stata代写-1500XXXX
时间:2022-08-09
Abstract:
Accompanied by the rapid popularity of the information technology, the imparity
of the access to the Internet has been raising people’s concern. As the largest developing
country in the world, the uneven distribution of the availability to the Internet has been
a significant problem in recent years. Following the existing literature, we examine the
effect of Internet use on people’s cognition ability, which is one of the key
measurements of human capital. Based on the 2010 and 2014 Chinese Family Panel
Studies dataset, we estimated the causal effect of using the net on one’s cognition,
measured by verbal and math test scores using instrumental variables indicating the
Internet use status in the respondent’s community. Also, we examined the channels by
which the net users benefit from the Internet. Furthermore, we use quantile regressions
to show that the difference of cognitive ability between net users and non-users is
largest for people with relatively poor cognitive ability, and we suggest that the digital
divide of cognition can become smaller if we can give more policy support for the
access to the Internet for people with low human capital.
Key words: Internet, cognition, digital divide
2
1. Introduction
Since 1990s, information technology such as Internet has been experiencing a rapid
growth, which brings us in an information era with endless possibilities. The
development of Internet is not only a technical progress, but a strong social, economic
and cultural force, making revolutionary changes to people’s lifestyle and social
development. As the Internet technology is playing a more and more important role in
the economy, the inequality of Internet use has been a significant problem. The people
with higher socioeconomic status tend to have more access to Internet. This kind of
imparity use of information technology is called “digital divide”. Since such a
conception was put forward (OECD, 2001), it has aroused great concern from various
disciplines, including economics, sociology, psychology and so on (Li and Xie, 2017).
There is much existing literature about the impact of information technology on the
wage inequality, and what we want to explain is the mechanism behind it. Human
capital is an essential issue in labor economics. Can Internet use impact the human
capital, then contributing to the income? Cognitive ability is a widely-use measurement
of one’s human capital, thus it is important to know whether the use of Internet
influence one’s cognitive ability, which is the main focus of this study.
While a large body of literature has shown that the use of Internet poses a significant
role to wage and human capital (e.g., Kim, 2012; Brynjolfsson, 2014; Park and Lee,
2015; DiMaggio and Bonikowski, 2008), knowledge about the potential outcome of
Internet use on cognitive ability is more limited. Poor cognitive function may have
profound social, economic and health implications (Lang et al, 2008). While some
studies have explored the link between Internet use and cognition (Jeremy et al., 2006;
Nicole et al, 2010; Johnson, 2011; Gáliková Tolnaiová and Gálik, 2015), several
challenges plague the empirical identifications.
First, omitted variables correlated with noth cognition and use of Internet may bias
estimation. Most studies, except for Ofer and Cristian (2011), do not account for
endogeneity problem and individual-level heterogeneity. For example, Poster (2013)
only controlled for several demographic variables such as marriage and race and
educational information. In this study, we are able to remove the individual-level
3
unobservable factors by using a longitudinal dataset — the China Family Panel Studies
(CFPS). Moreover, we try to use the IV method to estimate the casual effect between
Internet use and cognition.
Second, most cognitive tests in previous studies were administered to limited
cohorts, such as students (Ofer and Cristian, 2011; Johnson, 2011) or old people
(Jeremy et al., 2006; Nicole et al, 2010). It is not clear whether the findings inferred
from these specific groups hold true for the population as a whole. The cognitive tests
in our nationally representative sample cover nearly all ages above 10, which enable us
to test if there is age heterogeneity in cognition.
Third, most economic studies on this topic have been silent about the heterogeneity
among people with various cognitive ability levels. Although the gender gap in
cognitive performance has been studied (Johnson, 2011), the educational and
socioeconomic status heterogeneity of Internet premium on cognition are still unknown
to us. In this study, we use quantile regression method to see whether the effect of using
Internet varies with the cognition ability level.
2. Data
2.1 Data source
CFPS (China Family Panel Studies) is a nationally representative longitudinal
survey designed and implemented by the Institute of Social Science Surveys (ISSS) of
Peking University. This survey was conducted in 25 Chinese provinces (these provinces
jointly cover 95% of the Chinese population) in five years (2008, 2009, 2010, 2011,
2012). In each wave, the CFPS survey samples about 15,000 households nationwide
using the multi-stage probability proportional to size (PPS) sampling method, and all
family members in each sample household are included. The questionnaire collects
individual-, family-, and community-level information on the demographic,
socioeconomic and health-related variables.
