Unequal Loneliness in the Digitalized Classroom: Two Loneliness E ﬀ ects of School Computers and Lessons for Sustainable Education in the E-Learning Era

: Incorporating information and communication technology (ICT) into schooling has been one of the most conspicuous trends in education innovation for decades. Despite the education community’s optimistic consensus on the digitalization of the classroom, however, evidence-based research on the educational e ﬀ ectiveness of ICT is an unﬁnished task. In this situation, this study gives renewed attention to the socioemotional e ﬀ ects of school computers and draws lessons for sustainable education in the e-learning era. By analyzing the Trends in International Mathematics and Science Study (TIMSS) 2015, this study identiﬁes the causal link between school computer usage time (the independent variable) and satisfaction with peer relationships (the dependent variable) among elementary and middle school students: the loneliness deepening e ﬀ ect. Then, considering the issue of digital divide, it ﬁnds the positive interaction between the independent variable and academic performance (the moderating variable): the loneliness inequality e ﬀ ect. These two ﬁndings—summarized by the term “unequal loneliness”—call for critical reﬂections on the current use of school computers but do not support the Ludditish claim that wholly denies ICT’s educational values and potentials. Rather, the existence of the loneliness inequality e ﬀ ect additionally implies an opportunity to go beyond mere technological determinism and deliberate on human users’ capabilities for e ﬀ ective ICT usage.


Introduction
This study analyzes the educational effects of school computers on elementary and middle school students and draws lessons for sustainable education in the e-learning era. Such an approach can be classified as a part of the educational effectiveness research; however, while the majority of efforts to date have focused on the academic effects of schooling, this study explores the socioemotional effects [1]. By analyzing the Trends in International Mathematics and Science Study (TIMSS) 2015, a dataset of international students' assessment in the mid-2010s, this study identifies the negative effect of school computer usage time (the independent variable) on satisfaction with peer relationships (the dependent variable) among fourth and eighth graders, naming it the loneliness deepening effect. Furthermore, this study finds the positive interaction between the independent variable and academic performance (the moderating variable), naming it the loneliness inequality effect. The term "unequal loneliness" summarizes these two findings.
Then, criticisms with sharper tones have arisen since the 1990s [21][22][23][24][25]. The critics all pointed out the problem of distraction. The argument was that early and excessive exposure to the computer screen with fast-paced visual stimuli would not only lead to declines in literacy, thinking skills, and other academic abilities, but also negatively impact human relationships in the classroom.
With the analytic framework of the media effects research, the literature so far can be summarized into the three hypotheses [26]. The first is a relatively simple, physical one: the time displacement hypothesis: with more time spent in front of screens, the time for face-to-face communication in the classroom grows shorter.
However, the objection was soon made to it: it is too hasty to predicate that the total communication time has declined, as social media and other interactive tools enable extra communication outside the classroom. Another hypothesis was therefore added as a counter-counterargument: although the Internet ensures broad connectivity, the "weak ties" produced by the online communication cannot quantitatively or qualitatively replace the "strong ties" in the traditional classroom-verbal and nonverbal exchanges such as conversation, listening with attention, and empathy. This is the medium-specific hypothesis, which essentially states that computers and the Internet have a special media effect, going beyond being simple tools.
In addition, there is the content-specific hypothesis, which criticizes the quality of content delivered by computers and the Internet; for example, it was pointed out that educational software relies too much on visualization or gamification under the cloak of "edutainment." Critics argued that while bringing momentary fun, it would ultimately hurt the classroom atmosphere, negatively impacting both academic work and communication.
Problems were also raised with the concept of digital natives [24,27]. The critics insisted that while present-day children may be more familiar with digital culture than older generations, they should not be glamorized as digital natives given the underdevelopment of interpersonal skills, lack of empathy, and other side effects in socioemotional development brought on by over-immersion to digital devices. According to the critics, the digitalization of the classroom would be adding fuel to the fire, given the need for guarding against excessive influence of digital devices in the lives of young students.

