Analyze

Analyze
1
Test Hypotheses
Analyze List Vital Few Xs
www.effortsconsulting.com ‹#›
2
Learning Objectives
1. Learn how to analyze data to find root cause.
2. Understand how to build the null and alternative

hypotheses.
3. Determine if data are normal or non-normal.
4. Determine the proper test.
3
Types of Analysis
Individual Experience
Group Experience
Graphical Interpretation of Observed Data
Statistical Interpretation of Observed Data
Graphical Interpretation of Experimental Data
Statistical Interpretation of Experimental Data
4
Planning the Tests
For each theory (or group of theories), list all data required,
each tool to be used, the results that will support the theory,
and results that will rule out the theory.
Theory or group Tool to be Results that will Results that will
Data required
to be tested applied support theory rule out theory
5
Planning the Data Collection
FILE: ANALYZE-DATA COLLECTION PLAN.XLS
For each theory or group of theories identified, indicate the

data elements required and how they will be collected.
6
Stratification
7
Stratification
of Data
8
Learning Objectives
1. Define the concept of stratification.
2. Describe how to stratify data.
3. Interpret stratified data.
9
What Is Stratification?
Concept
Stratification is the separation of data. Its

most frequent use is during the Analyze
Phase to identify which categories (Xs)
contribute to the problem being solved (Y).
10
What Is Stratification?
Stratification answers the following TYPES of questions:
§ Is there a defect that occurs most frequently?

§ Which factors (Xs) contribute most to the project Y?
– Do the results differ across groups?
11
Example: Data on Invoice Errors
Invoice Number 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
Day of Week M Th T M T W T Th F F M W Th F T
Week of Month 1 4 3 4 4 4 1 2 4 3 4 1 4 3 4
Accountant A B A C B A B C A B C A B C A
Invoice Number 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Day of Week Th F F Th M T T T W T W Th F M M
Week of Month 4 2 4 4 4 4 1 3 4 4 4 4 3 4 1
Accountant A B A C C D D D D A B C D A A
Invoice Number 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
Day of Week M W W T F Th Th M W F Th T T W F
Week of Month 4 3 2 4 2 3 4 3 2 4 4 2 4 3 2
Accountant A C D D A C C D A C D B A B C
Invoice Number 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Day of Week F F Th W W W Th M T Th W F M F W
Week of Month 4 4 3 4 4 4 2 1 4 4 2 4 2 1 4
Accountant D D D C A D C D A C C C A D C
12
Errors on Invoices
20
Number of Errors
10
0
M T W Th F A B C D
Days of the Week Accountants
13
Errors by Week: Total and Percent in Error
Total Errors Percent in Error

40 4
30 3
20 2
10 1
0 0
1 2 3 4 1 2 3 4
Weeks Weeks
14
How to Stratify
1. Select the stratification variables

2. Establish categories
3. Sort observations into the categories
4. Calculate the phenomenon being measured for each
category
5. Display the results
6. Prepare and display the results for other stratification
variables
7. Plan for additional confirmation
15
How to Interpret Stratification
Basically, look across the categories of a variable to see

whether one or more of the categories stand out.
§ If results appear to indicate the likely source of the phenomenon being
studied, validate results by gaining further details on the underlying
cause.
§ If initial stratification does not yield useful results, there are two
possible courses of action.
– Conduct a two-stage stratification (stratify by a 2nd variable within
each category of a 1st variable).
– Stratify by other variables.
16
Pitfalls
Pitfalls and Traps

§ Concluding too much from data
§ Concluding the anomalous category is the cause

of the problem
§ When collecting new data, failing to collect

enough identifying information
17
Cascading Stratification
Pareto Chart Example
Frequency
Pareto A
Product Line
Pareto B
Location
Pareto C
Operator
18
Stratification—Histograms
What differences do you see between the output

from Supplier A and B?
Total
Frequency
Measure
Supplier A Supplier B
Frequency
Frequency
15
20
25
30
5
10
35
40
45
5
10
15
20
25
30
35
40
45
Measure Measure
19
Exercise: Interpretation
Agree or Disagree? Why?

1. If a team is solving a problem for which no data exist,
there is no point in making a list of variables or
characteristics to be used in stratifying data.
2. When a stratification does not indicate an obvious pattern,
the exercise has been a waste of time.
3. Stratification is a tool which applies well only to relatively
simple situations with a limited number of variables.
4. Stratification applies to situations dealing with methods
and procedures as well as with technology.
15 Minutes
20
Stratification – Summary
1. Stratification is the separation of data into categories.

2. Most frequent use is during the Analyze Phase to
identify which categories contribute to the problem being
solved.
3. Separate data into different categories. Data
observations in a given group or category share
common characteristics that define the category. This
sorting process is called stratification.
4. The characteristic used to separate the data is called a
stratification variable. Each stratification variable will
have two or more values.
5. Specific combinations, ranges or variables, define a
category. A category may be defined by more than one
variable.
6. Using a graphical tool to view the stratified data is
helpful.
21
Hypothesis Testing
22
Hypothesis
Testing Introduction
23
Learning Objectives
1. Introduce the terminology and methodology

required for Hypothesis Testing.
2. Use the Hypothesis Testing Roadmap as a guide.
3. Understand alpha and beta risk and their

association with Type I and Type II errors.
24
Hypothesis Testing Example
American Justice System Example

The American justice system can be used to illustrate the concept of
hypothesis testing.
In the American justice system innocence is assumed until guilt is
proven.
It requires strong evidence, “beyond a reasonable doubt,” to convict
the defendant. This corresponds to rejecting the null hypothesis and
accepting the alternative hypothesis.
Rejecting the null hypothesis, Ho, is a strong conclusion because it
requires enough evidence to find the defendant guilty. If there is
insufficient evidence, you fail to reject Ho and the verdict is not guilty.
Note that you never say innocent (you never accept Ho). Always say
you do not reject, or fail to reject Ho.
Null Hypothesis ( Ho ): Defendant is not guilty

Alternative Hypothesis ( Ha ): Defendant is guilty
25
Hypothesis Testing Example
The only four possible outcomes are:

1. An innocent defendant is set free Correct decision
2. An innocent defendant is jailed Type I error
a. The probability of this type of error

occurring is a
3. A guilty defendant is set free Type II error
a. The probability of this type of error

occurring is b
4. A guilty defendant is jailed Correct decision
26
Alpha and Beta Risk
Truth
Ho Ha
Note: Type II error
Type II is also called
Fail to Correct Error Consumers’ Risk
Decision
Reject Ho b
Decision
Type I
Error Correct
Reject Ho Decision
a
Note: Type I error
is also called
Producers’ Risk
27
Process Example
FILE: ANALYZE-HYPOTHESIS TESTING PROCESS 2.MTW
Process A Process B
Yield Yield
89.7 84.7
81.4 86.1
84.5 83.2
84.8 91.9
87.3 86.3
79.7 79.3
85.1 82.6
81.7 89.1
83.7 83.7
84.5 88.5
“Is there a real difference between and ?”

“Is there a real difference between the average yield of Process A and
Process B?”
28
Process Example
§ Practical Question: Did the modifications on Process B improve

yield when compared to the current process represented by
Process A?
– Stated another way using mean values, “Has the mean value
of yield improved?”
Descriptive Statistics
Variable Process Total Count Mean StDev
Process A Yield A 10 84.240 2.902
Process B Yield B 10 85.540 3.650
§ Graphical representation of the data: Does the observed

difference in the mean between Process A and Process B occur
by chance?
§ Statistical Question: Is the mean for Process B (85.54) different
enough from the mean for Process A (84.24) to be considered
statistically significant?
– Are the means close enough to have occurred just by chance?
29
Process Example
Process A Process B
Statistical Concept:
In actuality, do the yield
measurements from the
processes represent two
different populations? B B B B B BB BB B
A AA AAAA A
A
80.0 82.5 85.0 87.5 90.0 92.5
Or, do the yield

measurements from the
processes come from . . .. . . : ::. .. . . . . . .
one population?
----+---------+---------+---------+---------+---------+-
80.0 82.5 85.0 87.5 90.0 92.5
30
Concept: Formulating the “Null” and “Alternative” Hypotheses
Null Hypothesis (Ho) Alternative Hypothesis (Ha)

the default proposition what you want to prove
you hope to reject
§ Statistical Interpretation:
There is no difference in H o : µ a = µb § Statistical Interpretation: The
population means for Process
H a : µ a ¹ µb
population means for A and Process B are different.
Processes A and B.
§ Practical Interpretation: The
§ Practical Interpretation: average yield of Process B is
There is no difference in different than that of
the average yields of the Process A.
two processes, i.e., your
changes did not help.
Goal: You must show that the values you observed were so
unlikely to come from the same population that Ho must be wrong!
You want to reject Ho, thus defaulting to Ha.
31
Definitions
§ Null Hypothesis (Ho)

§ Alternative Hypothesis (Ha)
§ Type I Error
§ Alpha Risk
§ Significance Level
§ Type II Error
§ Beta Risk
§ Significant Difference
§ P-Value
§ Power
§ Test Statistic
32
Hypothesis Testing: P-Value
P-Value is the probability of being wrong if you say

there is a difference.
P-Value £ α: There is a
statistical difference
33
Steps in Hypothesis Testing
1. State the practical problem. 7. Use the appropriate

2. Establish the Hypotheses. graphical tool to explore
3. Decide on appropriate the data.
statistical test. 8. Check data assumptions
4. Set the alpha level. (change choice of tool if
appropriate).
5. Determine initial sample
size. Calculate the power. 9. Run the statistical test and
determine the statistical
6. Develop the sampling plan conclusion.
and collect the data.
10. Translate statistical
conclusion to a practical
conclusion.
Practical Statistical Statistical Practical
Problem Problem Conclusion Conclusion
34
Summary
1. Hypothesis testing follows specific steps, and requires

formulation of both a null and alternative hypothesis.
2. Hypothesis tests have four possible outcomes.
3. Type I error is the error in rejecting Ho when it is in fact true;
you set the maximum risk as α.
4. Type II error is the error in failing to reject Ho when it is in
fact false; the risk of making a Type II error is β risk.
5. P-Value is the probability of being wrong if you reject Ho; if
P-Value ≤ α, there is a statistically significant difference.
35
Confidence Intervals
36
Confidence
Intervals
37
Learning Objectives
1. Determine the range of values that includes the

true value of the estimated population parameter.
38
Confidence Intervals Agenda
§ Definition of a Confidence Interval (CI)

– Practical
– Statistical
§ Why do you need to study CIs?

