The Normal Distribution

Continuous random variables and the normal distribution. Densities and distribution functions; the normal density, with proofs that it integrates to 1 and has mean mu and variance sigma squared; standardization and the standard normal table; the 68-95-99.7 rule; checking normality with histograms, normal probability plots and the Shapiro-Wilk test; the log, reciprocal and Box-Cox transformations; and why the normal curve arises. Worked on the heights and systolic blood pressures of NHANES adults.

The 2,517 men examined in the National Health and Nutrition Examination Survey (NHANES) of 2017 to 2018 have a mean height of 173.53 cm and a standard deviation of 7.684 cm. Suppose all we know about them is these two numbers and the assumption that heights follow a normal curve. That is enough to predict that 1.61% of the men are taller than 190 cm; the data hold 1.55%. It predicts that 66.66% are between 165 and 180 cm, against 67.54% observed, and that the tallest tenth begins at 183.38 cm, against 183.50 cm observed. Two numbers and a curve reproduce the whole distribution of 2,517 measurements.

The same two numbers for systolic blood pressure do much worse. Among 5,247 adults the mean is 126.36 mm Hg and the standard deviation 20.125. The normal curve predicts that 0.38% of adults have a pressure of 180 or more, about 20 people; there are 95. It predicts 9.51% below 100; there are 4.67%. Why does the curve work so well for one measurement and so badly for the other, and how can we tell which case we are in?

The normal distribution is the bell-shaped distribution, fixed by its mean and standard deviation, that describes the heights. This article develops it in four steps. First we set up the mathematics of continuous random variables that it needs. Then we define the curve and prove its properties, and show how standardization, tables and software turn it into probabilities. Next we take up the practical questions: how to tell whether data are close enough to normal, and what to do when they are not. Finally we ask why so many measurements, and so many statistics, turn out to be normal at all.

The article assumes the expected values and variances of The Binomial Distribution. It is the course's home for the normal distribution: every later unit that assumes normal data, or a normal sampling distribution, builds on the facts proved here.

Continuous random variables

The counts of The Binomial Distribution and The Poisson Distribution take separate whole-number values, each with a positive probability. A measurement such as height can take any value in an interval. The probability that it equals one exact value, 173.5000... cm, is zero. So probabilities attach instead to intervals. They are described by a probability density function (pdf) $f$, a nonnegative function whose area over an interval is the probability of the interval:

$$P(a \le X \le b) = \int_a^b f(x)\,dx, \qquad \int_{-\infty}^{\infty} f(x)\,dx = 1.$$

The second condition says that the total probability is 1. A density is not itself a probability: $f(x)\,dx$ is approximately the probability of the short interval from $x$ to $x + dx$. Because single points have probability zero, it makes no difference whether the ends of an interval are included: $P(a < X < b) = P(a \le X \le b)$. The discrete distributions do not share this convenience.

Check your understanding. A probability can never exceed 1. Can a density?

Yes. A density is probability per unit of measurement. The normal model of men's heights peaks at $1/(\sigma\sqrt{2\pi}) = 0.0519$ per centimeter, which is 5.192 per meter. The areas, which are the probabilities, are the same in both units.

The cumulative distribution function is defined as for discrete variables, $F(x) = P(X \le x)$. It is now the area to the left of $x$:

$$F(x) = \int_{-\infty}^{x} f(t)\,dt, \qquad f(x) = F'(x),$$

the second by the fundamental theorem of calculus. The expected value and variance replace sums by integrals:

$$E[X] = \mu = \int_{-\infty}^{\infty} x f(x)\,dx, \qquad \operatorname{Var}(X) = \sigma^2 = \int_{-\infty}^{\infty} (x - \mu)^2 f(x)\,dx.$$

Every rule proved for discrete variables in The Binomial Distribution carries over with the same proofs (Grinstead and Snell, 1997, §5.2 and §6.3):

  • linearity of expectation;
  • $\operatorname{Var}(aX + b) = a^2\operatorname{Var}(X)$ and $\operatorname{Var}(X) = E[X^2] - \mu^2$;
  • the additivity of variances for independent variables.