Conflicting Promises within the Optimistic Consensus: Cooperativization versus Personalization of Education
The previous discussions can give the impression that they oversimplify the matter with the dichotomy of yes or no. However, aside from the aforementioned criticisms from techno-skepticism, two conflicting promises within the optimistic consensus added an issue: cooperativization versus personalization of education [28,29].
The cooperativization of education is as described in Section 2.1: computers encourage creative activities like project-based learning, the majority of which consist of cooperative learning based on teacher-student and student-student interactions. Among the technological factors that are expected to enable this are portable computer hardware (tablets, smartphones, etc.) as well as interactive Internet services (social media, video communications, etc.).
At the same time, however-and even in the same piece of literature-the personalization of education was also promised: while previous classes corresponded to a "standardized" education where teachers uniformly delivered the contents of textbooks, new classes under ICT would be a "personalized" education that respects the ability and aptitude of each student; in this way, it would overcome the limitations of one-to-many learning-also referred to as "the factory model"-and realize the long-cherished dream of one-to-one learning. Relevant technologies are adaptive learning programs applying artificial intelligence and immersive virtual and augmented realities [4].
However, as the personalization received praise, the alternative promise of cooperativization quietly faded. As explained in Section 2.1, the cooperativization of education involved utilizing the interactive nature of ICT to convert one-to-many to many-to-many learning. This naturally led to the question of how the realization of one-to-one learning, which began from the same criticism of one-to-many learning but constituted a very different-and at times contrasting-alternative, could coexist.
The dissonance was admitted in a recent piece of literature that speaks for the optimistic consensus [4]. It led to the addition of provisions such as "personalization is not to be confused with isolation" and "some of the best learning that takes place is not just personalized, but collaborative as well"; this is reminiscent of the concept of "networked individualism" in the social network theory [30]. However, it is hardly deniable that such provisions are not yet specific enough to resolve the conflict.

Digital Divide: ICT and Educational Inequality
As mentioned in Section 1, this study identifies the loneliness inequality effect together with the loneliness deepening effect. The framework reflects the problem of educational inequality, another cornerstone of the ICT discussion.
As is well known, the digital divide is a concept that symbolizes the problem. Since its first appearance in 1995, the concept has expanded to three levels of meaning. The first-level divide refers to the gap in the physical access to ICT resources across socioeconomic categories (class, race, etc.). Up to the early 2000s, the concept retained this simplest meaning. However, as the supply of ICT resources advanced to certain degrees, the first level naturally gave way to higher levels. From the mid-2000s, the gap in the usage of ICT appeared as the second-level divide. Then, the third-level divide, the gap in the outcomes from utilizing ICT, gained attention [31].
Obviously, this has been an issue that encompasses all of society; however, discussions on education have followed this path as well [11,32]. In the beginning, the first-level divide-differentiated opportunities to access ICT resources between advantaged and disadvantaged students-became a major problem. This re-highlighted the responsibility of schools for equality of opportunity under the cause of digital bridging, justifying massive investments in ICT since the 1990s. Then, as computers and the Internet gradually became ubiquitous in schools, the discussion shifted to the second-and third-level issues. Questions began to arise around how different schools or students may have different learning experiences and outcomes from computers and the Internet.
There has been some controversy over the questions, and the debate between the optimistic consensus and techno-skepticism (see Sections 2.1 and 2.2) has been reproduced [28]. Optimists expected that advantages of ICT such as the humanization, democratization, and cooperativization of education would provide extra learning opportunities to disadvantaged students who had been alienated from the traditional classroom, reducing educational inequality [33][34][35]. On the other hand, skeptics were concerned about the possibility of ICT failing to reduce-or worse, exacerbating-educational inequality. They have argued that side effects of ICT such as poor literacy or distraction may be even more destructive to disadvantaged children who lack the opportunity for supplemental learning outside of school [23,36].

