Introduction to Statistical Methods

Sampling, Sampling Distributions, and Estimation

📅 Published: 2026-07-06

🎓 Level: postgraduate

👥 Audience: Postgraduate students in statistics, data science, and machine learning

Prerequisite Knowledge

This lecture builds on the following concepts from earlier lectures. If any feel unfamiliar, review the linked notes before proceeding.

Previously Covered in This Subject

Measures of Central Tendency (Mean) — covered in Lecture 1
Variance and Standard Deviation — covered in Lecture 2
Population vs. Sample Notation — covered in Lecture 2
The Normal Distribution — covered in Lecture 8
Z-Transformation and Standard Normal Table — covered in Lecture 8
Random Variables and Probability Distributions — covered in Lectures 5-8

Sampling, Sampling Distributions, and Estimation

You already do statistics every day without knowing it. When you taste a spoonful of dal to decide if the whole pot needs salt, you are sampling. When you split your dataset 80-20 for training and testing, you are sampling. This lecture formalizes that instinct — it gives you the mathematical tools to say how good your estimate is and how sure you can be about it.

This lecture builds the bridge between describing data (what you did in earlier sessions with means, variances, and histograms) and inferring something about the world from it. The journey goes like this. You start with a population you cannot fully measure. You draw a sample, then compute a statistic from it. Using the sampling distribution and the Central Limit Theorem, you reason backward to say something about the population. Along the way, you will learn when to stratify, how to build a confidence interval, and why the bell curve shows up everywhere in statistics.

By the end of this lecture, you should be able to look at any sample mean and answer the two questions that matter: How precise is this estimate? and How confident should I be?

10.1 Population and Sample — The Core Idea

10.1.1 Definition and Intuition

Hook. You have a dataset of 10,000 customer reviews and you want to build a sentiment classifier. You don't train on all 10,000 — you set aside 2,000 for testing. Why? Because the only way to know if your model actually learned something is to test it on data it has never seen. That split is sampling. And the entire field of inferential statistics is built on this one move.

Intuition. Think of a population as the complete universe of things you care about — all customers, all patients, all manufactured bolts from a factory. A sample is a subset of that universe that you actually measure. You sample because measuring the whole population is expensive, time-consuming, or physically impossible.

The blood test analogy. A lab draws 10 mL of your blood. The report doesn't describe just those 10 mL — it describes your entire blood profile. The lab sampled, analyzed, and generalized. Your ~5 liters of blood is the population. The vial is the sample. The same logic applies to every train-test split you write.

The cooking analogy. Before tasting, your mom stirs the pot — she makes the population homogeneous. Then one spoonful tells her about the whole dish. Stirring is key: if the salt sits at the bottom, one spoonful from the top tells you nothing. This is why random (or stratified) sampling matters — we will get to that.

Where the analogies break. In the blood test, the 10 mL is physically identical in composition to the rest of your blood (it's homogeneous). Real-world data populations are rarely that uniform. That is why sampling method matters — a bad sample gives bad inferences.

Formal definition. A population is the complete set of all items or individuals under study. It has size $N$ (finite) or is treated as infinite. A sample is any subset of the population, of size $n$ , drawn for analysis.

Population mean: $μ = \frac{1}{N} \sum_{i = 1}^{N} x_{i}$ (a parameter — fixed, usually unknown)
Sample mean: $x ˉ = \frac{1}{n} \sum_{i = 1}^{n} x_{i}$ (a statistic — computed from data, varies across samples)

The population is what you want to know about. The sample is what you actually see. Statistical inference is the process of reasoning from the sample back to the population.

All of statistics — particularly inferential statistics — is built on this movement: from sample analysis back to population understanding. You compute a statistic from the sample and use it to estimate the corresponding population parameter.

