A Note on the Use of Item Parceling in Structural Equation Modeling with Missing Data

Abstract

Parceling is referred to as a procedure for computing sums or average scores across multiple items. Parcels instead of individual items are then used as indicators of latent factors in the structural equation modeling analysis (Bandalos 2002, 2008; Little et al., 2002; Yang, Nay, & Hoyle, 2010). Item parceling may be applied to alleviate some problems in analysis with missing data (e.g., MCAR, MAR, and MNAR) and/or nonnormal data. No simulation study has been conducted to examine whether using parceling leads to better (at least not worse) results than individual items when there are missing values and nonnormality issues in the dataset. The purpose of this study is to investigate how item parceling behaves under various simulated conditions in structural equation modeling with missing and non-normal distributed data. The design factors of the simulation study included sample size, missingness mechanism, percentage of missingness, degree of nonnormality for individual items, and magnitude of factor loadings. For each condition, 2000 datasets were generated. Each generated dataset was analyzed at both parcel and item levels using full information maximum likelihood estimation method. All analysis models were considered correctly specified. The results of the simulation showed that models based on parcels were less likely to be rejected than those based on individual items. Specifically, parcel analysis tended to result in smaller empirical alpha based on the chi-square test, greater CFI, and smaller RMSEA and SRMR. In addition, Parameter estimates from parcel level analysis performed equally well or slightly better than those from item level analysis in all conditions. In general, parcel level analysis yielded results as good as or better than those from item level analysis under any type of missing mechanisms, different degrees of nonnormality of data, and percentage of missingness.