Calling for Evidence-Based Research
Thus far, this study has outlined the discussions on the digitalization of the classroom focusing on the socioemotional effects. However, as they have largely relied on experiences and intuitions, it has been difficult to draw any definite conclusions therefrom. Evidence-based research is required, like that being attempted by this study.
As pointed out in Section 1, previous empirical studies on the educational effectiveness of ICT have mostly been based on students' academic performance as the dependent variable and have faced methodological and theoretical challenges. In this situation, this study expands the scope of investigation to the non-academic effects of ICT, hoping to resume and diversify the discussion. Few similar attempts have been made.
Slightly widening the range of the literature review, there have been intermittent attempts at empirical research in the psychological field, covering the effects of ICT on the psychological well-being Sustainability 2020, 12, 7889 6 of 20 of children [37]. However, as these studies have analyzed the effects of home computers, they are far removed from the interest of this study-the effects of school computers [12].
Moreover, the effects of home computers themselves remain unclear due to the endogeneity problem. Particularly at issue here is the possibility of reverse causality; usage of home computers itself is dependent on the current psychological state of each student. For example, a recent analysis showed that 15-year-olds who spend more than six hours a day on the Internet clearly feel lonelier than others; however, this fact does not shed light on whether the Internet makes them lonely or whether their loneliness takes them to the Internet [11] (pp. [43][44][45][46].
On the other hand, this study focuses on the non-academic byproducts of school computers, used only for the purpose of schoolwork in the highly controlled environment of the classroom. Such a research design makes it free from the reverse causality problem, one of the most problematic obstacles in the educational effectiveness research. Therefore, a more robust causal inference compared to psychological research on home computers is expected.
Furthermore, this study identifies the loneliness inequality effect caused by the interaction between school computers and student characteristics; such an approach is also rarely attempted. Revisiting the three levels of the digital divide, the vast majority of existing empirical studies on ICT and educational equality have focused on whether the opportunities for physical access are being equally guaranteed in classrooms [38], while there are only a few ones on the different usage of school computers across socioeconomic categories and even fewer on the different outcomes [32].

Peer Relationships and Loneliness: A New Frontier in the Educational Effectiveness Research
As mentioned in the beginning, this study is a part of the educational effective research: a field that derives policy implications for sustainable education by identifying school/teacher-level variables that have effects on students' learning outcomes. Naturally, the most representative outcome variable so far has been academic performance-the extent to which a student has achieved understanding and application ability in key subjects such as mathematics, science, and reading. However, over the past two decades, quite a few attempts have also been made to explore the non-academic-socioemotional or physical-outcomes (attitude, behavior, health, etc.) [1,15].
This study is a reflection of this latest trend. The point of departure is peer relationships, which form the basis of classroom life along with teacher-child relationships. It is generally agreed that positive peer relationships make critical contributions to the supportive school climate, having a profound effect on students' well-being and schools' connectedness [39,40].
However, there are two ways to approach the issue of relationships: addressing the objective experiences or the subjective feelings therefrom. This study's choice is the latter due to the data availability; thus, the dependent variable is not the state of peer relationships itself but the degree of satisfaction with it, and the negative effect of the independent variable on the very feeling is named with the concept of loneliness-rather than social isolation [41]. Such a shift of focus does not alter the implications of this study. Actually, chronic loneliness of children and adolescents is reported to be associated with weakened peer relationships, which again produces negative effects on children's socioemotional developments and experiences (e.g., less self-esteem, more social anxiety) [42].

Academic Performance as the Control/Moderating Variable
Meanwhile, academic performance is given new roles in this study. First, it functions as the control variable in the identification of the loneliness deepening effect. Although the research design prevents reverse causality, it cannot be concluded that endogeneity has been perfectly eliminated. Thus, academic performance is utilized to assess the possibility. Academic performance is the factor that teachers consider to be most important when introducing teaching tools and strategies. Furthermore, directly linked to the degree of participation in classes, it is likely to have a certain level of impact on students' psychological states. Therefore, if there are still some omitted variables that affect both the independent and dependent variables, the most likely influence would be academic performance; and if there is no significant difference in the estimation even after inserting academic performance as the control variable, we would safely determine the causal direction.
Second, academic performance works as the moderating variable in the identification of the loneliness inequality effect, and this function is actually more pivotal. The point is to consider the variable as an indicator of students' status in the classroom; students with higher academic performance are in a relatively advantageous position in the classroom, usually playing a leading role in schoolwork. In addition, they are more likely to be better positioned in terms of family background, such as parents' socioeconomic status (SES). Thus, once the negative effect of the independent variable on the dependent variable is identified at average, the simultaneous identification of the positive interaction between the independent variable and academic performance can be interpreted as another manifestation of the digital divide in education, where students with lower academic performance experience the negative effect more strongly, and vice versa.