CFPS contains a series of questions about Internet use status. CFPS (waives 2010,
2014 and 2016) collects the Internet use behavior of nearly all the respondents of no
less than 10 years old. The four measurements of the Internet use status are Internet
4
popularity rate, the length of time online, the activities on the Internet and the evaluation
of importance of surfing the Internet. In particular, the rate of Internet popularity is the
proportion of Internet users in the whole population; the length of time is the average
time Internet users spent online each week; the activities mainly contain the frequency
of study, work, social life, entertainment and commercial activity like shopping online;
the importance of Internet means how important the subject thinks Internet as an
information channel, and the importance of different online activities.
CFPS 2010 and CFPS 2014 contain the same cognitive ability module, i.e., 24
standardized mathematics questions and 34 word-recognition questions. All these
questions are obtained from standard textbooks and are sorted in ascending order of
difficulty. The starting question depends on the respondent’s education level. The test
ends when the individual incorrectly answers three questions in succession. The final
test score is defined as the rank of the hardest question a respondent is able to answer
correctly. If the respondent fails to answer any questions during the test, his or her test
score is assigned as the rank of the starting question minus one. For example, a
respondent with middle school education begins with the 9th question in the verbal test.
If the hardest question he is able to correctly answer is the 14th question, then his verbal
test scores would be 14. However, if he fails the 9th, 10th, and 11th questions
consecutively, his verbal test scores would be 8.
The CFPS is suitable for our study for several reasons. First, the survey includes
several standardized cognitive tests. Second, detailed information about the Internet use
behavior and attitude is available to us, enabling us to test the relation between Internet
and cognition comprehensively. Third, the longitudinal data allow us to remove
unobserved individual factors that may bias estimates. Further, the survey embodies
rich information at multiple levels, allowing us to control for a wide range of covariates.
Finally, because the cognitive tests are administered to all age cohorts older than 10, we
can study the effects of Internet on different age groups.
We restrict our sample to people above 10, and further drop the observations with
missing information on the key variables such as verbal test and math test score, Internet
use information, gender and age. Our final study sample contains 45,761 individuals,
5
with 71,087 observations covering China’s 25 provinces, providing a rather large
sample for relatively accurate estimations.
2.2 Descriptive analysis
Table 1 presents the key variables and their summary statistics. We include the two
key dependent variables in our research here, the respondent’s verbal scores and math
scores. We also put here some demographic and socioeconomic factors, such as one’s
age in years, gender, urban-rural status, education years and annual household income
which has been adjusted so as to be comparable between year 2010 and 2014.
[Insert Table 1 here]
According to Table 1, 49% of the observations are male, and 46% are from urban
areas. The average age of the whole sample is 43, while it is 29 for Internet users, and
48 for the other, which is quite consistent with the common sense. 25% of the
observations are net users, which is much lower than the average level in developed
countries about 70%, indicating the great developmental potential in the Internet
industry in our country.
Obviously, there is striking difference between the two subsample groups. The
Internet users have over 5 years more education than non-users. Moreover, the netizens
are more wealthy, more likely to live in urban area, and tend to have higher cognitive
ability. All these difference are statistically significant by t-test between the two groups,
which will get more accurate estimation in our following regression analysis.
3 Method
3.1 Baseline regression
According to the related literature and our descriptive statistic results, our baseline
econometric specification is as follows:
it it it it it i t itScore User F T X (1)
The dependent variable itScore is the cognition tests scores of respondent i in
6
year t . The key variable itUser indicates whether a person use the Internet, and
vector itF means the frequency of several online activities. itT is the length of time
online in the subject’s spare time. We control for a set of demographic correlates itX ,
including gender, age and its square terms, log form of household income that year,
years of education, marriage status and province dummies. i denotes individual
fixed effects. t indicates year fixed effects. it is the error term.
Since we have panel data, we use FE model to control for the individual
heterogeneity, which results will be compared to the RE results to make more robust
conclusions.
3.2 Instrumental variable (IV) regression
The above models simply assume that an individual’s cognitive ability level is
exogenous. However,the cognitive ability measure may suffer from endogeneity
problem because of the following reasons: (1) unobserved factors such as lifestyles and
motivation for study can affect both the cognitive ability and one’s use of the Internet;
(2) higher cognitive ability can contribute to higher efficiency to acquire information
and knowledge online, which in turn boosts one’s Internet use.