§ The Confidence Interval Equation
§ CI for the Population Mean (t-Distribution)
§ Example
39
Definition of the Confidence Interval
Practical
A range determined by sample data, which tells you where the true
population parameter is located with a certain degree of confidence.
Statistical
A (1- α) % confidence interval for a population parameter, μ or σ, is
a random interval that is located between upper and lower
confidence limits (CL's):
§ Probability [ Lower CL ≤ μ ≤ Upper CL ] = 1 - α
§ Probability [ Lower CL ≤ σ ≤ Upper CL ] = 1 - α
regardless of the value of the parameter.
What Is Alpha (α)?

The maximum risk or probability that the actual population
parameter is located outside the confidence interval (also known as
a Type I Error). This probability is always greater than zero, and is
frequently established at 5%.
40
Why Study CIs?
§ Sample statistics, such as the sample mean and the sample standard
deviation, are only estimates of the true population parameters, μ and s.
§ Because of inherent sample to sample variability in these estimates,
uncertainty is quantified using statistically based Confidence Intervals
(CI).
§ The automotive industry calculates 95% CI for their data.
§ Some medical applications calculate 99.9% CI for their data.
§ Confidence Intervals can be interpreted as follows:
– Approximately 95 samples out of 100 will yield a CI that contains the
true population parameter, or
– Approximately 5 samples out of 100 will yield a CI that does not
contain the true population parameter
– 95% certainty that the true population parameter is inside the interval
§ The CI provides a way to investigate sample to sample variation.
§ Use CIs to obtain statistical confidence for the population mean, standard
deviation, and Cp parameters.
41
The Confidence Interval Equation
§ Confidence Intervals take the general form:

C.I. = Statistic ± [K * (Sample Std Dev)]
Statistic = Sample Mean, or
= Sample Variance, or
= Sample Cp, etc...
K = Constant based on a statistical distribution
§ For estimating µ, K is selected from a family of “t” distributions.

§ For estimating s, K is selected from a family of Chi-Squared
distributions.
42
Confidence Interval for the Population Mean
When process performance is estimated, it is from a relatively small sample
of the process. To calculate, use estimates of the mean and standard
deviation taken from a sample. Both estimates have inherent variation. The
goal is to quantify certainty of where the true process is performing.
First, look at the confidence interval for the population mean.
Ø Example: Suppose you want to determine the 95% Confidence Interval
for the population mean from 10 samples (n = 10) off a reactor. Sample
the reactor and obtain the following sample statistics:
§ Sample Mean = x = 249.6
§ Sample Standard Deviation = s =14.15
With this sample information, use the general formula for Confidence
Intervals to estimate the population mean:
s s
x - t n -1,1-a /2 £ µ £ x + t n -1,1-a /2
n n
Where
§ x = sample mean
§ s = sample standard deviation
§ n = sample size
§ t n - 1, 1- a/2 = t value for n-1 Degrees of freedom and probability 1 - a / 2
43
What Is the t Distribution?
§ The t Distribution is a family of bell-shaped (normal-like)

distributions whose width is sample size dependent.
§ The smaller the sample size, the wider and flatter the
distribution.
§ To give an idea of the values of “t” for 95% CIs for different
sample sizes, you can look at the table below:
Z Distribution.
(approx. n > 30) Sample t value
5 2.78
10 2.26
t Distribution 20 2.09
(n=10)
30 2.05
t Distribution
(n=5) 100 1.98
1000 1.96
44
t Distribution vs. Normal Distribution
§ Use t-Table instead of Z-Table to compute the area in the

tails when sample size is less than or equal to 30.
§ Note: There is more area under the tail of the “t” as
compared to the “Z” distribution.
§ If s is unknown, s is the best estimate.
Normal
Distribution
( n > 30 )
t Distribution
( n ≤ 30 )
t
Z
X -µ if n > 30 X -µ
t n -1 = Z=
s/ n if n ≤ 30 s/ n
45
Back to the Example
Ø Example: Suppose you want to determine the 95% Confidence Interval
for the population mean from 10 samples (n = 10) off a reactor.
Sample the reactor and obtain the following sample statistics:
§ Sample Mean = X = 249.56
§ Sample Standard Deviation = S =14.15
s s
x - t n -1,1-a / 2 £ µ £ x + t n -1,1-a / 2
n n
14.15 14.15
249.56 - 2.262 * £ µ £ 249.56 + 2.262 *
10 10
t n-1, 1-a/2:
df (Degrees of Freedom) =n - 1 =9
1-a/2=1-.05/2=.975
t9, .975 =2.262
249.56 - 10.11 £ µ £ 249.56 + 10.11
239.45 £ µ £ 259.67
§ Solution: You are 95% confident that the actual process mean is
somewhere between 239.45 and 259.67.
§ Notice that the t value is very close to 2.00.
§ Do the problem in MINITAB®.
46
Example: Calculating CI for Population Mean
FILE: ANALYZE-PROCESS 1-T.MTW
Using MINITAB®
Confidence Intervals
Variable N Mean StDev SE Mean 95.0 % CI
Process 10 249.56 14.15 4.47 ( 239.44, 259.68 )
This result states with 95% certainty the actual process

mean is between 239.43 and 259.69. There is also a 5%
chance that you’re wrong.
47
Testing Hypothesis
48
Testing Hypotheses
When Y Is Categorical and
X Is Categorical
49
Learning Objectives
1. Compare a proportion to a target.
2. Compare 2 proportions.
3. Power and Sample Size.
4. Is Output Y dependent on Input X?
50
Test This Theory: Is it a “Fair” Coin?
§ If you flip a coin 10 times, and get 8 heads, is it a fair coin?

§ You want to be 95% confident that you are correct
(Confidence level: 95.0).
§ You want to test if the coin is fair, that is, is there a 50%
chance that you will get heads?
§ The Null Hypothesis is Ho: Proportion of heads = 0.5.
§ You have 3 alternatives: less than, not equal, or greater than.
The theory you are testing: Is the coin fair or not? Therefore,
choose “not equal.”
§ The Alternative Hypothesis is Ha: Proportion of heads ≠ 0.5.
§ Conduct a 1 Proportion test.
51
Is it a “Fair” Coin?
A “fair” coin will flip heads 50% of the time. In other words,
the proportion is 0.5.
§ Look at the 95% confidence interval. You can say that
based on 8 heads out of 10 flips, you are 95% confident
that the proportion of heads you can expect to get from
this coin lies between 0.44 to 0.97.
§ Why do you have such a large interval?
§ Since this interval (0.44 to 0.97) includes 0.5, it is
concluded that you do not have sufficient evidence to
reject Ho.
§ Before you conclude, you should evaluate the power of
this test. The power of this test is only 47% (0.4688),
which means that the ability of this test to detect a
difference when there is one is only 47% (the probability
of being correct in rejecting Ho is 0.4688). This test does
not have enough power due to the small sample size.
52
Is It a “Fair” Coin?
§ What if you have 100 coin flips and

80 result in heads?
§ Look at the 95% confidence interval.
§ Based on 80 heads out of 100 flips,
you are 95% confident that the
proportion of heads expected from
this coin lays between 0.71 to 0.87.
§ Since this interval (0.71 to 0.87) does
NOT include 0.5, there is sufficient
evidence to reject Ho and conclude
that the coin is NOT a fair coin.
§ Question: Why did you get a different
conclusion?
53
Is It a “Fair” Coin?
Compare:
Notice that as your sample size increased (from 10 flips to
100 flips), your 95% Confidence Interval decreased. In
other words, you are now more confident of your results
because you have more data.
Note: Rejecting Ho is a much stronger conclusion than

failing to reject Ho. You fail to reject Ho because you have
insufficient evidence to do so.
54
What You Have Learned
§ You have just learned how to compare a

proportion to a target value.
§ You will learn how to compare two proportions

next.
55
Are Office A and Office B Different?
§ Office A produced 16,000 items and 223 were found to

be defective. Office B produced 14,000 items and 150
were found to be defective.
§ Quality performance is measured as defective rate

expressed as a proportion.
§ Is there a statistical difference in the quality performance

of these two offices?
Ho: PA = PB or PA – PB = 0 (test difference of zero)
Ha: PA ¹ PB or PA- PB ¹ 0
§ Conduct a 2 Proportion Test.
56
Office A vs. Office B
§ You want to be 95% confident of your results. Since you

are testing for a difference, Ho is that the defect rates are
equal, and expect the difference to be 0.0.
§ Set test difference: 0.0 and Alternative: “not equal.”
§ Look at the 95% CI for p(1) – p(2). The interval is
0.00073 to 0.0057.
§ This interval does NOT include 0.0.
§ The P-Value = 0.012 means that there is a 1.2% chance
of being wrong in rejecting Ho.
§ You can conclude that there IS a difference between
office A and office B.
§ Try testing A<B. Why would you do that? Interpret the
results.
57
Exercise One: Test of Proportions
TEST OF PROPORTIONS
Repair rate for television sets
Problem:
The quality assurance department for the ABC television company
wants to estimate the proportion of their 35-inch television sets that
needed to be repaired within four years of purchase. The department is
particularly interested in whether the need for repairs for their brand of
television is different than all other brands currently on the market. It is
known that the overall repair rate for all other televisions currently on
the market is 6.8% or 0.068.
Market Research followed up on a sample of 2,856 35-inch televisions

purchased from ABC during a particular time period. Of the 2,856 sets,
a total of 236 needed repair within four years of purchase.
Determine whether or not the television has the same repair rate as the
average of all brands of televisions currently on the market.
15 Minutes
58
Exercise Two: Power and Sample Size
POWER AND SAMPLE SIZE
Repair rate for television sets
Problem:
For exercise one, the repair rate for 35-inch televisions from the ABC
television company was significantly different than the repair rate of
similar televisions from other manufacturers. You tested the following
hypotheses:
Ho: p=0.068
Ha: p ≠0.068 or p > 0.068
If you repeat this study many times, sampling 2,856 sets each time,
how often will the test show ABC to be different. . .
IF the true rate for ABC sets is in fact 0.068? 15 Minutes
IF the true rate for ABC sets is in fact 0.083 (as measured)?
59
What You Have Learned
§ 1-Proportion Test
§ 2-Proportions Test
§ Power and Sample Size
§ Next: Is Y independent of X?
– Chi-Square Contingency Tables Test
60
Contingency Table Description
§ A contingency table is used to analyze data via a two-way

classification. In a table, this can be described as a row
factor and as a column factor.
§ It tests the relationship between two variables/factors.
Ho: “{A} is independent of {B}”
Ha: “{A} is NOT independent of {B}”
§ A test statistic is calculated. This statistic uses a c2 (Chi-

