The Impact of Missing Data on the Detection of Nonuniform
Differential Item Functioning
W. Holmes Finch
Since prior work has demonstrated that the presence of missing data and the methods used to deal with them can have a marked impact on the ability of commonly used techniques to accurately identify the presence of uniform DIF and former studies mainly focus on uniform DIF, then it is important to check the impact of missing data on nonuniform DIF detection. The logistic regression, crossing SIBTEST, and the item response theory likelihood ratio test methods are evaluated in terms of hypothesis testing and effect size estimation.
There are three types of missing data: missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR). The methods to treat missing data are listwise deletion, zero imputation, multiple Imputation, Stochastic regression imputation.
The simulation study was conducted.
Simulation: 3PL model/20/40 items
Sample size:250/250,500/500,/1000/1000
Means: 0/-0.5 0/0.5
Percentage of missing data: 0%,10%,20%,30%
DIF:0, 0.4, 0.8, 1
The results show across all conditions presented, LR and CSIB produced very comparable Type I error rates, whereas IRTLR had somewhat higher rates than the other methods. LD produced results very similar to those obtained with the complete data.MI appears to be much preferable to SRI. However for different missing types, the results are different and inconsistent .
Comments:
1 The missing data are more important in small sample size situation, so the future study can consider this.
2 The unbalance sample design can be consider in the future.