (0.532) (0.567) (1.068) Female* everyday music downloader −2.199 −1.875** −3.037** (0.751) (0.820) (1.427) Everyday homework −10.751 −10.430 −10.572 (0.659) (0.912) (1.014) Female* everyday homework 5.334 5.096 3.963 (0.944) (1.182) (1.446) Home link to Internet 5.314 (11.198) 1.337† (0.710) 0.922 (0.916) Home computer 11.187 (10.579) 10.096 (0.716) 0.909 (0.967) School fixed effects Yes Yes Yes No. of schools 25,448 19,245 6,203 No. of observations 647,578 496,509 151,069 R-squared 0.56 0.49 0.52 Notes: In addition to school-year fixed effects, the regressions include control for individual level and family variables as indicated. Individual controls include age, grade level, immigrant status (two); family controls include father and mother education (three categories) and occupation (four categories), books in the home (five categories). Standard errors in parentheses are computed using the “repest” STATA procedure; all coefficients are statistically significant at the p < 0.01 level, unless otherwise indicated: **p < 0.05; *p < 0.1; †not statistically significant. The estimates of other variables are interesting. They show that the base effects of everyday homework are negatively associated with math test scores, however less so among girls, given the positive coefficient of the interaction between female and everyday homework. The general negative association between everyday homework and math test scores may be a sign that only struggling students have to do homework every day, but it is consistent with a “too smart to do homework” boys’ culture (Bishop, 2006) or reducing effort at school to be popular (Bursztyn & Jensen, 2015). ### 4.4. Decomposition Results The decomposition results focusing on the role of intense gaming are computed using Equation (6) for the salient results from previous tables. They are presented in Table 7: Panel A reports the results among all ICT countries, Panel B those among HI countries, and Panel C those among MI countries. The head of each column indicates from which table and column the estimated coefficients of the base gaming effect γ ̂ p h and the coefficient of the interaction with female γ ̂ f h = γ ̂ f γ ̂ p h are taken. Given the differences in sign between the coefficients of SPGs and MMOs found in Table 6, columns (5) and (6) present separately the effects of these types of games. The average level of intense gaming among boys G ¯ m and among girls G ¯ f are reported in the following rows for each panel. The raw difference in test scores, T ¯ m T ¯ f , to be accounted for is listed next; the numbers vary slightly because of changes in samples arising from the availability of control variables that differ across PISA waves. The two components of the decomposition attributable to gaming, the more traditional composition effects ( G ¯ m G ¯ f ) γ ̂ p h and the more controversial structural effects G ¯ f ( γ ̂ f γ ̂ p h ) , along with the total effects attributable to gaming are reported next. 31 Table 7. Decomposition Results of the Role of Everyday Gaming on the Gender Math Gap. (1) (2) (3) (4) (5) (6) PISA Waves 2003–2015 2003–2015 2003–2015 2003, 2009–2015 2009–2015 A: All countries using coefficients from Table 2, Col. 6 Table 2, Col. 7 Table 4, Col. 2 Table 5, Col. 2 Table 6, Col. 2 SPG MMO Proportion of everyday gamers Boys (Gm ) 0.404 0.404 0.406 0.411 0.322 0.316 Girls (Gf ) 0.142 0.142 0.139 0.129 0.115 0.063 Gender difference in math scores 13.308 13.308 14.104 14.849 14.450 14.450 Composition effect 0.115 0.692 1.265 1.765 0.574 1.058 Structural effect 2.274 2.104 0.897 0.729 −0.505 0.607 Total attributable to gaming 2.389 2.796 2.162 2.493 0.069 1.664 % of the gender gap 18 21 15 17 0.5 12 B: HI countries using coefficients from Table 4, Col. 4 Table 5, Col. 4 Table 6, Col. 4 SPG MMO Proportion of everyday gamers Boys (Gm ) 0.411 0.420 0.317 0.323 Girls (Gf ) 0.129 0.134 0.110 0.054 Gender difference in math scores 14.843 14.630 15.105 15.105 Composition effect 1.331 1.888 0.575 1.125 Structural effect 0.835 0.477 −0.482 0.520 Total attributable to gaming 2.166 2.364 0.094 1.645 % of the gender gap 15 16 1 11 C: MI countries using coefficients from Table 4, Col. 6 Table 5, Col. 6 Table 6, Col. 6 SPG MMO Proportion of everyday gamers Boys (Gm ) 0.398 0.405 0.330 0.304 Girls (Gf ) 0.155 0.149 0.123 0.054 Gender difference in math scores 9.322 9.747 9.726 9.726 Composition effect 1.078 1.584 0.733 0.507 Structural effect 1.383 1.276 −0.187 0.466 Total attributable to gaming 2.460 2.860 0.546 0.972 % of the gender gap 26 29 6 10 Note: As explained in the text, the composition effect is computed as ( G ¯ m G ¯ f ) γ ̂ p h and the structural effect is computed as G ¯ f ( γ ̂ p h γ ̂ f ) . A first important point to note is the size and stability of the total effects attributable to intense gaming across specifications: they are in the two points, statistically significant range, of the gender math gap discussed above. They amount to 15–29% of the gender math gap. Against the backdrop of the negative composition effects found in Table 2, and more generally in the literature on the math gender gap, this is notable. Starting with the coefficients of Table 4, the composition effects dominate, and exceed the effects attributed to country-level variables reported above. 32 Columns (5) and (6) which compare the role of SPGs and MMOs further indicate that two-thirds of the total effect in the latter years (PISA 2009–2015) come from MMOs hardcore players. The composition effects capture the fact that because they are so few female MMOs players, this lowers the average female benefits of MMOs on the math test scores. The structural effects capture the differences in returns to MMOs across genders and are arguably more sensitive to the issue of potential negative self-selection of girls into intense MMOs gaming. 33 We discuss the relative importance of self-selection into intense gaming next. ### 4.5. Determinants of Intense Gaming In this section, we explore the determinants of intense gaming by gender and their impact on our estimated coefficients. We begin in Table 8 by presenting the results of school-year fixed effects models of everyday gaming separately by gender and by HI/MI countries. These regressions also include the same host of individual and family variables as in previous tables, but we display only the variables where we find substantial differences by gender. Estimating within-school-year models imply that we are “holding constant” many of the determinants of gaming that arise from geographical location and the use of computers in schools. Differences in geographical location and Internet penetration have been used as exogenous determinants of computer use at home and its impact on academic achievement. 