Demand for MOOC - An Application of Big Data

Demand for MOOC - An Application of Big Data

Accepted Manuscript Demand for MOOC - an application of Big Data Tingting Tong, Haizheng Li PII: DOI: Reference: S1043-951X(17)30071-8 doi: 10.1016/...

879KB Sizes 10 Downloads 95 Views

Accepted Manuscript Demand for MOOC - an application of Big Data

Tingting Tong, Haizheng Li PII: DOI: Reference:

S1043-951X(17)30071-8 doi: 10.1016/j.chieco.2017.05.007 CHIECO 1060

To appear in:

China Economic Review

Received date: Revised date: Accepted date:

1 January 2017 5 May 2017 15 May 2017

Please cite this article as: Tingting Tong, Haizheng Li , Demand for MOOC - an application of Big Data, China Economic Review (2017), doi: 10.1016/ j.chieco.2017.05.007

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT Demand for MOOC- An Application of Big Data*

CR

IP

T

Tingting Tong (Corresponding author) Dongbei University of Finance and Economics International Business College No.217, Jianshan Street, Shahekou District, Dalian, 116025, China Email: [email protected]

M

AN

US

Haizheng Li Georgia Institute of Technology School of Economics 221 Bobby Dodd Way, Atlanta, GA 30332 Email: [email protected]

AC

CE

PT

ED

October 2016

*

Partial financial support is from the National Natural Science Foundation of China (Grant #71273288). We gratefully acknowledge valuable comments from Patrick McCarthy, Usha Nair-Reichert, and Olga Shemyakina, as well as from session participants at conferences and workshops.

ACCEPTED MANUSCRIPT Demand for MOOC-An Application of Big Data Abstract We evaluate factors affecting the demand for MOOC by estimating its demand function in OECD countries and in China. We apply a Big Data approach to construct a proxy for MOOC

T

demand using Google Trends for OECD and Baidu Index for China. Based on estimation results

IP

of various panel data models, we find that in both cases, higher unemployment promotes MOOC

CR

demand. However, in OECD countries, the proportion of individuals with high school level or higher education have positive and significant effects on MOOC demand, while in China, we

US

observe positive and significant effects from internet speed and average income.

AN

Keywords: Online education, MOOC, Big Data, Google Trends, Baidu Index

AC

CE

PT

ED

M

JEL Codes: I21, I28

ACCEPTED MANUSCRIPT I.

Introduction Massive open online course (MOOC) refers to tuition-free online courses, in which

anyone with internet access can enroll. The New York Times declared 2012 as “The Year of the MOOC.” In 2013, Georgia Institute of Technology launched an online Computer Science

T

Master’s degree by cooperating with the MOOC provider Udacity. In 2015, Arizona State

IP

University (ASU) started to offer MOOC-based academic credits from edX. In 2014, the number

CR

of universities offering MOOCs had increased to 400, and 22 of the top 25 U.S. universities on U.S. News & World Report’s rankings are now offering courses online for free. The total

US

number of courses in the MOOC platform has raised to 2,400.1

AN

The rapid development of MOOC raises a number of important questions. Will MOOC maintain its rapid growth in the future? Will it substitute the traditional face-to-face education

M

and revolutionize future education? The rising cost of traditional education has been one of the

ED

greatest challenges for educators and policy makers.2 Administrators have adopted several costmanagement actions such as increasing class size, raising teaching loads, and reducing support

PT

staff. However, the college tuition cost is still viewed as expensive by the general public (Bass

CE

2014).

The development of online education, especially MOOC, has been considered as a

AC

promising way to solve these problems. MOOC’s absence of tuition and fees challenges the economic model traditionally subscribed to colleges and universities. This new form of 1

Shah, D. 2014. Online Courses Raise Their Game: A Revie of MOOC Stats and Trends in 2014. https://www.class-central.com/report/moocs-stats-and-trends-2014/ 2 Among OECD countries, tuition fees charged by their tertiary institutions differ significantly. The tuition is particularly high in the United States. For the academic year 2016-2017, the average tuition and fees was $9,650 for public four-year colleges, and $33,480 for private non-profit four-year colleges. This accounts for 17% and 58% of its GDP per capita, respectively. Between 2011-12 and 2016-17, published tuition fee prices rose by 9% in the public four-year sector, and by 13% at private nonprofit four-year institutions, after adjusting for inflation (College Board 2016). In China, Li and Liu (2014) have shown that the average tuition per student accounted for 46% of GDP per capita in 2002. Financial constraints have hindered students from attending higher education in China (Wang et al. 2014).

1

ACCEPTED MANUSCRIPT education allows individuals to take free yet high quality courses online, even for an entire degree. Deming et al. (2015) found that institutions with more online students do charge lower prices. In addition, according to Christiansen and Horn (2011), the development of online education such as MOOC will bring a “disruptive innovation”, in which higher education will

T

eventually become more convenient, accessible, and significantly cheaper. Some scholars

IP

believe that MOOCs will bring a revolution to, or even substitute, traditional high-cost education

CR

(Barber et al. 2013).

In addition to the low cost, MOOC has the advantage of openness and flexibility, which

US

challenges the closed and privileged nature of the traditional higher education (Krause and Lowe

AN

2014). Individuals now have better access to education without the limit of geography, time, and financial constraints. As a result, MOOC is especially important for lifelong learners and

M

working professionals who have a high opportunity cost of time (Schuwer et al. 2015). MOOC

ED

may also speed up graduation in large institutions. By taking some courses online, students do not have to wait for another semester if they did not register successfully (McPherson and Bacow

PT

2015). Moreover, MOOC broadens the access to higher education by offering courses to

CE

individuals in distant locations. People from less developed areas now have access to courses from the most prestigious universities in the world, and thus can receive high quality education

AC

that facilitates their human capital accumulation (Clotfelter et al. 2010). For the learning effectiveness of MOOC, existing empirical studies (although very few) do not find statistically significant differences between online or hybrid courses and traditional courses (McPherson and Bacow 2015). For example, Bowen et al. (2014) compared learning outcomes of traditional and hybrid versions of a statistics course in a randomized trial, and they concluded that well-designed online courses can deliver equivalent educational outcomes. In

2

ACCEPTED MANUSCRIPT addition, in the study of the University of Maryland system that incorporated MOOC content into hybrid courses, students in hybrid sections did as well as or slightly better than students in traditional sections in terms of passing rates and learning assessments (Griffiths et al. 2014). There are, however, concerns about MOOC. More specifically, MOOC has a low

T

completion rate, and a significant number of MOOC students stop taking MOOC courses after

IP

the first one or two weeks.3 As class size expands, it is hard for professors to give individual

CR

attention and progressive feedback in online classes (Bass 2014). Moreover, MOOC-based learning lacks face-to-face interactions between students and instructors, as well as among

US

students themselves. It thus relies on a student’s self-discipline (Banerjee and Duflo 2014).

AN

Despite the optimism or concerns, the future of MOOC depends on public interest (or preference), the perceived value, and its explicit or implicit costs. In other words, its future will

M

be determined by the demand for online courses and degrees. Given the unique features of

ED

MOOC, its demand will be determined by many factors. Among them, the effect of unemployment is particularly interesting due to its policy implications. On one hand,

PT

unemployment could possibly reduce the demand for MOOC as unemployed individuals

CE

spending time looking for a new job. On the other hand, unemployed individuals have lower time cost, and free and flexible MOOCs may be an effective means to improve their skills. We

AC

conjecture that, given that other factors are constant, an increase in unemployment would have positive effects on the demand for MOOC. Additionally, MOOC demand structure is likely to be different between developed and developing countries. For instance, given the differences in relative income level and availability of education, the income effect of MOOC may be stronger in developing countries. We are also interested in testing the impact of a country’s general education level on the demand for MOOC. 3

Source: http://www.gse.upenn.edu/pdf/ahead/perna_ruby_boruch_moocs_dec2013.pdf

3

ACCEPTED MANUSCRIPT If the general education level increases, the potential user pool for MOOC is larger and thus may increase its demand. However, the effect may be different between developed and developing countries because of the relatively smaller pool of highly educated individuals in developing countries. Moreover, different internet infrastructure between developed and developing

T

countries may also generate different impacts on MOOC. Therefore, by evaluating MOOC

IP

demand in relatively more developed OECD countries and less developed China, an in-depth

CR

investigation into the above hypotheses will shed new lights on the demand for MOOC. Because MOOC is accessible from all over the world, we intend to study its demand from

US

an international perspective. In particular, we first study MOOC demand using cross country

AN

data from OECD countries; and then conduct a country level study using data from China. The comparison of MOOC demand between OECD countries and China may help generate some

M

insights about MOOC.