The normal density

A random variable $X$ has the normal distribution with mean $\mu$ and standard deviation $\sigma > 0$ if its density is

$$f(x) = \frac{1}{\sigma\sqrt{2\pi}}\, e^{-(x - \mu)^2/(2\sigma^2)}, \qquad -\infty < x < \infty.$$

Following the course's formula sheet, this is written $X \sim N(\mu, \sigma)$, with the standard deviation as the second argument. Many books and most software documentation write the variance there instead, $N(\mu, \sigma^2)$, so it is worth checking which convention a source uses. The case $\mu = 0$, $\sigma = 1$ is the standard normal distribution. Its density and distribution function have their own symbols:

$$\phi(z) = \frac{1}{\sqrt{2\pi}}\, e^{-z^2/2}, \qquad \Phi(z) = \int_{-\infty}^{z} \phi(t)\,dt.$$

The shape follows from the exponent:

  • The density depends on $x$ only through $(x - \mu)^2$, so it is symmetric about $\mu$, which is therefore also its median.
  • It is largest where the exponent is zero, at $x = \mu$, with height $1/(\sigma\sqrt{2\pi})$, which is 0.3989 for the standard normal.

Differentiating twice,

$$f'(x) = -\frac{x - \mu}{\sigma^2} f(x), \qquad f''(x) = \frac{(x - \mu)^2 - \sigma^2}{\sigma^4} f(x).$$

So the curve bends downward between $\mu - \sigma$ and $\mu + \sigma$ and upward outside them. The inflection points, where the bell turns into its tails, are exactly one standard deviation from the mean. Beyond them the density falls off faster than any exponential. Values more than a few standard deviations from the mean are very rare, though never impossible.

The next two proofs show that this is a density and that $\mu$ and $\sigma$ earn their names.

Proof that the total area is 1

The constant $1/\sqrt{2\pi}$ is what makes the area 1. The integral of $e^{-z^2/2}$ has no elementary antiderivative, so the proof uses a trick: square the integral, read the square as a double integral over the plane, and switch to polar coordinates, where it becomes easy.

Write $I = \int_{-\infty}^{\infty} e^{-z^2/2}\,dz$. Then

$$I^2 = \int_{-\infty}^{\infty}\int_{-\infty}^{\infty} e^{-(x^2 + y^2)/2}\,dx\,dy = \int_0^{2\pi}\int_0^{\infty} e^{-r^2/2}\, r\,dr\,d\theta = 2\pi \Big[-e^{-r^2/2}\Big]_0^{\infty} = 2\pi,$$

by changing to polar coordinates, $x^2 + y^2 = r^2$ and $dx\,dy = r\,dr\,d\theta$. So $I = \sqrt{2\pi}$ and $\int \phi = 1$. For general $\mu$ and $\sigma$, the substitution $z = (x - \mu)/\sigma$, $dx = \sigma\,dz$, reduces $\int f$ to $\int \phi$. $\square$

Proof that the mean is $\mu$ and the variance $\sigma^2$

For the standard normal, $z\phi(z)$ is an odd function, so its integral over the whole line is zero. (It converges, since $\phi$ falls off so fast.) So $E[Z] = 0$. For the variance, integrate by parts with $u = z$ and $dv = z\phi(z)\,dz$, for which $v = -\phi(z)$ because $\phi'(z) = -z\phi(z)$:

$$E[Z^2] = \int_{-\infty}^{\infty} z \cdot z\phi(z)\,dz = \Big[-z\phi(z)\Big]_{-\infty}^{\infty} + \int_{-\infty}^{\infty}\phi(z)\,dz = 0 + 1 = 1.$$

So $Z$ has mean 0 and variance 1. The general case follows from the next section: $X = \mu + \sigma Z$, so $E[X] = \mu$ and $\operatorname{Var}(X) = \sigma^2$. $\square$

The script checked all three integrals numerically for $\phi$, with the midpoint rule on $[-12, 12]$. It found 1.0000000000 for the area, zero to rounding error for the mean, and 1.0000000000 for the variance.