34 Table 8. Determinants of Everyday Computer Gaming 2003, 2012, and 2015 by Country Income Status and Gender: Accounting for Personality. Proportion Gaming Everyday (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) All Countries HI Countries MI Countries Boys: 0.429 Girls: 0.140 Boys: 0.450 Girls: 0.134 Boys: 0.403 Girls: 0.149 Everyday music downloader 0.272*** 0.272*** 0.123*** 0.123*** 0.250*** 0.251*** 0.110*** 0.111*** 0.309*** 0.309*** 0.141*** 0.141*** (0.012) (0.012) (0.010) (0.010) (0.007) (0.007) (0.008) (0.008) (0.022) (0.022) (0.022) (0.022) Home link to Internet 0.032* 0.032* 0.017* 0.017* 0.039*** 0.040*** 0.012† 0.012† 0.013† 0.013† 0.019† 0.019† (0.015) (0.015) (0.007) (0.007) (0.007) (0.007) (0.009) (0.009) (0.033) (0.033) (0.010) (0.010) Home computer 0.108* 0.109* 0.057† 0.059† 0.081* 0.083* 0.035† 0.037† 0.141† 0.140† 0.083† 0.084† (0.041) (0.041) (0.030) (0.030) (0.034) (0.034) (0.024) (0.024) (0.070) (0.070) (0.052) (0.053) Home books 11–15 0.007† 0.007† 0.007† 0.008† −0.001† 0.001† 0.001† 0.002† 0.011* 0.011* 0.011*** 0.011*** (0.004) (0.004) (0.006) (0.006) (0.008) (0.008) (0.013) (0.014) (0.005) (0.005) (0.002) (0.002) 26–100 0.029*** 0.030*** 0.008† 0.009† 0.017* 0.018** 0.003† 0.003† 0.041*** 0.042*** 0.014† 0.014† (0.005) (0.005) (0.007) (0.006) (0.007) (0.007) (0.010) (0.010) (0.004) (0.005) (0.009) (0.009) 101–200 0.022** 0.023** 0.018** 0.019*** 0.008† −0.009† 0.011† 0.011† 0.042*** 0.042*** 0.029*** 0.030*** (0.008) (0.008) (0.005) (0.005) (0.010) (0.010) (0.009) (0.009) (0.008) (0.008) (0.007) (0.007) 201 + 0.005† 0.005† 0.024* 0.024* −0.003† −0.003† 0.024† 0.023† 0.012† 0.012† 0.018† 0.017† (0.008) (0.007) (0.011) (0.011) (0.011) (0.010) (0.016) (0.015) (0.006) (0.006) (0.013) (0.013) Make friends easily −0.020† −0.023*** −0.034*** −0.034*** 0.003† −0.007† (0.011) (0.006) (0.008) (0.005) (0.016) (0.007) Feel I belong −0.017* −0.010 −0.026*** −0.011* −0.002† −0.007† (0.007) (0.006) (0.007) (0.005) (0.010) (0.015) Feeling awkward −0.004† 0.021*** 0.000† 0.022*** −0.006† 0.020*** (0.005) (0.003) (0.006) (0.004) (0.007) (0.004) School fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No. of schools 20,882 20,882 20,555 20,555 15,907 15,907 15,598 15,598 4,975 4,975 4,957 4,957 No. of observations 230,243 230,243 235,613 235,613 179,427 179,427 180,444 180,444 50,816 50,816 55,169 55,169 R-squared 0.205 0.205 0.096 0.097 0.180 0.181 0.091 0.094 0.244 0.244 0.104 0.104 Notes: Other included variables are age, grade level, immigrant status (two); family controls include father and mother education (three categories) and occupation (four categories), mother home. Standard errors clustered at the country-level in parentheses. ***p < 0.01; **p < 0.05; *p < 0.1; †not statistically significant. We find twice as strong correlations between everyday gaming and everyday music downloading among boys than among girls. The association between a home link to the Internet and a home computer is also much stronger among boys than among girls, but only in HI countries. This is consistent with Winn and Heeter (2009) who find that even if there is a computer in the home, girls may have less access or may have more chores to perform. 35 The number of books in the home, a variable that shows strong association with test scores, does not indicate that girl gamers are more negatively selected than boys, just the opposite. The next set of personality variables shows that popularity among boys and girls, measured as those who make friends easily, is negatively associated with math scores, consistent with Bursztyn and Jensen (2015). We also find a negative correlation between intense gaming and the sense of belonging to the school as measured by “Feel I belong” variable, but it is higher among boys than girls. However, the “Feel awkward” variable is not significant for boys, but show a positive correlation with intense gaming among girls in both HI and MI countries. We cannot infer a direction of causality: are girl gamers feeling awkward because they play computer games or are awkward girls attracted to hardcore computer gaming? In Table 9, we explore the effect of this potential indicator of negative selection on the stability of our estimated coefficients of intense gaming using school-year fixed effects model. Comparing the odd columns to the even columns shows that our personality or “sense of belonging” variables alter only slightly our estimates of everyday gaming on math test scores, especially the girl-specific effects which change a lot less than the coefficients of the female dummy. Thus if anything, the awkwardness of girl gamers is captured in the intercept rather than in the interaction with intense gaming. If other unobserved factors are similar to these personality variables, this suggests that our interaction estimates suffer from a limited omitted variable bias using an argument à la Altonji et al. (2005). Table 9. School-year Fixed Effects Estimates on PISA Math Scores 2003, 2012–2015 by Country Income Status: Accounting for Personality. (1) (2) (3) (4) (5) (6) Average Score All ICT Countries HI Countries MI Countries 492.35 521.23 447.17 Female −13.895 −14.429 −15.511 −15.947 −13.840 −14.426 (0.629) (0.630) (0.727) (0.722) (1.120) (1.120) Everyday gamer 8.275 8.276 8.881 8.874 7.144 7.098 (0.732) (0.721) (0.817) (0.800) (1.345) (1.356) Female* everyday gamer −5.651 −5.378 −3.135 −3.091 −7.741 −7.305 (1.077) (1.068) (1.120) (1.112) (2.147) (2.143) Everyday music downloader −12.430 −11.969 −15.022 −14.391 −6.880 −6.743 (0.715) (0.712) (0.709) (0.714) (1.629) (1.636) Female* everyday music downloader 0.552† 0.482† 0.030† −0.016† 0.612† 0.577† (0.890) (0.893) (0.961) (0.950) (1.878) (1.897) Home link to Internet 5.103 5.074 8.819 8.548 0.999† 1.197† (0.832) (0.831) (1.042) (1.047) (1.446) (1.451) Home computer 9.761 9.380 10.948 10.681 7.267 6.880 (0.920) (0.923) (0.924) (0.927) (1.398) (1.411) Make friends easily −10.965 −12.824 −7.286 (0.612) (0.640) (1.157) Feel I belong 9.044 9.477 6.367 (0.636) (0.641) (1.215) Feel awkward −13.407 −12.495 −12.641 (0.737) (0.873) (1.252) School fixed effects Yes Yes Yes Yes Yes Yes No. of schools 21,798 21,798 16,616 16,616 5,182 5,182 No. of observations 465,856 465,856 359,871 359,871 105,985 105,985 R-squared 0.55 0.55 0.49 0.49 0.50 0.50 Notes: In addition to school-year fixed effects, the regressions include control for individual level and family variables. Individual controls include age, grade level, immigrant status (two); family controls include father and mother education (three categories), mother home, and occupation (four categories), books in the home (five categories). Standard errors in parentheses are computed using the “repest” STATA procedure; all coefficients are statistically significant at the p < 0.01 level, unless otherwise indicated: **p < 0.05; *p < 0.1; †not statistically significant. ## 5. Conclusion In this chapter, we use several empirical strategies to estimate the effect of everyday computer gaming (by comparison with more casual gaming) on the gender gap in math test scores within the limitations of the PISA 2003–2015 surveys. This chapter complements recent studies which found mixed effects of various initiatives, implemented in many countries, to bridge the so-called digital divide in computer access among youths, either at home or in schools. Computer gaming is a new reality where another “divide” has appeared: boys engage daily in computer (or digital devices) gaming at three times the rate of girls. This chapter provides convincing international evidence of gender differences in the effect of everyday computer gaming on math test scores, as experienced in the field over 12 years. Our base specification includes an interaction between indicators for female and everyday computing gaming, along with base effects. In addition to a host of individual, family, school, country, and ICT variables, along with either country and year fixed effects or school-year fixed effects, it also controls for music downloading and its interaction with the female indicator to capture the potential distraction/displacement effect of these Internet-enabled activities. Overall intense computer gaming has a positive association with math test scores, as predicted by some authors (Adachi & Willoughby, 2013; Green & Bavelier, 2003, 2012). But the girl-specific effect, estimated by the interaction between the female and everyday gaming indicators, is always negative and significant. In many cases (in particular in MI countries), it is sufficiently large to dominate the positive base effect and yield a total negative effect among girls. To quantify the role of intense gaming in accounting for the gender math gap, we develop a pooled hybrid decomposition that extends the Fortin (2008) framework. In