—

Advanced Statistical Methods

Normal Distribution and Statistical Inference

Published: 2026-08-11

Level: postgraduate

Audience: Postgraduate students in Advanced Statistical Methods

Statistical inference is the bridge from data to decisions: we observe a sample and try to say something trustworthy about the whole population. The normal distribution is the workhorse under that bridge. This session reviews the normal (Gaussian) distribution and its properties, then uses it to walk into statistical inference — sampling distributions, estimation, hypothesis testing, and p-values.

The road map for the session is short and stacked: (1) what the normal distribution is and why its shape is special; (2) the standard normal variable $z$ and how to read probability from the tables; (3) the second big job of $z$ — flagging outliers; (4) the vocabulary of inference — population, sample, parameter, statistic; (5) the two branches of statistical inference, starting with estimation; and (6) the sampling distribution of the sample mean, where the central limit theorem makes the normal distribution do the heavy lifting for almost everything that follows in the course.

2.1 The Normal Distribution (Gaussian)

2.1.1 What the Normal Distribution Is

Why does almost every remaining topic in this course — sampling, estimation, hypothesis testing, p-values — stand on one particular curve? Because that curve, the normal distribution, describes what "average plus natural variation" looks like over and over again in real measurements, and because it has one property that makes hand calculation possible: perfect symmetry.

The normal distribution — also called the Gaussian distribution or the z distribution, and these are all one and the same — is one of the finest and most important continuous distributions in statistics. You likely met it in earlier probability or statistics courses, where discrete distributions also appear. This session spends the most time on the normal distribution because almost everything in statistical inference leans on it.

The two big shape facts. First, it is a bell-shaped curve. Second, it is a symmetric distribution: the two extremes are always rare, while the middlings are always common. Extreme low values and extreme high values carry very little probability; values near the center carry the most. Because the curve is symmetric, the area to the right of the center equals the area to the left of the center. That 50/50 split is what makes so many table computations fall out neatly.

An everyday picture helps fix the shape. Take the heights of every person in a large city and plot how many fall at each height. Most people cluster near the middle, and the count thins out smoothly in both directions — a few very short people, a few very tall people, nobody below zero or above, say, nine feet. That is the bell: rare extremes, common middlings. The analogy holds as long as the variation is natural and additive; it breaks the moment a process has a hard floor (incomes, which cannot go below zero, skew right) or one dominant cause of variation (a machine that drifts off setting produces a shifted bell, not a normal one).

The formal definition, the one that appears in any standard textbook: a continuous random variable $X$ is said to have a normal distribution with the parameters $μ$ and $σ$ . Two parameters give the complete picture of any distribution — discrete or continuous — as seen last session: $μ$ tells you where the distribution is centrally located (the population mean), and $σ$ tells you how it is spread (the population standard deviation). Last session's old-drug-versus-new-drug example made this concrete: to anticipate the efficacy of a drug — to judge whether one is the better drug — you compare both parameters, where the two distributions sit and how widely they spread. (Real-world: this location-and-spread comparison is exactly how pharmaceutical trials compare treatment effects.)

The normal distribution. A continuous random variable $X$ follows a normal distribution with mean $μ$ and standard deviation $σ$ , written $X \sim N (μ, σ^{2})$ , when its density curve is the bell described below. Three properties follow directly from the definition:

The highest point of the curve is at the mean, which is also the median and the mode — for a normal curve, all three centers coincide.
The curve is symmetric about $μ$ , so it has zero skewness: the mirror image of the left side is the right side.
The total area under the curve is 1, split exactly 0.5 to the left of $μ$ and 0.5 to the right.
The tails extend to infinity in both directions and never quite touch the horizontal axis, so extreme values are rare but never impossible.

One more label worth keeping straight: the normal distribution is a continuous random variable's distribution, in contrast to the discrete distributions from earlier sessions, where the probability sits at individual points. (Real-world: the normal model shows up in process times, measurements, prices, and almost any "average plus natural variation" setting, which is why it keeps appearing in every topic that follows.)

Scope: the normal model applies when a continuous quantity varies symmetrically around a central value and extremes are genuinely rare. It breaks down for bounded quantities with a floor (waiting times, prices of houses — skewed), for data with heavy tails (financial returns, which produce crashes far more often than the bell predicts), and for any process with a strong outside force such as a quota or a cap. When the bell does not fit, using normal-table probabilities anyway is a common and costly error.

2.1.2 The Probability Functions Behind the Normal Curve

Before the normal formula, here is a refresh on how probabilities are generated in the distributions you already know: with the parameter values, you can generate all the probabilities of that distribution.

If $X$ is a discrete random variable that behaves like a binomial distribution, the probability function is

$P (X = x) = (x n) p^{x} q^{n - x}$

where $n$ is the number of trials, $p$ is the probability of success on one trial, and $q = 1 - p$ . The parameters of the binomial distribution are $n$ and $p$ .

If $X$ follows a Poisson distribution, the probability function is

$P (X = x) = \frac{e ^{- λ} λ ^{x}}{x !}$

where $λ > 0$ (lambda) is the rate parameter — the average number of events in the interval. The single parameter of the Poisson distribution is $λ$ .

The pattern to notice: in each case, once you plug in the parameters, the function hands you the probability for any value $x$ . The normal distribution works the same way, but with one twist that makes it a density rather than a probability function. For a discrete distribution, plugging in $x$ gives the probability $P (X = x)$ directly. For a continuous distribution, the height $f (x)$ is not a probability at all — probability arrives only as the area under the curve between two values. That is why probabilities like $P (X = 11.5)$ exactly are always zero for a continuous variable, and only statements like $P (X \geq 11.5)$ make sense.

The normal distribution's probability density function (PDF) is the complicated-looking expression:

$f (x) = \frac{1}{σ 2 π} e^{- \frac{1}{2} (\frac{x - μ}{σ})^{2}}$

described in words as "1 by sigma root 2 pi, e power minus half, x minus mu by sigma, whole square." Here $x \in R$ is the value of the random variable, $μ \in R$ is the population mean, $σ > 0$ is the population standard deviation, $π \approx 3.14159$ , and $e \approx 2.71828$ is the base of the natural logarithm.

Read the formula piece by piece so it loses its scare factor. The factor $\frac{1}{σ 2 π}$ is a normalizing constant: it exists only so that the total area under the curve equals 1. The exponent $- \frac{1}{2} (\frac{x - μ}{σ})^{2}$ controls the shape: the quantity $\frac{x - μ}{σ}$ measures how far $x$ sits from the center in standard-deviation units, and the minus sign makes the density fall off as that distance grows. Squaring makes both sides (below the mean and above it) decay identically — that squaring is exactly where the symmetry of the bell comes from. At $x = μ$ the exponent is 0 and the curve reaches its peak $f (μ) = \frac{1}{σ 2 π}$ ; the larger $σ$ is, the lower and flatter that peak is.

Assumption: the PDF assumes a variable that is continuous and unbounded, with variation that is the sum of many small independent effects. The professor's advice stands: no need to mug it up — this formula is only for your reference. You will never be asked to evaluate the PDF by hand; you will be asked to use the two parameters to find areas, which is what the standard normal table does for you. What you must keep is the structure: two parameters, $μ$ for location and $σ$ for spread.

The refresh matters because these are the necessary distributions that help you understand the inference concepts much better.

2.1.3 Families of Normal Curves

There are many families of normal distributions: the mean may be the same while the variances are different; or both mean and variance are different; or the mean is different while the variance stays the same. Every one of these is still a normal distribution, because each one has the symmetric property — in every case the area to the right side is the same as the area to the left side. So "is it normal?" is never answered by the shape alone being a specific bell; any bell described by these two parameters defines a member of the normal family.