Standardization

Do we need a table for every $\mu$ and $\sigma$? No: one change of scale turns every normal distribution into the standard one.

Proposition. If $X \sim N(\mu, \sigma)$, then $Z = (X - \mu)/\sigma \sim N(0, 1)$. More generally, $aX + b \sim N(a\mu + b, |a|\sigma)$ for constants $a \neq 0$ and $b$.

Proof. For any $z$, using the substitution $t = (x - \mu)/\sigma$,

$$P(Z \le z) = P(X \le \mu + \sigma z) = \int_{-\infty}^{\mu + \sigma z} \frac{1}{\sigma\sqrt{2\pi}} e^{-(x - \mu)^2/(2\sigma^2)}\,dx = \int_{-\infty}^{z} \frac{1}{\sqrt{2\pi}} e^{-t^2/2}\,dt = \Phi(z).$$

The general statement follows in the same way, with the inequality reversed when $a < 0$. $\square$

So a linear change of units keeps a normal variable normal: heights in inches are normal if heights in centimeters are. And every normal probability reduces to a standard normal one:

$$P(X \le x) = \Phi\!\left(\frac{x - \mu}{\sigma}\right).$$

The quantity $z = (x - \mu)/\sigma$ is the z-score of Measures of Center and Spread, the number of standard deviations $x$ lies from the mean. What the normal model adds is a probability for each z-score.

The standard normal density, with the shares of the area within one, two and three standard deviations of the mean. The top axis shows the same curve in centimeters for the normal model of men's heights, mean 173.5 cm and standard deviation 7.68 cm. Notice that only the labels change: standardization is a change of scale.
The standard normal density, with the shares of the area within one, two and three standard deviations of the mean. The top axis shows the same curve in centimeters for the normal model of men's heights, mean 173.5 cm and standard deviation 7.68 cm. Notice that only the labels change: standardization is a change of scale.

Worked: men's heights

Back to the men's heights of the opening, with $\mu = 173.53$ and $\sigma = 7.684$. A height of 190 cm has $z = (190 - 173.53)/7.684 = 2.143$, and

$$P(H > 190) = 1 - \Phi(2.143) = 0.0161,$$

against the observed share of 39/2,517 = 0.0155. For the range from 165 to 180 cm, the z-scores are $-1.110$ and $0.842$, and

$$P(165 \le H \le 180) = \Phi(0.842) - \Phi(-1.110) = 0.6666,$$

against 0.6754 observed.

Questions can also run backward, from a probability to a value. The 90th percentile of the standard normal is $z_{0.90} = 1.2816$, so the model's 90th percentile of height is $\mu + 1.2816\sigma = 183.38$ cm; the data's is 183.50 cm. And the model gives 3.91% of men below 160 cm, where the data have 4.13%. Every prediction is within about one percentage point.

Check your understanding. Without a table, where does the model put the shortest tenth of the men?

At $\mu - 1.2816\sigma = 163.68$ cm, as far below the mean as the 90th percentile lies above it, by symmetry. The data's 10th percentile is 164.10 cm. Likewise, the model makes 157.06 cm or less as rare as 190 cm or more, 0.0161 each; in the data, 1.67% of men are that short.

Tables and software

$\Phi$ has no formula in elementary functions, so it must be computed numerically (NIST/SEMATECH, §1.3.6.6.1). For most of the twentieth century this meant a printed standard normal table of $\Phi(z)$ for $z$ from 0 to about 3.5 in steps of 0.01, and statistics courses still teach reading one. Three facts make a table of positive $z$ sufficient:

  • symmetry: $\Phi(-z) = 1 - \Phi(z)$, so the left tail beyond $-z$ equals the right tail beyond $z$;
  • complements: $P(Z > z) = 1 - \Phi(z)$;
  • intervals: $P(a < Z < b) = \Phi(b) - \Phi(a)$.

