Winsteps

In 1996, Yellow Freight Systems, a freight-hauling company with 13,000 trucks, had major financial problems. Since their start in the 1920s, YFS had used an "efficiency" model: maximize the use of drivers, trucks, freight terminals, etc. The result was that trucks ran full and cheaply, but customers had to schedule their own operations to fit with YFS schedules, and YFS schedules weren't reliable because if part of a load was delayed in loading for some reason, then the whole truck was delayed, and schedules for that truck were missed. Further, freight handling procedures were low priority, so freight was damaged. The result was that 40% of loads were picked up late, delivered late or delivered damaged according to http://www.fastcompany.com/magazine/54/yellow.html?partner=rss
So YFS was running a system in which they optimized their view of the data: trucks, drivers, etc., under the assumption that "most cost-effective resource management => supposedly-best customer satisfaction => highest profitability".

But YFS was failing financially because customers saw things differently. They didn't care about YFS efficiency. They wanted loads picked up on the customer's (not YFS's) schedule, also loads delivered on the customer's (not YFS's) schedule, and loads delivered intact (not damaged), and shipping charges were only a secondard consideration. For the customers, "best customer satisfaction => most profit-productive resource management => highest profitability".

In 1997, YFS switched to using the customer's perspective, and the company is thriving. Trucks are running 90% full, but customer satisfaction has increased from 60% to 95%.

It is the same in data management. The typical social-science philosophy is: "best-fitting data summarizing description" => "imagined-to-be-best basis of inference for decision making". In an attempt to capture 100% of the information in the data, idiosyncratic parameterizations of the data are welcomed, so long as they provide the statistical "'best fit", even if that means that the parameters are uninterpretable, counter-intuitive or the estimates are unstable across analyses.

The Rasch philosopy is: "conceptually-best basis of inference for decision making" => "most useful data summarizing description". So the parameterization is chosen to reflect the bases for inference that have proved to be most successful. In science and industry, these are linearity, parameter separability, ease of conceptualization and stability of the findings.

Metaphorically, Rasch uses 90% of the information to increase inferential utility from 60% to 95%.

Mike Neiss on http://www.tompeters.com/: "I remember Dr. Deming telling us at a [General Motors] management meeting that the head of Toyota wasn't afraid to share their 'secrets' with us because we couldn't even see what we needed to see and besides 'GM management will never do it.' "

We've been giving away Rasch 'secrets' for 40 years, and still social scientists can't see what is needed, and the big statistical packages, SPSS and SAS, still don't implement Rasch usefully.

But my suspicion is that the folks in Taiwan and other parts of Asia are getting the message ... Really funky Rasch applications and software are emerging ....

Can one have too much of a good thing? Certainly one can with loyalty. Everyone agrees that too little loyalty, i.e., disloyalty, undermines an organization. But too much loyalty means that doubtful instructions go unquestioned and new boat-rocking ideas ("creative dissent") are not presented, no matter how good they are.

Rasch fit statistics are much the same. Too much variation in the data, unmodeled noise, drowns out the measurement music. But too little variation and the music turns into the sound of marching boots.