10.1.2 The 80-20 Split as Sampling

Worked example — train-test split as sampling. You have a dataset of 1,000 labeled emails (700 spam, 300 not-spam). You run:

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

Here, the full 1,000 emails is your population. The 800 emails in X_train form one sample. The 200 in X_test form another. Your model learns patterns from the training sample. You evaluate on the test sample to estimate how well those patterns generalize. If your training accuracy is 98% but test accuracy is 72%, your sample-based inference was wrong — the training sample did not represent the population well enough. This is exactly the risk sampling always carries.

Scope: when the population-sample framework applies. This framework assumes the sample is drawn from the same population you want to make claims about. If your training data is all indoor photos and your test data is all outdoor photos, you have violated this assumption. The population changed between sampling events. The framework also assumes observations within the sample are independent — the blood test vial doesn't affect the next vial drawn. In time-series data, this independence often fails.

Visual intuition. Picture a large circle (the population of $N$ items). Inside it, draw a smaller circle (your sample of $n$ items). The arrow of inference points from the small circle back to the large one. The quality of that inference depends entirely on how the small circle was chosen. If you picked items only from the top-left corner of the large circle, your arrow points to the wrong place.

Pitfalls. 1. Confusing sample with population. Saying "the average is 72" without specifying whether it is $x ˉ$ or $μ$ is a common sloppiness. Train yourself: if you computed it from data, it's a statistic. If it's the true unknown value, it's a parameter. 2. Assuming bigger samples are always representative. A large biased sample is worse than a small random one. A survey of 10,000 people from one city tells you nothing about the whole country. 3. Ignoring how the sample was collected. The math of sampling distributions (coming up) assumes random sampling. If your sample was collected haphazardly, the formulas in this lecture do not apply.

Recap. A population is the whole you care about; a sample is the part you measure. Every statistic you compute from a sample is a guess about a population parameter — and the quality of that guess depends on how the sample was drawn. Next: we name these two families of quantities — parameters and statistics — with their own Greek and Roman notation.

Real-world & domain connection. The population-sample framework underpins every ML pipeline. Your training set is a sample; your test set is a sample; every mini-batch in SGD is a sample. When a self-driving car company reports "99.9% detection accuracy," they are extrapolating from a test sample to the population of all possible driving scenarios. Whether that extrapolation is valid depends on whether their test sample actually represents all road conditions — night, rain, snow, construction zones. Sampling theory is what lets you quantify the gap between the number you measured and the number you actually care about.

10.2 Parameters and Statistics — Two Vocabularies

10.2.1 Definition

Hook. You run df.describe() and get a mean of 72.5. Is that number a fact about your dataset, or a fact about the world? The answer changes everything — and statistics has two different names for these two kinds of numbers.

Intuition. A parameter is the truth — the actual average height of every human alive, which nobody can measure exactly. A statistic is your best guess — the average height of the 500 people you actually measured. Parameters live in the population and wear Greek letters. Statistics live in the sample and wear Roman letters. The whole game of inferential statistics is: "Given my statistic, what can I say about the parameter?"

Formal definitions. A parameter is any numerical characteristic of a population — it is fixed (though usually unknown). A statistic is any numerical characteristic computed from a sample — it is a random variable because different samples give different values.

Concept	Population (parameter)	Sample (statistic)
Mean	$μ$ (mu)	$x ˉ$ (x-bar)
Variance	$σ^{2}$ (sigma squared)	$s^{2}$
Standard deviation	$σ$	$s$
Proportion	$P$ (capital P)	$p^$ (p-hat)

The notation is not arbitrary. When a problem gives you $μ = 100$ , it is telling you about the whole population. When it gives you $x ˉ = 98.5$ , it is telling you about one particular sample. Never mix them up in an exam answer.

10.2.2 The N vs. N−1 Question

Why divide by $n - 1$ for sample variance? The sample variance formula is:

s^{2} = \frac{\sum _{i = 1}^{n} ( x _{i} - x ˉ ) ^{2}}{n - 1}

not $n$ in the denominator. The reason traces back to the parameter-statistic distinction.

