Scribd
Upload
3 views

Module-7-2015

Uploaded by

Dame Yenyetou
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
3 views

Module-7-2015

Uploaded by

Dame Yenyetou
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 10

IE 361 Module 7

Measurement and Two Sample Inferences Part 1

Reading: Section 2.2.2 Statistical Methods for Quality Assurance

ISU and Analytics Iowa LLC

(ISU and Analytics Iowa LLC) IE 361 Module 7 1 / 10


Basic Two Sample Inference Formulas
Parallel to the one sample inference formulas are the two sample formulas
of elementary statistics. These are based on a model that says that
y11 , y12 , . . . , y1n1 and y21 , y22 , . . . , y2n2
are independent samples from normal distributions with respective means
µ1 and µ2 and respective standard deviations σ1 and σ2 . In this context,
the so-called "Satterthwaite approximation" gives limits
s
s12 s2
y 1 y 2 t̂ + 2 for estimating µ1 µ2 (1)
n1 n2
where appropriate "approximate degrees of freedom" for t̂ are
2
s12 s2
+ 2
n1 n2
ν̂ =
s14 s24
+
(n1 1)n12 (n2 1)n22
(ISU and Analytics Iowa LLC) IE 361 Module 7 2 / 10
Basic Two Sample Inference Formulas
This degrees of freedom formula is one that you may not have seen in an
elementary statistics course, where sometimes only methods valid when
one assumes that σ1 = σ2 are presented. It turns out that above
min ((n1 1 ) , ( n2 1)) ν̂
so that a simple conservative version of this method (that you are welcome
to employ in IE 361) uses degrees of freedom
ν̂ = min ((n1 1 ) , ( n2 1))
Further, in the two-sample context, standard elementary con…dence limits
for comparing standard deviations are
s1 1 s1 1 σ1
q and q for (2)
s2 F(n 1 s2 F(n 1 σ2
1 ),(n 2 1 ),upper 1 ),(n 2 1 ),lower

(and be reminded that F(n1 1 ),(n2 1 ),lower = 1/F(n2 1 ),(n1 1 ),upper so that
standard F tables giving only upper percentage points can be employed).
(ISU and Analytics Iowa LLC) IE 361 Module 7 3 / 10
Basic Two Sample Inference Formulas

If these formulas do not look familiar, you should immediately stop and
review them. Their use can, for example, be seen in Chapter 6 of
Vardeman and Jobe’s Basic Engineering Data Collection and Analysis or in
the Stat 231 text. Here we will consider a variety of applications of them
to problems that arise in metrological studies for quality assurance. Our
basic objective continues to be to illustrate (and help you develop the
thought process necessary to successfully employ) the basic insight that
How sources of physical variation interact with a data collection
plan governs what of practical importance can be learned from a
data set, and in particular, how measurement error is re‡ected in
the data set.

(ISU and Analytics Iowa LLC) IE 361 Module 7 4 / 10


Measurements of a Single Measurand From Two Devices
Application to Two Samples Consisting of Repeat Measurements of a Single Measurand
Made Using Two Di¤erent Devices

One way to create "two samples" of measurements is to measure the same


item repeatedly with two di¤erent devices. This possibility is illustrated in
the next …gure.

Figure: Cartoon Illustrating Repeat Measurement of the Same Measurand With


Two Devices
(ISU and Analytics Iowa LLC) IE 361 Module 7 5 / 10
Measurements of a Single Measurand From Two Devices

Direct application of the two-sample con…dence interval formulas show


that the two-sample Satterthwaite approximate t interval (1) provides
limits for
µ1 µ2 = (x + δ1 ) (x + δ2 ) = δ1 δ2
(the di¤erence in device biases), while the F interval (2) provides a way of
comparing device standard deviations σdevice1 and σdevice2 . This data
collection plan provides for straightforward comparison of the
characteristics of the two devices.

(ISU and Analytics Iowa LLC) IE 361 Module 7 6 / 10


Application to Two Samples Consisting of Single
Measurements Made With Two (Linear) Devices On
Multiple Measurands From a Stable Process (Only One
Device Used for a Given Measurand)

There are quality assurance contexts in which measurement is destructive


and cannot be repeated for a single measurand, and nevertheless one
needs to somehow compare two di¤erent devices. In such situations, the
only thing that can be done is to take items from some large pool of items
or from some stable process and (probably after randomly assigning them
one at a time to one or the other of the devices) measure them and try to
make comparisons based on the resulting samples. This possibility is
illustrated in the next …gure.

(ISU and Analytics Iowa LLC) IE 361 Module 7 7 / 10


Single Measurements From Two Linear Devices On
Multiple Measurands From a Stable Process

Figure: Cartoon Illustrating Two Samples Composed of Single Measurements on


Di¤erent Measurands From a Stable Process
(ISU and Analytics Iowa LLC) IE 361 Module 7 8 / 10
Single Measurements From Two Linear Devices On
Multiple Measurands From a Stable Process
Direct application of the two-sample Satterthwaite approximate t interval
(1) provides limits for
µ1 µ2 = ( µx + δ1 ) ( µx + δ2 ) = δ1 δ2
(the di¤erence in device biases). So, even in contexts where measurement
is destructive, it is possible to compare device biases. It’s worth
contemplating, however, the di¤erence between the present scenario and
the immediately preceding one (represented by panel 5).

The measurements y in panel 5 are less variable than are the


measurements y here in panel 8. That is evident in the standard
deviations shown on the …gures and follows from the fact that in the
present case (unlike the previous one) measurements are a¤ected by
unit-to-unit/measurand-to-measurand variation. So all else being equal,
one should expect limits (1) applied in the present context to be wider/less
informative than when applied to data collected as in the last application.
(ISU and Analytics Iowa LLC) IE 361 Module 7 9 / 10
Single Measurements From Two Linear Devices On
Multiple Measurands From a Stable Process

That is completely in accord with intuition. One should expect to be able


to learn more useful to comparing devices when the same item(s) can be
remeasured than when it (they) can not be remeasured.

Notice that if the F limits (2) are applied here, one winds up with only an
indirect comparison of σdevice1 and σdevice2 , since what can be easily
estimated is the ratio q
σ2x + σ2device1
q
σ2x + σ2device2

(ISU and Analytics Iowa LLC) IE 361 Module 7 10 / 10

You might also like