ENGI 4421
Descriptive Statistics
A Brief Definition of Statistics:
Page 1-01
[Navidi Chapter 1 & Devore Chapter 1]
Statistics is the science of making decisions in the absence of certainty. .
Some more definitions:
A population is a set of objects.
A sample is a subset of the population.
A simple random sample of size n is a
sample chosen by a method such that
each collection of n items in the population
is equally likely to be in the sample as every
other collection of n items
(as in a fair lottery).
Probability uses knowledge of the population to predict the contents of a sample when the
sample is drawn randomly from the population.
Inferential Statistics uses knowledge of a random sample to deduce some properties of the
population from which the sample was drawn. A good foundation in probability theory is
needed before one can study inferential statistics effectively.
The design of this course is:
Descriptive Statistics
Devore Ch. 1; Navidi Ch. 1
Lecture Notes Ch. 1
Probability
Inferential Statistics
Devore Ch. 2-6;
Navidi Ch. 2-4
Devore Ch. 7-9, 12, 14;
Navidi Ch. 5-7
Notes Ch. 2-11
Notes Ch. 11-15
There is also some coverage of decision trees, an introduction to probability through the concept
of a fair bet and Bayesian confidence intervals. These topics are not in the textbooks. The
Devore textbook does not cover the topic of propagation of error.
ENGI 4421
Random Samples
Descriptive Statistics
Page 1-02
(Navidi textbook: Chapter 1)
Example 1.01
A quality control inspector selects every one hundredth item from a production line for testing, in
order to deduce whether or not the production line is operating properly. With a daily production
run of 10,000 items, there are 100 items in the sample. Is the sample so drawn a simple random
sample?
Example 1.02
An engineer wishes to judge the breaking strength of a set of 10 kg beams that have been
delivered in a shipment of 500 beams stacked on a pallet in a large cube. The engineer selects
three beams from the top of the stack for testing.
Is the sample so drawn a simple random
sample?
Example 1.03
The output from one days run of a production line is labelled from 1 to 10,000. The inspector
has a list of 100 distinct numbers between 1 and 10,000, provided from a random number
generator and selects the 100 items that match the 100 numbers for the sample to be tested. Is the
sample so drawn a simple random sample?
In all of the cases above, the population is tangible it consists of a finite number of real objects.
A random sample of repeated measurements of the same object or process under identical
conditions is drawn from a conceptual population: the infinite set of all values that could have
been observed.
ENGI 4421
Descriptive Statistics
Page 1-03
Summarizing Data
Given a set of observations { x1, x2, ..., xn }, one can summarize the observations using various
types of graphical display (time series plot, bar chart, histogram, pie chart, pictogram, etc.) or by
numerical measures (stem and leaf diagram, frequency table) or single numbers measuring typical
value, spread of values, etc.
Example 1.04 Course text (Navidi, third edition), table 1.2 p. 21
[This data set is available from the course web site, at
[Link]/~ggeorge/4421/demos/t1/[Link]]
The data set below consists of observations on particulate matter emissions x (in g/gal) for 62
vehicles driven at high altitude.
7.59
23.38
7.18
3.84
9.52
5.91
1.83
6.28 6.07
9.24 3.22
6.95 18.64
2.37 3.81
2.68 10.14
5.60 5.61
7.56
5.23
2.06
7.10
5.32
9.20
1.50
5.54 3.46 2.44 3.01 13.63 13.02
4.04 17.11 12.26 19.91 8.50 7.81
6.04 5.66 8.86 4.40 3.57 4.35
5.84 2.89 4.68 1.85 9.14 8.67
7.31 2.09 6.32 6.53 6.32 2.01
6.46 5.29 5.64 2.07 1.11 3.32
By itself, this table is not very helpful as we try to grasp the overall picture of emissions.
