58 Simple Statistical Methods for Software Engineering
Seeing Process Drift
If the median is not on a process target value, we can say that the process has drifted.
e amount of drift can be easily seen if we draw a target line on the box plot.
Figure 4.3 shows a box plot of bug repair time. e corporate goal is to x bugs within
a maximum of 16 days. e goal is marked on the box plot for easy interpretation.
Detecting Skew
e box plot is an eloquent way of expressing problems in process. One can see
clearly if the process results are skewed. If the median is in the middle of the box,
data are not skewed. If the median shifts to the right, data are left skewed. If the
median shifts to the left, data are right skewed. Another sign of skew is the length
of whisker. If the right whisker is longer, as seen in Figure 4.3, the process is skewed
to the right.
Seeing Variation
e width of the box is a measure of process variation. Box width shows variation
with 50% confidence level. e whisker-to-whisker width also expresses variation,
perhaps with better clarity and more dramatic effect. e whisker-to-whisker range
is an expression of variation with confidence levels more than 90%. In Figure 4.3,
the whisker-to-whisker range is from 6 to 45 days. e variation is far in excess of
what is anticipated.
Goal
Bugs must be repaired
within 16 days
Corporate goal
0 20 40 60 80
Figure 4.3 Box plot of bug repair time, days.
Tukey’s Box Plot 59
Risk Measurement
If we plot specification lines on the box plot, we can easily see the risk element
in the process. If the entire whiskers stay outside the specification lines, the
process is very risky. Risk is proportional to the portion of the box plot that
stays outside process specifications. Bug repair, shown in Figure 4.3, surely has a
schedule risk. We cannot quantify risk using a box plot, but we can qualitatively
say the risk is very high. Risk management is one area where qualitative judg-
ment is good enough and often more dependable than sophisticated quantita-
tive analysis.
Outlier Detection
A very useful result from the box plot is the detection of outliers. e rules
applied in the box plot do not assume any mathematical distribution function
for the process. e box plot way of detecting outliers diers from the control
chart way of detecting outliers. In control charts, we use the probability density
function that corresponds to the inherent distribution of data. e box plot
rules do not apply any distribution formula. e box plot uses a distribution-
free judgment that is more universal and robust enough to engage all kinds of
data.
Comparison of Processes
Box plots are used to compare process results. All the three elements of processes
can be visually compared:
Central tendency
Dispersion
Outliers
is visual comparison performs the functions of three tests: t test for process
mean, F test for process variation, and control chart tests for process outliers. is
comparison is discussed later in this chapter.
Improvement Planning
Process improvement planning is well supported by box plot analysis. A box plot
defines the problem with a picturesque essay of three dimensions of the process:
central tendency, dispersion, and outliers. A box plot is an empirical problem state-
ment. If we think that a well-dened problem is a problem half solved, then we
stand to gain immensely by the box plot way of problem definition.
60 Simple Statistical Methods for Software Engineering
An approximate answer to the right problem is worth a good
deal more than an exact answer to an approximate problem.
John W. Tukey
Box plots help us to identify and define the right problems.
e productivity box plot shown in Figure 4.1 highlights three opportunities
for innovation:
1. Removal of outliers: is is the easiest innovation. ere is no outlier in the
lower side of the plot. at is good news. e outliers with higher values
might appear as welcome outcomes. Here is the good old question of specify-
ing an upper limit even for the better side of events. It may be suspected that
extreme value of productivity is the result of a compromise, a slow acting
fuse, that might show up later somewhere as an issue. Although we need to
understand all the outliers, the outliers beyond the outer fence may be stud-
ied in detail. e presence of more outliers on the right side also indicates the
possible existence of a tail or skew in the distribution.
2. Shifting the median toward higher levels: is means the expected value of
productivity can be improved.
3. Reduction of IQR as well as whisker-to-whisker width: Process variation, depicted
both by the IQR and whisker-to-whisker width, can be reduced to minimize
variation.
e three innovations could coexist. Improving the median may be accompanied
by reduction in outliers, and vice versa. It is a good strategy to take up one at a time, in
the previously mentioned order, and take the beneficial side effects in the other two.