ED

OECD countries include major developed countries such as the United States and the United Kingdom, while China is the largest developing country. The average GDP per capita in

PT

OECD countries in 2015 is $36,520, but in China it is only $12,310. The difference in the

CE

economic development stage can affect the demand structure for MOOC. In addition, the average percentage of people with tertiary education out of the population (age between 25 and

AC

64) in OECD countries is 30.57%, while it is only 9.68% in China (in 2010). Such a large difference in higher education could have different impacts on the preferences for MOOC. When higher education is relatively scarce, people may value a traditional college education more because of the signaling effect or the prestigious social status. In China, for example, it is found that a tertiary degree obtained via part-time schooling has a lower return, and thus may be valued less compared to a regular college degree (Li et al. 2017). Moreover, the average internet speed

4

ACCEPTED MANUSCRIPT in OECD countries is 28.63 Mbit/s but only 6.01 Mbit/s in China (in 2015). When the speed is low, an improvement may have a larger impact on MOOC demand; however, its effect may vanish when the overall speed is already high enough like in OECD countries. The relevant findings will have different policy implications.

T

It is challenging to measure the demand for MOOC worldwide. One reason is that

IP

MOOC platforms (i.e., websites) can be accessed from all over the world at any time. As a result,

CR

traditional measures of education demand, such as the number of college applicants, cannot be used (e.g., Glewwe and Jacoby 2004). In recent years, the use of search engine data has drawn

US

increasing attention. With the rapid development of information technology, the number of

AN

internet users account for an increasing proportion of the world population. Their online search behavior has been stored and analyzed by search engines. For instance, researchers have used

M

such data to forecast various phenomena such as diseases, unemployment rates, tourist volumes,

ED

and the housing market (e.g., Ginsberg et al. 2009, Choi and Varian 2012, Yang et al. 2014). In this study, we adopt a Big Data approach by using internet search engine data to

PT

construct a proxy for the demand of MOOC in different geographic regions. More specifically,

CE

we use a search-volume index of MOOC related keywords generated by internet search engines such as Google. It is likely that the more searches for MOOC related words within a geographic

AC

region, the higher its demand. There is concern, though, that many of those searches may not represent serious interest and thus are not a good indicator of the demand. However, those who are seriously interested (i.e., those with true demand) will almost surely do the search. Moreover, because we use panel data, the systematic difference caused by searches will be differenced out if the distributions of search seriousness across years in a region are consistent (e.g., the probability of serious search and non-serious search is comparable across years). Therefore, we

5

ACCEPTED MANUSCRIPT believe that the search-volume index is a reasonable proxy for the demand for MOOC, especially when alternative measures are not available. This study adds to the literature by employing Big Data techniques to study the demand for MOOC, when alternative data are not available. This study is also among the first to estimate

T

factors affecting the demand for MOOC in both developed and developing countries.4 Our study

IP

aims to shed new light on the future of new online education delivery mechanisms. A good

CR

understanding of MOOC demand can help to develop future education policies and strategies for both governments and educational institutions in the era of information technology.

US

The rest of the paper is organized as follows. In Section II, we give a general description

AN

of MOOC. Section III illustrates factors that affect MOOC demand. In Section IV and V, we

II.

ED

respectively. Section VI concludes.

M

discuss empirical results of the demand for MOOC in OECD countries and in China,

About MOOC

PT

The term “MOOC” was first used in 2008 to describe the open online course

CE

“Connectivism and Connective Knowledge” (also known as CCK08) designed by George Siemens and Stephen Downes.5 CCK08 was presented to 25 tuition-paying students at the

AC

University of Manitoba in Canada, and was provided to 2,200 members of general public participants who took the class simultaneously for free with computers and internet. In 2011, two professors from Stanford University, Sebastian Thrun and Peter Norvig, designed the first American MOOC, “Introduction to Artificial Intelligence”, which attracted more than 160,000

4

In a traditional demand function, the key element is its own price, among other factors such as income, related goods, market size, and preference. MOOCs, however, are essentially free. Therefore, the demand function for MOOC differs from a traditional demand function. 5 Source: https://sites.google.com/site/themoocguide/3-cck08---the-distributed-course

6

ACCEPTED MANUSCRIPT students. Later, Thrun started up the company Udacity, a platform offering MOOC. Within a year, two other Standard professors, Andrew Ng and Daphne Koller, launched another MOOC platform called Coursera. In 2012, MIT and Harvard launched edX. In recent years, Coursera, Udacity, and edX have become the leading MOOC providers and are considered the “Big Three”

T

MOOC providers.6

IP

According to a report from Class Central, Coursera, Udacity, and edX offer more than

CR

half of all MOOCs in 2015.7 As of February 2016, Coursera had more than 1,800 courses and 18 million users from more than 190 counties. Coursera partners with over 140 schools all over the

US

world, such as Yale University in the U.S., University of Edinburgh in U.K., and Peking

AN

University in China; edX offers more than 850 courses and partners with over 90 institutions such as UC Berkeley and The University of Texas System; while Udacity offers more than 120

M

courses and partners with institutions such as San Jose State University and AT&T. 8 Among

ED

three platforms, Udacity concentrates on courses that provide in-demand skills for the workplace, such as computer science, programming, and software learning, while Coursera and edX offer

PT

courses in a variety of subjects. Most courses are offered in English with English subtitles, and

CE

many of them have subtitles in other languages such as Chinese, French, and Spanish. MOOC platforms have appeared all around the world, not just in the U.S. Well-known

AC

platforms include Eliademy in Finland, openHP in Germany, FutureLearn in UK, and OpenClassrooms in France. In China, the most popular MOOC platforms include edX, Coursera,

6

Coursera website: https://www.coursera.org/, Udacity website: https://www.udacity.com/, edX website: https://www.edx.org/. Article about the Big Three MOOC providers: http://www.nytimes.com/2012/11/04/education/edlife/the-big-three-mooc-providers.html?_r=0 7 Shah, D. 2015. Less Experimentation, More Iteration: A Review of MOOC Stats and Trends in 2015. https://www.class-central.com/report/moocs-stats-and-trends-2015/ 8 The numbers presented here are obtained from the official websites of Coursera, Udacity, and edX in February, 2016.

7

ACCEPTED MANUSCRIPT as well as local platforms such as XuetangX, IMOOC, and Chinese University MOOC.9 Specifically, XuetangX is the first Chinese MOOC platform offering courses in all subjects from top universities in China such as Tsinghua University and Nankai University. IMOOC is another popular Chinese MOOC platform offering courses mainly about information technology. While

T

international platforms such as Coursera provide courses to the Chinese from prestigious

IP

universities around the world, local platforms offer courses in the Chinese language.

CR

The format of MOOC is different from online video and TV courses. Most MOOC courses are separated into different sessions such as modules or weeks, and sometimes there are

US

quizzes following each session to keep students’ attention. Normally each session is composed of

AN

a series of short videos with various production styles, including presentation slides, or a classroom, studio, or office desk setup. For most courses, there are midterm and final tests, and

M

students can get a certificate of completion only if they pass the tests. Additionally, most MOOC

ED

platforms offer discussion forums where professors/teaching assistants can interact with students, and students are encouraged to form their own study groups on other social networks such as

PT

Facebook.

CE

Most MOOC courses are accessible for free. However, MOOC platforms charge fees if students need certificates to confirm that certain courses have been completed and related

AC

knowledge/skills have been mastered. For instance, edX offers Verified Certificates for certain courses with a fee ranging from $50 to $100 per course. 10 Similarly, Coursera offers Course Certificates with a fee ranging from $20 to more than $200 per course.11 Udacity charges fees for 9

XuetangX website: http://www.xuetangx.com/, IMOOC website: http://www.imooc.com/ Chinese University MOOC: http://www.icourse163.org/ 10 In addition to Verified Certificates, edX offers XSeries certificates, which are available when users successfully complete a series of verified courses that make up an XSeries https://support.edx.org/hc/en-us/articles/206212058-What-types-of-certificates-does-edX-offer11 In addition to Course Certificates, Coursera offers Course Specializations, which requires the successful completion of a series of courses and a final capstone project.

