5. Normal Distributions: The Bell Curve
I couldn’t claim that I was smarter than sixty-five other guys—but the
average of sixty-five other guys, certainly.
—Richard Feynman
Distributions constitute part of the core knowledge base for any modeler.
Later, we use distributions to construct and analyze models of path
dependence, random walks, Markov processes, search, and learning. We
also require a working knowledge of distributions to measure inequality in
power, income, and wealth and to perform statistical tests. Our treatment of
distributions unfolds over two short chapters—one each for normal and
power law (long-tailed) distributions—in which we take the perspective of
modelers rather than statisticians. As modelers, we are interested in two big
questions: Why do we see the distributions we do, and why do distributions
matter?
To address the first big question, we need to reacquaint ourselves with what
distributions are. A distribution mathematically captures variation
(differences within a type) and diversity (difference across types) by
representing them as probability distributions defined over numerical
values or categories. A normal distribution takes the familiar bell curve
shape. Heights and weights of most species satisfy normal distributions.
They are symmetric around their means and do not include particularly
large or small events. We do not encounter many six-foot-long ants or fourpound elk. We can rely on the central limit theorem to explain the
prevalence of normal distributions. It tells us that when we add up or
average random variables, we can expect to obtain a normal distribution.
Many empirical phenomena, in particular any aggregate like sales data or
vote totals, can be written as sums of random events.
Not all event sizes are normal. Earthquakes, war deaths, and book sales
exhibit long-tailed distributions: they consist mostly of tiny events but
include the occasional whopper. Californians experience over 10,000
earthquakes each year. Unless you are staring at the quivering petals of a

jasmine blossom, you would not notice them. Occasionally, though, the
earth opens up, highways collapse, and cities tremble.
Knowing whether a system produces a normal or long-tailed distribution
matters for any number of reasons. We want to know whether a power grid
will suffer massive outages, or whether a market system will produce a
handful of billionaires and billions of poor people. With knowledge of
distributions, we can predict the likelihood of floodwaters that exceed a
levee’s walls, the probability that Delta flight 238 arrives in Salt Lake City
on time, and the odds that a transportation hub costs double its budgeted
amount. Knowledge of distributions is also relevant in design. Normal
distributions imply no large deviations, so airplane designers need not
create leg space for the eighteen-foot human. An understanding of
distributions can also guide actions. As we learn later, preventing riots
depends less on reducing average levels of discontent than on appeasing
people at the extreme.
In this chapter, we adopt a structure-logic-function organization. We define
normal distributions, describe how they arise, and then ask why they
matter. We apply our knowledge of distributions to explain why good
things come in small samples, to test for significance of effects, and to
explain Six Sigma process management. We then go back to the logic
question and ask what happens if we multiply rather than add random
variables. We learn that we obtain a lognormal distribution. Lognormal
distributions include larger events and are not symmetric about their means.
It follows that multiples of effects lead to more inequality, an insight that
has implications for how policies for increasing salaries affect income
distributions.

The Normal Distribution: Structure
A distribution assigns probabilities to events or values. The distribution of
daily rainfall, test scores, or human height assigns a probability to every
possible value of the outcomes. Statistical measures condense the
information contained in a distribution into single numbers, such as the
mean, the average value of the distribution. The mean height of a tree in
Germany’s Black Forest might be eighty feet, and the mean time spent in

the hospital following open-heart surgery might be five days. Social
scientists rely on means to compare economic and social conditions across
countries. In 2017, the United States per capita GDP of $57,000 exceeded
that of France, which equaled $42,000, while mean life expectancy in
France exceeds that of the United States by three years.
A second statistic, variance, measures a distribution’s dispersion: the
average of the squared distance of the data to the mean.1 If every point in a
distribution has the same value, the variance equals zero. If half of the data
have value 4 and half have value 10, then, on average, each point lies
distance 3 from the mean, and the variance equals 9. The standard
deviation of a distribution, another common statistic, equals the square root
of the variance.
The set of possible distributions is limitless. We could draw any line on a
piece of graph paper and interpret it as a probability distribution.
Fortunately, the distributions we encounter tend to belong to a few classes.
The most common distribution, the normal distribution, or bell curve, is
shown in figure 5.1.
image
Figure 5.1: Normal Distribution with Standard Deviations
Normal distributions are symmetric about their mean. If the mean equals
zero, the probability of a draw larger than 3 equals the probability of a draw
less than -3. A normal distribution is characterized by its mean and
standard deviation (or, equivalently, its variance). In other words, graphs of
normal distribution all look identical, with approximately 68% of all
outcomes within one standard deviation of the mean, 95% of all outcomes
within two standard deviations, and more than 99% lying within three
standard deviations. Normal distributions allow for any size outcome or
event, though large events are rare. An event five standard deviations from
the mean occurs about once in every 2 million draws.
We can exploit the regularity of normal distributions to assign probabilities
to ranges of outcomes. If houses in Milwaukee, Wisconsin, have a mean
square footage equal to 2,000 with a standard deviation of 500 square feet,

