Dr. Linacre, I am a Data Analytics undergraduate student and researcher at Ohio State, and am applying the Rasch model to survey data I have on childhood stressors. This helped me to understand the model so much better. Thank you very much! Elizabeth Gilbert
This is a great lecture, Dr. Linacre. Would you upload your lectures on Marginal Maximum Likelihood Estimation as well, if you have any recordings of those?
Very nice presentation. I would assume that it would be difficult to show JMLE (the unconditional) in action. Dr. Linacre, is there a real improvement in terms of the Marginal Maximum likelihood worked out by Darrell Bock that IRT programs use?; when I ran simulations the differences were really negligible. Peter Paprzycki
Thanks, Peter. To see JMLE in action, there is Mark Moulton's Excel spreadsheet at www.rasch.org/moulton.htm and, yes, in practice differences between estimation methods are usually inconsequential.
Thank you! I can imagine that one needs at least Excel to do these things! I can also see how the calibration can get murky if the counts between categories are grossly disproportionate. Does the Winsteps program actually use the sum of the items' variance in the division to compute the distance in the next step of the NR method?
You are right, Peter. Newton-Raphson does not work well when estimating Andrich thresholds. Winsteps uses a curve-fitting estimation method instead. This is implemented in the polytomous Excel spreadsheet linked from www.rasch.org/moulton.htm
Dr. Linacre, I also noticed on sample datasets that the estimated Rasch discrimination problems when I request, discrim = yes, corresponds almost perfectly with Infit problems. Would this be a case that Infit/Outfit can be an index of problems with discrimination?
John Michael Linacre I think this is a very good argument for folks that claim we need extra parametrizations. I speculated about it for a long time, and this is the best proof! Thank you for pointing to the reading.
Peter: extra parameterizations -> extra assumptions. When estimating item discrimination parameters, the assumptions usually include a normal theta distribution, ceiling and floor values for the item discriminations, and also some other limit to prevent the item discrimination estimates all becoming those floor and ceiling values. And, of course, the nice mathematical, statistical and inferential properties of the Rasch model are lost :-(
John Michael Linacre Yes, and then the fit should not be the most decisive index as Ben Wright explained. I like when he said that he wants to have the most efficient model that explains the most of the situation at hand. Thank you for the readings.
It's a pleasure to understand how Rasch Measurement works from one of the giants of the field himself! Thank you Dr. Linacre for sharing this lecture!
Dr. Linacre,
I am a Data Analytics undergraduate student and researcher at Ohio State, and am applying the Rasch model to survey data I have on childhood stressors. This helped me to understand the model so much better. Thank you very much!
Elizabeth Gilbert
Excellent, Elizabeth. This video is an oldie but a goodie :-)
This is a great lecture, Dr. Linacre. Would you upload your lectures on Marginal Maximum Likelihood Estimation as well, if you have any recordings of those?
Great explanation really, i wish the whole lecture is uploaded. Thanks a lot for the great video.
Hi - Parts 2 and 3 of the lecture are on TH-cam, Search for "Rasch model estimation"
@@johnmichaellinacre4960 oh, that is amazing. You are the best.
Dr. Linacre,
Where could I find one of these tables that you use in this lecture?
Scott, thank you for watching this lecture. At what time-point in the lecture is a relevant table?
It was at about 8:45
In the Winsteps manual and Help file, see www.winsteps.com/winman/whatisalogit.htm
Very nice presentation. I would assume that it would be difficult to show JMLE (the unconditional) in action. Dr. Linacre, is there a real improvement in terms of the Marginal Maximum likelihood worked out by Darrell Bock that IRT programs use?; when I ran simulations the differences were really negligible. Peter Paprzycki
Thanks, Peter. To see JMLE in action, there is Mark Moulton's Excel spreadsheet at www.rasch.org/moulton.htm and, yes, in practice differences between estimation methods are usually inconsequential.
Thank you! I can imagine that one needs at least Excel to do these things! I can also see how the calibration can get murky if the counts between categories are grossly disproportionate. Does the Winsteps program actually use the sum of the items' variance in the division to compute the distance in the next step of the NR method?
You are right, Peter. Newton-Raphson does not work well when estimating Andrich thresholds. Winsteps uses a curve-fitting estimation method instead. This is implemented in the polytomous Excel spreadsheet linked from www.rasch.org/moulton.htm
John Michael Linacre Thank you for the spreadsheets! I enjoy the presentations!
Dr. Linacre, I also noticed on sample datasets that the estimated Rasch discrimination problems when I request, discrim = yes, corresponds almost perfectly with Infit problems. Would this be a case that Infit/Outfit can be an index of problems with discrimination?
Well observed, Peter. See "Item Discrimination and Infit Mean-Squares" - www.rasch.org/rmt/rmt142a.htm
John Michael Linacre I think this is a very good argument for folks that claim we need extra parametrizations. I speculated about it for a long time, and this is the best proof! Thank you for pointing to the reading.
Peter: extra parameterizations -> extra assumptions. When estimating item discrimination parameters, the assumptions usually include a normal theta distribution, ceiling and floor values for the item discriminations, and also some other limit to prevent the item discrimination estimates all becoming those floor and ceiling values. And, of course, the nice mathematical, statistical and inferential properties of the Rasch model are lost :-(
John Michael Linacre Yes, and then the fit should not be the most decisive index as Ben Wright explained. I like when he said that he wants to have the most efficient model that explains the most of the situation at hand. Thank you for the readings.
Thanks fort sharing thi