Hi there,

I'm undertaking Rasch analysis, using Winsteps, on two separate scales, using survey data. This is for a conference presentation next month (ACSPRI/RC33 in Sydney), and I hope to eventually publish the results in a journal article. The first scale I am working with is dichotomous true/false data with 14 items. It's going well, apart from I can't work back from the person measure to the expectation for each item (using the output observation file XFILE data). The person measure seems to be very close to the mean of the log(expectation/1-expectation) values for the items for each person, but not exactly the same.

I would love to be able to say which output values can be used to substitute in for the original True/False values to run further analyses, say structural equation modelling or factor analysis, on the results. But these require item-level results for each person and not a total score. Do I just divide the log(expectation/1-expectation) by the number of items to get the score to use for each individual item?

Any comments appreciated. This is just for extension for the audience, and in case I get questions.

Michelle, the Winsteps XFILE= contains lots of information about each observation.

It sounds like you want a Rasch person measure corresponding to each observation, this is XFILE column PPMEAS "Predicted person measure from this response alone"

Thanks for the reply Mike, I have been pondering it.

I've gone through the person measures for 4 subjects, and I can see that the value for a "correct" response on the same item is the same for each person, and the value for an "incorrect" response is also the same.

Is there somewhere I can read up on how one gets from the Measure Difference to the Predicted Person Measure? There appears to be a negative correlation between the two, from the small sample of data I have looked at.

I've attached a subset of my data, the first 5 of 1127 subjects. I've used colour coding so I can see where subject sets of data end and start.

Michelle, the "measure difference" is the difference in logits between the ability of the person and difficulty of the item.

The "predicted person measure" is the same for every person who responds in the same way to the same item. It is an estimate of the ability of the person based on the response to that one item.

If you are looking for the person's ability based on the responses to all the items, then that is the "MEASURE" in Table 18. If you are looking for the person's ability based on only one item, then that is the "predicted person measure....".

Are you looking for something else?