8

ACCEPTED MANUSCRIPT its Nanodegree program such as the Android Basics and Intro to Programming, which generally cost $199/month with most programs expected to take between 6-12 months. However, students do not have to enroll in the program to access course materials. MOOCs have attracted an increasing number of students. The total number of students

T

who signed up for at least one MOOC course reached 35 million in 2015, which is double the

IP

number in 2014.12 Although there is no formal study on the demand for MOOC, some

CR

quantitative information on MOOC users is available. In particular, among individuals taking MOOC, the most important reasons for taking a MOOC include gaining knowledge and skills

US

and promoting career development. Based on data from a random end-of-course survey in MIT’s

AN

first MOOC course, Circuits and Electronics, Breslow et al. (2013) reported that 55.4% of the surveyed students were participating for the knowledge and skills, 25.5% for personal challenge,

M

and 8.8% for employment or job advancement opportunities.

ED

In addition, most MOOC students are highly educated. Christensen et al. (2013) examined the data from an online survey of students enrolled in at least one of the University of

PT

Pennsylvania’s 32 MOOCs offered on the Coursera platform. The results showed that MOOC

CE

takers are highly educated, with 83.0% having a post-secondary degree, 79.4% having a Bachelor’s degree or higher, and 44.2% having education levels higher than a Bachelor’s degree.

AC

Moreover, MOOC registrants are more likely to be male and tend to be young. For example, based on data from four MOOCs operated by The University of London International Programmes on Coursera in June 2013, the male to female ratio is 64:36, and the average age of MOOC users is 34 (Grainger 2013).

https://learner.coursera.help/hc/en-us/articles/208280296 12 Shah, D. 2015. Less Experimentation, More Iteration: A Review of MOOC Stats and Trends in 2015. https://www.class-central.com/report/moocs-stats-and-trends-2015/

9

ACCEPTED MANUSCRIPT With the increasing number of MOOC students, educational institutions have started to accept MOOC-based credits.13 For instance, Arizona State University (ASU) and edX announced the Global Freshman Academy in 2015. Students who pass online courses through edX can obtain college credits from ASU.14 In addition, it is also possible to get a degree based entirely

T

on MOOC. In 2013, the Georgia Institute of Technology launched an online Computer Science

IP

Master’s degree by cooperating with Udacity.15 Other online programs include University of

CR

Illinois’ online MBA degree on Coursera and MIT’s (half-MOOC, half-on campus) supply chain

Demand for MOOC

AN

III.

US

management Master’s degree program on edX.16

Based on microeconomic principle, the demand function of a good includes its own price,

M

the price of related goods, income, market size, and preference towards it. We include all of

ED

these variables in our estimation, except for the price of MOOC, as MOOC is essentially free. In particular, we include wage to estimate the income effect, traditional education tuition to identify

PT

the complementarity or substitutability between MOOC and traditional education,

CE

unemployment to capture the effect of opportunity cost and preference towards MOOC, general education level to represent the size effect and preference, and internet infrastructure to capture

AC

the size effect (i.e., number of users), cost effect (i.e., waiting time), and preference. Therefore, our specification is generally consistent with the theory.17

13

Lequerica A. 2016. MOOCs for Credit https://www.class-central.com/report/moocs-for-credit/#take-a-mooc 14 ASU’s program: http://asuonline.asu.edu/ 15 Georgia Tech’s program: https://www.omscs.gatech.edu/ 16 University of Illinois’s program: https://onlinemba.illinois.edu/, MIT’s program: http://micromasters.mit.edu/ 17 The estimation of the demand function can be very complicated due to various endogeneity problems embodied in the demand system. Fortunately, because MOOC is still in a relatively early stage of development, the extent of endogeneity should be relatively small, especially when its own price is almost zero. For example, college tuition is

10

ACCEPTED MANUSCRIPT Among these factors, internet speed plays an important role. A slow speed not only increases the opportunity cost of time, but also decreases the effectiveness of online educational delivery. Thus, a fast and reliable internet connection is the premise of the development of MOOC. The Federal Communications Commission (FCC) provides a Broadband speed guide

T

about the minimum internet speed required for certain online activities.18 According to the report,

IP

streaming movies or university lectures at HD quality requires a minimum download speed of 4

CR

Mbit/s.

The discussion about the relationship between traditional face-to-face education and

US

MOOC boils down to an empirical question. That is, how does MOOC demand respond to

AN

college tuition (i.e., the price of the related good), given that its own direct cost is almost zero? On one hand, high tuition may motivate individuals to substitute traditional education for MOOC

M

(i.e., substitutability). On the other hand, high tuition can decrease MOOC demand if most

ED

MOOC users are college students (i.e., complementarity). The cross-price elasticity of MOOC demand can provide useful information on their relationship.

PT

Income is another factor that affects the demand for education, which is at a higher

CE

hierarchy based on Maslow’s theory (Maslow 1943). Studies have shown a positive income effect for traditional education (e.g., Glewwe and Jacoby 2004). An increase in income should

AC

increase the demand for MOOC, as it is a form of education that helps achieve a higher level of self-fulfillment. In addition, higher income makes technology more affordable, and thus high income individuals are more active in technology use. Studies have shown that low-income individuals face greater challenges when using online resources because they have lower perceived technical skills (DiSalvo et al. 2016). unlikely to respond to the demand for MOOC at this stage. Moreover, because we employ panel data and Fixed Effects estimation methods, we are able to mitigate the potential endogeneity problem caused by omitted variables. 18 The guide can be accessed from https://www.fcc.gov/reports-research/guides/broadband-speed-guide

11

ACCEPTED MANUSCRIPT Preference towards MOOC can also be affected by education level, because highly educated individuals are more motivated to learn and are more open to new technology. They usually have more challenging occupations and jobs that require more frequent updates of skills and knowledge. Studies have shown that most MOOC students are well educated and are seeking

T

to advance their careers (Emanuel 2013). Additionally, it is generally believed that one’s

IP

education is positively correlated with one’s ability and motivation. Card (1999) showed that

CR

higher earnings of better educated workers are partially because individuals with higher ability have a stronger motivation to learn and choose to acquire more schooling. Heckman et al. (2006)

US

also indicated that education is highly correlated with noncognitive abilities such as persistence

AN

and motivation. In addition, since most MOOC users tend to be highly educated, as discussed before, the overall education level of a country also represents a size effect (i.e. user pool) of the

M

demand.

ED

However, to study the demand for MOOC, it is critical to find a measure for it. Such a measure is difficult to get, partially because MOOC is relatively new. It is almost impossible to

PT

get data for MOOC demand for multiple countries and years using any traditional measures.

approach.

CE

Therefore, in this study, we construct a proxy to measure MOOC demand based on a Big Data

AC

In particular, we measure MOOC demand from the perspective of consumer information search and purchase behavior through the internet. With the rapid development of the internet, an individual’s demand can be reflected by internet searches. According to McGaughey and Mason (1998), internet influences buyer behavior through each step of the classical buyer decision process, including problem recognition, information search, evaluation of alternatives, purchase decision, and post-purchase behavior. In our case, individuals first realize the need for additional

12

ACCEPTED MANUSCRIPT education/courses (i.e., problem recognition) and then gather related information (i.e. information search). During this search process, they use search engines and search keywords that can lead them to their specific areas of interest. For example, individuals who want to improve their knowledge and decide to choose online education may search for more general

T

keywords such as “(free) online education” or “(free) online course” at the beginning and learn

IP

of the existence of MOOC. Then an in-depth search of MOOC will expose individuals to various

name of certain MOOC platforms to find out more details.

CR

platforms such as Coursera, Udacity, and edX (i.e., alternatives). Later on, they search for the

US

When it comes to course variety, Coursera and edX have a wide range of courses choices

AN

in all subjects, while Udacity concentrates more on job market skills. Based on personal interest, preference, and need, individuals will choose products that best satisfy their needs. The above

M

internet search behaviors are captured, stored, and analyzed by the search engine. Such

ED

information can help us estimate how many individuals in a certain area have searched for certain keywords.

PT

Therefore, with a proper choice of keywords, search volumes can properly reflect the

CE

potential MOOC demand. Search volume for a more general term such as “MOOC” may only reflect how many people are interested or just curious about what MOOCs are. However, the

AC

search volume for certain platforms such as “Coursera” can capture individuals who are seriously interested in MOOC and are more likely to enroll. With a combination of general and specific keywords, we can use the search index generated by search engines to measure the demand for MOOC.