A few values recur so often in inference that they are worth knowing:

$z$ $\Phi(z)$ Use
1.2816 0.9000 80% two-sided, 90% one-sided
1.6449 0.9500 90% two-sided, 95% one-sided
1.9600 0.9750 95% two-sided
2.5758 0.9950 99% two-sided

Software computes $\Phi$ to full machine precision, usually through the error function, $\Phi(z) = \tfrac{1}{2}\left[1 + \operatorname{erf}(z/\sqrt{2})\right]$. It also computes the inverse, the quantile function, for any probability. Rounding a z-score to two decimals to look it up in a table loses a little accuracy. Using software loses none, and it is what the rest of the course uses. The table remains useful for understanding, because it shows at a glance how quickly the tail areas shrink.

The 68–95–99.7 rule

Integrating $\phi$ between $-k$ and $k$ gives the share of a normal distribution within $k$ standard deviations of its mean:

Within Probability Outside
$\mu \pm \sigma$ 0.6827 0.3173
$\mu \pm 2\sigma$ 0.9545 0.0455
$\mu \pm 3\sigma$ 0.9973 0.0027

Rounded, these are the 68–95–99.7 rule, or empirical rule. Exactly 95% lies within $\pm 1.96\sigma$, which is where the 1.96 of every 95% confidence interval comes from. The rule holds for every normal distribution, whatever its mean and standard deviation, because standardization reduces each one to the standard normal.

Measures of Center and Spread tests the rule on NHANES data. Men's heights follow it closely, while skewed measurements have too many values within one standard deviation and too many beyond three. For data of any shape, Chebyshev's inequality, proved there, guarantees only that at least 75% lie within two standard deviations.

Checking normality

The heights passed every comparison so far, and blood pressure failed. Every method in this course that assumes normal data needs some way to tell whether the assumption is reasonable: the $t$ interval and $t$-tests, analysis of variance, and the prediction intervals of regression. Three tools are in common use.

Histograms

Histograms of men's heights and of systolic blood pressure for NHANES adults, with the normal density of the same mean and standard deviation. The heights follow the curve closely. Notice that the pressures rise steeply on the left and trail off to the right, so the normal curve puts too much probability below 100 mm Hg and far too little above 170. The jagged bars in the pressure histogram come from readings recorded as even numbers.
Histograms of men's heights and of systolic blood pressure for NHANES adults, with the normal density of the same mean and standard deviation. The heights follow the curve closely. Notice that the pressures rise steeply on the left and trail off to the right, so the normal curve puts too much probability below 100 mm Hg and far too little above 170. The jagged bars in the pressure histogram come from readings recorded as even numbers.

A histogram with the fitted normal curve overlaid shows gross departures such as skewness or two peaks. But its appearance depends on the choice of bins. And the tails, where departures matter most for inference, are hard to judge on it.

Normal probability plots

A normal probability plot, the normal case of the quantile–quantile (Q–Q) plot, looks at the tails directly. Here is the idea in one sentence: if the data are normal, the $i$-th smallest value should sit about where the $i$-th smallest of a normal sample would, so plotting one against the other should give a straight line.

The construction. The plot compares the ordered data with the values a normal sample of the same size would be expected to have (NIST/SEMATECH, §1.3.3.21):

  1. Sort the $n$ observations, $x_{(1)} \le \cdots \le x_{(n)}$.
  2. For each rank $i$, compute a standard normal quantile $m_i$ that a normal sample's $i$-th smallest value would typically match. The NIST handbook uses Filliben's medians of the uniform order statistics, $u_i = (i - 0.3175)/(n + 0.365)$ for $1 < i < n$, with $u_n = 0.5^{1/n}$ and $u_1 = 1 - u_n$, and then $m_i = \Phi^{-1}(u_i)$. The simpler $u_i = (i - 0.5)/n$ gives nearly the same picture.
  3. Plot $x_{(i)}$ against $m_i$.

