Documente Academic
Documente Profesional
Documente Cultură
Mario Vianello
vianello.clm@tin.it
PRODUCT
QUALITY DESIGN
01OFHLO
These slides can not be reproduced or distributed without permission in writing from the author. 1
COURSE CONTENTS
These slides can not be reproduced or distributed without permission in writing from the author. 2
PRODUCT QUALITY DESIGN
2
FUNDAMENTALS
OF APPLIED STATISTICS
These slides can not be reproduced or distributed without permission in writing from the author. 3
2. Fundamentals of Applied Statistics
Presentation
These slides can not be reproduced or distributed without permission in writing from the author. 4
CONTENTS OF THIS CHAPTER
INTRODUCTION
- Population, elements, characters, characteristics
- The three main Chapters of Statistics
DESCRIPTIVE STATISTICS
- Histogram
- Cumulative curves and frequency density curves
- Bar graph (differences with the histogram)
- Prediction of defective percentage
- Pareto’s principle of “factor sparsity”
- Measures of Central Location and Measures of Dispersion
- Measures of Central Location: mean, median, mode
- Measures of Dispersion: range, average deviation, standard deviation
- Gauss Normal distribution and examples of applications
- Normal Probability Plot (NPP)
- Limits of Descriptive Statistics estimations
PROBABILITY THEORY
- Sampling
- Basic definitions and relationships about probability
- Examples of probability calculations
These slides can not be reproduced or distributed without permission in writing from the author. 5
CONTENTS OF THIS CHAPTER
2.1
I N T R O D U C T I O N
These slides can not be reproduced or distributed without permission in writing from the author. 7
2.1. Introduction
These slides can not be reproduced or distributed without permission in writing from the author. 8
2.1. Introduction
NEED OF A STATISTICAL APPROACH
Let’s suppose we want to know the mean life
of a lot of 10,000 bulbs.
These slides can not be reproduced or distributed without permission in writing from the author. 10
2.1. Introduction
These slides can not be reproduced or distributed without permission in writing from the author. 11
2.1. Introduction
2.1.1
Population, elements,
characters, characteristics
These slides can not be reproduced or distributed without permission in writing from the author. 12
2.1.1. Population, elements, characters, characteristics
Some words of current language
have a special meaning in STATISTICS:
• a statistical population (e.g. the lot of 10,000 bulbs) is done
by a set of elements (e.g. each of the 10,000 bulbs),
• these elements have one or more common characters,
which are the features by which we can determine whether
or not each element belongs to that population (e.g. the pro-
duction and packaging conditions that make the bulbs belonging
to the above lot).
Every element of a population has also some peculiar characteristics,
measurable or not and usually different from each other,
which are the object of inquiry (e.g. bulb lifetime).
These slides can not be reproduced or distributed without permission in writing from the author. 13
2.1.1. Population, elements, characters, characteristics
These slides can not be reproduced or distributed without permission in writing from the author. 14
2.1. Introduction
2.1.2
The three main Chapters
of Statistics
These slides can not be reproduced or distributed without permission in writing from the author. 15
2.1.2. The three main Chapters of Statistics
STATISTICS is divided into 3 big CHAPTERS
2.2
D E S C R I P T I V E
S T A T I S T I C S
These slides can not be reproduced or distributed without permission in writing from the author. 17
2.2. Descriptive Statistics
• building a histogram;
• calculating the mean and/or other position measurements;
• calculating standard deviation and/or other dispersion
measurements.
These slides can not be reproduced or distributed without permission in writing from the author. 19
2.2. Descriptive Statistics
2.2.1
Histogram
These slides can not be reproduced or distributed without permission in writing from the author. 20
2.2.1. Histogram
Histogram - Definitions
Frequency
Class
Width
Frequency
Characteristic
Limits of the class
±0,10
[ mm ]
Ø 5
1 2 3 4 5
5.06 5.00 4.95 5.00 5.05
DIAMETERS MEASURED ON 80 ROLLS MADE BY A LATHE
5.05 2 5.02
5,05 4.94
5,02 4,945.01 5,01 4.95 4,95 minimum = 4,87
4.94 3 4.99
4,94 5.03
4,99 5,035.04 5,04 4.95 4,95 minimum = 4.87
5.03 4 5.00
5,03 4.88
5,00 4,885.04 5,04 5.02 5,02 maximum = 5,15
4.99 5 4,99
5.04 5,04
5.01 5,015.06 5,06 5.02 5,02 maximum = 5.15
4.98 6 4,98
4.95 4,95
5.07 5,075.03 5,03 4.95 4,95 range = 0,28
5.05 7 5,05
5.01 5,01
5.04 5,045.15 5,15 4.98 4,98 range = 0.28
4.94 8 4,94
4.97 4,97
4.97 4,975.02 5,02 5.02 5,02
9 4,99 5,02 4,99 4,94 5,03
4.99 5.02 4.99 4.94 5.03
10 5,07 5,03 5,03 4,97 5,09
5.07 5.03 5.03 4.97 5.09
11 5,02 4,95 5,06 4,90 4,96 mean = 4,99988
5.02 4.95 5.06 4.90 4.96
12 4,99 4,99 5,03 4,87 5,05
4.9913 4.99
4,97 5.03
5,02 5,014.87 5,08 5.05 4,91 mean= =
st.dev
4.99988
0,049 5.000
4.9714 5.02
4,97 5.01
5,02 5,015.08 4,97 4.91 5,01
4.9715 5.02
4,94 5.01
5,00 5,004.97 5,03 5.01 4,93 st. dev. ==
median
0.04887 5,01 0.049
4.9416 5,03
5.00 4,93
5.00 5,015.03 4,96 4.93 5,01
5.03 4.93 5.01 4.96 5.01
These slides can not be reproduced or distributed without permission in writing from the author. 22
2.2.1. Histogram
Building a Histogram
1. Collect data
2. Calculate the data Range: R = (Xmax-Xmin)
3. Define the number of classes Nc
4. Calculate the classes width (rounding the value, if
necessary): h = R/Nc
These slides can not be reproduced or distributed without permission in writing from the author. 23
2.2.1. Histogram The number of
Numero di classi INSUFFICIENTE, per es. 3 classi
TOO FEW classes (3 classes)
classes must be
chosen in a very
smart way
The ideal number of classes is
Frequency
Frequenza
Good information
(showing a clearly
symmetrical distribution)
Frequency
Frequency
Frequenza
Frequenza
Real blanks
or few data ?
Caratteristica
Characteristic Caratteristica
Characteristic
These slides can not be reproduced or distributed without permission in writing from the author. 24
2.2.1. Histogram
To build
To a histogram
construct it’s always
a histogram recommended
it is always to follow
advisable to use the
thefollowing
followingrules:
rules.
The number of classes Nc must be linked to the number n of data, in order to avoid n Nc
too many blank classes because of fewer data. The following Sturges semiempirical 5 3
relationship is often used: 10 4
Nc
10
= 1 + -------
3
.
Log10 n
(*) In the Table on the right, the minimum
values of n to increase Nc are showed. 12
23
5
6
(*) Instead of Sturges relationship, it is possible to use the simpler (but not so good) expression: Nc = n
These slides can not be reproduced or distributed without permission in writing from the author. 25
2.2.1. Histogram
classes frequency
%
# from to centre check off absolute % cumulative
These slides can not be reproduced or distributed without permission in writing from the author. 26
2.2.1. Histogram
Range of Variability
Expected value
These slides can not be reproduced or distributed without permission in writing from the author. 27
2.2.1. Histogram
These slides can not be reproduced or distributed without permission in writing from the author. 28
2.2.1. Histogram
Irregular Shapes
"island“ shape
F
Look for causes of the few
isolated data using stratification
Island
X
“tooth comb“ shape
F There is something strange in
the classes intervals:
Too many classes
compared with the
number of data
Wrong rounding of
measures
Errors
X
These slides can not be reproduced or distributed without permission in writing from the author. 29
2.2.1. Histogram
Irregular Shapes
“precipice“ shape
F
Measurement errors or distortes
values (e.g. due to the tool
used or to missing/rejected
data)
F X
"asymmetric“ shape
X
These slides can not be reproduced or distributed without permission in writing from the author. 30
2.2.1. Histogram
F
"Bimodal"
Merging of non
homogenous data
+
X X
These slides can not be reproduced or distributed without permission in writing from the author. 31
2.2.1. Histogram
Skewness
Histogram of Skewed to the left Histogram of Skewed to the right
140 250
120
200
100
150
Frequency
Frequency
80
60
100
40
50
20
0 0
70 80 90 100 110 120 130 0 12 24 36 48 60 72
Skewed to the left Skewed to the right
Skewed distribution (to the left) are Skewed distribution (to the right) may
typically found when measuring e.g. occur because the process has a physical
times. This because there is usually a upper limit or only has an upper spec limit.
natural limit to how fast a process can be
completed.
These slides can not be reproduced or distributed without permission in writing from the author. 32
2.2. Descriptive Statistics
2.2.2
Cumulative curves and
frequency density curves
These slides can not be reproduced or distributed without permission in writing from the author. 33
2.2.2. Cumulative curves and frequency density curves
These slides can not be reproduced or distributed without permission in writing from the author. 34
2.2.2. Cumulative curves and frequency density curves
Cumulative frequency
These slides can not be reproduced or distributed without permission in writing from the author. 35
2.2.2. Cumulative curves and frequency density curves
These slides can not be reproduced or distributed without permission in writing from the author. 36
2.2.2. Cumulative curves and frequency density curves
60%
line respect to the histogram) have
to be equivalent.
50%
40% 37,50%
30% 26,25%
27,50%
20%
10%
3,75%
2,50%
1,25% 1,25%
0%
4,750 4,775 4,800 4,850 4,900 4,950 5,000 5,050 5,100 5,150 5,200 5,225 5,250
DIAMETERS [mm]
These slides can not be reproduced or distributed without permission in writing from the author. 37
2.2.2. Cumulative curves and frequency density curves
50% 50%
40% 37,50% 8%
= = 10%
% OF ROLLS IN EACH CLASS
FREQUENCY DENSITY
30% 27,50% 6%
26,25%
25% 5%
20% 4%
15% 3%
10% 2%
5% 3,75% 1%
2,50%
1,25% 1,25%
0% 0%
485 490 495 500 505 510 515
ROLL DIAMETER [hundredths of a millimeter]
These slides can not be reproduced or distributed without permission in writing from the author. 39
2.2.2. Cumulative curves and frequency density curves
corresponding to an interval
60% 1200%
between two preset limits, can be
almost always approximated by a
50% trapezoid (or 2 if useful) and 1000%
means the probability to find a
40% diameter of a value inside the 37,50% 800%
considered interval.
30% 26,25%
27,50% 600%
20% 400%
10% 200%
3,75%
2,50%
1,25% 1,25%
0% 0%
4,750 4,775 4,800 4,850 4,900 4,950 5,000 5,050 5,100 5,150 5,200 5,225 5,250
DIAMETERS [mm]
These slides can not be reproduced or distributed without permission in writing from the author. 40
2.2.2. Cumulative curves and frequency density curves
These slides can not be reproduced or distributed without permission in writing from the author. 41
2.2.2. Cumulative curves and frequency density curves
These slides can not be reproduced or distributed without permission in writing from the author. 42
2.2.2. Cumulative curves and frequency density curves
Frequency density distribution
The distribution of the heights of a population
(including children) can never have a value less Adult males
0,045
than zero, while it can be “open” on the right, for (only)
possible “giants”. As a consequence it will be Maschi adulti
mean = 171 cm
0,040 dissimmetric and so not normal. (soltanto)
approximately
media = 171 cm
Of course, if we limit ourselves to consider a mo- normal
distribuzione
0,035
DI PROBABILITÀ
0,025
FREQUENCY
WholePopolazione
population
0,020
completa
(children included)
DENSITÀ
(inclusi
mean bambini)
= 147 cm
0,015
non-normalcm
media = 147
distribuzione
0,010 distribution
non-normale
0,005
0,000
0 20 40 60 80 100 120 140 160 180 200 220 240
HEIGHTs [cm]
ALTEZZE [cm]
These slides can not be reproduced or distributed without permission in writing from the author. 43
2.2.2. Cumulative curves and frequency density curves
Cumulative distribution
1,0
0,9
maschi adulti
adult males
0,8 whole population
popolazione completa
CUMULATA
FREQUENCY
0,7
These two (cumulative) distributions
0,6
clearly represent two very different
populations, but both have the same
PROBABILITÀ
0,5
s-shape, which does not highlight
CUMATIVE
0,4
the differences in meaning as well as
the frequency density distributions of
0,3 the previous slide.
0,2
0,1
0,0
0 20 40 60 80 100 120 140 160 180 200 220 240
ALTEZZE
HEIGHTs [cm]
These slides can not be reproduced or distributed without permission in writing from the author. 44
2.2.2. Cumulative curves and frequency density curves
Therefore:
cumulative distributions represent the trend of calcula-
tions, but they are all S-shaped curves, very similar to
each other, and therefore unable to give an immediate
idea of the distribution of the population’s characteristic;
on the opposite, frequency density distributions (which
are the derivatives of the previous ones) have a shape
similar to the histogram and give an immediate graphical
perception of how the investigated characteristic is di-
stributed in the population;
the frequency density distributions can be not symme-
trical and have very different shapes.
These slides can not be reproduced or distributed without permission in writing from the author. 45
2.2.2. Cumulative curves and frequency density curves
Examples of frequency density distribution
Red curves have shapes obtainable by Weibull distribution
DECREASING
SYMMETRICAL
(e.g.: Normal )
RIGHT-SKEWED
RECTANGULAR
or UNIFORM
With a constant
failure rate
costante
These slides can not be reproduced or distributed without permission in writing from the author. 46
2.2. Descriptive Statistics
2.2.3
Bar graph
(differences with the histogram)
These slides can not be reproduced or distributed without permission in writing from the author. 47
2.2.3. Bar graph (differences with the histogram)
These slides can not be reproduced or distributed without permission in writing from the author. 48
2.2.3. Bar graph (differences with the histogram)
But a class distribution can be also used in some other application,
not seen before and which do not respect the 3 basic conditions.
In these cases, the columns become simply bars, and they are repre-
sented one far from the other, to remember that a following interpolation
by a frequency density curve would not have any meaning and the dia-
gram loses the name of histogram to assume that of bar graph.
50%
45%
40% For instance, a bar graph
40%
like that on the left indicates
35%
30% only the different volumes
Percentages
Percentuali
30%
25%
25% of supply in 2004: an inter-
20%
polation by a frequency
15%
density curve does not
10%
have any meaning.
5%
5%
0%
Forn. A A
Supplier Forn. B
Supplier B Forn. C
Supplier C Altri
Others
Forniture 2004
SUPPLIES 2004
These slides can not be reproduced or distributed without permission in writing from the author. 49
2.2. Descriptive Statistics
2.2.4
Evaluation of
defectives percentage
These slides can not be reproduced or distributed without permission in writing from the author. 50
2.2.4. Evaluation of defectives percentage
These slides can not be reproduced or distributed without permission in writing from the author. 51
2.2.4. Evaluation of defectives percentage
If we can accept the hypothesis that
the histogram closely represents the population and
that the situation remains unchanged over time,
then, with the help of one of the many software available on the market,
we can estimate the percentage of defectives as the sum of the areas
of the 2 tails outside the tolerance range (4.90 and 5.10) corresponding
to 4.10%. 40 6
Prescribedprescritte
Tolleranze tolerance
(histogram)
4
CLASSI
25
IN EACH
PEZZI NELLE
20 3
OFDIITEMS
15
2
Predicted Predicted
NUMERO
10 Previsione Previsione
defectives
frazione half defectives
frazionehalf
NUMBER
dipercentage
non conformi percentage
di non conformi
1
5
0 0
4,70 4,75 4,80 4,85 4,90 4,95 5,00 5,05 5,10 5,15 5,20 5,25 5,30
QUOTE MISURATE [ [mm]
MEASUREMENTS mm ]
These slides can not be reproduced or distributed without permission in writing from the author. 52
2.2.4. Evaluation of defectives percentage
Also if all the items of the sample are inside the tolerance range,
by the Statistics, it is possible to estimate the percent of defectives
which will be present in the whole production in the field.
This is a very big benefit, mostly when we have a very little percentage of defectives, in
which case, to find defectives, it would require to sample an unacceptable amounts of items.
These slides can not be reproduced or distributed without permission in writing from the author. 53
2.2. Descriptive Statistics
2.2.5
The Pareto principle
of “factor sparsity”
These slides can not be reproduced or distributed without permission in writing from the author. 54
2.2.5. The Pareto principle of “factor sparsity”
“Vital few”
really important things are very few
I M P A C T
F A I L U R E S specific cumulative
CUMULATIVE % IMPACT
I M P A C T ( number of failures )
5 :: 99,99%
10% 1029 :::::::::::::::::::::::::::::::::::: 0,02 0,01% 150 100,00%
Total 150 100,00%
0 0
External rear wiew mirror
Pipes/sleeve fastening
Headlamps
Seat rails
Windscreen wiper
Others
These slides can not be reproduced or distributed without permission in writing from the author. 56
2.2.5. The Pareto principle of “factor sparsity”
These slides can not be reproduced or distributed without permission in writing from the author. 57
2.2. Descriptive Statistics
2.2.6
Position and spread
measurements
These slides can not be reproduced or distributed without permission in writing from the author. 58
2.2.6. Position and spread measurements
These slides can not be reproduced or distributed without permission in writing from the author. 59
2.2. Descriptive Statistics
2.2.7
Position measurements:
mean, median, mode
These slides can not be reproduced or distributed without permission in writing from the author. 60
2.2.7. Position measurements: mean, median, mode
In a value distribution:
The mean is the most known and widely used among position
measurements. It is the value obtained by the ratio between the sum
of data values and total number of data (arithmetic average). Its value
represents the center of gravity of all values.
The mode is the value (or values: if the distribution has more than
one mode) where frequency (or probability) density has its maximum:
in other words, it corresponds to the maximum(s) of the frequency (or
probability) density distribution.
These slides can not be reproduced or distributed without permission in writing from the author. 61
2.2.7. Position measurements: mean, median, mode
These slides can not be reproduced or distributed without permission in writing from the author. 62
2.2.7. Position measurements: mean, median, mode
The classic “position measure” is the MEAN
DESCRIPTIVE Statistics
TRUE S xi
when we examine all
elements of a
m = ---------
population n
MEAN
INFERENCIAL Statistics
ESTIMATED S xi
when we examine
only a sample taken
m
^ = x = --------
from the population n
These slides can not be reproduced or distributed without permission in writing from the author. 63
2.2.7. Position measurements: mean, median, mode
(1) If all the values appear the same number of times (i.e. with the same frequency), the mode does not exist,
but if there are more different values with the maximum frequency, then there are more than one mode.
These slides can not be reproduced or distributed without permission in writing from the author. 64
2.2.7. Position measurements: mean, median, mode
On the opposite, the mean uses all the true values of data.
We can illustrate this concept saying that the mean has a
greater density of information than the median.
These slides can not be reproduced or distributed without permission in writing from the author. 65
2.2.7. Position measurements: mean, median, mode
These slides can not be reproduced or distributed without permission in writing from the author. 66
2.2.7. Position measurements: mean, median, mode
Positioning with continuous distributions
With symmetric
distributions
(e.g. normal distribution)
mean, median and mode
These slides can not be reproduced or distributed without permission in writing from the author. 67
2.2.7. Position measurements: mean, median, mode
Positioning with continuous distributions
With asymmetric distributions,
median
mean, median and mode
mode are always arranged
mean
in the order shown in the left
(respect to the skewed portion
which causes asymmetry).
If the distribution would be skewed in the right
(and not in the left as in Figure),
mean, median and mode
would be arranged in the reverse order.
50% 50%
These slides can not be reproduced or distributed without permission in writing from the author. 68
2.2.7. Position measurements: mean, median, mode
These slides can not be reproduced or distributed without permission in writing from the author. 69
MEAN
Left Right
1x4= 4 1x5= 5
2x3= 6 1x4= 4
3x2= 6 2x3= 6
5x1= 5 2x2= 4
2x1= 2
21 21
to the
These slides can not be reproduced or distributed without permission in writing from the author. 70
2.2.7. Position measurements: mean, median, mode
These slides can not be reproduced or distributed without permission in writing from the author. 71
EXAMPLES OF APPLICATION OF MEDIAN AND PERCENTILES
Selection CLOTHES SIZE CRASH test impact against barrier
70
60
forn.1
forn.2
forn.3
forn.4
forn.5
These slides can not be reproduced or distributed without permission in writing from the author. 73
2.2. Descriptive Statistics
2.2.8
Spread measurements:
Range, average deviation
and standard deviation
These slides can not be reproduced or distributed without permission in writing from the author. 74
2.2.8. Spread measurements: Range, average deviation and standard deviation
Examples :
if the depth of a river is on average less than 1 meter, this
does not exclude the presence of points with dangerous
depth !
two towns with the same average annual temperature of
18°C can have ranges winter-summer very different, e.g.
one from -10°C to +35°C and the other from +14°C to
+24°C: two climates completely different !
These slides can not be reproduced or distributed without permission in writing from the author. 75
2.2.8. Spread measurements: Range, average deviation and standard deviation
These slides can not be reproduced or distributed without permission in writing from the author. 76
DESCRIPTIVE Statistics 2.2.8. Spread measurements: Range, average deviation and standard deviation
TRUE (1)
Standard
deviation
Statistical INFERENCE
ESTIMATED (2)
These slides can not be reproduced or distributed without permission in writing from the author. 77
2.2.8. Spread measurements: Range, average deviation and standard deviation
These slides can not be reproduced or distributed without permission in writing from the author. 78
2.2.8. Spread measurements: Range, average deviation and standard deviation
V. To use n - 1 as denominator means to recognize that, to estimate the popu-
lation variability, we need a sample of at least 2 elements (which is the mini-
mum to make a comparison), otherwise the denominator is zero.
VI. If we want to describe expression (2) using the concept of degrees of free-
dom (abbreviated as dof: see Par. 4.3.5.2, introduction to ANOVA), we can consi-
der standard deviation as the square root of the ratio between the sum of
the squares of deviations and the degrees of freedom. Thus we can say
that the degrees of freedom are n - 1 , because one degrees of freedom was
spent in calculating the mean x. In other words, as we have before evaluated
the mean, we have reduced the possibility of variation of the observations as
they was n-1 and not n.
VII. If we apply expression (1) to the data of a sample to calculate its standard de-
viation, s, to be used as an estimate of the true standard deviation of the whole
population, st, we would use a biased estimator, which, with the increase of
the number of tested samples (all of size n), converges (see next slide) towards
a value sb smaller than the true value st, according to the following relationship:
sb = √ (n-1) / n . st
Of course the problem is reduced if the size n of the sample is large, because
the ratio indicated above becomes more and more close to 1 when n grows.
The mathematical justification of this will be provided few slides later (under the title DEMONSTRA-
TION OF THE CORRECTION FOR AN “UNBIASED” ESTIMATOR), immediately after the presenta-
tion of two new expressions as an alternative to the previous expressions (1) and (2). For the mo-
ment, we limit to the graphical illustration shown in the following slide.
These slides can not be reproduced or distributed without permission in writing from the author. 79
2.2.8. Spread measurements: Range, average deviation and standard deviation
Graphical illustration of the previous Point VII
Let’s consider small samples of assigned size n (e.g.: n = 5) and let's start by extracting a sample
from the population and then by calculating its standard deviation s1 with the expression (1), i.e.
dividing by n (instead of by n-1). Let’s extract a second sample and calculate, for it too, the stan-
dard deviation s2, with the expression (1) as before, and then calculate the mean of s1 and s2.