Data
This study analyzes the dataset of the year 2015 from TIMSS, which evaluates the mathematics and science performances of fourth and eighth graders worldwide and surveys their family/school lives every four years [43]. Considering the rapid changes in technology, it is regrettable that TIMSS 2019 has not yet been released; nevertheless, TIMSS 2015 also takes into account the recent trend of e-learning and retains its timeliness. The data file can be downloaded from the official website [44]. STATA 15.1 is used for all statistical work.
Of the 57 countries that participated in TIMSS 2015, 31 had both their Grade 4 and 8 populations participate. Observations with missing values are removed by listwise deletion before the analyses, so from a total of 259,095 fourth graders and 257,685 eighth graders in the original data file, 251,860 and 251,288 observations remain, respectively. The ratio of missing data does not exceed 3%.

Variables
The dependent variable is satisfaction with peer relationships of each student, sourced from four-point Likert item-responses in the TIMSS 2015 Student Questionnaire [45]. Items (a) to (g) state satisfaction with school life in various aspects, and responses (a) to (g) ask students' degree of agreement-from "Agree a lot" to "Disagree a lot"-on the respective items (see Figure 1). Sustainability 2020, 12, x FOR PEER REVIEW 7 of 20

Data
This study analyzes the dataset of the year 2015 from TIMSS, which evaluates the mathematics and science performances of fourth and eighth graders worldwide and surveys their family/school lives every four years [43]. Considering the rapid changes in technology, it is regrettable that TIMSS 2019 has not yet been released; nevertheless, TIMSS 2015 also takes into account the recent trend of elearning and retains its timeliness. The data file can be downloaded from the official website [44]. STATA 15.1 is used for all statistical work.
Of the 57 countries that participated in TIMSS 2015, 31 had both their Grade 4 and 8 populations participate. Observations with missing values are removed by listwise deletion before the analyses, so from a total of 259,095 fourth graders and 257,685 eighth graders in the original data file, 251,860 and 251,288 observations remain, respectively. The ratio of missing data does not exceed 3%.