If the data are normal with mean $\mu$ and standard deviation $\sigma$, then $x_{(i)} \approx \mu + \sigma m_i$. The points then lie close to a straight line with intercept $\mu$ and slope $\sigma$.

Normal probability plots of men's heights, systolic blood pressure, its logarithm and 1000 divided by it, with the line mean plus SD times the normal quantile. Heights lie on the line. Notice that blood pressure bends upward like a smile: the lower tail is too short and the upper tail too long for a normal distribution, the signature of right skew. The log reduces the bend; the reciprocal nearly removes it. PPCC is the correlation of the points.
Normal probability plots of men's heights, systolic blood pressure, its logarithm and 1000 divided by it, with the line mean plus SD times the normal quantile. Heights lie on the line. Notice that blood pressure bends upward like a smile: the lower tail is too short and the upper tail too long for a normal distribution, the signature of right skew. The log reduces the bend; the reciprocal nearly removes it. PPCC is the correlation of the points.

The shape of the departure tells what kind it is:

  • a curve bending upward at both ends, like the blood pressure plot, means right skew; downward at both ends, left skew;
  • an S shape, with points below the line on the left and above it on the right, means heavier tails than the normal; the reverse S means lighter tails;
  • isolated points far off the line at the ends are outliers.

The correlation between the $x_{(i)}$ and the $m_i$, the probability plot correlation coefficient (PPCC), condenses the plot into one number, 1 for a perfect line. For men's heights it is 0.99979, for systolic pressure 0.97232.

The number must be judged against the sample size. For 30 random samples of 25 values drawn from a true normal distribution, the PPCC had median 0.9858 (interquartile range 0.9819 to 0.9897) and fell as low as 0.9580. Small normal samples look less straight than large ones. A plot of 25 points needs a far wider allowance for wiggles than one of 2,500.

Formal tests

Tests of normality turn the question into a hypothesis test, with normality as the null hypothesis. The most widely used is the Shapiro–Wilk test. It compares a weighted combination of the ordered values, with weights derived from the expected values of normal order statistics, to the sample variance. Its statistic $W$ is at most 1, and small values are evidence against normality (NIST/SEMATECH, §7.2.1.3). Applied to all 2,517 men's heights, it gives $W = 0.99945$ and $p = 0.709$: even with thousands of observations, no departure from normality is detected.

Check your understanding. A researcher draws 50 adults, tests their systolic pressures with Shapiro–Wilk, gets $p = 0.2$, and concludes that blood pressure is normally distributed. What is wrong with the conclusion?

A nonsignificant test does not show that data are normal, only that the sample was too small to detect the departure. For each of 30 seeds the script drew 50 adults' systolic pressures, from a distribution whose skewness of 1.04 is plain in every figure above. The $p$-values had median 0.014 (interquartile range 0.003 to 0.139), and only 57% of the samples were rejected at the 5% level. Samples of 50 men's heights were rejected 7% of the time, close to the nominal 5%.

So the tests are weakest where they matter most:

  • in small samples, where methods depend most on normality, tests often miss real skewness;
  • in very large samples, where the central limit theorem makes many methods insensitive to it, they can flag departures too small to matter.

The plot, which shows what kind of departure there is and how large, is generally the more useful tool, with a test as a supplement.

Transformations

What can be done about skewed data? A transformation can often make them close to normal. For right-skewed positive measurements, the usual choices compress the large values more than the small ones: the square root, the logarithm and the reciprocal, in increasing order of strength. Box and Cox (1964) organized them into one family with a parameter $\lambda$,

$$T(y) = \frac{y^{\lambda} - 1}{\lambda} \quad (\lambda \neq 0), \qquad T(y) = \ln y \quad (\lambda = 0).$$

In this family $\lambda = 1$ leaves the shape unchanged, $\lambda = 1/2$ is essentially the square root, $\lambda = 0$ the logarithm and $\lambda = -1$ the reciprocal. The subtraction of 1 and the division by $\lambda$ make the family continuous at 0 and keep the order of the values. The Box–Cox normality plot of the NIST handbook chooses $\lambda$ by computing the PPCC of the transformed data for a range of $\lambda$ and taking the largest (NIST/SEMATECH, §1.3.3.6).