We can continue in this way, extracting many other samples, always calculating the standard de-
viation of the last sample and then averaging it with those of all previous samples. As the number
of samples increases, the average value of the standard deviations of all samples converges to-
wards its limit, which can be either:
the true value, st, if standard deviations were calculated using the expression (2);
a biased value, sb, with a bias related to the ratio (n-1)/n of the previous slide, if standard de-
viations were calculated with the expression (1): this latter aspect is shown below.
s biased s true
bias
These slides can not be reproduced or distributed without permission in writing from the author. 80
2.2.8. Spread measurements: Range, average deviation and standard deviation
The numerator S(xi - m)2 of previous expression (1) can also be
rewritten in the following way:
n n n n n
S(xi - m)2 = S(xi2 - 2 m xi + m2) = Sxi2 - 2 m Sxi + n m2 = Sxi2 - 2 m n m + n m2 =
i=1 i=1 i=1 i=1 i=1
remembering the
definition of mean
replacing m with
his expression
With the same steps, we can write in a similar way also the expres-
sion (2), where, instead of m, there is x.
These slides can not be reproduced or distributed without permission in writing from the author. 81
2.2.8. Spread measurements: Range, average deviation and standard deviation
Replacing into (1) and (2) the numerators in the form just elaborated,
we obtain the following formulas equivalent to the preceding:
n . S (xi2) - (S xi)2
s = ------------------------ (1 bis)
n2
n . S(xi2) - (S xi)2
s
^ = s = ----------------------- (2 bis)
n . (n - 1)
These slides can not be reproduced or distributed without permission in writing from the author. 82
2.2.8. Spread measurements: Range, average deviation and standard deviation
These slides can not be reproduced or distributed without permission in writing from the author. 84
2.2.8. Spread measurements: Range, average deviation and standard deviation
These slides can not be reproduced or distributed without permission in writing from the author. 85
2.2.8. Spread measurements: Range, average deviation and standard deviation
Regardless of the square root sign, we can rewrite (1 bis) in terms of statistical inference:
n n n n
S xi 2 S xi S xi 2 S xi 2 S x i xj
i=1 i=1 i=1 i=1 i<j
s2 = --------- - (-------)2 = --------- - -------- - 2 ------------ =
n n n n2 n2
n n
n -1 S xi xj n-1 n - 1 S xi 2 n-1
n i<j n i=1
= -------- . S xi2 - 2 (-----------)2 = -------- . S xi2 = ------- . ------- = -------- . s2
i=1 i=1
n 2 n 2 n 2 n n n
a b
a) As we suppose that the n observations are independent (i.e. not influencing each other),
n
we can demonstrate that S xi xj = 0. Indeed, given two discrete random variable X e Y
i< j
with means respectively mx e my, we can define la covariance Kxy as:
Kxy = S S (xi - mx) . (yj - my) . pxi pyj = S (xi - mx) pxi . S (yj - my) pyj
i j i j
where pxi and pyj are the probabilities related to the values xi and yj and the sums are
extended to all values where random variables are defined. In this situation, if random
variables are independent (what means that the pyj are not influenced by the pxi and vice
versa), the sum S (yj - my) pyj is null by definition.
b) As we assumed that the origin of the values of observations is placed at the true value of
the mean, m, it is by definition: s2 = (S xi2)/n.
These slides can not be reproduced or distributed without permission in writing from the author. 86
2.2.8. Spread measurements: Range, average deviation and standard deviation
As the median, also the range, of all the data values uses:
• only the extreme values (minimum and maximum observed);
• of the other values, it takes only the position in order to
establish the extreme values.
These slides can not be reproduced or distributed without permission in writing from the author. 87
2.2.8. Spread measurements: Range, average deviation and standard deviation
We wonder why we do not consider the average
deviation, defined as the mean of differences between
every value and their mean:
d = ( S |xi - x| ) / n
Of course we have to consider absolute values, other-
wise (by definition of mean) this sum would be always
equal to zero.
We can realize the
reason for this, for
example considering
the linear interpola-
tion of 3 not aligned
points.
These slides can not be reproduced or distributed without permission in writing from the author. 88
2.2.8. Spread measurements: Range, average deviation and standard deviation
Which of the following two alternatives is preferable to consider as the best
interpolating line ?
the line which minimizes the sum of distances in absolute value (average
deviation)
or the line which minimizes the sum of squares of distances (standard
deviation) ?
On this base, the old rule to remove the extreme (minimum &
maximum) values of a sample data must be reversed.
“Variability” acquires primary importance and therefore extre-
me values, just as representative of relatively infrequent events,
can not be eliminated because they provide valuable information
about variability.
These slides can not be reproduced or distributed without permission in writing from the author. 90
2.2. Descriptive Statistics
2.2.9
Gauss Normal Distribution
These slides can not be reproduced or distributed without permission in writing from the author. 91
2.2.9. Gauss Normal Distribution
2.2.9.1
Meaning
These slides can not be reproduced or distributed without permission in writing from the author. 92
2.2.9.1. Gauss Normal Distribution: MEANING
These slides can not be reproduced or distributed without permission in writing from the author. 93
2.2.9.1. Gauss Normal Distribution: MEANING
These slides can not be reproduced or distributed without permission in writing from the author. 94
2.2.9.1. Gauss Normal Distribution: MEANING
These slides can not be reproduced or distributed without permission in writing from the author. 95
2.2.9.1. Gauss Normal Distribution: MEANING
• values obtained with the launch of two dices (with a little straining
for lack of continuity).
These slides can not be reproduced or distributed without permission in writing from the author. 96
2.2.9.1. Gauss Normal Distribution: MEANING
These slides can not be reproduced or distributed without permission in writing from the author. 97
2.2.9.1. Gauss Normal Distribution: MEANING
The meaning and the importance of normal distribution
can be better understood considering
the sum of the values obtained by the launch of 2 dice
POSSIBILI
sommasum
6
POSSIBLE
with the same
5
5
con la medesima
4 4
ESITI
DI OF
3
3
NUMBER
OUTCOMES
2 2
NUMERO
1
1
2 3 4 5 6 7 8 9 10 11 12
SUM OF THE
SOMMA VALUES
DELLE OF DI
FACCE 2 dice: possible
2 DADI: esitioutcomes
possibili
These slides can not be reproduced or distributed without permission in writing from the author. 98
2.2.9.1. Gauss Normal Distribution: MEANING
27 27
25 25 0,12
25
of combinations
21 21
0,10
frequenza
Densitydifrequency
20
N° combinazioni
0,08
15 15
15
Densità
0,06
Number
10 10
10
0,04
6 6
5
3 3 0,02
1 1
0 0,00
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
SumSomma
of the outcomes of 3the
punteggi dei 3 dice
dadi
These slides can not be reproduced or distributed without permission in writing from the author. 99
2.2.9.1. Gauss Normal Distribution: MEANING
A complex phenomenon,
as the behaviour of a machinery,
can be thought as the result
of the combination of a very large number
of elementary random phenomena,
similar to the sum of the outcomes
of the launch of several dice.
These slides can not be reproduced or distributed without permission in writing from the author. 100
2.2.9.1. Gauss Normal Distribution: MEANING
Mathematical expression of
CUMULATIVE NORMAL DISTRIBUTION
x
(x - m)2
- ------------
1 2 . s2
F(x) = ----------------- .
e dx
s. 2. -
These slides can not be reproduced or distributed without permission in writing from the author. 101
2.2.9.1. Gauss Normal Distribution: MEANING
2 parameters distribution:
mean, m, which places the (mean of) distribution along the abscissa
axis, without changing its shape;
standard deviation, s, which gives to the bell (always symmetrical),
a shape large and flat (if it has high values) or high and narrow (if it has
low values).
These slides can not be reproduced or distributed without permission in writing from the author. 102
2.2.9.1. Gauss Normal Distribution: MEANING
Mathematical expression of
DENSITY NORMAL DISTRIBUTION
The density frequency distribution is the derivatives of the previous
cumulative distribution and is obtained from it simply removing the in-
tegral sign:
(x - m)2
- -------------
1 2 . s2
f(x) = ------------------ . e
s. 2.
These slides can not be reproduced or distributed without permission in writing from the author. 103
2.2.9. Gauss Normal Distribution
2.2.9.2
Standard
normal distribution
These slides can not be reproduced or distributed without permission in writing from the author. 104
2.2.9.2. STANDARD normal distribution
F(x) = ------------ .
e -(x /2)
dx
2. -
These slides can not be reproduced or distributed without permission in writing from the author. 105
2.2.9.2. STANDARD normal distribution
x2
- -----
1 2
f(x) = ------------ . e
2.
These slides can not be reproduced or distributed without permission in writing from the author. 106
2.2.9.2. STANDARD normal distribution
0,35
0,30
0,25
0,20
0,15
0,10
0,05
0,00
-4 -3 -2 -1 0 1 2 3 4
± 1 s 68.26% of data
± 2 s 95.45% of data
± 3 s 99.73% of data
These slides can not be reproduced or distributed without permission in writing from the author. 107
2.2.9.2. STANDARD normal distribution
It seems particularly meaningful to remember that
Measurement Theory provides the following definitions.
• Likely errors: errors in the range mean ± 1 s.
These errors are generally done and accep-
ted. The range includes 68.26% of data. Just to
get a visual idea, in weighing by means of a Chemist’s
balance, the “likely error" in the reading is mainly due to
a parallax error and it is usually assumed equal to half
the interval between two successive nicks on the scale.
These slides can not be reproduced or distributed without permission in writing from the author. 108
2.2.9.2. STANDARD normal distribution
These slides can not be reproduced or distributed without permission in writing from the author. 109
2.2.9.2. STANDARD normal distribution
Table of standard
normal distribution z
1-F(z)
z 0,00 0,01 0,02 0,03 0,04 0,05 0,06 0,07 0,08 0,09
0,0 0,500000 0,496011 0,492022 0,488033 0,484047 0,480061 0,476078 0,472097 0,468119 0,464144
0,1 0,460172 0,456205 0,452242 0,448283 0,444330 0,440382 0,436441 0,432505 0,428576 0,424655
0,2 0,420740 0,416834 0,412936 0,409046 0,405165 0,401294 0,397432 0,393580 0,389739 0,385908
0,3 0,382089 0,378281 0,374484 0,370700 0,366928 0,363169 0,359424 0,355691 0,351973 0,348268
0,4 0,344578 0,340903 0,337243 0,333598 0,329969 0,326355 0,322758 0,319178 0,315614 0,312067
0,5 0,308538 0,305026 0,301532 0,298056 0,294598 0,291160 0,287740 0,284339 0,280957 0,277595
0,6 0,274253 0,270931 0,267629 0,264347 0,261086 0,257846 0,254627 0,251429 0,248252 0,245097
0,7 0,241964 0,238852 0,235762 0,232695 0,229650 0,226627 0,223627 0,220650 0,217695 0,214764
0,8 0,211855 0,208970 0,206108 0,203269 0,200454 0,197662 0,194894 0,192150 0,189430 0,186733
0,9 0,184060 0,181411 0,178786 0,176186 0,173609 0,171056 0,168528 0,166023 0,163543 0,161087
1,0 0,158655 0,156248 0,153864 0,151505 0,149170 0,146859 0,144572 0,142310 0,140071 0,137857
1,1 0,135666 0,133500 0,131357 0,129238 0,127143 0,125072 0,123024 0,121001 0,119000 0,117023
1,2 0,115070 0,113140 0,111233 0,109349 0,107488 0,105650 0,103835 0,102042 0,100273 0,098525
1,3 0,096801 0,095098 0,093418 0,091759 0,090123 0,088508 0,086915 0,085344 0,083793 0,082264
1,4 0,080757 0,079270 0,077804 0,076359 0,074934 0,073529 0,072145 0,070781 0,069437 0,068112
1,5 0,066807 0,065522 0,064256 0,063008 0,061780 0,060571 0,059380 0,058208 0,057053 0,055917
1,6 0,054799 0,053699 0,052616 0,051551 0,050503 0,049471 0,048457 0,047460 0,046479 0,045514
1,7 0,044565 0,043633 0,042716 0,041815 0,040929 0,040059 0,039204 0,038364 0,037538 0,036727
1,8 0,035930 0,035148 0,034379 0,033625 0,032884 0,032157 0,031443 0,030742 0,030054 0,029379
1,9 0,028716 0,028067 0,027429 0,026803 0,026190 0,025588 0,024998 0,024419 0,023852 0,023295
2,0 0,022750 0,022216 0,021692 0,021178 0,020675 0,020182 0,019699 0,019226 0,018763 0,018309
2,1 0,017864 0,017429 0,017003 0,016586 0,016177 0,015778 0,015386 0,015003 0,014629 0,014262
2,2 0,013903 0,013553 0,013209 0,012874 0,012545 0,012224 0,011911 0,011604 0,011304 0,011011
2,3 0,010724 0,010444 0,010170 0,009903 0,009642 0,009387 0,009137 0,008894 0,008656 0,008424
2,4 0,008198 0,007976 0,007760 0,007549 0,007344 0,007143 0,006947 0,006756 0,006569 0,006387
2,5 0,006210 0,006037 0,005868 0,005703 0,005543 0,005386 0,005234 0,005085 0,004940 0,004799
2,6 0,004661 0,004527 0,004397 0,004269 0,004145 0,004025 0,003907 0,003793 0,003681 0,003573
2,7 0,003467 0,003364 0,003264 0,003167 0,003072 0,002980 0,002890 0,002803 0,002718 0,002635
2,8 0,002555 0,002477 0,002401 0,002327 0,002256 0,002186 0,002118 0,002052 0,001988 0,001926
2,9 0,001866 0,001807 0,001750 0,001695 0,001641 0,001589 0,001538 0,001489 0,001441 0,001395
3,0not be 1,350E-03
These slides can reproduced or distributed without permission in writing from the author. 110
Other Table of
standard
normal distribution
These slides can not be reproduced or distributed without permission in writing from the author. 111
2.2.9.2. STANDARD normal distribution
with mean m = 30
These slides can not be reproduced or distributed without permission in writing from the author. 112
x - mx - 30
z = ---------- = -----------
s 5
m=0
z
-4 -3 -2 -1 0 1 2 3 4
m = 30 s=5
x
15 20 25 30 35 40 45
These slides can not be reproduced or distributed without permission in writing from the author. 113
2.2.9.2. STANDARD normal distribution
These slides can not be reproduced or distributed without permission in writing from the author. 114
In a specific normal distribu-
tion, the distance between the
mean and a set limit, expressed in
standard deviation units z, in
order to refer to standard normal
m=0
z
-4 -3 -2 -1 0 1 2 3 4
m = 30 s=5
15 20 25 30 35 40 45
42
These slides can not be reproduced or distributed without permission in writing from the author. 115
2.2.9.2. STANDARD normal distribution
Table of standard
normal distribution z
1-F(z)
z 0,00 0,01 0,02 0,03 0,04 0,05 0,06 0,07 0,08 0,09
0,0 0,500000 0,496011 0,492022 0,488033 0,484047 0,480061 0,476078 0,472097 0,468119 0,464144
0,1 0,460172 0,456205 0,452242 0,448283 0,444330 0,440382 0,436441 0,432505 0,428576 0,424655
0,2 0,420740 0,416834 0,412936 0,409046 0,405165 0,401294 0,397432 0,393580 0,389739 0,385908
0,3 0,382089 0,378281 0,374484 0,370700 0,366928 0,363169 0,359424 0,355691 0,351973 0,348268
0,4 0,344578 0,340903 0,337243 0,333598 0,329969 0,326355 0,322758 0,319178 0,315614 0,312067
0,5 0,308538 0,305026 0,301532 0,298056 0,294598 0,291160 0,287740 0,284339 0,280957 0,277595
0,6 0,274253 0,270931 0,267629 0,264347 0,261086 0,257846 0,254627 0,251429 0,248252 0,245097
0,7 0,241964 0,238852 0,235762 0,232695 0,229650 0,226627 0,223627 0,220650 0,217695 0,214764
0,8 0,211855 0,208970 0,206108 0,203269 0,200454 0,197662 0,194894 0,192150 0,189430 0,186733
0,9 0,184060 0,181411 0,178786 0,176186 0,173609 0,171056 0,168528 0,166023 0,163543 0,161087
1,0 0,158655 0,156248 0,153864 0,151505 0,149170 0,146859 0,144572 0,142310 0,140071 0,137857
1,1 0,135666 0,133500 0,131357 0,129238 0,127143 0,125072 0,123024 0,121001 0,119000 0,117023
1,2 0,115070 0,113140 0,111233 0,109349 0,107488 0,105650 0,103835 0,102042 0,100273 0,098525
1,3 0,096801 0,095098 0,093418 0,091759 0,090123 0,088508 0,086915 0,085344 0,083793 0,082264
1,4 0,080757 0,079270 0,077804 0,076359 0,074934 0,073529 0,072145 0,070781 0,069437 0,068112
1,5 0,066807 0,065522 0,064256 0,063008 0,061780 0,060571 0,059380 0,058208 0,057053 0,055917
1,6 0,054799 0,053699 0,052616 0,051551 0,050503 0,049471 0,048457 0,047460 0,046479 0,045514
1,7
1,8
0,044565
0,035930
0,043633
0,035148
2,4
0,042716
0,034379
0,041815
0,033625
0,008198
0,040929
0,032884
0,040059
0,032157
0,039204
0,031443
0,038364
0,030742
0,037538
0,030054
0,036727
0,029379
1,9 0,028716 0,028067 0,027429 0,026803 0,026190 0,025588 0,024998 0,024419 0,023852 0,023295
2,0 0,022750 0,022216 0,021692 0,021178 0,020675 0,020182 0,019699 0,019226 0,018763 0,018309
2,1 0,017864 0,017429 0,017003 0,016586 0,016177 0,015778 0,015386 0,015003 0,014629 0,014262
2,2 0,013903 0,013553 0,013209 0,012874 0,012545 0,012224 0,011911 0,011604 0,011304 0,011011
2,3 0,010724 0,010444 0,010170 0,009903 0,009642 0,009387 0,009137 0,008894 0,008656 0,008424
2,4 0,008198 0,007976 0,007760 0,007549 0,007344 0,007143 0,006947 0,006756 0,006569 0,006387
2,5 0,006210 0,006037 0,005868 0,005703 0,005543 0,005386 0,005234 0,005085 0,004940 0,004799
2,6 0,004661 0,004527 0,004397 0,004269 0,004145 0,004025 0,003907 0,003793 0,003681 0,003573
2,7 0,003467 0,003364 0,003264 0,003167 0,003072 0,002980 0,002890 0,002803 0,002718 0,002635
2,8 0,002555 0,002477 0,002401 0,002327 0,002256 0,002186 0,002118 0,002052 0,001988 0,001926
2,9 0,001866 0,001807 0,001750 0,001695 0,001641 0,001589 0,001538 0,001489 0,001441 0,001395
3,0not be 1,350E-03
These slides can reproduced or distributed without permission in writing from the author. 116
2.2.9. Gauss Normal Distribution
2.2.9.3
Classical applications
These slides can not be reproduced or distributed without permission in writing from the author. 117
2.2.9.3. Gauss Normal Distribution: CLASSICAL APPLICATIONS
These slides can not be reproduced or distributed without permission in writing from the author. 118
2.2.9.3. Gauss Normal Distribution: CLASSICAL APPLICATIONS
The steps listed below are usually followed:
calculation of data mean and standard deviation;
constructing the histogram and superimposing a
normal distribution with the data mean and standard
deviation: if we see that a normal distribution can not
reasonably interpolate the data, we must:
• apply another interpolating function, e.g. a Weibull function
(see next Chapter);
• if there is not a satisfactory interpolating function, we can
interpolate manually on the histogram (graphical interpolation).
calculation of the percentage corresponding to the
distribution tail or to the zlim value, according to the
request; we can do that in two ways:
• by using TABLES;
• by using EXCEL functions.
These slides can not be reproduced or distributed without permission in writing from the author. 119
2.2.9.3. Gauss Normal Distribution: CLASSICAL APPLICATIONS
EXERCISE
typical problem of risk assessment
These slides can not be reproduced or distributed without permission in writing from the author. 120
2.2.9.3. Gauss Normal Distribution: CLASSICAL APPLICATIONS
First of all, let us calculate the distance z,
measured by units of standard deviations s(= 8 gr = 0.08 hg),
of the limit value (8.00 hg) from the mean (8.10 hg).
8.00 - 8.10
z = ---------------- = - 1.25
0.08
-4 -3 -2 -1 0 1 2 3 4
8,1 hg
8,0 hg
Now we have to find the tail area on the left of the 8.0 hg limit, e.g. by
using the standard normal distribution Table.
These slides can not be reproduced or distributed without permission in writing from the author. 121
2.2.9.3. Gauss Normal Distribution: CLASSICAL APPLICATIONS
Table of standard
normal distribution z
1-F(z)
z 0,00 0,01 0,02 0,03 0,04 0,05 0,06 0,07 0,08 0,09
0,0 0,500000 0,496011 0,492022 0,488033 0,484047 0,480061 0,476078 0,472097 0,468119 0,464144
0,1 0,460172 0,456205 0,452242 0,448283 0,444330 0,440382 0,436441 0,432505 0,428576 0,424655
0,2 0,420740 0,416834 0,412936 0,409046 0,405165 0,401294 0,397432 0,393580 0,389739 0,385908
0,3 0,382089 0,378281 0,374484 0,370700 0,366928 0,363169 0,359424 0,355691 0,351973 0,348268
0,4 0,344578 0,340903 0,337243 0,333598 0,329969 0,326355 0,322758 0,319178 0,315614 0,312067
0,5 0,308538 0,305026 0,301532 0,298056 0,294598 0,291160 0,287740 0,284339 0,280957 0,277595
0,6 0,274253 0,270931 0,267629 0,264347 0,261086 0,257846 0,254627 0,251429 0,248252 0,245097
0,7 0,241964 0,238852 0,235762 0,232695 0,229650 0,226627 0,223627 0,220650 0,217695 0,214764
0,8 0,211855 0,208970 0,206108 0,203269 0,200454 0,197662 0,194894 0,192150 0,189430 0,186733
0,9 0,184060 0,181411 0,178786 0,176186 0,173609 0,171056 0,168528 0,166023 0,163543 0,161087
1,0 0,158655 0,156248 0,153864 0,151505 0,149170 0,146859 0,144572 0,142310 0,140071 0,137857
1,1 0,135666 0,133500 0,131357 0,129238 0,127143 0,125072 0,123024 0,121001 0,119000 0,117023
1,2 0,115070 0,113140 0,111233 0,109349 0,107488 0,105650 0,103835 0,102042 0,100273 0,098525
1,3 0,096801 0,095098 0,093418 0,091759 0,090123 0,088508 0,086915 0,085344 0,083793 0,082264
1,4 0,080757 0,079270 0,077804 0,076359 0,074934 0,073529 0,072145 0,070781 0,069437 0,068112
1,5 0,066807 0,065522 0,064256 0,063008 0,061780 0,060571 0,059380 0,058208 0,057053 0,055917
1,6 0,054799 0,053699 0,052616 0,051551 0,050503 0,049471 0,048457 0,047460 0,046479 0,045514
1,7 0,044565 0,043633 0,042716 0,041815 0,040929 0,040059 0,039204 0,038364 0,037538 0,036727
1,8 0,035930 0,035148 0,034379 0,033625 0,032884 0,032157 0,031443 0,030742 0,030054 0,029379
1,9 0,028716 0,028067 0,027429 0,026803 0,026190 0,025588 0,024998 0,024419 0,023852 0,023295
2,0 0,022750 0,022216 0,021692 0,021178 0,020675 0,020182 0,019699 0,019226 0,018763 0,018309
2,1 0,017864 0,017429 0,017003 0,016586 0,016177 0,015778 0,015386 0,015003 0,014629 0,014262
2,2 0,013903 0,013553 0,013209 0,012874 0,012545 0,012224 0,011911 0,011604 0,011304 0,011011
2,3 0,010724 0,010444 0,010170 0,009903 0,009642 0,009387 0,009137 0,008894 0,008656 0,008424
2,4 0,008198 0,007976 0,007760 0,007549 0,007344 0,007143 0,006947 0,006756 0,006569 0,006387
2,5 0,006210 0,006037 0,005868 0,005703 0,005543 0,005386 0,005234 0,005085 0,004940 0,004799
2,6 0,004661 0,004527 0,004397 0,004269 0,004145 0,004025 0,003907 0,003793 0,003681 0,003573
2,7 0,003467 0,003364 0,003264 0,003167 0,003072 0,002980 0,002890 0,002803 0,002718 0,002635
2,8 0,002555 0,002477 0,002401 0,002327 0,002256 0,002186 0,002118 0,002052 0,001988 0,001926
2,9 0,001866 0,001807 0,001750 0,001695 0,001641 0,001589 0,001538 0,001489 0,001441 0,001395
3,0not be 1,350E-03
These slides can reproduced or distributed without permission in writing from the author. 122
2.2.9.3. Gauss Normal Distribution: CLASSICAL APPLICATIONS
These slides can not be reproduced or distributed without permission in writing from the author. 123
2.2.9.3. Gauss Normal Distribution: CLASSICAL APPLICATIONS
If we suppose to accept
no more than 2% of underweight cans,
we can ask:
• which value must have the mean, m, if we do not
change the current standard deviation (i.e. if we
keep the existing production process as it is) ?
