[Rasch] Modeling response time

Ricardo Primi rprimi at mac.com
Wed Feb 1 08:00:55 AEDT 2017


Hi Stuart
I guess that you don’t want to model interaction . You want to model main effects of item and person to and then inspect the residuals trying to see who is far off the predicted RT from 
this model considering additive main effects  right? Besides including interaction make this model  non-identified because you have an interaction parameter for each combination of personX item (minos 1)

Usually IRT models are just main effects models (person as a random variable and item as fixed or sometimes random variables). Interactions are used to assess DIF. In this case you have several items from the same group and this makes the identification of interaction possible (you have the same interaction parameter for th items of the same group)

So i think the model would be :

lmer(log(answer_time) ~ qid + (1|examinee), data=test.DF, REML=FALSE)

And then examine the residuals after fitting this model 

Best

Ricardo

> Em 31 de jan de 2017, à(s) 14:56, Stuart Luppescu <lupp at uchicago.edu> escreveu:
> 
> On Tue, 2017-01-31 at 09:13 +0000, Finlayson, Ian wrote:
>> Hello Stuart,
>> 
>> Stepping away from IRT/Rasch, one possibility that jumps out to me
>> would be to fit a mixed effect regression to the log-transformed
>> response times, using persons and items, and then examine the
>> standardised residuals for observations greater than, say, 2.5 SD
>> from zero. That should highlight response times which are longer than
>> might be expected given the person's average response time and the
>> item's average response time.
> 
> Hi Ian, That is exactly what I tried first. I included random effects
> for the interaction of person and item, but the problem is that it is
> an overspecified model, and so it doesn't run. Here's the model:
> 
> lmer(log(answer_time) ~ qid + (1|examinee/qid), data=test.DF, REML=FALSE)
> 
> When I asked about this on the R-sig-mixed-models list, I received this
> reply:
> 
> in this sample data set, there is a single response per question.  This
> will make the qid-with-examinee random effect variance almost
> impossible to estimate (strongly confounded with the observation-level
> residual variance)
> 
> Because of this problem I considered using Rasch analysis.
> -- 
> Stuart Luppescu
> Chief Psychometrician (ret.)
> UChicago Consortium on School Research
> http://consortium.uchicago.edu
> 
> ________________________________________
> Rasch mailing list
> email: Rasch at acer.edu.au
> web: https://mailinglist.acer.edu.au/mailman/options/rasch/rprimi%40mac.com



More information about the Rasch mailing list