The present research uses the public released data sets collected by Taiwan Education Panel Survey (TEPS) in 2001 and 2003 (Chang 2003). In 2001, with the support and authorization of the Ministry of Education, National Council of Science, and Academia Sinica in Taiwan, TEPS using multistage stratified sampling method surveyed 20,004 7th graders in 333 junior high schools. These sampled students were surveyed again in their 9th grade. The follow-up sample size is 18, 903. The sample size of the public released data is 70% of the surveyed students.5 TEPS data were collected by administering the ability test and student’s questionnaire in the classroom under a standardized condition. Each surveyed students was also asked to take home a copy of parent’s questionnaire for one of his or her parents or guardians to answer, and the answered questionnaire was taken back for field staff to collect. Surveys were also administered to surveyed students’ homeroom teachers, Chinese language
teachers, English teachers, and Math teachers.
For the present research, I used 2001 and 2003 student data, 2001 parent data, and 2001 data of math teacher’s evaluation of surveyed students. The sample size of each data is slightly different. The sample size of the 2001 public released student data is 13,978. After merging this data with the follow-up student data and deleting cases that have no information about math ability test scores, cram schooling for math in 2003, gender, and ability grouping, the sample size reduces to 12,025. For all other
5 Please refer to http://www.teps.sinica.edu.tw/introduction.htm for basic information about TEPS (in Chinese).
17
variables used in the analysis, I use either mode or median to replace their missing values to reduce the loss of the sample size.6 The sample size for the present research is further reduced to 11,373, which is the analytical sample with common support after the matching.
The present research focuses on estimating the effect of math cramming
undertaken in the first semester of the 9th grade. The choice of estimating the effect of math cramming in the 9th grade is due to the fact that 9th graders would face their first senior high school entrance examination held in the second semester. Hence, the main purpose of cram schooling for math should be for this examination. The math ability of these students was also measured by TEPS in the later half of the first semester in 2003. Another reason to study the effect of math cramming is because math learning is largely taken place at school or cram schools rather than at home. Moreover,
coaching for math standardized tests is also shown to be more effective (Becker 1990).
The original math ability scores given by TEPS are IRT scores (Yang, Tam, and Huang 2004). For the ease of presentation and understanding, I transform the IRT scores into normal curve equivalent (NCE) scores for the sample before matching (N
= 12,025), which range from 1 to 99 with a mean of 50 and a standard deviation of 21.06. Table 1 shows that the mean of the math ability NCE scores for the matched sample (N=11,373) is 64.564 and the standard deviation is 29.858. The variable of math cramming in the 9th grade is coded as a dummy variable with 1 indicating the student undertaking math cramming during the first semester of the 9th grade and 0 for no math cramming. Table 1 shows that about 47% of student included for the present research undertook cram schooling in the 9th grade and a similar percentage of students have the experience of attending cram schools or seeking private tutoring
6 I also use listwise deletion to deal with missing values and the final result of analysis is very similar to the findings presented in this paper.
18
ever since they were 5th graders.
Other than variables of the math ability scores in the 9th grade, the experience of math cramming in the first semester of the 9th grade, and student’s own initiative in undertaking math cramming, and if students attends a high ability class in the 9th grade, all other variables used for matching are obtained from the TEPS 2001 data. In total, I used 27 variables for propensity score matching. Most of these other variables are variables often considered in previous empirical studies of cram schooling in Taiwan or studies of coaching in other countries (Stevenson and Baker 1992; Sun and Hwang 1996; Powers and Rock 1999; Baker, Akiba, and Wiseman 2001; Briggs 2001;
Lin and Chen 2006; Liu 2006). These matching variables can be grouped into three types: (1) student’s individual characteristics which include gender, learning habits, and math achievement in the 7th grade, (2) family backgrounds, and (3) school and class characteristics, which include school types and class environment regarding academic competition and ability grouping. Variables such as student’s learning habits, prior math achievement, and school or class characteristics are rarely available for previous studies of cram schooling in Taiwan. The following are more detailed description of the measurement and coding the variables used in propensity score matching. Summary statistics of variables are presented in Table 1 for both sample cases included in and excluded from the present study.