When you compute $\sum (x_{i} - x ˉ)^{2}$ , you are measuring deviations from the sample mean $x ˉ$ , not from the true population mean $μ$ . And $x ˉ$ is always, by definition, the value that minimizes the sum of squared deviations for that particular sample. So $\sum (x_{i} - x ˉ)^{2}$ is always slightly smaller than $\sum (x_{i} - μ)^{2}$ would be. Dividing by $n$ would systematically underestimate $σ^{2}$ . Dividing by $n - 1$ (Bessel's correction) compensates for this bias and makes $s^{2}$ an unbiased estimator — meaning if you repeated the sampling infinitely many times, the average of all your $s^{2}$ values would equal $σ^{2}$ .

Worked illustration. Take a tiny population: $2, 4, 6, 8$ . The true population variance $σ^{2}$ (using $N = 4$ in denominator) = 5.0. Now draw all possible samples of size $n = 2$ without replacement (there are 6 such samples). Compute the variance for each sample: - Sample $2, 4$ : $x ˉ = 3$ , using $n$ (2) → $s^{2} = 1.0$ ; using $n - 1$ (1) → $s^{2} = 2.0$ - Sample $2, 6$ : $x ˉ = 4$ , using $n$ → $s^{2} = 4.0$ ; using $n - 1$ → $s^{2} = 8.0$ - ... and so on for all 6 samples.

Average of the six $n$ -denominator variances: 2.5 (underestimates 5.0). Average of the six $n - 1$ -denominator variances: 5.0 (exactly right). This is Bessel's correction at work.

Scope. Bessel's correction ( $n - 1$ ) is specifically for estimating $σ^{2}$ from a sample. When you are simply describing a dataset (not inferring a population parameter), dividing by $n$ is fine — that is what df.var(ddof=0) gives you. The correction matters only when your goal is inference.

Visual intuition. Imagine firing arrows at a target. $μ$ is the bullseye. Each sample mean $x ˉ$ is an arrow hole somewhere on the target. The sample variance $s^{2}$ (with $n - 1$ ) measures the spread of your arrows. If you used $n$ instead, your spread estimate would shrink — making you overconfident about how close your arrows cluster.

Pitfalls. 1. Using $n - 1$ blindly for everything. The correction applies to variance, not to the mean or proportion formulas. Do not subtract 1 from $n$ when computing $x ˉ$ . 2. Forgetting that as $n$ grows, the difference between $n$ and $n - 1$ shrinks. With $n = 10, 000$ , the correction is negligible. With $n = 5$ , it matters a lot. 3. Confusing descriptive and inferential contexts. Pandas .describe() uses $n - 1$ by default (ddof=1) because it assumes you are doing inference. If you want the actual spread of your specific dataset, set ddof=0.

Recap. Parameters (Greek) are population truth; statistics (Roman) are sample estimates. The $n - 1$ in sample variance is not a quirk — it corrects for the fact that $x ˉ$ is already fitted to your sample. Next: how you draw the sample determines whether your statistic is any good.

Real-world & domain connection. The parameter-statistic distinction is why A/B testing platforms report "confidence intervals" rather than point estimates. When Google runs an experiment on search result ranking, the click-through rate they measure on 1% of users is $p^$ . The true effect on all users is $P$ . The entire experiment is one giant exercise in going from $p^$ back to $P$ , with the sample size determining how tight the bound can be.

10.3 Sampling Methods

Hook. You train a fraud detection model. It gets 99.9% accuracy. Impressive, right? Then you realize only 0.1% of transactions are fraud — your model just learned to say "not fraud" every time and was right 99.9% of the time. The problem was not the model. The problem was how you sampled.

10.3.1 Random Sampling

Intuition. Random sampling is like a lottery — every ticket (record) has the same chance of being picked. No favorites, no exclusions. This is the gold standard because it is the only sampling method that probability theory can fully describe. Every formula from this point forward in the lecture assumes random sampling unless stated otherwise.

Definition. In simple random sampling, every record in the population has an equal probability of being selected, and every subset of size $n$ is equally likely to be the sample. This is what train_test_split with random_state approximates.

