The Impacts of Nonnormality, Missing Mechanisms, and Multicollinearity on Bollen-Stein Bootstrapping Technique with Longitudinal Data

국문 초록 (Abstract)

  본 연구의 목적은 결측치를 포함한 비정규분포 종단 데이터 하에서 Bollen-Stein 부트스트랩 기법의 효과성을 검증하는 것에 있다. Bollen-Stein 부트스트랩 기법은 비정규분포성에 의해 발생하는 표준오차, 모형 적합도의 편향을 보정해 주기 위해 최근 새롭게 개발된 기법이다. 전통적 부트스트랩 기법과 달리, Bollen-Stein 부트스트랩 샘플들의 공분산구조는 원 데이터의 공분산 구조와 일치하도록 조정되어 편향되지 않는 모수 값을 제공할 수 있다고 하였다. 이 연구에서는 비정규분포, 결측치 메커니즘 (임의적 결측치 생성 및 체계적 결측치 생성), 다중공선성, 그리고 다양한 표본 수 조건들 하에서 로버스트 최대우도추정량기법, 점근성 자유 분포 기법, 그리고 Bollen-Stein 부트스트랩 기법들의 모수값, 표준오차, 모형 적합도 편향 정도를 검증해 보았다. 결과는 로버스트 최대우도추정량 기법과 Bollen-Stein 부트스트랩 기법이 모수 값과 표준오차 값에서 비슷한 결과를 나타내었으나 다중공선성의 이유로 표준오차 값은 과대 추정되었다. 모형 적합도에서는 로버스트 최대우도추정량기법의 결과치가 가장 우수한 것으로 보고되었다. 일반적으로 체계적 결측치 메커니즘, 다중공선성, 그리고 비정규분포는 모수값 및 모형적합도 등의 통계적 결과치들을 심각하게 왜곡하였다. 자세한 설명과 여러 이슈들에 대해 토의하였다.

다국어 초록 (Multilingual Abstract)

  This study investigates the effects of Bollen-Stein bootstrapping technique on statistical influences in nonnormal and incomplete longitudinal data. Recently, Bollen-Stein bootstrap (BSB) has been proposed to correct for standard error and fit statistic bias due to nonnormality. Unlike a naive bootstrapping, covariance structure of the bootstrap samples in BSB is consistent with that of parent data matrix. The purpose of this article is to compare the test statistics of three different estimation techniques (robust maximum likelihood method; Robust ML, asymptotic distribution free method; ADF, BSB). The effects of various research conditions were investigated on parameter estimates, standard error estimates, and model fit statistics. The longitudinal research design included two missing mechanisms (missing at random; MAR, missing not at random; MNAR), three types of sample sizes (N = 200, 500, 1000), multicollinearity, and severe nonnormality (skewness=5.0 & kurtosis=70.0). Results indicated that BSB and robust ML yielded similar parameter and standard error estimates, while both procedures provided overestimated standard errors due to multicollinearity. For the analysis of model fit, results noted that robust ML outperformed BSB and ADF across all cell designs. In general, nonrandomly missing (MNAR), multicollinearity, and nonnormality increased the degree of bias toward the test statistics. The detailed explanations and other practical issues are discussed.

목차 (Table of Contents)

• Ⅰ. Background and Purpose of Study
  Ⅱ. Bollen-Stein Bootstrap
  Ⅲ. Methods
  Ⅳ. Results
  Ⅴ. Discussion
  References
  저자소개
  〈요약〉
  

참고문헌 (Reference)

1 Gold. M. S., "Treatment of missing data: A Monte Carlo comparison of RBHDI, Interactive Stochastic Regression Imputation, and Expected Maximization" 7 : 319-355, 2002

2 Biesanz, J. C., "The role of coding time in estimation and intercepting growth curve models" 9 : 30-52, 2004

3 Curran, P. J., "The robustness of test statistics to nonnormality and specification error in confirmatory factor analysis" 1 : 16-29, 1996

4 Olsson, U. H., "The performance of ML, GLS, and WLS estimation in structural equation modeling under conditions of misspecification and nonnormality" 7 : 557-595, 2000

5 Enders,K.C., "The impact of nonnormalilty on full information maximum-likelihood estimation for structural equation models with missing data" 6 : 352-370, 2001

6 Vasu, E. S., "The effect of multicollinearity and the violation of the assumption of normality on the testing of hypotheses in regression analysis" 1975

7 West, S. G., "Structural equations with non-normal variables: Problems and remedies in:Structural Equation Modeling: Issues and applications" Sage 56-75, 1995

8 EQS, "Structural equation modeling software (version of 6.1 B91), Multivariate Software, Inc"

9 Chou, C. P., "Scaled test statistics and robust standard for non-normal data in covariance structure analysis: A Monte Carlo study" 44 : 359-368, 1991

10 Nevitt, J., "Performance of bootstrapping approaches to model test statistics and parameter standard error estimation in structural equation modeling" 8 : 353-377, 2001

11 Grewal, R., "Multicollinearity and Measurement Error in in Structural Equation Models: Implications for Theory Testing" 23 : 519-529, 2004

12 Muth?,B.O., "Goodness of fit with categorical and other nonnormal variables in:Testing structural equation models" Sage Publications 1993

13 Chou, C. P., "Estimates and tests in structural equation modeling in:Structural equation modeling: Issues and applications" Sage 37-55, 1995

14 Shin, T., "Effect of missing data methods in Structural Equation Modeling with nonnormal longitudinal data" 2008

15 Fouladi, R. T., "Covariance structure analysis techniques under conditions of multivariate normality and nonnormality-Modified and bootstrap based test statistics" 1998

16 Shin,T., "Comparison of three growth modeling techniques in multilevel analysis of longitudinal academic achievement scores: Latent growth modeling, Hierarchical linear modeling, and Longitudinal profile via multidimensional scaling" 8 (8): 262-275, 2007

17 Hu, L., "Can test statistics in covariance structure analysis be trusted" 112 : 351-362, 1992

18 Bollen, K. A., "Bootstrapping goodness-of-fit measures in structural equation models" 21 : 205-229, 1992

19 Beran, R., "Bootstrap tests and confidence regions for functions of a covariance matrix" 13 : 95-115, 1985

20 Browne,M.W., "Asymptotically distribution-free methods for the analysis of linear latent variates models" 41 : 193-208, 1984

21 Enders,K.C., "Applying the Bollen-Stein bootstrap for goodness-fit measures to structural equation models with missing data" 37 : 359-377, 2002

22 Enders,K.C., "An SAS Macro for implementing the modified Bollen-Stein bootstrap for missing data: Implementing the bootstrap using existing structural equation modeling software" 14 : 620-641, 2005

23 Browne, M. W., "Alternative ways of assessing model fit in:Testing Structural Equation Models" Sage 136-162, 1993

24 Savalei, V., "A statistically justified pairwise ML method for incomplete nonnormal data: A comparison with direct ML and pairwise ADF" 12 : 183-214, 2005

25 Cool A.L., "A review of methods for dealing with missing data" 2000

26 Fleishman.A.I., "A method for simulating non-normal distributions" 43 : 521-532, 1978

27 Shin,T., "A comparison of missing data methods on statistical influence in latent growth modeling" 19 (19): 193-214, 2006

28 Gold, M, S., "A comparison of maximum-likelihood and asymptotically distribution-free methods of treating incomplete nonnormal data" 10 : 47-79, 2003