equivalent analyses for Winsteps and Conquest

[gsbeglia]

My lab is in the process of trying to move from Winsteps to Conquest and we need to make sure we are running equivalent analyses in the two programs. The first question I have is about the ICC figures, which have different x axis labels and look somewhat different in Winsteps and Conquest (I am unable to attach a document with ICC figures because of their size but the probabilities for selecting each answer option are somewhat different for the two programs). I want to determine if both programs have the same axes and can be interpreted the same way. The  x-axis in Conquest is "latent trait (logits)" and in Winsteps the x-axis is "person measure minus item measure".

The second question I have is about determining dif. Using these item*gender estimates in ConQuest, I would like to apply the same cutoffs for determining the presence of dif that we have historically used for Winsteps outputs. In a Winsteps dif analysis, there are estimates given for each gender that we then subtract from one another to determine the magnitude of dif between males and females. Are these Winsteps estimates equivalent to the item*gender estimate that ConQuest generates? If so, can I subtract the item*gender estimates in ConQuest to determine the magnitude of dif?

Gena

[dan_c]

Can you provide some example syntax and output for your diff analysis? The estimates for each item*gender interaction should provide you an estimate and std error plus fit statistics that you can use to judge diff. You could also use the mh command:

https://www.acer.org/files/Conquest-Notes-7-UsingMantel-HaenszelStatistics.pdf

[gsbeglia]

Hi Dan,

I attached the two outputs for my dif analysis on race. The second output appears to not have converged well. Is it appropriate to subtract the item*race estimates from one another for each race and item and use the resulting number to determine the magnitude of dif? We have done something like this in Winsteps but I am not sure if the estimate values are equivalent.

Relatedly, it is possible that the data are three dimensional. How would I run a dif for three dimensional data? Would I run dif on each dimension separately?

Gena

title dif_race Partial credit model for DIF analysis for Race;

datafile pre_test_121917.dat;
format sex 12 race 13 responses 14-37;

labels <<labels_sex_and_race.lab;

model item+race+item*race+race*item*step;

estimate !fit=no,stderr=quick;
show !table=1:2 >> sample_race_dif_output.shw;
plot expected!gins=1:2, overlay=yes, legend=yes;

reset;

datafile pre_test_121917.dat;
format sex 12 race 13 responses 14-37;

labels << labels_sex_and_race.lab;

model item+race+item*race+item*step;

estimate !fit=no,stderr=empirical;
show !table=1:2 >> sample_race_dif_output.shw;
plot expected!gins=1:2, overlay=yes, legend=yes;