For a finite population of size $N$ , a simple random sample of size $n$ means each of the $(n N)$ possible samples has probability $1/ (n N)$ of being drawn.

The random_state parameter controls the seed of the pseudorandom number generator. The value 42 is a pop-culture convention — you can use any integer. Change it and your train-test composition changes, which can shift your model's performance. This sensitivity is a reminder: your results depend partly on the split, not just on the model.

10.3.2 Stratified Sampling — Fixing the Class Imbalance Problem

The problem random sampling cannot solve. Suppose you have 1,000 medical records: 200 with diabetes (yes) and 800 without (no). A plain random 80-20 split could easily put 195 "yes" records in training and only 5 in testing. Your test set now has almost no positive cases — making your evaluation meaningless. Worse, the training set is dominated by "no" records, so the model learns to predict "no" and achieves high accuracy by doing nothing useful.

A stratum (plural: strata) is a homogeneous subgroup of the population. In stratified sampling, you divide the population into strata and then draw a random sample from each stratum in proportion to its size. For classification, the target variable $y$ defines the strata.

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, stratify=y, random_state=42
)

With stratify=y, each stratum gets its fair share in both train and test. If the full dataset is 20% "yes", then training gets ~20% "yes" and testing gets ~20% "yes". The selection within each stratum remains random — stratification and randomness work together.

Q: Can we use stratify=X instead of stratify=y? A: No. The imbalance we care about is in the target class. Homogeneity is defined in terms of the output $y$ for classification. The algorithm stratifies on $y$ .

Q: How do you stratify for regression when $y$ is continuous? A: For regression, class imbalance in the target is less of a concern since $y$ is continuous. If you need stratification (e.g., for demographic balance across age groups), the algorithm bins or clusters the continuous $y$ into homogeneous chunks and samples proportionately from each. For most ML classification work, stratification is about discrete target classes.

Q: Can stratification ever hurt? A: In ML, no — it prevents class-imbalance bias. In business, beware of purposive sampling — cherry-picking favorable examples to make results look good. That is not stratification; it is fraud.

10.3.3 K-Fold Cross-Validation as Sampling

K-fold cross-validation is sampling applied repeatedly. Split the data into $K$ folds. Train on $K - 1$ folds, test on the held-out fold. Rotate through all $K$ folds as the test set. Each fold is a sample. The cross-validation score — the average of $K$ test scores — is a more strong estimate of generalization than a single train-test split. It averages out the randomness of any one split.

Scope: when each method applies. - Random sampling: the default. Works when the population is about homogeneous with respect to what you are predicting. - Stratified sampling: essential when the target class distribution is skewed (fraud detection, rare disease diagnosis, churn prediction). Without it, your evaluation metrics are unreliable. - K-fold CV: use when you need a strong performance estimate and can afford $K$ rounds of training. Typical $K$ is 5 or 10. For imbalanced data, use StratifiedKFold.

Visual intuition. Picture a jar with 800 white marbles ("no diabetes") and 200 red marbles ("yes diabetes"). A random handful of 200 marbles might contain only 2 red ones. Stratified sampling is like first separating reds and whites into two jars, then taking 20% from each — you get exactly 160 white and 40 red. K-fold CV is like splitting the jar into 5 equal portions, using 4 for practice and 1 for the real test, cycling through all 5.

Pitfalls. 1. Using random sampling on imbalanced data. Your "99% accuracy" means nothing if 99% of records are one class. Always check class distribution before choosing a sampling strategy. 2. Setting random_state once and forgetting it. If your conclusions change when you change the seed, your model is not stable. Report mean ± std across multiple random splits. 3. Forgetting that stratification applies to the target, not the features. stratify=y, never stratify=X for classification. 4. Using too few folds in CV. $K = 3$ gives high-variance estimates. $K = 5$ or $10$ is standard. $K = n$ (leave-one-out) is computationally expensive and rarely worth it.