this decomposition, the composition effects evaluate the impact of girls’ lower participation in intense gaming at the base returns, and the structural effects evaluate the contribution of the girl-specific negative effects given this lower participation. We thus find that the total effects of everyday gaming, the sum of the composition and structural effects, account for 13–29% of the gender gap in math test scores. These results point to a plausible channel by which the increasing use of ICT technologies introduces a substantial “swimming upstream” factor to the closing of the gender math gap. This complements the recent literature (e.g., Guiso et al., 2008; Nollenberger et al., 2016) that links gender equality across countries to the gender math gap by identifying a specific pathway. Our within-school-year models also seek to advance our understanding of the mechanisms behind these gender differences, in particular, the distraction and gaming network effects channels, which have received less attention in the literature. Considering the distraction channel, it is interesting that the negative effects of music downloading are always larger than those of intense gaming, and that the former gender-specific effects often lack significance: this highlights the gender-specificity of computer gaming. Although the differences are small, the base effects of intense gaming are enhanced and the negative girl effects mitigated when student behavioral issues are controlled for, also consistent with a distraction/displacement mechanism. The comparison of the estimates of SPGs versus MMOs is likely the most revealing regarding gaming network effects. The base effects of both SPGs and MMOs, as well as the girl-specific effects of SPGs are always positive, although the latter lack significance in MI countries. But the girl-specific effects of MMOs are always negative and sufficiently large to overtake the positive base effects of MMOs. These results are consistent with stronger positive effects of gaming for boys when they connect with the gaming community. For girls, the negative effects may reflect some negative social interactions described in the gaming literature. Even, if we were to dismiss the negative girl-specific effects on the basis unclear self-selection of girls into this activity, we would be left with a substantial male–female advantage in math scores attributable to playing MMOs, as shown in Table 7. Because the array of skills in building network of team members in the context of MMOs are transferable in the world of global Internet businesses (Martin, 2010; Werbach & Hunter, 2012), this male advantage will likely become nontrivial in the future. However, closing that gaming gap by enticing girls to play more may not be advisable. Issues with gaming content and the gaming community should be addressed first. Despite some positive correlations between intense gaming and math test scores found in this chapter, concerned parents should check that their children are not losing sleep or skipping classes to play electronic games. To the extent that computing gaming increases some types of specialized human capital, from visual-spatial skills to online networking, large gender differences in intense gaming will contribute to exacerbating these gender differences in the labor market. However, it is also possible that computer gaming will remain a dominant leisure activity for young adult males that will limit their participation in the labor market. For the United States, Aguiar, Bils, Charles, and Hurst (2017) calculate that increases in computer gaming since 2004 explain 38–79% of the four percentage points higher decline in hours of work among men aged 21–30 relative to men aged 31–55. ## Notes We would like to acknowledge Mingyi Hua for her outstanding research assistance on this project. We are grateful to the editors and referees for their careful reading and helpful suggestions. We would like to thank George Akerlof, Martha Bailey, Paul Beaudry, Vincent Boucher, Steve Durlauf, Susan Dynarski, David Green, Brian Jacob, Philip Oreopoulos, Thomas Lemieux, and seminar participants at CIFAR Workshop on SIIWB, the 2015 CEA Meetings, the 2015 EALE/SOLE Meetings, the Workshop on Social Identity and Social Interactions in Economics, the University of Michigan, the University of Calgary, the University of Waterloo, and UBC-Okanagan for helpful comments and discussions on an earlier version of the chapter. Financial support was provided by CIFAR and SSHRC Grants #410-2011-0567. 2 We omit PISA 2000 when considering math scores because of comparability issues. 3 In PISA 2012 and PISA 2015, students were asked in the ICT supplement how frequently they use a computer (or digital devices in 2015) outside of school to play one-player games (also known as single-player game–SPG) and collaborative online games: (1) never or hardly ever, (2) once or twice a month, (3) once or twice a month, (4) almost every day, and (5) every day. We focus on the most intense use (4) and (5) which we call “everyday.” 4 The question initially asked in 2003 and 2006 on how often do you use computers at home to “download music from the Internet” has evolved to include “downloading music, films, games, or software from the Internet” in later waves. Whether this should include music or film “streaming” is left to the respondent. We use this variable to denote an Internet-enabled nongaming activity. 5 Given that ICT participating countries vary by year, not much should be inferred from the time paths of the male advantage. Also, the normalization to the average of 500 is done over a set of 30 OECD countries, not necessary those participating in the ICT supplements. 6 We present some complementary evidence on reading test scores. 7 For example, Angrist and Lavy (2002) analyze the effects of a large-scale computerization policy in elementary and middle schools in Israel, based on a comparison between schools that received funding and those that did not. Beuermann et al. (2013) present the results of an RCT conducted in the context of the One Laptop per Child (OLPC) initiative, where laptops for home use were randomly provided to children attending primary schools in Lima, Peru. 8 MMOs require high-speed Internet access and are more fun to play with more gamers around, although this does not apply to SPGs. 9 The term was originally coined by Blau and Kahn (1997) in the context of increasing residual inequality. 10 Bohnet (2016) gives numerous examples of experiments showing that when women act against gender norms it often backfires. 11 Recent papers (Fairlie, 2015; Pabilonia, 2016) on the interaction between computer/media use and schooling outcomes have focused on the distraction effects studying whether computer/media use displaces homework time. We introduce frequency of homework in some specifications below but argue that it is not a clear predictor of academic outcomes (e.g., weaker students may have to study more). 12 In 2012, The SIMS 3, a role-playing game more popular with girls ranked 4th. When understood in the context of potentially steering young male gamers away from watching porn, this portrayal can be seen as part of the business model. 13 Clearly games such as the Candy Crush Saga, a match-three puzzle video that came out in 2012, stay clear of these controversial issues. Research also shows that girls are more attracted to puzzle games than to fighting games. 14 However, the establishment of specialized treatment centers for problematic gaming in South-East Asia, Europe, and the United States reflects a growing need for professional help. 15 The average gross domestic product (GDP) per capita is HI countries is around$35,000, while it is only around \$8,000 in MI countries.