• which value must have the standard deviation, s,
if we do not change the current mean (i.e. if we
keep the current margin from the limit) ?
z 0,00 0,01 0,02 0,03 0,04 0,05 0,06 0,07 0,08 0,09
0,0 0,500000 0,496011 0,492022 0,488033 0,484047 0,480061 0,476078 0,472097 0,468119 0,464144
0,1 0,460172 0,456205 0,452242 0,448283 0,444330 0,440382 0,436441 0,432505 0,428576 0,424655
0,2 0,420740 0,416834 0,412936 0,409046 0,405165 0,401294 0,397432 0,393580 0,389739 0,385908
0,3 0,382089 0,378281 0,374484 0,370700 0,366928 0,363169 0,359424 0,355691 0,351973 0,348268
0,4 0,344578 0,340903 0,337243 0,333598 0,329969 0,326355 0,322758 0,319178 0,315614 0,312067
0,5 0,308538 0,305026 0,301532 0,298056 0,294598 0,291160 0,287740 0,284339 0,280957 0,277595
0,6 0,274253 0,270931 0,267629 0,264347 0,261086 0,257846 0,254627 0,251429 0,248252 0,245097
0,7 0,241964 0,238852 0,235762 0,232695 0,229650 0,226627 0,223627 0,220650 0,217695 0,214764
0,8 0,211855 0,208970 0,206108 0,203269 0,200454 0,197662 0,194894 0,192150 0,189430 0,186733
0,9 0,184060 0,181411 0,178786 0,176186 0,173609 0,171056 0,168528 0,166023 0,163543 0,161087
1,0 0,158655 0,156248 0,153864 0,151505 0,149170 0,146859 0,144572 0,142310 0,140071 0,137857
1,1 0,135666 0,133500 0,131357 0,129238 0,127143 0,125072 0,123024 0,121001 0,119000 0,117023
1,2 0,115070 0,113140 0,111233 0,109349 0,107488 0,105650 0,103835 0,102042 0,100273 0,098525
1,3 0,096801 0,095098 0,093418 0,091759 0,090123 0,088508 0,086915 0,085344 0,083793 0,082264
1,4 0,080757 0,079270 0,077804 0,076359 0,074934 0,073529 0,072145 0,070781 0,069437 0,068112
1,5 0,066807 0,065522 0,064256 0,063008 0,061780 0,060571 0,059380 0,058208 0,057053 0,055917
1,6 0,054799 0,053699 0,052616 0,051551 0,050503 0,049471 0,048457 0,047460 0,046479 0,045514
1,7 0,044565 0,043633 0,042716 0,041815 0,040929 0,040059 0,039204 0,038364 0,037538 0,036727
1,8 0,035930 0,035148 0,034379 0,033625 0,032884 0,032157 0,031443 0,030742 0,030054 0,029379
1,9 0,028716 0,028067 0,027429 0,026803 0,026190 0,025588 0,024998 0,024419 0,023852 0,023295
2,0 0,022750 0,022216 0,021692 0,021178 0,020675 0,020182 0,019699 0,019226 0,018763 0,018309
2,1 0,017864 0,017429 0,017003 0,016586 0,016177 0,015778 0,015386 0,015003 0,014629 0,014262
2,2 0,013903 0,013553 0,013209 0,012874 0,012545 0,012224 0,011911 0,011604 0,011304 0,011011
2,3 0,010724 0,010444 0,010170 0,009903 0,009642 0,009387 0,009137 0,008894 0,008656 0,008424
2,4 0,008198 0,007976 0,007760 0,007549 0,007344 0,007143 0,006947 0,006756 0,006569 0,006387
2,5 0,006210 0,006037 0,005868 0,005703 0,005543 0,005386 0,005234 0,005085 0,004940 0,004799
2,6 0,004661 0,004527 0,004397 0,004269 0,004145 0,004025 0,003907 0,003793 0,003681 0,003573
2,7 0,003467 0,003364 0,003264 0,003167 0,003072 0,002980 0,002890 0,002803 0,002718 0,002635
2,8 0,002555 0,002477 0,002401 0,002327 0,002256 0,002186 0,002118 0,002052 0,001988 0,001926
2,9 0,001866 0,001807 0,001750 0,001695 0,001641 0,001589 0,001538 0,001489 0,001441 0,001395
3,0not be 1,350E-03
These slides can reproduced or distributed without permission in writing from the author. 126
2.2.9.3. Gauss Normal Distribution: CLASSICAL APPLICATIONS
Accepting
(to avoid interpolation, even if we could use EXCEL)
a percentage of underweight cans equal to 2.018%
(instead of the target equal to 2.00%),
we obtain: z = 2.05, which becomes z = - 2.05 to take into
account that the limit is lower than the mean
(and therefore on its left).
zlim - m
From: z = --------------- = - 2.05
s
Applying the formula in the reverse way, we get:
1) m = 8.00 + (2.05 . 0.08) = 8.164 hg
2) s = (8.00 - 8.10) / (-2.05) = 0.04878 hg = 4.88 gr
These slides can not be reproduced or distributed without permission in writing from the author. 127
2.2.9. Gauss Normal Distribution
2.2.9.4
Normal Probability Plot
(NPP)
These slides can not be reproduced or distributed without permission in writing from the author. 128
2.2.9.4. Gauss Normal Distribution: NPP
These slides can not be reproduced or distributed without permission in writing from the author. 129
2.2.9.4. Gauss Normal Distribution: NPP
These slides can not be reproduced or distributed without permission in writing from the author. 131
2.2.9.4. Gauss Normal Distribution: NPP
DIAMETER MEASUREMENTS OF 80 ROLLS PRODUCED BY A LATHE
comparison with expected values from a normal distribution
35
31,5
Paired columns 30 Measured values
30 graph Expected values
Number of rolls in each class
25
22
21
20
19,4 19,3
15
10
4,5 4,5
5
3
2
1 0,4 1
0 0,0 0,4 0 0,0
0
4,80 4,85 4,90 4,95 5,00 5,05 5,10 5,15 5,20
Diameter [mm]
We can immediately perceive that the patterns of the two series of columns are similar,
So we can reasonably think that rolls diameters are producing according to a normal distribution
(with some uncertainty about the tails) .
These slides can not be reproduced or distributed without permission in writing from the author. 132
2.2.9.4. Gauss Normal Distribution: NPP
In order to answer to
these questions, we
can also use another
graphical method,
but more refined
than just observing
the columns of a
histogram. This is
the Normal
Probability Plot
(NPP).
The NPP construction
consists of transfor-
ming the cumulative
normal distribution
function into a straight We can choose the straight line as
line, as shown on the we prefer (better if around 45°)
right.
These slides can not be reproduced or distributed without permission in writing from the author. 133
2.2.9.4. Gauss Normal Distribution: NPP
These slides can not be reproduced or distributed without permission in writing from the author.
The word probit (“probability unit”) indicates the random variable:
Z-m
Y = ---------- + 5
s
where Z is a random variable normally distributed with mean m and standard
deviation s. Values transformed as probit are interesting because it is im-
possible that Y variable get negative values. Besides they are a helpful
reference in the graphical construction of the NPP chart.
2.2.9. Gauss Normal Distribution
2.2.9.5
Limits
of Gauss normal distribution
These slides can not be reproduced or distributed without permission in writing from the author. 137
2.2.9.4. Gauss Normal Distribution: LIMITS
These slides can not be reproduced or distributed without permission in writing from the author. 138
2.2.9.4. Gauss Normal Distribution: LIMITS
Adult males
0,045
(only)
Maschi adulti
mean = 171 cm
0,040 (soltanto)
approximately
media = 171 cm
normal
distribuzione
0,035
DI PROBABILITÀ
distribution
normale
DENSITY
0,030
0,025
FREQUENCY
WholePopolazione
population
0,020
completa
(children included)
DENSITÀ
(inclusi
mean bambini)
= 147 cm
0,015
media = 147 cm
non-normal
distribuzione
distribution
0,010 non-normale
0,005
0,000
0 20 40 60 80 100 120 140 160 180 200 220 240
HEIGHTs [cm]
ALTEZZE [cm]
These slides can not be reproduced or distributed without permission in writing from the author. 139
2.2.9.4. Gauss Normal Distribution: LIMITS
ADDITIONAL COMMENTS
For example, the dotted normal distribution of the next slide has
been obtained:
requiring that the mean is equal to the mode of the actual distribution
and roughly optimizing standard deviation (assumed equal to 1.58
times the standard deviation of only males, which is equal to 9 cm).
These slides can not be reproduced or distributed without permission in writing from the author. 140
2.2.9.4. Gauss Normal Distribution: LIMITS
Normale with:
Normal distribution con : Popolazione
mean =media
mode =ofmoda popolazione
the whole completa
population
Non-normal
0,025 . completa
distribution
scarto
and standard tipo ==1,58
deviation 1.58quello deistd.
(males "maschi"
dev.)
DI PROBABILITÀ
(inclusi
on bambini)
the heights of
DENSITY
mediawomen
men, = 147 cm
and
0,020
children (already
distribuzione
discussed)
non-normale
FREQUENCY
0,005
0,000
0 20 40 60 80 100 120 140 160 180 200 220 240
ALTEZZE [cm]
HEIGHT [cm]
These slides can not be reproduced or distributed without permission in writing from the author. 141
2.2. Descriptive Statistics
2.2.10
Limits
of Descriptive Statistics
“estimations”
These slides can not be reproduced or distributed without permission in writing from the author. 142
2.2.10. Limits of Descriptive Statistics “estimations”
These slides can not be reproduced or distributed without permission in writing from the author. 143
2.2.10. Limits of Descriptive Statistics “estimations”
These slides can not be reproduced or distributed without permission in writing from the author. 144
2. Fundamentals of Applied Statistics
2.3
P R O B A B I L I T Y
T H E O R Y
These slides can not be reproduced or distributed without permission in writing from the author. 145
2.3. Probability Theory
2.3.1
Sampling
These slides can not be reproduced or distributed without permission in writing from the author. 146
2.3.1. Sampling
These slides can not be reproduced or distributed without permission in writing from the author. 147
2.3.1. Sampling
MAIN CATEGORIES OF SAMPLING
Random: every population element has an equal chance of being part of
the sample.
Systematic: it is selected an item every n (it is practical in the production
processes, but we must pay great attention to any phenomenon which
may occur at regular intervals).
Stratified: having identified a stratification criterion, the elements of the
sample will be drawn randomly from the different “strata”, in the same pro-
portion in which they appear in the whole population. For example, if we want
to determine the mean weight of an adult population composed 55% of women and
45% of men, as we know that women are lighter than men, we will obtain a better
efficiency if we choose a random sample (but) with the bond to have 55% of women
and 45% of men.
In bunches: we divide the population in many sub-populations (without in-
tersections), according to a determined criterion (e.g.: different geographic
zones). These sub-populations can be regarded as similar to the berries of a
bunch of grapes (which represents the full set); we will choose randomly
only some berries in which to carry our investigation (the others will be ex-
cluded from sampling).
These slides can not be reproduced or distributed without permission in writing from the author. 148
2.3.1. Sampling
Random sampling
x x x x Each population
element has an equal
x x x xxxx
x x chance of being
x selected as part of
x x x
x the sample.
x
These slides can not be reproduced or distributed without permission in writing from the author. 149
2.3.1. Sampling
Random sampling
We can resort to :
cards in a hat well shuffled before the drawing;
random numbers Tables: see next slide;
generation of random numbers by a software (e.g.
EXCEL);
others …..
These slides can not be reproduced or distributed without permission in writing from the author. 150
2.3.1. Sampling
Extract of RANDOM NUMBERS TABLES A random numbers Table is
composed of many pages, in
which digits from 0 to 9 are
printed in a random way (as if
every digit has the same pro-
bability to be drawn).
Example of use - We want to
randomly extract 20 engineers
of 600 members of the Engi-
neering Association of a coun-
try. At every Engineer we can
associate a number of 3 digit
from 001 to 600. On the ran-
dom numbers Table, we arbi-
trarily choose a page, a row
and three columns from which
we start to read and select the
numbers (of 3 digits), going
from the top to the bottom (or
vice versa). Of course we ex-
clude numbers greater than
600 and repeated values.
These slides can not be reproduced or distributed without permission in writing from the author. 151
2.3.1. Sampling
A B B
A B B B The sample is
A A B B B random and
AABBBBCDDD
C D D proportional to each
C D group
D D
D
These slides can not be reproduced or distributed without permission in writing from the author. 152
2.3.1. Sampling
Systematic sampling
Population
or process Sample Rule
These slides can not be reproduced or distributed without permission in writing from the author. 153
2.3.1. Sampling
Systematic sampling
we risk to find
always equal value items
These slides can not be reproduced or distributed without permission in writing from the author. 154
2.3.1. Sampling
Sampling by subgroups
Process Sample Rule
These slides can not be reproduced or distributed without permission in writing from the author. 155
2.3.1. Sampling
In bunches sampling
These slides can not be reproduced or distributed without permission in writing from the author. 156
2.3.1. Sampling
NO NO YES
2.3.2
Basic
definitions and relationships
about probability
These slides can not be reproduced or distributed without permission in writing from the author. 158
2.3.2. Basic definitions and relationships about probability
2.3.2.1
Probability definitions
These slides can not be reproduced or distributed without permission in writing from the author. 159
2.3.2.1. Probability definitions
1. Classical
2. Frequentist
3. Axiomatic
4. Bayesian
These slides can not be reproduced or distributed without permission in writing from the author. 160
2.3.2.1. Probability definitions
EXAMPLES
Possible
Game Description Values
events
Probability that in a throw tails 1/2 0,5 50%
COIN 2: head or tail. comes out.
DIE 6: each of the Probability that in a throw:
panes num- pane 5 comes out 1/6 0,167 17%
bered from 1 an even pane comes out 3/6 0,5 50%
to 6.
Probhability to draw from the deck:
DECK OF 54 54: each of the a JOKER 2/54 0,037 4%
54 cards. a King
CARDS 4/54 0,074 7%
a card of value from 7 to 10 16/54 0,296 30%
(with 2 Jokers)
These slides can not be reproduced or distributed without permission in writing from the author. 161
2.3.2.1. Probability definitions
CLASSICAL probability definition
The classical definition needs a clarification: all results
(or alternatives) must be “equally possible”: that means
having equal probability of occurring. We fall into a vicious
circle because equi-probability is a particular case of the sa-
me probability that we want to define.
These slides can not be reproduced or distributed without permission in writing from the author. 162
2.3.2.1. Probability definitions
2. FREQUENTIST probability definition
“Frequentists” define the probability as the
Ratio between the number of “successes” and the
number of experiments, when the number of
experiments tends to the infinity.
These slides can not be reproduced or distributed without permission in writing from the author. 163
2.3.2.1. Probability definitions
FREQUENTIST probability definition
What we said can be written in a mathematical expression as:
nS
P(S) = lim -------
n n
These slides can not be reproduced or distributed without permission in writing from the author. 164
2.3.2.1. Probability definitions
3. AXIOMATIC probability definition
Mathematics states that to apply the probability concept to
calculations, it is enough to accept the following 3 postulates or
axioms (with some simplification):
1) the probability of an event E is a real number, positive or null:
P(E) 0;
2) if S is the “sample space” (or “universal sample space” or
“universe”) of all possible outcomes of an experiment, the
associated probability is 1: P(S) = 1;
3) if A and B are mutually exclusive events, then the probability
that can happen A or B (AB) is equal to the sum of the
individual probabilities of the events A and B:
P(AB) = P(A) + P(B) ;
the above expression can be extended to an infinite set of events
E1, E2, E3, … mutually exclusive.
The definition of mutually exclusive events will be given in the following. Here we
want only to state the few postulates necessary to the Mathematics in order to elabo-
rate the theory.
These slides can not be reproduced or distributed without permission in writing from the author. 165
2.3.2.1. Probability definitions
These slides can not be reproduced or distributed without permission in writing from the author. 166
2.3.2.1. Probability definitions
BAYESIAN probability definition
(*) The following slides about Bayesian Statistics are taken from the Report: Guida M., Pulcini G. - Elementi
di Inferenza e Decisione Bayesiana (Fundamentals of Bayesian Inference and Decision) - C.N.R. Istituto
Motori (Engines Institute), Naples, November 6th-7th 2000.
These slides can not be reproduced or distributed without permission in writing from the author. 167
2.3.2.1. Probability definitions
BAYESIAN probability definition
So if we say that:
• an event is more probable than another one
These slides can not be reproduced or distributed without permission in writing from the author. 168
2.3.2.1. Probability definitions
BAYESIAN probability definition
These slides can not be reproduced or distributed without permission in writing from the author. 169
2.3.2.1. Probability definitions
BAYESIAN probability definition
The most common use of forecast Bayesian approach is based on the fol-
lowing 3 steps:
1. a first estimate, prior (before), based on the best side information (both
objective and subjective) available;
2. a second estimate, posterior (after), based on experimental data;
3. the two estimates are then combined together, giving them different "weights",
in relation to the different situations.
E X A M P L E S
To estimate the in field failure frequency of a new product, a Company can: 1. initially
use failure frequency data of similar products already in the market, after “correcting” them in
function of improvements/criticalities predictable on the basis of modifications planned for the
new model under development (prior); 2. as soon as it is possible to test the new product,
failure frequency can be estimated experimentally (posterior); 3. then, after verifying that
the “posterior” is not too much different from the “prior”, both estimates will be combined
together, quantifying, in some way, their different relative “weight” (open problem).
To define the insurance premium of a new kind of risk (never managed before), an
Insurance Company will initially use data of “similar” risks, after recalibrating them on the
basis of subjective considerations about differences (side information prior). Then, gradually
over the years that will capture data on the specific risk (posterior), the relative weight of the
side information will be reduced respect to the “specific” one and will become zero after a
period long enough.
These slides can not be reproduced or distributed without permission in writing from the author. 170
2.3.2.1. Probability definitions
Example 2
Insurance
premiums
relating to
insurance
against
a new type
of risk:
the wall
of death
These slides can not be reproduced or distributed without permission in writing from the author. 171
2.3.2.1. Probability definitions
Probability BAYESIAN definition
For example, let us suppose that an Insurance Company want to ensure the motorcyclists
of the wall of death (fun fair), but they do not have any data about this kind of risk.
A first estimation of the average yearly cost for an insured of this category (cost from
which premium can be evaluated) can be obtained by the Bayesian method, exploiting the
similarities.
Let us suppose that we can subjectively state that the risks of the wall death
motorcyclists are 60% similar to those (known) of the firemen and 40% to those (also
known) of the trapezists. Defined cm, cf e ct the relative yearly costs, we can write, at the
moment:
cm = 0.60 cf + 0.40 ct
If we know that, for a good costs estimate, we need data of at least 10,000 polices, once
we have achieved actual data about wall death motorcyclists, the preceding subjective
estimation will lower its importance to leave greater importance to the true data of motor-
cyclists. Indicating by nm their number, we can write:
These slides can not be reproduced or distributed without permission in writing from the author. 172
2.3.2. Basic definitions and relationships about probability
2.3.2.2
Other definitions
These slides can not be reproduced or distributed without permission in writing from the author. 173
2.3.2.2. Other definitions
“EVENTS” definitions
EVENT DEFINITION P(E) EXAMPLES
These slides can not be reproduced or distributed without permission in writing from the author. 174
2.3.2.2. Other definitions
CONDITIONED PROBABILITY
Let us imagine that in a box there are a white ball and a black
ball. We want to know the probability to drawn the white ball at
the second drawing.
Done the 1st drawing ...
If we do not know the outcome of
the 1st drawing, the probability that
the left ball be, for instance, white
is always 0.5.
If the outcome of the 1st drawing is
known, then the 2nd drawing is
always certain: P(W\W) = 0;
P(W\B) = 1.
So the knowledge of a certain event can “condition” the estimation of
the probability that another event (linked to the previous one) happens.
These slides can not be reproduced or distributed without permission in writing from the author. 175
2.3.2. Basic definitions and relationships about probability
2.3.2.3
Relationships
These slides can not be reproduced or distributed without permission in writing from the author. 176
2.3.2.3. Probability relationships
GENERAL RULE OF SUM
Let us imagine that at a party there are 100 girls, of whom:
• 55 blond P(A) ;
• 40 with blue eyes P(B);
• 20 blond and with blue eyes;
• 25 neither blond nor with blue eyes.
E V E N T S
ANY TYPE
RULES PARTICULAR
(general expressions)
sum Mutually exclusive
probability that one P(AB) = P(A) + P(B) - P(AB)
of two events happens
(one or the other, no matter what) P(AB) = P(A) + P(B)
These slides can not be reproduced or distributed without permission in writing from the author. 178
2.3.2.3. Probability relationships
• 55 blond P(A) ;
EXAMPLES
35 20 20 • 40 with blue eyes P(B);
about 100 girls • 20 blond and with blue eyes;
25
• 25 neither blond nor with blue eyes.
Venn diagram
These slides can not be reproduced or distributed without permission in writing from the author. 179
2.3.2. Basic definitions and relationships about probability
2.3.2.4
Sample space
These slides can not be reproduced or distributed without permission in writing from the author. 180
2.3.2.4. Sample space
These considerations might also appear a bit pedantic, but they must
be considered in the probability calculations, because they can chan-
ge when the referring sample space changes.
These slides can not be reproduced or distributed without permission in writing from the author. 181
2.3.2.4. Sample space
For simplicity, let us also assume that the occurrences of these kinds
of failure are mutually independent events, since the occurrence of
any of them does not affect in any way the occurrence of the other
two. This implies that the occurrence of a type of failure should not facilitate
the occurrence of any of the other two kinds, both if they are related to a
single component or to different components.
These slides can not be reproduced or distributed without permission in writing from the author. 182
2.3.2.4. Sample space
These slides can not be reproduced or distributed without permission in writing from the author. 183
2.3.2.4. Sample space
In the case that only a single type of failure can occur, we can use
a binary logic: the failure does happen or does not happen and
these two alternatives, mutually exclusive, are the whole sample
space.
In our example of 3 different kinds of failure, the description of
every outcome of the experiment must include the state (present or
absent) of each of them. The set of all possible outcomes, listed in
the following Table, represents the sample space.
These slides can not be reproduced or distributed without permission in writing from the author. 184
2.3.2.4. Sample space
States of every kind of failure
A B C
Occurrence, F 0.45 0.40 0.35
FW = FA . RB . RC + FB . RA . RC + FC . RA . RB + FA . FB . RC +
+ FA . FC . RB + FB . FC . RA + FA . FB . FC
These slides can not be reproduced or distributed without permission in writing from the author. 186
2.3.2.4. Sample space
Operating simplifications
approximation and acceptability threshold
With very low failure frequencies, their products (of two and three
factors) can be neglected and the reliabilities R are very close to 1.
