MUHAMMAD NAEEM SARWAR, ZAHIDA JAVED, SYEDA NAUREEN MUMTAZ, ZUBAIR YOUNAS, SYEDA RABIA BASRI, MUHAMMAD ASIF SHAHZAD, SUMERA BARI, SHAFAQ NAZ
DOI: https://doi.org/Missing data can compromise the validity, reliability, and generalizability of psychometric assessment results, especially in mainstream school populations. This study examines the impact of missing data on educational assessments of students in mainstream education schools. A mathematics achievement test was administered to 275 students, and missing data patterns (MCAR, MAR, MNAR) were analyzed. Missing data were dealt with different approaches including classical techniques (mean imputations, listwise deletions), statistical techniques (Bayesian estimations, multiple imputations), and machine learning approaches (k-nearest neighbors, random forests). Using root mean square error, bias and recovery of psychometric parameters within IRT frameworks evaluated the techniques. Bayesian estimation and multiple imputations generally surpassed the other techniques, on most of the evaluated criteria. The results support the need of making educational assessments more reliable for underrepresented students and the need of data-informed decision making for more representative mainstream school populations.