Squared) distribution to test hypotheses about the
frequency of occurrence of some event.
§ Given a and the degrees of freedom associated with each
factor, a critical value is determined.
§ If the test statistic is larger than the critical value, Ho
is rejected. This will easily be seen by comparing the
P-Value to alpha. If P-Value is ≤ a then the probability
of committing a Type I error is low and the conclusion
is to reject the null hypothesis.
61
Chi-Square Test for Independence
Use frequencies to determine if samples are independent

(from the same population).
Invoices Invoices
(no errors) (with errors)
Actual/observed 1st shift fo=20 fo=50 Total= T-row =70
frequencies
2nd shift fo=40 fo=70 Total=T-row=110
Total Total N= Grand Total=180

= T-col= 60 = T-col=120
{% of {% of
“Expected” T-col T-row T-col X T-row
f e = total in X total in X {Total} = X XN =
N
frequency column} row} N N
Invoices Invoices
(no errors) (with errors)
1st shift fe=23.33 fe=46.67
2nd shift fe=36.67 fe=73.33
62
Χ2 Critical Table
Test for Independence:

DF=(R-1)(C-1)
R=rows; C=columns
ά value = 0.05
df .250 .100 .050 .025 .010 .005 .001

1 1.323 2.706 3.841 5.024 6.635 7.879 10.828
2 2.773 4.605 5.991 7.378 9.210 10.597 13.816
3 4.108 6.251 7.815 9.348 11.345 12.838 16.266
4 5.385 7.779 9.488 11.143 13.277 14.860 18.467
5 6.626 9.236 11.070 12.832 15.086 16.750 20.515
6 7.841 10.645 12.592 14.449 16.812 18.548 22.458
7 9.037 12.017 14.067 16.013 18.475 20.278 24.322
8 10.219 13.362 15.507 17.535 20.090 21.955 26.125
9 11.389 14.684 16.919 19.023 21.666 23.589 27.877
10 12.549 15.987 18.307 20.483 23.209 25.188 29.588
11 13.701 17.275 19.675 21.920 24.725 26.757 31.264
12 14.845 18.549 21.026 23.337 26.217 28.300 32.909
13 15.984 19.812 22.362 24.736 27.688 29.819 34.528
Degrees of freedom
14 17.117 21.064 23.685 26.119 29.141 31.319 36.123

15 18.245 22.307 24.996 27.488 30.578 32.801 37.697
16 19.369 23.542 26.296 28.845 32.000 34.267 39.252
17 20.489 24.769 27.587 30.191 33.409 35.718 40.790
18 21.605 25.989 28.869 31.526 34.805 37.156 43.312
19 22.718 27.204 30.144 32.852 36.191 38.582 43.820
20 23.828 28.412 31.410 34.170 37.566 39.997 45.315
21 24.935 29.615 32.671 35.479 38.932 41.401 46.797
22 26.039 30.813 33.924 36.781 40.289 42.796 48.268
23 27.141 32.007 35.172 38.076 41.638 44.181 49.728
24 28.241 33.196 36.415 39.364 42.980 45.558 51.179
25 29.339 34.382 37.652 40.646 44.314 46.928 52.620
26 30.434 35.563 38.885 41.923 45.642 48.290 54.052
27 31.528 36.741 40.113 43.194 46.963 49.645 55.476
28 32.620 37.916 41.337 44.461 48.278 50.993 56.892
29 33.711 39.087 42.557 45.722 49.588 52.336 58.302
30 34.800 40.256 43.773 46.979 50.892 53.672 59.703
40 45.616 51.805 55.758 59.342 63.691 66.766 73.402
50 56.334 63.167 67.505 71.420 76.154 79.490 86.661
60 66.981 74.397 79.082 83.298 88.379 91.952 99.607
70 77.577 85.527 90.531 95.023 100.425 104.215 112.317
80 88.130 96.578 101.879 106.629 112.329 116.321 124.839
90 98.650 107.565 113.145 118.136 124.116 128.299 137.208
100 109.141 118.498 124.342 129.561 135.807 140.169 149.449
63
Chi-Square Test for Independence
FILE: ANALYZE-CHI SQUARE—INVOICE EXAMPLE.MTW
Pass/Fail by Shift Calculations
(20-23.33)2 (50-46.67)2
23.33 +
g
(fo-fe)2 46.67
c2calc = S
j=1 fe
=
(40-36.67)2 (70-73.33)2
= 1.169
36.67 + 73.33
Since
c2calc < c2crit
c2critical = 3.841 Fail to reject H0. Quality is
independent of shift.
Also, P-Value > α
64
Steps in Chi-Squared Hypothesis Testing
1. State the practical problem. 7. Use the appropriate graphical

2. Establish the hypotheses. tool to explore the data.
a. Ho: There is no difference 8. Check data assumptions
b. Ha: There is a difference (change choice of tool if
appropriate).
3. Decide on appropriate statistical
test. 9. Run the statistical test and
4. Set the alpha level. conclusion.
5. Determine initial sample size. a. If p alpha, reject Ho*
6. Develop the sampling plan and b. Otherwise, fail to reject Ho
collect the data.
10. Translate statistical conclusion
to a practical conclusion.
* If test is performed manually, reject if c2 (calc) > c2 (critical)

65
Chi-Square Test: Exercise
FILE: ANALYZE-CHI SQUARE—MOLDING.MTW
A team seeking to reduce the defects in a molded plastic product

wants to test the theory that the defects are related to the cavity
within the mold that the specific item came from. They are
concerned with three types of defects: Short (insufficient material),
flash (extra material adhering to the item), and deformed
(misshapen). Can they accept this theory?
Good Short Flash Deform

Cavity 001 180 3 10 7
Cavity 002 183 13 1 3
Cavity 003 182 4 8 6
Cavity 004 182 3 6 9
Cavity 005 180 4 6 10
Cavity 006 184 4 5 7 25 Minutes
Are defect types independent of the cavity from which the

item came?
66
Summary
1. Hypothesis tests of categorical Y and X data include:

a. Comparing a proportion to a target (MINITAB®: 1
Proportion)
b. Comparing 2 proportions (MINITAB®: 2
Proportions)
c. Independence of Y and X (MINITAB®: Chi-Square
Test [Two-Way Table in worksheet])
2. Increasing sample size increases statistical power,
allowing smaller differences to be detected.
67
Normality Test
68
Normality Tests
69
Learning Objectives
1. Determine when to use a Normality test.
2. Use the proper steps in performing a Normality

test.
3. Correctly interpret a Normality test.
70
Normality Test Method
FILE: ANALYZE-NORMALITY TESTS.MTW
Step 1: State the null and alternative hypotheses.

a. Ho : the data are Normally distributed
b. Ha: the data are not Normally distributed
Step 2: Use the appropriate graphical tool to
evaluate the data; if sample sizes are small,
it will often be difficult to graphically evaluate
normality.
Step 3: In MINITAB®: Stat > Basic Statistics >
Normality Test
Step 4: Enter your data in the column(s).
Step 5: Select the columns containing the data.
Step 6: Choose the Anderson-Darling Normality test.
Step 7: Run the test.
71
Normality Test Example
For this discussion, you will use some data called

Population 1.
Population 1
85.3
97.1
67.9
77.6
99.3
84.0
82.4
86.1
88.8
72
Normality Test Example
Here is an output of a Normality test called a Probability Plot.
Note the P-Value in the summary to the right of the plot. Because the P-Value
(0.640) is greater than the alpha level of 0.05, you fail to reject the null
hypothesis. In other words, the data may be treated as Normal. The points fall
reasonably close to the reference line indicating that the data follow a normal
distribution.
73
Alternative MINITAB® for Steps 3, 4, and 5
Summary for Population 1

A nderson-D arling N ormality Test
A -S quared 0.25
P -V alue 0.640
M ean 85.389
S tDev 9.488
V ariance 90.026
S kew ness -0.284629
Kurtosis 0.508642
N 9
M inimum 67.900
1st Q uartile 80.000
M edian 85.300
3rd Q uartile 92.950
70 75 80 85 90 95 100 M aximum 99.300
95% C onfidence Interv al for M ean
78.096 92.682
95% C onfidence Interv al for M edian
78.694 95.208
95% C onfidence Interv al for S tDev
9 5 % C onfidence Inter vals
6.409 18.177
Mean
Median
80 84 88 92 96
74
Exercise
§ Using the data supplied, determine if each sample is

normally distributed.
§ Can you predict the normality of the distribution from just
looking at the numbers?
Steps for Normality Test

MINITAB®: Stat > Basic Statistics > Normality Test
§ Variable: (Samples in Column)
§ Test for Normality: Anderson-Darling 15 Minutes
75
Summary
1. Continuous data should be normally distributed

for ease of analysis.
2. Many tests are robust to the assumption of

normality, however, it is better to verify normality
by performing a Normality test.
76
Test for
Equal Variance
77
Learning Objectives
1. Understand why testing for equal variances is

important.
2. Learn how to test for equality of two or more

variances.
3. Be able to select the proper test statistic based

on the normality or non-normality of the data.
78
Comparing Two Variances
Is there a significant difference in the VARIABILITY of

Outputs?
BEFORE AFTER
CHANGE CHANGE
or A or B
Change or Difference in Variability

of Output
79
Equal Variances (MINITAB®)
§ If the normal probability plot or Anderson Darling Test indicates

that you are dealing with normally distributed data, then use
Bartlett’s (or F) Test Statistic presented by MINITAB®.
§ If the data are not normally distributed, then use Levene’s Test
Statistic presented by MINITAB®.
MINITAB®: Stat > ANOVA > Test for Equal Variances