IV.

Demand for MOOC in OECD countries

13

ACCEPTED MANUSCRIPT In recent years, the European Commission funded a number of MOOC projects, such as the HOME project (i.e., Higher Education Online: MOOCs the European Way), to develop and strengthen an open network for European cooperation.19 With political support, registrants from OECD countries account for a significant portion of worldwide MOOC students. Based on a

T

survey by Christensen et al. (2013), more than 65% of surveyed MOOC students originate from

IP

an OECD country, with 34.3% coming from the United States and 14.8% from a BRICS country

CR

(i.e., Brazil, Russia, India, China, and South Africa). In addition, according to Coursera, four of its top ten countries in terms of Coursera learners are OECD countries, including the United

US

States, Canada, the United Kingdom, and Spain. Other countries on the list include China,

AN

Mexico, etc.20

Our data include 34 OECD countries from 2012 to 2015. As discussed above, we use

M

Google Trends data to proxy MOOC demand in each country for each year.21 Google Trends

ED

provides a (nearly) real-time Google search volume index for certain keywords from 2004 to the present. Specifically, in this study, we use a combination of keywords “MOOC”, “Coursera”,

PT

“Udacity”, and “edX”. The last three keywords represent the three most popular MOOC

CE

platforms in the world.22 They have enough search volume for Google Trends to generate search indexes for each country. For other platforms, Google Trends cannot provide an index because of

AC

the low search volume in many countries. Figure 1 is a snapshot of Google Trends showing the relative popularity of these keywords. Consistent with the fact that MOOC took off in 2012, the popularity of MOOC and its 19

Source: http://home.eadtu.eu/ Source: https://blog.coursera.org/post/142363925112 21 Google Trends can be accessed from https://www.google.com/trends/ 22 There are other possible combinations of key words to measure the demand for MOOC such as “online courses” and “online education”. However, for those who are serious about taking MOOC, they will be more likely to use one of the three major platforms. Therefore, we believe that the search using names of those three platforms reflects serious interest and thus is a better proxy of MOOC demand. 20

14

ACCEPTED MANUSCRIPT platforms surged during 2012. The high points usually fall in January to February or September to October, which corresponds to the start of the spring or fall semester. The low points are usually in December, which is the holiday season. Not surprisingly, the combined search term has the highest search index among all keywords, and Coursera has the highest search index

T

among the three platforms. In addition, from 2012 to 2015, the search index of Coursera

IP

increased nearly three times, the fastest among the three platforms.

CR

However, Google Trends does not report the actual search volume; rather, it reports a

of searches (i.e.,

) divided by the total number

US

search index, which equals the keywords’ search volumes (i.e.,

) done on Google in a specific location. Moreover, in order to have a range

AN

of 0 to 100, all these search indexes are divided by the highest index value among them and then multiplied by 100 (see Google Trends Help Center).23 More specifically, the Google Trends ) for all n countries and T years included in the data is

ED

reported as:

M

index in country and year (i.e.,

KSVit KSVit ) / (max( |(i 1,2,...,n,t 1,2,...,T ) )) , TSVit TSVit

is the search volume for the above four keywords in country and year ,

is

CE

where

PT

GTit  100  (

total search volume in country and year .

AC

When collecting data, we first obtain monthly indexes of Google Trends and then calculate the yearly average in each country.24 One particular problem for the Google Trend index

is that an increase in the Google Trends could either result from more searches for the

keywords or from fewer other searches, and thus a particular value from Google Trends cannot be specifically interpreted. To avoid this problem, we convert the Google Trends index (i.e.,

)

23

Website: https://support.google.com/trends/?hl=en#topic=4365599 Google Trends will report an index of zero if search volume is very low. In the year 2012, some countries have search indexes of zero during the first a few months when MOOC begins to appear. 24

15

ACCEPTED MANUSCRIPT into an absolute search volume (i.e.,

). However, in the equation above,

for each

country is unknown. Because the data on the total search volume of the world is available, we estimate each country’s total search volume based on its share of internet users in the world.25 is estimated as:

is the world total search volume at time ,

in country and time , and to

is number of internet users in the world at time , and

is the

among all the and for the four keywords defined above.

US

highest ratio of

represents number of internet users

CR

where

IP

T

More specifically, the absolute search volume

Therefore, we use the estimated absolute keyword search volume

to measure the

AN

demand of MOOC in country and year for all OECD countries from 2012 to 2015.26 We

M

construct a panel dataset with the MOOC demand measure and other possible explanatory

ED

variables for each country. The panel data enable us to do a much more reliable regression analysis than cross-sectional or time-series data.

PT

Table 1 lists the descriptive statistics.27 MOOC demand measured by the

index

CE

increases significantly from 2012 to 2015, and overall, the demand in 2015 is approximately 2.6 times higher than in 2012. Figure 2 displays the search volume index of each OECD country in

AC

2015 to show country level variations, with darker colors indicating higher demand. The United States has the highest demand, with a search volume index of 776, which is nearly 12 times higher than the average index. In addition, the country with fastest increase in MOOC search is France, with the index increasing from 19.22 to 205.92 from 2012 to 2015.

25

The implicit assumption is that average search amounts of each internet user in OECD countries are close. can be viewed as a search index that represents absolute search volume. It contains an unknown constant 27 Percent of college and high school graduates for 2014 and 2015 are linearly predicted based on data from 2010 to 2013. Wage data for 2015 are linearly predicted based on historical data from 2000 to 2014. Internet speed data for 2015 are linearly predicted based on historical data from 2012 to 2014. 26

16

ACCEPTED MANUSCRIPT As discussed above, MOOC demand in a particular country is affected by its costs, income level, and population size, in addition to history, culture, and the preference for education. Among them, internet speed is an important factor. When studying MOOC courses, individuals can either download short lectures in each module or watch it online. For example, for a 10-

T

minute video, it takes approximately 6 minutes to download with an internet speed of 2 Mbit/s.

IP

For some courses, users can have access to all lectures; while some courses are posted on a

CR

weekly basis. All the course-related materials, such as the syllabus, announcements, assignments, and exams, are posted online. Thus, a fast internet speed is highly desirable for MOOC users.

US

Specifically, data on internet speed are obtained from a series of OECD reports (2014,

AN

2015). It is measured by download speed (Mbit/s) for the first quarter of the year.28 Based on Table 1, internet speed increases significantly during the period, from an average of 14.35 Mbit/s

M

in 2012 to 28.63 Mbit/s in 2015. For instance, in 2015, internet speed in OECD countries such as

ED

Japan and Mexico is three times higher than the speed in 2012. In 2015, countries with the highest internet speed include Korea, Japan, Switzerland, and Sweden, and countries with the

PT

lowest internet speed include Italy, Greece, Turkey, and Mexico. Internet speed in South Korea

CE

is the highest (i.e., 58.42 Mbit/s), which is nearly seven times higher than in Italy (i.e., 8.66 Mbit/s). The United States only has an intermediate internet speed.29

AC

In general, it is believed that college tuition has important implications for the demand for MOOC. College tuition in OECD countries differs significantly. For example, in some OECD countries, such as Denmark, Finland, and Iceland, public tertiary institutions charge no tuition fees for residents. On the other hand, U.S. has one of the most expensive higher education 28

The data are collected by Ookla, a company specializes in broadband testing and web-based network diagnostic applications. Website: http://www.ookla.com/ 29 Miller, C.C. 2014. Why the U.S. Has Fallen Behind in Internet Speed and Affordability. http://www.nytimes.com/2014/10/31/upshot/why-the-us-has-fallen-behind-in-internet-speed-andaffordability.html?_r=0

17

ACCEPTED MANUSCRIPT systems in the world. For the academic year 2016-2017, the average tuition and fees was $9650 for public four-year colleges, and $33,480 for private non-profit four-year colleges, which accounted for 17% and 58% of its GDP per capita, respectively (College board 2016). However, we cannot find complete tuition data for all OECD countries in all years in our sample. Therefore,

T

we rely on panel data fixed effects to partially capture the tuition effect. In the study for China in

IP

next section, we have tuition data to estimate the cross-price demand elasticity for MOOC.