These slides can not be reproduced or distributed without permission in writing from the author. 187
2.3.2.4. Sample space
2.3.3
Examples of
probability calculations
These slides can not be reproduced or distributed without permission in writing from the author. 189
2.3.3. Examples of probability calculations
EXAMPLE 1
The probability to obtain 1 TAIL and 2 HEADS in 3 throws regardless of the order of
appearance is then:
P(1 TAIL and 2 HEADS) = 3 . P(THH) = 3 . q p2 = 3 . 0.53 = 3 . 0.125 =
= 0.375 37.5%
These slides can not be reproduced or distributed without permission in writing from the author. 191
2.3.3. Examples of probability calculations
R U L E
The number of all possible outcomes with n throws and with k successes,
regardless of the order of their appearance,
is given by binomial coefficients (see later about binomial distribution),
which have the following expression:
n n! n (n - 1) (n - 2) ... (n - k + 1)
( ) = ---------------- = ------------------------------------------
k (n - k)! . k! k!
These slides can not be reproduced or distributed without permission in writing from the author. 192
2.3.3. Examples of probability calculations
In the above example, if we were interested in knowing the proba-
bility to obtain, with 3 throws, one TAIL and two HEADS regar-
dless of the order of appearance, instead of to list all possible
dispositions as before, we can proceed as follows.
Calculate the probability of occurence of any of the wanted
dispositions (as before), for example:
P(THH) = q p2 = 0.53 = 0.125 12.5%
These slides can not be reproduced or distributed without permission in writing from the author. 194
2.3.3. Examples of probability calculations
1 1 1 1 1 1
So the calculation: ---- + ---- + ---- + ---- + ---- = ---------- 5.683% is wrong!
90 89 88 87 86 17.595
These slides can not be reproduced or distributed without permission in writing from the author. 195
2.3.3. Examples of probability calculations
Probability that a number is drawn in a lottery drum
It was very easy to
DRAWING imagine that every-
PROBABILITY CALCULATIONS
# position
one of the 5 events
1
has the same pro-
1 l ----- . 1 . 1 . 1 . 1 = 1/90 = 0.011111
90
bability, because
89 1
there is no reason
2 l ----- ----- . 1 . 1 . 1
.
= 1/90 = 0.011111 that a number is
90 89 more likely to be
89 88 1 drawn in a fixed
3 l ----- . ----- . ----- . 1 . 1 = 1/90 = 0.011111 drawing (e.g. the
90 89 88 first one) than in
89 88 87 1 another.
4 l ----- ----- ----- ----- . 1
. . .
= 1/90 = 0.011111
90 89 88 87 The right result
89 88 87 86 1 is 1/18
5 l ----- . ----- . ----- . ----- . ----- = 1/90 = 0.011111 against the
90 89 88 87 86
previous
S = 5/90 = 1/18 = 0.055556
1/17.595.
These slides can not be reproduced or distributed without permission in writing from the author. 196
2.3.3. Examples of probability calculations
EXAMPLE 5: Probability that a DOUBLE be drawn in a lottery drum
The calculation must be performed by a Table as shown below (and similar to that of
the previous example). For the first number drawn, the chances are 2 (one for both
the number of a double), while, after the first number was drawn, the second number
has only one chance.
R U L E
These slides can not be reproduced or distributed without permission in writing from the author. 198
2.3.3. Examples of probability calculations
More often we are interested to the reverse problem, i.e. to determine the mi-
nimum sample size able to detect at least one defective, with a probability not
less than a fixed value. We get:
n = Log (1 - Pone or more) / Log (1 - p)
These slides can not be reproduced or distributed without permission in writing from the author. 199
2.3. Probability Theory
2.3.4
A brief mention of
the statistical distributions
of current use
These slides can not be reproduced or distributed without permission in writing from the author. 200
2.3.4. Statistical distributions of current use
These slides can not be reproduced or distributed without permission in writing from the author. 201
2.3.4. Statistical distributions of current use
These slides can not be reproduced or distributed without permission in writing from the author. 202
2.3.4. Statistical distributions of current use
2.3.4.1
Discrete distributions
These slides can not be reproduced or distributed without permission in writing from the author. 203
2.3.4.1. Discrete distributions
Binomial distribution;
Poisson distribution.
These slides can not be reproduced or distributed without permission in writing from the author. 204
2.3.4. Statistical distributions of current use
2.3.4.2
Binomial distribution
These slides can not be reproduced or distributed without permission in writing from the author. 205
2.3.4.2. Binomial distribution
These slides can not be reproduced or distributed without permission in writing from the author. 206
2.3.4.2. Binomial distribution
Let us imagine that in a box there are 75 yellow balls and 25 green balls.
Let us state to call:
• not-success the drawing of a yellow ball
• and success the drawing of a green ball .
If we do many drawings with the rule that, after each drawing, the drawn ball
be put in the box again and all the balls are shuffled (drawings with reinsta-
tement).
At every drawing:
the probability q of drawing a yellow ball (not-success) is equal to 75%;
the probability p of drawing a green ball (success) is equal to 25%.
These slides can not be reproduced or distributed without permission in writing from the author. 207
2.3.4.2. Binomial distribution
Obviously, expression (1) is valid for all sequences with the same number of
yellow (3) and green (2) balls, regardless of the order in which they occur:
that means that all outcomes of 5 drawings with 3 yellow balls and 2 green
balls have the same probability of occurrence given by (1).
These slides can not be reproduced or distributed without permission in writing from the author. 208
2.3.4.2. Binomial distribution
The expression:
bu(k) = pk qn-k = pk (1-p)n-k (2)
expresses the probability of each single fixed outcome with k successes
and n - k not-successes in n drawings: n and p are parameters, while k is
the independent variable which can never have values greater than n.
These slides can not be reproduced or distributed without permission in writing from the author. 209
2.3.4.2. Binomial distribution
If, as usually, we are not interested in the order of successes and not-suc-
cesses, but only in their number, it is sufficient to multiply the unitary proba-
bility bu, given by (2), by the number of all possible dispositions in which can
be present k successes in n drawings. This number is given by the following
expression, which, as we saw before, expresses binomial coefficients:
n n! n (n - 1) (n - 2) ... (n - k + 1)
( ) = ---------------- = ------------------------------------------ (3)
k (n - k)! . k! k!
These slides can not be reproduced or distributed without permission in writing from the author. 210
2.3.4.2. Binomial distribution
Binomial coefficients were already defined (at the Par. 2.3.3., where we
exposed some examples of probability calculations) as:
For k = 0, bino-
n n! n (n - 1) (n - 2) ... (n - k + 1) mial coefficients
( ) = ---------------- = ------------------------------------------ are always equal
k (n - k)! . k! k! to 1, by definition.
Tartaglia/Pascal triangle for binomial coefficients
0,1,1,2,2,3,3,……
con kk = 0,
k ( (with ) )
0 1
1 1 1
2 1 2 1
3 1 3 3 1
4 1 4 6 4 1
5 1 5 10 10 5 1
6 1 6 15 20 15 6 1
n
7 1 7 21 35 35 21 7 1
8 1 8 28 56 70 56 28 8 1
9 1 9 36 84 126 126 84 36 9 1
10 1 10 45 120 210 252 210 120 45 10 1
11 1 11 55 165 330 462 462 330 165 55 11 1
12 1 12 66 220 495 792 924 792 495 220 66 12 1
e c ac ne d
t e sr ao on
n = 30; p = 0,02
f(k)
probabilitàf(k)
0,5
n = 30; p = 0,04
distribution
n = 60; p = 0,02
0,4
n = 60; p = 0,04
0,3
Probability di
Distribuzione
0,2
0,1
0,0
0 1 2 3 4 5 6 7
Numeroof
Number disuccesses
successi kk
These slides can not be reproduced or distributed without permission in writing from the author. 212
2.3.4.2. Binomial distribution
1,0
0,9
F(k)
cumulata F(k)
0,8
Probabilitàprobability
0,7
0,6
0,5
n = 30; p = 0,02
Cumulative
0,1
0,0
0 1 2 3 4 5 6 7
Numeroofdisuccesses
Number successi kk
These slides can not be reproduced or distributed without permission in writing from the author. 213
2.3.4.2. Binomial distribution
BINOMIAL DISTRIBUTION
The probability b of k successes in n tests
with unitary success probability p The probability P
is given by: to detect at least a failure (= one or more )
n with a field true frequency equal to p,
b(k) = ( ) . pk . (1 - p)n-k testing a sample of n elements,
k can be evaluated by the following expression,
The probability of zero successes is given by which refers to binomial distribution :
P = 1 - (1 - p)n
the previous expression with k = 0 :
n
b(0) = ( ) . p0 . (1 - p)n-0 = 1 . 1 . (1 - p)n = (1 - p)n This probability can also be regarded as the
0 fraction (or %) of different kinds of failure
The probability P of " one or more " successes (with roughly the same fixed field-frequency)
can be written as the complement to 1 detectable by a sample of n elements.
of the preceding:
P = 1 - b(0) = 1 - (1 - p)n .
These slides can not be reproduced or distributed without permission in writing from the author. 214
2.3.4.2. Binomial distribution
CLASSES OF FAILURE PERCENTTO
PROBABILITY OF DIFFERENT
DETECT FAILURE
AL LEAST KINDS
ONE OCCURENCE
FREQUENCY in detectable
a sample ofinassigned size
the sample
[ repairs / 100 units ] sample size n
from to class
centers
5 10 20 50 100 500
(not-included) (included)
27,50 32,50 30,000 83,193% 97,175% 99,920% 100,000% 100,000% 100,000%
22,50 27,50 25,000 76,270% 94,369% 99,683% 100,000% 100,000% 100,000%
17,50 22,50 20,000 67,232% 89,263% 98,847% 99,999% 100,000% 100,000%
12,50 17,50 15,000 55,629% 80,313% 96,124% 99,970% 100,000% 100,000%
7,50 12,50 10,000 40,951% 65,132% 87,842% 99,485% 99,997% 100,000%
5,50 7,50 6,500 28,541% 48,936% 73,925% 96,528% 99,879% 100,000%
4,50 5,50 5,000 22,622% 40,126% 64,151% 92,306% 99,408% 100,000%
3,50 4,50 4,000 18,463% 33,517% 55,800% 87,011% 98,313% 100,000%
2,50 3,50 3,000 14,127% 26,258% 45,621% 78,193% 95,245% 100,000%
1,70 2,50 2,100 10,068% 19,123% 34,589% 65,395% 88,025% 99,998%
1,30 1,70 1,500 7,278% 14,027% 26,086% 53,031% 77,939% 99,948%
1,10 1,30 1,200 5,858% 11,372% 21,451% 45,318% 70,098% 99,761%
0,90 1,10 1,000 4,901% 9,562% 18,209% 39,499% 63,397% 99,343%
0,70 0,90 0,800 3,937% 7,718% 14,840% 33,076% 55,211% 98,198%
0,50 0,70 0,600 2,964% 5,841% 11,340% 25,985% 45,218% 95,066%
0,35 0,50 0,425 2,107% 4,170% 8,165% 19,181% 34,682% 88,111%
0,25 0,35 0,300 1,491% 2,960% 5,832% 13,949% 25,952% 77,737%
0,15 0,25 0,200 0,996% 1,982% 3,925% 9,525% 18,143% 63,249%
0,05 0,15 0,100 0,499% 0,996% 1,981% 4,879% 9,521% 39,362%
0,05 0,050 0,250% 0,499% 0,995% 2,470% 4,878% 22,125%
These slides can not be reproduced or distributed without permission in writing from the author. 215
2.3.4.2. Binomial distribution
PROBABILITY TO DETECT
PERCENTAGE OFAL LEAST ONE
DIFFERENT OCCURRENCE
KINDS OF FAILURE (ONE OR MORE)
in function
detectable of the
in function oftrue field
the true failure
field failurefrequency andofofthe
frequency and the sample
sample sizesize
100%
% OF PROBABILITY OF DETECTING AL LEAST ONE OCCURRENCE
90%
80%
% OF DIFFERENT KINDS OF FAILURE
70%
detectable in the sample
60%
50%
40%
30%
5
Sample size n
10
20% 20
50
10% 100
500
0%
0 1 2 3 4 5 6 7 8 9 10
TRUE
TRUE FIELD FIELD FREQUENCY
FAILURE FREQUENCY [ [repairs
repairs/ /100
100units
units] ]
These slides can not be reproduced or distributed without permission in writing from the author. 216
2.3.4.2. Binomial distribution
PROBABILITY TOPrecentuali
DETECT AL LEASTdiverse
di tipologie ONE OCCURRENCE
di guasto (ONE OR MORE)
in function
rilevabili of della
in funzione the true field failure
frequenza frequency
vera in rete and of the
e delle dimensioni delsample size
campione
OCCURRENCE
100% 5
DEL CAMPIONE
Sample size n
NUMEROSITÀ
10
20
90%
50
DIVERSE
100
80%
500
AL LEAST ONE
RILEVABILI DA CAMPIONE
70%
DI INCONVENIENTE
60%
OF DETECTING
50%
40%
% TIPOLOGIE
30%
% OF PROBABILITY
20%
10%
0%
0,0 0,5 1,0 1,5 2,0
TRUEFREQUENZA VERA
FIELD FAILURE IN RETE [inconvenienti
FREQUENCY / 100
[ repairs / 100 unità
units ] ]
100%
90%
80%
OF FAILURE
70%
detectable in the sample
LEAST
more)
60%
DETECT AT KINDS
or
50%
DIFFERENT
(one
40%
% OFTO
30%
5
Sample size n
PROBABILITY
10
20% 20
50
10% 100
500
0%
0 1 2 3 4 5 6 7 8 9 10
TRUE
TRUE FIELD
FIELD FREQUENCY
FAILURE [ repairs
FREQUENCY / 100/ 100
[ repairs units ] ]
units
These slides can not be reproduced or distributed without permission in writing from the author. 218
2.3.4.2. Binomial distribution
Now let us imagine that the different kinds of failure with a frequency around 2
repairs/100 units (as “windscreen wiper blades: noise”) are 12. We want to know
which is the percentage (of this 12 kinds of failure) that can be expected to appear
with at least one occurence in a sample of 10 cars ( 0.19.12 = 2.28) and in a
sample of 50 cars ( 0.63 .12 = 7.56). Same procedure as before, but with a broader
PERCENTAGE OF DIFFERENT KINDS OF FAILURE
meaning. detectable in function of the true field failure frequency and of the sample size
100%
90%
80%
% OF DIFFERENT KINDS OF FAILURE
70%
detectable in the sample
60%
50%
40%
30%
5
Sample size n
10
20% 20
50
10% 100
500
0%
0 1 2 3 4 5 6 7 8 9 10
TRUE
TRUE FIELD
FIELD FREQUENCY
FAILURE [ repairs
FREQUENCY / 100/ 100
[ repairs units ] ]
units
These slides can not be reproduced or distributed without permission in writing from the author. 219
2.3.4.2. Binomial distribution
We know, at least for the main car models already marketed, the
distribution in frequency classes of the different kinds of failure (i.e.
component & defect). In other words, we know how many different kinds
of failure fall in a stated range of frequency (class). Hence (with the seen
use of the binomial distribution), we can evaluate the probability to get at
least one failure (one or more) in a sample of fixed size, or (which is the
same) we can evaluate the percentage of different kinds of failure we can
expect to be detectable in every frequency class, during the test on this
sample.
As an example, the next slide presents one of these analysis, invented, but
indicative of what you might find on a car of upper-middle class in the mid
80s. At the present time, the defects were reduced by about 10 times !
These slides can not be reproduced or distributed without permission in writing from the author. 220
2.3.4.2. Binomial distribution
736
TOTAL
Repairs/100 cars
in the class
Number of different failure kinds 1,216
TOTAL
1,216 different failure kinds 379.8
Repairs/100 cars (total of each class)
379.8 repairs/100 cars
85
57
46
38,8
43
37,6
34,3
33,3
28,5
27,2
26,0
25,4
24,9
20,0
23
17,9
17,0
19
13,8
13,6
17
17
11,4
10,1
12
9
3
2
0,05 0,10 0,20 0,30 0,43 0,60 0,80 1,00 1,20 1,50 2,10 3,00 4,00 5,00 6,50 10,00
These slides can not be reproduced or distributed without permission in writing from the author. 221
2.3.4.2. Binomial distribution
Blue columns indicate the number of different kinds of
failure frequency in the class (defined by the x-axis).
Multiplying each class center for the relative number of kinds
of failure, we obtain an approximate value for the red co-
lumns: in fact, the exact result is a bit different, because it
was obtained by the actual frequency of each kind of failure of
the class.
Frequency classes do not have the same width, but they
must be chosen so that the values of their failure frequencies
(red columns) are quite close: low value of the class center
and large number of different kinds of failure in the class
balance each other.
We see that we have a total of about 1,200 different kinds of
failure mode, of which 736 have a failure frequency less than
0,05% (i.e. 0,5 per thousand). Generally we must not to fix
these kinds of failure with very low frequency, because of the
risk of produce other kinds of failure with greater failure
frequency.
These slides can not be reproduced or distributed without permission in writing from the author. 222
2.3.4.2. Binomial distribution
Knowing:
• the detectability of the kinds of failure in each class, using a
sample of given size (by means of the binomial distribution),
• and the number of different kinds of failure in each class,
which have the same detectability (obtained by the previous
histogram),
These slides can not be reproduced or distributed without permission in writing from the author. 223
2.3.4.2. Binomial distribution
Detectable percentage of DIFFERENT KINDS OF FAILURE and of REPAIRS/100 CARS
CLASSES OF FAILURE
FREQUENCY Data measured IN FIELD as a function of the sample size n
(warranty) sample size n
[ repairs / 100 units ] 10 20 50 100
from Average Average Average Average
to class N° of failure Repairs/100 cars No. failure kinds Repairs/100 cars No. failure kinds Repairs/100 cars No. failure kinds Repairs/100 cars No. failure kinds Repairs/100 cars
(not-in- detectability detectability detectability detectability
(included) centers kinds in the class mildly detectable mildly detectable mildly detectable mildly detectable mildly detectable mildly detectable mildly detectable mildly detectable
cluded) (binomial) (binomial) (binomial) (binomial)
7,50 12,50 10,000 3 26,0 65,13% 1,95 16,93 87,84% 2,64 22,84 99,48% 2,98 25,87 100,00% 3,00 26,00
5,50 7,50 6,500 5 33,3 48,94% 2,45 16,30 73,92% 3,70 24,62 96,53% 4,83 32,14 99,88% 4,99 33,26
4,50 5,50 5,000 2 10,1 40,13% 0,80 4,05 64,15% 1,28 6,48 92,31% 1,85 9,32 99,41% 1,99 10,04
3,50 4,50 4,000 6 25,4 33,52% 2,01 8,51 55,80% 3,35 14,17 87,01% 5,22 22,10 98,31% 5,90 24,97
2,50 3,50 3,000 12 34,3 26,26% 3,15 9,01 45,62% 5,47 15,65 78,19% 9,38 26,82 95,24% 11,43 32,67
1,70 2,50 2,100 17 37,6 19,12% 3,25 7,19 34,59% 5,88 13,01 65,40% 11,12 24,59 88,03% 14,96 33,10
1,30 1,70 1,500 19 28,5 14,03% 2,67 4,00 26,09% 4,96 7,43 53,03% 10,08 15,11 77,94% 14,81 22,21
1,10 1,30 1,200 9 11,4 11,37% 1,02 1,30 21,45% 1,93 2,45 45,32% 4,08 5,17 70,10% 6,31 7,99
0,90 1,10 1,000 17 17,9 9,56% 1,63 1,71 18,21% 3,10 3,26 39,50% 6,71 7,07 63,40% 10,78 11,35
0,70 0,90 0,800 23 20,0 7,72% 1,78 1,54 14,84% 3,41 2,97 33,08% 7,61 6,62 55,21% 12,70 11,04
0,50 0,70 0,600 43 27,2 5,84% 2,51 1,59 11,34% 4,88 3,08 25,99% 11,17 7,07 45,22% 19,44 12,30
0,35 0,50 0,425 57 24,9 4,17% 2,38 1,04 8,17% 4,65 2,03 19,18% 10,93 4,78 34,68% 19,77 8,64
0,25 0,35 0,300 46 13,8 2,96% 1,36 0,41 5,83% 2,68 0,80 13,95% 6,42 1,92 25,95% 11,94 3,58
0,15 0,25 0,200 85 17,0 1,98% 1,68 0,34 3,92% 3,34 0,67 9,53% 8,10 1,62 18,14% 15,42 3,08
0,05 0,15 0,100 136 13,6 1,00% 1,35 0,14 1,98% 2,69 0,27 4,88% 6,64 0,66 9,52% 12,95 1,29
0,05 0,050 736 38,8 0,50% 3,67 0,19 1,00% 7,33 0,39 2,47% 18,18 0,96 4,88% 35,90 1,89
Total 1.216 379,8 33,67 74,24 61,28 120,11 125,29 191,82 202,29 243,42
%: 100,0% 100,0% 2,8% 19,5% 5,0% 31,6% 10,3% 50,5% 16,6% 64,1%
These slides can not be reproduced or distributed without permission in writing from the author. 224
2.3.4.2. Binomial distribution
To plan the experimental tests for a new car model, it is necessary to esti-
mate the detectability of its failures with samples of a fixed size.
These slides can not be reproduced or distributed without permission in writing from the author. 225
2.3.4.2. Binomial distribution
These slides can not be reproduced or distributed without permission in writing from the author. 226
2.3.4.2. Binomial distribution
100%
91,6%
No. failure kinds mildly detectable
90%
86,1%
% OF FAILURES DETECTABLE IN THE SAMPLE
70%
64,1%
60% 58,0%
50,5%
50%
41,9%
40%
31,6% 32,1%
30%
25,5%
19,5%
20% 16,6%
10,3%
10%
5,0%
2,8%
0%
These slides can not be reproduced or distributed without permission in writing from the author. 227
2.3.4. Statistical distributions of current use
2.3.4.3
Poisson Distribution
These slides can not be reproduced or distributed without permission in writing from the author. 228
2.3.4.2. Poisson distribution
Special case of the Binomial distribution with: small p; large n and n . p = l constant.
Poisson’s distribution has only one parameter l,
because it does not change, also if n and/or p change, provided that their product l = n . p does not change.
A random variable is distributed according to Poisson if the probability that it takes the value
x (0, 1, 2, 3, 4, ...) is given by density function showed below:
Probability
Funzionedistribution
densità Cumulative
Funzione distribution
ripartizione
lx x li
f(x, l) = e-l ------ F(x, l) = S e-l ------
x! i=0
i!
mean = m = l; standard deviation = s = l; variance = s2 = l
It is mainly used when we want to evaluate the probability of events that occur each after the
other and the time between two successive events is exponentially distributed.
Examples
Andamenti of Poisson
esemplificativi distribution
della shapes
distribuzione di POISSON
(probability distribution)
(distribuzione di probabilità)
0,6
lambda = 0,6
f(k)
lambda = 1,0
di probabilitàf(k)
0,5
lambda = 1,2
Probability distribution,
0,4
lambda = 2,0
lambda = 2,4
0,3
Distribuzione
0,2
0,1
0,0
0 1 2 3 4 5 6 7
Numerokdiofsuccessi
Number k
successes
These slides can not be reproduced or distributed without permission in writing from the author. 230
2.3.4.2. Poisson distribution
Examples
Andamenti of Poisson
esemplificativi distribution
della distribuzioneshapes
di POISSON
(cumulated probability)
(probabilità cumulata)
1,0
0,9
F(k)
cumulata F(k)
0,8
Probabilitàprobability,
0,7
0,6
0,5
Cumulated
0,0
0 1 2 3 4 5 6 7
Numerok di
Number ofsuccessi k
successes
These slides can not be reproduced or distributed without permission in writing from the author. 231
2.3.4.2. Poisson distribution
These slides can not be reproduced or distributed without permission in writing from the author. 232
2.3.4.2. Poisson distribution SLIDE
VERY
IMPORTANT
1st EXAMPLE OF INDEPENDENT APPLICATION
TYPICAL AUTOMOTIVE CASE
Based on data resulting from Customer Service, a car manufacturer found that his cars,
during the warranty period of two years, incurred an average of l = 0,80 failures per car.
According to this mean, Poisson distribution gives the distribution of the percent of cars
with 0, 1, 2, 3, … failures (each).
Where lindicates the mean number of defects per produced unit (related to the total
warranty period) and k indicates the (hypothesized) number of defects per unit:
lk e-l
p(k) = --------------- (with k = 0, 1, 2, 3, ..... )
k!