To address the above endogeneity concern due to the variable omission or reverse
causality, we use the community level Internet use status as our instrumental variable,
which specifically means whether the community has its own service website. On the
one hand, the Internet availability is clustered at the community level, which reflects
the typical characteristic in the rapid Internet penetration process in China. A village or
a community is often the basic unit of the penetration of Internet, where most household
becoming available to the Internet almost at the same time. Moreover, the more people
surfing Internet in a community and the more frequent they use net, the more likely a
community has its own website for the everyday service online. Thus, the existing of
the community service website is highly correlated with individuals’ net use behavior
in that community. On the other hand, whether one’s community has its service website
is not directly correlated to his cognition development, except for the channel of his
7
Internet use decision. In a word, we argue that the service website status of the
individual’s community is a good IV for our estimation.
Naturally, however, it is impossible to control for all possible variables that might
be correlated with community net use status and one’s use of Internet. We deal with
such problem by using a simple overidentification test use measures of the penetration
rate of the Internet as additional instruments. We then use overidentification tests to
detect whether the community net use status has a direct effect on people’s cognitive
ability.
3.3 Quantile regression
To deal with the heterogeneity further and show clearer policy implications, we will
use the quantile regression method to examine the effect of Internet use on different
levels of cognitive ability. Is the availability to the net more important for those who
have high cognitive ability or those with lower cognitive ability? And where does the
inequity mainly come from? To try to give answers to these questions, we try to do
quantile regressions on the .10th, .25th, .50th, .75th, and .90th quantile points.
4 Results and discussion
4.1 Baseline regression results
Table 2 reports the results of the baseline regression on the two measures of the
cognitive ability. The first four columns reports the results for verbal scores and the
other four for math scores. The key explanatory variable is whether the individual uses
Internet, which is a dummy variable. The results consistently show that there is a
significantly positive relationship between the use of Internet and the cognitive ability
measures.
[Insert Table 2 here]
For verbal test scores, column (1) and (2) both only contain the key explanatory
variables, while column (3) and (4) add some control variables including one’s age,
gender, urban-rural status, the log of the household annual income, education years and
8
also province dummies. We can see from the results that in FE model in column (1),
which controls for individual heterogeneity, the coefficient is quite smaller than that in
the RE model in column (2). After control for other variables in column (3) and (4), the
coefficients of the two model are similar, and at the same time similar to that in column
(1), which indicates the results of these three columns are more convincing than column
(2). From the coefficient in the FE model with control variables, net users’ verbal score
is 1.6 higher than the non-users on average, which is about 0.15 standard deviation
(according to Table 1).
Similarly, we can detect the same pattern of the coefficients in column (5) to (8),
which contain results for FE and RE regression results on math test scores. In column
(7), the coefficient is 0.439, which indicates that holding other factors constant, net-
users have 0.44 higher math scores than non-users, which is about 0.07 standard
deviation. The effect of Internet use on math is much lower than that on verbal scores,
which is quite consistent to the reality, since compared those who are not available to
Internet, the net users have more convenient channel for information and they read more
words and learn many new words, while the improve on math ability is relatively
smaller and not so direct as verbal ability.
With regard to other variables, we find that the education years are significantly
positively correlated with one’s cognitive ability, whether in FE models or in RE models.
The effect of age is negative, which may be explained by the biology mechanism of
one’s cognition development and the improvement in people’s education in China as
well. The effect of the household income is not that consistent, which is significantly
positive in FE model and not significant in RE for verbal scores, while negative for
math scores in RE model.
4.2 IV regression results analysis
Table 3 reports the IV regression results. Here our key explanatory variable is
whether a person use Internet, and we use two instruments, whether the community
where the person lives has its own service website, and the Internet penetration rate of
that community, which measures the percentage of net users among all people there.
Since we have panel data here, we use the systematic GMM method for our IV
9
estimation, which can provide consistent estimation. Note that the systematic GMM
model with IV only apply to the balanced panel data, so the observation here is much
smaller than that in the baseline regression, restricted to the individuals whose
information are available for both 2010 and 2014 dataset. In systematic GMM model,
we also controlled for individual fixed effect, which allow us to compare these results
with the FE results in baseline regressions.