Variables
The dependent variable is satisfaction with peer relationships of each student, sourced from four-point Likert item-responses in the TIMSS 2015 Student Questionnaire [45]. Items (a) to (g) state satisfaction with school life in various aspects, and responses (a) to (g) ask students' degree of agreement-from "Agree a lot" to "Disagree a lot"-on the respective items (see Figure 1). Of the seven item-responses, item-response (d) ("I like to see my classmates at school") is the key to this study's interest. However, using this as is leads to the issue of response bias from the subjective tendencies or attitudes of the respondents-an inherent weakness of the Likert scale [46]. Therefore, this study adopts the following scoring strategy to minimize the effects of response bias: let y , y , y , and y be four-point scores coded from responses (a), (b), (c), and (d), respectively, and Of the seven item-responses, item-response (d) ("I like to see my classmates at school") is the key to this study's interest. However, using this as is leads to the issue of response bias from the subjective tendencies or attitudes of the respondents-an inherent weakness of the Likert scale [46]. Therefore, this study adopts the following scoring strategy to minimize the effects of response bias: let y a , y b , y c , and y d be four-point scores coded from responses (a), (b), (c), and (d), respectively, and let y be the average of y a , y b , and y c (if there is any missing value among y a , y b , and y c , y is calculated by the remaining ones); then y = y d − y is the dependent variable defined in this study, while responses (e), (f), and (g) are not considered as they may be affected by the psychological aftereffect from response (d).
This strategy is justified given the difference in contents between items (a), (b), and (c) on the one hand and item (d) on the other. The former items ask about general satisfaction associated with "school", whereas the latter asks about specific satisfaction with "classmates". In fact, Tables 1 and 2 indicate that the correlations among y a , y b , and y c are more than 0.4, while the remaining correlations with y d consistently fall below it.  Constructing y = y d − y can be seen as an effort to identify momentary fluctuations of emotions that students may experience from the following two-step question while forfeiting the measurement of an absolute level of satisfaction: "Is school going well for you? ( . . . ) How are your relationships with friends?" For example, a student who responds "Agree a lot" to items (a), (b), and (c) may suddenly respond "Agree a little" to item (d), which indicates y = −1. However, a student who responds "Disagree a lot" to items (a), (b), and (c) may suddenly respond "Disagree a little" to item (d); this would indicate y = 1. As an analogy, students who have earned a C in a class averaging at D would be evaluated higher than a student who earned a B in a class averaging at A.
The independent variable is school computer usage time of each school. This is again sourced from an item-response in the TIMSS 2015 Student Questionnaire [45]: the question about how often the student uses a computer or tablet for schoolwork at school (see item (b) in Figure 2). However, as digitalization is an educational policy imposed on the whole school, it is more appropriate to prepare a school-level aggregate variable for it. Thus, the four categories of response (b)-from "Every day or almost every day" to "Never or almost never"-are coded into number of days in a week, 5, 1.5, 0.375, and 0, respectively, and the school averages are calculated, weighted by the student sampling weight (TOTWGT) in the TIMSS 2015 dataset.
The control/moderating variable is academic performance of each student. A key subject's test score can be used as a proxy for it, and this study adopts the first of the five "plausible values" of mathematics performance (ASMMAT01, BSMMAT01) estimated by the multiple imputation technique in the TIMSS 2015 dataset. The score indicates how well each student understands the curriculum Sustainability 2020, 12, 7889 9 of 20 contents of Grade 4 (number, geometric shapes and measures, data display) and Grade 8 (number, algebra, geometry, data and chance) [43].
Sustainability 2020, 12, x; doi: FOR PEER REVIEW www.mdpi.com/journal/sustainability would be evaluated higher than a student who earned a B in a class averaging at A. The independent variable is school computer usage time of each school. This is again sourced from an item-response in the TIMSS 2015 Student Questionnaire [45]: the question about how often the student uses a computer or tablet for schoolwork at school (see item (b) in Figure 2). However, as digitalization is an educational policy imposed on the whole school, it is more appropriate to prepare a school-level aggregate variable for it. Thus, the four categories of response (b)-from "Every day or almost every day" to "Never or almost never"-are coded into number of days in a week, 5, 1.5, 0.375, and 0, respectively, and the school averages are calculated, weighted by the student sampling weight (TOTWGT) in the TIMSS 2015 dataset.   Table 3 shows the descriptive statistics of the three variables, calculated by grade and country. The most notable is the dependent variable; the sample means in all countries are close to zero, with sample standard deviations between 0.6 and 0.7. As most observed values are concentrated around zero, this can be said to be a very conservative variable. This is a natural result, as responses (a), (b), and (c) would be somewhat similar to response (d) in Figure 1. It is an inevitable price of using the scoring strategy to minimize the response bias. The point is to see whether the statistically significant effects could be identified despite the conservative nature of the dependent variable. n/a n/a n/a n/a Lebanon n/a n/a n/a n/a 0.  In terms of the independent variable, the impression may be that most figures are quite a bit smaller than expected, overshadowing decades of mass investment in ICT. For example, New Zealand is the only country showing two days or more per week of school computer usage time in Grade 4, although the list somewhat expands in Grade 8 (Australia, Canada, New Zealand, Sweden, and the United States). However, it may be too cursory to determine that Table 3 indicates the contradictory reality of "overinvestment" and "underuse" of ICT resources in schools [47]. It should be remembered that the independent variable is derived from the item-response asking students about their using computers for schoolwork at school (see Figure 2). Thus, there is a possibility that some teacher-led computer activities were not counted as a part of school computer usage time. Fortunately, such an underestimation does not affect the consistency of the measurement.
In terms of the control/moderating variable, considerable performance gaps between countries do exist, but it is beyond the scope of this study to discuss such gaps. Further, using the values from Table 3, the variable is standardized by grade and country before the analyses. This manipulation has no statistical effects on the results in Section 4, being only for ease of interpretation, particularly in Section 4.2.