Core Benefits of Box Plot
Tukey’s box plot contains sufficient statistical strategies and yet retains its intended
simplicity. Many attempts are being made to enhance the information content in box
plots and make them colorful as well. We focus on the simple box plot in this chapter
and find that it has great potential.e box plot can be applied to the following:
Provide a visual summary of data
See process variation
Detect outliers
Detect skew in data
Tukey’s Box Plot 61
See process drift
Compare processes
Plan process improvement
Twin Box Plot
Let us take the case of reestimating software development effort. Teams are reluc-
tant to do a second estimate and are in a hurry to move forward with development.
However, it is a best practice to do a second estimate after a fortnight into the
project when many project details become visible. We get to know the require-
ments better, teams communicate better, risks are seen with clarity, and we are
enlightened by the early lessons. e second estimate is expected to be more accu-
rate. We wish to compare the second estimates with rst estimates and study the
improvement.
e box plot can be eminently used to compare the two results. In Figure 4.4, a
twin box plot is shown comparing two sets of effort variance data.
e twin box plot offers what might be called a visual test, a preliminary analy-
sis before we start rigorous tests. Visual judgment of the following can provide vital
clues regarding the differences between two results:
1. Is there a difference between the whisker-to-whisker widths?
2. Is there a difference between the box widths?
3. Is there a relative shift in the position of the median?
If the answer is yes to any one of these questions, we need to take a deeper
look at the box plots. Sometimes the presence or absence of outliers could make
a difference. Sometimes the skew of the median line inside a box could provide
a clue.
If the difference is significant, the boxes in the two plots may be completely
disjointed. ey may not overlap. Using the box plot representation, it is rather easy
to see if the new result is different from the old.
EVA 1
EVA 2
EVA
–80 –60 –40 –20 0 20 40 60 80
Figure 4.4 Comparison of two estimates using box plots.
62 Simple Statistical Methods for Software Engineering
If results due to innovation show improvement, one or more of the following
visual clues may be present:
e overall length of the box plot would have decreased
Outliers might have disappeared
e central line might show a favorable shift
e box might have shrunk
e box might be relocated in a favorable region
e unfavorable whisker might have diminished
If an improvement is not visible in a box plot, it may not be an improvement in
the first place. e question of looking for significance does not arise.
However, in most cases, people take pains to go through lengthy procedures
to execute signicant tests to check dierences, without box plot visual
checks. In some cases even after box plot rejections, people go through the
ritual of signicance tests.
Holistic Test
e twin box plot test is a holistic approach; it can compare two populations (two
groups) in a complete balanced fashion that no other test can offer. e price we
pay for completeness is loss of rigor. It so happens that rigorous tests have narrower
scope than robust tests; approximate analysis can sweep more terrain than precise
analysis. We need such a holistic test before we go into more sophisticated tests.
e twin box plot shown in Figure 4.4 offers a holistic comparison described in
the following paragraphs.
First, it compares the median values. e median of the rst estimate is 4.67%,
and the median of the second estimate is 1.27%. Comparing medians is more
robust than comparing means, which makes sense even with nonnormal data. is
is a comment on central tendency.
en dispersion is compared at two levels; the rst IQR is 23.45 and the sec-
ond and improved value is 8.54. It is evident that the core of the estimation process
covering 50% of results shows less dispersionan order of magnitude less. e
new dispersion is one-third the old. e old whisker-to-whisker range is 86.48,
whereas the new whisker-to-whisker range is 20.32, four times less. It is evident
that the dispersion is reduced in the new estimation technique; it is more reliable.
e box plot provides an order of magnitude test before we resort to p values for
judgment.
e box plot identifies outliers in the second group; not every estimate has been
well performed. e best practice must spread. e second process has philosophi-
cal problems called statistical outliers. However, in a practical sense, even the outli-
ers are better than the first process.
..................Content has been hidden....................

You can't read the all page of ebook, please click here login for view all page.
Reset
3.22.241.228