(for 2 or more variances)
Ho: s2pop1= s2pop2

Ha: s2pop1 ¹ s2pop2
MINITAB®: Stat > Basic Statistics > 2 Variances

(for 2 variances only)
80
Example
FILE: ANALYZE-PROCESS 2.MTW
Practical Problem: Suppose the

process is modified. You want to
see if you have “significantly”
improved the yield of the process
(measured in kilograms) with
these modifications before you
spend significant time and
resources modifying the rest. After
sampling the yield, how do you
determine if there is a “real”
difference between the two yields?
81
Example: Comparing 2 Variances
FILE: ANALYZE-PROCESS 2.MTW
Ho: s2processA= s2processB
Ha: s2processA ¹ s2processB
Fail to reject the

null; variances are
assumed equal.
82
Summary
1. Testing for equal variances may be done:

a. To compare variation in the output of two or more
processes
b. To test the assumption of equal variances before
proceeding with additional hypothesis tests (e.g.,
ANOVA)
2. When comparing variances, select the proper test statistic:
a. Normal data: F-test (2 variances), Bartlett’s (2 or more
variances)
b. Non-normal data: Levene’s (2 or more variances)
83
Testing Hypothesis
84
Testing Hypotheses
When Y Is Continuous
and X Is Categorical
85
Learning Objectives
1. Compare the mean of a population to a target.
2. Compare the mean of two populations.
3. Perform 1-Sample and 2-Sample t-Tests.
86
1-Sample t-Test
§ In a 1-sample t-Test you are asking the practical question:

How does my population’s mean compare to a target
value?
– Ho: population mean = target value
– Ha: population mean not equal to target value (or less than or
greater than target value depending on the nature of the problem)
§ The statistical approach to answering this question is:
– What is the probability that the difference between the mean and
target value would occur by chance?
– Does the confidence interval for the population mean include the
target value (or null hypothesis)?
§ In order to run a t-Test, you need to determine whether
the data are normally distributed. If the data are not
Normal, you should not use the t-Test. Other tests would
be more reliable.
87
Steps in Hypothesis Testing
1. State the practical 7. Use the appropriate

problem. graphical tool to explore
2. Establish the hypotheses. the data.
3. Decide on appropriate 8. Check data assumptions
statistical test. (change choice of tool if
4. Set the alpha level. appropriate).
5. Determine initial sample a. Normality
6. Develop the sampling determine the statistical
plan and collect the data. conclusion.
conclusion.
88
Example: 1-Sample t-Tests
FILE: ANALYZE-1-SAMPLE t POPULATIONS.MTW
Using the data from example “Population 1,”

does the confidence interval for the
population mean include the target value
85, i.e., null Hypothesis?
Ho: population mean = 85
Ha: population mean ≠ 85
Individual Value Plot of Population 1
(with Ho and 95% t-confidence interval for the mean)
_
X
Ho
70 75 80 85 90 95 100
Population 1
89
Example: 1-Sample t-Tests (continued)
Using the data from example “Population 1,” does the

confidence interval for the population mean include the
target value 75, i.e., null Hypothesis?
Ha: population mean ≠ 75
_
X
Ho
70 75 80 85 90 95 100
Population 1
90
1-Sample t-Test
You also can ask, is the population mean greater than the
target value? This creates a one-sided test.
Ha: population mean > 75
_
X
Ho
70 75 80 85 90 95 100
Population 1
91
1-Sample t-Test
Or, is the population mean less than the target value?

Ha: population mean < 75

_
X
Ho
70 75 80 85 90 95 100
Population 1
92
Exercise: 1-Sample t-Tests
FILE: ANALYZE-1-SAMPLE t POPULATIONS.MTW
Determine if the population mean for Population 2 and 3 is greater

than, equal to, or less than 75, using a 95% confidence level.
Population 1 Population 2 Population 3

85.3 72.3 76.1
97.1 97.3 86.4
67.9 91.3 75.8
77.6 72.7 88.7
99.3 86.7 77.8
84.0 102.2
82.4 85.1
86.1 73.5
88.8 74.4
96.6 25 Minutes
90.2
84.4
84.9
81.0
87.5
93
What You Have Learned: 1-Sample t-Tests
Step 1: State the practical problem and objectives.

Step 2: State hypotheses:
a. Ho: There is no difference
b. Ha: There is a difference—Direction of difference
Step 3: Determine the statistical test to use.
Step 4: Set alpha level (usually 0.05).
Step 5: Determine sample size and calculate the power (1-β).
Step 6: Develop data-collection plan and collect data.
Step 7: Use graphical tools to explore the data.
Step 8: Test for normality (Anderson-Darling test).
Step 9: Run test and state statistical conclusion:
a. If p ≤ alpha, reject Ho
b. Otherwise, fail to reject Ho
Step 10: Translate statistical conclusion into practical conclusion.
94
Comparing 2 Means
Is there a significant difference in the Output means?
Difference in
Output Means
BEFORE CHANGE AFTER CHANGE

Or Process A Or Process B
95
Why Test Population Parameters Against Each
Other?
You never know the true population

parameters. Thus, use statistics Sample 1
µ1
from samples to determine how
σ1
likely it is that these two samples µ2
came from the same population. Sample 2
σ2
These comparisons allow you to

compare the performance of two
OR
process means and standard
deviations. Sample 1
µ Sample 2
Then you can conclude, with some
σ
degree of statistical confidence, that
these parameters came from one
population or two.
96
2-Sample t-Tests: Example
FILE: ANALYZE-2 SAMPLE T.MTW
Do these two populations have the same

mean?
Population A Population B
81.29 75.6
86.56 73.2
86.20 71.1
85.33 75.3
80.95 70.3
87.76 75.0
88.45 79.1
70.0
80.1
74.5
82.2
97
Steps in 2-Sample t-Tests
1. State the practical problem. 7. Use the appropriate graphical tool

2. Establish the hypotheses: to explore the data.
a. Ho: There is no difference 8. Check data assumptions (change
b. Ha: There is a difference— choice of tool if appropriate).
Direction of difference a.Normality
3. Decide on appropriate statistical b.Equal Variances (check box
test. appropriately)
4. Set the alpha level (usually 0.05). 9. Run the statistical test and
5. Determine initial sample size. determine the statistical
Calculate the power. conclusion. If P-value ≤ a reject
Ho .
6. Develop the sampling plan and
collect the data. 10. Translate statistical conclusion to
a practical conclusion.
98
Population A Population B
81.29 75.6
86.56 73.2
86.20 71.1
85.33 75.3
80.95 70.3
87.76 75.0
88.45 79.1
70.0
80.1
74.5
82.2
Are these two groups from the same population or from

different populations? That is, do they have the same
population mean?
99
Graphical Analysis Options

Dotplot of Population A, Population B
Population A
Population B
70.0 72.5 75.0 77.5 80.0 82.5 85.0 87.5
Data
www.effortsconsulting.com 100
‹#›
Test for Normality

Probability Plot of Population B
Normal
99
Mean 75.13
StDev 4.009
95 N 11
AD 0.290
Mean 75.13 90
P-Value 0.542
StDev 3.985 80
70
N 11
Percent
60
50
AD 0.292 40
30
P-Value 0.542 20
10
1
65 70 75 80 85
Population B
Probability Plot of Population A
Normal
99
Mean 85.22
StDev 2.980
95 N 7
AD 0.452
90
P-Value 0.185
80
70
Percent
60
50
40 Mean 85.22
30
20
StDev 2.980
10 N 7
5 AD 0.452
1
P-Value 0.185
78 80 82 84 86 88 90 92 94
Population A
‹#›
Analytic Analysis (separate variances)

Two-Sample t-Test and CI: Population A, Population B
Two-sample t for Population A vs. Population B

N Mean StDev SE Mean
Population A 7 85.22 2.98 1.1
Population B 11 75.13 4.01 1.2
Difference = mu (Population A) – mu (Population B)

Estimate for difference: 10.09
95% CI for difference: (6.57; 13.62)
t-Test of difference = 0 (vs. not =):
t-Value = 6.11 P-Value = 0.000 DF = 15
Since the P-Value is lower than alpha, the null must go!
‹#›
Exercise: 2-Sample t-Tests
FILE: ANALYZE-2 SAMPLE T EXERCISE.MTW
1. Consider the following Before and After yield

data. At the 95% confidence level, has there
been a statistically significant increase in yield?
Before: (95.8, 94.7, 94.7, 94.6, 95.1, 94.9, 94.9)
After: (98.0, 97.3, 96.8, 98.3, 97.4, 95.6, 95.5)
2. Conduct appropriate graphical analysis.*
3. Using the data, now answer the question: At the
95% confidence level, are the average yields
the same?
15 Minutes
*Black Belts use at least three.
‹#›
What You Have Learned: 2-Sample t-Test
1. State the practical problem. 7. Use the appropriate graphical tool

2. Establish the hypotheses. to explore the data.
a. Ho: There is no difference 8. Check data assumptions (change
b. Ha: There is a difference— choice of tool if appropriate).
Direction of difference a. Normality
3. Decide on appropriate statistical b. Equal Variances (check box
test. appropriately)
4. Set the alpha level (usually 0.05). 9. Run the statistical test and
5. Determine initial sample size. determine the statistical
Calculate the power. conclusion. If P-value ≤ a reject
6. Develop the sampling plan and Ho .
collect the data. 10. Translate statistical conclusion to
a practical conclusion.
‹#›
Analysis of Variance
105
(ANOVA)
‹#›
Learning Objectives
1. When to use a one-way ANOVA.
2. Describe why you will use ANOVA.
3. Practice running an ANOVA test.
‹#›
Analysis of Variance Highlights
§ ANOVA tests the equality of population means by analyzing variances

between groups and within groups (F-test).
§ The OUTPUT or RESPONSE variable is continuous (yield,

temperature, volts, % Impurities, cycle time, etc.).
§ You want to know whether specific discrete INPUTS or FACTORS have

a significant effect on the response.
§ A factor has two or more LEVELS. Levels are discrete, so a continuous

treatment such as line pressure must be classified into discrete
intervals like low and high.
§ A ONE-WAY ANOVA is used when there is only one treatment variable,

but it can have many levels.
Is there a real difference between m1 & m2 & m3 & . . . ,
where 1,2,3 . . . are different treatments and levels?
‹#›
ANOVA Method
Step 1: State the practical problem.