16

We exclude Shanghai, China, because these macro variables are not available for this area.

17

Before 2015, we construct our school/year identifier by concatenation of a three-letter country code and the SCHOOLID provided by PISA. There is a wide dispersion in the number of students per school: there are 117 schools with more than 100 students and 231 schools with only one student.

18

To obtain the replicate weights, PISA uses the Fay variant of the Balanced Repeated Replication method with a Fay coefficient equal to 0.5 and 80 replicates.

19

The nonuser category is too collinear with Internet access and home computer to be entered successfully.

20

Densities for other years are available upon request.

21

These school variables are largely similar to those used in Fuchs and Wößmann (2008). Some of the earlier school variables are however not available in PISA 2015.

22

The detailed questions common across PISA waves are reported in Table AIII.

23

All factors have a scale reliability coefficient over 0.80.

24

As explained in Fortin (2008), the corresponding advantage of men is obtained by running the pooled regression with the male dummy.

25

Although decompositions are often viewed as simple accounting exercises, Fortin et al. (2011) show that they can be given more of a structural interpretation under several conditions, in particular, the conditional independence assumption. This assumption is slightly weaker than the independence assumption as it allows the covariates and error terms to be correlated, provided that the correlation is the same across groups. We also note that structural effects are sensitive to the choice of omitted or base category, an issue to which we return below.

26

Algan and Fortin (2016) report gender-specific results for a wide range of models similar to those presented here.

27

28

When we include observations from earlier waves of the survey, we have sufficient variations to estimate school-specific variables in addition to the school fixed effects. Including school-year fixed effects nullifying potentially unequal unobserved treatment by gender across schools.

29

When entered together the teachers and students hindrance to learning factors are of opposite signs reflecting a high degree of collinearity between these variables, but the student factors dominate.

30

More students per teacher – a higher student–teacher ratio – should impede learning opportunities and reduce test scores.

31

An undesirable feature of the structural effects, also referred to as the unexplained component, is that the effect of a categorical variable depend on the omitted group. This can be problematic in the case of categorical variables with unclear rankings, such as field of study, or industry. In the case of the union/nonunion effect, what is estimated is clear. Similarly, here the omitted group is casual gamers.

32

We note that the country-level effects reported in Section 4.2 suffer from an omitted variable bias, but in the previous literature (e.g., Fryer & Levitt, 2010) they were the only ones seeming to account for the gender math gap.

33

On the other hand, the pooled coefficients used to compute the composition effects are identified from the sample of all intense gamers, a much larger group and more average group.

34

Vigdor et al. (2014) use the availability of broadband Internet at the ZIP code level and follow students over time using a North Carolina administrative database. They do not have information on students’ gaming habits, but in student fixed-effect specifications, they find that the increased availability of high-speed Internet is associated with less frequent self-reported computer use for homework.

35

It would have been interesting to include the presence of siblings by gender, but this information was not available.

## References

Adachi & Willoughby (2013) Adachi, P. J. , & Willoughby, T. (2013). More than just fun and games: The longitudinal relationships between strategic video games, self-reported problem solving skills, and academic grades. Journal of Youth and Adolescence, 42(7), 10411052.

Aguiar, Bils, Charles, & Hurst (2017) Aguiar, M. , Bils, M. , Charles, K. K. , & Hurst, E. (2017). Leisure luxuries and the labor supply of young men. NBER Working Paper No. w23552. National Bureau of Economic Research, Cambridge, MA.

Algan, Cahuc, & Shleifer (2013) Algan, Y. , Cahuc, P. , & Shleifer, A. (2013). Teaching practices and social capital. American Economic Journal: Applied Economics, 5(3), 189210.

Algan & Fortin (2016) Algan, Y. , & Fortin, N. M. (2016). Computer gaming and test scores: Cross-country gender differences among teenagers. IZA Discussion Paper no. 10433. IZA Institute of Labor, Bonn.

Altonji, Elder, & Taber (2005) Altonji, J. G. , Elder, T. E. , & Taber, C. R. (2005). Selection on observed and unobserved variables: Assessing the effectiveness of Catholic schools. Journal of Political Economy, 113(1), 151184.

Angrist & Lavy (2002) Angrist, J. , & Lavy, V. (2002). New evidence on classroom computers and pupil learning. Economic Journal, 112(482), 735765.

Avvisati & Keslair (2017) Avvisati, F. , & Keslair, F. (2017). REPEST: Stata module to run estimations with weighted replicate samples and plausible values. Statistical Software Components. Boston College, Chestnut Hill, MA.

Beasley & Collins Standley (2002) Beasley, B. , & Collins Standley, T. (2002). Shirts vs. skins: Clothing as an indicator of gender role stereotyping in video games. Mass Communication & Society, 5(3), 279293.

Beuermann, Cristia, Cruz-Aguayo, Cueto, & Malamud (2013) Beuermann, D. W. , Cristia, J. P. , Cruz-Aguayo, Y. , Cueto, S. , & Malamud, O. (2013). Home computers and child outcomes: Short-term impacts from a randomized experiment in Peru. NBER Working Paper No. 18818. National Bureau of Economic Research, Cambridge, MA.

Biagi & Loi (2013) Biagi, F. , & Loi, M. (2013). Measuring ICT use and learning outcomes: Evidence from recent econometric studies. European Journal of Education, 48(1), 2842.

Bishop (2006) Bishop, J. H. (2006). Drinking from the fountain of knowledge: Student incentive to study and learn – externalities, information problems and peer pressure. In E. Hanushek & F. Welch (Eds.), Handbook of the economics of education (Vol. 2, pp. 909944). Amsterdam: Elsevier - North Holland.

Blau & Kahn (1997) Blau, F. D. , & Kahn, L. M. (1997). Swimming upstream: Trends in the gender wage differential in the 1980s. Journal of Labor Economics, 15(1), 142.

Blau & Kahn (2017) Blau, F. D. , & Kahn, L. M. (2017). The gender wage gap: Extent, trends, and explanations. Journal of Economic Literature, 55(3), 789865.

Bohnet (2016) Bohnet, I. (2016). What works: Gender equality by design. Cambridge, MA: Harvard University Press.

Bursztyn & Jensen (2015) Bursztyn, L. , & Jensen, R. (2015). How does peer pressure affect educational investments? Quarterly Journal of Economics, 130(3), 13291367.

