In low-stakes assessment, there is often a substantive amount of missing data observed from omitted and not-reached items. Missing item responses are often MNAR and lead to biased parameter estimates if not handled properly. This study evaluated and compared the accuracy of parameter estimates between classical approaches and model-based approaches. The study contains an empirical data analysis and a simulation study. For the empirical data analysis, a subset of data from PISA 2018 reading portion of test was selected. Item responses from nine countries were calibrated using traditional approaches and two of the model-based approaches and subsequently compared. It was found that person parameter estimates of the model-based approach are the closest to the model scoring omits as wrong and not-reached as administered. The item parameter estimates from the model-based approaches coincided with the estimates from the model currently adopted by the program for most countries. For the simulation study, it was found that scoring both omitted and not-reached items as not-presented was sufficient to retrieve unbiased parameter estimates in most conditions, even when the rate of missing item responses was extreme.