CR

As discussed before, a country’s overall education level is another important variable that affects the demand for MOOC. In order to capture the impact of education, we use two measures

US

of education level for each country. The measures are the proportion of college graduates and

AN

high school graduates of the population aged 25-64, because those two groups are the main users of MOOC.30 Income data are obtained from OECD Employment and Labor Market Statistics and

M

are measured in constant prices at 2014 US dollars.31 The data for unemployment are obtained

ED

from Labor Market Statistics.32 To capture the size effect of MOOC demand, we use population size from 20 to 44 years old in a country, as a majority of MOOC users are in this age range.33

PT

In conducting regression analysis, we first use the regular Fixed Effects (FE) model with

CE

country fixed effects for the OECD panel data, because country fixed effects are likely to be correlated with other country level variables. We then move to an expanded two-way FE model

AC

with year dummies included, in order to control for year specific effects across all OECD countries. Additionally, given the rising trend of MOOC demand and some possible trended explanatory variables, it is desirable to add a time trend to de-trend variables and avoid spurious

30

Data are obtained from Education dataset: Population who attained tertiary education by sex and age group. Data source: https://data.oecd.org/eduatt/adult-education-level.htm. 31 Data source: https://data.oecd.org/earnwage/average-wages.htm 32 Data source: https://data.oecd.org/unemp/harmonised-unemployment-rate-hur.htm 33 Ideally, we should use the number of Internet users. Because data are not available, we use population data, which are obtained from: http://stats.oecd.org/Index.aspx?DatasetCode=POP_FIVE_HIST

18

ACCEPTED MANUSCRIPT regression. The trend model has another advantage of saving the degree of freedom compared to the model with year dummies, as well as reducing the multicollinearity among regressors when year dummies are included. 34 The results for all models are reported in Table 2.35 Model 1 presents the regular Fixed Effects (FE) model, and shows that internet speed has

T

a positive impact on MOOC demand, and it is statistically significant. Presumably, a higher

IP

internet speed enables a fast and reliable access to MOOC, and reduces the time cost. It also

CR

increases the study efficiency because of easier access to online class lectures and other resources such as discussion forums. However, after we control for year dummies and time

US

trends in Model 2 and 3, internet speed becomes insignificant. It is possible that internet speed is

AN

highly correlated with time, and thus it captures some annual effects. In addition, for all OECD countries, the minimum internet speed far exceeds the recommended speed for online video

M

streaming (i.e., 4 Mbit/s). For example, the minimum average internet speed in any OECD

ED

country in 2015 is 8.66 Mbit/s. Therefore, at this speed, it is plausible that internet speed is not a significant factor affecting MOOC demand.

PT

For all models, unemployment appears to have positive and significant effect.

CE

Specifically, based on Model 2 and 3, a one percentage point increase in unemployment rate will result in an increase of MOOC demand by 8.0-9.0%. The positive and significant impact from

AC

unemployment supports the point view that MOOC serves an important role during the process of re-employment. 36 Many MOOC courses provide or upgrade in-demand skills, and therefore help individuals on the job market. For instance, in 2015, courses about computer science and 34

We estimate models with a linear trend and a quadratic trend. We also run the model with the linear trend only. Considering quadratic model is more general, and the nonlinear trend is significant, we only present the quadratic trend model. Moreover, for the quadratic trend model, the adjusted R-square is significantly higher than the linear trend model (0.653 vs. 0.869), indicating that the quadratic trend model fits better. 35 We also run the basic OLS estimation and the Random Effect model. However, the Hausman test rejects the Random Effect model. 36 See the interview with Coursera CEO Richard Levin, who served as the president of Yale University for 20 years. Source: http://www.mckinsey.com/industries/social-sector/our-insights/coursera-takes-aim-at-unemployment

19

ACCEPTED MANUSCRIPT programming account for nearly 20% of all the MOOC courses offered, followed by business and management, which account for 16.8%. Those skills are useful in the job market and can help unemployed workers find jobs. For instance, Udacity has several hiring partners such as Google and AT&T that are regularly hiring their talented graduates. Graduates from the

T

Nanodegree Plus program provided by Udacity are guaranteed to get a job within six months of

CR

incentive to take MOOC, and thus increase its demand.

IP

their graduation date. 37 As a result, it is not surprising to see that unemployment can provide

Similarly, for all model specifications, education shows positive effect on the demand. A

US

one percentage point increase in the proportion of college graduates is expected to increase

AN

MOOC demand by approximately 8.0-9.0% (Model 2 and 3). The positive and significant effect is consistent with the literature, as most MOOC takers have a Bachelor’s degree or higher.

M

Interestingly, the effect of high school graduates on MOOC is almost identical to that of college

ED

graduates, positive and statistically significant with similar magnitudes. For high school graduates, MOOC offers an opportunity to receive a college-level education.

PT

The average wage in a country, which reflects the income effect, does not show a

CE

statistically significant effect on MOOC. One possible reason is that most OECD countries are developed countries with high income, and thus an increase in wage does not have a significant

AC

impact on the demand for MOOC.38 For example, in 2015, the average wage in OECD countries is approximately five times higher than in China. Population size does not show a significant effect on MOOC demand either, perhaps because the size effect has been mostly captured by the relative size of people with different levels of education in a country. Additionally, Model 3 37

Flipkart, an e-commerce company headquartered in India, starts to hire students based on Udacity’s Nanodegree programs without interviews. Source: http://timesofindia.indiatimes.com/tech/tech-news/Flipkart-hires-withoutinterviews/articleshow/50756104.cms 38 Other possible explanations are that a country’s internet speed/infrastructure and education level all capture some income effect.

20

ACCEPTED MANUSCRIPT shows a clear pattern of quadratic trend for MOOC, a general upward trend, but the increase speed slows down with time. The above study based on OECD data provides interesting results on the demand for MOOC. However, the results are mostly based on developed countries. For developing countries,

T

there exist significant differences in economic development, educational institutions, and internet

IP

infrastructure. Therefore, MOOC may have very different implications for higher education. For

CR

example, will MOOC become a low-cost and effective method to improve the availability and quality of higher education in developing countries? Will those who were denied the opportunity

US

of higher education choose MOOC as an alternative?

AN

Additionally, due to the vast differences across countries, a unified demand model across countries is more restrictive than at a country level. In particular, the coefficients in the demand

M

model may differ across countries, but this is less of a concern for a country level model.

ED

Therefore, in the next section, we will study the factors affecting the demand for MOOC in

Demand for MOOC in China

CE

V.

PT

China.

China is a developing country with the largest higher education market and the largest

AC

number of internet users in the world. Given its size, the demand for MOOC in China has important implications not only for China itself but also for the future of MOOC in general. China is currently among the top countries in terms of Coursera registered learners, although English is not their official language. In July 2015, Coursera announced that it has more than 1 million registrations from China, making the country their second largest user base after the

21

ACCEPTED MANUSCRIPT United States.39 MOOC has also gained political support from the Chinese government, and the Ministry of Education encourages the government, colleges, and society to work together to promote the development of online education platforms. It recommends that, by the year of 2020, China should have at least 3000 nationally recognized high quality online courses. 40

T

In this section, we study the demand for MOOC in China, using provincial panel data for

IP

31 provinces from 2012 to 2015 to compare the demand with OECD countries and gain an in-

CR

depth understanding of the demand for online education. The proxy for MOOC demand originates from the largest internet service provider in China, Baidu.41 In 2015, 86.7% of the

US

Chinese internet users searched via Baidu.42 We use the Baidu Index to represent the demand for

AN

MOOC. It is a similar service to Google Trends, and provides Baidu query volume data from June 2006 to the present on a daily basis in each province.43 Moreover, in the OECD analysis,

M

Google Trends could only provide the relative search popularity, and we needed to estimate the

ED

absolute search volume. For China, however, Baidu provides absolute search volume, which generally reflects people’s search behavior more directly and precisely.

