Prepared by: Dr.

Shazarel

Tutorial :
Table 16 : Tensile strength of paper

Hardwood Observations
Concen. (%) 1 2 3 4 5 6 Total Avg
5 7 8 15 11 9 10 60 10.00
10 12 17 13 18 19 15 94 15.67
15 14 18 19 17 16 18 102 17.00
20 19 25 22 23 18 20 127 21.17
383 15.96
a = 4, n = 6, N = an = 24.
Please find :

(i) SSF, SSE and SST

(ii)degrees of freedom of SSF and degrees of freedom of SSE
(iii)MSF and MSE
(iv)Fo
1
 Prepared by: Dr. Shazarel

Tutorial :

2
 Prepared by: Dr. Shazarel

SUMMARY: ANOVA for Single Factor Experiment

1.Decide the factor to be investigated and the following two parameters:

1) number of levels a;
2) replicate n.
2.Carry out the experiments, obtain N (= an) observations yij by a random manner.
3.Check if the residuals (estimates of errors) satisfy the requirements.
4.Calculate sum of squares SST, SSF and SSE.

3
 Prepared by: Dr. Shazarel

Cont.

5. Calculate the degrees of freedom for each of SST, SSF, SSE

6. Calculate the mean squares MSF and MSE.
7. Calculate the ratio F0 between MSF and MSE.
8. Specify a type I error , and find the control value F , , from the F
distribution Tables (1 = a - 1, 2 = a (n – 1) ).
9. Conclude if the factor has significant effect on the response.

4
 Prepared by: Dr. Shazarel

6.4 MULTIPLE COMPARISON TECHNIQUES

Test on Individual Level Means

▪ ANOVA results in a single parameter F0, which can tell whether a factor has effect
on the response, or whether the mean response values will be different at
different levels of the factor from an overall viewpoint.

▪ However, ANOVA cannot decide the direction of influence of the factors, nor
identify which factor level mean is different from another level mean.

5
 Prepared by: Dr. Shazarel

Figure 6.6 Box plots of paper strength example

6
 Prepared by: Dr. Shazarel

▪ The plots indicate that changing the hardwood concentration has a strong effect on
the paper strength; specifically, higher hardwood concentrations produce higher observed
paper strength.

▪ Box plots show the variability of the observations within a factor level and the
variability between factor levels.

▪ However, a less subjective approach is required to test the individual level means.

7
 Prepared by: Dr. Shazarel

Multiple Comparison Procedures

▪ It is the techniques for making comparisons between two or more level means
subsequent to an analysis of variance.
▪ When we run an analysis of variance and obtain a significant F0 value, we have
shown simply that the overall null hypothesis is false. We do not know which of a
number of possible alternative hypotheses is true.
H1: µ1 ≠ µ2 ≠ µ3 ≠ µ4 ≠ µ5
or H1:µ1 ≠ µ2 = µ3 = µ4 = µ5
▪ Multiple comparison techniques allow us to investigate hypotheses that involve
means of individual levels. For example, we might be interested in whether µ1 is
different from µ2, or whether µ2 is different from µ3.

8
 Prepared by: Dr. Shazarel

▪ One of the major problems with making comparisons among level means is that
unrestricted use of these comparisons can lead to an excessively high probability of a Type I
error.
▪ For example, if we have 10 levels in which the complete null hypothesis is true (with α =
0.05)
Ho: µ1 = µ2 = µ3 = … = µ10
▪ A series t tests between all pairs of level means will make the possibility of making at least
one Type I error very high.

▪ The probability of making at least one Type I error increases as we increase the number of
independent t tests we make between pairs of level means. While it is nice to find significant
differences, it is not nice to find ones that are not really there. We need to find some way to make
the comparisons we need but keep the probability of incorrect rejections of Ho under control.

9
 Prepared by: Dr. Shazarel

Least Significant Difference (LSD) Test is a useful tool to test the differences between the
individual level means.

▪ The basic requirement for a LSD test is that the F0 for the overall analysis of
variance (ANOVA) must be significant. If the F0 was not significant, no comparisons
between level means are allowed. You simply declare that there are no differences
among the level means and stop right there.

