Application and Comparison of Missing Data Methods for Factor Analysis and Multinomial Logistic Regression

Abstract

Higher education institutions collect a plethora of data on their students each year from admissions, to learning management systems, to satisfaction surveys. Computer software has made collecting data from participants ubiquitous in the college environment. Despite having the ability to collect a large amount of data, missing data values for certain fields remains a challenge for successful data analysis, particularly with surveys covering sensitive topics. Traditionally, missing data would lead to removing observations which makes analysis easier but could compromise estimates. This methodological paper will review and compare conventional and advanced missing data methods of listwise deletion, regression imputation, full information maximum likelihood (FIML) and multiple imputation (MI) in the context of confirmatory factor analysis and multinomial logistic regression analysis. The data set used for these analyses is a campus climate survey conducted at a 2-year college in Washington State. The confirmatory factor analysis will cover the construct of campus climate, while the multinomial logistic regression will make use of the demographic variables. The results of the present study showed first that, in the confirmatory factor analysis model, the five missing data handling methods resulted in similar loading estimates and fit indices, except that the categorical multiple imputation technique performed better than the other methods in regard to precision of estimates and standard errors. With respect to the multinomial logistic regression model results, use of listwise deletion resulted in a larger confidence interval range for coefficients and higher standard errors than the imputation methods. In short, we recommend use of multiple imputation across all sample sizes, and for larger samples, use of FIML is also appropriate. Listwise deletion is not recommended.