Recap. Random sampling is the foundation. Stratified sampling fixes class imbalance by ensuring each stratum gets proportional representation. K-fold cross-validation reduces the variance of your performance estimate by averaging across multiple splits. Pick the right method for your data, or your model's numbers will lie to you. Next: once you have a sample, what can you say about the statistic you computed from it?

Real-world & domain connection. Stratified sampling is mandatory in medical AI. A skin cancer classifier trained on a dataset where 99.5% of images are benign will "achieve" 99.5% accuracy by predicting "benign" for every image — and kill patients. The FDA requires stratified evaluation on balanced test sets for any AI-assisted diagnostic tool. In finance, fraud detection faces the same issue: 0.1% of transactions are fraudulent, and an unstratified evaluation will hide a useless model.

10.4 Sampling Distribution

10.4.1 Why Multiple Samples?

Hook. You flip a coin 10 times and get 7 heads. Is the coin biased? You can't tell from one set of 10 flips. But if you repeated the 10-flip experiment 1,000 times and always got 7 or more heads, you would be suspicious. The pattern across repeated samples is what lets you judge a single result — and that pattern has a name.

Intuition. A single sample gives you one $x ˉ$ . It is just one data point. To understand how much $x ˉ$ can vary just by chance, you need to see many $x ˉ$ values from many samples. The sampling distribution is the distribution of all possible values of a statistic (like $x ˉ$ or $p^$ ) across all possible samples of the same size drawn from the same population.

Think of it this way: the population is a fixed thing. Each sample is like taking a photograph of it from a slightly different angle. The sampling distribution is the album of all possible photographs — it tells you which views are common and which are rare.

10.4.2 Key Results

The sampling distribution of the mean — two fundamental results.

Let $X_{1}, X_{2}, \dots, X_{n}$ be a random sample from a population with mean $μ$ and standard deviation $σ$ . Define the sample mean:

X ˉ = \frac{1}{n} i = 1 \sum n X_{i}

Then:

μ_{x ˉ} = E [X ˉ] = μ

σ_{x ˉ} = Var (X ˉ) = \frac{σ}{n}

The first result says the sample mean is unbiased — on average, it hits the population mean. The second result says the variability of $x ˉ$ shrinks as $n$ . To cut the standard error in half, you need four times the sample size — not twice.

The quantity $σ_{x ˉ} = σ / n$ is called the standard error of the mean. It is the standard deviation of the sampling distribution, and it measures how much a typical $x ˉ$ deviates from $μ$ .

Worked example — standard error in action. A population has $μ = 100$ and $σ = 20$ . You draw a sample of size $n = 25$ . The standard error is:

σ_{x ˉ} = \frac{20}{25} = \frac{20}{5} = 4

Now increase to $n = 100$ :

σ_{x ˉ} = \frac{20}{100} = \frac{20}{10} = 2

Quadrupling the sample size halved the standard error. With $n = 25$ , about 95% of sample means fall within $100 \pm 2 \times 4 = [92, 108]$ . With $n = 100$ , that tightens to $100 \pm 2 \times 2 = [96, 104]$ . Larger $n$ → tighter spread around $μ$ → more precise estimates.

Scope. These results ( $μ_{x ˉ} = μ$ and $σ_{x ˉ} = σ / n$ ) hold for any population with finite variance, regardless of the population's shape, as long as the sample is random. They do not require normality. They are exact mathematical consequences of expectation and variance rules, not approximations. The normality part comes next — that is the CLT.

Visual intuition. Imagine the population as a big cloud of points. Draw one sample of size $n$ and compute its mean — that is one dot on a new axis. Draw another sample, another dot. After thousands of samples, the dots form a distribution. That distribution (the sampling distribution) is always centered at $μ$ (unbiased) and its spread is $σ / n$ . The shape of that distribution — whether it looks normal or not — is what the CLT addresses.