The Poisson distribution provides the probability of finding k failures in an unit. For exam-
ple, in the case of passenger cars, it provides the probability (or frequency, or percentage)
with which, at the end of the warranty period, you will find cars with 0 failures (no repair
required to the Customer Office), 1 failure, 2 failures, 3 failures and so on. This probability
corresponds to the searched percentage of cars (see histogram in the next slide) that pre-
sents the number (of course integer!) of failures marked in the ordinate (y-axis).
It is evident the importance of the histogram on the next page to immediately highlight the
economic consequences (in terms of warranty costs at the Customer Office) of the unre-
liability of a car model. This is one of the key reasons that drive us to study and use
the Poisson distribution in the automotive field!
These slides can not be reproduced or distributed without permission in writing from the author. 233
2.3.4.2. Poisson distribution
VERY
IMPORTANT
SLIDE
0,45 0,45
0,40
0,36
Probability distribution
44,93% 0,35
45%
0,30
0,25
40% 0,20
0,05
percentage of cars
0,04
30% 0,00 0,01 0,00 0,00 0,00 0,00
0 1 2 3 4 5 6 7 8
20%
15% 14,38%
10%
5% 3,83%
0,77%
0,12% 0,02% 0,00% 0,00%
0%
0 1 2 3 4 5 6 7 8
These slides can not be reproduced or distributed without permission in writing from the author. 235
2.3.4.2. Poisson distribution
We see that less than half of the produced cars will not have any
failure in the warranty period, while 36% will have one, 14% will
have two and so on, with percentages quickly decreasing as the
number of failures per car increases.
These slides can not be reproduced or distributed without permission in writing from the author. 236
2.3.4.2. Poisson distribution
The Poisson’s hypothesis (that the standard deviation is equal to the square root of the mean)
is true only if all lots of the various components comprising the considered system have
failure frequencies also different between them but constant over time. As it is highly
unlikely that this will continue to happen for some time in a plant, the Poisson distribution is, in
practice, a convenient indication of what would happen with a variance to its lower limit.
In other words, if all the failure frequencies of the subsequent batches of the same component
(e.g.: xAi, xBi, xCi, etc.) remain substantially constant for all subsequent deliveries, then the Pois-
son’s law is valid and the system variance is equal to the average: s2S = mS . On the opposite, if
the values of the failure frequencies of the successive batches of even a single component (e.g.:
B) fluctuate from a batch to another, then the system variance increases and becomes:
s2S > mS .
These slides can not be reproduced or distributed without permission in writing from the author. 237
2.3.4.2. Poisson distribution
2nd EXAMPLE OF INDEPENDENT APPLICATION: SIMPLE
After discussing one of the main automotive independent applications of Poisson distribution,
we present now an additional (very simple) example of more general use.
Example. In a large company, the number of employees who need the medical room in a
day is on average equal to 3.8: what is the probability that in a single day there are exac-
tly 4 employees who use medical room?
Denoted by l= 3,8 the daily mean number of employees in the medical room, the
Poisson’s distribution:
lx e-l
p(x) = --------------- (with x = 0, 1, 2, 3, ..... )
x!
gives the probability to detect x employees in the medical room within a day.
To solve the proposed question, just put in the above equation:
l = 3,8; k = 4; k! = 4! = 24
and you get: f(k) = (3,84 . e-3,8) / 24 = 0,19436 ≈ 19%. The histogram in the next slide
shows the result just found and those corresponding to other values of k.
Worth to note that, from a strictly mathematical point of view, a random variable having the Poisson distribu-
tion can assume infinite values between zero and infinity, while, in practice, the chances become negligible
after a fairly small set of values of k.
These slides can not be reproduced or distributed without permission in writing from the author. 238
2.3.4.2. Poisson distribution
0,25
0,2046
0,20 0,1944
0,1615
probability
0,1477
0,15
0,10 0,0936
0,0850
0,0508
0,05
0,0224 0,0241
0,0102
0,0039 0,0013 0,0004
0,00
0 1 2 3 4 5 6 7 8 9 10 11 12
number of employees
These slides can not be reproduced or distributed without permission in writing from the author. 239
2.3.4.2. Poisson distribution
These slides can not be reproduced or distributed without permission in writing from the author. 240
2.3.4.2. Poisson distribution
3rd EXAMPLE OF INDEPENDENT APPLICATION: STRUCTURED
We present now another example of Poisson general use:
more structured than the previous one.
Example. In a production process are randomly produced also defective parts. Let us
assume that the probability of producing a fixed number of defective units in a time-inter-
val is not influenced by the number of defective units produced in other time-intervals.
Then, indicating by lthe average number of defective units per day of production,
the Poisson’s distribution:
lx e-l
p(x) = --------------- (with x = 0, 1, 2, 3, ..... )
x!
gives the probability to detect x defective units in a day.
If the actual numbers of defective units detected in the production do not match with the
frequencies predicted by Poisson distribution, we should conclude that the defective units
do not occur randomly but may be related to some special cause.
The following Table shows data of a sample of 201 production days, during which 150
defective units have been collected, giving l = 150/201 = 0,746.
These slides can not be reproduced or distributed without permission in writing from the author. 241
2.3.4.2. Poisson distribution
F I T T I N G
x n Poisson Total
number of number of observed distribution number of
defective days with frequency defective
predicted
units [%] units
x defective frequency
produced
units [%] [x . n]
in a day
0 102 50.75% 47.413% 0
1 59 29.35% 35.383% 59
2 31 15.42% 13.203% 62
3 8 3.98% 3.284% 24
4 0 0.00% 0.613% 0
5 1 0.50% 0.091% 5
6 0 0.00% 0.011% 0
77oromore
più 0 0.00% 0.000% 0
Totali 201 100.000% 99.9987% 150
Mean ,l = 150 / 201 = 0.74627
These slides can not be reproduced or distributed without permission in writing from the author. 242
2.3.4.2. Poisson distribution
COMPARISON BETWEEN OBSERVED AND PREDICTED DISTRIBUTIONS
55%
50,7%
50%
47,4% Observed frequency
45% Poisson distribution predicted frequency
% of days with x defective units
40%
35,4%
35%
29,4%
30%
25%
20%
15,4%
15% 13,2%
10%
5% 4,0%
3,3%
2.3.4.4
Continuous distributions
These slides can not be reproduced or distributed without permission in writing from the author. 244
2.3.4.4. Continuous distributions
We will limit to :
Lognormal distribution;
Weibull distribution
Exponential distribution.
These slides can not be reproduced or distributed without permission in writing from the author. 245
2.3.4. Statistical distributions of current use
2.3.4.5
Rectangular or Uniform
distribution
These slides can not be reproduced or distributed without permission in writing from the author. 246
2.3.4.5. Rectangular or Uniform distribution
Called a and b the lower and upper limits of the definition interval,
the probability density function is given by:
1
f(x) = ----------
b-a
These slides can not be reproduced or distributed without permission in writing from the author. 247
2.3.4.5. Rectangular or Uniform distribution
Ordinates axis for CUMULATIVE distribution
cumulative function
a b
The above diagram shows the typical trend of the probability density function, from which
we immediately understand why an uniform distribution is also called rectangular.
This distribution is symmetric around is central vertical axis so that mean and median are
the same.
Of course cumulative distribution has linear trend and its value is 0 in a and 1 in b.
These slides can not be reproduced or distributed without permission in writing from the author. 248
2.3.4.5. Rectangular or Uniform distribution
These slides can not be reproduced or distributed without permission in writing from the author. 249
2.3.4.5. Rectangular or Uniform distribution
For instance, the 90 numbers of the lottery have all the same probability
to be drawn; their distribution is defined only for integer numbers between
1 and 90 (limits included).
The single probability to be drawn, equal for all numbers, is:
1/90 = 0.01111 1.1% .
These slides can not be reproduced or distributed without permission in writing from the author. 250
2.3.4. Statistical distributions of current use
2.3.4.6
Gauss normal distribution
These slides can not be reproduced or distributed without permission in writing from the author. 251
2.3.4.6. Gauss normal distribution
These slides can not be reproduced or distributed without permission in writing from the author. 252
2.3.4. Statistical distributions of current use
2.3.4.7
Lognormal distribution
These slides can not be reproduced or distributed without permission in writing from the author. 253
2.3.4.7. Lognormal distribution
These slides can not be reproduced or distributed without permission in writing from the author. 254
2.3.4.7. Lognormal distribution
mean-ln = 2.0
0,4
0,3
0,2
0,1
0,0 t
0 1 2 3 4
These slides can not be reproduced or distributed without permission in writing from the author. 255
2.3.4.7. Lognormal distribution
LOGNORMAL distrib.: examples of cumulative functions
(s always equal to 1)
1,1
1,0
0,9
Cumulative probability
0,8
0,7
0,6
0,5
0,4
mean-ln = 0.0
0,3
mean-ln = 0.5
0,2 mean-ln = 1.0
0,1
mean-ln = 2.0
0,0 t
0 2 4 6 8 10 12
These slides can not be reproduced or distributed without permission in writing from the author. 256
2.3.4.7. Lognormal distribution
s l2
ml + ------
2 (2 ml + sl2) s l2
mt = e s t2 = e . e -1
s t2 s t2
ml = ln (mt) - ------ s l2 = ln ( ------ + 1 )
2 mt2
These slides can not be reproduced or distributed without permission in writing from the author. 257
2.3.4.7. Lognormal distribution
Lognormal distribution
is frequently used to represent
the steel strength distribution,
when applying the methods of Probabilistic Design,
which will be discussed in the next Chapter 3.
These slides can not be reproduced or distributed without permission in writing from the author. 258
2.3.4. Statistical distributions of current use
2.3.4.8
Weibull distribution
These slides can not be reproduced or distributed without permission in writing from the author. 259
2.3.4.8. Weibull distribution
h = const. = 50,000
Probability density
These slides can not be reproduced or distributed without permission in writing from the author. 261
2.3.4.8. Weibull distribution
Shapes of cumulative Weibull distribution
h = const. = 50,000
b
t
- (-----)
h
Setting t = h in the previous F(t) = 1 - e , we obtain F(t) = 1 - e-1 = 0,632 whatever
the value of b. This explains the important feature of the Weibull cumulative distributions
(see Figure). All intersect at the same point, defined by the value of h and by the value of its
cumulative probability, which always corresponds to 63,2%, as just proved. Hence the name of
characteristic life for h is justified.
These slides can not be reproduced or distributed without permission in writing from the author. 262
2.3.4.8. Weibull distribution
These slides can not be reproduced or distributed without permission in writing from the author. 263
2.3.4.8. Weibull distribution
Starting from basic expression of the cumulative Weibull distribution:
b
t
- (-----)
h
F(t) = 1 - e
we can write: b
t
- (-----)
h1
b 1 - F(t) = e = --------------
t b
(-----) t
h1 (----)
e = -------------- h
1 - F(t) e
b
t 1
(-----) 1 b . ln t - b . ln h = ln ln --------------
h = ln -------------- 1 - F(t)
1 - F(t)
1
ln ln -------------- = b . ln t - constant
1 - F(t)
which, in appropriate coordinates (Weibull Chart), is the equation of a straight line.
These slides can not be reproduced or distributed without permission in writing from the author. 264
2.3.4.8. Weibull distribution
Cumulative probability
Distance covered [km]
PDF, cumulative functions and
corresponding Weibull’s straight lines
Probability density
These slides can not be reproduced or distributed without permission in writing from the author.
Distance covered [km] 265
2.3.4.8. Weibull distribution
Weibull straight lines of previous examples
Car model: PANDA 30 HP
Component: Braking System / DISK Defect: EARLY WEAR-OUT (replacement)
Car model: UNO 45 HP
Position: Front Tested components = 74
Component: Braking System / DISK Defect: EARLY WEAR-OUT (replacement)
Kind of test: C.C.C. (FIAT Customer Constant Control) of which failed = 12 Position: Front Tested components = 79
b = 3.71 h = 47,823 km Average life = 43,164 km Kind of test: C.C.C. (FIAT Customer Constant Control) of which failed = 6
Date 12/10/1984 b = 2.15 h = 78,243 km Average life = 69,292 km
Date 05/25/1985
These slides can not be reproduced or distributed without permission in writing from the author. 266
2.3.4.8. Weibull distribution
Weibull straight lines of previous examples
Car model: DELTA
Car model: REGATA (gasoline motors)
Component: HORN Defect: FAULTINESS (replacement)
SENSOR OF THE ROD
Position: - Tested components = 88 Component: Defect: FAULTINESS (replacement)
OF ENGINE OIL LEVEL
Kind of test: TEST 4 YEARS (LANCIA) of which failed = 11 Position: - Tested components = 58
b = 1.03 h = 219,239 km Average life = 216,791 km Kind of test: C.C.C. (FIAT Customer Constant Control) of which failed = 5
Date 05/25/1986
These slides can not be reproduced or distributed without permission in writing from the author. 267
2.3.4.8. Weibull distribution
Probability density
Densità di probabilità
0 10.000 20.000 30.000 40.000 50.000 60.000 70.000 80.000 90.000 100.000
Percorrenza
Distance [km][km]
covered
These slides can not be reproduced or distributed without permission in writing from the author. 268
TYPE OF x = failed;
SAMPLE o = suspended SAMPLE DEFINITIONS
x
COMPLETE n = number of elements
x
x g=n in the sample;
x
x
x
(t = undefined) g = number of failed ele-
x If g = n ≥ 5, we are in the scope of the
Weibull.
ments, during the
t test;
TRUNCATED (end time)
g* = minimum number of
t = t* failed elements
S A M P L E S
x
x
x (t = undefined)
O If g ≥ 5, we are in the scope of the The term “incomplete sample” is
x Weibull. not the opposite of the term “com-
plete sample”, because this “opposi-
t te" is represented by any one of the
INCOMPLETE types of "suspended sample" (one
of which is the “incomplete sam-
(with some “suspended” before the ple”, which is characterized by some
last failed) item removed from the test still wor-
x
O 0<g<n king, before the last “failed"). A spe-
cial case of "opposite“ (in the sense
x
x
x (t = undefined) of "antipodes”) of the complete sam-
x If g ≥ 5, we are in the scope of the ple is the truncated sample with
O Weibull. zero “failed” (success run).
t
These slides can not be reproduced or distributed without permission in writing from the author. 269
2.3.4.8. Weibull distribution
SITUATIONS AND TOOLS RELATED
WITH WEIBULL DISTRIBUTION EXPRESSIONS
PROBLEMS TOOLS EXPRESSIONS
COMPLETE or TRUNCATED
MEDIAN RANKS (M.R.) For complete , truncated
or CENSORED SAMPLES
and censored samples
The expression on the right is valid Median ranks (MR.) give a good (stabilized) esti- (= without suspended before last failed):
only if suspended elements (if there mation of cumulative distribution. They are obtai-
are) are all successive or simulta- ned, by interpolation, from BETA distribution. The i - 0.3
M.R. =
neous to the last failed element. most used approximation is the expression in the N + 0.4
right. For more precise calculations, TABLES are where:
Complete sample = N failed elements ;
Truncated sample = test ends after a stated number of km,
available in Literature. i = data sequential index (in this case,
hours or cycles: failed number < N;
it concerns only the failed elements;
Censored sample = test ends when a stated number of k
elements have been failed, with k < N.
N = total number of tested elements.
These slides can not be reproduced or distributed without permission in writing from the author. 270
2.3.4.8. Weibull distribution
1. example of calculations with a COMPLETE SAMPLE: needs of median ranks
PROBLEMS TOOLS EXPRESSIONS
COMPLETE or TRUNCATED
MEDIAN RANKS (M.R.) For complete , truncated
or CENSORED SAMPLES
and censored samples
The expression on the right is valid Median ranks (MR.) give a good (stabilized) esti- (= without suspended before last failed):
In a complete sample, as shown in the Table below, the simple only if suspended elements (if there
are) are all successive or simulta-
neous to the last failed element.
mation of cumulative distribution. They are obtai-
ned, by interpolation, from BETA distribution. The
most used approximation is the expression in the
M.R. =
i - 0.3
N + 0.4
cumulative percentages of elements failed during the test (full Complete sample = N failed elements ;
Truncated sample = test ends after a stated number of km,
hours or cycles: failed number < N;
right. For more precise calculations, TABLES are where:
available in Literature. i = data sequential index (in this case,
blue curve in the next slide) correctly represent what has happe-
it concerns only the failed elements;
Censored sample = test ends when a stated number of k
elements have been failed, with k < N.
N = total number of tested elements.
ned in the test sample, but can not be assumed as a general fai- (at least one failed is preceded
by one or more suspended)
The elements removed from the test
and ORDER INCREMENT (O.I.)
When there are suspended (obviously before a
O.I. =
N+2-i
lure law to be used for predictions, if only because, repeating the before they fail (after different test
times and for different reasons) are
called suspended .
failed ), the data sequential index , i, includes, not
only the failed (as before), but also the su-
spended . In the expression of Median Ranks it is
O.N. = O.N.previous + O.I.
O.N. - 0.3
experiment with other samples in nominally equal conditions (dot- substituted by the Order Number O.N. evaluated
by Order Increment O.I.. Thus, the Median Ranks where:
calculations (always conducted only for the failed
M.R. =
N + 0.4
ted red curves in the next slide), the time at which the last element elements) take into account of suspended too.
The O.N. starts with a value of zero.
ed and suspended in the order in which
they occur.
Current Percentage
Test time
data Cumulative percent of failed elements
for each
sequen- of failed elements calculated through
element
tial index the median ranks
[hours]
i Formula % Formula %
1 - 0,3
1 30 1/7 14,29% 9,46%
7 + 0,4
2 - 0,3
2 60 2/7 28,57% 22,97% Instead, by calculating the correct percenta-
7 + 0,4
3 - 0,3 ges according to the median ranks, the fai-
3 90 3/7 42,86% 36,49%
7 + 0,4 lure cumulative percentage corresponding to
4 195 4/7 57,14%
4 - 0,3
50,00%
the last failed element is always less than
7 + 0,4 100% and it may prove that the general fai-
5 215 5/7 71,43%
5 - 0,3
63,51%
lure law (useful for predictions) thus obtained
7 + 0,4 is the best possible among those deductible
6 240 6/7 85,71%
6 - 0,3
77,03% from the data of the single tested sample.
7 + 0,4
7 - 0,3
7 275 7/7 100,00% 90,54%
7 + 0,4
These slides can not be reproduced or distributed without permission in writing from the author. 271
2.3.4.8. Weibull distribution
Graphical display of the concepts mentioned in the previous slide:
1. classical cumulative curve (100% of failed at 275 hours);
2. examples of different results that can occur when the experiment is repeated several times in nominally equal conditions;
3. curve of median ranks.
FAILED ELEMENTS
100%
OF CONSUNTIVATI
90%
80%
70%
DI ROTTI
60%
PERCENTAGE
50%
PERCENTUALE
40%
30%
CUMULATIVE
20%
10%
0%
0 50 100 150 200 250 300 350 400
TEST
ORETIME [hours]
DI PROVA
These slides can not be reproduced or distributed without permission in writing from the author. 272
2.3.4.8. Weibull distribution
These slides can not be reproduced or distributed without permission in writing from the author. 273
2.3.4.8. Weibull distribution
2. example of calculations with an INCOMPLETE SAMPLE PROBLEMS
COMPLETE or TRUNCATED
or CENSORED SAMPLES
TOOLS
INCOMPLETE SAMPLES
this case, there are also 5 suspended elements removed ORDER NUMBER (O.N.) N + 1 - O.N.previous
(at least one failed is preceded O.I. =
and ORDER INCREMENT (O.I.) N+2-i
by one or more suspended)
The elements removed from the test When there are suspended (obviously before a
from the test at the times indicated in the Table below. before they fail (after different test
times and for different reasons) are
called suspended .
failed ), the data sequential index , i, includes, not
only the failed (as before), but also the su-
spended . In the expression of Median Ranks it is
substituted by the Order Number O.N. evaluated
O.N. = O.N.previous + O.I.
M.R. =
O.N. - 0.3
N + 0.4
by Order Increment O.I.. Thus, the Median Ranks where:
Current calculations (always conducted only for the failed
elements) take into account of suspended too.
i = data sequential index including both fail-
ed and suspended in the order in which
Test time Failed (F) The O.N. starts with a value of zero. they occur.
data
for each or
sequential
element Suspended
index
[hours] (S)
i
Wanting at all costs to avoid recourse to the ORDER
1 25 S
NUMBER and the ORDER INCREMENT, we could:
2 25 S
3 30 F
1. ignore all suspended, giving up some useful infor-
mation (and already paid!); we would fall in the pre-
4 60 F vious case, but obtaining a worst case estimate;
5 70 S 2. imagine that any suspended element was working
6 90 F
when the last element failed; which is equivalent to
attribute to all suspended elements an equal test time of
7 110 S 275 hours, that is, to transform the incomplete sample
8 180 S in a censored sample; obviously, in doing so, we would
obtain a best case estimate.
9 195 F
10 215 F
In concluding this topic, it will make a comparison be-
tween these results and the correct ones, obtainable by
11 240 F using the ORDER NUMBER and ORDER INCREMENT.
12 275 F
These slides can not be reproduced or distributed without permission in writing from the author. 274
2.3.4.8. Weibull distribution
2a. example of calculations with a CENSORED SAMPLE
Current Percentage
Test time
data Cumulative percent of failed elements
for each
sequen- of failed elements calculated through
element
tial index the median ranks
i
[hours] Current Percentage
Formula % Formula % Test time
1 - 0,3 data Cumulative percent of failed elements
1 30 1/7 14,29% 9,46% for each
7 + 0,4 sequen- of failed elements calculated through
2 60 2/7 28,57%
2 - 0,3
22,97%
element
7 + 0,4 tial index the median ranks
[hours]
3 90 3/7 42,86%
3 - 0,3
36,49% i Formula % Formula %
7 + 0,4
4 - 0,3 1 - 0,3
4 195 4/7 57,14%
7 + 0,4
50,00% 1 30 1/12 14,29% 5,65%
12 + 0,4
5 - 0,3
5 215 5/7 71,43% 63,51%
7 + 0,4 2 - 0,3
6 - 0,3 2 60 2/12 28,57% 13,71%
6 240 6/7 85,71% 77,03% 12 + 0,4
7 + 0,4
3 - 0,3
7 275 7/7 100,00%
7 - 0,3
90,54% 3 90 3/12 42,86% 21,77%
7 + 0,4 12 + 0,4
4 - 0,3
4 195 4/12 57,14% 29,84%
12 + 0,4
The present application is 5 215 5/12 71,43%
5 - 0,3
12 + 0,4
37,90%
very similar to the previous 6 240 6/12 85,71%
6 - 0,3
45,97%
one: the only difference is 12 + 0,4
7 - 0,3
7 275 7/12 100,00% 54,03%
that, having imagined that 12 + 0,4
have reached the end of 9 Placed at the end, the "suspended" ele-
ments only serve to change the denomi-
the test, the percentages 10 nator, which runs from 7 to 12 (the cal-
no longer refer to a total 11
culation of the percentages of failed ele-
ments stops with the last failed ).
number of elements equal
12
to 7 but equal to 12.
These slides can not be reproduced or distributed without permission in writing from the author. 275
2.3.4.8. Weibull distribution
These slides can not be reproduced or distributed without permission in writing from the author. 276
2.3.4.8. Weibull distribution
2.b. example of correct calculations for an INCOMPLETE SAMPLE
Current
Test time Failed (F)
data
for each or Order Increment Order Number Median Rank
sequential
index
element Suspended O.I. O.N. M.R.
