The use of hypotesis tests such as the "t-test", in quality improvement has
many pitfalls. Firstly, the assumptions in any test must be validated,
including the probablity model used and the degree of randomness of
samples.
Secondly, the confidence interval chosen for the hypotesis has no specific
validity. Most im****tantly, such tests dispose of the time element of the
data. The results are purely historical and make no prediction as to
future
behaviour, as do control charts. It is for such reasons Deming made his
statements about the poor teaching of statistics and that hypothesis tests
should never be used for analytical problems, such as in the study of
processes.
Hypotesis testing should be confined to enumerative studies, such as in
psychological tests ... the area where six sigma's leader, Mikel Harry
holds
his qualifications.
Dr Tony Burns
www.q-skills.com
"Raymond J. Johnson Jr." <Rayjay@[EMAIL PROTECTED]
> wrote in message
news:qU46i.4461$C96.2954@[EMAIL PROTECTED]
> gn228169 wrote:
>> Raymond J. Johnson Jr. wrote:
>>> gn228169 wrote:
>>>
>>>> When analyzing a new method or process, we usually use a t-test to
see
>>>> if the new method is equivalent. If it is not, we look at the
variance
>>>> and target values to see if it failed because it was "better" Is
there
>>>> a statistical test to check for "better" without checking first for
>>>> equivalence and then for variance and target?
>>>> Thanks, Jim
>>>
>>>
>>> You've got some 'splainin' to do. What's the difference between a
>>> "process" and a "method"? You put "better" in scare quotes, apparently
>>> because you understand that it's a relative concept, but you don't
tell
>>> us about what makes "better" better. What are your criteria? Are you
>>> considering only output values? Are there fiscal considerations (i.e.,
>>> given equivalent output, is one process better because it's less
>>> costly)? What, exactly, are you trying to do?
>>
>> A process and a method are the same thing, sorry. Better is less
>> variation and closer to target. A one sided t-test will not tell me if
>> the new method has less variation and is closer to the target as one
>> person suggested.
>> I think cpk,ppk,cpm etc. will work but I still have a question. If a
cpk,
>> cpm score is greater using the new method, I would assume the new
method
>> is "better". My question is, How much of an increase in Ppm is
>> significant? Example: If my Ppm value was a 0.90 and increased to a
0.99
>> using the new method, is this due to common noise or is this a truly
>> "better"?
>
> It's only "better" if it saves money. Nitpicking the numbers *wastes*
> money. It's not clear whether you're trying to solve a particular
problem
> or just dealing with theory, but if you're talking about a difference of
> .09 PPM, it's hard to believe it's worth worrying about.
|