[Insert Table 3 here]
Here it is easy to see that after control for some co-variables, the effect of using
Internet are larger than that when there are no control variables, for both test scores.
But maybe it is not comparable between column (1) and (2), and between column (3)
and (4), for we lose many observations when adding controls since the number of
individuals with information for all these variables are fewer. The two results based on
different samples, which make them likely to be not comparable. But all the coefficients
here for the key explanatory variable are significantly positive, which indicate there
does exists a positive effect from the net use on one’s cognitive ability. We can compare
column (1) in Table 3 with column (1) in Table 1, both in FE models and with no
controls. We can find that the coefficients are similar. The same logic applies to column
(3) in Table 3 and column (5) in Table 1, which tells us the same story. Since the IV
results are quite similar to the baseline results, our IV estimation is more convincing.
Also, the weak-IV test F-value in all the four models are rather high, which indicates
that there is no weak IV problem. (Stock & Yogo, 2002) Moreover, in the
overidentification test, the p-values for the models are all higher than 0.1, proving that
our two instruments are both exogenous statistically.
4.3 Channel analysis
In the baseline results and after we addressed the endogenous problem in IV
regression, we can conclude that being exposed to the Internet has positive effect on
one’s cognitive ability, measured by the word test and math test scores. The following
question is: how can the use of Internet impact one’s cognitive ability, and what is the
10
mechanism behind the causal relationship? This is the question that we want to give
answer to in the following part.
In the questionnaire of CFPS 2014, the Internet part is richer than waive 2010,
maybe because of the revolutionary change and rapid penetration of Internet in China
these years, especially the popularity of the mobile Internet and smart phone. In 2014
questionnaire, there are some questions about the concrete usages of the net, which are
not contained in waive 2010 but can be used to examine the channel mechanism of the
relationship between the net use and cognitive ability. There are five questions about
the frequency of the usages of Internet, studying, working, social activity, entertainment
and business activity. For example, “how often do you use the Internet for study?” The
answers contains 7 options: almost every day, 3 to 4 times a week, once or twice a week,
two or three times a month, once a month, once several months, and never. We denoting
the frequency as 1 to 7 from low to high. So we have five consecutive variables here
indicating the frequency of different usages of the net.
We use the five frequency variables as independent variables, and the scores of
word test and math test as dependent variables, with the sample restricted to the net-
users. The results are reported in Table 4, part A and part B containing the regression
results on word scores and math scores, respectively. In the first five columns of each
part, we add only one frequency variable, and we put all the five variables in the last
column.
[Insert Table 4 here]
It is easy to see that for word scores, in column (1) to (5), all the coefficients are
significant except for the business activity. The coefficient of study is significant and
keeps consistent both in value and significance in column (1) and (6), which matches
our intuition very much, since study itself can improve one’s cognitive ability, and the
use of the net provides more study resources and infinite knowledge. The coefficient of
social activity is also keeping highly significant in both column (1) and (6), since the
more people enjoy social life online, the more they are exposed to the new words and
11
try to use more words which maybe fashionable. The coefficients of entertainment is
also highly significant, which can be explained by similar mechanism to the effect of
social life online. Especially in these years, as smart phone and social software such as
QQ and Wechat are more and more common used in China, the social life and
entertainment are fixed together to some content, which also leads some people who
seldom read paper books to read more online, improving their cognitive ability
measured by word test scores. In column (6), however, the coefficients of work and
business is not significant, which may because there is relatively few people in our
sample who works or does business work online frequently and thus the variation is
somehow small. Also, there is another possibility that for those who can use the net for
work and business, they are most likely to be white-collar workers, who have been well
educated. Their cognition ability is rather high already, and using net more frequently
may not improve their word test scores further.
The results for math scores are very different from the above verbal scores’ results.