One way to improve visibility is simply to rearrange these data into ascending order:
1.11 1.50
2.68 2.89
4.35 4.40
5.66 5.84
6.95 7.10
9.14 9.20
19.91 23.38
(a)
1.83
3.01
4.68
5.91
7.18
9.24
1.85 2.01 2.06 2.07 2.09 2.37 2.44
3.22 3.32 3.46 3.57 3.81 3.84 4.04
5.23 5.29 5.32 5.54 5.60 5.61 5.64
6.04 6.07 6.28 6.32 6.32 6.46 6.53
7.31 7.56 7.59 7.81 8.50 8.67 8.86
9.52 10.14 12.26 13.02 13.63 17.11 18.64
Construct a bar chart of the data, using class intervals of equal width, with the first
interval having lower limit 1.0 (inclusive) and upper limit 3.0 (exclusive).
ENGI 4421
Descriptive Statistics
Page 1-04
Example 1.04 (continued)
Let us first generate a frequency table for emissions manually.
Interval
for x
Frequency
f
1x< 3
3x< 5
5x< 7
7x< 9
9 x < 11
11 x < 13
13 x < 15
15 x < 17
17 x < 19
19 x < 21
21 x < 23
23 x < 25
12
11
18
9
5
1
2
0
2
1
0
1
_____
62
Total:
The bar chart then follows, with the height of each bar equal to the frequency of that class:
Bar Chart of Emissions
20
Frequency
15
10
11
13
15
Emissions
17
19
21
23
25
ENGI 4421
Descriptive Statistics
Page 1-05
Example 1.04 (continued)
However, the tail on the right has a rough appearance. It can be smoothed out by combining
some adjacent class intervals. However, that would cause a misleading appearance of greater
numbers in that tail of the bar chart:
Bar Chart of Emissions
20
Frequency
15
10
11
15
Emissions
25
There is a subtle difference between a bar chart and a histogram. A bar chart is used for
discrete (countable) data (such as number of defective items found in one run of a process) or
nonnumeric data (such as engineering major chosen by students). The bars are drawn with
arbitrary (often equal) width. No two bars should touch each other. The height of each bar is
proportional to the frequency.
A histogram is used for continuous data (such as shear stress or weight or time, where
between any two possible values another possible value can always be found). [A histogram can
also be used for discrete data.] Each bar covers a continuous interval of values and just touches
its neighbouring bars without overlapping. Every possible value lies in exactly one interval.
Unlike a bar chart, it is the area of each bar that is proportional to the frequency in that interval.
Only if all intervals are of equal width will the histogram have the same shape as the bar chart.
The relative frequency in an interval is the proportion of the total number of observations that
fall inside that interval. A relative frequency histogram can then be generated, with bar height
given by
ENGI 4421
Descriptive Statistics
Page 1-06
Example 1.04 (continued)
bar height =
The total area of all bars in a relative frequency histogram is always 1. In chapter 6 we will see
that the relative frequency histogram is related to the graph of a probability density function, the
total area under which is also 1.
Extending the frequency table for the sample of 62 observations of particulate matter emissions:
Interval
for x
Frequency Rel. Freq.
f
r
1x< 3
3x< 5
5x< 7
7x< 9
9 x < 11
11 x < 15
15 x < 25
12
11
18
9
5
3
4
_____
62
Total:
.19355
.17742
.29032
.14516
.08065
.04839
.06452
_______
1.00000
Density (= bar heights)
d
0.09677
0.08871
0.14516
0.07258
0.04032
0.01210
0.00645
Histogram of Emissions
0.16
0.14
0.12
Density
0.10
0.08
0.06
0.04
0.02
0.00
11
15
Emissions (g/gal)
25
ENGI 4421
Descriptive Statistics
Page 1-07
These diagrams will be generated during the first tutorial session using Minitab.
All Minitab sessions are available from the course web site at
" [Link]/~ggeorge/4421/demos/[Link] ".
Measures of Location
The mode is the most common value.
In example 1.04 the mode is
From the frequency tables, the modal class is
A disadvantage of the mode as a measure of location is
The sample median ~
x (or the population median ~ ) is the halfway value in an ordered set.
For n data, the median is the (n + 1)/2 th value if n is odd.
The median is the semi-sum of the two central values if n is even,
(that is median = [ (n/2 th value) + ((n/2 + 1)th value) / 2 ).