PT

For comparison purposes, we use the same keywords as with the OECD study, including

CE

MOOC, Coursera, Udacity, and edX. In this case, the demand proxy for MOOC in China is mainly for English courses, which represents only part of the demand from those who can take

AC

courses in English.44 Figure 3 shows the snapshots of Baidu Index of four keywords. In China,

39

Shah, D. 2015. How Coursera Cracked the Chinese Market. https://www.class-central.com/report/coursera-cracked-chinese-market/ 40 Original document: http://www.moe.edu.cn/publicfiles/business/htmlfiles/moe/s7056/201504/186490.html 41 Website: https://www.baidu.com/ 42 Google exited Chinese market in 2010, and as a result, Google search share decreased dramatically from more than 30% in 2010 to less than 2% in 2015. http://gs.statcounter.com/search-engine-market-share/all/china/#yearly-2010-2016 43 Baidu index can be accessed from http://index.baidu.com/ 44 We attempt to study the MOOC demand for local Chinese platforms, which include some well-known Chinese local MOOC platforms, XuetangX, IMOOC, and Chinese University MOOC. However, these local platforms only started recently, and do not have comparable data with the international platforms. We still obtain the search volume

22

ACCEPTED MANUSCRIPT MOOC and its three most popular platforms start to gain search volume beginning in 2012, grow steadily in 2013, and become relatively stable in the years 2014 and 2015. Search volume has several peaks, especially in June, August, and September for keyword “MOOC” in year 2014, which is likely due to some important MOOC-related news being released during that time. The

T

low points often fall on the month of the Spring Festival, the most important holiday in China.

IP

Coursera has the highest average search volume among three platforms. From 2012 to 2015,

CR

search indices grew approximately 13 times for Udacity and edX, and 11 times for Coursera. Table 3 presents the descriptive statistics.45 The Baidu index increases significantly, and

US

searches in 2015 are 26 times higher than 2012. Figure 4 displays the index of each province in

AN

2015, with darker colors indicating higher demand. In general, more developed coastal provinces such as Jiangsu and Zhejiang have higher MOOC demand, while less developed inland provinces

M

such as Qinghai and Xinjiang have lower MOOC demand. Beijing has the highest MOOC

ED

demand, with an index of 1302, which is nearly three times higher than the average index. In China, internet speed has increased significantly over the past a few years, from 0.98

PT

Mbit/s in 2012 to 6.01 Mbit/s in 2015, representing a six-fold increase.46 In 2015,

CE

provinces/cities with the highest internet speed include Beijing, Shanghai, and Tianjin, and provinces with the lowest internet speed include Xizang, Guangdong, and Gansu. Internet speed

AC

in Beijing is the highest (i.e., 7.10 Mbit/s), which is significantly higher than that in Xizang (Tibet, i.e., 5.01 Mbit/s). Even with such increases, compared to OECD countries, internet speed

for local MOOC platforms and conduct similar demand analysis. The results are presented in Appendix Table A1. Most results are similar to that for English platforms. 45 Due to data limitation, tuition per capita, percent of college and high school education, unemployment rate, GDP per capita, and internet users in 2015 are linear predictions based on available historical data. Internet speed data for 2012 are predicted based on 2013-2015 data. 46 Internet speed is the average connection speed obtained from the China Internet Speed Report from the Broadband Development Alliance. Data source: http://www.chinabda.cn/xzzq/index.shtml

23

ACCEPTED MANUSCRIPT in China is significantly lower. The average speed in 2012 is much slower than the recommended minimum speed of 4 Mbit/s for video screaming. Unlike in the OECD analysis above, tuition data are available.47 In earlier years, college tuition in China was relatively high compared to family income. For example, in 2002, the

T

average tuition per student accounted for 46% of GDP per capita (Li and Liu 2014). It then

IP

declined gradually, and until 2012, tuition per capita in China was, on average, around 7,274

CR

yuan, which accounted for 17% of GDP per capita, and dropped to 14% of GDP per capita in 2015.48 This ratio is relatively high compared to some OECD countries, especially for those with

US

free college education.

AN

The number of individuals with education at the high school level or higher has increased rapidly due to the fast expansion of higher education in China.49 From 2012 to 2015, the percent

M

of college graduates increased from 14.14% to 16.36%; and high school graduates increased

ED

from 19.70% to 21.04%.50 Moreover, the return to schooling in China is still rising, from 1-3% in the 1980s to 8-12% after 2000 (Li and Liu 2014). Such rising returns to schooling will bring

PT

strong incentive for individuals to invest in schooling and pursue more education.

CE

To capture the additional size effect, we use the number of internet users since data are

47

AC

available in China.51 Table 3 indicates that the number increases by approximately 4 million

In our study, tuition per capita data are calculated based on the ratio of government income from tuition and fees and the number of students in college and universities in China. Government income from tuition and fees is obtained from China Education Finance Statistical Yearbook Table 3-7 (2011, 2012, 2014), and number of students is obtained from China Statistical Yearbook (2012, 2013, 2014, 2015). Due to data limit, government income from tuition and fees in 2013 is the average of data from 2012 and 2014. 48 The numbers are authors’ calculation based on various issues of China Educational Finance Statistical Yearbook and China Statistical Yearbook. 49 Number of college graduates, high school graduates, and labor force data are obtained from China Statistical Yearbook (2012, 2013, 2014, 2015). 50 Note that the maximum percentage of college graduates is higher than that of high school graduates in both 2012 and 2015. The reason is that in Beijing, the percent of college graduates is much higher than high school graduates in both years. 51 Data are obtained from the Statistical Report of internet Development in China by the China Internet Network Information Center (CNNIC).

24

ACCEPTED MANUSCRIPT during the period, from 18.19 million in 2012 to 22.40 million in 2015. In 2015, Guangdong province had the highest number of internet users (i.e., 76 million), while Xizang had the lowest number of users (i.e., 1.35 million). Table 4 displays the result for the demand analysis in China.52 The one-way FE model

T

shows the positive effect of internet speed, wage and unemployment; while in the two-way FE

IP

model with year dummies, all explanatory variables (except for year dummies) are statistically

CR

insignificant. It is possible that, compared to cross-country models, annual effects are more likely to be highly correlated with other variables within a country, due to common shocks such as

US

policies from the central government. The dramatic changes in the estimated coefficients

AN

between the one-way and two-way FE models also indicate a high degree of multi-collinearity. In fact, a F-test on the joint significance of the explanatory variables excluding the year dummies

M

in the two-way FE model shows a significant effect of those variables as a group. Therefore, the

ED

trend model is generally preferable given the potential trending of the dependent variable and some explanatory variables and its parsimonious nature. Our discussion will mostly be based on

PT

Model 3.53

CE

Differing from the OECD model, as Model 3 indicates, a higher internet speed will significantly promote MOOC demand. More specifically, if internet speed increases by 1%,

AC

MOOC demand will increase by 1.181%. A main reason is that China’s internet speed is among the slowest in the world.54 In 2012, even the province with the fastest internet speed could not meet the recommended speed for video streaming. Even in 2015, the minimum average speed for

52

We also run the basic OLS estimation and the Random Effect model. The Hausman test rejects the Random Effect model. 53 The linear trend term is not statistically significant. Thus there is no need to use the quadratic trend specification; however, the linear trend term still helps model specification due to trending variables. 54 Akamai. 2015. State of the Internet Report. https://content.akamai.com/PG5641-Q4-2015-SOTI-Connectivity-Report.html

25

ACCEPTED MANUSCRIPT any province was 5.01 Mbit/s, only slightly above the recommended speed. An additional complication related to internet speed is that China enforces strict internet content censorship through the so called “Great Firewall”, which blocks and filters content, especially from overseas websites. Most internet communication between China and the rest of the world is routed

T

through a very small number of fiber-optic cables, via one of the three international gateways:

IP

the Beijing-Qingdao-Tianjin area in the north, Shanghai on the east coast, and Guangzhou in the

CR

south.55 The firewall and international gateways slow down internet from the normal speed. Because the MOOC demand we studied is for international platforms in US, a higher speed is

US

much more desirable for MOOC users to get access to the international platforms.56

AN

In addition, the income effect of the demand is positive and statistically significant. If the average wage increases by one percent, the demand for MOOC will increase by 6.603 percent,

M

which is highly elastic. This effect is different from that of the OECD countries. As hypothesized

ED

before, income level in China is much lower, and thus has a stronger effect. Many low-income Chinese, especially in rural areas, lack basic equipment and skills for using online resources, and

PT

a higher income can bring a significant impact to them. Based on this result, it is expected that

CE

the demand for MOOC will continue to rise as the Chinese economy grows. Similar to OECD countries, unemployment also has a positive effect on MOOC demand

AC

in China, and the effect is much larger. A one standard deviation increase in unemployment rate will increase MOOC demand by approximately 70%. One reason for this much larger effect is that, in China, unemployment rate is very low and highly stable (i.e., in the range of 3%) with a small standard deviation (i.e., 0.6-0.7 percentage points), and thus an increase in unemployment

55

Source: http://www.theatlantic.com/magazine/archive/2008/03/the-connection-has-been-reset/306650/ It would be interesting to see how the internet speed affects the MOOC demand for local platforms. Due to the current incomplete data and small sample size, we leave it for future research. 56

26

ACCEPTED MANUSCRIPT rate represents a much larger change than in OECD countries.57 Additionally, given the size of the population and labor force, a slight increase in unemployment will result in a large number of people without a job, which represents another size effect for demand. We do not observe a statistically significant effect on MOOC from the relative size of

T

college or high school graduates. Moreover, the estimated coefficient for college tuition is not

IP

statistically significant either. Therefore, there is no clear-cut result about the complementarity or

CR

substitutability between MOOC and traditional education in China based on our data. One possible reason is that MOOC is still developing in China, so it may not be familiar to a lot of

US

people, especially in rural and/or poor regions. In addition, there is no university in China

AN

offering MOOC-based academic credits, so it is more difficult for Chinese students to substitute a college degree with a MOOC based degree. Another possible reason is that, as discussed before,

M

college education in China is still scarce. Thus, Chinese people still have strong preferences

Conclusions

PT

VI.