[hours] (S)
i
1 25 S
2 25 S
12 + 1 - 0 1.182 - 0.3
3 30 F = 1.182 0 + 1.182 = 1.182 = 0.071 0.071
12 + 2 - 3 12 + 0.4
12 + 1 - 1.182 2.364 - 0.3
4 60 F = 1.182 1.182 + 1.182 = 2.364 = 0.166 0.166
12 + 2 - 4 12 + 0.4
5 70 S
12 + 1 - 2.364 3.693 - 0.3
6 90 F = 1.330 2.364 + 1.330 = 3.694 = 0.274 0.274
12 + 2 - 6 12 + 0.4
7 110 S
8 180 S
12 + 1 - 3.693 5.555 - 0.3
9 195 F = 1.861 3.693 + 1.861 = 5.555 = 0.424 0.424
12 + 2 - 9 12 + 0.4
12 + 1 - 5.555 7.416 - 0.3
10 215 F = 1.861 5.555 + 1.861 = 7.416 = 0.574 0.574
12 + 2 - 10 12 + 0.4
12 + 1 - 7.416 9.277 - 0.3
11 240 F = 1.861 7.416 + 1.861 = 9.277 = 0.724 0.724
12 + 2 - 11 12 + 0.4
12 + 1 - 9.277 11.139 - 0.3
12 275 F = 1.861 9.277 + 1.861 = 11.138 = 0.874 0.874
12 + 2 - 12 12 + 0.4
NOTE - If there are not the two suspended at the beginning, the first failed element would have the data sequential index, i, equal to 1, an Order In-
crement, O.I., equal to (12+1)/(12+2-1) = 1, an Order Number equal to 1 (i.e. equal to the Order Increment) and a Median Rank, M.R., equal
to (1-0.3)/(12+0.4) = 0.056 (5.6%).
These slides can not be reproduced or distributed without permission in writing from the author. 277
2.3.4.8. Weibull distribution
N O T E
In the previous example, we can observe that, when there are not "suspended" elements between two "failed",
the value of the Order Increment (O.I.) remains constant.
This can be justified putting the Order Increment of a failed equal to k:
N + 1 - O.N.0
= k
N+2-i
The Order Increment of the failed immediately following (if there are no suspended elements between the two) will be given by:
(N + 1 - O.N.0) . (N + 2 - i - 1)
(N + 2 - i) (N + 1 - O.N.0) . (N + 2 - i - 1)
= = = k which
which was to be
be proved.
proven
(N + 2 - i - 1) (N + 2 - i) . (N + 2 - i - 1)
This reasoning may be repeated in a similar way for all next failed elements (without suspended elements interposed).
These slides can not be reproduced or distributed without permission in writing from the author. 278
2.3.4.8. Weibull distribution
These slides can not be reproduced or distributed without permission in writing from the author. 279
2.3.4.8. Weibull distribution
These slides can not be reproduced or distributed without permission in writing from the author. 280
As shown in the Weibull Chart at the
side, the more reliable failure law
(full black line), obtained with the cor-
rect calculation that uses the order
number and the order increment, is
intermediate with respect to those
obtainable with the two simplifica-
tions discussed above:
1. to completely neglect the suspen-
ded elements: in this way, the
incomplete sample becomes a
complete sample and obviously
we get a pessimistic failure law
(dashed blue line);
2a. to imagine that all suspended ele-
ments would have arrived at the
end of test without failing: in this
way, the incomplete sample be-
comes a censored sample and
obviously we get an optimistic
failure law (dotted red line).
These slides can not be reproduced or distributed without permission in writing from the author. 281
2.3.4.8. Weibull distribution
WEIBULL CALCULATED
BY CLASSES OF TIME/MILEAGES
These slides can not be reproduced or distributed without permission in writing from the author. 282
2.3.4.8. Weibull distribution
For WEIBULL BY CLASSES, in the absence of anything better, we propose
the EXCEL sheet WeibullBYclasses.xls, located in the directory MiniSW.
h estimation AUXILIARY CALCULATIONS
EXPERIMENTAL WEIBULL
EXPERIMENTAL data NORMAL distribution WEIBULL distribution Searching among cumulative h data data
WARNING corrected fractions estimated
Classes Frequency Cum. Cumulative value from
calculations for calculations for
Probability Cumulative Probability Cumulative AREA
n° of abs. corrected experimental
Lower Upper Class Cumulative (if any)
clas- Absolute Fraction Density density fraction density fraction decision making data MEAN STD.DEV. MEAN STD.DEV.
limit limit center fraction values fraction
ses
1 0 100 50 0 0 0,00000 0,00000 0 0,00000 0,00000 0,00000 0,00000 0,00000 no 0,00 0,00 0,00 0,13
2 100 200 150 0 0 0,00000 0,00000 0 0,00000 0,00000 0,00010 0,00000 0,00053 no 0,00 0,00 0,05 55,93
3 200 300 250 0 0 0,00000 0,00000 0 0,00000 0,00002 0,00343 0,00007 0,00862 no 0,00 0,00 1,69 645,32
4 300 400 350 0 0 0,00000 0,00000 0 0,00000 0,00036 0,04521 0,00048 0,06095 no 0,00 0,00 16,84 2.092,36
5 400 500 450 25 0,29411765 0,00294 0,29412 25 0,28923 0,00199 0,24724 0,00188 0,25376 no 11.250,00 346.020,76 84,71 2.217,31
6 500 600 550 30 0,35294118 0,00353 0,64217 55 0,64052 0,00397 0,62807 0,00402 0,64238 found 597,5760 16.500,00 9.342,56 221,23 29,29
7 600 700 650 20 0,23529412 0,00235 0,87581 75 0,87471 0,00285 0,90934 0,00318 0,94895 13.000,00 135.640,14 206,40 2.656,54
8 700 800 750 10 0,11764706 0,00118 0,99235 85 0,99180 0,00074 0,99053 0,00037 0,99943 7.500,00 332.525,95 27,53 1.345,43
9 800 900 850 0 0 0,00000 0,99180 85 0,99180 0,00007 0,99961 0,00000 1,00000 0,00 0,00 0,09 9,29
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
n° of classes = 9 S= 85 Mean = 567,65 Dmeans = -9,11 Mean = 558,53 597,5760 48.250,00 823.529,41 558,53 9.051,60
Although a little more sophisticated, the basic setting of the calculations is very similar to that
which will be soon explained when discussing the parametric method of interpolation.
These slides can not be reproduced or distributed without permission in writing from the author. 283
2.3.4.8. Weibull distribution
ADDITIONAL DETAILS
1. The calculation of the mean and variance of the Weibull distri-
bution, starting from shape (b) and scale (h) parameters:
m = h . G(1 + 1/b); s2 = h2 . G(1 + 2/b) - m2
requires the use of the Gamma function (G), whose exposure
exceeds the limits of this Course.
NOTE - In EXCEL we found the Gamma distribution, f(x, a, q), which is not the
Gamma function, G(a), of our interest. This last is defined as:
G(a) = xa-1 e-x dx
0
If a is a positive integer equal to k, we have: G(k) = (k -1)! .
If a is greater than 1, we have: G(a) = (a -1) . G(a -1) .
For large values of a, we can use Sterling approximate formula:
G(a +1) = 2 . . a . (a/e)a (where e is the basis of natural logarithms)
In an EXERCISE we will see an example of interpolation with the Weibull distribution using
the Solver of EXCEL.
These slides can not be reproduced or distributed without permission in writing from the author. 284
2.3.4.8. Weibull distribution
These slides can not be reproduced or distributed without permission in writing from the author. 285
2.3.4. Statistical distributions of current use
2.3.4.9
Exponential distribution
These slides can not be reproduced or distributed without permission in writing from the author. 286
2.3.4.9. Exponential distribution
m = h
s = h
These slides can not be reproduced or distributed without permission in writing from the author. 287
2.3.4.9. Exponential distribution
2.3.4.10
Summary of main
statistical distributions
These slides can not be reproduced or distributed without permission in writing from the author. 289
2.3.4.10. Short summary of main statistical distributions
Short summary of the key concepts about the most important statistical distributions
DISTRIBUTION TYPE NUMBER OF PARAMETERS FEATURES MAIN APPLICATIONS MAIN RESULTS REMARKS
It is the most widely used distribution A small difference between the mean and the median
2 parameters: Simmetry: i.e. equiprobability of results
Normal continouos Distribution of sample means. (sometimes, for convenience, even of a sample ensures that data are acceptably distributed
m and s. by defect or by excess . according to the normal distribution.
inappropriately!).
Bernoulli's experiments Probability of seeing in the The probabilities of success , p, and of not-success ,
2 parameters: Criteria for designing
Binomial discrete (with only 2 possible outcomes): sample at least one defect (one q = 1 - p, may be different (and not the same as in the
p and n. reliability experimental tests. case of a fair coin).
e.g. Head & Tail . or more) of a given type.
These slides can not be reproduced or distributed without permission in writing from the author. 290
2.3.4.10. Short summary of main statistical distributions
FEATURES MEAN VARIANCE TYPICAL APPLICATIONS R E M A R K S
Continuous
Symmetric
(Gauss)
m and s
distribution, rapidly sloping down as you move away from the mean perly used. It is the best known and most widely used distribution: sometimes
mean)
m s2 (e.g.: the measurement errors when those "by defect" are used even improperly. Mean, median and mode coincide and therefore a small
equiprobable to those "by excess"). difference between the mean and the median (of the sample data) provides the
It is the typical distribution of sample means. acceptability of their interpretation according to the normal law.
Confidence intervals with small samples (n 30). This distribution For large values o f n, the Student's t distribution coincides with the normal
(or d.o.f. = n-1)
Continuous
m, sandn
Symmetric
takes into account the greater uncertainty due to the fact that thesof distribution.
mean)
2 the sample is no longer a good estimate of that one of the For greater precision, it is common to resort to Student's t distribution even for
m s
t
population. It is as if it were used a normal distribution more and values of n greater than 30. If you are working manually (using Tables), the
more lower and wider, while the sample size decreases from n = 30 t-distribution is not used beyond n = 50, whereas if you use your computer, the
to n = 2. t-distribution may also be used beyond n = 100.
The domain may also be made up of several separate segments on the abscis-sa
(constant) is given by: y =
a and b extremes of the
definition domain on the
2 parameters:
RECTAN-
Continuous
(e.g. in the case of integer numbers, for which is performed the continuity
(b - a)2
GULAR
a+b
a and b
Situations of "equiprobability": eg., in the case of the 90 numbers of correction (1)), but in any case it must be limited both left and right.
1/(b-a).
a batch of established size. This is useful for batch control (in Quality noulli's experiment). Examples: outcomes of heads or tails in successive tos-
for p = 0,5 simmetric
n.p.q ses of a coin, presence of a face ≤ 2 in the throw of a die, extraction of a white
Discrete
.
and
n p verifications. ball from a box containing also black balls in known proportion, extraction from a
batch of an operating or defective component, etc.. In general, the Binomial di-
p
(with q = 1-p)
It is also the basis for the definition of confidence intervals for stribution gives the probability of a predetermined number of " successes " in
proportions: e.g. the number of defective units per million a given number of experiments.
(dpm), regardless of the number of defects in each unit.
The Poisson distribution is the limit of the Binomial when p → 0 and n → ∞. Since
Right skewed
1 parameter:
POISSON
the variance is equal to the mean, it is not necessary to evaluate s from the data.
l=n p
Discrete
l l known the mean l, we can calculate the percentage of units (e.g. However there is good correlation with reality only if the variability of successive
cars) with 0, 1, 2, 3, ... defects (each). batches of each component remains substantially constant over time, but if it
fluctuates, as often happens, the variability is higher than that assumed by the
Poisson's model.
1 parameter:
NENTIAL
Exponential
Continuous
ESPO-
h h2 which have constant the hazard rate : It is obtainable from the Weibull distribution for b = 1.
h
Continuous
to 63.2%;
by the zero
b and h
the help of the Gamma " Failure laws " of different shapes
b = shape parameter ; if it is:
distribution (usually indicated depending on components,
b < 1 : infant mortality failures
with a G) or by computerized usually represented in Weibull Cards .
recursive procedures b = 1 : random or useful life failures (esponential distribution )
b > 1 : wear-out failures.
(1) The continuous distributions can also be used (always in a continuous manner) to address problems in the field of discrete. For example, having to do with balls, numbered from 1 to 10, extracted from a box containing them in varying
quantities, the probability that is extracted a ball with the number 7 (or any other assigned number) is given by the area under the curve between the abscissas 6.5 and 7.5: this way of operating is called " continuity correction " .
These slides can not be reproduced or distributed without permission in writing from the author. 291
2.3. Probability Theory
2.3.5
A brief mention of
interpolating problems
These slides can not be reproduced or distributed without permission in writing from the author. 292
2.3.5. Interpolating problems
STATISTICAL DISTRIBUTION
These slides can not be reproduced or distributed without permission in writing from the author. 293
2.3.5. Interpolating problems
These slides can not be reproduced or distributed without permission in writing from the author. 294
2.3.5. Interpolating problems
These slides can not be reproduced or distributed without permission in writing from the author. 295
2.3.5. Interpolating problems
These slides can not be reproduced or distributed without permission in writing from the author. 296
2.3.5. Interpolating problems
Typical supply problem
A restaurant specializing in
"T-bone steaks” recorded 41 52 46 42 46 36 44 68 58 44
during 50 successive Sun- 49 48 48 65 52 50 45 72 45 43
days, the number of steaks 47 49 57 44 48 49 45 47 48 43
requested and served, as 45 56 61 54 51 47 42 53 41 45
58 55 43 63 38 42 43 46 49 47
shown in the table to the
right.
These slides can not be reproduced or distributed without permission in writing from the author. 297
2.3.5. Interpolating problems
These slides can not be reproduced or distributed without permission in writing from the author. 298
2.3.5. Interpolating problems
45%
40% Istogramma
Histogram
Normal
Normale
35% 34,0%
the classes
CLASSI
32,0%
30%
NELLE
25%
DOMENICHEin
%%ofDISundays
20%
15%
12,0%
10%
8,0%
6,0%
5% 4,0%
2,0% 2,0%
0%
40 45 50 55 60 65 70 75
Number of requested
NUMERO “T-bone
DI FIORENTINE steaks”
RICHIESTO (esposti estremivalues
(abscissa referdelle
superiori classi)
to class upper limits)
These slides can not be reproduced or distributed without permission in writing from the author. 299
2.3.5. Interpolating problems
Parametric Method
These slides can not be reproduced or distributed without permission in writing from the author. 302
To better evidence, the centers of the upper bases of the columns of the above histogram have been replaced with
large diamond-shaped points.
PROBABILITY DENSITY
0,08
Experimental data
0,07 Normal distribution
Weibull distribution
0,06
Probability density
0,03
0,02
0,01
0,00
35 40 45 50 55 60 65 70 75
Abscissa
These slides can not be reproduced or distributed without permission in writing from the author. 303
2.3.5. Interpolating problems
CUMULATIVE PROBABILITY
1,00
0,90
0,80
On the right side of the graph,
Cumulative Probability
0,40
0,30
Abscissa
These slides can not be reproduced or distributed without permission in writing from the author. 304
2.3.5. Interpolating problems
But the Parametric method is not the only one, there is also the
Maximum Likelihood Estimation (MLE) method,
which is based on cumulative functions.
M A X I M U M L I K E L I H O O D E S T I M A T I O N (M.L.E.) I N T E R P O L A T I O N S
EXPERIMENTAL data NORMAL distribution WEIBULL distribution
Classes Frequency Cumulated Cumulative
Optional Probability Cumulative Optional Probability Cumulative
n° of
Lower Upper Class absolute correct(1) D2 D2
clas- Absolute Fraction Density WEIGHTS density fraction WEIGHTS density fraction
ses limit limit center values fraction
1 35 40 37,50 2 0,04 0,00800 2 0,03373 1,0 0,01850 0,12484 0,00830 1,0 0,01880 0,15295 0,01421
2 40 45 42,50 17 0,34 0,06800 19 0,37103 1,0 0,04307 0,33832 0,00107 1,0 0,03709 0,33851 0,00106
3 45 50 47,50 16 0,32 0,06400 35 0,68849 1,0 0,05850 0,62437 0,00411 1,0 0,05464 0,60720 0,00661
4 50 55 52,50 6 0,12 0,02400 41 0,80754 1,0 0,04635 0,85338 0,00210 1,0 0,05144 0,85841 0,00259
5 55 60 57,50 4 0,08 0,01600 45 0,88690 1,0 0,02143 0,96288 0,00577 1,0 0,02356 0,97839 0,00837
6 60 65 62,50 3 0,06 0,01200 48 0,94643 1,0 0,00578 0,99412 0,00227 1,0 0,00338 0,99920 0,00278
7 65 70 67,50 1 0,02 0,00400 49 0,96627 1,0 0,00091 0,99943 0,00110 1,0 0,00008 1,00000 0,00114
8 70 75 72,50 1 0,02 0,00400 50 0,98611 1,0 0,00008 0,99997 0,00019 1,0 0,00000 1,00000 0,00019
Initialize (as a
experi-mental
(1) i.e.: ---------------
the experimental
general rule)
h = mean of
weight a value of 0, the point will be excluded from the calculation (as if there was n + 0,4
not). Obviously, if all weights are 1, all points are kept in the same account. Often
b = 2;
greater weight is given to the central points or to points of the side where we are
with:
data
data
interested in the value of the tail area.
In this case, the SOLVER function must be used
once for the Normal and once for the Weibull.
These slides can not be reproduced or distributed without permission in writing from the author. 305
2.3.5. Interpolating problems
CUMULATIVE PROBABILITY
1,0
0,9
0,8
Cumulative probability
0,4
0,3
Abscissa
These slides can not be reproduced or distributed without permission in writing from the author. 306
To better evidence, the centers of the upper bases of the columns of the above histogram have been replaced with
large diamond-shaped points.
PROBABILITY DENSITY
0,08
Experimental data
0,07 Normal distribution
Weibull distribution
0,06
Probability density
0,03
0,02
0,01
0,00
35 40 45 50 55 60 65 70 75
Abscissa
These slides can not be reproduced or distributed without permission in writing from the author. 307
2.3.5. Interpolating problems
These slides can not be reproduced or distributed without permission in writing from the author. 308
2.3.5. Interpolating problems
Answer to the 2nd question Answer to the 1st question
100% 100%
98%
96%
90% 90%
classes
82%
80%
the CLASSI
Sunday, at least
60 – 10 = 50
DI DOMENICHE
50% 60 steaks
40%
34% 38%
32%
% of
30%
%
20%
12%
10% 8%
6%
4%
4% 2% 2%
0%
35 40 45 50 55 60 65 70 75
Number
NUMERO DI of requested
FIORENTINE “T-bone
RICHIESTO steaks”
IN UNA DOMENICA
These slides can not be reproduced or distributed without permission in writing from the author. 309
2.3.5. Interpolating problems
These slides can not be reproduced or distributed without permission in writing from the author. 310
2.3.5. Interpolating problems
2.3.6
Possible deductions when we can not use
a theoretical distribution:
Tchébychev’s inequality
These slides can not be reproduced or distributed without permission in writing from the author. 312
2.3.6. Tchébychev’s inequality
These slides can not be reproduced or distributed without permission in writing from the author. 313
2.3.6. Tchébychev’s inequality
These slides can not be reproduced or distributed without permission in writing from the author. 314
2.3.6. Tchébychev’s inequality
mean
1
fraction of data > 1 - ------
k2
m
m-ks m+ks
These slides can not be reproduced or distributed without permission in writing from the author. 315
2.3.6. Tchébychev’s inequality
This theorem is true for any kind of data: no matter if they relate
to populations or to samples, to continuous or to discrete distri-
butions. Moreover it is always true regardless of the shape of
the distribution.
These slides can not be reproduced or distributed without permission in writing from the author. 316
2.3.6. Tchébychev’s inequality
1 1 1
1 - ----- = 1 - ----- = 1 - ----- = 1 - 0.25 = 0.75 75%
k2 22 4
This can also be expressed by saying that in no more than 25% of the
months the production will be outside the range 2,200÷3,800 cars.
These slides can not be reproduced or distributed without permission in writing from the author. 317
2.3.6. Tchébychev’s inequality
In conclusion, in the above example, we cannot say that in each tail there are less than
25%/2 = 12.5% of data, because, in general, we cannot say that data are distributed a
half in a tail and the other half in the other tail ! The only thing we can say is that the
sum of two tails contains a data fraction less than 25% and that this value applies, a
fortiori, for each tail, which surely contains less than 25% of data (this is the same thre-
shold value of the sum of two tails and not the half corresponding to 12.5%).
Of course, if the random variable is defined only for positive values (produced units,
number of operations, etc.), if the value m - k s is negative, the Tchébychev lower li-
mit must be put equal to zero.
These slides can not be reproduced or distributed without permission in writing from the author. 318
2.3.6. Tchébychev’s inequality
1
k = ----------
1 - P
k = 1 / P
These slides can not be reproduced or distributed without permission in writing from the author. 319
2.3.6. Tchébychev’s inequality
These slides can not be reproduced or distributed without permission in writing from the author. 320
2.3.6. Tchébychev’s inequality
TCHÉBYCHEV's INTERVAL NORMAL DISTRIBUTION
or for a comparison
upper percentile with the same upper percentile
1 k z % of data
lower upper
P = 1 - ----- [units of standard [units of within
limit limit
k2 deviation s] standard deviation s ] mz s
These slides can not be reproduced or distributed without permission in writing from the author. 321
2. Fundamentals of Applied Statistics
2.4
S T A T I S T I C A L
I N F E R E N C E
These slides can not be reproduced or distributed without permission in writing from the author. 322
2.4. Statistical inference
The word inference is the most general form for "reasoning", which
can be referred both to deductions and to inductions. “Statistical
inference” is the process that, starting from information collected
from a sample, allows to reach conclusions about the population
from which the sample was taken.
These slides can not be reproduced or distributed without permission in writing from the author. 323
2.4. Statistical inference
2.4.1
Estimates
These slides can not be reproduced or distributed without permission in writing from the author. 324
2.4.1. Estimates
2.4.1.1
Punctual estimates
These slides can not be reproduced or distributed without permission in writing from the author. 325
2.4.1.1. Punctual estimates
Without going into details, we simply remember that
Statistics proves that:
the best “estimator” of the “true” mean of a population is the
mean of the sample;
the best “estimator” of the “true” proportion in a population
is the proportion detected on the sample, provided that the
proportion in the sample is enough far away from 0 and from
1 (which is practically the same situation)(1);
the best “estimator” of the “true” standard deviation of a
population is the standard deviation of the sample(2).
(1) Even when the proportion of "successes" detected on the sample is zero, it is plausible that the true propor-
tion of successes in the whole population (but also simply in a larger sample) is greater than zero (see also
at the end of the discussion of the PROPORTIONS, Par. 2.4.1.8). It is also obvious that something similar hap-
pens when the proportion is very close to 1, because, by simply exchanging the term "success" with the
term “not-success“, you fall in the previous case (with the proportion close to 0).
(2) Provided that we have calculated the sample standard deviation dividing the sum of squares or deviance
(numerator) by n - 1. If we would divide the sum of squares by n (instead of by n-1), in order to obtain an
“unbiased” estimator of the standard deviation of the population, we need to multiply this value of the sam-
ple standard deviation by the corrective coefficient: n / (n – 1) , (where n is the sample size, as usual).
These slides can not be reproduced or distributed without permission in writing from the author. 326
2.4.1. Estimates
2.4.1.2
Distribution of sample means
These slides can not be reproduced or distributed without permission in writing from the author. 327
2.4.1.2. Distribution of sample means
The probability that, drawing a ticket at random, we find that one with
the number 3 is the same as finding a ticket with any other number.
DRI S
E TCRTIAB N
UG Z IUOLNAER R D
E ITST TAR
N IGBOULTAI ROEN
(equally probable outcomes)
o dell'equiprobabilità
0,24 This probability,
0,22
0,20 the same
density
Densità di probabilità
0,18
0,16 Probability density is equal
Densità di probabilità to i0.142857
per tutti bigliettinifor allathe
pari tickets
0,142857 for all tickets,
0,14
is equal to
Probability
0,12
0,10
0,08
0,06
1/7 0.143
0,04
0,02 corresponding to
0,00
0 1 2 3 4 5 6 7 8 14.3%
B ITGI L
CI K
EET TTI S
NI
These slides can not be reproduced or distributed without permission in writing from the author. 328
2.4.1.2. Distribution of sample means
These slides can not be reproduced or distributed without permission in writing from the author. 329
2.4.1.2. Distribution of sample means
These slides can not be reproduced or distributed without permission in writing from the author. 330
2.4.1.2. Distribution of sample means
In total there are 35 different samples with 3 tickets each C L
samples means samples means samples means samples means samples means
15
The histogram shows that the
14
means of samples 13 … well approximated by a
13 with 3 tickets each NORMAL DISTRIBUTION ...