▪ On the other hand, if the overall F0 is significant, you can proceed to make any or all
pair wise comparisons between individual level means by the use of a t test.

10
 Prepared by: Dr. Shazarel

• To test between the ith level mean (µi) and jth level mean (µj),

Ho: µi = µj

H1: µi ≠ µj we calculate (6.20)

(6.21)

• The value of MSE has already been obtained during the overall ANOVA. The DOF
of t is equal to the DOF of MSE, i.e., a(n-1) or (N-a).
• Then, using a two-tailed t test for a specified level of type I error α
• (e.g., α = 0.05), we can decide a pair of two-sided control limits
• ±tα(dof) from the t table.
11
 Prepared by: Dr. Shazarel

12
 Prepared by: Dr. Shazarel

▪ If the t value determined by Eq (6.21) falls within the limits ±tα(dof), we cannot reject
the null hypothesis in (6.20). We will therefore conclude that there is no difference
between µi and µj.

▪ On the other hand, if the t value falls beyond the limits ±tα(dof), we will reject the null
hypothesis and conclude that there is a difference between µi and µj.

13
 Prepared by: Dr. Shazarel

The Example of Tensile Strength of Paper

Table 18
Hardwood Observations
Concen. (%) 1 2 3 4 5 6 Total Avg
5 7 8 15 11 9 10 60 10.00
10 12 17 13 18 19 15 94 15.67
15 14 18 19 17 16 18 102 17.00
20 19 25 22 23 18 20 127 21.17
a=4, n=6

y1 =10.00, y2=15.67, y3=17.00, y4=21.17

14
 Prepared by: Dr. Shazarel

Problem:
Are hardwood concentrations 5% and 20% produce the same paper strength ?

Hypothesis: Ho : 1 = 4, H1 : 1 = 4

Solutions (Five Steps):

(1) Use Eq (6.21)

y1 =10.00, y4=21.17

(Note: MSE = 6.51 was calculated during ANOVA, which is always carried out before
the test on individual level means)

15
 Prepared by: Dr. Shazarel

(2) Determine the DOF, DOF = a (n – 1) = 4 × (6 – 1) = 20

(3) Decide a type I error level. We use α = 0.01 as in the ANOVA for this example.

(4) Find the limits from the t table: ±tα(dof) = ±t0.01(20) = 2.845≈ 3.00

(5) Make conclusion: since the calculated t value (= -7.58) falls beyond the limits
±3.00, we will reject the null hypothesis.

Conclusion: There is a significant difference in paper strength between using

hardwood concentrations 5% and 20%. Moreover, since t < 0, µ1 < µ4.

y1 = µ1 = 10.00 y4 = µ4 = 21.17

16
 Prepared by: Dr. Shazarel

17
 Prepared by: Dr. Shazarel

General Steps to Test the Individual Level Means

(1) Calculate t using Eq (6.21) and find the MSE value obtained during ANOVA.
(2) Determine the DOF, DOF = a (n – 1).
(3) Specify a type I error α.
(4) Find the control limits from the t table: ±tα(dof) based on α and DOF.
(5) Make conclusion by comparing the calculated t value against the control limits.

18
 Prepared by: Dr. Shazarel

6.5 RESIDUAL ANALYSIS & MODEL ADEQUACY CHECKING

Requirements on Error  ij
▪ Error ij results from measurement error and the effects of the nuisance factors that
are not considered in the experiment.
▪ ij should satisfy the following requirements, so that the F0 ratio will follow an F
distribution and the control limit F ,1 ,2 can be easily determined from the associated
Table.
(1) normal distribution.
(2) independent distribution.
(3) zero mean value.
(4) constant variance 2 at all runs.
▪ In actual model check, the residual eij, instead of the error ij, will be considered.

19
 Prepared by: Dr. Shazarel

▪ Residual Analysis and Model Check

Tensile strength of paper (From example 1)

Hardwood Observations
Concen. (%) 1 2 3 4 5 6 Total Avg
5 712 84 157 112 915 1021 60 10.00
10 1219 171 1313 1814 1917 156 94 15.67
15 1410 1820 1911 1724 1622 185 102 17.00
20 199 2516 2223 233 188 2018 127 21.17

Error:  ij = yij − i (theoretical error)

Residual: eij = yij - ȳi. (estimated error) (a1)

20
 Prepared by: Dr. Shazarel

▪ Since i  yi. then, ij  eij . Residual eij is the estimate of error ij.