then 68% of houses have between 1,500 and 2,500 square feet and 95%
have between 1,000 and 3,000 square feet. If the 2019 fleet of Ford Focuses
can travel, on average, 40 miles per gallon with a standard deviation of 1
mile per gallon, then more than 99% of Focuses will get between 37 and 43
miles per gallon. As much as a consumer might hope, her new Focus will
not run 80 miles on a gallon of gasoline.

The Central Limit Theorem: Logic
No end of phenomena exhibit normal distributions: physical sizes of flora
and fauna, student test scores on exams, daily sales at convenience stores,
and the life spans of sea urchins. The central limit theorem, which states
that adding or averaging random variables produces a normal distribution,
explains why (see box).

Central Limit Theorem
The sum of N ≥ 20 random variables will be approximately a normal
distribution provided that the random variables are independent, that each
has finite variance, and that no small set of the variables contributes most
of the variation.2
One remarkable aspect of this theorem is that the random variables
themselves need not be normally distributed. They could have any
distribution so long as each has finite variance and no small subset of them
contributes most of the variance. Suppose that data on the purchasing
behaviors of the people in a small town of population 500 shows that each
person spends on average $100 a week. Some of those people might spend
$50 one week and $150 the next. Others might spend $300 every third
week, while others might spend random amounts between $20 and $180
each week. So long as each person’s spending has finite variation and no
small subset of people contribute most of the variation, the sum of the
distributions will be normally distributed with a mean of $50,000.
Aggregate weekly spending will also be symmetric: as likely to be above
$55,000 as it is below $45,000. By the same logic, the number of bananas,

quarts of milk, or boxes of taco shells that people buy will also be normally
distributed.
We can also apply the central limit theorem to explain the distribution of
human heights. A person’s height is determined by a combination of
genetics, the environment, and interactions between the two. The genetic
contribution could be as high as 80%, so we will assume that height
depends only on genes. At least 180 genes contribute to human height.3
One gene may contribute to having a longer neck or head and another to a
longer tibia. Though genes interact, to a first approximation, we can assume
that each contributes independently. If height equals the sum of the
contributions of the 180 genes, then heights will be normally distributed.
By the same logic, so too will the weights of wolves and the length of
pandas’ thumbs.

Applying Our Knowledge of Distributions:
Function
Our first application of the normal distribution reveals why exceptional
outcomes occur far more often in small populations, why the best schools
are small, and why the counties with the highest cancer rates have small
populations. Recall that in a normal distribution 95% of outcomes lie
within two standard deviations and 99% lie within three standard
deviations, and that by the central limit theorem, the mean of a collection of
independent random variables will be normally distributed (with the
caveats about variance). It follows that we can be pretty confident that
population averages on test scores and the like will be normally distributed.
The standard deviation of the average of the random variables, however,
does not equal the average of the variables’ standard deviations, nor does
the standard deviation of the sum equal the sum of the standard deviations.
Instead, those formulae depend on the square roots of the population sizes
(see box).