The only one significant is the coefficient of study. Naturally, there are some students
in our sample who use the resources online to study, and their cognitive ability
improved as a whole, whether measured by verbal scores or math scores. Also, since
here it is only OLS results, there may be endogenous problem, for those who have high
cognitive ability tend to use the net to study more frequently. Here we just observe the
positive correlation and we cannot give the explanation for the causal relationship. But
the key point here is those who are not available to Internet cannot use it for study, and
maybe that limits their potential for further improvement on the cognitive ability
measured by math scores. It is interesting to note that in column (6), after putting all
the five frequency variables together, we find that the frequency of use Internet for work
is negatively correlated with the math scores. We have analyzed in the above part for
the verbal scores that using Internet more cannot have further marginal significant effect
on one’s verbal ability for those who can use Internet for work. While here we even
find that there may be negative effect from the maybe overuse of the Internet on one’s
math scores. This kind of results are reflected in some research in psychology and
sociology fields about the negative effect of the overuse of the Internet (e.g. Davis,
12
2001; Bozoglan et al.,2014).
4.4 Quantile regression results
Since we can use quantile regression to analyze the effect of Internet on cognitive
ability on different cognitive ability levels, we can find the source of the imparity of
Internet use, “digital divide”, more concretely. The results of the quantile regressions
are reported in Table 5.
[Insert Table 5 here]
It is very clear that the coefficient of using Internet is most large for those with
almost lowest cognitive ability level, .10th quantile point, measured by both verbal tests
and math tests. As for the quantile points after the median, the most coefficients are
also significant but rather small. So the effect of the availability to Internet is almost
significantly positive for all the people. However, the people with poor cognition are
where the rub is. These people may contain those who are less educated, those who are
in rural areas, and those who does simply physical work. Due to the low income to pay
and the bad condition of the infrastructure including the net penetration condition where
they live, it is hard for them to be exposed to the internet, which in turn limits their
accumulation of human capital and increase the imparity of the income distribution.
5 Conclusions
This paper examines the effect of using Internet on people’s cognitive ability in
China, contributing to the growing literature on the effect of ICT (information and
communication technology) on human capital and income distribution. As far as we
know, this paper provides the first nationally representative estimates on the cognitive
ability improvement induced by using Internet in China, the largest developing country
in the world. Based on the data obtained from 2010 and 2014 Chinese Family Panel
Study, our research provides evidence on the existence of the digital divide in cognitive
ability with several major findings:
First, the estimated results show that the availability to the Internet has significant
13
and positive impact on one’s cognitive ability, measured by verbal and math test scores.
Moreover, these impacts are not just a correlation, but can also be interpreted as a causal
relationship. By employing the instrumental variable regressions, our study finds that
the positive effect still holds and remains statistically significant after considering the
endogeneity of one’s cognition test scores.
Second, we examined the channels by which net users can benefit from the use of
Internet. By using a series of variables indicating frequencies of different usages, we
find that the effects are significant for the usages of study, social life and entertainment
for the verbal scores, and only study for math scores.
Third, by quantile regression, we find that the effect of Internet use on one’s
cognitive ability are not uniform across people with different cognitive ability levels.
The effect is the largest for those whose verbal and math test scores are relatively low,
who are the people with highest potential of cognition improvement after being
available to the Internet.
According to the above findings, we can give several policy implications as follows:
first, the divide of human capital caused by the uneven distribution of the access to the
Internet does exist in China, and we should try to improve the penetration rate of the
net in the whole country, both by technical improvement and policy support, which
would generate more human capital and decrease the inequity of the income distribution.
Second, it is the time to guide people use the Internet more properly and purify the
online environment, making a healthy atmosphere for online study and communication
and making the net playing a more positive and important role in people’s all-round
development. Third, we should especially shed light on those with lower human capital
and relatively poor cognitive ability, whose human capital development has been
limited due to the inaccessibility to the Internet. Give more policy and financial support
for the less wealthy rural area and underdeveloped regions such as the west part of
China, boosting the construction of ICT infrastructure and making the penetration rate
of Internet higher, and that will decrease the regional and urban-rural imparity of human
capital and income distribution.
14
Reference:
Bozoglan B., Demirer V. and Sahin I.: Problematic Internet use: Functions of use,
cognitive absorption, and depression. Computers in Human Behavior 2014, 37: 117-
123.
Brynjolfsson, E.: The Great Decoupling: How Information Technology Is
Affecting Productivity, Wages and Employment. American Association for the
Advancement of Science 2014.
Davis R.A. A cognitive-behavioral model of pathological Internet use. Computers
in Human Behavior 2001, 17: 187-195.
DiMaggio, P. and Bonikowski, B. Make Money Surfing the Internet? The impact
of Internet Use of Earnings on the U.S workers. American Sociology Review 73: 227-
250.