Picture three bells on one set of axes. The horizontal axis is the value of $x$ ; the vertical axis is the density $f (x)$ . Curve A: $μ = 0$ , small $σ$ — tall and narrow, squeezed tight around 0. Curve B: $μ = 0$ , large $σ$ — short and wide, spread far from 0. Curve C: $μ = 20$ , same $σ$ as B — identical in shape to B but shifted right so its peak sits at 20. All three are normal. The mean alone fixes where the bell stands; the standard deviation alone fixes how fat it is.

Family member	Mean $μ$	Standard deviation $σ$	What changes
Same center, different spread	same	different	Curve gets taller and narrower, or shorter and wider; the peak moves up and down
Different center, same spread	different	same	The whole bell slides left or right; shape is untouched
Both different	different	different	Location and spread both change; the bell shifts and reshapes at once

The takeaway for identification: the family is defined by the two parameters, not by any single bell. Given any $μ$ and any $σ > 0$ , the formula in section 2.1.2 draws a perfectly valid member of the normal family.

2.1.4 Worked Example: Probability That Assembly Takes at Least 11.5 Minutes

The first and most fully worked example of the session — the machinery assembly problem.

Problem. The time required to assemble a piece of machinery is a random variable having roughly a normal distribution with mean 12.9 and standard deviation 2. Find the probability that the assembly of a piece of machinery of this kind will take at least 11.5 minutes.

Step 1 — define the random variable. This is the usual practice whenever you deal with distributions: name your random variable first. Here $X$ = time required to assemble a piece of machinery, in minutes. The behavior of $X$ is given: $X \sim N (μ = 12.9, σ = 2)$ roughly.

Step 2 — translate the question into a probability statement. "At least 11.5 minutes" means $P (X \geq 11.5)$ .

Step 3 — the general method is integration. For discrete distributions you use summations; for continuous distributions you use integration. The general way to write this probability is

$P (X \geq 11.5) = \int_{11.5}^{\infty} \frac{1}{σ 2 π} e^{- \frac{1}{2} (\frac{x - μ}{σ})^{2}} d x$

with $μ = 12.9$ and $σ = 2$ . By looking at the structure of this improper integral you can tell it is really difficult to solve for probabilities: you need transformations, simplifications, and integration formulas, and only after all that work do you arrive at the answer. It is a very tedious way to do the task. (This integral has no elementary closed form — the exponential inside cannot be integrated by the standard rules you learned in calculus — which is exactly why nobody does it this way.)

Step 4 — transform into the standard normal variable. The trick that avoids all the calculus: if $X$ is a normal random variable with mean $μ$ and standard deviation $σ$ , convert it into its standard normal variable

$Z = \frac{X - μ}{σ}$

Then

$P (X \geq 11.5) = P (Z \geq \frac{11.5 - 12.9}{2}) = P (Z \geq - 0.7)$

The arithmetic: $11.5 - 12.9 = - 1.4$ , and $- 1.4/2 = - 0.7$ .

Step 5 — read the standard normal table. The normal tables give cumulative probabilities (everything from the lowest value up to the point you look up). The table gives $P (Z \leq - 0.7) = 0.2420$ . Because we want everything to the right of $- 0.7$ , we use the complement:

$P (Z \geq - 0.7) = 1 - P (Z \leq - 0.7) = 1 - 0.2420 = 0.7580$

Sense-check. 11.5 minutes sits below the mean of 12.9, so more than half of all assemblies should finish by 11.5 — and 0.7580 is indeed above 0.5. If the answer had come out below 0.5 for a value below the mean, a sign or direction error slipped in.

The takeaway repeated at the end: instead of the complicated integrals, it is always better to transform the given variable into its standard normal variable and then extract the probability.

Exam note: you will be expected to read values like this from the table, so the transformation and the table are the skills to practice. The routine is always the same three moves: standardize with $Z = (X - μ) / σ$ , decide which side of the point the question wants, and apply the complement rule $P (Z \geq a) = 1 - P (Z \leq a)$ when the question asks for the right side.

2.1.5 Student Questions and Answers

Several students pushed back on the integral during the session; their questions cluster into three confusion points.

Q: Solving that integral looks like a lot of work. Is there any way to avoid the tedious calculus?

A: Yes — nobody applies calculus to get normal probabilities. What we do instead is transform the given normal random variable into its standard normal variable and then read the normal tables. The tables are what everyone uses.

Q: What is that variable that the normal tables are built on?

A: It is z. For a normal $X$ with mean $μ$ and standard deviation $σ$ , set $Z = (X - μ) / σ$ . The moment you convert, the central line is zero, the total probability is one, and because the distribution is symmetric the area to the right of the center is 0.5 and the area to the left is 0.5.

Q: What does "cumulative" mean in these tables?

A: An exact probability asks about one value — for example, if $X$ equals 2, then $P (X = 2)$ is the exact probability. A cumulative probability asks about everything up to a value — you should say $X \leq 2$ , and compute from the lowest value to that point. The entire set of calculations in the tables is cumulative in this sense: for any point, the table value is the probability from the lower tail up to that point.

2.2 The Standard Normal Distribution and Z-Scores

2.2.1 Standardizing the Variable

The standard normal variable

$Z = \frac{X - μ}{σ}$

is $z$ — how many standard deviations the value $x$ sits away from the mean $μ$ . Standardizing is a change of units: whatever bell you started with, the transformed curve is centered at zero, carries total probability one, and is symmetric about zero. The two halves each hold probability 0.5. Instead of working out the tedious improper integral with calculus, you look up the answer in the table.

Think of standardizing as converting prices into a single currency so they can be compared. A price of 1,150 rupees and a price of 12 dollars cannot be compared directly until both are converted to the same currency; likewise, a value from a bell centered at 100 with spread 20 and a value from a bell centered at 0 with spread 1 cannot be compared until both are converted to "standard-deviation units." That is all $Z$ does: it re-labels every value by its distance from the mean, measured in $σ$ -units. The analogy breaks where the units matter for interpretation — a z-score tells you the distance but not the original measurement's size — which is why real problems always convert back at the end if an answer must be stated in original units.

Formally, the standardization $Z = (X - μ) / σ$ produces a random variable with:

mean $E (Z) = 0$ — the center of every bell is re-labeled 0;
standard deviation 1 — the spread of every bell is re-scaled to one $σ$ -unit;
the same bell shape — subtracting a constant shifts the curve, dividing by a constant stretches it, and neither operation destroys the symmetry.

The standard normal distribution is the special member of the normal family with $μ = 0$ and $σ = 1$ , written $Z \sim N (0, 1)$ .

A z-score of $- 0.7$ , for example, reads as "0.7 standard deviations below the mean"; a z-score of $+ 2$ reads as "2 standard deviations above the mean." Because every normal distribution standardizes to the same $N (0, 1)$ curve, one single table of areas serves every possible bell — that is the entire reason the standard normal exists.

2.2.2 Cumulative Probabilities and the Complement Rule

The tables you meet in textbooks give cumulative probabilities: the value at any point $z_{0}$ is $P (Z \leq z_{0})$ , the area under the curve from the far left up to $z_{0}$ . You read entries like "Z less than or equal to minus 0.7," "Z less than or equal to 2.9," and so on.