Models and Estimations
In identifying the loneliness deepening effect, this study uses the simplest form of the linear regression model (Model 1). The coefficient β 1 , the influence of the independent variable X i on the dependent variable Y i , is estimated by grade and country.
More sophisticated identification strategies [48] are not used, because the definitions of the independent and dependent variables largely prevent the issue of endogeneity, as mentioned in Section 2.5. The independent variable is the frequency of using school computers for schoolwork under the instruction of teachers, while the dependent variable is an unintended emotional consequence that is irrelevant to schoolwork. Therefore, in principle, this study is free from reverse causality.
Further, as explained in Section 2.7, this study utilizes academic performance as the control variable to detect the residual chance of the endogeneity. Let Model 2 be an expansion of Model 1, with the control variable C i and its squared, cubic terms added. The coefficient β 1 is then estimated by grade and country, each estimate being compared with the previous one from Model 1.
In identifying the loneliness inequality effect, academic performance now functions as the moderating variable, as explained in Section 2.7. Let Model 3 be another expansion of Model 1, with the moderating variable M i and the interaction term X i M i added. The coefficients β 1 and β 3 , which together form the slope of the independent variable X i , are then estimated by grade and country.
As examined in Section 4.1, there would be no significant difference in the estimation of β 1 between Models 1 and 2. This implies that X i could be considered an exogenous variable. However, it is theoretically still possible for M i to be an endogenous variable. Therefore, X i M i , the product of the two, could be an endogenous variable. Fortunately, some simulation results indicate that it is possible to estimate the coefficient of an interaction term between endogenous and exogenous variables in a fairly robust manner by simply using the linear regression model [49].
The ordinary least squares (OLS) estimator is used in the estimations. Although the weighted least squares (WLS) estimator reflects the different sampling probabilities of the observations, its actual benefit is less significant in this study, as simply confirming the statistical significance and sign of each estimate, rather than its detailed value, is sufficient. Furthermore, WLS estimates usually have larger standard errors than OLS ones, with decreased statistical powers [50,51].
Considering that the data are clustered by schools, this study calculates the cluster-robust standard errors. As the TIMSS 2015 was conducted in more than 100 schools in each country, it meets the asymptotic assumption of the cluster-robust standard errors. The vce (cluster clustvar) option in the regress command in STATA 15.1 is used for the calculations [48,52]. Table 4 shows the estimation of the coefficient β 1 in Models 1 and 2 by grade and country. First, the estimates for Model 1 show statistically significant results for 14 out of 48 countries (29.2%) in Grade 4 and 13 out of 40 countries (32.5%) in Grade 8. −0.07 * −0.08 ** n/a n/a Belgium (Flemish) 0.06 0.05 n/a n/a As discussed in Section 3.2, the statistical insignificance for about 70% of the countries can be interpreted as a result of the conservative nature of the dependent variable. However, the signs of the estimates are consistently negative for the remaining 30% with statistical significance. In Grade 4, all 14 countries show negative effects, and in Grade 8, 12 of the 13 countries do (Lithuania being the exception).

Identification of the Loneliness Deepening Effect
The situation is invariant even after the control variable is added. Apart from minor changes in estimates or p-Values, the estimation results for Models 1 and 2 remain almost the same, only the United Arab Emirates in Grade 8 shifting from a statistically insignificant effect to a statistically significant negative one. As explained in Section 2.7, this implies that the causal direction from school computer usage time to satisfaction with peer relationships can be determined quite safely.
The results are also similar between Grades 4 and 8. If the comparison is narrowed down to the 31 countries that have data for both grades, the estimates for Model 1 show statistically significant negative effects for eight countries (25.80%) in Grade 4 and nine countries (29.03%) in Grade 8. The two counts are almost the same, considering the offsetting existence of one positive effect in Grade 8 (Lithuania). In terms of countries, only one (Australia) shows negative effects in both grades; seven countries show negative effects only in Grade 4 and eight countries only in Grade 8. In short, no trend is found between the two grades.
Thus, Table 4 indicates that the digitalization of the classroom is likely to have a loneliness deepening effect on elementary and middle school students: when school computer usage time rises, there is a strong possibility that satisfaction with peer relationships goes down.
Nevertheless, it is inconceivable that school computers' effects work exactly to the same degree and in the same direction for every student. Thus, in Section 4.2, it is additionally examined how each student's academic performance moderates the loneliness deepening effect. Table 5 shows the estimation of the coefficients β1 and β3 in Model 3 by grade and country. In particular, the estimates of β3, the coefficient of the interaction term, show statistically significant results for only 5 out of 48 countries (10.42%) in Grade 4 and 6 out of 40 countries (15.00%) in Grade 8.