For the example above,
sample median
In the table of grouped values, the 31st and 32nd values fall in the same class.
The median class is therefore
The sample arithmetic mean x (or the population mean ) is the ratio of the sum of the
observations to the number of observations.
From individual observations,
and from a frequency table,
ENGI 4421
Descriptive Statistics
Page 1-08
For Example 1.04 above, from the 62 raw data (not from the frequency table),
The relative advantages of the mean and the median can be seen from a pair of smaller samples.
Example 1.05
Let A = { 1, 2, 3, 4, 5 }
and
B = { 1, 2, 3, 23654, 5 } .
Then
~
x =
for set A
and ~
x =
for set B , while
for set A and x
for set B.
Note that the mode is not well defined for either set.
A disadvantage of the mean as a measure of location is
Advantages of the mean over the median include
the median uses only the central value(s) while the mean uses all values.
For a symmetric population, the mean and the median ~ will be equal. If the mode is
unique, then it will also be equal to the mean and median of a symmetric population.
ENGI 4421
Descriptive Statistics
Page 1-09
The data set in Example 1.04 is far from being symmetric. It shows a strong positive skew.
Signs of positive skew include:
Signs of negative skew include:
While conceptual populations can be perfectly symmetric, random samples rarely display perfect
symmetry they may be only approximately symmetric, with mean median and tails of
approximately equal length.
ENGI 4421
Descriptive Statistics
Page 1-10
Measures of Variation
The simplest measure of variation is the range = (largest value smallest value).
A disadvantage of the sample range is
A disadvantage of the population range is
The effect of outliers can be eliminated by using the distance between the quartiles of the data as a
measure of spread instead of the full range.
The lower quartile qL is the { (n + 1) / 4 }th smallest value.
The upper quartile qU is the { 3(n + 1) / 4 }th smallest value.
[Close relatives of the quartiles are the fourths.
The lower fourth is the median of the lower half of the data, (including the median if and only if
the number n of data is odd).
The upper fourth is the median of the upper half of the data, (including the median if and only if
the number n of data is odd).
In practice there is often little or no difference between the value of a quartile and the value of the
corresponding fourth.]
The interquartile range is IQR = qU qL and
the semi-interquartile range is SIQR = (qU qL) / 2
Example 1.04 (again):
n = 62
(n + 1) / 4 = 15.75
and 3 (n + 1) / 4 = 47.25
qL = value 3/4 of the way from x15 to x16
qU = value 1/4 of the way from x47 to x48
The semi-interquartile range is then
ENGI 4421
Descriptive Statistics
Page 1-11
The boxplot illustrates the median, quartiles, outliers and skewness in a compact visual form.
The boxplot for example 1.04, as generated by version 14 of MINITAB, is shown below.
[and will be generated in the tutorial session.]
Unequal whisker lengths reveal skewness. The whiskers extend as far as the last observation
before the inner fence. The fences are not plotted by MINITAB.
The inner fences are 1.5 interquartile ranges beyond the nearer quartile, at
xL 1.5 IQR (lower)
and
xU + 1.5 IQR (upper) [3.411 and +14.819 here]
The outer fences are twice as far away from the nearer quartile, at
xL 3 IQR (lower)
and
xU + 3 IQR (upper) [9.942 and +21.655 here]
Any observations between inner & outer fences are mild outliers, which would be indicated by an
open circle (or, in MINITAB, by an asterisk). There are three mild outliers in this example.
Any observations beyond outer fences are extreme outliers, which would be indicated by a closed
circle (or, in MINITAB, by an asterisk or a zero). There is one extreme outlier in this example.
If you encounter an extreme outlier, then check if the measurement is incorrect or is from a
different population. If the observation is genuine from a population that is not strongly skewed,
then it is a rare event (< 0.01% in most populations, though not for Example 1.04!).
Measures of variability based on quartiles are not easy to manipulate using calculus methods.