Cumulative probability (probability up to a value), written $P (Z \leq z_{0})$ , is the shaded area under the standard normal curve lying to the left of $z_{0}$ . Two landmarks anchor every reading: $P (Z \leq 0) = 0.5$ because the curve is symmetric about 0, and $P (Z \leq \infty) = 1$ because the total area is 1.

The complement rule. Many questions ask for the other side — "at least", "greater than", "to the right of". For those, flip the question: $P (Z \geq a) = 1 - P (Z \leq a)$ . In the assembly example,

$P (Z \geq - 0.7) = 1 - P (Z \leq - 0.7) = 1 - 0.2420 = 0.7580$

The interval rule. A question about the strip between two values subtracts the smaller cumulative probability from the larger one:

$P (a \leq Z \leq b) = P (Z \leq b) - P (Z \leq a)$

These two moves — complement for right-side questions, subtraction for between questions — cover nearly every table exercise in the course.

2.2.3 Worked Examples: Reading Table Values

We practice reading the table with a few quick values before moving on to outliers. "Now you tell me all these particular probabilities — it is only a practice now, and in the exam point of view also you should know how to retrieve the probability values."

Example 1 — a left-side probability. $P (Z \leq 1.5) = 0.9332$ . The center is 0, the value is 1.5, and this is the entire portion to the left of 1.5. Read it by finding the row 1.5 and the column .00 in the cumulative table: the entry is 0.9332. In the same way, $P (Z \leq 1) = 0.8413$ . Sense-check: both values sit above 0.5, as they must — 1 and 1.5 are both to the right of the center, so more than half the area lies to their left.

Example 2 — the strip between two values. $P (1 \leq Z \leq 1.5)$ : the strip between 1 and 1.5. Subtract the smaller cumulative probability from the larger:

$P (1 \leq Z \leq 1.5) = P (Z \leq 1.5) - P (Z \leq 1) = 0.9332 - 0.8413 = 0.0919$

Sense-check: the strip is narrow (0.5 of a standard deviation) and sits in the near-center region, so a small area around 9% is reasonable; the full right half is 0.5, so any strip inside it must be smaller.

Example 3 — the strip from 0 to 2.5. $P (0 \leq Z \leq 2.5)$ : the strip between 0 and 2.5. Here the left endpoint is the center itself, where the cumulative value is exactly 0.5. The table gives $P (Z \leq 2.5) = 0.9938$ , so

$P (0 \leq Z \leq 2.5) = P (Z \leq 2.5) - P (Z \leq 0) = 0.9938 - 0.5000 = 0.4938$

Sense-check: 2.5 is far into the right tail, so the strip from the center out to 2.5 covers almost all of the right half — but not all of it, so the answer must be just under 0.5. It is: 0.4938.

Example 4 — the right tail via the complement. $P (Z \geq 1.58)$ : read $P (Z \leq 1.58) = 0.9429$ from the table, then $P (Z \geq 1.58) = 1 - 0.9429 = 0.0571$ . Sense-check: 1.58 is well into the right side, so only a small tail area of about 5.7% should remain — and it does.

A habit worth building now: before reading any table, sketch the standard normal curve, mark the point or points, and shade the area the question asks for. The shaded picture tells you whether to take the table entry directly, subtract two entries, or use the complement — and it catches most sign and direction errors before they happen.

Exam note: knowing how to see the normal tables helps in attempting quizzes and exam questions of this type. The skills that get tested are: reading $P (Z \leq z_{0})$ from the row/column intersection, subtracting for strips, complementing for right tails, and checking each answer against the 0.5 landmark.

2.2.4 Student Questions and Answers

Q: How do we read a value like $P (Z \leq 1.5)$ from the tables?

A: Look up 1.5 in the cumulative probability table: $P (Z \leq 1.5) = 0.9332$ . In the same way, $P (Z \leq 1) = 0.8413$ . The row gives the first decimal of $z$ , the column gives the second, and the body of the table holds the area to the left of that point.

Q: What if the probability is between two values, such as $P (1 \leq Z \leq 1.5)$ ?

A: Subtract the smaller cumulative probability from the larger one: $P (Z \leq 1.5) - P (Z \leq 1) = 0.9332 - 0.8413 = 0.0919$ . You are removing the left part of the curve up to 1 and keeping only the slice between 1 and 1.5.

Q: What is the value of $P (0 \leq Z \leq 2.5)$ ? A student called out 0.5 during the session — is that right?

A: Not quite. The value 0.5 is the entire right half of the curve, from the center all the way to infinity. The strip asked for runs only from 0 up to 2.5, and 2.5 stops well short of infinity — the tiny remaining tail beyond 2.5 is excluded. The correct reading is $P (Z \leq 2.5) = 0.9938$ minus the left-half area $P (Z \leq 0) = 0.5000$ , which gives 0.4938: almost the whole right half, but slightly less than 0.5. The one-line check: any strip that stops at a finite value inside the right half must be smaller than 0.5.

2.3 Outlier Detection with Z-Scores

2.3.1 What Outliers Are

The z-score has a second big job beyond probability extraction: detecting outliers. Outliers are unusual values — "sometimes the data set will have one or more observations with unusually large or unusually small values; these extreme values are called outliers." In a raw data set, abnormal entries can definitely impact your study and your inferences, so detecting outliers and addressing them is very important.

Why outliers matter — the example: for a class, if you want to calculate the average age of, say, first-class or second-class students, it will be somewhere around eight years. Now suddenly add a teacher's age into that whole data set — a far higher value. You can easily see there is an impact on the entire average: it certainly is not that the average age of first-class students is around 20 or 30; the average got dragged up because of that one outlier present in the data set. So a single extreme entry can change the meaning of an average.

The size of the drag depends on how many normal values the outlier is competing with. In a class of five students averaging 8 years (total $5 \times 8 = 40$ years), adding one teacher aged 40 years changes the average to $(40 + 40) /6 \approx 13.3$ years — the class average jumps by more than five years because of one person. In a class of 500, the same teacher barely moves the average. The mechanism is identical either way: the outlier adds a large number to the numerator while counting as only one more item in the denominator. Averages are sensitive to extreme values, and that sensitivity is why outliers must be found before they silently change conclusions.

Where outliers come from. Many reasons exist. One is wrong entries — human errors in data entry. But there are other situations: in any manufacturing industry producing a particular product, at one point the machine may show abnormal behavior and then suddenly restore itself to normal behavior. That is not a wrongly entered value — it is something we require intervention for: there is an anomaly happening with the machine and we should rectify it. So outliers are not always wrongly entered data; there are also unusual behaviors. We cannot simply ignore the outliers; we try to justify how significant the outlier is. (Real-world: z-score monitoring in manufacturing quality control is precisely this — flagging the point where the process went abnormal so the intervention happens at the right moment.)

Assumption: the z-score outlier rule presumes the data come from a bell-shaped distribution, so "unusual" can be measured in standard-deviation units. When the distribution is strongly skewed, the same rule can mislabel legitimate large values as outliers or miss small-value outliers entirely; in those settings, distribution-specific checks (or the quartile rule of section 2.3.2) are safer. Note also the opposite failure mode: computing the mean and standard deviation including the suspect value lets a large outlier inflate both, shrink the z-score of every other value, and hide itself — so detection should always be separate from estimation.

2.3.2 Detecting Outliers: Box Plots and Z-Scores

There are many approaches to detecting outliers. A box plot is one such interesting thing — a simple pictorial way of doing it. A data set may have one or more observations with unusually large or unusually small values; the box plot shows those extreme values visually. The textbook discusses this in detail in Chapter 3.