12 are distributed as a
11 symmetric bell ... We saw before that the
10 mean remain the same (4)
9 9
9
0 1 2 3 4 5 6 7 8
These slides can not be reproduced or distributed without permission in writing from the author. 333
2.4.1.2. Distribution of sample means
15
14
13
13
12
11
10
9 9
9
4
… and the same Of course the mean of
3 happens, on the left the biggest sample (6)
side, for the smaller 2 2 is smaller than the big-
2 gest ticket (7) …
sample
1
0 1 2 3 4 5 6 7 8
These slides can not be reproduced or distributed without permission in writing from the author. 334
2.4.1.2. Distribution of sample means
Moving from counting the number of elements (tickets)
0,50 and of the number of samples
0,45
to analyzing their statistical distributions, we obtain:
0,40 Sample
PROBABILITÀ
0,20
0,15
0,25
i.e. in this case
the ticket with the number 5
DENSITÀ
0,20
0,15
0,10
0,05
0,00
0 1 2 3 4 5 6 7 8
Valori of
Values deiTICKETS
BIGLIETTINI
or ofothe
delle MEDIE
MEANS ofdei campioni
samples withda
33 bigliettini
tickets each
These slides can not be reproduced or distributed without permission in writing from the author. 336
2.4.1.2. Distribution of sample means
These slides can not be reproduced or distributed without permission in writing from the author. 337
2.4.1.2. Distribution of sample means
0,50
0,45
0,40
PROBABILITÀ
0,35
DENSITÀ DIDENSITY
0,30
0,25
PROBABILITY
0,20
0,15
0,10
0,05
0,00
0 1 2 3 4 5 6 7 8
Values
Valori of
deiTICKETS or ofothe
BIGLIETTINI MEANS
delle ofdei
MEDIE samples withda
campioni 33tickets each
bigliettini
These slides can not be reproduced or distributed without permission in writing from the author. 338
2.4.1.2. Distribution of sample means
PROBABILITY DENSITY
DENSITÀ DI PROBABILITÀ
DENSITÀ DI PROBABILITÀ
0,15 0,15
n=5 n=5
PROBABILITY
0,10 0,10
0,05 0,05
0,00 0,00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
These slides can not be reproduced or distributed without permission in writing from the author. 339
Additional examples…
These slides can not be reproduced or distributed without permission in writing from the author. 340
2.4.1.2. Distribution of sample means
sN - n
sx = ------ . ------ for finite populations with N elements.
n N-1
These slides can not be reproduced or distributed without permission in writing from the author. 341
2.4.1.2. Distribution of sample means
In practice, we often consider
the populations as “infinite”,
also because the value of the corrective coefficient
(see previous slide) is very often close to 1
(and therefore of low importance)
(e.g.: with N = 1000 and n = 100, it is 0.949 ≈ 1).
We use the correction only when a sample
contains a significant percentage of the whole population.
For example, in the case of tickets, we saw that:
standard deviation s of single ements is equal to 2;
standard deviation sx of the distribution of the means of all the
samples with 3 tickets is equal to 0.9428;
the corrective coefficient with N = 7 and n = 3 is 0.8165;
sdivided by 3 and multiplied by the corrective coefficient is
equal to (2/1.7321) . 0.8165 = 0.9428, just as we calculated using
the means of all possible different samples.
These slides can not be reproduced or distributed without permission in writing from the author. 342
2.4.1.2. Distribution of sample means
D I S T R I B U T I O N S
of of sample means
single elements (samples of n elements)
Infinite
populations
sx = s / n
Standard
Finite s
deviation
populations s N-n
of N
sx =
elements n N-1
These slides can not be reproduced or distributed without permission in writing from the author. 343
2.4.1.2. Distribution of sample means
D I S T R I B U T I O N S
of of sample means
single elements (samples of n elements)
Infinite
populations
sx = s / n
Standard
Finite s
deviation
populations s N-n
of N
sx =
elements n N-1
These slides can not be reproduced or distributed without permission in writing from the author. 344
2.4.1.2. Distribution of sample means
An additional demonstration
of the expression
of the variance of a sample
to obtain an unbiased estimator
These slides can not be reproduced or distributed without permission in writing from the author. 345
2.4.1.2. Distribution of sample means
Let’s consider a whole population of N elements with mean m and variance s2. As already
seen, the mathematical expression of s2 is:
N
S (xi - m)2
i=1
s2 = ---------------
N
Adding and subtracting, at the numerator, any two values, that however we regard as the
mean, x, of a generic sample of n items taken from the population, you can write:
N
S (xi - m)2 1
N
1
N
--------------- = ----- S (xi - x + x - m) = ----- S [(xi - x) + (x - m)]2 =
i=1
s2 = . 2 .
N N i=1 N i=1
1 N 1 N
= ----- S [(xi-x) + (x-m)] = ----- .S [(xi-x)2 + (x-m)2 + 2 (xi-x) . (x-m)] =
. 2
N i=1 N i=1 constant
(independent of i)
1 N 1 N 2 N
= ----- . S (xi - x)2 + ----- .S (x - m)2 + ----- .S [(xi - x) . (x - m)] =
N i=1 N i=1 constant N i=1 zero (by definition of mean)
1 N N (independent of i) 2 N
= ----- S (xi - x) + ----- (x - m) + ----- (x - m) S (xi - x) =
. 2 . 2 . .
N i=1 N N i=1
These slides can not be reproduced or distributed without permission in writing from the author. 346
2.4.1.2. Distribution of sample means
Ultimately, taking into account that the last term LEGENDA
(the 3rd one) of the second member is 0, it is N = size of the whole population
easy to rewrite the last expression found as:
n = size of the sample
1 N
s2 = ----- . S (xi - x)2 + (x - m)2 k = number of all possible samples (of n elements)
extractable from the whole population (of N ele-
N i=1 ments)
Indicating by s2 the 1st term of the 2nd mem- m = (true) mean of the whole population
ber, we get: x = mean of a sample of n elements
s2 = s2 + (x - m)2 s2 = (true) variance of the whole population
s2 = variance of all elements of the population, but
hence, we get s2 as: calculated with reference to the mean of a sample
s2 = s2 - (x - m)2 of n elements
Let’s consider all possible samples of n elements extractable from the whole population (of N
elements) and denote by k their number. We can now calculate the means of the three
terms of the previous equation extended to all k samples of n elements:
k k k
S s2j
j=1
S s2
j=1
S (xj - m)2
j=1
--------- = --------- - ---------------
k k k
s2 True variance of the Variance
whole population of the k “sample means”
of the samples of n elements
(whose general mean is
always equal to m)
These slides can not be reproduced or distributed without permission in writing from the author. 347
About the last shown equation, the following remarks can be done:
1. the first member is the mean of the k values of s2 and may be indicated by s2;
2. as s2 is here a constant, the first term of the second member is given by s2 multiplied by k and
divided by k, that is s2;
3. the second term of the second member is, by definition, the variance of the sample means, sx2,
which, as has been seen(1), is equal to the variance s2 of single elements of the whole population,
divided by the size n of the sample.
At this point, the previous equation may be written in the following way:
s2 n-1
s2 = s 2 - s x2 = s2 - ----- = ---------- . s2
n n
The final result is:
n-1
s2 = ---------- . s2
n
The variance s2 is the mean of the k variances of all the N data of the whole population with
respect to each of the k sample means (of all possible samples of n elements each). If, in-
stead of this mean value, we would use, as is usually done, the variance of a unique sample
(of n elements, instead of N), we would introduce an additional source of deviation in the esti-
mate of the true variance, s2, but that's part of the “samplig randomness”.
The reason why variances s2 and s2 do not coincide resides in the fact that the variance s2
is calculated with respect to the sample means of all possible samples of size n, instead than
respect to the overall (or true) mean m.
(1) From the relationship between standard deviations sx = s / √ n it is immediate to obtain the analogous
relationship between variances: sx2 = s2 / n.
These slides can not be reproduced or distributed without permission in writing from the author. 348
The last highlighted expression implies that, taking successively, one after the other, many
samples of size n and calculating the variance, s2, for each of them and then calculating the
mean of all the variances gradually obtained, the value to which these means tend while the
number of samples (of size n) increases, is not the population’s true variance, s2true, but a
value a little smaller, s2biased, equal to:
n-1
s biased = --------- . s2true
2
n
s2 biased s2 true
bias
It is said that s2 is a biased estimator of s2. The coefficient (n -1) / n represents the bias of
the estimator (of s2) s2.
It is worth to remark that the bias coefficient tends to 1 (and therefore its importance decrea-
ses) with increasing the size n of the sample, while the bias coefficient may be essential with
low values of n (i. e. with small samples).
These slides can not be reproduced or distributed without permission in writing from the author. 349
2.4.1.2. Distribution of sample means
In order to obtain an unbiased (= corrected) estimator of s2 …
n-1
… let’s consider the above equation (penultimate and last slide): s2 = ---------- . s2
n
and solve it with respect to s2; we removed the upper dash of s2
because we assumed to calculate this variance for any sample n
of n elements (and no longer on the whole population of N ele- s
^2 = ---------- . s2
ments), thus the second member of the equation becomes an
estimate of the variance s2 (in itself biased, but corrected by n-1
multiplying it by the bias coefficient) indicated by the circumflex
s2
accent on s2.
2.4.1.3
Confidence interval
for the estimation of a mean
(and Student’s t distribution)
These slides can not be reproduced or distributed without permission in writing from the author. 351
2.4.1.3. Confidence interval for a mean (and t distribution)
These slides can not be reproduced or distributed without permission in writing from the author. 352
2.4.1. Estimates
2.4.1.4
Confidence intervals for a mean
if s is known
These slides can not be reproduced or distributed without permission in writing from the author. 353
2.4.1.4. Confidence intervals for a mean if s is known
General Purpose
To estimate the mean of a population,
trying to quantify the uncertainty of the estimate.
These slides can not be reproduced or distributed without permission in writing from the author. 354
2.4.1.4. Confidence intervals for a mean if s is known
E X A M P L E
We want to evaluate the mean of a population with
standard deviation known: s = 60
by a sample of n = 36 elements
(e.g. because we cannot afford a bigger one)
These slides can not be reproduced or distributed without permission in writing from the author. 355
2.4.1.4. Confidence intervals for a mean if s is known
These slides can not be reproduced or distributed without permission in writing from the author. 356
2.4.1.4. Confidence intervals for a mean if s is known
The red area under the curve is proportional to the probability of finding a
sample (of n = 36 elements) with its mean falling inside to the interval in
abscissa.
Now let us consider a "tail" (yellow in Figure) of 2.5%. The distance E
between the true mean mand the initial limit of the tail (according to the
normal distribution Table, see Par. 2.2.9.2), is given by:
E = 1.96 . s = 1.96 . 10 = 19.6
and the limit value is equal to m + E = 119.6 .
This means that only 2.5% of the
samples have their means greater than
n = 36
119.6 and therefore far from the true
E = 19.6 mean m over a value of E (19.6).
sx = 10
Taking into account that, on the left of
the graph, there is a further 2.5% of
samples with their means less than:
m - E = 100 - 19.6 = 80.4
2.5% and therefore far from the true mean m
over a value of E in absolute value, we
m = 100 can say that there is, in all, a 5% of
samples with their mean far from the
m + 19.6 = 119.6 true mean m over an absolute value
equal to E.
These slides can not be reproduced or distributed without permission in writing from the author. 357
2.4.1.4. Confidence intervals for a mean if s is known
These slides can not be reproduced or distributed without permission in writing from the author. 358
2.4.1.4. Confidence intervals for a mean if s is known
Samples that can be extracted are many and different
and it is not possible to know which of them happened ...
true mean
n = 36
E = 19,6
sx = 10
2,5% 2,5%
m = 100
m + 19,6 = 119,6
These slides can not be reproduced or distributed without permission in writing from the author. 360
2.4.1.4. Confidence intervals for a mean if s is known
In conclusion,
in the conditions of the previous example, we can say that,
although we based the construction
of the distribution of sample means on a random sample,
there is a risk less than 5%
that the 95% confidence interval, built on the sample,
does not contain the true mean.
These slides can not be reproduced or distributed without permission in writing from the author. 361
2.4.1.4. Confidence intervals for a mean if s is known
0,45
0,40
95% of data are within ± 1.960 s.
0,30
(e.g.) that only 5% of samples of n = 36
0,25 elements has a mean far from the true
0,20
0,15
mean m (unknowable) more than a distan-
0,10
ce E, called half confidence interval.
Known s = 60 and using the normal distri-
0,05
0,00
-4 -3 -2 -1 0 1 2 3 4
± 1 s 68.26% of data
± 2 s 95.45% of data
bution properties, we can evaluate this di-
± 3 s 99.73% of data stance E as:
s 60
E = |Li - x| = za/2 -------- = 1.96 --------- = 19.6
n 36
Accuracy of the esti-
mate or maximum ac-
ceptable uncertainty.
It is the 1st of the two Value obtained from the
initial choices. normal distribution Ta-
ble, depending on the va-
lue chosen for a/2, i.e. for
the risk of failing.
Significance or risk to 2.5% 2.5%
fail. It is the complement
to 1 of the confidence
level and is the 2nd initial E E
choice.
m – 19.6 m + 19.6
These slides can not be reproduced or distributed without permission in writing from the author. 363
2.4.1.4. Confidence intervals for a mean if s is known
Evaluation of the Improving of the estimate’s
variability of the accuracy by increasing the
investigated phe- sample size (= time & mo-
nomenon. ney invested).
Accuracy (uncertainty)
of the estimate.
s 60
E = |Li - x| = za/2 -------- = 1,96 --------- = 19,6
n 36
Significance or risk to
fail. It is the complement
to 1 of the confidence
level and can be chosen
regardless of the physica-
lity of the investigated 2.5% 2.5%
phenomenon, but only
according to the risk of
failing that we accept to E E
run. m – 19.6 m + 19.6
These slides can not be reproduced or distributed without permission in writing from the author. 364
2.4.1.4. Confidence intervals for a mean if s is known
On this basis, it is easy to establish a relationship (with the same sample
size n) between:
the confidence level (= credibility of the estimate)
and the half confidence interval, E (= accuracy of the estimate)
100%
90%
80%
level
60%
Confidence
50%
40%
30%
n = 18
20% n = 36
10%
n = 72
0%
0 5 10 15 20 25 30 35 40
2.4.1.5
Confidence intervals for a mean
if s is unknown
(Student’s t distribution)
These slides can not be reproduced or distributed without permission in writing from the author. 368
2.4.1.5. Confidence intervals for a mean if s is unknown (t distrib.)
0,20
0,15
0,10
0,05
0,00
-4,0 -3,0 -2,0 -1,0 0,0 1,0 2,0 3,0 4,0
Units of standard deviation from the mean
These slides can not be reproduced or distributed without permission in writing from the author. 370
Significance livel a
n d.o.f. 25% 20% 10% 5% 2,5% 1% 0,5% 0,10%
2 1 1,000 1,376 3,078 6,314 12,706 31,821 63,657 318,309
3 2 0,816 1,061 1,886 2,920 4,303 6,965 9,925 22,327
4 3 0,765 0,978 1,638 2,353 3,182 4,541 5,841 10,215
5 4 0,741 0,941 1,533 2,132 2,776 3,747 4,604 7,173
Table of
6 5 0,727 0,920 1,476 2,015 2,571 3,365 4,032 5,893
7 6 0,718 0,906 1,440 1,943 2,447 3,143 3,707 5,208
8 7 0,711 0,896 1,415 1,895 2,365 2,998 3,499 4,785
9 8 0,706 0,889 1,397 1,860 2,306 2,896 3,355 4,501
Student’s t
10 9 0,703 0,883 1,383 1,833 2,262 2,821 3,250 4,297
11 10 0,700 0,879 1,372 1,812 2,228 2,764 3,169 4,144
12 11 0,697 0,876 1,363 1,796 2,201 2,718 3,106 4,025
13 12 0,695 0,873 1,356 1,782 2,179 2,681 3,055 3,930
distribution
14 13 0,694 0,870 1,350 1,771 2,160 2,650 3,012 3,852
15 14 0,692 0,868 1,345 1,761 2,145 2,624 2,977 3,787
16 15 0,691 0,866 1,341 1,753 2,131 2,602 2,947 3,733
17 16 0,690 0,865 1,337 1,746 2,120 2,583 2,921 3,686
18 17 0,689 0,863 1,333 1,740 2,110 2,567 2,898 3,646
19 18 0,688 0,862 1,330 1,734 2,101 2,552 2,878 3,610
20 19 0,688 0,861 1,328 1,729 2,093 2,539 2,861 3,579
21 20 0,687 0,860 1,325 1,725 2,086 2,528 2,845 3,552
22 21 0,686 0,859 1,323 1,721 2,080 2,518 2,831 3,527
23 22 0,686 0,858 1,321 1,717 2,074 2,508 2,819 3,505 n = 30 (d.o.f. = 29)
Normal distribu-
24 23 0,685 0,858 1,319 1,714 2,069 2,500 2,807 3,485 tion: n =
n = 10 (d.o.f. = 9)
25 24 0,685 0,857 1,318 1,711 2,064 2,492 2,797 3,467
26 25 0,684 0,856 1,316 1,708 2,060 2,485 2,787 3,450
27 26 0,684 0,856 1,315 1,706 2,056 2,479 2,779 3,435
28 27 0,684 0,855 1,314 1,703 2,052 2,473 2,771 3,421
29 28 0,683 0,855 1,313 1,701 2,048 2,467 2,763 3,408
30 29 0,683 0,854 1,311 1,699 2,045 2,462 2,756 3,396
31 30 0,683 0,854 1,310 1,697 2,042 2,457 2,750 3,385
32 31 0,682 0,853 1,309 1,696 2,040 2,453 2,744 3,375
n = 5 (d.o.f. = 4)
33 32 0,682 0,853 1,309 1,694 2,037 2,449 2,738 3,365
34 33 0,682 0,853 1,308 1,692 2,035 2,445 2,733 3,356 1- F(t)
n = 2 (d.o.f. = 1)
35 34 0,682 0,852 1,307 1,691 2,032 2,441 2,728 3,348
36 35 0,682 0,852 1,306 1,690 2,030 2,438 2,724 3,340
37 36 0,681 0,852 1,306 1,688 2,028 2,434 2,719 3,333 -4,0 -3,0 -2,0 -1,0 0,0 1,0 2,0 t 3,0 4,0
38 37 0,681 0,851 1,305 1,687 2,026 2,431 2,715 3,326
39 38 0,681 0,851 1,304 1,686 2,024 2,429 2,712 3,319
40 39 0,681 0,851 1,304 1,685 2,023 2,426 2,708 3,313
41 40 0,681 0,851 1,303 1,684 2,021 2,423 2,704 3,307
42 41 0,681 0,850 1,303 1,683 2,020 2,421 2,701 3,301
43 42 0,680 0,850 1,302 1,682 2,018 2,418 2,698 3,296
44 43 0,680 0,850 1,302 1,681 2,017 2,416 2,695 3,291
45 44 0,680 0,850 1,301 1,680 2,015 2,414 2,692 3,286
46 45 0,680 0,850 1,301 1,679 2,014 2,412 2,690 3,281
47 46 0,680 0,850 1,300 1,679 2,013 2,410 2,687 3,277
48 47 0,680 0,849 1,300 1,678 2,012 2,408 2,685 3,273
49 48 0,680 0,849 1,299 1,677 2,011 2,407 2,682 3,269
50 49 0,680 0,849 1,299 1,677 2,010 2,405 2,680 3,265
These slides can not be reproduced or distributed without permission in writing from the author.
0,674 0,842 1,282 1,645 1,960 2,326 2,576 3,090 371
2.4.1.5. Confidence intervals for a mean if s is unknown (t distrib.)
Rischio del
SUPPLIER's
FORNITORE
Let’s
Supponiamo
assume che
that un lotto sia
a batch considerato
is considered:
[%]
risk
acceptable
batch is accepted[%]
absolutely
assolutamente inaccettabile
unacceptable if its se la per-
true sua
difettosità
centage of(vera) è ≥units
defective 6,5%iso≥più. 6.5%.
Probability thatdi
Probabilità
2.4.1.6
Example of evaluation
of the confidence interval
for a mean
These slides can not be reproduced or distributed without permission in writing from the author. 376
2.4.1.6. Example of evaluation of the confidence interval for a mean
A biologist wants to estimate the average length of trout one year aged,
who are reared in a basin.
These slides can not be reproduced or distributed without permission in writing from the author. 377
2.4.1.6. Example of evaluation of the confidence interval for a mean
Table of
standard normal distribution z
1-F(z)
z 0,00 0,01 0,02 0,03 0,04 0,05 0,06 0,07 0,08 0,09
0,0 0,500000 0,496011 0,492022 0,488033 0,484047 0,480061 0,476078 0,472097 0,468119 0,464144
0,1 0,460172 0,456205 0,452242 0,448283 0,444330 0,440382 0,436441 0,432505 0,428576 0,424655
0,2 0,420740 0,416834 0,412936 0,409046 0,405165 0,401294 0,397432 0,393580 0,389739 0,385908
0,3 0,382089 0,378281 0,374484 0,370700 0,366928 0,363169 0,359424 0,355691 0,351973 0,348268
0,4 0,344578 0,340903 0,337243 0,333598 0,329969 0,326355 0,322758 0,319178 0,315614 0,312067
0,5 0,308538 0,305026 0,301532 0,298056 0,294598 0,291160 0,287740 0,284339 0,280957 0,277595
0,6 0,274253 0,270931 0,267629 0,264347 0,261086 0,257846 0,254627 0,251429 0,248252 0,245097
0,7 0,241964 0,238852 0,235762 0,232695 0,229650 0,226627 0,223627 0,220650 0,217695 0,214764
0,8 0,211855 0,208970 0,206108 0,203269 0,200454 0,197662 0,194894 0,192150 0,189430 0,186733
0,9 0,184060 0,181411 0,178786 0,176186 0,173609 0,171056 0,168528 0,166023 0,163543 0,161087
1,0 0,158655 0,156248 0,153864 0,151505 0,149170 0,146859 0,144572 0,142310 0,140071 0,137857
1,1 0,135666 0,133500 0,131357 0,129238 0,127143 0,125072 0,123024 0,121001 0,119000 0,117023
1,2 0,115070 0,113140 0,111233 0,109349 0,107488 0,105650 0,103835 0,102042 0,100273 0,098525
1,3 0,096801 0,095098 0,093418 0,091759 0,090123 0,088508 0,086915 0,085344 0,083793 0,082264
1,4 0,080757 0,079270 0,077804 0,076359 0,074934 0,073529 0,072145 0,070781 0,069437 0,068112
1,5 0,066807 0,065522 0,064256 0,063008 0,061780 0,060571 0,059380 0,058208 0,057053 0,055917
1,6 0,054799 0,053699 0,052616 0,051551 0,050503 0,049471 0,048457 0,047460 0,046479 0,045514
1,7 0,044565 0,043633 0,042716 0,041815 0,040929 0,040059 0,039204 0,038364 0,037538 0,036727
1,8 0,035930 0,035148 0,034379 0,033625 0,032884 0,032157 0,031443 0,030742 0,030054 0,029379
1,9 0,028716 0,028067 0,027429 0,026803 0,026190 0,025588 0,024998 0,024419 0,023852 0,023295
2,0 0,022750 0,022216 0,021692 0,021178 0,020675 0,020182 0,019699 0,019226 0,018763 0,018309
2,1 0,017864 0,017429 0,017003 0,016586 0,016177 0,015778 0,015386 0,015003 0,014629 0,014262
2,2 0,013903 0,013553 0,013209 0,012874 0,012545 0,012224 0,011911 0,011604 0,011304 0,011011
2,3 0,010724 0,010444 0,010170 0,009903 0,009642 0,009387 0,009137 0,008894 0,008656 0,008424
2,4 0,008198 0,007976 0,007760 0,007549 0,007344 0,007143 0,006947 0,006756 0,006569 0,006387
2,5 0,006210 0,006037 0,005868 0,005703 0,005543 0,005386 0,005234 0,005085 0,004940 0,004799
2,6 0,004661 0,004527 0,004397 0,004269 0,004145 0,004025 0,003907 0,003793 0,003681 0,003573
2,7 0,003467 0,003364 0,003264 0,003167 0,003072 0,002980 0,002890 0,002803 0,002718 0,002635
2,8 0,002555 0,002477 0,002401 0,002327 0,002256 0,002186 0,002118 0,002052 0,001988 0,001926
2,9 0,001866 0,001807 0,001750 0,001695 0,001641 0,001589 0,001538 0,001489 0,001441 0,001395
3,0not be 1,350E-03
These slides can reproduced or distributed without permission in writing from the author. 379
2.4.1.6. Example of evaluation of the confidence interval for a mean
These slides can not be reproduced or distributed without permission in writing from the author. 380
2.4.1. Estimates
2.4.1.7
Confidence interval
of the estimation
for a standard deviation
( short mention )
These slides can not be reproduced or distributed without permission in writing from the author. 381
2.4.1.7. Confidence interval for the estimation of a std. deviation
(n - 1) . s2 s2
c2 = -------------- = ---------
s2 s2
called chi-square , follows the well-known distribution of the same
name (which is available in Excel). For it the quantity n - 1 is a
parameter, called degrees of freedom.
