Abstract
Item parceling procedure may be applied to alleviate some difficulties in analysis with missing data and/or nonnormal data in structural equation modeling. A simulation study was conducted to investigate how item parceling behaves under various conditions in structural equation model with missing and nonnormal distributed data. Design factors included missing mechanism, percentage of missingness, distribution of item data, and sample size. Results showed that analysis conducted at the parcel level yielded lower model rejection rates than analysis based on the individual items, and the patterns were consistent across missing mechanism, percentage of missing, and distribution of item data. In addition, parcel-level analyses resulted in comparable parameter estimates to item-level analyses.
Highlights
AND LITERATURE REVIEWStructural equation modeling (SEM) has been frequently used in empirical data analysis to examine hypothesized relationships among a set of variables
A commonly used estimation method in SEM, maximum likelihood (ML), requires the sample size be sufficiently large and observed variables be multivariate normally distributed. Violation of these assumptions results in inaccurate model chisquare statistic, fit indices, parameter estimates, and standard errors associated with parameter estimates (e.g., Bollen, 1989; Curran, West, & Finch, 1996; Chou, Bentler, &Satorra, 1991), and the degree of bias tends to increase as the model complexity increases
MLR corrects for the positive bias in model chi-square statistic and the negative bias of standard errors associated with parameter estimates
Summary
AND LITERATURE REVIEWStructural equation modeling (SEM) has been frequently used in empirical data analysis to examine hypothesized relationships among a set of variables. A commonly used estimation method in SEM, maximum likelihood (ML), requires the sample size be sufficiently large and observed variables be multivariate normally distributed. Violation of these assumptions results in inaccurate model chisquare statistic, fit indices, parameter estimates, and standard errors associated with parameter estimates (e.g., Bollen, 1989; Curran, West, & Finch, 1996; Chou, Bentler, &Satorra, 1991), and the degree of bias tends to increase as the model complexity increases. MLR corrects for the positive bias in model chi-square statistic and the negative bias of standard errors associated with parameter estimates Another alternative is weighted least square (WLS). WLS requires very large sample sizes even when the data are Eğitimde ve Psikolojide Ölçme ve Değerlendirme Dergisi, Cilt 7, Sayı 1, Yaz 2016, 59-72
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
