Longitudinal studies involve multiple observations of the same individuals over a period of time.
Observations taken repeatedly across individuals tend to be similar, just as is the case for observations taken from individuals within a cluster (O'Connell & McCoach, 2008). Multilevel models, such as HLM, can not only model the nested data but can also examine how individuals change or grow over time, and identify factors related to such growth or change. HLM accommodates intra-individual
correlation among individuals within the longitudinal study to serve as the nesting unit or cluster for each individual (O'Connell & McCoach, 2008), and thus minimizes the treat of unit heterogeneity: one expects more similarity for observations of the same unit at different times than for simultaneous observations of different units (Halaby, 2003). Numerous traditional longitudinal approaches, such as repeated-measures MANOVA, cannot easily handle unbalanced longitudinal data, missing data, or uneven time points (Luke, 2004). Applying HLM to longitudinal research can not only accurately estimate causal effects, but can also describe how the trajectories of development and growth over time vary systematically across groups with different life course experiences (Halaby, 2003). Latent curve analysis (LCA) is another approach for modeling growth, and the two approaches of HLM and LCA are appealing because they both model individual growth as a function of time, compare different growth rates across different groups, and yield comparable results (Chou, Bentler, & Pentz, 2009). Both approaches have different advantages, and HLM can more easily specify models and yield statistical results (Chou, et al., 2009).
CHAPTER 3 METHODOLOGY
Data
Data were obtained from the Taiwan Education Panel Survey (TEPS). TEPS is a nationally representative longitudinal dataset containing data on Taiwanese adolescents from seventh grade to twelfth grade. TEPS was designed, administrated, and supported primarily by Academia Sinica, which is the highest academic institution in Taiwan. TEPS is intended to assess the educational performance of adolescents, including in mathematics, reading, science, and problem-solving. Students answered both the assessment questions and a questionnaire on their background, including learning habits, friends and family. Parents and teachers also completed a questionnaire on the socioeconomic status of their family and the performance of their children or students. TEPS included four waves of surveys - Wave 1 in 2001, Wave 2 in 2003, Wave 3 in 2005 and Wave 4 in 2007 (Table 1). Stratified random sampling was used. The strata were based on urbanization (namely urban, rural etc.), private or public school, and school level (namely junior high school, senior school etc.).
The sample was divided into three groups during the administration of the first survey wave in 2001: junior high school student sample (n=20,077), senior high school student sample (n=14,610), and junior college student sample (4,679). Two waves (2001 and 2003) were conducted for the sample of senior high school student, and only part of junior high school student sample (n=4,000) were followed up to wave 3 and wave 4 surveys (core panel). New sample (n=16,000) were included in wave 3 and wave 4 surveys for the group of junior high school student sample to make the wave 3 and wave 4 samples more comparable with wave 1 and wave 2 samples.
The completion rates for the senior high school students in the wave 1 and wave 2 surveys were 99.6% and 99.7%, respectively; meanwhile, the rates for the junior high school students in waves 1 to 4
were 99.6%, 99.1%, 96.5% and 99.3%, respectively; and the rates for the junior college student sample were 95.8%, 99.4%, 90.8% and 96.9%, respectively. This study selected only the core panel (students from the junior high school sample who had completed all four survey waves).
Survey Research Data Archive (ARDA) provides three database versions: public, members only, and restricted. This study used five data files of member only versions, also used in previous studies (Chang, 2003, 2005, 2007, 2009), obtained from SRDA website
(https://srda.sinica.edu.tw/group/scigview/2/8): “w1_j_s_lv6.0” (wave 1 students), “w2_j_s_lv6.0”
(wave 2 students), “w3_sf_s_cp_lv6.0” (wave 3 students), “w4_sf_s_cp_lv6.0” (wave 4 students), and
“w1_j_p_lv6.0” (wave 1 parents). Five data files were merged, and all results were weighted using the variable “w1stwt3”, an approach suggested by ARDA that can be inferred to all 7th grade students in 2001.
This study investigates the effect of pre-divorce parental marital conflict and parental divorce on adolescent mental health following parental divorce. Adolescents who (1) completed surveys in wave 1
~ wave 4, and (2) whose parents had either never divorced or divorced before or during elementary school were analyzed in SPSS (n=3,958).
Table 1. Data collection in TEPS
year 2001 2002 2003 2004 2005 2006 2007
Several logical errors in the data file were corrected: (1) three respondents checked both “never divorced” and “divorced during elementary school” in the w4 questionnaire. After comparison with previous answers from the w1 questionnaire, these answers were corrected to “never divorce” to maintain consistency with previous answers; (2) one respondent checked both “never divorced” and
“divorced during senior high school”, and after comparison with the w1 questionnaire for the same respondent the answer “never divorced” was corrected to “divorced during senior high school”. (3) three respondents checked both “no parental conflict” and “parental conflict” as their answers for a single time period, and in these cases the answers were recoded as missing values. (4) ten respondents checked parental divorced more than one periods of school (such as before elementary school and during elementary school), and these answers were corrected to only check the initial period of school.
Because HLM will delete the cases with missing level-2 variables, and respondents with missing vales of in level-2 variables, including “parental education”, “household income”, and “parental conflict” were automatically deleted in HLM. The final sample obtained via HLM thus comprised 3,886 subjects (including respondents whose parents had never divorced, and those whose parents divorced during and before elementary school).
Hypotheses