The Square Root Rules

The standard deviations of the mean σμ and of the sum σΣ of N independent
random variables each with standard deviation σ are given by the following
formulae:4
image
The formula for the standard deviation of the mean implies that large
populations have much lower standard deviations than small ones. From
this, we can infer that we should see more good things and more bad things
in small populations. And in fact we do. The safest places to live are small
towns, as are the least safe. The counties with the highest rates of obesity
and cancer have small populations. These facts can all be explained by
differences in standard deviations.
Failure to take sample size into account and inferring causality from
outliers can lead to incorrect policy actions. For this reason, Howard
Wainer refers to the formula for the standard deviation of the mean the
“most dangerous equation in the world.” For example, in the 1990s the
Gates Foundation and other nonprofits advocated breaking up schools into
smaller schools based on evidence that the best schools were small.5 To see
the flawed reasoning, imagine that schools come in two sizes—small
schools with 100 students and large schools with 1,600 students—and that
student scores at both types of schools are drawn from the same distribution
with a mean score of 100 and a standard deviation of 80. At small schools,
the standard deviation of the mean equals 8 (the standard deviation of the
student scores, 80, divided by 10, the square root of the number of
students). At large schools, the standard deviation of the mean equals 2.
If we assign the label “high-performing” to schools with means above 110
and the label “exceptional” to schools with means above 120, then only
small schools will meet either threshold. For the small schools, an average
score of 110 is 1.25 standard deviations above the mean; such events occur
about 10% of the time. A mean score of 120 is 2.5 standard deviations
above the mean; an event of that size should occur about once in 150
schools. When we do these same calculations for large schools, we find that
the “high-performing” threshold lies five standard deviations above the
mean and the “exceptional” threshold lies ten standard deviations above the

mean. Such events would, in practice, never occur. Thus, the fact that the
very best schools are small is not evidence that smaller schools perform
better. The very best schools will be small even if size has no effect solely
because of the square root rules.

Testing Significance
We also use the regularity of the normal distribution to test for significant
differences in mean values. If an empirical mean lies more than two
standard deviations from a hypothesized mean, social scientists reject the
hypothesis that the means are the same.6 Suppose we advance a hypothesis
that commute times in Baltimore equal those in Los Angeles. Suppose that
our data show that commute times in Baltimore averaged 33 minutes,
compared to 34 minutes in Los Angeles. If both data sets have standard
deviations of the mean equal to 1 minute, then we could not reject the
hypothesis that the commute times are the same. The means differ, but only
by a single standard deviation. If instead commute times in Los Angeles
averaged 37 minutes, then we would reject the hypothesis because the
means differ by four standard deviations.
Physicists, though, might not reject the hypothesis, at least not if the data
came from a physics experiment. Physicists impose stricter standards
because they have larger data sets—there are a lot more atoms than people,
and cleaner data. The evidence physicists relied on for the existence of the
Higgs boson in 2012 would occur randomly less than once in 7 million
trials were the Higgs boson not to exist.
The drug approval process used by the United States Food and Drug
Administration (FDA) also uses tests of significance. If a pharmaceutical
company claims that a new drug reduces the severity of eczema, that
company must run two randomized controlled trials. To construct a
randomized controlled trial the company would create two identical
populations of eczema sufferers. One of the populations receives the drug.
The other population receives a placebo. At the end of the trial, the average
severity as well as average rates of negative side effects are compared. The
company then runs statistical tests. If the drug significantly reduces eczema
(measured in standard deviations) and does not significantly increase side

effects, the drug can be approved. The FDA does not use a hard-and-fast
two-standard-deviation rule. The statistical bar will be lower for a drug that
cures a fatal disease and exhibits only minor side effects than for a drug
that cures toenail fungus but has a higher-than-expected incidence of bone
cancer associated with its usage. The FDA also cares about the power of
the statistical test—the probability that the test shows that the drug works.

Six Sigma Method
As our final application, we show how normal distributions inform quality
control through the Six Sigma method. Developed in the mid-1980s by
Motorola, the Six Sigma method reduces errors. The method models
product attributes as drawn from a normal distribution. Imagine a company
that produces bolts for door handles that must fit snugly into knobs made
by another manufacturer. Specifications call for the bolts to be 14
millimeters in diameter, though any bolt between 13 and 15 millimeters in
diameter will function properly. If the diameters of the bolts are normally
distributed with a mean of 14 millimeters and a standard deviation of 0.5
millimeter, then any bolt that differs by more than two standard deviation
fails. Two-standard-deviation events occur 5% of the time—far too high a
rate for manufacturers.
The Six Sigma method involves working to reduce the size of a standard
deviation to lower the probability of a failure. Companies can reduce error
rates by tightening quality control. On February 26, 2008, Starbucks closed
down over seven thousand shops for over three hours to retrain employees.
Similarly, checklists used by airlines and now hospitals reduce variation.7
Six Sigma reduces the standard deviation so that even a six-standarddeviation error avoids a malfunction. In our bolt example, that would
require reducing the standard deviation of a bolt’s diameter to one-sixth of
a millimeter. Six standard deviations implies an error rate of 2 per billion
cases. The actual threshold used assumes an unavoidable rate of one and a
half standard deviations. Thus, a six-sigma event actually corresponds to a
four-and-a-half sigma event, and an allowable error rate of about 1 per 3
million.