Gáliková Tolnaiová, S., & Gálik, S: Influence of the internet on the cognitive
abilities of man. phenomenological and hermeneutical approach 2015, 12.
Jeremy, F., Salvador, R. and Eszter, H.: Cognitive ability and Internet use among
older adults. Poetics 2006, 34: 236–249.
Johnson, G.M.: Internet Activities and Developmental Predictors: Gender
Differences Among Digital Natives. Journal of Interactive Online Learning 2011, 10(2):
64-76.
Kim, S.: Two faces of comprehensive information technology, knowledge
acquisition, and polarized wage structure. Journal of Economic Research 2012, 17(2),
139-158.
Lang, IA., Llewellyn, DJ., Langa, KM., Wallace, RB., Huppert, FA., and Melzer,
D: Neighborhood deprivation, individual socioeconomic status, and cognitive function
in older people: Analyses from the English Longitudinal Study of Ageing. Journal of
the American Geriatrics Society 2008, 56: 191–198.
Li, W. and Xie, Y. Internet Inequality. In China's Livelihood Development Report
2016, edited by Xie Yu et al. Beijing: Peking University Print.
Nicole, W., Khaled H., and Milena H.: Computer use by older adults: A multi-
disciplinary review. Computers in Human Behavior 2010, 26: 870-882.
15
Park, J., & Lee, B.: Wage inequality in the digital economy: disentangling the
impacts of information technology and communication technology. Social Science
Electronic Publishing 2015.
Ofer, Malamud and Cristian, Pop-Eleches: Home Computer Use and the
Development of Human Capital. The Quarterly Journal of Economics 2011, 126(2):
987-1027.
Organization for Economic Co-operation and Development (OECD):
Understanding the Digital Divide. Paris FR 2001. (http:
//www.oecd.org/sti/1888451.pdf)
Stock, J. H., & Yogo, M. (2002). Testing for weak instrument in linear IV
regression. NBER Technical Working Paper, 14, 80-108.
16
Table 1: Summary statistics
Variable Definition All Sample Using Internet Not Using Internet
Verbal scores Verbal scores
17.50 26.34 14.55
(10.80) (5.85) (10.46)
Math scores Math scores
10.14 15.55 8.33
(6.56) (4.70) (6.08)
age age in years
42.63 28.60 47.76
(18.62) (12.09) (17.60)
urban
urban-rural status
(urban=1)
0.46 0.65 0.40
(0.50) (0.48) (0.49)
male gender (male=1)
0.49 0.54 0.47
(0.50) (0.50) (0.50)
log(income)
log form of annual
household income
4.92 6.46 4.51
(5.04) (5.10) (4.93)
edu education years
7.13 10.93 5.83
(4.65) (3.36) (4.29)
observations 71087 17784 53303
Note: 1) Data Source: China Family Panel Studies, 2010 and 2014. 2) The reported statistics are
coefficients with heteroscedasticity robust standard errors in the parentheses.
17
Table 2 Baseline Results
Verbal test Math test
(1) (2) (3) (4) (5) (6) (7) (8)
FE RE FE RE FE RE FE RE
Using internet 1.867*** 8.371*** 1.598*** 1.358*** 0.846*** 4.219*** 0.439*** 0.482***
(0.1130) (0.0769) (0.1555) (0.0723) (0.0689) (0.0526) (0.0843) (0.0386)
age -0.105*** -0.135*** -0.060*** -0.042***
(0.0191) (0.0019) (0.0086) (0.0009)
male 0.465 0.996*** -0.104 0.428***
(1.7469) (0.0612) (0.7844) (0.0303)
urban 0.021 0.295*** 0.062** 0.159***
(0.0429) (0.0308) (0.0255) (0.0157)
ln_inc 0.036*** 0.002 0.010* -0.006**
(0.0110) (0.0059) (0.0051) (0.0029)
eduyear 0.764*** 1.424*** 0.603*** 1.082***
(0.0601) (0.0080) (0.0384) (0.0042)
Province dummies No No Yes Yes No No Yes Yes
Constant 17.032*** 15.472*** 14.463*** 9.927*** 9.927*** 9.159*** 7.030*** 4.334***
(0.0283) (0.0533) (0.3141) (0.0172) (0.0172) (0.0319) (0.9595) (0.1718)
Observations 71,087 71,087 56,093 56,093 71,085 71,085 56,091 56,091
R-squared 0.010 0.010 0.041 0.036 0.008 0.008 0.074 0.069
N of individuals 45,761 45,761 41,975 41,975 45,761 45,761 41,975 41,975
Note:
1) Data Source: China Family Panel Studies, 2010 and 2014. 2) The
reported statistics are coefficients with heteroscedasticity robust
standard errors in the parentheses.