The numerical approach uses the standard z-score:

$Z = \frac{X - μ}{σ}$

The simple logic: if the z-score of an observation is less than minus three or above plus three — that is $Z < - 3$ or $Z > 3$ — we can say that particular observation is an outlier. A quick example: data like 20, 40, 2000, 30, 60, 60, and so on, from a normal distribution with a given $μ$ and $σ$ . By simple inspection, without any calculation, you can say the 2000 is an abnormal entry — an outlier anyway. But to give numerical justification it is better to compute the z-score for that observation and compare it with the limits; the z-score is a very basic way to justify numerically how to identify outliers.

Why $\pm 3$ ? The empirical rule says almost all bell-shaped data lie within three standard deviations of the mean — so a value that lands beyond $\pm 3$ sits where, in a genuinely normal population, only about 3 values in 1,000 would be found. The rule is so a filter, not a verdict: a flagged observation is a candidate to review for accuracy and for whether it belongs in the data set at all. The textbook recommends treating any data value with a z-score less than $- 3$ or greater than $+ 3$ as an outlier.

The box plot gives a second, purely pictorial route based on quartiles rather than standard deviations. A box plot marks the median, the first quartile $Q_{1}$ , and the third quartile $Q_{3}$ as a box, draws whiskers out to the most extreme values inside the limits below, and plots anything beyond the limits as a separate dot:

$Lower limit = Q_{1} - 1.5 \times IQR, Upper limit = Q_{3} + 1.5 \times IQR$

where $IQR = Q_{3} - Q_{1}$ is the interquartile range. Any observation below the lower limit or above the upper limit is flagged as an outlier. The two methods do not always agree — a value can pass the z-score rule yet fail the quartile rule, or the other way around — and either one may be used.

Worked mini-example — the professor's data. Data: 20, 40, 2000, 30, 60, 60. Sorted: 20, 30, 40, 60, 60, 2000. For the box-plot route, $Q_{1} = 30$ (median of the lower half 20, 30, 40) and $Q_{3} = 60$ (median of the upper half 60, 60, 2000), so $IQR = 60 - 30 = 30$ . The limits are

$Lower limit = 30 - 1.5 (30) = - 15, Upper limit = 60 + 1.5 (30) = 105$

The value 2000 is greater than 105, so the box-plot rule flags 2000 as an outlier — matching what the eye saw instantly. The other five values all fall inside $[- 15, 105]$ . Sense-check: with five values clustered between 20 and 60, any number in the thousands must lie far outside the pattern of its neighbors; both numerical and pictorial methods agree.

2.3.3 The Empirical Rule

Whenever a distribution is bell-shaped — the data behaves like a normal distribution — the empirical rule lets us conclude:

about 68 percent of the observations lie within one-sigma limits,
95 percent of the observations lie within two-sigma limits,
99.7 percent of the observations lie within three-sigma limits.

So out of 20 observations, roughly 68 percent of them — about 14 values — would sit inside the one-sigma band. This behavior follows straight from the bell shape of the distribution: it is exactly what the symmetric, peaked curve guarantees.

The empirical rule is the bridge between the z-score and the normal curve's areas. "Within one-sigma limits" means the band $μ \pm σ$ , which in z-units is the strip $- 1 \leq Z \leq + 1$ ; the area of that strip is $P (Z \leq 1) - P (Z \leq - 1) = 0.8413 - 0.1587 = 0.6826 \approx 68%$ . The two-sigma band $μ \pm 2 σ$ holds about 95%, and the three-sigma band $μ \pm 3 σ$ holds about 99.7%. The more precise values from the normal curve are 68.3%, 95.4%, and 99.7%; the rounded 68-95-99.7 form is the one to quote. A rule of thumb worth keeping: the professor's version and the textbook's version are the same rule rounded differently.

Because the bell is symmetric, the bands split neatly in half: about 34% of the data lies between the mean and one standard deviation above it, and another 34% between the mean and one below. The same splitting trick answers questions like "how much of the data lies above the mean plus two sigma?": half of 95% leaves 2.5% in each tail, so about 2.5% lies above $μ + 2 σ$ and about 2.5% below $μ - 2 σ$ .

Worked mini-example — detergent cartons. A filling line packs liquid detergent with mean weight 16 ounces and standard deviation 0.25 ounces, and the fill weights are bell-shaped. The empirical rule says: about 68% of cartons weigh between $16 \pm 1 (0.25)$ , i.e., between 15.75 and 16.25 ounces; about 95% between $16 \pm 2 (0.25)$ , i.e., between 15.50 and 16.50 ounces; and almost all between 15.25 and 16.75 ounces. By symmetry, about 34% weigh between 16.00 and 16.25 ounces (half of the 68% band), and about 2.5% weigh below 15.50 (half of the 5% outside the two-sigma band). Sense-check: the bands grow wider as the sigma count grows — 15.75–16.25 is inside 15.50–16.50 is inside 15.25–16.75 — exactly as the rule promises.

2.3.4 Worked Example: California Backyard Structures

A note on the problem's naming convention first: from now on, when a problem is referenced by three numbers, the first position is the chapter, the second is the page number, and the third is the problem number. This one is Chapter 3, page 132, problem 42 — referenced as 3.132.42 — from the 13th edition.

Problem. Many families in California are using backyard structures for home offices, art studios, and hobby areas, as well as for extra storage. Suppose that the mean price for a customized wooden single backyard structure is $μ = $3, 100$ and the standard deviation is $σ = $1, 200$ .

Part a — the z-score for $x = 2300$ :

$Z = \frac{2300 - 3100}{1200} = \frac{- 800}{1200} \approx - 0.67$

Part b — the z-score for $x = 4900$ :

$Z = \frac{4900 - 3100}{1200} = \frac{1800}{1200} = 1.5$

Part c — comment on whether the second price is an outlier: not an outlier. Its z-score of 1.5 lies between $- 3$ and $+ 3$ , so by the z-score rule the value is not an outlier. A price 1,800 dollars above the average sits one and a half standard deviations up the curve — unusual but entirely plausible for this market.

Part d — a check for $x = 13, 000$ :

$Z = \frac{13000 - 3100}{1200} = \frac{9900}{1200} = 8.25$

Since $8.25 > 3$ , this value is an outlier. A price nearly ten thousand dollars above the mean is more than eight standard deviations away — statistically implausible for a market whose prices are roughly normal.

Sense-check. The two ordinary prices, 2,300 dollars and 4,900 dollars, both stay inside $\pm 3$ z-units and are accepted; the extreme quote of 13,000 dollars sails past the $+ 3$ boundary. The rule gives the same verdict the eye would give, but with a number behind it.

That is the very basic, numerically justified way to decide outliers: compare the z-score with the minus-three / plus-three limits. (Real-world: for a market with mean price 3,100 dollars and spread 1,200 dollars, a 13,000-dollar quote is statistically implausible — a flag worth investigating rather than accepting as typical. A contractor quoting 13,000 while the market centers on 3,100 should be asked why; the number is either a different product, an error, or something worth a second look.)

Exam note: the skills tested here are the z-score computation $Z = (X - μ) / σ$ with real numbers, the comparison against the $- 3$ / $+ 3$ limits, and the empirical rule 68-95-99.7 for bell-shaped data. Practice the four-step routine — state $μ$ and $σ$ , compute $Z$ , compare with the limits, state the verdict and why.

2.3.5 Student Questions and Answers

Q: What are the different approaches to detecting outliers?