Pitfalls. 1. Confusing standard deviation ( $σ$ ) with standard error ( $σ / n$ ). $σ$ describes the spread of individual data points. $σ / n$ describes the spread of sample means. The standard error is always smaller than $σ$ for $n > 1$ . 2. Thinking $σ_{x ˉ} = σ / n$ requires the CLT. It does not — it follows directly from $Var (X ˉ) = Var (\sum X_{i} / n) = (1/ n^{2}) \cdot n σ^{2} = σ^{2} / n$ . No normality needed. 3. Forgetting that standard error shrinks slowly. The $n$ in the denominator means diminishing returns: going from $n = 10$ to $n = 40$ cuts SE in half, but going from $n = 100$ to $n = 400$ (much more data) also only cuts SE in half.

Recap. The sampling distribution of $x ˉ$ has mean $μ$ and standard deviation $σ / n$ . These two facts are exact — they need only random sampling and finite variance. The shape of this distribution is what we turn to next: the Central Limit Theorem tells us that shape is a bell curve when $n$ is large enough.

Real-world & domain connection. Polling organizations report "margin of error ±3%" — that margin is just $Z_{α /2} \times$ standard error. When a poll of 1,000 voters has a margin of ±3%, the standard error of the proportion is about $0.03/1.96 \approx 0.0153$ . The same logic applies to every A/B test, clinical trial, and quality control chart: the standard error is what separates signal from noise.

10.5 The Central Limit Theorem

10.5.1 The Big Idea

Hook. Roll a single die. The outcome is uniform — 1 through 6, each equally likely. Now roll 30 dice and take the average. Do this 10,000 times and plot the averages. The histogram looks like a bell curve. The original population was a flat rectangle. The sampling distribution is a perfect bell. This is not a coincidence — it is a theorem.

Intuition. The CLT is the reason statisticians can sleep at night. You almost never know the true distribution of your population. Customer spend is not normal. Click counts are not normal. Lifetimes are not normal. But the CLT says: it doesn't matter. Once you average enough data points ( $n \geq 30$ ), the average behaves as if it came from a normal distribution. The shape of the population washes out, and the bell curve takes over.

Think of it like blending a smoothie. You throw in strawberries (lumpy), bananas (curved), and yogurt (flat). Blend long enough and the result is smooth — the individual shapes disappear. The CLT is the blender for probability distributions: with enough sample size, the sampling distribution of the mean becomes smooth and bell-shaped, no matter what went in.

The Central Limit Theorem (informal). Let $X_{1}, X_{2}, \dots, X_{n}$ be independent and identically distributed (i.i.d.) random variables from any population with mean $μ$ and finite variance $σ^{2}$ . Then, as $n \to \infty$ :

X ˉ_{n} \sim N (μ, \frac{σ ^{2}}{n})

Equivalently, the standardized mean converges to the standard normal:

Z = \frac{X ˉ - μ}{σ / n} d N (0, 1)

Practical rule: For $n \geq 30$ , treat $X ˉ$ as about normal.

10.5.2 Why This Matters

The CLT is the foundation of nearly every confidence interval, hypothesis test, and A/B test you will ever run. It is not that "everything is normally distributed." It is that sample means are, once $n$ is large enough. And most statistics of interest — averages, proportions, regression coefficients — are ultimately sample means of something.

If someone challenges your use of a Z-test: "Who said the data is normal?" — the answer is: Central Limit Theorem. The CLT gives you permission to use normal-based methods whenever $n \geq 30$ , regardless of the population's shape.

10.5.3 The Two-Part Summary

Population is normal → the sampling distribution of $x ˉ$ is exactly normal for any $n$ , even $n = 2$ .
Population is not normal (or unknown) → the sampling distribution approaches normal as $n$ increases. At $n \geq 30$ , it is effectively normal for practical purposes.