ENGI 4421
Descriptive Statistics
Page 1-12
The deviation of the ith observation from the sample mean is ( xi x ) . At first sight, one might
consider that the sum of all these deviations could serve as a measure of variability.
However:
An alternative is the mean absolute deviation from the mean, defined as
1
n
MAD =
Unfortunately, the function f ( x) = xi x
i =1
xi x
is not differentiable at the one point where the
derivative is most needed, at x = x . Instead, the mean square deviation from the mean is used:
The population variance 2 for a finite population of N values is given by
and the sample variance s
1
=
N
(x )
i =1
of a sample of n values is given by
1
s =
n 1
2
(x x)
i =1
The square root of a variance is called the standard deviation and is positive (unless all values are
exactly the same, in which case the standard deviation is zero). The reason for the different divisor
2
(n 1) in the expression for the sample variance s will be explained later.
The MINITAB output for various summary statistics for example 1.04 is shown here:
Variable
Emissions
N
62
Mean
6.596
StDev
4.519
Variance
20.421
Minimum
1.110
Q1
3.425
Median
5.750
Q3
7.983
Maximum
23.380
ENGI 4421
Descriptive Statistics
Page 1-13
When calculating a sample variance by hand or on some simpler hand-held calculators, one of the
following shortcut formul may be easier to use:
n
s2 =
i =1
xi 2
1
xi
n i = 1
or
n 1
s2 =
xi
nx 2
i =1
or
n 1
2
n
n xi
x
i =1 i
i
=
1
.
s2 =
n ( n 1)
n
When all values of x are integers, the last of these three formul allows the sample variance to
be expressed exactly as a fraction. The formul for data taken from a frequency table with m
classes are similar:
m
1
s =
n 1
2
f (x x)
i =1
or
s =
2
or
s =
2
i =1
fi xi
1
fi xi
fi xi
n i = 1
n 1
2
nx 2
i =1
or
n 1
i =1
s =
2
fi xi
fi xi
i =1
n ( n 1)
2
where, in each case, n =
fi
i =1
and
x =
fi xi
i =1
fi
i =1
However, all of the shortcut formul are more sensitive to round-off errors than the definition is.
Example 1.06:
Find the sample variance for the set { 100.01, 100.02, 100.03 } by the definition and by one of the
shortcut formul, in each case rounding every number that you encounter during your
2
computations to six or seven significant figures, (so that 100.01 = 10002.00 to 7 s.f.). The
2
correct value for s in this case is .0001, but rounding errors will cause all three shortcut formul
to return an incorrect value of zero. (Try it!).
ENGI 4421
Descriptive Statistics
Page 1-14
Example 1.07:
Find the sample mean and the sample standard deviation for
x = the number of service calls during a warranty period, from the frequency table below.
xi
fi
65
30
Sum:
x =
fi xi
f
fi xi
100
n fi xi 2 ( fi xi )
s =
n ( n 1)
2
or
s2 =
1
n 1
fi ( xi x )
fi xi2
ENGI 4421
Descriptive Statistics
Page 1-15
For any data set:
3/4 of all data lie within two standard deviations of the mean.
8/9 of all data lie within three standard deviations of the mean.
(1 1/k2 ) of all data lie within k standard deviations of the mean (Chebyshevs inequality).
For a bell-shaped distribution (for which population mean = population median = population
mode and the probability distribution is exactly or nearly Normal [Gaussian]):
~ 68% of all data lie within one standard deviation of the mean.
~ 95% of all data lie within two standard deviations of the mean.
> 99% of all data lie within three standard deviations of the mean.
ENGI 4421
Descriptive Statistics
Page 1-16
Misleading Statistics - Example 1.08
Both graphs below are based on the same information, yet they seem to lead to different
conclusions.
Our profits rose enormously in the
last quarter.
vs.
Our profits rose by only 10%
in the last quarter.
Example 1.09 (pictograms)
Number of cameras sold in 2013
Visual displays can be very misleading.
data include,
for graphs:
Number of cameras sold in 2014
Questions to ask when viewing visual summaries of
for bar charts / pictograms :
[End of the chapter Descriptive Statistics]