Probability plot correlation of systolic blood pressure after the Box-Cox transformation with parameter lambda, for lambda from -2 to 2. No transformation (lambda = 1) gives 0.972, the logarithm 0.992 and the reciprocal 0.999. Notice that the curve peaks close to lambda = -1, the reciprocal.
Probability plot correlation of systolic blood pressure after the Box-Cox transformation with parameter lambda, for lambda from -2 to 2. No transformation (lambda = 1) gives 0.972, the logarithm 0.992 and the reciprocal 0.999. Notice that the curve peaks close to lambda = -1, the reciprocal.

For systolic pressure, the best $\lambda$ is $-1.15$, with a PPCC of 0.99911, and the reciprocal is nearly as good, 0.99902. The skewness falls from 1.038 for the raw pressures to 0.525 for their logarithm and $-0.067$ for their reciprocal.

The transformation also repairs the model's tail probabilities. Fit a normal distribution on the transformed scale, and transform the cut-offs to match:

Event Observed Normal on raw scale On log scale On reciprocal scale
SBP $\ge$ 140 0.2121 0.2490 0.2259 0.2109
SBP $\ge$ 180 0.0181 0.0038 0.0081 0.0163
SBP $<$ 100 0.0467 0.0951 0.0720 0.0552

Read the row for 180 or more. The raw normal model gave 0.0038, about 20 adults where there are 95; on the reciprocal scale it gives 0.0163, close to the observed 0.0181. Body mass index behaves similarly: its skewness of 1.206 falls to 0.343 on the log scale, and its PPCC rises from 0.96694 to 0.99670.

A transformation has a cost in interpretation. Means and differences on a log scale become geometric means and ratios on the original scale, which are often natural for biological measurements. The reciprocal of a blood pressure has no clinical meaning, and results computed on that scale must be transformed back before anyone can use them. A modest transformation with a clear meaning, usually the logarithm, is often preferred to the statistically best one.

Transforming is not the only remedy. The methods of Nonparametric Methods avoid the normal assumption altogether. And for inference about means, large samples make it less important, as the next unit shows.

Why the normal distribution arises

Why should one curve fit any real measurement at all?

The normal curve first appeared in 1733, in Abraham de Moivre's approximation to binomial probabilities for a large number of trials (O'Connor and Robertson, 2004). It is the de Moivre–Laplace theorem of The Normal Approximation to the Binomial. In the nineteenth century it became the standard model for errors of measurement, which is why it is also called the Gaussian distribution, after Carl Friedrich Gauss.

The connection between binomial counts and errors of measurement is the reason the curve is everywhere. A binomial count is a sum of many small independent contributions, the Bernoulli trials. A Poisson count with a large mean is likewise a sum of many independent smaller counts. In both cases the skewness, $(1 - 2p)/\sqrt{np(1 - p)}$ and $1/\sqrt{\lambda}$, shrinks toward zero as the number of contributions grows. The central limit theorem states that this is no accident. The sum or mean of many independent contributions with finite variance has a distribution close to normal, whatever the distribution of the contributions. It is the subject of The Central Limit Theorem. It is what makes the normal distribution the basis of inference about means in The Sampling Distribution of the Mean and every unit after it.

The theorem suggests which measurements will be close to normal: those built up additively from many small, independent influences of similar size, and measured within a homogeneous group. Adult height within one sex fits that description. Pooling the two sexes does not. The heights of all 5,184 adults have an excess kurtosis of $-0.496$, a flatter top and shorter tails than the normal, because they mix two normal distributions 13.7 cm apart. Their PPCC, 0.99707, is lower than that of either sex alone (0.99979 for men, 0.99896 for women).

Blood pressure fails for other reasons. Its influences, such as age, weight and the stiffening of arteries, tend to act multiplicatively rather than additively. And the population mixes people with and without hypertension. Multiplicative effects are additive on the log scale, which is one reason the logarithm helps.