[Table 1 is about here]
1. Student’s individual characteristics (1) Gender: 1 is male and 0 is female.
(2) The homework subject that spends most time in doing: 1 is math and 0 is other subjects.
19
(3) Never let anything distract doing homework since elementary school: This is an ordinal variable treated as a continuous variable in the analysis. The variable ranges from 1 to 4 with 1 indicating and 4 indicating strongly disagree.
(4) Always review course lessons after school since elementary school: Same as (3).
(5) Always try to solve difficult problems in learning since very young: Same as (3).
(6) Math is always a headache: Same as (3).
(7) Can keep up with math teaching: This ordinal variable and the next two are math teacher’s evaluation of surveyed students. The variable is treated as a continuous variable in the analysis and ranges from 1 to 4. 1 means that the student’s learning pace is way ahead of teaching, 2 means that the student can keep up with teaching, 3 means that the student cannot keep up with teaching, and 4 means that the student is way behind.
(8) Math homework performance: This ordinal variable is math teachers to evaluate if students always, sometimes, rarely, or never late in turning in math homework. 1 means always and 4 means never.
(9) Attend special class for gifted students in the 7th grade: Students are asked if they attend special classes for students who are evaluated to be gifted in certain academic subjects such as math, science, and Chinese or English languages. The variable is dummy coded with 1 meaning yes or 0 meaning no.
(10) Make own decision about undertaking cram schooling: A dummy coded variable with 1 indicating yes and 0 indicating no.
(11) Cram schooling experiences: This variable is coded into four categories
20
including before 4th grade or never, from 5th grade to 7th grade, 7th grade only, and from 5th to 6th grade. The category of “from 5th grade to 7th grade” is the reference group in the regression analysis.
(12) Math ability IRT scores in the 7th grade: This variable is a continuous variable.
2. Family backgrounds
(1) Ethnicity: Four ethnic groups are constructed according to parents’ answer about their ethnicity. They are Minnan, Hakka, Mainlander, and Aborigine.
Minnan is the reference group in the regression analysis.
(2) Parental education level: Three parental education levels are constructed according to the highest level of education attained by either parent. Three levels are high school, college, and graduate school. The level of high school is the reference group in the regression analysis.
(3) Parent occupation: The three types of parental occupation constructed are professional or clerical workers, sales and service workers, and other. The classification is based on the differentiation of white collar and blue collar jobs as well as the consideration of the sample size of each category.
(4) Monthly family income: The monthly family income is divided into less than NT$20,000, NT$20,000 to less than NT$50,000, NT$50,000 to less than NT$100,000, and NT$100,000 or above.7
(5) Living with both biological parents: This variable is dummy coded with 1 indicating yes and 0 indicating no.
(6) Sib size: This variable is constructed from student’s answers to four questions regarding the number of younger and older sisters and brothers. This number
7 The annual average exchange rate between NT dollars and US dollars is about34.999 to 1 in 2001 (Directorate-General of Budget, Account, and Statistics, Executive Yuan 2007).
21
of siblings is double checked and corrected with five questions about whether living with siblings, the number of siblings under 18 years old, if parents are partial towards a particular sibling, and the relationship between siblings.
(7) Parental educational expectation: This variable is coded into three levels of educational expectation. They are expectation of getting a high school diploma, getting a college degree, and getting a graduate degree. The reference group in the regression analysis is getting a high school diploma.
3. School and class characteristics
(1) Attend private school: 1 means yes and 0 means no.
(2) The 9th grade class is a high ability class: 1 means yes and 0 means no.
(3) Poor class climate for learning: This is an ordinal variable ranges from 1 to 4 with 1 indicating strongly agree and 4 indicating strongly disagree.
(4) Class average grade is good: Same as (3).
(5) Classmates often discuss homework or study together: Same as (3).
(6) Intense academic competition among classmates: Same as (3).
(7) Classmates often discuss about the entrance examination: Same as (3).
(8) School location: This variable is about the level of urbanization of the school location. The three levels are rural, small city, and major city. Rural is the reference group in the regression analysis.