Checchi & Flabbi (2013) Checchi, D. , & Flabbi, L. (2013). Intergenerational mobility and schooling decisions in Germany and Italy: The impact of secondary school tracks. Rivista di politica economica, 3, 757.

Colecchia & Schreyer (2002) Colecchia, A. , & Schreyer, P. (2002). CT investment and economic growth in the 1990s: Is the United States a unique case? a comparative study of nine OECD Countries. Review of Economic Dynamics, 5(2), 408442.

Cortes & Pan (2016) Cortes, P. , & Pan, J. (2016). When time binds: returns to working long hours and the gender wage gap among the highly skilled. IZA Discussion Paper No. 9846. IZA Institute of Labor, Bonn.

Croson & Gneezy (2009) Croson, R. , & Gneezy, U. (2009). Gender differences in preferences. Journal of Economic Literature, 47(2),448474.

Czernich, Falck, Kretschmer, & Wößmann (2011) Czernich, N. , Falck, O. , Kretschmer, T. , & Wößmann, L. (2011). Broadband infrastructure and economic growth. Economic Journal, 121(552), 505532.

De San Román & de la Rica Goiricelaya (2016) De San Román, A. G. , & de la Rica Goiricelaya, S. (2016). Gender gaps in PISA test scores: The impact of social norms and the mother’s transmission of role attitudes. Estudios de Economia Applicada, 24, 79108.

Entertainment Software Association (ESA) (2013) Entertainment Software Association (ESA) . (2013). Essential facts about the computer and video game industry: 2012 sales, demographic and usage data. Washington, DC.

Entertainment Software Association (ESA) (2017) Entertainment Software Association (ESA) . (2017). Essential facts about the computer and video game industry: 2017 sales, demographic and usage data. Washington, DC.

Fairlie (2015) Fairlie, R. W. (2015). Do boys and girls use computers differently, and does it contribute to why boys do worse in school than girls? IZA Discussion Paper no. 9302. IZA Institute of Labor, Bonn.

Fairlie & London (2012) Fairlie, R. W. , & London, R. A. (2012). The effects of home computers on educational outcomes: Evidence from a field experiment with community college students. Economic Journal, 122(561), 727753.

Fairlie & Robinson (2013) Fairlie, R. W. , & Robinson, J. (2013). Experimental evidence on the effects of home computers on academic achievement among school children. American Economic Journal: Applied Economics, 5(3), 211240.

Fortin (2005) Fortin, N. M. (2005). Gender role attitudes and the labour market outcomes of women across OECD countries. Oxford Review of Economic Policy, 21(3), 416438.

Fortin (2008) Fortin, N. M. (2008). The gender wage gap among young adults in the United States: The importance of money versus people. Journal of Human Resources, 43(4), 884918.

Fortin, Bell, & Böhm (2017) Fortin, N. M. , Bell, B. , & Böhm, M. (2017). Top earnings inequality and the gender pay gap: Canada, Sweden, and the United Kingdom. Labour Economics, 47, 107123.

Fortin, Lemieux, & Firpo (2011) Fortin, N. M. , Lemieux, T. , & Firpo, S. (2011). Decomposition methods. In O. Ashenfelter & D. Card (Eds.), Handbook of labor economics (Vol. 4A, Chap. 1, pp. 1102). Amsterdam: North-Holland.

Fryer & Levitt (2010) Fryer, R. G. , & Levitt, S. D. (2010). An empirical analysis of the gender gap in mathematics. American Economic Journal: Applied Economics, 2(2), 210240.

Fuchs & Wößmann (2004) Fuchs, T. , & Wößmann, L. (2004). Computers and student learning: Bivariate and multivariate evidence on the availability and use of computers at home and at school. Brussels Economic Review, 47(3–4), 359386.

Fuchs & Wößmann (2008) Fuchs, T. , & Wößmann, L. (2008). What accounts for international differences in student performance? A re-examination using PISA data. Physica-Verlag HD. 209–240.

Gelbach (2016) Gelbach, J. B. (2016). When do covariates matter? And which ones, and how much? Journal of Labor Economics, 34(2), 509543.

Gentile (2011) Gentile, D. A. (2011). The multiple dimensions of video game effects. Child Development Perspectives, 5(2), 7581.

Goolsbee & Guryan (2006) Goolsbee, A. , & Guryan, J. (2006). The impact of internet subsidies in public schools. Review of Economics and Statistics, 88, 336347.

Green & Bavelier (2003) Green, C. S. , & Bavelier, D. (2003). Action video games modify visual attention. Nature, 423, 534537.

Green & Bavelier (2012) Green, C. S. , & Bavelier, D. (2012). Learning, attentional control, and action video games. Current Biology, 22(6), R197R206.

Guiso, Monte, Sapienza, & Zingales (2008) Guiso, L. , Monte, F. , Sapienza, P. , & Zingales, L. (2008). Culture, gender, and math. Science, 320(5880), 11641165.

Hamlen (2010) Hamlen, K. R. (2010). Re-examining gender differences in video game play: Time spent and feelings of success. Journal of Educational Computing Research, 43(3), 293308.

Hartmann & Klimmt (2006) Hartmann, T. , & Klimmt, C. (2006). Gender and computer games: Exploring females’ dislikes. Journal of Computer-Mediated Communication, 11(4), 910931.

Hoeft, Watson, Kesler, Bettinger, & Reiss (2008) Hoeft, F. , Watson, C. L. , Kesler, S. R. , Bettinger, K. E. , & Reiss, A. L. (2008). Gender differences in the mesocorticolimbic system during computer game-play. Journal of Psychiatric Research, 42(4), 253258.

Jackson, Von Eye, Witt, Zhao, & Fitzgerald (2011) Jackson, L. A. , Von Eye, A. , Witt, E. A. , Zhao, Y. , & Fitzgerald, H. E. (2011). A longitudinal study of the effects of internet use and videogame playing on academic performance and the roles of gender, race and income in these relationships. Computers in Human Behavior, 27(1), 228239.

Kendall-Morwick (2015) Kendall-Morwick, J. (2015, February). The need for voices in CS to address the# GamerGate controversy. In Proceedings of the 46th ACM Technical Symposium on Computer Science Education (pp. 514–514). Kansas City: Association for Computing Machinery.

Kretschmer (2012) Kretschmer, T. (2012). Information and communication technologies and productivity growth. OECD Digital Economy Papers, No. 195. OECD Publishing, Paris.

Machin, McNally, & Silva (2007) Machin, S. , McNally, S. , & Silva, O. (2007). New technology in schools: Is there a payoff? Economic Journal, 117(522), 11451167.

Malamud & Pop-Eleches (2011) Malamud, O. , & Pop-Eleches, C. (2011). Home computer use and the development of human capital. Quarterly Journal of Economics, 126(2), 9871027.

Martin (2010) Martin, O. Jr. (2010). There’s an app for that (or soon will be): Smart phones, social networking, and internet gaming. Gaming Law Review and Economics, 14, 171174.