Identification of the Loneliness Inequality Effect
Actually, it is a natural consequence that fewer than half the countries are statistically significant compared with the previous identification of the loneliness deepening effect. It is well known that when an interaction term is added and tested, a sample size that is at least several times bigger than the original is required to maintain the statistical power [53]. Indeed, interactions are much harder to identify than main effects.
However, it must be noted that the signs of the estimates are consistently positive for the small number of countries with statistical significance. In Grade 4, all five countries show positive interactions, and in Grade 8, five of the six countries do (Thailand being the exception). In other words, students with higher academic performance experience gentler negative slopes-or even positive ones-of the independent variable.
To take Australia in Grade 4 as an example, as the standardized score of academic performance (M i ) increases along the five points of −2, −1, 0, 1, and 2, the slope (β 1 + β 3 M i ) also increases along the five corresponding points of −0.082, −0.058, −0.034, −0.010, and 0.014, respectively. The value for which the sign of the slope switches (−β 1 /β 3 ) is 1.42. Assuming that academic performance is normally distributed in each country, 92.22% of the students continue to experience the negative socioemotional effects of ICT, while only 7.78% enjoy the positive ones.
The sign-switching values vary across countries. Excepting Thailand in Grade 8, which shows a negative interaction, and regarding statistically insignificant estimates of β 1 as simply zero, the value goes from 0 (Lithuania, Spain, the United States) via 1.14 (Slovakia) to 1.42 (Australia) in Grade 4 and from 0 (Japan, Korea) via 1.42 (the United Arab Emirates) and 1.80 (Hungary) to 2.19 (Malaysia) in Grade 8. 0.050 0.000 n/a n/a As explained in Section 3.3, these figures were calculated from the unweighted estimates and thus lack absolute accuracy. Nevertheless, once these figures are taken as given, and the normal distribution of academic performance is assumed again, at most 50% or at least 1.43% of the students are high performers in each country who experience the loneliness deepening effect's reversal-to the loneliness alleviating effect.
In this sense, despite the greater difficulty in identifying statistically significant results, Table 5 does indicate the loneliness inequality effect of the digitalization of the classroom: for elementary and middle school students with lower academic performance, school computer usage time is likely to have a stronger negative effect on satisfaction with peer relationships, and vice versa.

Implications of the Loneliness Deepening Effect
The identification of the loneliness deepening effect in Section 4.1 does provide an inescapable, critical message for the education community. It confirms-in the relatively invisible socioemotional sphere-that school computers currently produce a certain degree of negative educational outcomes in the global dimension, beyond national contexts.
In the debate between the optimistic consensus and techno-skepticism, it can be regarded as a factor that raises the latter's score (see Sections 2.1 and 2.2). Or, it can be understood to demonstrate that the conflict between the two promises within the optimistic consensus will not be resolved, with the negative superiority of the personalization over the cooperativization being dominant in the end (see Section 2.3). By either explanation, it becomes clear that the present-day use of ICT in the classroom falls short of the optimistic consensus.
Some debaters, in an attempt to present a third option between the two poles of optimistic consensus and techno-skepticism, have hypothesized that there may be a threshold grade or age where computers' negative effects reverse to positive [22] (pp. 205-206). However, as examined in Section 4.1, the estimation results are similar between Grades 4 and 8, with negative effects being universal in both. In other words, insofar as the TIMSS 2015 is concerned, the hypothesis cannot be confirmed; such a threshold either does not exist or does at some point after Grade 8.

Implications of the Loneliness Inequality Effect
The identification of the loneliness inequality effect in Section 4.2 can be regarded as more evidence for critical opinions of ICT's educational effectiveness. It shows that the current usage of school computers has the additional risk of expanding the scope of educational inequality to the socioemotional dimension.
In particular, it is highly plausible that the conflict between the two promises of cooperativization and personalization is playing a decisive role (see Section 2.3). Actually, one of the few prior empirical studies addressing the second-level digital divide in education-the problem of differential usage-provides a link to support this interpretation (see Sections 2.4 and 2.5).
According to Warschauer and Matuchniak (2010), it is frequently observed that students use school computers in quite different patterns depending on their family background (SES, race, etc.) and academic performance. Disadvantaged students often seem to use computers for the purpose of "remedial drills and practices," whereas advantaged ones utilize them for "higher-order skills" like modeling, presentation, and other applications. The former is usually done within the scope of preplanned learning courses, allowing only limited variations through adaptive one-to-one feedback between the individual student and computer software or website. In that sense, it corresponds well to the personalization of education. On the contrary, the latter often encourages constructivist activities, open-ended discussions, and other collective works among teachers and students, which can be appropriately described as the cooperativization of education [32].
Thus, a mechanism can be conceived that differential school computer usage because of students' background or academic performance-high-SES/high-performing students' cooperativized use versus low-SES/low-performing ones' personalized use-leads to different socioemotional outcomes. This mechanism may not be the only cause of the loneliness inequality effect, but it seems to be the most reasonable explanation for now.
It is regrettable that this partly hypothetical explanation cannot be tested further, for the independent variable is measured solely on a quantitative dimension, without any information about differential usage. It is the limitation of this study, which has been already shared by many prior empirical studies on the educational effectiveness of ICT [54]. It will be a major challenge for follow-up studies.
Actually, the issue of cooperativization versus personalization is not that new in the field of sociology of education. As a prime example, the classical proposition of James Coleman, a sociologist and pioneer in the educational effectiveness research, assumes the irreconcilability of the two [55,56]. Regarding schools as a social organization, he demonstrated that the effectiveness of a learning process-even with identical content and quality-can differ depending on whether it is based on individuals' isolated activities or the whole classroom community's collective ones. His conclusion was that the latter is more effective than the former, and this study's finding reaffirms this. Unfortunately, the ideal of networked individualism is not yet realized, at least in the digitalized classroom [30].

