Anselmi, P., Robusto, E., & Cristante, F.
Analyzing missingness at the level of the individual respondent: Comparison of five statistical tests
To date, missingness has been investigated at the level of the entire sample of respondents, on data samples deriving from a unique mechanism (either MCAR, or MAR, or MNAR). In this work, missingness is investigated at the level of the individual respondent, on data samples consisting of patterns of missingness deriving from different mechanisms. Five tests (runs test, Wilcoxon rank sum test, point-biserial correlation, t-test, standardized outfit) are used for detecting whether the particular pattern of missingness displayed by an individual respondent has been generated by a random (MCAR, MAR) or nonrandom (MNAR) mechanism. For all the tests, the rejection of the null hypothesis for a particular pattern is taken as an indication that the pattern might have been caused by a nonrandom mechanism. A simulation study and a real data application example showed that Wilcoxon rank sum test and t-test outperform the other tests in identifying the MNAR patterns.
Testing, Psychometrics, Methodology in Applied Psychology, 2018, Vol. 25, pp. 379-394, DOI: 10.4473/TPM25.3.4