This paper involved the two interesting issues, DIF and missing data. In practical data, It is hard to avoid missing data. there are several method to deal with missing data introduced by this paper, such as listwise deletion (LD), Omitted as incorrect (ZI), Multiple imputation (MI), and Stochastic Regression Imputation (SRI). LD method is directly removing the examinee with missing value. ZI is scoring zero to substitute for missing value. MI and SRI are creating difference data set by sampling and then combining the analysis result from those data set. The different between MI and SRI is that SRI include the regression in the step two for drawing the value filling the missing data. The analysis here is DIF detection. The nonuniform DIF is emphasized in this paper because this issue was less investigated. The result shows that the MI and LD methods are recommended.
1. In the table 1 and 2, sometimes the LD method can perform better than complete data. I think it corresponded to the influence of sample size in Table 2. The size must decrease in LD, and It cause the Type I error of the LR detection increased.
2. I think it is better to report the effective sample size adopted in LD method.
3. Imputation based on Rasch model is mentioned in the suggestion of the future study. Giving a proportion of correct response or giving a score is considered. Which one is reasonable?
4. MNAR is simulated as the missing data were taken from an incorrect response. I think the more reasonable simulation situation is that missing data should be taken from the probability of correct response at or below 50 %. It corresponds with the conception of an examinee who did not know the correct answer.