The normal distribution is therefore a model, not a law of nature. It is exact for no real measurement: heights cannot be negative, and every normal distribution gives negative values some probability. It is very close for some measurements and a poor fit for others. Its greatest importance is not as a description of raw data at all, but as the distribution of means and other statistics computed from many observations.

Corrections to the course notes

The unit's notes give the normal density, the standard normal, z-scores and the empirical rule with the values 68.27%, 95.45% and 99.73%, all as here; the rounded "95% within two standard deviations" is exact at 1.96 standard deviations.

  • They list the heights of individuals as an example of a normal distribution. More precisely, heights are close to normal within a sex, and the pooled heights of men and women are a mixture of two normal distributions with a flatter shape.
  • They also list blood pressure among measurements that are often normal. Among NHANES adults systolic pressure is skewed to the right, as shown above, and close to normal only after a transformation.
  • Their recommendation to check normality with Q–Q plots and with the Shapiro–Wilk and Kolmogorov–Smirnov tests rather than assume it is developed above, with the caution that a nonsignificant test in a small sample is weak evidence of normality.

Summary

Quantity Formula
Density $f(x) = \dfrac{1}{\sigma\sqrt{2\pi}}e^{-(x - \mu)^2/(2\sigma^2)}$; notation $N(\mu, \sigma)$
Probability of an interval $P(a \le X \le b) = \int_a^b f(x)\,dx = \Phi\big(\frac{b - \mu}{\sigma}\big) - \Phi\big(\frac{a - \mu}{\sigma}\big)$
Standardization $Z = (X - \mu)/\sigma \sim N(0, 1)$; $aX + b \sim N(a\mu + b, \lvert a\rvert\sigma)$
Symmetry $\Phi(-z) = 1 - \Phi(z)$
Quantiles $z_{0.90} = 1.2816$, $z_{0.95} = 1.6449$, $z_{0.975} = 1.9600$, $z_{0.995} = 2.5758$
Empirical rule 0.6827, 0.9545, 0.9973 within 1, 2, 3 SD
Inflection points $\mu \pm \sigma$
Probability plot ordered data against normal quantiles; straight line if normal
Box–Cox $(y^\lambda - 1)/\lambda$, or $\ln y$ at $\lambda = 0$

What we have learned

Probabilities for a continuous random variable are areas under its density. The normal density is fixed by its mean and standard deviation, and standardization reduces every normal probability to one table. For men's heights, two numbers and the curve predict every share within about a percentage point. For systolic blood pressure they do not, because the pressures are skewed. A normal probability plot shows the skew as a smile, and a Shapiro–Wilk test on a small sample can easily miss it. A logarithm or reciprocal brings the pressures close to normal, at some cost in interpretation.

A mean of many independent contributions tends to be normal even when the contributions are not. Blood pressure is skewed, but what about the mean blood pressure of 25 adults drawn at random? How much does it vary from sample to sample, and what shape does its distribution take? That is the next unit, beginning with The Sampling Distribution of the Mean.

In Python

from statistics import NormalDist
from scipy import stats

H = NormalDist(173.53, 7.684)              # men's heights, cm
print(1 - H.cdf(190))                      # P(H > 190): about 0.0161
print(H.cdf(180) - H.cdf(165))             # about 0.6666
print(H.inv_cdf(0.90))                     # 90th percentile: about 183.38
print(NormalDist().inv_cdf(0.975))         # 1.959964

# normal probability plot and Shapiro-Wilk test of a sample x:
# (osm, osr), (slope, intercept, r) = stats.probplot(x, dist='norm')
# W, p = stats.shapiro(x)

scipy.stats.probplot uses Filliben's medians, as above, and returns the PPCC as r. Every number in this article is computed by tools/experiments/applied-statistics/normal-distribution.py in the site's repository, from the NHANES files loaded by descdata.py; the Shapiro–Wilk results used scipy 1.15.2.

References