Miller & Summers (2007) Miller, M. K. , & Summers, A. (2007). Gender differences in video game characters’ roles, appearances, and attire as portrayed in video game magazines. Sex Roles, 57(9–10), 733742.

Mincer & Polachek (1974) Mincer, J. , & Polachek, S. W. (1974). Family investments in human capital: earnings of women. Journal of Political Economy, 82(2, pt. 2), 76108.

Nollenberger, Rodríguez-Planas, & Sevilla (2016) Nollenberger, N. , Rodríguez-Planas, N. , & Sevilla, A. (2016). The math gender gap: The role of culture. American Economic Review, 106(5), 257261.

OECD (2012) OECD . (2012). PISA 2012 technical report. Paris: OECD Publishing.

OECD (2014) OECD . (2014). PISA 2012 results: Creative problem solving: Students skills in tackling real-life problems (Vol. V). Paris: OECD Publishing.

Ogletree & Drake (2007) Ogletree, S. M. , & Drake, R. (2007). College students video game participation and perceptions: Gender differences and implications. Sex Roles, 56(7–8), 537542.

Pabilonia (2016) Pabilonia, S. W. (2016). Childrens media use and homework time. In C. Kalenkoski & G. Foster (Eds.) The Economics of multitasking (pp. 91109). New York, NY: Palgrave Macmillan.

Rehbein, Psych, Kleimann, Mediasci, & Möße (2010) Rehbein, F. , Psych, G. , Kleimann, M. , Mediasci, G. , & Möße, T. (2010). Prevalence and risk factors of video game dependency in adolescence: Results of a German nationwide survey. Cyberpsychology, Behavior, and Social Networking, 13(3), 269277.

Richard (2013) Richard, G. T. (2013). Designing games that foster equity and inclusion. In Proceedings of the CHI2013 Workshop: Designing and Evaluating Sociability in Online Video Games (pp. 83–88). Paris, France.

Schott & Horrell (2000) Schott, G. R. , & Horrell, K. R. (2000). Girl gamers and their relationship with the gaming culture. Convergence: The International Journal of Research into New Media Technologies, 6(4), 3653.

Skoric, Teo, & Neo (2009) Skoric, M. M. , Teo, L. L. C. , & Neo, R. L. (2009). Children and video games: Addiction, engagement, and scholastic achievement. Cyberpsychology & Behavior, 12(5), 567572.

Smyth (2007) Smyth, J. M. (2007). Beyond self-selection in video game play: An experimental examination of the consequences of massively multiplayer online role-playing game play. CyberPsychology & Behavior, 10(5), 717721.

Spencer, Steele, & Quinn (1999) Spencer, S. J. , Steele, C. M. , & Quinn, D. M. (1999). Stereotype threat and women’s math performance. Journal of Experimental Social Psychology, 35(1), 428.

Spiezia (2010) Spiezia, V. (2010). Does computer use increase educational achievements? Student-level evidence from PISA. OECD Journal: Economic Studies, 2010(1), 122.

Spreckelmeyer et al. (2009) Spreckelmeyer, K. N. , S. Krach , G. Kohls , L. Rademacher , A. Irmak , K. Konrad , T. Kircher , & G. Gründer . (2009). Anticipation of monetary and social reward differently activates mesolimbic brain structures in men and women. Social Cognitive and Affective Neuroscience, 4(2), 158165.

Turner, Reynolds, Lee, Subasic, & Bromhead (2014) Turner, I. , Reynolds, K. J. , Lee, E. , Subasic, E. , & Bromhead, D. (2014). Well-being, school climate, and the social identity process: A latent growth model study of bullying perpetration and peer victimization. School Psychology Quarterly, 29(3), 320.

United Nations (UN) (2012) United Nations (UN) . (2012). World economic situation and prospects 2012. New York, NY: United Nations Publication.

Vigdor, Ladd, & Martinez (2014) Vigdor, J. L. , Ladd, H. F. , & Martinez, E. (2014). Scaling the digital divide: Home computer technology and student achievement. Economic Inquiry, 52(3), 11031119.

Werbach & Hunter (2012) Werbach, K. , & Hunter, D. (2012). For the win: How game thinking can revolutionize your business. Wharton Digital Press. Philadelphia, PA: The Wharton School, University of Pennsylvania.

Winn & Heeter (2009) Winn, J. , & Heeter, C. (2009). Gaming, gender, and time: Who makes time to play? Sex Roles, 61(1–2), 113.

## Appendix 1. Complementary Results and Variables Definitions

### Fig. A1.

Density of PISA 2015 Mathematics Test Score by Gender and Countries Income Status.

Table AI.

Countries Participating in the ICT Survey.

PISA ICT Participation
Country Code High Income 2000 2003 2006 2009 2012 2015
Australia AU Yes Yes Yes Yes Yes Yes Yes
Austria AT Yes No Yes Yes Yes Yes Yes
Belgium BE Yes Yes Yes Yes Yes Yes No
Brazil BR No Yes No No No No Yes
Bulgaria BG No Yes No Yes Yes No Yes
Canada CA Yes Yes Yes Yes Yes No No
Chile CL No Yes No Yes Yes Yes Yes
China CN No No No No No Yes No
Colombia CO No No No Yes No No Yes
Costa Rica CR No No No No No Yes Yes
Croatia HR Yes No No Yes Yes Yes Yes
Czech Republic CZ Yes Yes Yes Yes Yes Yes Yes
Denmark DK Yes Yes Yes Yes Yes Yes Yes
Estonia EE Yes No No No Yes Yes No
Dominican Rep. DO No No No No No No Yes
Finland FI Yes Yes Yes Yes Yes Yes Yes
France FR Yes No No No No No Yes
Germany DE Yes Yes Yes Yes Yes Yes Yes
Greece GR Yes No Yes Yes Yes Yes Yes
Hong Kong HK Yes No No No Yes Yes Yes
Hungary HU Yes Yes Yes Yes Yes Yes Yes
Iceland IS Yes No Yes Yes Yes Yes Yes
Ireland IE Yes Yes Yes Yes Yes Yes Yes
Israel IS Yes Yes No No Yes Yes Yes
Italy IT Yes No Yes Yes Yes Yes Yes
Japan JP Yes No Yes Yes Yes Yes Yes
Jordan JO No No No Yes Yes Yes No
Latvia LV No Yes Yes Yes Yes Yes Yes
Liechtenstein LI Yes Yes Yes Yes Yes Yes No
Lithuania LT No No No Yes Yes No Yes
Luxembourg LU Yes Yes No No No No Yes
Macao, China MO Yes No No Yes Yes Yes Yes
Mexico MX No Yes Yes No No Yes Yes
Netherlands NL Yes No No Yes No Yes Yes
New Zealand NZ Yes Yes Yes Yes Yes Yes Yes
Norway NO Yes Yes No Yes Yes Yes No
Panama PA No No No No Yes No No
Peru PE No No No No No No Yes
Poland PL Yes No Yes Yes Yes Yes Yes
Portugal PT Yes No Yes Yes Yes Yes Yes
Qatar QA No No No Yes Yes No No
Russia RU No Yes Yes Yes Yes Yes Yes
Serbia RS No No Yes Yes Yes Yes No
Singapore SG Yes No No No Yes Yes Yes
Slovakia SK Yes No Yes Yes Yes Yes Yes
Slovenia SI Yes No No Yes Yes Yes Yes
South Korea KR Yes No Yes Yes Yes Yes Yes
Spain ES Yes No No Yes Yes Yes Yes
Sweden SE Yes Yes Yes Yes Yes Yes Yes
Switzerland CH Yes Yes Yes Yes Yes Yes Yes
Taiwan TW Yes No No No No Yes Yes
Thailand TH No Yes Yes Yes Yes No No
Trinidad and Tobago TT No No No No Yes No No
Tunisia TN No No Yes No No No Yes
Turkey TR No No Yes Yes Yes Yes Yes
United Kingdom GB Yes Yes Yes No No No Yes
United States US Yes Yes Yes No No No No
Uruguay UR No No Yes Yes Yes Yes Yes