Lessons for Sustainable Education in the E-Learning Era
The findings of this study suggest that the overall stance of the education community on ICT needs to be more cautious than now. In other words, a more rigorous-and sometimes even self-censoring-approach is required before bringing computers and the Internet into the classroom. Similarly, candid questions about the current logic dominating the digitalization of the classroom are called for. Has the introduction of ICT so far been done with a purely educational focus? Has the commercialism of ICT companies not been involved? Or has ICT not been misunderstood as an expedient tool to replace teachers' teaching efforts rather than promote them [10,47]?
To be sure, the findings cannot be linked to the Ludditish claim that ICT should be eliminated from the classroom, nor the technological determinism that ICT's negative effects are beyond our control. Clearly, ICT still has the potential to be a truly beneficial tool for young students. The very existence of the loneliness inequality effect confirms this. While the effect does indicate another aspect of the digital divide in education, it also shows the possibility that the ultimate effects of ICT can be moderated by human users' own capabilities. According to the explanation in Section 5.2, high-performing students seem to possess the ability to utilize computers in fairly autonomous ways, mobilizing them in their own constructive, creative, and collective works, while low-performing ones do not.
Then, the problem is not technology itself, but the subjective conditions of users that determine its effectiveness and sustainability in the last instance. For example, Jane Healy, a child psychologist who can be grouped as a techno-skeptic but who also strongly rejects any Ludditish temptations, has presented a list of prerequisites for using ICT in educational purposes in genuinely wise and effective manners. Although the list was made more than two decades ago, it still contains considerable validity and timeliness at this point, characterized by the near-universalization of ICT resources and e-learning practices [22]  if parents and teachers are willing to provide a human scaffold for technology-assisted learning.
Only if such preconditions are met may young students then profit from wise choices in this emerging field of e-learning, and ICT can finally be an integral part of sustainable educationmaximizing the opportunity to learn and offering adequate support for all students-both in terms of academic performance and socioemotional development [15].

Lessons for the Whole Society
The primary focus of this study has been educational. Nevertheless, some lessons can also be discussed on the broader dimension of whole society, especially from the angle of digital sociology. While convenience is certainly enhanced as digital technology's mediating role in our public and daily lives is becoming more profound, its adverse effects are the other side of the coin. Accordingly, the critical reflection that the "digital society" is also the "digital risk society" is increasingly salient, and issues such as Internet addiction disorder, privacy infringement, cyberbullying, and other forms of digital violence are cited as the "wicked problems" of the digital risk society [57,58].
The findings of this study, as an example of the weakened socioemotional ties caused by digitalization, can be understood as another manifestation of those potential risks. Furthermore, the identification of the loneliness inequality effect leads us to the issue of digital divide, another wicked problem of the digital risk society.
For sustainable use of digital technology, a balanced and realistic consideration of its opportunities and risks is essential. Such an effort can be a cornerstone of effective policy interventions and wise social practices to maximize technology's advantages and minimize its adverse effects. This study is also expected to stimulate meaningful discussions on this broader dimension, beyond the narrow, technical field of elementary and secondary education.