1) When I first view the abstract of this paper, I immediately had a question that how can we do the comparison between the abilities which are based on different cognitive process (dimensions) and different set of difficulty parameter. After I read the result, I found the author didn't compare the extent of absolute location between abilities or difficulties. It is reasonable that the location can't be compare directly. The conclusions are majorly based on the correlation and variance of parameters. The author examine the model fit Indices such as AIC, BIC, chi square and likelihood ratio, and the results are used to inference the cognitive process. 3P3I was the best model. It was explained as the evidence of that the slow and fast responses are caused by different processes.
2) I considered about that why those conclusions based on the result of correlation and variance needed to be investigated by this complicated model. The same conclusions can be simply obtained by the correlation and variance of raw score or real time data. But the complicated model indeed is more accurate approach than the traditional analysis by raw score. It is because the parametric model can give us the standard error of parameters and the accurate correlation value which included the error of latent traits.
3) There are two kinds of definition of slow and fast, within persons and within items split. The author divided the time date into the fast 50% and slow 50% within persons or items. Then this categorical data can be analyzed by item response theory, the person and item parameter can be scaled. However, my argument is that why did the continuous data need to be transformed to the categorical data and then latent trait model is used to obtain the continuous scale? My suggestion is that we just need to use real data of time ( or log-time) then a continuous latent trait model can be used for obtaining the item and person's parameter. it is because making the cut point for transforming data from continuous to categories must loss the information which original data had. At the same time, the problem of location of cut points (e.g. why to choose 50% for two categories rather than 33.3% for three categories?) can be avoided.