Notes: The high income/middle income classification is from the United Nations (2012). The Gulf States are not classified as high-income countries because higher incomes may not extend to entire population surveyed.

Table AII.

Means of Individual Variables by Country-income Group, Gender, and Gaming – PISA 2003–2015.

Individual Variables High Income Middle Income
Boys Girls Boys Girls
Everyday Gamers Others Everyday Gamers Others Everyday Gamers Others Everyday Gamers Others
Home link to Internet 0.910*** 0.835 0.881 0.854†† 0.700*** 0.413 0.639*** 0.471††
Home computer 0.923** 0.832 0.912* 0.875 0.828*** 0.506 0.810*** 0.568
Percent of individuals with Internet 0.698 0.675 0.674 0.684 0.396*** 0.333 0.360 0.356
Percent of households with computer 0.718 0.720 0.693*** 0.722 0.417*** 0.339 0.382*** 0.365
Age 15.77** 15.79 15.77 15.78 15.80 15.80 15.79 15.80
International grade 9.92 9.99 9.94 9.98 9.55 9.67 9.67 9.74††
First-generation immigrant 0.074 0.062 0.071 0.069 0.055*** 0.034 0.040** 0.031
Second-generation immigrant 0.031 0.033 0.029 0.032 0.023 0.018 0.024 0.017
Father secondary education 0.538** 0.509 0.534 0.531 0.478** 0.386 0.479** 0.411
Father tertiary education 0.254* 0.289 0.232** 0.259 0.272*** 0.189 0.233* 0.196††
Father white-collar high skill 0.365 0.374 0.338* 0.373† 0.298*** 0.235 0.277* 0.245††
Father white-collar low skill 0.117 0.133 0.115 0.130 0.095 0.098 0.096 0.095
Father blue-collar high skill 0.220 0.215 0.230 0.221 0.204** 0.256 0.220 0.246†
Mother secondary education 0.583 0.580 0.601 0.607 0.449* 0.368 0.478** 0.404
Mother tertiary education 0.231 0.241 0.201 0.214†† 0.310*** 0.195 0.255*** 0.200†
Mother white-collar high skill 0.347 0.337 0.333 0.352 0.351*** 0.235 0.326** 0.272
Mother white-collar low skill 0.282 0.295 0.305 0.307† 0.206** 0.174 0.217* 0.192
Mother blue-collar high skill 0.048 0.049 0.045 0.051 0.054 0.078 0.075 0.085
Home books 11–15 0.148 0.139 0.151 0.138 0.221** 0.259 0.218** 0.246
26–100 0.321 0.306 0.318 0.318 0.309*** 0.259 0.311 0.285
101–200 0.187 0.193 0.193 0.204† 0.143*** 0.106 0.152** 0.125
201 +  0.232 0.260 0.240 0.254 0.151*** 0.103 0.162** 0.128
Log GDP per capita 10.29 10.33 10.34 10.32 8.931 8.805 8.823 8.840†††
Youth male unemployment 17.76 16.68 17.48 17.25 15.08 13.68 14.59 13.70
Youth female unemployment 17.34 15.93 17.33 16.60 17.85 16.53 17.96 16.52
Female LFP (15 + ) 0.52 0.51 0.52 0.52 0.52 0.48 0.52 51.82

Notes: Proportion of everyday gamers are reported in Table 6. Statistical significance of differences in means across everyday gamers and others indicated as ***p < 0.01; **p < 0.05; *p < 0.1. Statistical significance of differences in means across boy everyday gamers and girl everyday gamers indicated as †††p < 0.01; ††p < 0.05; †p < 0.1.

Table AIII.

School Climate Factors Common across PISA 2003, 2012, and 2015.

 A. Factors hindering learning: In your school, to what extent is the learning of students hindered by the following phenomenon? Answers coded 1 to 4: Not at all, Very little, To some extent, A lot Teachers’ Behavior Students’ Behavior a) Teachers not meeting individual students’ needs b) Student absenteeism c) Teacher absenteeism d) Students skipping classes e) Staff resisting change f) Student use of alcohol or illegal drugs g) Teachers being too strict with students h) Students intimidating or bullying other students B. Sense of belonging: My school is a place where? Answers originally coded 1 to 4: Strongly agree, Agree, Disagree, Strongly disagree are recoded: 0 (Disagree, Strongly Disagree) and 1 (Strongly agree, Agree) a) I make friends easily b) I feel like I belong c) I feel awkward and out of place

Notes: Summary measures of the teachers’ behavior and of students’ behavior, and of students' sense of belonging are constructed using the Cronbach’s alpha coefficient of internal consistency (popular in psychology). All factors have a scale reliability coefficient over 0.80.

## Appendix 1. Complementary Results and Variables Definitions

### Fig. A1.

Density of PISA 2015 Mathematics Test Score by Gender and Countries Income Status.

Table AI.

Countries Participating in the ICT Survey.