In this expression:
s is the sample standard deviation
s is the population standard deviation (estimable, but, in general,
unknowable).
These slides can not be reproduced or distributed without permission in writing from the author. 382
2.4.1.7. Confidence interval for the estimation of a std. deviation
Without going in deeper details, (in the same way we did for the
confidence interval of a mean) we can write:
(n - 1) . s2
c21-a/2 < -------------- < c2a/2
s2
and hence:
(n - 1) . s2 (n - 1) . s2
---------------- < s < ----------------
c2a/2 c21-a/2
These slides can not be reproduced or distributed without permission in writing from the author. 383
2.4.1.7. Confidence interval for the estimation of a std. deviation
So we can write:
s s
------------------- < s < -------------------
za/2 za/2
1 + --------- 1 - ---------
2n 2n
These slides can not be reproduced or distributed without permission in writing from the author. 384
2.4.1. Estimates
2.4.1.8
Confidence interval
of the estimation
for a proportion
These slides can not be reproduced or distributed without permission in writing from the author. 385
2.4.1.8. Confidence interval for the estimation of a proportion
These slides can not be reproduced or distributed without permission in writing from the author. 389
2.4.1.8. Confidence interval for the estimation of a proportion
Looking at the two previous diagrams, we can clearly see that the pairs of curves
for the same sample size, n, define confidence intervals almost symmetrical
with respect to the midline (drawn with dots and dashes) only when they are in the
right side of the diagram, i.e. for values of the proportion not too smaller and not
too greater than 0.5 (that is far enough away from 0 and 1, or 0% and 100%). On
the opposite, in the left side of the diagram, these intervals become dissymme-
trical, because the lower curves progressively concentrate as they approach
(coming from right) to the zero (of the proportion) at the origin of the x-axis; while
the upper curves tend to proportionately expand with respect to the midline
(drawn with dots and dashes).
The shape of these curves does not allow to calculate the confidence intervals
using the normal distribution (which is symmetric) in the left side of these
diagrams, while it is not impossible to do that in the right side. Looking at the
diagrams, we can see that, as we move from right to left, the first pairs of curves
which become dissymmetrical are those relating to smaller samples, while those
relating to larger samples remain longer (almost) symmetrical. We can conclude
that the limit value of the proportion, within which we can evaluate the confidence
intervals of the diagram by means of the normal distribution, can be pushed the
further to the left, the larger is the sample and, of course, the greater is the
approximation that we are willing to accept for the confidence interval !
These slides can not be reproduced or distributed without permission in writing from the author. 390
2.4.1.8. Confidence interval for the estimation of a proportion
The above diagrams are easy to use, but only solve the most common problems. If we
want to estimate a confidence interval with a confidence level different from 95% and from
99%, or if we need to enlarge the area of the graph in the bottom left, because the detec-
ted frequency is very close to zero (or to 1)(1), it is necessary to conduct calculations on
their own, and then it comes in handy what will be discussed in the following slides.
1. BINOMIAL DISTRIBUTION
Points of curves of the previous diagrams can be correctly calculated using the bino-
mial distribution, of which (as already seen) the basic expression is:
n
P= ( ) pk (1 - p)n-k
k
where P is the probability of obtaining k successes(1) in n experiments and p is the
(Bernoulli’s) probability of success in a single experiment. In the diagrams for propor-
tions, the two-sided (or bilateral) Confidence Interval is displayed on the pair of curves
(related to the sample size n) by their intersections with a vertical line that crosses the
x-axis at the value k/n, corresponding to the proportion detected on sample. Therefore,
each point of the upper curve represents the upper limit of the Confidence Interval
(with Confidence Level equal 95% in the 1st of the two previous diagrams and equal to
99% in the 2nd one).
These points express the probability of success, pu, corresponding to a probability
Pu(2). And similarly each point of a lower curve represents the lower limit (of the Confi-
dence Interval) which is given by the probability, pl, corresponding to a probability Pl(3).
(1) Of course, k will always be integer and never greater than n.
(2) Equal to 95+(100-95)/2 = 97,5% in the 1st previous diagram and equal to 99+(100-99)/2 = 99,5% in the 2nd previous diagram.
(3) Equal to 5/2 → 2,5% in the 1st previous diagram and equal to 1/2 = 0,5% in the 2nd previous diagram.
These slides can not be reproduced or distributed without permission in writing from the author. 392
2.4.1.8. Confidence interval for the estimation of a proportion
For example, suppose we find k = 3 defectives in a sample of n = 12 items. The de-
tected proportion is k/n = 3/12 = 0,25 i.e. 25%. With reference to a confidence level
of 95% (1st diagram), the previous expression of the bino-mial distribution becomes
the following(1):
1 - 0,95 12
Pa/2 = ------------ = ( ) p3 (1 - p)9 = 220 . p3 (1 - p)9
2 3
0,30
A limitation on the use of the binomial distribution is the need to calculate the bino-
mial coefficients which, in turn, require the “factorial” operation. Generally, the po-
cket calculators calculate the factorial, n!, until n=69, and not for larger values! How-
ever, especially in the last versions, the function of EXCEL binomial distribution
works even up to values of n much larger.
Anyway, with very large samples, the manual use of the binomial distribution may lead
to considerable complications.
These slides can not be reproduced or distributed without permission in writing from the author. 394
2.4.1.8. Confidence interval for the estimation of a proportion
2. POISSON DISTRIBUTION
We already mentioned that the Poisson distribution is the limit of the binomial
distribution when n tends to infinity and p tends to zero. So, if the sample is
very large (e.g. with more than 100 items), the binomial distribution can be
replaced by the Poisson distribution: especially where p is small, i.e. in the
left area of the diagrams where the dissymmetry mainly occurs. That's just
what we need, because if the sample is small, we have already seen that we
can use the binomial distribution without problems.
In conclusion, when the sample is very large, the application of the Poisson
distribution is unsuitable only in the case where p is not small enough.
For example, imagine finding k = 2 defective items in a sample with n = 1000
items. The detected proportion is k/n = 2/1000 = 0,002 i.e. 2‰.
With reference to the usual Confidence Level equal to 95%, we can calculate,
by successive approximations, the lower and upper confidence limits
shown in the Table of the following slide.
These slides can not be reproduced or distributed without permission in writing from the author. 395
2.4.1.8. Confidence interval for the estimation of a proportion
Lower
Limit
0,025 0,00025387 =POISSON(2;1000*pLower;FALSE) 0,025000 0,00025396 -0,035%
Upper
Limit
0,025 0,00684187 =POISSON(2;1000*pUpper;FALSE) 0,025000 0,00682650 0,225%
As you can see in the above Table, in this situation, the results obtained with
the Poisson distribution are substantially the same as those obtainable using
the binomial distribution.
These slides can not be reproduced or distributed without permission in writing from the author. 396
2.4.1.8. Confidence interval for the estimation of a proportion
3. NORMAL DISTRIBUTION
If the sample is large (say, n ≥ 100) and p is far enough away from 0 and 1,
and then we are on the right side of the diagram where the pairs of curves
are fairly symmetrical with respect to median line (drawn with dots and da-
shes), we can use even the normal distribution with a mean value of the pro-
portion equal to k/n and with a standard deviation of the proportion given by:
sp = √ p (1 - p)
For example, whith n = 200, k = 142, the detected proportion is:
pr = k/n = 142/200 = 0,71 i.e. 71%.
As usual, with reference to a Confidence Level of 95%, but this time bilateral
(= one-sided), we obtain za/2 = 1,96. We can now rewrite the same expres-
sion used in the Confidence Intervals for a mean, except change the stan-
dard deviation, which must consist of that of the proportion divided by √ n
(because it is "as if" we were working with sample means):
k sp k pr (1 - pr)
Limit = ----- ± za/2 ------- = ----- ± za/2 --------------
Upper n √n n n
Lower
with the example data, the previous expressions become:
0,71 (1 - 0,71) 0,773
Limit = 0,71 ± 1,96 ---------------------- = 0,71 ± 1,96 0,0321 = 0,71 ± 0,06289
.
Upper 200 0,647
Lower
These slides can not be reproduced or distributed without permission in writing from the author. 397
2.4.1.8. Confidence interval for the estimation of a proportion
We see that, in this situation, the approximation with the normal distribution is
more than acceptable since it shows (last column on the right) a deviation
less than ± 3% compared to the results obtainable with the binomial distribu-
tion, which is considered here as the correct reference. Clearly the only ad-
vantage of the use of the normal distribution is the ease of use.
The calculation of confidence intervals for proportions using the normal
distribution is particularly common in Attribute Control Charts.
These slides can not be reproduced or distributed without permission in writing from the author. 398
2.4.1.8. Confidence interval for the estimation of a proportion
#. COMPARISON AMONG THE RESULTS OF ALL SITUATIONS
The Table below shows the results obtained with the 3 distributions in all three examples
discussed above: the first two columns on the left characterize the context (generalized
situation), while the succeeding three (numerically) describe the 3 examples and the latest
right column proposes, situation to situation, comparative remarks among the 3 distributions.
The development of calculations is exactly the same we have already seen.
OBTAINED RESULTS THROUGH THE 3 DISTRIBUTIONS
TYPICAL SITUATIONS (in the same way as in the previous slides)
Generalized description Examples of application BINOMIAL POISSON NORMAL
Detected
D% D%
Limits
Sample size values
proportion n k pdet=k/n (assumed to be the values compared to values compared to
n pd correct reference) the binomial the binomial COMMENTS
In this situation, only the binomial di-
LLower 0,05793300 0,05522200 -4,680% 0,00500000 -91,369% stribution provides valid results, while the
other two distributions just capture the or-
whatever whatever 12 3 0,250 der of magnitude. Be noted that Poisson
LUpper 0,55695000 0,68458300 22,916% 0,49500000 -11,123% distribution is better than normal in the
assessment of the Lower Limit.
LLower 0,66555000 0,66547750 -0,011% 0,64711179 -2,770% This is one of the situations where the
far away normal distribution gives the best of
large 200 142 0,710
from 0 and 1 himself. In this case, however, the Poisson
LUpper 0,75191000 0,75646550 0,606% 0,77288821 2,790% distribution provides even better results.
Note that, in the first example with n = 12 and k = 3, the normal distribution does not provide
good results (D% equal to -91% and to -11%), due to the too small sample (n = 12 instead of
n ≥ 100), but it captures at least the order of magnitude, especially for the Upper Limit: this pro-
ves that its applicability is far from restricted (e.g. Attribute Control Charts).
These slides can not be reproduced or distributed without permission in writing from the author. 399
2.4.1.8. Confidence interval for the estimation of a proportion
These slides can not be reproduced or distributed without permission in writing from the author. 400
2.4.1.8. Confidence interval for the estimation of a proportion
40 6
Prescribedprescritte
Tolleranze tolerance
(histogram)
(istogramma)
5
proportion is zero,
CLASS
4
CLASSI
not even a defective 25
IN EACH
was detected in the
PEZZI NELLE
20 3
10 Previsione Previsione
defectives half defectives half
actually zero, but frazione frazione
NUMBER
2.4.1.9
Confidence interval
bilateral
and unilateral (one-sided)
These slides can not be reproduced or distributed without permission in writing from the author. 402
2.4.1.9. Confidence interval bilateral and unilateral
90% 95%
confidence confidence
interval interval
estimated estimated
value value
sample mean
sample mean
90% of times the confidence interval contains the true value (as shown in the above 95% of times the confidence interval contains the true value (as shown in the above
Figure), while 10% of times the estimated value will shift (5% right, and 5% left) Figure), while 5% of times the estimated value will shift left enough so that, under the
enough for the confidence interval does not contain the true value. true value, there is the shaded area (outside the confidence interval).
2.4.1.10
Determination of
sample size as a function
of confidence interval
These slides can not be reproduced or distributed without permission in writing from the author. 405
2.4.1.10. Determination of sample size as a function of C.I.
Mean
Starting from: LU = x + za/2 . s / n , we can write:
E = LU - x = za/2 . s / n ,
These slides can not be reproduced or distributed without permission in writing from the author. 406
2.4.1.10. Determination of sample size as a function of C.I.
These slides can not be reproduced or distributed without permission in writing from the author. 407
2.4.1.10. Determination of sample size as a function of C.I.
Sample size for confidence interval of a
M E A N (small sample t distribution)
Enter the input data in yellow cells (light gray on printed paper),
after reading the instructions about optional inputs.
Proportion
(the proof is similar to that of the mean, but in this case it is omitted for brevity)
za/2
n = p . (1 - p) . ( -------- )2
sp2 E
where, as for the mean, E is the maximum admitted error (for up or down,
i.e. the half confidence interval).
The product p . (1 - p) takes its maximum value equal to ¼ when p = 1/2,
which is the value currently assumed without best information. We highlight
that this condition is the worst case only if the absolute error E is kept con-
stant (regardless of p), but usually E decreases with p (or with 1 - p), be-
cause we try to maintain the percentage error approximately the same.
These slides can not be reproduced or distributed without permission in writing from the author. 409
2.4.1.10. Determination of sample size as a function of C.I.
Sample size for confidence interval of a
P R O P O R T I O N
(valid only if the assumptions for using normal distribution are verified)
Enter the input data in yellow cells (light gray on printed paper),
after reading the instructions about optional inputs.
2.4.2
Test of
a Statistical Hypothesis
These slides can not be reproduced or distributed without permission in writing from the author. 411
2.4.2. Test of a Statistical Hypothesis
These slides can not be reproduced or distributed without permission in writing from the author. 412
2.4.2. Test of a Statistical Hypothesis
These slides can not be reproduced or distributed without permission in writing from the author. 413
2.4.2. Test of a Statistical Hypothesis
2.4.2.1
Type I and Type II errors
These slides can not be reproduced or distributed without permission in writing from the author. 414
2.4.2.1. Type I and Type II errors
The problem is always addressed
first of all considering the credibility
of the null hypothesis H0,
namely that there is no real difference
between the two compared populations.
In the following, we will refer to the previous example of trout,
which,
after having been fed for a year with the traditional feedstuff,
have
a mean, ^ m0, estimated equal to 77.4 mm
and a standard deviation, ^ s0, estimated equal to 10.0 mm
(estimates are based on data from a sample of n = 50 trout).
The corresponding distribution
is normal, because we are working with sample means
and the null hypothesis H0
is that trout fed with the new feedstuff are of the same population .
These slides can not be reproduced or distributed without permission in writing from the author. 415
2.4.2.1. Type I and Type II errors
The most sensitive assumption is the definition of
an acceptance limit for H0.
It is often convenient to refer
to an alternative hypothesis, called H1,
which is the "reasonably“ alternative to the occurrence of
the null hypothesis H0 (i.e. the fact that it does not occur).
For example, if a trout length (at this stage of growth)
greater than or equal to 81.0 mm is considered sufficient
in order to opt for the new feedstuff,
we can consider,
in addition to the sample means distribution
with mean m0 = 77,4 mm and standard deviation s0 = 10,0 mm
corresponding to the null hypothesis H0 (on the left in the next Figure)
another distribution with mean m1 = 81.0 mm
and standard deviation a bit bigger, let us say equal to s1 = 12,0 mm
corresponding to the alternative hypothesis H1
(on the right in the next Figure):
it indicates that if the actual distribution is this (or even moved further right),
the null hypothesis H0 is surely false.
These slides can not be reproduced or distributed without permission in writing from the author. 416
2.4.2.1. Type I and Type II errors
Acceptance
0,30
accettazione
If the true mean vera is greater
limite di
limit
than this value, we accept H1
0,25
If
SeHH0 0isè true,
vera, as true (and H0 as false)
di probabilità
HH11 is false
è falsa
density
0,20
seen as "break-even"
Probability
0,15
Densità
0,10
0,05 b
a
0,00
70 75 80 85 90
Lunghezza trote[mm]
Trout length [mm]
If the null hypothesis H0 is true, there is a certain probability (equal to the si-
gnificance a) that the mean m1 of the 2nd sample (new feedstuff, but under
this hypothesis equivalent to the traditional feedstuff), is on the right of the
threshold. On the other hand, if the alternative hypothesis H1 is true (i.e. if
the true mean of trout fed with the new feedstuff is greater than or equal to
m1), there is still a certain probability (expressed by the value b) that the
mean of a sample taken from the second tank is on the left of the threshold.
These slides can not be reproduced or distributed without permission in writing from the author. 418
2.4.2.1. Type I and Type II errors
Risks are H0 H0
of two kinds true false
We accept the Right decision TYPE II error
null hypothesis
b
H0 (1 - a) probability of erroneously
confidence level accepting the null
hypothesis
These slides can not be reproduced or distributed without permission in writing from the author. 419
Acceptance
limit
di probabilità
Probability density
To reduce
both a and b risks togheter,
Densità
These slides can not be reproduced or distributed without permission in writing from the author. 420
2.4.2.1. Type I and Type II errors
It is important to keep in mind that (when possible)
we demonstrate
(with a risk of failing equal to the significance a)
that two populations are different;
while it is never possible to prove
that two populations are equal.
It is also worthwhile to note that, in most cases,
to consider two populations as significantly
(with risk of failing = a)
different implies an economic investment
(for example: the adoption of the new feedstuff)
in order to exploit the benefits of new knowledge.
Instead, not being able to detect
that the two populations are different (with risk of failing = b)
results in a loss of business
(but without immediate disbursements).
For this reason it is common to assume a < b.
These slides can not be reproduced or distributed without permission in writing from the author. 421
2.4.2. Test of a Statistical Hypothesis
2.4.2.2
Chi-square test for
the goodness of fit
These slides can not be reproduced or distributed without permission in writing from the author. 422
2.4.2.2. Chi-square test for the goodness of fit
These slides can not be reproduced or distributed without permission in writing from the author. 423
2.4.2.2. Chi-square test for the goodness of fit
Let us imagine to actually throw 3 coins 200 times and to obtain the results
in the column “OBSERVED frequencies" of the Table.
OBSERVED EXPECTED EXPECTED
Number of HEADs
frequencies probabilities frequencies
0 18 1/8 = 0.125 25
1 63 3/8 = 0.375 75
2 84 3/8 = 0.375 75
3 35 1/8 = 0.125 25
Total 200 8/8 = 1.000 200
90 OBSERVED frequencies 84
EXPECTED frequencies
80 75 75 Paired
70
63 columns
histogram
Frequencies
60
50
40 35
30 25 25
20
18
10
0 1 2 3
Number of HEADs
We note that there are discrepancies between observed and expected
frequencies.
These slides can not be reproduced or distributed without permission in writing from the author. 424
2.4.2.2. Chi-square test for the goodness of fit
In order to test if these discrepancies are so small that they can be attribu-
ted to the random variability (or not), let us indicate by “f” the observed fre-
quencies and by “e” the expected ones, and then determine the value of the
statistical variable c2 (chi-square) using the following expression:
( f - e )2
c2 = S -------------
e
calculating the ratio (f-e)2/e separately for each of the 4 distribution classes.
These slides can not be reproduced or distributed without permission in writing from the author. 425
2.4.2.2. Chi-square test for the goodness of fit
If the obtained value is too large (compared to the threshold obtained from
chi-square Table), we reject the null hypothesis that the binomial distribu-
tion with n = 3 and p = 1/2 adequately represents the examined experiment.
In fact, the rejection of the null hypothesis implies either that the coins are
not well balanced or that they were not thrown in randomly.
By substituting the values of expected and observed frequencies, we get:
(18 - 25)2 (63 - 75)2 (84 - 75)2 (35 - 25)2
c2 = --------------- + -------------- + -------------- + --------------- = 8.96
25 75 75 25
To decide if this value is large enough to reject the null hypothesis (at the
level of significance a), we only need to check whether it exceeds the c2
value related to the degrees of freedom of this case.
Generally, in this kind of test, the number of degrees of freedom is given by:
• the number of terms (f-e)2/e that have been added to obtain c2,
• subtracting the number of quantities that are used in calculating the ex-
pected frequencies.
These slides can not be reproduced or distributed without permission in writing from the author. 426
2.4.2.2. Chi-square test for the goodness of fit
These slides can not be reproduced or distributed without permission in writing from the author. 427
2.4.2.2. Chi-square test for the goodness of fit
These slides can not be reproduced or distributed without permission in writing from the author. 428
2.4.2.2. Chi-square test for the goodness of fit
Examples
• In the most common case, in which we want to know if the data are ar-
ranged according to a known distribution, we take as denominator the
expected frequencies e;
• On the opposite, if we want to know which of a set of known statistical
distributions, is best suited to interpret the experimental data, the most
reliable reference is given by experimental data and therefore we take
as denominator the observed frequencies f.
These slides can not be reproduced or distributed without permission in writing from the author. 429
2.4.2.2. Chi-square test for the goodness of fit
TESTING THE FIT OF A NORMAL (or other) DISTRIBUTION
Problem 1: experimental data are arranged according to a normal distribution ?
31,5
30 Measured values
30
Expected values
Number of rolls in each class
25
22
21
19,4 19,3
20
Paired
columns
15 histogram
10
4,5 4,5
5
3
2
1 0,4 1
0 0,0 0,4 0 0,0
0
4,80 4,85 4,90 4,95 5,00 5,05 5,10 5,15 5,20
Diameter [mm]
These slides can not be reproduced or distributed without permission in writing from the author. 430
2.4.2.2. Chi-square test for the goodness of fit
f e
C L A S S E S EXPECTED (f-e)2
OBSERVED frequencies ----------
Class
# from to frequencies (normal e
center distribution)
1 4,825 4,875 4,85 1 0,393 0,939
2 4,875 4,925 4,90 3 4,521 0,512
3 4,925 4,975 4,95 21 19,417 0,129
4 4,975 5,025 5,00 30 31,462 0,068
5 5,025 5,075 5,05 22 19,319 0,372
6 5,075 5,125 5,10 2 4,476 1,370
7 5,125 5,175 5,15 1 0,387 0,972
Total 80 79,975 4,361
Mean 4,99987 (f-e)2
Std. dev. 0,048567 c2 = S ---------
e
2.4.2.3
Summary of criteria
to check if available data
are distributed according to
the normal distribution
These slides can not be reproduced or distributed without permission in writing from the author. 433
2.4.2.2. Summary of criteria (to check if available data are distributed according to the normal distribution)
Because of their importance, all available criteria to test whether a series of experimental
data can be properly interpreted by a normal distribution are summarized below.
MATHEMATICAL
VISUAL APPLICATION
PRINCIPLE APPLICATION
(subjective)
(measured)