Step 3: Select appropriate statistical test.
Step 4: Set alpha level.
Step 5: Determine sample size.
Step 6: Develop sampling plan and collect data.
Step 7: Use the appropriate graphical tool to evaluate the data.
Step 8: Do the assumptions for the model hold?
a. Response means independent and response variable
normally distributed within each factor level.
b. Population variances are equal across all levels of the factor.
Step 9: Construct ANOVA table and interpret statistical conclusion.
a. Errors independent and normally distributed about mean zero.
b. Interpret the P-Value (or the F-statistic) for the factor effect.
Step 10: Translate statistical conclusion into practical conclusion.
‹#›
Example Setup
FILE: ANALYZE-ANOVA DIET DATA.MTW
§ Twenty-four animals receive

DIET A DIET B DIET C DIET D
one of four diets.
62 63 68 56
§ The type of diet is the factor
of interest. 60 67 66 62
§ Blood coagulation time is 63 71 71 60
the response (output). 59 64 67 61
§ During the experiment, 65 68 63
diets were assigned
66 68 64
randomly to animals. Blood
samples were taken and 63
tested in random order. 59
Why?
* Adapted from Box, Hunter & Hunter, Statistics for Experimenters, Wiley
‹#›
Example: Steps 1 and 2
Step 1: State the practical problem.

a. We, as diet researchers, would like to know if diet can
affect blood coagulation time.
Step 2: State the null and alternate hypotheses
Ho: m diet1 = m diet2 = m diet3 = m diet4 (or) Ho : t’s =0
Ha: at least one diet is different (or) Ha : t’s ≠0

Treatment (diet) effect = t
§ Interpretation of the null hypothesis: The average blood

coagulation time of each diet is the same (or what the animal eats
does NOT affect blood coagulation time).
§ Interpretation of the alternative hypothesis: At least one diet

affects the average blood coagulation time differently than another
(or what type of diet the animal eats does affect blood coagulation
time).
‹#›
Example: Steps 3-6 (ANOVA Analysis)
Step 3: Select appropriate test.

a. ANOVA has more power and information than simple t-
Tests.
Step 4: Set alpha level.
a. Usually 5% (0.05).
Step 5: Determine sample sizes.
a. As more factors and levels of factors are added to the
test, the sample size needs to increase to support
them.
Step 6: Develop sampling plan and collect data.
a. Ensure that treatments and treatment levels are
assigned randomly.
b. Balance the design among treatments and levels.
‹#›
Example: Step 7
Step 7: Use the appropriate graphical tool to evaluate

the data.
a. It’s good practice to ALWAYS plot the data.
Box pl ot of Di et A, Di et B, Di et C, Di et D
72.5
70.0
67.5
65.0
Dat a
62.5
60.0
57.5
55.0
Diet A Diet B Diet C Diet D
‹#›
Example: Step 8
Step 8: Do the assumptions for the model hold?

a. Response means independent and response variable normally distributed
within each factor level.
b. Population variances are equal across all levels of the factor.
Ho:
Ha : Ho:
Ha :
Probability Plot of Diet A
Normal
99 Probability Plot of Diet B
Mean 61
Normal StDev 1.826
95 99 Probability Plot of Diet C N 4
AD Mean
0.219 66
90 Normal Test for Equal Variances for CoagTi me
P-ValueStDev
0.608 2.828
80
95 99 Probability Plot of Diet D N 6
Mean 68
90 Normal AD 0.288 Bartlett's Test
70 P-ValueStDev
0.493 1.673
Diet A Test Statistic 1.67
Percent
60 95 99 N 6
80 Mean 61 P-Value 0.644
50 AD 0.492
70 90 StDev 2.619
40 P-Value 0.129 Levene's Test
Percent
60 95 N 8
30 80 Test Statistic 0.65
50 AD 0.279
20 70 90 P-Value 0.593
40 P-Value 0.543
Diet B
Percent
30 60 80
10
50
20 70
Diet
5 40
Percent
30 60
10 50
20
1 5 40 Diet C
56 57
10
58
30 59 60 61 62 63 64 65
20 Diet A
1 5
60 10 62 64 66 68 70 72 74
Diet B Diet D
1 5
64 65 66 67 68 69 70 71 72
Diet C
1
55.0 57.5 60.0 62.5 65.0 67.5 0 2 4 6 8 10 12
Diet D 95% Bonferroni Confidence Int ervals for St Devs
‹#›
Example: Step 9
Step 9: Construct and run ANOVA Table in MINITAB®.

One-way ANOVA: CoagTime versus Diets
SourceDF SS MS F P
Diets 3 228.00 76.00 13.57 0.000
Error 20 112.00 5.60
Total 23 340.00
S = 2.366 R-Sq = 67.06% R-Sq(adj) = 62.12%
Individual 95% CIs For Mean Based on Pooled StDev
Pooled StDev = 2.366

‹#›
What Are Fits and Residuals?
Row Diet Fits1 Coag Time Resi1

1 A 61 62 1
The Fit is just the 2 A 61 60 -1
Mean of each Diet 3 A 61 63 2
4 A 61 59 -2
5 B 66 63 -3
6 B 66 67 1
7 B 66 71 5
8 B 66 64 -2
9 B 66 65 -1
10 B 66 66 0
11 C 68 68 0
12 C 68 66 -2
13 C 68 71 3
14 C 68 67 -1
15 C 68 68 0
The residual is 16 C 68 68 0
(Raw Score-Mean) 17 D 61 56 -5
18 D 61 62 1
19 D 61 60 -1
20 D 61 61 0
21 D 61 63 2
22 D 61 64 3
23 D 61 63 2
24 D 61 59 -2
‹#›
Example: Step 9a
Step 9a: Do the assumptions for the errors hold (residual

analysis)?
a. Errors of the model are independent and normally distributed.
Are the
residuals This graph investigates
Residual Plots for Coag Time
normally whether the
Normal Probability Plot of the Residuals Residuals Versus the Fitted Values
distributed?
99
5.0
mathematical model fits
90
2.5
equally for low to high
Residual
values of the fits.
Percent
50 0.0
-2.5
10
-5.0
1
-5.0 -2.5 0.0 2.5 5.0 60 62 64 66 68 This graph investigates
Residual Fitted Value
how the Residuals
Histogram of the Residuals Residuals Versus the Order of the Data
5.0
behave across the
4.8
2.5
experiment. This is
3.6
Frequency
probably the most

Residual
0.0
2.4
-2.5
important graph, since
1.2
-5.0
it will signal that
0.0
-4 -2 0 2 4 2 4 6 8 10 12 14 16 18 20 22 24 something outside the
Residual Observation Order
experiment may be
Histogram - bell Individual residuals - operating. Nonrandom
curve? Ignore for patterns are warnings.
small datasets (<30).
trends? or outliers?
‹#›
Example: Step 9b
Step 9b: Interpret the P-Value (or the F-statistic)

for the factor effect.
a. Assuming the residual assumptions are satisfied:
― If P-Value is less than or equal to 0.05, then REJECT Ho
― Otherwise, fail to reject null If P-Value is less than or
hypothesis equal to 0.05 then at
least one group mean is
Analysis of Variance on CoagTime different. In this case,
reject the null
Source DF SS MS F P _ hypothesis that all the
Diet 3 228.00 76.00 13.57 0.000 group means are equal.
At least one Diet mean
Error 20 112.00 5.60 is different.
Total 23 340.00
F-test is close to An F-test result this
1.00 when group large translates to a
means are similar. In very small p-value of
this case, The F-test 0,000 in this case.
is much greater.
‹#›
Example: Step 10
Calculate R2 for the treatment terms

SS Diets 228
R =2
= = 0.67
SSTotal 340
67% of the Variation in CoagTime

can be explained by Diet
Step 10: Translate the statistical conclusion into a practical conclusion.

a. Diet type DOES affect blood coagulation time. In particular,
if you would like to increase blood coagulation time, try
diets B or C. If the goal is to decrease blood coagulation
time, try diets A or D.
‹#›
Exercise
FILE: ANALYZE-ANOVA HOTELS.MTW
City Hotel Stars Price
LA NEW OTANI 3 119
LA HILTON 3 150
LA BEVERLY PLZA 3 110
LA
LA
HOL INN CONV
LE DUFY
2
2
79
145
Are Hotels In New York City
LA BILTMORE 4 140 More Expensive Than Hotels
LA LE PARC 2 165
LA SHERATON GRD 3 175 In Other Cities ?
SF HOL INN FIN 2 99
SF STOUFFER 5 185
SF MANDARIN 4 265
SF DIVA 2 109
SF GRAND HYATT 4 169
SF HOL INN GATE 2 99
SF NOB HILL LAM 2 175
SF INN AT OPERA 3 110
DC LOMBARDY 2 115
DC SHERATON 2 185
DC HILTON 3 166
DC GRAND HYATT 3 189
DC ONE WASH CIR 3 125
DC COMFORT INN 1 64
DC CAPITOL HILL 1 120 25 Minutes
DC RAD PRK TERR 3 119
NY EASTGATE 1 170
NY HELMSLEY MID 2 135
NY HOL INN CRWN 2 185
NY THE MARK 3 250
NY PENINSULA 4 250
NY WARWICK 2 170
NY GRAND HYATT 3 210
NY THE REGENCY 4 215
‹#›
Non Parametric Tests
121
Nonparametric Tests
‹#›
Learning Objectives
1. Apply nonparametric tests when data are ordinal

and/or not normal.
2. Perform tests.
§ Wilcoxon
§ Mann-Whitney
§ Kruskal-Wallis
‹#›
Agenda
§ Nonparametric definition and test efficiency

§ 1-Sample tests (comparing the median of one sample to
a target)
- 1-Sample Sign
- Wilcoxon
§ 2-Sample tests (comparing the medians of two
samples)
- Mann-Whitney
§ More than two samples (comparing the medians of two
or more samples)
- Kruskal-Wallis
‹#›
Definition of Nonparametric
A statistical method is nonparametric when it does not rely on

computing the “parameters” of a specific probability distribution.
Nonparametric methods are used for:
1. Data with a nominal (classes or categories) measurement
scale—e.g. red, yellow, blue.
2. Data with an ordinal measurement scale (ordered data). Only
comparisons of “greater,” “less,” or “equal” between
measurements are possible—e.g. low, medium, high.
3. Continuous data when the distribution is either unknown or is
known to violate an assumption for one of the parametric
methods—e.g. is not normally distributed.
This course deals with #3, which is the most common in

projects.
‹#›
Why Study Nonparametric Tests?
§ Some populations are not normal.