PISA ICT Participation
Country Code High Income 2000 2003 2006 2009 2012 2015
Australia AU Yes Yes Yes Yes Yes Yes Yes
Austria AT Yes No Yes Yes Yes Yes Yes
Belgium BE Yes Yes Yes Yes Yes Yes No
Brazil BR No Yes No No No No Yes
Bulgaria BG No Yes No Yes Yes No Yes
Canada CA Yes Yes Yes Yes Yes No No
Chile CL No Yes No Yes Yes Yes Yes
China CN No No No No No Yes No
Colombia CO No No No Yes No No Yes
Costa Rica CR No No No No No Yes Yes
Croatia HR Yes No No Yes Yes Yes Yes
Czech Republic CZ Yes Yes Yes Yes Yes Yes Yes
Denmark DK Yes Yes Yes Yes Yes Yes Yes
Estonia EE Yes No No No Yes Yes No
Dominican Rep. DO No No No No No No Yes
Finland FI Yes Yes Yes Yes Yes Yes Yes
France FR Yes No No No No No Yes
Germany DE Yes Yes Yes Yes Yes Yes Yes
Greece GR Yes No Yes Yes Yes Yes Yes
Hong Kong HK Yes No No No Yes Yes Yes
Hungary HU Yes Yes Yes Yes Yes Yes Yes
Iceland IS Yes No Yes Yes Yes Yes Yes
Ireland IE Yes Yes Yes Yes Yes Yes Yes
Israel IS Yes Yes No No Yes Yes Yes
Italy IT Yes No Yes Yes Yes Yes Yes
Japan JP Yes No Yes Yes Yes Yes Yes
Jordan JO No No No Yes Yes Yes No
Latvia LV No Yes Yes Yes Yes Yes Yes
Liechtenstein LI Yes Yes Yes Yes Yes Yes No
Lithuania LT No No No Yes Yes No Yes
Luxembourg LU Yes Yes No No No No Yes
Macao, China MO Yes No No Yes Yes Yes Yes
Mexico MX No Yes Yes No No Yes Yes
Netherlands NL Yes No No Yes No Yes Yes
New Zealand NZ Yes Yes Yes Yes Yes Yes Yes
Norway NO Yes Yes No Yes Yes Yes No
Panama PA No No No No Yes No No
Peru PE No No No No No No Yes
Poland PL Yes No Yes Yes Yes Yes Yes
Portugal PT Yes No Yes Yes Yes Yes Yes
Qatar QA No No No Yes Yes No No
Russia RU No Yes Yes Yes Yes Yes Yes
Serbia RS No No Yes Yes Yes Yes No
Singapore SG Yes No No No Yes Yes Yes
Slovakia SK Yes No Yes Yes Yes Yes Yes
Slovenia SI Yes No No Yes Yes Yes Yes
South Korea KR Yes No Yes Yes Yes Yes Yes
Spain ES Yes No No Yes Yes Yes Yes
Sweden SE Yes Yes Yes Yes Yes Yes Yes
Switzerland CH Yes Yes Yes Yes Yes Yes Yes
Taiwan TW Yes No No No No Yes Yes
Thailand TH No Yes Yes Yes Yes No No
Trinidad and Tobago TT No No No No Yes No No
Tunisia TN No No Yes No No No Yes
Turkey TR No No Yes Yes Yes Yes Yes
United Kingdom GB Yes Yes Yes No No No Yes
United States US Yes Yes Yes No No No No
Uruguay UR No No Yes Yes Yes Yes Yes

Notes: The high income/middle income classification is from the United Nations (2012). The Gulf States are not classified as high-income countries because higher incomes may not extend to entire population surveyed.

Table AII.

Means of Individual Variables by Country-income Group, Gender, and Gaming – PISA 2003–2015.

Individual Variables High Income Middle Income
Boys Girls Boys Girls
Everyday Gamers Others Everyday Gamers Others Everyday Gamers Others Everyday Gamers Others
Home link to Internet 0.910*** 0.835 0.881 0.854†† 0.700*** 0.413 0.639*** 0.471††
Home computer 0.923** 0.832 0.912* 0.875 0.828*** 0.506 0.810*** 0.568
Percent of individuals with Internet 0.698 0.675 0.674 0.684 0.396*** 0.333 0.360 0.356
Percent of households with computer 0.718 0.720 0.693*** 0.722 0.417*** 0.339 0.382*** 0.365
Age 15.77** 15.79 15.77 15.78 15.80 15.80 15.79 15.80
International grade 9.92 9.99 9.94 9.98 9.55 9.67 9.67 9.74††
First-generation immigrant 0.074 0.062 0.071 0.069 0.055*** 0.034 0.040** 0.031
Second-generation immigrant 0.031 0.033 0.029 0.032 0.023 0.018 0.024 0.017
Father secondary education 0.538** 0.509 0.534 0.531 0.478** 0.386 0.479** 0.411
Father tertiary education 0.254* 0.289 0.232** 0.259 0.272*** 0.189 0.233* 0.196††
Father white-collar high skill 0.365 0.374 0.338* 0.373† 0.298*** 0.235 0.277* 0.245††
Father white-collar low skill 0.117 0.133 0.115 0.130 0.095 0.098 0.096 0.095
Father blue-collar high skill 0.220 0.215 0.230 0.221 0.204** 0.256 0.220 0.246†
Mother secondary education 0.583 0.580 0.601 0.607 0.449* 0.368 0.478** 0.404
Mother tertiary education 0.231 0.241 0.201 0.214†† 0.310*** 0.195 0.255*** 0.200†
Mother white-collar high skill 0.347 0.337 0.333 0.352 0.351*** 0.235 0.326** 0.272
Mother white-collar low skill 0.282 0.295 0.305 0.307† 0.206** 0.174 0.217* 0.192
Mother blue-collar high skill 0.048 0.049 0.045 0.051 0.054 0.078 0.075 0.085
Home books 11–15 0.148 0.139 0.151 0.138 0.221** 0.259 0.218** 0.246
26–100 0.321 0.306 0.318 0.318 0.309*** 0.259 0.311 0.285
101–200 0.187 0.193 0.193 0.204† 0.143*** 0.106 0.152** 0.125
201 +  0.232 0.260 0.240 0.254 0.151*** 0.103 0.162** 0.128
Log GDP per capita 10.29 10.33 10.34 10.32 8.931 8.805 8.823 8.840†††
Youth male unemployment 17.76 16.68 17.48 17.25 15.08 13.68 14.59 13.70
Youth female unemployment 17.34 15.93 17.33 16.60 17.85 16.53 17.96 16.52
Female LFP (15 + ) 0.52 0.51 0.52 0.52 0.52 0.48 0.52 51.82

Notes: Proportion of everyday gamers are reported in Table 6. Statistical significance of differences in means across everyday gamers and others indicated as ***p < 0.01; **p < 0.05; *p < 0.1. Statistical significance of differences in means across boy everyday gamers and girl everyday gamers indicated as †††p < 0.01; ††p < 0.05; †p < 0.1.

Table AIII.

School Climate Factors Common across PISA 2003, 2012, and 2015.

 A. Factors hindering learning: In your school, to what extent is the learning of students hindered by the following phenomenon? Answers coded 1 to 4: Not at all, Very little, To some extent, A lot Teachers’ Behavior Students’ Behavior a) Teachers not meeting individual students’ needs b) Student absenteeism c) Teacher absenteeism d) Students skipping classes e) Staff resisting change f) Student use of alcohol or illegal drugs g) Teachers being too strict with students h) Students intimidating or bullying other students B. Sense of belonging: My school is a place where? Answers originally coded 1 to 4: Strongly agree, Agree, Disagree, Strongly disagree are recoded: 0 (Disagree, Strongly Disagree) and 1 (Strongly agree, Agree) a) I make friends easily b) I feel like I belong c) I feel awkward and out of place

Notes: Summary measures of the teachers’ behavior and of students’ behavior, and of students' sense of belonging are constructed using the Cronbach’s alpha coefficient of internal consistency (popular in psychology). All factors have a scale reliability coefficient over 0.80.