3) The dependent variable in
column (1) to (4) is the respondent’s verbal test score, and in column
(5) to (8) it is math test score.
18
Table 3 IV Regression Results
Verbal test Math test
(1) (2) (3) (4)
Using internet
1.478*** 5.924*** 0.307* 2.345***
(0.3508) (0.5979) (0.1743) (0.2847)
Controls No Yes No Yes
Observations 50,652 28,236 50,648 28,232
Number of individuals 25,326 14,118 25,324 14,116
Weak IV test F-value 1660.25 591.60 1660.86 592.19
Overidentification test p-value 0.581 0.107 0.890 0.374
Note: 1) Data Source: China Family Panel Studies, 2014. 2) The reported statistics are coefficients
with heteroscedasticity robust standard errors in the parentheses. 3) The dependent variable in
column (1) and (2) is the respondent’s verbal test score, and in column (3) and (4) it is math test
score.
19
Table 4 Channel Analysis
Part A: Verbal test score
(1) (2) (3) (4) (5) (6)
Study 0.164*** 0.154***
(0.0285) (0.0302)
Work 0.079*** 0.010
(0.0279) (0.0304)
Social activity 0.118*** 0.089***
(0.0323) (0.0340)
Entertainment 0.098*** 0.061*
(0.0349) (0.0372)
Business 0.056 -0.017
(0.0368) (0.0387)
Constant 19.085*** 19.478*** 18.809*** 19.001*** 19.500*** 18.374***
(0.4253) (0.4207) (0.4604) (0.4604) (0.4160) (0.4798)
Controls Yes Yes Yes Yes Yes Yes
Observations 7,188 7,186 7,186 7,187 7,185 7,181
R-squared 0.279 0.276 0.277 0.276 0.276 0.281
Part B: Math test score
(1) (2) (3) (4) (5) (6)
Study 0.135*** 0.159***
(0.0160) (0.0168)
Work -0.006 -0.058***
(0.0160) (0.0173)
Social activity -0.016 -0.019
(0.0178) (0.0187)
Entertainment 0.011 0.020
(0.0192) (0.0202)
Business -0.019 -0.031
(0.0216) (0.0226)
Constant 4.212*** 4.546*** 4.652*** 4.498*** 4.598*** 4.225***
(0.3045) (0.3064) (0.3207) (0.3228) (0.3096) (0.3319)
Controls Yes Yes Yes Yes Yes Yes
Observations 7,187 7,185 7,185 7,186 7,184 7,180
R-squared 0.578 0.574 0.574 0.574 0.574 0.580
Note: 1) Data Source: China Family Panel Studies, 2014. 2) The reported statistics are coefficients with
heteroscedasticity robust standard errors in the parentheses. 3) The dependent variable in Part A is the respondent’s
verbal test score, and in Part B it is math test score. 4) The main independent variables are the frequency of different
usages of the Internet.
20
Table 5 Quantile Regression Results
Part A: Verbal scores
(1) (2) (3) (4) (5)
.10th .25th .50th .75th .90th
Using internet 6.294*** 2.709*** 0.395*** 0.313** 0.685***
(0.1273) (0.0975) (0.0965) (0.1229) (0.1258)
Controls Yes Yes Yes Yes Yes
Observations 56,093 56,093 56,093 56,093 56,093
Part B: Math scores
(1) (2) (3) (4) (5)
.10th .25th .50th .75th .90th
Using internet 2.548*** 0.289*** 0.181*** 0.077 0.288***
(0.0951) (0.0510) (0.0427) (0.0505) (0.0677)
Controls Yes Yes Yes Yes Yes
Observations 56,091 56,091 56,091 56,091 56,091
Note: 1) Data Source: China Family Panel Studies, 2014. 2) The reported statistics are quantile regression
coefficients with standard errors in the parentheses.