– Tests that assume normality may yield misleading
results, e.g., incorrect confidence intervals.
§ Some data only specify order or counts of events.
– Nonparametric methods are the only possibility.
§ Medians may be more useful representations of
central tendency than means.
– This happens when population is skewed.
‹#›
Nonparametric Tests
Nonparametric Test Parametric (Normal) Test
§ Levene’s Test § Bartlett’s Test
§ 1-Sample Sign § 1-Sample t-Test

§ 1-Sample Wilcoxon
§ Mood’s Median § 2-Sample t-Test

§ Mann-Whitney
§ Mood's Median § ANOVA

§ Kruskal-Wallis
‹#›
Hypothesis Testing Method: Non-normal Data

problem. graphical tool to explore the
2. Establish the hypotheses. data.
a. Not normal
5. Determine initial sample
6. Develop the sampling plan
and collect the data. conclusion.
conclusion.
Problem Problem Solution Solution
‹#›
Concept: Wilcoxon
A 1-Sample Wilcoxon test is a statistical test used to

compare a sample median to a target value in order to
identify if there is a significant difference between them.
This test can be used when the sample data are not
normally distributed, but it should be from a population
whose measurement scale is at least ordinal or continuous
and the distribution is assumed to be symmetrical.
The 1-Sample Wilcoxon is a robust test for non-normal

data, especially for smaller sample sizes, but is not a good
test to use if there are many outliers.
‹#›
1-Sample Wilcoxon in MINITAB®
FILE: ANALYZE-MEDIAN HOUSING PRICES.MTW
Probability Plot of Boston Suburb

Normal
99
Mean 313.4
StDev 151.3
95 N 20
AD 0.737
90
P-Value 0.046
80
70
Percent
60
50
40
30
20
10
1
0 100 200 300 400 500 600 700 800
Boston Suburb
‹#›
1-Sample Wilcoxon in MINITAB®
‹#›
1-Sample Wilcoxon Conclusion
Test of Median = 206.0 versus Median ≠ 206.0
Wilcoxon Signed Rank Test: Boston Suburb

Sometimes
Test of median = 206.0 versus median not = 206.0 MINITAB® shows
the P-Value as p.
N for Wilcoxon Estimated Do not confuse
N Test Statistic p Median this with a
measure of
Boston Suburb 20 20 180.0 0.005 292.5 proportions.
MINITAB®
eliminates the
Estimated Achieved Conf. Interval sample values
N Median Confidence Lower Upper that = the
hypothesized
Boston Suburb 20 293 95.0 235 373 median and
reduces N
accordingly.
State the Practical Conclusion
Conclude that the median housing price in Boston is not equal to

$206K at a significance level of .005.
‹#›
Concept: Mann-Whitney
The Mann-Whitney test is a statistical test used to compare the

medians from two random samples in order to determine if they are
from the same population (i.e. the medians are equal) or from different
populations (the medians are different).
It requires:
§ The measurement scale is at least ordinal
§ The two sample distributions are of similar shape
§ The two variances are not significantly different
But...
§ Does not assume normality
§ Distributions do not need to be symmetrical
§ The two samples do not need to be the same size
MINITAB® uses the Wilcoxon formulation to analyze Mann-Whitney

and is often called the Wilcoxon Mann-Whitney.
‹#›
Example: Mann-Whitney
FILE: ANALYZE-MANN-WHITNEY 2 PROCESS LOSS.MTW
In this example you want to

test whether changing from Process 1 Process 2
Process 1 to Process 2 Loss Loss
270 218
reduces the amount of 236
225
234
214
material lost per batch. 130
280
116
200
272 276 Test that Process 2
160 146
For a Mann-Whitney test you 220 182 has been improved
130 238
need 2 samples: 242 288 and that its median
186 190 loss is less than the
§ Not necessarily equal 266 236
216 244 losses from Process 1.
318 258
§ Each with similar 294
282
240
294
continuous distributions, 234
258
220
200
but not necessarily 276
282
220
186
normal or symmetric 390
290
352
202
280 218
278 248
The batches at the right are 288 278
288 248
the results before and after 244 270
236 242
the change.
‹#›
Hypothesis Testing Method: Non-Normal Data

problem. graphical tool to explore
2. Establish the hypotheses. the data.
5. Determine initial sample a. Not Normal
6. Develop the sampling determine the statistical
plan and collect the data. conclusion.
conclusion.
Problem Problem Solution Solution
‹#›
Step 1: State the Practical Problem: You want to compare the medians of
two processes to see if one reduces loss.
Step 2: State the null and alternate P robabili ty P lot of P r ocess 1 L oss
hypotheses. 99
Normal
Mean 252.5
Ho: Median 1 = Median 2 95
StDev
N
54.91
28
Ha: Median 1 > Median 2

AD 0.944
90
P-Value 0.014
80
or Process 2 reduces loss 70
Pe r cent
60
Step 3:
50
Test medians, original not 40
30
normal. 20
Step 4:
10
Alpha level 0.05. 5
1
100 150 200 250 300 350 400
Steps 5-6: Collect Data: You have data Process 1 Loss
from Process 1 and you need to collect data from Process 2.

You do not need the same number of data points.
‹#›
Step 7: Use the appropriate tool to explore the data.
Boxplot of Process 1 Loss, Process 2 Loss

400
350
300
Data
250
200
150
100
Process 1 Loss Process 2 Loss
‹#›
Step 8: Check distribution
assumptions.
§ Normality
§ Similar shape—if not
use Mood’s Median
Summary for Process 1 Loss Summary for Process 2 Loss

A nderson-D arling N ormality Test A nderson-D arling N ormality Test
A -Squared 0.94 A -Squared 0.30
P -V alue 0.014 P -V alue 0.547
M ean 252.54 M ean 230.64
S tDev 54.91 S tDev 46.97
V ariance 3015.00 V ariance 2205.94
S kew ness -0.39980 S kew ness 0.01944
Kurtosis 1.39314 Kurtosis 1.34430
N 28 N 28
M inimum 130.00 M inimum 116.00
1st Q uartile 227.25 1st Q uartile 200.50
M edian 268.00 M edian 235.00
3rd Q uartile 282.00 3rd Q uartile 255.50
160 240 320 400 M aximum 390.00 150 200 250 300 350 M aximum 352.00
95% C onfidence Interv al for M ean 95% C onfidence Interv al for M ean
231.24 273.83 212.43 248.85
95% C onfidence Interv al for M edian 95% C onfidence Interv al for M edian
236.00 280.00 215.79 246.21
95% C onfidence Interv al for S tDev 95% C onfidence Interv al for S tDev
9 5 % Confidence Inter vals 9 5 % Confidence Inter vals
43.41 74.74 37.13 63.93
Mean Mean
Median Median
230 240 250 260 270 280 210 220 230 240 250
‹#›
Step 9a: Run the test.
For confidence interval
For significance test
‹#›
Mann-Whitney Signed Rank Test
Step 9b: State statistical conclusions.

Mann-Whitney Test and CI: Process 1 Loss, Process 2 Loss
N Median
Process 1 Loss 28 268.00
Process 2 Loss 28 235.00
Point estimate for ETA1-ETA2 is 28.00

95.2 Percent CI for ETA1-ETA2 is (-0.01,48.00)
W = 921.5
Test of ETA1 = ETA2 vs. ETA1 > ETA2 is significant at 0.0219
The test is significant at 0.0219 (adjusted for ties)
Step 10: State the Practical Conclusion.

Sample 1 Median is > Sample 2 Median so that Process 2 does
reduce the loss.
‹#›
Exercise: Mann-Whitney
FILE: ANALYZE-DEPARTMENT CLAIMS—KRUSKAL WALLIS.MTW
You have claim resolution time data (in days) from 2 claims-
processing departments. You desire to prove if one
department is processing claims faster than the other.
Department 1 Department 2
150 130
335 315
148 128
146 126
160 140
156 136
152 132 10 Minutes
148 128
144 124
55 35
‹#›
Concept: Kruskal-Wallis
The Kruskal-Wallis test is a statistical test used to compare medians from

two or more random samples to see if they come from the same or different
populations. It is an extension of the Wilcoxon Mann-Whitney test.
It requires:
§ The measurement scale is at least ordinal
§ The distributions are of similar shape
§ Distribution is any continuous one
But . . .
§ Does not assume normality
§ Distribution does not need to be symmetric
§ The samples do not need to be the same size
The Kruskal-Wallis is the most robust test to use when you are testing
medians from more than 2 samples.
‹#›
Example: Kruskal-Wallis
FILE: ANALYZE-DEPARTMENT CLAIMS—KRUSKAL-WALLIS.MTW
A project team is testing theories on causes for long cycle

times to resolve claims in an insurance company.
§ Three different departments process claims in the

company.
§ Some of the theories relate to which of the three
departments does the processing.
§ The team wishes to test the theory that there is a
difference among the three departments in claim
processing time.
‹#›
Step 1: State the Practical Problem; Does Department 3

resolve claims faster than Department 1 and
Department 2?
Ho = There is no difference among Department
resolution time
Ha = There is a difference in Department
resolution time
Step 3: Believed non-normal, hence Kruskal-Wallis.
Step 4: α = 0.05 Confidence Level = 95% (1- α = 0.95).
‹#›
FILE: ANALYZE-DEPARTMENT CLAIMS—KRUSKAL-WALLIS.MTW
Step 5-6: Determine sample size (n=10 each), plan, and collect
data.
Department 1 Department 2 Department 3
150 130 125
335 315 310
148 128 123
146 126 121
160 140 135
156 136 131
152 132 127
148 128 123
144 124 119
55 35 30
Step 7: Chart data to explore.

Step 8: Check normality—Run Kruskal-Wallis as indicated.
‹#›
Step 9: Run and interpret the results.