Actual Data Residuals for the Paper Strength

Experiment

▪ The total number of residuals is equal to the total number of

observations, i.e., N = an.

a = 4, n =6 , therefore N=(a)(n)=(4)(6)=24

21
 Prepared by: Dr. Shazarel

4 Steps for Model Check:

1) The normality assumption is checked by plotting the residuals on the normal probability
paper.
2) The independent assumption is checked by plotting the residuals against the time or run
order in which the observations were taken.
3) Zero mean assumption is satisfied automatically if equation (a1) is used to calculate the
residuals.
4) In practice, constant 2 is indicated by constant range. Therefore, the assumption of
equal variances is checked by plotting the residuals against the levels and
comparing the spreads of the residuals. Range is the difference between the max.
and min. residuals in a run.

22
 Prepared by: Dr. Shazarel

Procedures
1.Rank the residuals in ascending order, ek, k = 1, 2, ....., N (ek  ek+1).

2.Estimate the cumulative distribution function value for each ek:

(6.22)

3.Plot all N points (ek, Pk) on the normal probability paper.

4.If the points cluster to a straight line, the residual distribution is close to normal.

23
 Prepared by: Dr. Shazarel

Paper Strength example

k ek Pk
1 -3.67 0.0208
2 -3.17 0.0625
3 -3 0.1042
4 -3 0.1458
5 -2.67 0.1875
6 -2.17 0.2292
7 -2 0.2708
8 -1.17 0.3125
9 -1 0.3542
10 -1 0.3958
11 -0.67 0.4375
12 0 0.4792
13 0 0.5208
14 0.83 0.5625
15 1 0.6042
16 1 0.6458
17 1 0.6875
18 1.33 0.7292
19 1.83 0.7708
20 2 0.8125
21 2.33 0.8542
22 3.33 0.8958
23 3.83 0.9375
24 5 0.9792

24
 Prepared by: Dr. Shazarel

(1) Check normal distribution.

*If the points cluster to a straight line, the residual distribution is close to normal.

25
 Prepared by: Dr. Shazarel

Run order eij

1 1.33
2 1
3 1.83
(2) Check independent distribution. 4
5
-2
1
6 -0.67
7 5
8 -3.17
9 -2.17
10 -3
11 2
12 -3
13 -2.67
14 2.33
15 -1
16 3.83
17 3.33
18 -1.17
19 -3.67
20 1
21 0
22 -1
23 0.83
24 0

Plot of residuals versus the time or run order

26
 Prepared by: Dr. Shazarel

(3) Check zero mean value

Use equation (a1)

Residual
Conc. 1 2 3 4 5 6 Total (eij/n)
5 -3 -2 5 1 -1 0 0.00
10 -3.67 1.33 -2.67 2.33 3.33 -0.67 0.00
15 -3 1 2 0 -1 1 0.00
20 -2.17 3.83 0.83 1.83 -3.17 -1.17 0.00

That is, zero mean assumption is satisfied automatically.

27
 Prepared by: Dr. Shazarel

(4) Check constant variance 2 at all levels.

Plot of residuals versus the runs

28
 Prepared by: Dr. Shazarel

Summary
▪ In general, moderate departures from the basic assumptions on the errors ij (or
residuals) are of little concern in ANOVA.

▪ However, if any of these assumptions is seriously violated, the results from ANOVA
cannot be relied on.

▪ Some techniques (e.g., variance-stabilizing transformation, randomization test) may be

used to deal with such problems.

29
 Prepared by: Dr. Shazarel

References:

1. Montgomery, D. C., & Runger, G. C. (2014). Applied statistics and

probability for engineers. John Wiley and Sons.
2. Montgomery, D. C. (2017). Design and analysis of experiments. John
wiley & sons.

30
 Prepared by: Dr. Shazarel

thinkSTATISTICS

31
 Prepared by: Dr. Shazarel

THANK
YOU

32