ED

towards traditional higher education, possibly due to its prestigious signaling value.

CE

Facing various challenges in traditional education, many educators consider MOOC as a possible solution. However, whether MOOC can bring a revolution to traditional education

AC

depends on whether such a demand exists and how it evolves. In this study, we investigate factors affecting the demand for MOOC in OECD countries and in China. Due to the difficulty in measuring the demand for MOOC, we find its proxy by applying Big Data techniques to collect data from internet search engines. In particular, we use search-volume data of MOOCrelated keywords as a proxy for MOOC demand, and then estimate a demand function.

57

China’s official unemployment rates have been very low with very small variations over years. http://fortune.com/2015/08/20/china-unemployment/

27

ACCEPTED MANUSCRIPT Our results, based on the estimation using both OECD and China data, are generally consistent with our predictions. However, their demand structures for MOOC exhibit several differences. We find that in both OECD countries and China, an increased unemployment rate will promote the demand for MOOC. By providing valuable platforms for skill-improvement and

T

re-employment, MOOC appears to become a possible solution to the unemployment problem.

IP

We find a positive and significant income effect in China, but not in OECD countries.

CR

Additionally, internet speed has a positive effect on the demand for MOOC in China, but does not show an effect in OECD countries. Moreover, we find that the general education level,

US

measured by the share of high school and college graduates, only has a significant effect on

AN

MOOC demand in OECD countries.

We do not have a clear-cut conclusion about the complementarity or substitutability

M

between MOOC and traditional education. One reason is that the data limit our results. Because

ED

the tuition data, which represent the price of a related good (traditional education), are not available for some OECD countries, we are unable to estimate the effect of tuition on the

PT

demand for MOOC. In China, tuition data are available, but tuition does not seem to have a

CE

significant impact on MOOC demand. One possible explanation is that, because MOOC is relatively new in China (started in 2012), its impact on traditional education may need several

AC

more years to emerge. However, MOOC can reduce the cost of education, provide great flexibility for individuals to learn, and has the potential to become more interactive with the advance of information technology. Therefore, a combination of MOOC and traditional education may be able to take advantage of both educational delivery mechanisms, and thus generate better education outcomes.

28

ACCEPTED MANUSCRIPT Reference Banerjee, A. V., and E. Duflo. 2014. (Dis)organization and Success in an Economics MOOC. American Economic Review 104(5): 514-518. Barber, M., K. Donnelly, and S. Rizvi. 2013. An Avalanche is Coming: Higher Education and

T

the Revolution Ahead. Institute for Public Policy Research. London.

IP

Bass, S.A. 2014. Simple Solutions to Complex Problems – MOOCs as a Panacea? The Journal

CR

of General Education 63(4): 256-268.

Bowen, W. G., M. M. Chingos, K. A. Lack, and T. I. Nygren. 2014. Interactive Learning Online

AN

Analysis and Management 33(1): 94-111.

US

at Public Universities: Evidence from a Six-Campus Randomized Trial. Journal of Policy

Breslow, L., D.E. Pritchard, J. DeBoer, G.S. Stump, A.D. Ho, and D. Seaton. 2013. Studying

ED

Practice in Assessment 8: 13–25.

M

Learning in the Worldwide Classroom: Research into edX’s First MOOC. Research &

Card, D., 1999. The causal effect of education on earnings. In: Ashenfelter, O., Card, D. (Eds.),

PT

Handbook of Labour Economics, vol. 3. North-Holland, Amsterdam, 1801–1863.

CE

Christensen, C., and M. Horn. 2011. Colleges in Crisis: Disruptive Change Comes to American Higher Education. Harvard Magazine, July-August, 7(21): 40-43.

AC

Christensen, G., A. Steinmetz, B. Alcorn, A. Bennet, D. Woods, and E.J. Emmanuel. 2013. The MOOC Phenomenom: Who Takes Massive Open Online Courses and Why? University of Pennsylvania. Choi, H., and H. Varian. 2012. Predicting Present with Google Trends. Economic Record 88(S1): 2-9.

29

ACCEPTED MANUSCRIPT Clotfelter, C.T., H.F. Ladd, and J.L. Vigdor. 2010. Teacher Credentials and Student Achievement in High School. A Cross-Subject Analysis with Student Fixed Effects. Journal of Human Resources 45: 655-681. College board. Trends in Higher Education. Trends in College Pricing. 2016.

T

Deming, D. J., C. Goldin, L. F. Katz, and N. Yuchtman. 2015. Can Online Learning Bend the

IP

Higher Education Cost Curve? American Economic Review 105(5): 496-501.

CR

DiSalvo, B., P.K. Roshan, and B. Morrison. 2016. Information Seeking Practices of Parents: Exploring Skills, Face Threats and Social Networks. CHI Conference on Human Factors in

US

Computing Systems, California, USA.

AN

Emanuel, E.J. 2013. Online Education: MOOCs taken by Educated Few. Nature 503: 342. Ginsberg, J., M.H. Mohebbi, R.S. Patel, L. Brammer, M.S. Smolinski, and L. Brilliant.

M

2009. Detecting Influenza Epidemics Using Search Engine Query Data. Nature

ED

457(7232): 1012-1014.

Glewwe, P. and H.G. Jacoby. 2004. Economic Growth and the Demand for Education: Is there A

PT

Wealth Effect? Journal of Development Economics 74(1): 33-51.

CE

Grainger, B. 2013. Massive Open Online Course (MOOC) Report 2013. University of London International Programmes.

AC

Griffiths, R., M. Chingos, C. Mulhern, and R. Spies. 2014. Interactive Online Learning on Campus. ITHAKA. Heckman, J.J., J. Stixrud, and S. Urzua. 2006. The Effects of Cognitive and Noncognitive Abilities on Labor Market Outcome and Social Behavior. Journal of Labor Economics 24(3): 411-482.

30

ACCEPTED MANUSCRIPT Krause, S., and C. Lowe. 2014. Invasion of the MOOCs: The Promise and Perils of Massive Open Online Courses. San Francisco: Parlor Press. Li, H. and Q. Liu. 2014. Human Capital: Schooling. Oxford Companion to the Economics of China. Oxford University Press.

T

Li, H, Q, Liu, and P. Ederer. 2017. Regular Schooling versus Lifelong Learning: Does Regular

IP

Employment Affect the Effects of Higher Education? Working paper.

CR

Maslow, A. 1943. A Theory of Human Motivation, Psychological Review 50: 370-96.

Marketing Theory and Practice 6(3): 1-11.

US

McGaughey, R.E., and K.H. Mason. 1998. The Internet as a Marketing Tool. Journal of

Economic Perspectives 29(4): 135-154.

AN

McPherson M.S. and L.S. Bacow. Online Higher Education: Beyond the Hype Cycle. Journal of

M

OECD. 2014. Access Network Speed Tests. OECD Digital Economy Papers, No. 237, OECD

ED

Publishing.

OECD. 2015. OECD Digital Economy Outlook 2015, OECD Publishing, Paris.

PT

Schuwer, R., I. Gil-Jaurena, C.H. Aydin, E. Costello, C. Dalsgaard, M. Brown, D. Jansen, and A.

CE

Teixeira. 2015. Opportunities and Threats of the MOOC Movement for Higher Education: The European Perspective. International Review of Research in Open and Distributed

AC

Learning 16(6): 20-38.