Kruskal-Wallis Test: C6 versus C5
Kruskal-Wallis Test on C6
C5 N Median Ave Rank Z
Department 1 10 149.0 22.1 2.90
Department 2 10 129.0 14.1 -0.62
Department 3 10 124.0 10.3 -2.29
Overall 30 15.5
H = 9.36 DF = 2 P = 0.009
H = 9.37 DF = 2 P = 0.009 (adjusted for ties)
Since the P-Value is less than 0.05, reject H0.

Step 10: State the Practical Conclusion.
There is a difference in Department claim resolution time, but
is Department 3 faster than Department 2?
‹#›
You now need to compare Department 3 against Department 2
since their medians are close.
Repeat Step 9: Perform a Mann-Whitney test on
Dept 2 > Dept 3.
N Median
Department 2 10 129.0
Department 3 10 124.0
Point estimate for ETA1-ETA2 is 5.0

95.5 Percent CI for ETA1-ETA2 is (-5.0,15.0) W = 124.0
Test of ETA1 = ETA2 vs. ETA1 > ETA2 is significant at 0.0810
The test is significant at 0.0808 (adjusted for ties)
Since the P-Value is significant at 0.0810, you can reject Ho at 90% level
but not at 95%.
Repeat Step 10: State the Practical Conclusion.
You are 90% sure that Department 3 is faster than

Department 2.
‹#›
Exercise: Three or More Medians
FILE: ANALYZE-3 OR MORE SAMPLES.MTW
1. A Trucking company wishes to test the lifetime of

tires (as determined from tread wear) from 4
different manufacturers. Each of 4 types of tire has
been tested on six similar light trucks. Determine if
there is a significant difference between tire
manufacturers.
7 Minutes
‹#›
Summary
1. Nonparametric tests are commonly used when one or more of

the samples being tested come from a distribution other than
normal.
2. Graphing the data along with testing for normality are
necessary steps prior to running the tests.
3. The 1-Sample Wilcoxon test is a more robust test than the 1-
Sample Sign when testing median to a target.
4. The Mann-Whitney test is a more robust test than Mood's
Median when comparing medians from 2 populations.
5. The Kruskal-Wallis test along with Mood's Median are used to
compare medians from two or more samples. Kruskal-Wallis
is the preferred test.
‹#›
Testing Hypothesis
150
Testing Hypotheses When Y
Is Continuous and X Is
Continuous
Correlation and Regression
and
Scatter Diagrams
‹#›
Learning Objectives
1. Define the concept of scatter diagrams.
2. Describe how to use and create scatter diagrams.
3. Interpret the patterns shown in scatter diagrams.
4. Use correlation and regression tools to narrow the

list of continuous input variables.
5. Perform correlation/regression analysis.
6. Deliver an updated list of critical Xs.
‹#›
Scatter Diagram
Concept
A scatter diagram is a graphic representation of the
relationship between two variables.
Variable (Y)
Output
Process Variable (X)
‹#›
Scatter Diagrams and Cause-Effect Relationships
RH V RH V
10 43.1 30 46.0
39.9 43.2
41.3 45.5
42.2 45.8
40.4 44.1
20 45.2 40 49.1
42.9 45.0 Scatterplot of Volts vs RH
42.6 48.4 60
44.3 48.9
55
vs. Volts
50
45
40
0 20 40 60 80 100
RH
‹#›
Constructing Scatter Diagrams
To construct a scatter diagram you need:

§ A good theory
§ Correctly paired data
§ Accuracy
§ Complete information
§ Representative data
‹#›
Steps in Constructing and Interpreting a Scatter
Diagram
1. Develop a plausible and relevant theory about the
suspected relationship between two variables of interest.
2. Obtain the table of raw data, and determine the high and
low values for each variable.
3. Decide which variable will be plotted on the horizontal axis.
4. Draw and label the horizontal and vertical axes.
5. Plot the paired data.
6. Title the chart, and provide other appropriate notations.
7. Identify and classify the pattern of correlation.
8. Check for potential pitfalls in your analysis. Consider
confounding factors and other possible explanations for
the observed pattern, and decide on the team’s next steps.
‹#›
Typical Patterns of Correlation
Patterns of Correlation
Y Y Y
X X X
Strong, Positive Strong, Negative Complex
Y Y Y
X X X
Weak, Positive Weak, Negative None
‹#›
Confounding Factors
Database: Errors Vs. Experience

Average Errors / Week in May
22
20
Average Errors Per Week
18
16
14
12
10
8
6
4
2
0
0 12 24 36 48 60 72
Months of Service
‹#›
Stratified: Errors vs. Experience
Average Errors / Week in May

22 Group A
20 Group B
Average Errors Per Week
18
16
14
12
10
8
6
4
2
0
0 12 24 36 48 60 72
Months of Service
‹#›
Shelf Life Study 1
150 Specified
Weight of Active Ingredient (mg)
Shelf
Life
140
130
Minimum
120 Effective
Weight
110
100
0 5 10 15 20 25
Time Since Production (months)
‹#›
Shelf Life Study 1
150 Specified
Shelf
Life
140
130
Minimum
120 Effective
Weight
110
100
0 5 10 15 20 25
‹#›
Shelf Life Study by Supplier
Supplier A
150 Supplier B Specified
Shelf
Life
140
130
Minimum
120 Effective
Weight
110
100
0 5 10 15 20 25
‹#›
Example: Scatter Diagrams Using MINITAB®
FILE: ANALYZE-STICKY.MTW
A “Sticky” Diagnostic Journey
A manufacturer of heavy machinery used in road

construction was receiving complaints from the field
about erratic operation of hydraulic systems. The
hydraulic systems, used to raise and lower the bucket of
a front-end loader, for example, had a tendency to “stick”
(cease to operate) in certain positions near the extremes
of their travel…(continued below)
‹#›
‹#›
Scatterplot of Volts vs RH
60
55
Volts
50
45
40
0 20 40 60 80 100
RH
‹#›
Scatter Diagrams
Warning! Correlation Does Not Imply Causation
100 200 300
80 80
70 70
Population
(In Thousands)
60 60
50 50
100 200 300
Number of Storks
Source: Box, Hunter, Hunter. Statistics For Experimenters. New York, NY: John Wiley & Sons. 1978
‹#›
Scatter Diagrams: Pitfalls
Pitfalls:
§ Assuming correlation = causation
- Rum prices and ministers’ salaries are known to be correlated.
Does an increase in salary cause a rum price increase? Or the
other way around?
- Patient days have increased without staffing increases while
computer budget has increased. Does more spending on
computers boost nursing staff productivity?
§ Mixing up the “X” and the “Y” variables

§ Data are not paired correctly
§ Improper scaling: Scatter diagram is not square or plots only part of
range
§ A plausible and convincing physical explanation is all that is needed;
so data and analysis are not done
‹#›
Pitfalls and Problems in Interpretation
Stratification and Range of Data
50 A
50 A
Y
45
B
Y
45
40
0 25 50 75 100
X
True Relationship between X & Y 40
0 20 40
Complex Correlation X
Examining Only Part of Range
Yields False Conclusion of Strong
Positive Correlation
‹#›
Pitfalls and Problems in Interpretation
Range of Operation and Effect of Scale
50 B 100 C
49
75
48 Y
Y
50
47
25
46
45 0
30 50 70 0 25 50 75 100
X X
No correlation exists if process only Scale change affects the appearance

operates from X = 30 to 70 of the pattern of correlation
‹#›
When to Use Scatter Diagrams
When to Use
§ Testing hypotheses and identifying vital few Xs
§ Designing remedies and controls
‹#›
Summary—Scatter Diagrams
1. Cause-effect diagrams help develop theories; scatter

diagrams help analyze data to prove or disprove those
theories.
2. The graphic nature of the diagram shows the
relationships between variables.
3. A relationship does not necessarily mean cause.
4. Analysis through scatter diagrams is a 4-step process.
a. Develop theory.
b. Collect paired data and create the graph.
c. Identify and classify the pattern of correlation.
d. Revisit the theory in light of the data and potential
pitfalls.
‹#›
Correlation and Regression Topics
§ Why do you need to use Correlation and Regression?

§ Correlation description
§ How to analyze data using Correlation
§ Simple Regression Analysis description
§ How to analyze data using Simple Regression
§ Summary
§ In-class exercises
‹#›
Definitions
§ Correlation
§ Regression Equation
§ Coefficient of Determination
‹#›
Why are These Tools Used?
§ They can be used on the

process in a natural operating All
state without artificially Possible
inducing variation (as with Xs
DOE).
§ Scatter diagrams can show

complex relationships that
help avoid statistical mistakes.
Vital Few Xs
§ Correlation provides RegressionPlot
quantification of the strength 60
of the relationship between 50
two variables. OUT-1 40
30
20
§ Regression can provide 10
Y=-10.3333+7.75X
R- Squar ed=0.941
prediction equations with 0
which independent variables -10

Regr essi on
95%CI
can be controlled. -20 95%PI
0 1 2 3 4 5 6 7 8
KNOB-1
‹#›
How Much Does X Impact the Y?
What Is the Size of the Correlation?
You can quantify the association between X

and Y with the Correlation Coefficient r.
( xi - x )( yi - y )
The Correlation Formula: r= å
å (x i
- x ) 2 å ( yi - y ) 2
30 30 30 30
29 29 29 29
28 28 28 28
27 27 27 27
26 26 26 26
25
C1
25 25 25
C1
C1
C1
24 24 24 24
23 23 23 23
22 22 22 22
21 21 21 21
20 20 20 20
20 21 22 23 24 25 26 27 28 29 30 20 21 22 23 24 25 26 27 28 29 30 20 21 22 23 24 25 26 27 28 29 30 20 21 22 23 24 25 26 27 28 29 30
C2 C2 C2
Corr of C1 and C2 Corr of C1 and C2 Corr of C1 and C2 Corr of C1 and C2

0.057 0.819 0.985 1.00
‹#›
Correlation Relationship Significance Table
Decision Decision
n point
n point
5 0.878 18 0.468
6 0.811 19 0.456
7 0.754 20 0.444
8 0.707 22 0.423
Guideline: 9 0.666 24 0.404
§ If | r | > 0.80, then 10 0.632 26 0.388
relationship is 11 0.602 28 0.374
important. 12 0.576 30 0.361
13 0.553 40 0.312
§ If | r | < 0.20, then
14 0.532 50 0.279
relationship is not
significant. 15 0.514 60 0.254
16 0.497 80 0.22
17 0.482 100 0.196
‹#›
Correlation Data Requirements
§ A correlation study must have:

