Investigation of Psychometric Properties of Scales with Missing Data Techniques for Different Sample Sizes and Missing Data Patterns

Abstract

The purpose of this study is to investigate the psychometric properties of scales with different missing data techniques. For this purpose 100 data sets were generated under different conditions of sample sizes (250, 500 and 1000) and number of items (10 and 15), respectively. Data points were deleted under missing completely at random, missing at random and missing not at random conditions by two, five and ten percent. Listwise deletion, similar response pattern imputation based on Euclidian distance, stochastic regression imputation, expectation - maximization algorithm and multiple imputation were carried out on incomplete data sets. Bias of Cronbach alpha, McDonald omega and Omega(W) coefficients were investigated for reliability estimates. Extracted variances and D-2 statistic obtained by principal component analysis and different indices obtained by confirmatory factor analysis are investigated for validity. Results show that listwise deletion, which is often applied as a default missing data technique, may cause serious problems. On the other hand expectation - maximization algorithm and multiple imputation generally outperformed but none of the techniques are the best for all conditions.