In both cases: $μ_{x ˉ} = μ, σ_{x ˉ} = \frac{σ}{n}$

10.5.4 Visual Intuition

The CLT in four frames. Start with an exponential population (heavily right-skewed — long tail to the right). Draw 10,000 samples:

$n = 2$ : The histogram of $x ˉ$ values is still skewed right — the CLT hasn't kicked in yet.
$n = 5$ : The skew starts softening. The right tail shrinks. A hint of symmetry appears.
$n = 15$ : The histogram is noticeably more symmetric. The peak is near $μ$ .
$n = 30$ : A clean, symmetric bell curve centered at $μ$ , with spread $σ / n$ . The exponential shape is gone.

Now repeat with a uniform population (flat rectangle). At $n = 2$ , the sampling distribution is triangular (the sum of two uniforms). At $n = 5$ , it is already bell-like. At $n = 30$ , it is indistinguishable from a normal curve. The CLT washes away the parent distribution's shape and replaces it with normality — sample size is the only knob you need to turn.

Scope: when the CLT applies and when it fails. - Requires independence. The CLT assumes observations are i.i.d. If your data has strong autocorrelation (time series, spatial data), the CLT may not apply without modification. - Requires finite variance. The CLT fails for heavy-tailed distributions with infinite variance (e.g., Cauchy distribution). For these, the sample mean never converges to normality — it stays heavy-tailed no matter how large $n$ is. - $n \geq 30$ is a guideline, not a guarantee. For extremely skewed populations, you may need $n = 50$ or more. For nearly symmetric populations, $n = 15$ may suffice. When in doubt, plot a histogram of bootstrap sample means. - The CLT describes the distribution of $x ˉ$ , not the distribution of the data. Your raw data can be as skewed as ever; only the average becomes normal.

Pitfalls. 1. Applying the CLT to individual observations. The CLT is about $x ˉ$ , not about $x$ . A single data point from an exponential distribution is still exponential, not normal. 2. Assuming $n \geq 30$ magically fixes everything. If your data has outliers, they still affect $x ˉ$ . The CLT ensures normality of the sampling distribution, not strongness of the estimator. 3. Forgetting that the CLT needs independent samples. If you sample without replacement from a small finite population, use the finite correction factor (Section 10.7). 4. Confusing "the distribution is normal" with "the data looks normal." A histogram of your raw data tells you about the population. A histogram of repeated $x ˉ$ values (which you rarely see in practice) tells you about the sampling distribution. The CLT is about the second one.

Recap. The CLT says: sample means are about normal when $n \geq 30$ , no matter what the population looks like. This is the single most important fact in inferential statistics — it is why Z-scores, confidence intervals, and hypothesis tests all use the normal distribution. Next: we put the CLT to work with real numerical examples.

Real-world & domain connection. Every A/B test at every tech company rests on the CLT. When Netflix tests a new recommendation algorithm on 50,000 users, they compute the mean watch time in each group and compare them. The watch time distribution is wildly non-normal (most users watch 0-5 minutes, a few binge for hours). But with $n = 50, 000$ , the CLT guarantees the average watch time per group is normally distributed — so a simple Z-test is valid. Without the CLT, modern data-driven decision-making would be mathematically impossible.

10.6 Worked Examples — Applying the CLT

Hook. You now have the CLT. But a theorem is useless until you can compute with it. The three examples below are the exact pattern that shows up on exams: given $μ$ , $σ$ , and $n$ , find the probability that $x ˉ$ falls in some range. Master this five-step recipe and you can solve any CLT problem in under two minutes.

The five-step recipe for CLT problems.

Confirm $n \geq 30$ (or that the population is stated to be normal) → CLT applies.
Write down $μ_{x ˉ} = μ$ and $σ_{x ˉ} = σ / n$ .
For each boundary value, compute $Z = (x ˉ - μ) / (σ / n)$ .
Look up the Z-values in the standard normal table.
Combine areas to get the requested probability.

10.6.1 Example 1 — Vehicle Age

Problem: The average age of a vehicle registered in the United States is 8 years (96 months). Assume the standard deviation is 16 months. A random sample of 36 vehicles is selected. Find the probability that the mean age of the sample is between 90 and 100 months.