A: A box plot is one such interesting approach — a simple pictorial way: the box carries the quartiles and whiskers stretch to the limits $Q_{1} - 1.5 (IQR)$ and $Q_{3} + 1.5 (IQR)$ , with anything beyond plotted as a separate point. In addition, the standard z-score approach works numerically: compute the z-score of the observation and compare it with the limits. Chapter 3 of the textbook has the detailed discussion of this.

Q: Is the 4,900 price an outlier?

A: No. Its z-score is 1.5, which lies between minus three and plus three, so that particular value is not an outlier. It is higher than the typical price — 1.5 standard deviations above the mean — but not so extreme that the data model calls it abnormal.

2.4 Population, Sample, Parameter, and Statistic

2.4.1 Parameter versus Statistic

Statistical inference is the next concept, and before talking about it there are two more terminologies to know, after population and sample.

A parameter is a characteristic of the population. For a given population, the value of that parameter is fixed — the parameter is always a constant. A statistic is a characteristic of the sample. The classic pair: if $μ$ is the population mean — the parameter — then the sample mean $X ˉ$ (x bar) is the statistic, the value we extract from the sample data. The key distinction to remember: a statistic is a random variable, whereas a parameter is a constant.

	Parameter	Statistic
Describes	the whole population	a sample drawn from it
Symbol (mean)	$μ$ (mu)	$X ˉ$ (x bar)
Value	fixed, one constant	changes from sample to sample
Known?	usually unknown — the target of inference	computed directly from the data

Every population has exactly one value of its mean $μ$ — it is what it is, whether we know it or not. Every sample of 20 students has its own $X ˉ$ , and a different sample has a different $X ˉ$ . The whole of statistical inference is the game of using the random statistic to say something about the fixed parameter.

Other pairs follow the same pattern: $σ$ (population standard deviation, a parameter) versus $s$ (sample standard deviation, a statistic); $p$ (population proportion) versus $p^$ (sample proportion). Once the pair is recognized, the constant-versus-random-variable rule settles which is which.

2.4.2 Why the Sample Mean Varies from Sample to Sample

To get clarity on why a statistic is a random variable, take the same example with different sets. Say the class has 445 students and the entire class average is about 70. Now take a set of 20 students and compute their average: you get 70.2. Take another different set of 20 students: the average is 60.4. Another: 78.4, and so on. At each and every set of sample values the average is going to vary. That is exactly the point: a statistic is a random variable, whereas a parameter is a constant. And because $X ˉ$ is a random variable, it must have a probability distribution — that is what we call the sampling distribution, the topic that follows.

Worked mini-example — a tiny population. Take a population of five values: ${2, 4, 6, 8, 10}$ . The population mean is the constant $μ = (2 + 4 + 6 + 8 + 10) /5 = 6$ . Now draw every possible sample of size 2 and average each one:

Sample	$X ˉ$	Sample	$X ˉ$
(2, 4)	3	(4, 10)	7
(2, 6)	4	(6, 8)	7
(2, 8)	5	(6, 10)	8
(2, 10)	6	(8, 10)	9
(4, 6)	5	(4, 8)	6

The sample means take values 3, 4, 5, 6, 7, 8, 9 — they vary all over the place even though the population mean never moves from 6. Some samples hit the target exactly ( $X ˉ = 6$ ), most land near it, and a few land far from it. This is the same phenomenon as the professor's 70.2, 60.4, 78.4 averages: sampling is a random draw, so the statistic it produces is a random variable. Sense-check: the sample means cluster around 6 — the center of the population — which is the first hint of what the sampling distribution of $X ˉ$ will turn out to be.

2.4.3 Student Questions and Answers

Q: When we move from the population level to the sample level, are we bound to make errors?

A: Yes. It is always understood that once you come down from population level to sample level, we are bound to commit mistakes; we are bound to have some errors. So instead of eliminating the errors, we try to minimize them — that is the whole scenario. The errors come from the fact that a sample is only part of the population: no sample, however large, is a perfect mirror of the whole. The practical response is not to seek a perfect sample (there is none) but to design samples and estimators that keep the error small and measurable — which is precisely what the sampling distribution in section 2.6 will let us quantify.

2.5 Statistical Inference: Estimation

2.5.1 The Two Branches of Statistical Inference

Statistical inference is broadly classified into two areas: the theory of estimation, and testing of hypothesis. Statistical inference always starts with two major problems. The first: estimating the parameters — when you do not know anything about the population parameter, what are you going to do? The theory of estimation is the area where you try to find the best possible replacement for the population parameter. Testing of hypothesis is the second major problem and comes in the upcoming sessions.

The situation that forces estimation: we want a number such as the mean spending of every customer in a country, but we can only ever see a handful of customers. The population mean $μ$ is real and fixed, yet invisible; the only raw material we have is the sample, and the sample mean $X ˉ$ is the natural candidate to stand in for $μ$ . Estimation is the branch of inference that turns this candidate into a defensible claim: it builds a rule — called an estimator — that turns sample data into a value (or a range) for the unknown parameter, and it tells us how good that value is likely to be. Hypothesis testing, coming later in the course, answers a different question: not "what is the parameter?" but "is the parameter this value or not?"

Estimation and testing are not two unrelated procedures; they are two questions asked about the same unknown $μ$ . Estimation says "give me your best guess." Testing says "decide between two competing claims about $μ$ ." Both rest on the same foundation — the sampling distribution of the estimator — which is why the course reviews it here before either branch goes further.

2.5.2 Point Estimation and Interval Estimation

Suppose the population mean $μ$ is unknown. It can be replaced using the sample values in two ways: with a single number — that is point estimation; or over an interval — that is interval estimation. Among the median, mode, and arithmetic mean, the sample mean $X ˉ$ is the best replacement — the sample average gives the best estimate — a point we will rely on. The textbook chapters 6, 7, 8 and 9 take care of all these things — sampling distribution, estimation, hypothesis testing, and p-values — and after this session you can look into them to some extent.

	Point estimation	Interval estimation
Output	one single number, e.g., $X ˉ = 62, 400$	a range of plausible values, e.g., $(61, 500, 63, 300)$
Question answered	"What is your best single guess at $μ$ ?"	"How wrong might that guess be?"
Strength	simple, easy to report	carries information about uncertainty
Weakness	no information about how far it may miss $μ$	wider statements require more machinery (covered in the interval-estimation chapters)

The arithmetic mean wins the "best replacement" title among the three averages for a reason that the coming sections make precise: among common estimators of $μ$ , the sample mean has the smallest spread around the true value — its sampling distribution is the tightest around $μ$ . The median and the mode are legitimate estimators in other settings, but for a symmetric, bell-shaped population the mean is the one to use, and it is the one the course builds on.

Exam note: the roadmap for the next chapters — chapters 6, 7, 8 and 9 cover sampling distribution, estimation, hypothesis testing, and p-values. After this session you should be able to look into them: the sampling distribution of $X ˉ$ is the engine, and both estimation and testing are applications of it.

2.6 The Sampling Distribution of the Sample Mean

2.6.1 The Mean and Variance of the Sample Mean

We anticipated that $X ˉ$ is a random variable; once it is a random variable, it must have a probability distribution — that is the sampling distribution. The sampling distribution of the sample mean is governed by its own two parameters:

$E (X ˉ) = μ$

$Var (X ˉ) = \frac{σ ^{2}}{n}$

where $n$ is the sample size.

