Date of Award
Morgridge College of Education, Research Methods and Information Science, Research Methods and Statistics
“Modern” methods of addressing missing data using full-information maximum-likelihood (FIML) have become mainstays in SEM analyses. FIML allows the inclusion of auxiliary variables which carry information that is related to missing values and can reduce bias in parameter estimates. Past research has illustrated the benefits of auxiliary variable inclusion under different missingness conditions (MCAR and MNAR; e.g., Enders, 2008), missingness proportions (e.g., Collins et al., 2001), and although limited, missingness patterns (e.g., Yoo, 2009) in FIML analyses. While past studies have focused on the effects of either continuous or ordinal auxiliary variables, no study has included both types in their study despite the fact that their inclusion would more likely reflect real-world conditions. Using a substantive model based on a study by Forber and Krueger (2019) which studied the impact of recession related events on mental health stemming from the 2008 recession, population parameters were drawn from MIDUS 2004-05 (Ryff et al., 2004-2006) and 2013-14 (Ryff et al., 2011-2014). A Monte Carlo simulation tested 3 levels of overall sample size (100, 250, 500), auxiliary variable proportions of missingness (10, 25, and 50%) and auxiliary variable patterns of missingness (Balanced, MAR-linear, and MAR-convex). Factorial ANOVAs were conducted to assess the effects of the three design factors and their interactions. The results found that overall, sample size had the most consistent significant effects on parameter estimates, standard errors, mean squared errors, and relative bias with all performance measures decreasing with increased sample size. High proportions of missingness of up to 50% did not affect the performance measures supporting the theoretical robustness of MAR missingness. Patterns of missingness in the auxiliary variable did not have a main effect, either. This study showed that with smaller sample sizes (i.e., n = 100), while there were some interactions of sample size/pattern/missingness on performance measures but overall, these did not negatively impact bias or model performance measures. The findings from this study were generally in line with past research in that addition of a combination of ordinal and continuous auxiliary variables can provide at least neutral to positive benefits to models with missing data.
Copyright is held by the author. User is responsible for all copyright compliance.
Salina Wu Whitaker
Received from ProQuest
Whitaker, Salina Wu, "The Combined Impact of Continuous and Ordinal Auxiliary Variables on Missing Data Imputation in SEM" (2021). Electronic Theses and Dissertations. 2016.