Wang, X, B.M. Fleisher, H. Li, and S. Li. 2014. Access to College and Heterogenous Returns to Education in China. Economics of Education Review 42: 78-92. Yang, Y., B. Pan, and H. Song. 2014. Predicting hotel demand using destination marketing organization's WEB traffic data. Journal of Travel Research, 53(4): 433-447.

31

ACCEPTED MANUSCRIPT

Mean 65.87 7.97 39.74 28.63

Max 776.46 25.47 61.53 58.42

52.60

34.91

T

Mean 26.33 8.71 38.24 14.35

Year 2012 S.D. Min 70.78 1.06 5.04 3.18 12.80 12.71 6.72 4.77

Max 406.54 24.81 58.33 32.05

32.57

10.03

15.30

10.46

17.07

55.10

43.88

14.11

18.57

73.17

43.67

13.21

18.77

71.99

12.70

20.10

0.11

105.51

20.30

0.11

106.91

IP

Variable KSV Unemployment Wage Internet speed Percent of college graduates Percent of high school graduates Population

Year 2015 S.D. Min 137.13 3.57 4.76 3.40 13.11 12.54 12.70 8.66

CR

Table 1. Descriptive Statistics for OECD countries

12.67

AN

US

Note: 1. The KSV index represents absolute search volumes of MOOC. 2. Wage is measured by one thousand dollars at a constant price of 2014 US dollars. Internet speed is measured by download speed in Mbit/s. Education level is represented by percentage of college and high school graduates between the age group of 25-64 in OECD countries. Population is measured by millions of population between 20 to 44 years old.

AC

CE

PT

ED

M

Table 2. Factors affect MOOC demand in OECD countries FE Dummy Internet speed 0.766 (-4.02) *** 0.100 (-0.64) Wage 0.549 (-0.26) 0.501 (-0.38) Unemployment 0.144 (-2.91) *** 0.076 (-2.92) *** Percent of college graduates 0.207 (-3.41) *** 0.079 (-2.18) ** Percent of high school graduates 0.106 (-1.72) * 0.082 (-2.52) ** Population -1.103 (-0.36) 0.749 (-0.48) Year Dummy 2013 0.926 (-13.04) *** Year Dummy 2014 0.982 (-8.36) *** Year Dummy 2015 0.917 (-5.47) *** Trend Trend2 Constants -12.424 (-1.23) -8.116 (-1.28) N 119 119 Adjusted R-square 0.429 0.847 F-Statistics 20.593 76.774

Trend 0.001 (-0.01) 0.878 (-0.60) 0.092 (-3.19) *** 0.086 (-2.11) ** 0.083 (-2.29) ** 0.488 (-0.28)

1.541 (-12.92) *** -0.249 (-11.58) *** -10.414 (-1.48) 119 0.81 67.304

Note: 1. The dependent variable is the logarithm of the KSV. Independent variables such as internet speed, wage, and population are in log forms. 2. Standard errors are in the parenthesis. Superscripts ***, **, and * represent significance at the 1%, 5%, and 10% level, respectively.

32

ACCEPTED MANUSCRIPT

Mean 0.18

S.D. 0.11

7.27 3.32 41.05 0.98

2.72 0.64 10.63 0.55

2.60 1.27 31.30 0.09

16.33 4.23 75.59 2.83

7.81 3.12 55.22 6.01

14.14

7.41

5.32

43.37

16.36

19.70 18.19

4.33 13.41

6.40 1.01

26.70 66.27

S.D. 2.71 3.45 0.72 13.86 0.50

21.04 22.40

Year 2015 Min Max 0.31 13.02 2.76 1.21 42.18 5.01

19.13 4.44 102.27 7.10

1.65

48.83

4.31 1.35

28.32 76.31

T

Mean 4.76

IP

Variable Baidu College tuition per capita Unemployment Wage Internet speed Percent of college graduates Percent of high school graduates Internet users

Year 2012 Min Max 0.00 0.47

CR

Table 3. Descriptive Statistics for China

8.62

5.07 15.79

AN

US

Note: 1. Baidu represents the Baidu search index. 2. College tuition per capita data are measured in thousand yuan. Wage is measured by thousand yuan with 2012 as the base year. Internet speed is measured by average connection speed in Mbit/s. Unemployment and education levels are measured in percentages (%). Internet users are measured in millions.

AC

CE

PT

ED

M

Table 4. Factors affect MOOC demand in China FE Internet speed 1.181 (-6.43) *** College tuition per capita 1.191 (-0.91) Wage 6.603 (-3.56) *** Unemployment 1.088 (-2.82) *** Percent of college graduates -0.065 (-1.41) Percent of high school graduates -0.071 (-1.54) Internet users -0.348 (-0.14) Year dummy 2013 Year dummy 2014 Year dummy 2015 Trend Constants -21.453 (-1.61) N 123 Adjusted R-square 0.831 F-Statistics 90.938

Dummy 0.054 (-0.81) -0.096 (-0.24) 0.163 (-0.2) -0.156 (-1.27) -0.009 (-0.67) -0.018 (-1.31) 1.353 (-1.65) 2.407 (-20.15) *** 3.073 (-17.05) *** 2.986 (-11.94) *** -3.94 (-0.81) 123 0.985 822.48

Trend 1.178 (-6.36) *** 1.298 (-0.95) 7.123 (-2.69) *** 1.078 (-2.76) *** -0.065 (-1.38) -0.069 (-1.49) 0.001 (0.00)

-0.068 (-0.28) -23.981 (-1.48) 123 0.829 78.716

Note: 1. The dependent variable is the logarithm of the Baidu Index. Independent variables such as internet speed, college tuition per capita, wage, and internet users are in log forms. 2. Standard errors are in the parenthesis. Superscripts ***, **, and * represent significance at the 1%, 5%, and 10% level, respectively.

34

AC

CE

PT

ED

M

AN

Figure 1. A screenshot of search in Google Trends

US

CR

IP

T

ACCEPTED MANUSCRIPT

34

ACCEPTED MANUSCRIPT

T P

I R

C S

U N

A

D E

M

T P

E C

C A

Figure 2. KSV Index for OECD countries in 2015

35

IP

T

ACCEPTED MANUSCRIPT

AC

CE

PT

ED

M

AN

US

CR

Figure 3. A screenshot of search in the Baidu Index

Figure 4. Baidu Index for provinces in China in 2015

36

ACCEPTED MANUSCRIPT Appendix Trend -1.951 (-1.01) 1.388 (-2.26) ** 1.466 (-0.43) -3.297 (-2.86) *** -0.057 (-0.98) -0.017 (-0.27) -6.76 (-1.39)

AN

US

CR

IP

T

Table A1. Factors affect MOOC demand in local platforms in China FE Dummy College tuition per capita 0.274 (-0.10) -0.439 (-0.91) Unemployment 1.491 (-1.69) * -0.066 (-0.41) Wage 17.988 (-4.75) *** 1.851 (-2.16) ** Internet speed 2.387 (-1.91) * 0.171 (-0.55) Percent of college graduates -0.035 (-0.41) 0.003 (-0.21) Percent of high school graduates -0.078 (-0.89) -0.024 (-1.54) Internet users -5.225 (-0.75) -1.449 (-1.18) Year dummy 2013 2.722 (-22.7) *** Year dummy 2014 3.331 (-12.43) *** Trend Constants -21.374 (-0.67) 3.573 (-0.61) N 91 91 R-squared 0.755 0.993 F 44.828 1342.026

3.526 (-7.56) *** 24.849 (-1.08) 91 0.881 87.873

AC

CE

PT

ED

M

Note: 1. The dependent variable is the logarithm of the Baidu Index for Chinese local platforms. Independent variables such as internet speed, college tuition per capita, wage, and internet users are in log forms, and the rest of the variables are at levels. 2. Number of observations for local platforms are smaller than international platforms because local platforms appear one year after international platforms. 3. In our three-year local platform model, controlling for year dummies is equivalent to controlling for trend and trend2, so nonlinear trend specification is not presented. 4. Standard errors are in the parenthesis. Superscripts ***, **, and * represent significance at the 1%, 5%, and 10% level, respectively.

37

ACCEPTED MANUSCRIPT Highlights We proxy MOOC demand for OECD countries and China using Big Data techniques.



Higher unemployment promotes MOOC demand.



Education level positively affects MOOC demand in OECD countries.



Internet speed and average income positively affect MOOC demand in China.

AC

CE

PT

ED

M

AN

US

CR

IP

T



38