Where $Var (X ˉ) = σ^{2} / n$ comes from — the full derivation. Let $X_{1}, X_{2}, \dots, X_{n}$ be the $n$ observations of the sample, each drawn independently from the same population with mean $μ$ and variance $σ^{2}$ . The sample mean is the average

$X ˉ = \frac{X _{1} + X _{2} + \dots + X _{n}}{n}$

Step 1 — expectation. Expectation is linear, so

$E (X ˉ) = \frac{1}{n} (E (X_{1}) + E (X_{2}) + \dots + E (X_{n})) = \frac{1}{n} (n μ) = μ$

Every observation carries the population mean $μ$ , and there are $n$ of them, so the $n$ 's cancel. The sampling distribution is centered exactly on the population mean.

Step 2 — variance. Variance does not behave linearly: a constant factor comes out squared, and the independent pieces add:

$Var (X ˉ) = Var (\frac{X _{1} + X _{2} + \dots + X _{n}}{n}) = \frac{1}{n ^{2}} Var (X_{1} + X_{2} + \dots + X_{n})$

Because the observations are independent, the variance of the sum is the sum of the variances:

$Var (X_{1} + X_{2} + \dots + X_{n}) = Var (X_{1}) + Var (X_{2}) + \dots + Var (X_{n}) = n σ^{2}$

Combining the two lines:

$Var (X ˉ) = \frac{n σ ^{2}}{n ^{2}} = \frac{σ ^{2}}{n}$

Step 3 — the square root. Standard deviation is the square root of variance, so the spread of the sampling distribution is

$σ_{X ˉ} = Var (X ˉ) = \frac{σ}{n}$

This quantity has its own name: the standard error of the mean. The professor's two phrasings — "variance of x bar equals sigma square by n" and "variance is sigma by root n" — are both correct, but they describe different quantities: $σ^{2} / n$ is the variance of $X ˉ$ , and $σ / n$ is the standard deviation (standard error) of $X ˉ$ . Mixing the two is one of the most common slips in this topic; if the quantity is squared, use $σ^{2} / n$ ; if it is not, use $σ / n$ .

The key property to remember: if $X$ follows a normal distribution with mean $μ$ and standard deviation $σ$ , then $X ˉ$ follows a normal distribution with mean $μ$ and variance $σ^{2} / n$ :

$X ˉ \sim N (μ, \frac{σ ^{2}}{n})$

This is the difference you observe: the sample mean keeps the population center but its spread shrinks — the standard deviation of $X ˉ$ is $σ / n$ , smaller than the population's $σ$ by a factor of $n$ .

Scope: the formula $Var (X ˉ) = σ^{2} / n$ assumes independent observations drawn from a population with variance $σ^{2}$ . Two standard limits: (1) sampling without replacement from a small finite population — the correction factor $(N - n) / (N - 1)$ multiplies the variance, where $N$ is the population size; in the tiny five-value population from section 2.4.2, $σ^{2} = 8$ , so samples of size 2 taken without replacement have variance $8/2 \times (5 - 2) / (5 - 1) = 3$ , matching the enumerated sample means exactly. When the sample is at most 5% of the population, the correction is negligible and $σ^{2} / n$ is used. (2) If the population variance $σ^{2}$ is unknown — the usual case — it must be estimated from the sample, and the estimated standard error $σ^/ n$ is what later chapters work with.

2.6.2 Standardizing the Sample Mean

In the same sense that you converted a single normal random variable $X$ into $Z = (X - μ) / σ$ , you standardize the sample mean:

$Z = \frac{X ˉ - μ}{σ / n}$

The change you should observe: for a single observation, divide by $σ$ ; when you are talking about the sample mean, divide by $σ / n$ . Everything else stays the same — the mean stays $μ$ , the shape is still normal.

The pattern to keep straight: a single value $x$ is one draw from the population, and its natural distance scale is one population standard deviation $σ$ . The sample mean $X ˉ$ is a different animal — it is an average of $n$ draws, and its distance scale is the standard error $σ / n$ . Dividing by the wrong one changes every z-score you compute. Remember it as: single observation $\to$ divide by $σ$ ; sample mean $\to$ divide by $σ / n$ . The $n$ in the denominator is the only difference — and it is the whole reason bigger samples give sharper statements about $μ$ .

	Single observation $X$	Sample mean $X ˉ$
Mean	$μ$	$μ$
Standard deviation	$σ$	$σ / n$ (standard error)
Standardization	$Z = \frac{X - μ}{σ}$	$Z = \frac{X ˉ - μ}{σ / n}$
Shrinks with $n$ ?	no	yes — spread $\to 0$ as $n \to \infty$

The last row is the practical payoff: averaging shrinks randomness. The more observations go into $X ˉ$ , the tighter the sampling distribution around $μ$ , and the more confident a statement like " $X ˉ$ is close to $μ$ " becomes.

2.6.3 The Central Limit Theorem

Two cases. If the population itself is normal, then $X ˉ$ will follow a normal distribution no matter the sample size — it is immaterial whether the sample size is small or large; any size works. But look at the other diagram: $X$ is not normal at all — it is some bimodal distribution. If $X$ behaves like that and is not normal, and you still want to use the normal distribution, then the only requirement is that you maintain the sample size large enough — that is what we generally call the central limit theorem. Irrespective of whatever your parent distribution is — binomial, Poisson, any theoretical distribution — $X ˉ$ will behave like a normal distribution provided $n$ is large enough.

Central limit theorem (CLT). In selecting random samples of size $n$ from a population, the sampling distribution of the sample mean $X ˉ$ can be approximated by a normal distribution as the sample size becomes large. Two consequences:

Normal population: the sampling distribution of $X ˉ$ is exactly normal for any sample size — small, large, it makes no difference.
Non-normal population: the sampling distribution of $X ˉ$ is roughly normal when $n$ is sufficiently large, with the same mean $μ$ and standard error $σ / n$ as always.

General practice treats $n \geq 30$ as enough for most populations; heavily skewed populations may need $n \geq 50$ . For binomial counts and other discrete data, the required $n$ depends on how near the probability is to 0 or 1 — a topic the course returns to with proportions.

The picture the professor drew is worth keeping. Imagine three very non-normal parent distributions: a flat uniform one, a symmetric "rabbit-eared" one with probability piled in the tails, and a skewed one with a long right tail. Take samples of size 2 from each: the sampling distributions still look like their parents. Take samples of size 5: they start to round toward a bell. Take samples of size 30: all three sampling distributions are roughly normal — the same bell, at the same center $μ$ , with the same shrinking spread $σ / n$ . The parent distribution leaves its fingerprints on the shape of small-sample results, but the CLT washes them out as $n$ grows.

Worked mini-example — the EAI salary survey. A population of 2,500 managers has salary standard deviation $σ = $4, 000$ . With a sample of size $n = 30$ , the standard error of the sample mean is

$\frac{σ}{n} = \frac{4000}{30} \approx 730.3$

so a typical sample mean lands within about $\pm $730$ of the true mean salary. Increase the sample to $n = 100$ :

$\frac{σ}{n} = \frac{4000}{100} = 400$

and the typical error drops to $\pm $400$ . Doubling precision required four times the sample size — the $n$ in the denominator is why: to halve the standard error you must quadruple $n$ . Sense-check: larger samples are more informative, but the return shrinks as the square root — the first 30 observations help more than the next 30.

Scope: the CLT guarantees approximate normality of $X ˉ$ in the center of the distribution for large $n$ — it does not guarantee that extreme tails match the normal curve perfectly, and it does not apply to other statistics (a single extreme observation in a sample of 30 from a skewed population still appears as an outlier in the data; the sampling distribution of the mean is normal, the data themselves are not). For tiny samples ( $n < 30$ ) from strongly skewed populations, the normal approximation can be poor, and the standard-error statements of this section need the sample-size caveat.