The application of the central limit theorem (and therefore an implicit
model of additive error) in the Six Sigma method is so subtle as to almost
go unnoticed. The bolt manufacturer likely does not perform a precise
measurement of the diameter of every bolt. It may sample a few hundred.
From that sample, it estimates a mean and a standard deviation. Then, by
assuming that variations in diameter result from the sum of random effects
such as machine vibrations, variation in the quality of metals, and
fluctuations in the temperature and speed of a press, they can invoke the
central limit theorem and infer a normal distribution of diameters. The
manufacturer then has a benchmark standard deviation that it can seek to
reduce.

Lognormal Distributions: Multiplying Shocks
The central limit theorem requires that we add or average independent
random variables in order to get a normal distribution. If the random
variables are not added but interact in some way, or if they fail to be
independent, then the resulting distribution need not be normal. In fact,
generally it will not be. For example, random variables that are the product
of independent random variables produce lognormal rather than normal
distributions.8 Lognormal distributions lack symmetry because products of
numbers larger than 1 grow faster than sums (4 + 4 + 4 + 4 = 16, but 4 × 4
× 4 × 4 = 256) and multiples of numbers less than 1 decrease faster than
sums ( image, but image). If we multiply sets of twenty random
variables with values uniformly distributed between zero and 10, their
product will consist of many outcomes near zero and some large outcomes,
creating the skewed distribution shown in 5.2.
image
Figure 5.2: A Lognormal Distribution
The length of the tail in a lognormal distribution depends on the variance of
the random variables multiplied together. If they have low variance, the tail
will be short. If they have high variance, the tail can be quite long because,
as noted, multiplying together a sequence of large numbers produces a very
large number. Lognormal distributions arise in a wide range of examples,

including the sizes of British farms, the concentration of minerals in the
earth, and the time from infection with a disease to the appearance of
symptoms.9 Income distributions within many countries approximate
lognormal distributions, though many deviate from lognormal at the upper
end by having too many people with high incomes.
A simple model that can explain why income distributions are closer to
lognormal than normal links policies about salary increases to their implied
distributions. Most organizations assign raises by percentages. People who
perform above average receive high-percentage raises. People who perform
below average receive low-percentage raises. Instead, organizations could
assign raises by absolute amounts. The average employee could receive a
$1,000 raise. Those who perform better could receive more, and those who
perform worse could receive less. The distinction between percentages and
absolute amounts may appear semantic, but it is not.10 Allocating raises by
percentages based on employee performance when performances from year
to year are independent and random produces a lognormal distribution.
Differences in income become exacerbated in future years even with
identical subsequent performance. An employee who has performed well in
the past and earns $80,000 will receive $4,000 from a 5% raise. Another
employee, who earns only $60,000, receives only $3,000 from the same 5%
raise. Inequality begets more inequality even with identical performance.
Had the organization allocated raises by absolute amounts, the two
employees would receive the same raise and the resulting distribution of
incomes would be closer to a normal distribution.

Summary
In this chapter, we covered the structure, logic, and function of normal
distributions. We saw that normal distributions can be characterized by a
mean and a standard deviation. We described the central limit theorem,
which shows how normal distributions arise whenever we add up or
average independent random variables with finite variance. And we
described formulae for the standard deviations of the mean and sum of
random variables. We then showed the consequences of those properties.
We learned that small populations will be far more likely to produce

exceptional events and how when we lack that insight we make improper
inferences and take unwise actions. We learned how assumption of
normally distributed random variables allows scientists to make claims
about the significance and power of statistical tests, and how process
management can predict the likelihood of failure using an assumption of
normality.
Not every quantity can be written as the sum, or the average, of
independent random variables. Thus, not every distribution will be normal.
Some quantities are products of independent random variables and will be
lognormally distributed. Log-normal distributions only take on positive
values. They also have longer tails, which means more large events and
many more very small events. Those tails become long when random
variables multiplied together have high variance. Long-tailed distributions
imply less predictability, whereas normal distributions imply regularity. As
a rule, we prefer regularity to the potential for large events. Therefore, we
benefit from knowing the logic that creates the various distributions. In
general, we would prefer that we add random shocks rather than multiply
them together so as to reduce the likelihood of large events.