– BIVARIATE DATA: The term is (Response) (Factor)
used to describe two pieces of
data that are connected Y (output) X (input)
presumably in a relationship
Y1 X1
§ Expressed mathematically, bivariate Y2 X2

data comprise ordered pairs: X and
Y (X, Y)
Y3 X3
Y4 X4
§ It is customary to call X, the input Y5 X5
variable, (independent) and Y, the
output variable, (dependent) Y_nth X_nth
NOTE: Correlation study can also be done with more than one X.
‹#›
Correlation Example
FILE: ANALYZE-TEMP VS. MORTALITY.MTW
Region Temperature Mortality

Is there evidence of a
1 31.8 67.3
2 34.0 52.5 relationship between the
3 40.2 68.1 mean annual
4 42.1 84.6
5 42.3 65.1 temperature in the
6 43.5 72.2
7 44.2 81.7 region and the mortality
8 45.1 89.2 index?
9 46.3 78.9
10 47.3 88.6
11 47.8 95.0
12 48.5 87.0
13 49.2 95.9
14 49.9 104.5
15 50.0 100.4
16 51.3 102.5
‹#›
Identical Values of Correlation Coefficient
5 5
4 4
3 3
Y 2 Y 2
1 1
0 0
0 1 2 3 4 5 0 1 2 3 4 5
X X
5 5
4 4
3 3
Y 2 Y 2
1 1
0 0
0 1 2 3 4 5 0 1 2 3 4 5
X X
‹#›
Regression
What if you could determine a model…
§ Time to Close an Order

o n
si
= 5.2 x size of order
+3 x number of service calls this year
e s
-2.7 x sales rep’s experience
g r
OR
Re
§ Variability of Forecast
= 3.8 x week of the month
+ 2.5 x production change of past month
+ 1.7 x variability of past quarter
…to predict and control future performance?
‹#›
Linear Relationship
b = Y intercept = the Y value when
the line intersects Y axis at
X=0 rise
m = slope =
run
Y
rise
run
b
0 X
A simple linear relationship can be described
mathematically by Y = b+mX
‹#›
Regression Example (Fitted Line Plot)
FILE: ANALYZE-SCATT 39.MTW
Obtain a prediction equation for mass of

fiber as a function of speed of discharge
from spinneret.
Fitted Line Plot
Mass = 18.00 + 4.784 Speed
44 Regression
95% C I
95% PI
42
S 1.35473
R-Sq 69.5%
40
R-Sq(adj) 67.9%
38
Mass
36
34
32
MINITAB® Output: 30
The regression equation is 3.50 3.75 4.00
Speed
4.25 4.50 4.75
Mass = 18.00 + 4.784 Speed

S = 1.35473 R-Sq = 69.5% R-Sq(adj) = 67.9%
Source DF SS MS F P
Regression 1 79.610 79.6096 43.38 0.000
Error 19 34.870 1.8353
Total 20 114.480
‹#›
Regression Example (Residuals)
FILE: ANALYZE-SCATT 39.MTW
Investigate the residuals plot.

Exercise: Regression
Residual Plots for Mass

Normal Probability Plot Versus Fits
99
2
90
1
Residual
Percent
50 0
-1
10
-2
1
-3.0 -1.5 0.0 1.5 3.0 34 36 38 40
Residual Fitted Value
Histogram Versus Order

8
2
6 1
Frequency
Residual
4 0
-1
2
-2
0
-3 -2 -1 0 1 2 2 4 6 8 10 12 14 16 18 20
Residual Observation Order
‹#›
Simple Regression Practice Exercises
FILE: ANALYZE-EROSION.MTW
Tungsten steel erosion shields are fitted to the low pressure

blading in steam turbines. The most important feature of a shield is
its resistance to wear. This is difficult, expensive and destructive to
measure directly in terms of abrasion loss, but it is known to be
associated with the hardness of steel.
1. What do you want to know?
Can you quantify the relationship between hardness (the X) and
abrasion loss (the Y) so that you can can predict the erosion
shield’s resistance to wear (abrasion loss)? Measuring hardness
would be more convenient and non-destructive which would
result in a cost savings.
2. Start with some graphs
Predict the nature of the relationship. Draw a scatter plot. Is there
a relationship? Does it match your prediction?
3. What tool should you use?
Simple Linear Regression.
Interpret the statistical results. Complete an analysis of the
residuals.
45 Minutes
‹#›
Simple Regression Practice Exercises, Solutions
Start with some graphs

You could predict that the abrasion loss would decrease as the
hardness increases. Drawing a scatter plot shows this to be the
case.
Problem 1
Scatterplot of Abrasion vs Hardness
650
600
550
500
Abrasion
450
400
350
300
650 675 700 725 750

Hardness
‹#›
Simple Regression Practice Exercises, Solutions
Regression Analysis
The regression equation is Problem 1, MINITAB®
Abrasion = 2693 - 3.16 Hardness Analysis
Predictor Coef StDev T p
Constant 2692.8 242.9 11.09 0.000
Hardness -3.1607 0.3462 -9.13 0.000
S = 41.9371 R-Sq = 78.4% R-Sq(adj) = 77.4%
Source DF SS MS F p
Regression 1 146569 146569 83.34 0.000
Error 23 40451 1759
Total 24 187020
The P-Value is ≤ 0.05 and R-sq explains 78% of
the variation. Check the residuals.
‹#›
Summary
1. Correlation is the measure of the relationship between

two quantitative variables. Be careful in assuming
causality.
2. Regression analysis seeks to find a relationship between

the variables in the form of a prediction equation. The
equation may be the answer desired or it may be the
means to the prediction that is desired.
‹#›
Checkpoints
1. How were potential causes identified?

2. How were the key factors identified? How were they verified?
3. Root causes of the problem have been determined along with
how conclusions were drawn.
4. Your approach to analyzing your data and why the approach
was chosen has been documented.
5. How have the data been analyzed to identify the root causes
of process variation and how have they been documented?
Data stratification was documented if used. The cause-effect
diagram tree was clear. Did you consider running
experiments?
6. The financial opportunity represented by addressing the
problem has been documented, along with the impact on
customer satisfaction, loyalty, and retention.
‹#›
Project Plan Deliverables
Project Plan Deliverables

Data-collection plan for Analyze Phase
■ Each theory tested (proven or disproven)
displayed as follows: Theory, Ho, Ha, MINITAB®
printout, graph printout, statistical conclusion and
practical conclusion/implication. Photos optional.
■ Summary, listing proven theories and disproven
theories.
List of (proven Xs) root causes, i.e., Y= f (X1,

X2,…Xn)
‹#›
Additional Summary Exercises
FILE: ANALYZE-FINAL EXERCISES.MTW
Homework – additional practice: Use the 10 step

Hypothesis Testing procedure to analyze the five scenarios
described in your book.
45 Minutes
‹#›
Effort Consulting
TRAINING / COACHING / CONSULTING
q Our services
• Business Transformation
• Enterprise Management System Get in touch with us
• Strategy & Digital Marketing
Efforts Consulting Pvt. Ltd.
• IRMM (Digital Factories) 508, Palladium Business Centre,
Opp. 4D Square Mall,
• Lean Six Sigma Near IIT Engg. College,
• Time and Motion Study Gandhinagar Road, Ahmadabad – 382424
(+91) 9879391004
• MFCA
www.effortsconsulting.com
• Productivity Norms
‹#›

Analyze

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Analyze

Uploaded by

Copyright:

Available Formats

Analyze

1. Learn how to analyze data to find root cause.

2. Understand how to build the null and alternative

3. Determine if data are normal or non-normal.

4. Determine the proper test.

Graphical Interpretation of Observed Data

Statistical Interpretation of Observed Data

Graphical Interpretation of Experimental Data

Statistical Interpretation of Experimental Data

For each theory or group of theories identified, indicate the

1. Define the concept of stratification.

2. Describe how to stratify data.

3. Interpret stratified data.

Stratification is the separation of data. Its

Stratification answers the following TYPES of questions:

§ Is there a defect that occurs most frequently?

Days of the Week Accountants

Total Errors Percent in Error

1. Select the stratification variables

Basically, look across the categories of a variable to see

Pitfalls and Traps

§ Concluding the anomalous category is the cause

§ When collecting new data, failing to collect

Pareto Chart Example

What differences do you see between the output

Agree or Disagree? Why?

1. Stratification is the separation of data into categories.

1. Introduce the terminology and methodology

2. Use the Hypothesis Testing Roadmap as a guide.

3. Understand alpha and beta risk and their

American Justice System Example

Null Hypothesis ( Ho ): Defendant is not guilty

The only four possible outcomes are:

2. An innocent defendant is jailed Type I error

a. The probability of this type of error

3. A guilty defendant is set free Type II error

a. The probability of this type of error

4. A guilty defendant is jailed Correct decision

“Is there a real difference between and ?”

§ Practical Question: Did the modifications on Process B improve

§ Graphical representation of the data: Does the observed

Or, do the yield

Null Hypothesis (Ho) Alternative Hypothesis (Ha)

§ Null Hypothesis (Ho)

P-Value is the probability of being wrong if you say

1. State the practical problem. 7. Use the appropriate

1. Hypothesis testing follows specific steps, and requires

1. Determine the range of values that includes the

§ Definition of a Confidence Interval (CI)

§ Why do you need to study CIs?

What Is Alpha (α)?

§ Confidence Intervals take the general form:

§ For estimating µ, K is selected from a family of “t” distributions.

§ The t Distribution is a family of bell-shaped (normal-like)

§ Use t-Table instead of Z-Table to compute the area in the

This result states with 95% certainty the actual process

1. Compare a proportion to a target.

3. Power and Sample Size.

4. Is Output Y dependent on Input X?

§ If you flip a coin 10 times, and get 8 heads, is it a fair coin?

§ What if you have 100 coin flips and

Note: Rejecting Ho is a much stronger conclusion than

§ You have just learned how to compare a