2.6.4 Student Questions and Answers

Q: Since $X ˉ$ is a statistic and so a random variable, does it follow a normal distribution?

A: Yes. If the population itself is normal, the sample mean follows a normal distribution for any sample size — small or large, it is immaterial. If the population is not normal, the sample mean is still roughly normal when the sample size is large enough; that is the central limit theorem, and it works for binomial, Poisson, or any other parent distribution. In both cases the sampling distribution is centered at the population mean $μ$ , and its spread is the standard error $σ / n$ .

Exam Guidance Summary

The table and standardization skills from this session are direct exam material. Collecting the guidance given throughout the lecture:

Read the standard normal table fluently: values like $P (Z \leq 1.5) = 0.9332$ , $P (Z \leq 1) = 0.8413$ , and interval values such as $P (1 \leq Z \leq 1.5) = 0.0919$ — "in the exam point of view also, you should know how to retrieve the probability values."
Table reading helps with quizzes too, not only the exam.
Know the standardization step $Z = (X - μ) / σ$ and the complement rule for right-side questions.
Know the outlier rule: a z-score below $- 3$ or above $+ 3$ flags an outlier, plus the empirical rule 68-95-99.7 for bell-shaped data.
Know the sampling-distribution results: $E (X ˉ) = μ$ , $Var (X ˉ) = σ^{2} / n$ , the standard error $σ / n$ , and the standardization $Z = (X ˉ - μ) / (σ / n)$ .
The normal density formula is reference only — no need to memorize it.
Chapter 3 has the detailed outlier discussion; chapters 6, 7, 8 and 9 cover sampling distribution, estimation, hypothesis testing, and p-values.
Remember the problem-reference convention chapter.page.problem (for example, 3.132.42) from the 13th edition.

A compact revision routine for exam week: for any normal problem, (1) name the random variable and its two parameters; (2) convert the question into a probability statement; (3) standardize; (4) decide which side or strip is wanted; (5) read or subtract or complement; (6) sense-check the answer against 0.5 and the direction of the question.

Key Industry Applications

Each concept in this session maps onto a concrete industry use. Where a problem from the lecture appears, its domain placement is given alongside.

Manufacturing — assembly-time modeling. The machinery assembly problem: process times as normal random variables with a stated mean and spread let you compute probabilities like "at least 11.5 minutes," which feed into production planning. A plant that knows assembly takes about 12.9 minutes on average with a 2-minute spread can compute how often a job misses an 11.5-minute target — and schedule capacity and delivery promises around that probability. The same pattern applies to order-fulfillment times, machine cycle times, and service call durations across operations.
Quality control — z-score monitoring. Z-score monitoring flags abnormal machine behavior that restores itself; an outlier that is not a data-entry mistake signals an anomaly that needs intervention. Statistical process control in manufacturing runs exactly this logic — when a measurement strays beyond three sigma from the process mean, the process is stopped and investigated even if the next measurement looks fine again. The same reasoning drives anomaly detection in sensor networks, network traffic, and bank transactions.
Pricing — California backyard structures. With mean price 3,100 dollars and standard deviation 1,200 dollars, the z-score tells whether a quoted price is typical (2,300 and 4,900) or statistically implausible (13,000). Real-estate and procurement analysts run the same check on market data: a price more than three standard deviations from the market mean is investigated rather than accepted — it may be a data error, a different product class, or a misquote.
Education and survey statistics — the teacher's-age example. One extreme value can drag an average, so check for outliers before reporting summaries. Survey research, census work, and any reporting of means (average income, average sales, average test scores) applies this rule before publishing — a single miscoded figure can change a headline number and the decision built on it.
Pharma — comparing an old drug and a new drug on location and spread. As recapped from the previous session, anticipating efficacy means comparing both parameters — where the distributions sit and how widely they spread. Clinical development compares treatment and placebo groups on mean response and variability exactly this way, and the sampling-distribution machinery from section 2.6 is what turns a sample of trial patients into a statement about the whole patient population.

ASM Lecture 2 notes · Normal Distribution and Statistical Inference

Advanced Statistical Methods· postgraduate· 2026-08-11

Sections Breakdown

12.1 The Normal Distribution (Gaussian)

What the normal distribution is, its two parameters, the density function and its structure, families of normal curves, and the machinery assembly worked example.

22.2 The Standard Normal Distribution and Z-Scores

Standardizing variables into z-scores, cumulative probabilities, the complement and interval rules, and table-reading practice.

32.3 Outlier Detection with Z-Scores

Why outliers matter, the box-plot and z-score detection rules, the empirical rule, and the California backyard structures worked example.

42.4 Population, Sample, Parameter, and Statistic

Parameter versus statistic, why the sample mean varies from sample to sample, and the unavoidable error of sampling.

52.5 Statistical Inference: Estimation

The two branches of statistical inference and the contrast between point estimation and interval estimation.

62.6 The Sampling Distribution of the Sample Mean

Mean and variance of the sample mean, the standard error, standardizing the sample mean, and the central limit theorem.

7Exam Guidance Summary

The professor's exam guidance: fluent table reading, standardization, the complement rule, the outlier rule, and the sampling-distribution formulas.

8Key Industry Applications

Industry placements of the session's concepts: manufacturing assembly-time modeling, quality control, pricing, survey statistics, and pharmaceutical trials.

Postgraduate students in Advanced Statistical Methods

Exam Revision Notes

Below is the distilled, exam-ready core. Every entry comes from the full explanation above. Use this section for rapid review; return to the main notes when a point needs more context.

The Normal Distribution (Gaussian)

Must-know: The normal distribution is a symmetric bell-shaped continuous distribution with parameters mu (location) and sigma (spread); probabilities come from standardizing to Z = (X - mu)/sigma and reading the cumulative table; the density formula is reference only, not for memorization.

$f (x) = \frac{1}{σ 2 π} e^{- \frac{1}{2} (\frac{x - μ}{σ})^{2}}$

⚠️ Top pitfall: Confusing the density height f(x) with a probability: for a continuous variable the probability at a single point is zero, and only areas (ranges like P(X >= 11.5)) carry probability.

Self-check: For the machinery assembly problem (mu = 12.9, sigma = 2), what is P(X >= 11.5)? Answer: P(Z >= -0.7) = 1 - 0.2420 = 0.7580.

Connects to: 2.2 The Standard Normal Distribution and Z-Scores, 2.6 The Sampling Distribution of the Sample Mean.

The Standard Normal Distribution and Z-Scores

Must-know: Read cumulative values like P(Z <= 1.5) = 0.9332 and P(Z <= 1) = 0.8413; strips subtract (P(1 <= Z <= 1.5) = 0.0919); right tails use the complement P(Z >= a) = 1 - P(Z <= a); P(0 <= Z <= 2.5) = 0.9938 - 0.5000 = 0.4938.

$Z = \frac{X - μ}{σ}$

⚠️ Top pitfall: Taking a table entry directly for a right-tail question: P(Z >= a) needs the complement 1 - P(Z <= a); a strip answer must be smaller than the 0.5 half-area landmark.

Self-check: What is P(0 <= Z <= 2.5)? Answer: 0.9938 - 0.5000 = 0.4938 (just under the 0.5 half, because the tail beyond 2.5 is excluded).

Connects to: 2.1 The Normal Distribution (Gaussian), 2.3 Outlier Detection with Z-Scores.

Outlier Detection with Z-Scores

Must-know: Outlier rule: z-score below -3 or above +3 flags an outlier; empirical rule 68-95-99.7 for bell-shaped data; California prices (mu = 3100, sigma = 1200): 2300 gives z = -0.67, 4900 gives z = 1.5 (not an outlier), 13000 gives z = 8.25 (an outlier).

