![](https://d3ilqtpdwi981i.cloudfront.net/CGCZ24Rkoq9wSkwfBjYPQmbqz_g=/425x550/smart/https://bepress-attached-resources.s3.amazonaws.com/uploads/ed/15/02/ed150263-33df-445d-adc3-83f51556a499/thumbnail_c0f54ec6-89f5-4614-8142-926cffc2818c.jpg)
This study used Monte Carlo simulations to evaluate the item parameter recovery from ACER ConQuest 3 software (Adams, Wu, & Wilson, 2012) for the dichotomous Rasch model. The authors’ primary focus was the comparison of its estimation methods, joint maximum likelihood (JML), marginal maximum likelihood (MML) with a normal distribution assumption and MML with a discrete distributions assumption when the populations were in fact non-normal. The simulation data sets were generated with two test lengths (10 and 50 items) and four alternative true population distributions for the abilities: normal, bimodal, uniform, and chi-square. As expected, results showed that MML-Normal was the best method when the assumption of ability distribution was matched, regardless the test length. However, the accuracy or MML-Normal decreased with the violation level of the assumption of normal distribution of the latent ability. The MML-Discrete estimation could overcome well the weakness of the MML-Normal when the normality of the ability distribution was violated. The estimates of the corresponding standard errors produced by ACER ConQuest 3 were also being examined and discussed.
Available at: http://works.bepress.com/ray_adams/30/