$Z = \frac{X - μ}{σ}, Z < - 3 or Z > 3 \Rightarrow outlier$

⚠️ Top pitfall: Computing the mean and standard deviation including the suspect value lets a large outlier inflate both, shrink other z-scores, and hide itself; detection should be separate from estimation.

Self-check: Is the 4,900 price an outlier for mu = 3100 and sigma = 1200? Answer: no — z = 1.5 lies between -3 and +3.

Connects to: 2.2 The Standard Normal Distribution and Z-Scores, 2.1 The Normal Distribution (Gaussian).

Population, Sample, Parameter, and Statistic

Must-know: Parameter = constant characteristic of the population; statistic = random variable computed from the sample. Classic pair: mu (parameter) versus X-bar (statistic). Moving from population to sample always carries errors — we minimize them, never eliminate them.

⚠️ Top pitfall: Treating X-bar as if it had a single fixed value: every different sample produces a different sample mean, which is exactly why X-bar needs its own probability distribution (the sampling distribution).

Self-check: In the tiny population {2, 4, 6, 8, 10}, does every sample of size 2 give the same mean? Answer: no — sample means range from 3 to 9 while the population mean stays 6.

Connects to: 2.6 The Sampling Distribution of the Sample Mean, 2.5 Statistical Inference: Estimation.

Statistical Inference: Estimation

Must-know: Inference has two branches: theory of estimation and testing of hypothesis. Estimation replaces an unknown parameter with a single number (point estimation) or a range (interval estimation). Among median, mode and arithmetic mean, the sample mean X-bar is the best replacement for mu.

⚠️ Top pitfall: Reporting a point estimate without any idea of how far it may miss the parameter: point estimates carry no uncertainty information, which is what interval estimation and the sampling distribution add.

Self-check: What are the two branches of statistical inference? Answer: theory of estimation and testing of hypothesis.

Connects to: 2.4 Population, Sample, Parameter, and Statistic, 2.6 The Sampling Distribution of the Sample Mean.

The Sampling Distribution of the Sample Mean

Must-know: E(X-bar) = mu, Var(X-bar) = sigma^2/n, standard error = sigma/sqrt(n), X-bar ~ N(mu, sigma^2/n) for normal populations; CLT: non-normal parent distributions still give an approximately normal X-bar when n is large; standardize with Z = (X-bar - mu)/(sigma/sqrt(n)).

$Z = \frac{X ˉ - μ}{σ / n}$

⚠️ Top pitfall: Conflating variance sigma^2/n with standard deviation sigma/sqrt(n) — the two phrasings both appeared in the lecture; also dividing by sigma instead of sigma/sqrt(n) when standardizing the sample mean.

Self-check: If sigma = 4000 and n = 30, what is the standard error of the sample mean? Answer: 4000/sqrt(30) ≈ 730.3; with n = 100 it drops to 400.

Connects to: 2.5 Statistical Inference: Estimation, 2.2 The Standard Normal Distribution and Z-Scores, 2.4 Population, Sample, Parameter, and Statistic.

Exam Guidance Summary

Must-know: Table reading, standardization Z = (X - mu)/sigma, complement rule, outlier rule -3/+3, empirical rule 68-95-99.7, and the sampling-distribution formulas E(X-bar) = mu, Var(X-bar) = sigma^2/n are the exam-critical skills.

$P (Z \geq a) = 1 - P (Z \leq a)$

⚠️ Top pitfall: Reading the table for the wrong side of the question — right-tail questions need the complement.

Self-check: What does 3.132.42 stand for? Answer: chapter 3, page 132, problem 42.

Connects to: 2.1 The Normal Distribution (Gaussian), 2.2 The Standard Normal Distribution and Z-Scores, 2.3 Outlier Detection with Z-Scores, 2.6 The Sampling Distribution of the Sample Mean.

Key Industry Applications

Must-know: Each concept maps to an industry use: normal process-time modeling feeds production planning; z-score monitoring flags machine anomalies needing intervention; z-scores flag statistically implausible prices; outlier checks protect reported averages; pharma compares treatment distributions on location and spread.

⚠️ Top pitfall: Reporting averages or acting on single extreme values without an outlier check first.

Self-check: Why would a 13,000-dollar backyard-structure quote be investigated? Answer: its z-score of 8.25 far exceeds the +3 outlier boundary for a market centered at 3,100 with spread 1,200.

Connects to: 2.3 Outlier Detection with Z-Scores, 2.1 The Normal Distribution (Gaussian).

Previous LectureLecture 01 · Introduction to Advanced Statistical Methods

Next LectureLecture 03 · Statistical Inference and Hypothesis Testing

Was this lecture useful?

Loading comments…

🤖

BitsNotes AI Assistant

Subject Notes Assistant

Configure AI Chat

Choose how to access the chatbot

Have your own API key?

Switch to "Bring Your Own Key" tab above for unlimited access with any OpenAI-compatible provider.

AI Provider Preset

API Endpoint URL (OpenAI-Compatible)

API Key

Model Name

🔑 Enter API key above to fetch live models from provider, or enter model name manually.

Remember key for this session (until tab closes)

OpenAI-Compatible API Support

Choose any provider preset (Gemini, DeepSeek, Kimi, GLM, MiniMax, Qwen, OpenAI, Groq, Ollama, etc.) or enter a custom endpoint URL.

Security & Privacy First

Your API key is sent directly from your browser to your specified provider. BitsNotes servers never store or see your key.

2.1 The Normal Distribution (Gaussian)

2.1.1 What the Normal Distribution Is

2.1.2 The Probability Functions Behind the Normal Curve

2.1.3 Families of Normal Curves

2.1.4 Worked Example: Probability That Assembly Takes at Least 11.5 Minutes

2.1.5 Student Questions and Answers

2.2 The Standard Normal Distribution and Z-Scores

2.2.1 Standardizing the Variable

2.2.2 Cumulative Probabilities and the Complement Rule

2.2.3 Worked Examples: Reading Table Values

2.2.4 Student Questions and Answers

2.3 Outlier Detection with Z-Scores

2.3.1 What Outliers Are

2.3.2 Detecting Outliers: Box Plots and Z-Scores

2.3.3 The Empirical Rule

2.3.4 Worked Example: California Backyard Structures

2.3.5 Student Questions and Answers

2.4 Population, Sample, Parameter, and Statistic

2.4.1 Parameter versus Statistic

2.4.2 Why the Sample Mean Varies from Sample to Sample

2.4.3 Student Questions and Answers

2.5 Statistical Inference: Estimation

2.5.1 The Two Branches of Statistical Inference

2.5.2 Point Estimation and Interval Estimation

2.6 The Sampling Distribution of the Sample Mean

2.6.1 The Mean and Variance of the Sample Mean

2.6.2 Standardizing the Sample Mean

2.6.3 The Central Limit Theorem

2.6.4 Student Questions and Answers

Exam Guidance Summary

Key Industry Applications

ASM Lecture 2 notes · Normal Distribution and Statistical Inference

Sections Breakdown

Exam Revision Notes

The Normal Distribution (Gaussian)

The Standard Normal Distribution and Z-Scores

Outlier Detection with Z-Scores

Population, Sample, Parameter, and Statistic

Statistical Inference: Estimation

The Sampling Distribution of the Sample Mean

Exam Guidance Summary

Key Industry Applications

Was this lecture useful?

BitsNotes AI Assistant

Configure AI Chat

Sign in to use AI Chat