Given: $μ = 96$ months, $σ = 16$ months, $n = 36$ . Find $P (90 \leq x ˉ \leq 100)$ .

Step 1 — Distribution of $x ˉ$ . $n = 36 \geq 30$ , so by CLT, $x ˉ$ is about normal with: $μ_{x ˉ} = 96, σ_{x ˉ} = \frac{16}{36} = \frac{16}{6} \approx 2.667$

Step 2 — Z-scores. $Z_{90} = \frac{90 - 96}{2.667} = \frac{- 6}{2.667} \approx - 2.25$ $Z_{100} = \frac{100 - 96}{2.667} = \frac{4}{2.667} \approx 1.50$

Step 3 — Probability. $P (90 \leq x ˉ \leq 100) = P (- 2.25 \leq Z \leq 1.50)$ From the standard normal table: $F (1.50) = 0.9332, F (- 2.25) = 0.0122$ $P (- 2.25 \leq Z \leq 1.50) = 0.9332 - 0.0122 = 0.9210$

Answer: 0.9210 (about 92.1%). Sense-check: 90 and 100 are both within about 2.25 standard errors of 96, so most samples should land in this range. A 92% probability is reasonable.

10.6.2 Example 2 — Customer Expenditure

Problem: The population mean expenditure per customer is \$85 with $σ = $9$ . A random sample of $n = 40$ customers is taken. Find $P (x ˉ \geq 87)$ .

Step 1 — CLT applies ( $n = 40 \geq 30$ ). $x ˉ$ is about normal with: $μ_{x ˉ} = 85, σ_{x ˉ} = \frac{9}{40} \approx 1.423$

Step 2 — Z-score. $Z = \frac{87 - 85}{1.423} = \frac{2}{1.423} \approx 1.41$

Step 3 — Probability. $P (x ˉ \geq 87) = P (Z \geq 1.41) = 1 - F (1.41)$ From the table, $F (1.41) = 0.9207$ . $P (x ˉ \geq 87) = 1 - 0.9207 = 0.0793$

Answer: 0.0793 (about 7.9%). Sense-check: 87 is 1.41 standard errors above the mean. The right-tail area beyond 1.41 should be small — about 8% feels right. Only about 1 in 13 samples would show an average this high by chance.

10.6.3 Example 3 — Shopping Hours

Problem: The average number of shoppers per hour is 448 with $σ = 21$ . A random sample of $n = 49$ hours is taken. Find $P (441 \leq x ˉ \leq 446)$ .

Step 1 — CLT applies ( $n = 49 \geq 30$ ). $σ_{x ˉ} = \frac{21}{49} = \frac{21}{7} = 3$

Step 2 — Z-scores. $Z_{441} = \frac{441 - 448}{3} = \frac{- 7}{3} \approx - 2.33$ $Z_{446} = \frac{446 - 448}{3} = \frac{- 2}{3} \approx - 0.67$

Step 3 — Probability. $P (441 \leq x ˉ \leq 446) = P (- 2.33 \leq Z \leq - 0.67)$ From the standard normal table: $F (- 0.67) = 0.2514, F (- 2.33) = 0.0099$ $P (- 2.33 \leq Z \leq - 0.67) = 0.2514 - 0.0099 = 0.2415$

Answer: 0.2415 (about 24.2%). Sense-check: both values are below the mean, in the left tail. The area between Z = −2.33 and −0.67 should be moderate — about a quarter of the distribution. Reasonable.

Q: What if $μ$ were 100,000 instead of 448 — would the distribution change? A: No. The decision to use the normal distribution rests entirely on $n \geq 30$ , not on the magnitude of $μ$ . Whether $μ$ is 448 or 100,000, the CLT still applies. The Z-score formula automatically handles the scale.

Q: If the population is already normally distributed, can we use $Z = (x ˉ - μ) / σ$ instead of $Z = (x ˉ - μ) / (σ / n)$ ? A: No — never. Even when the population is normal, $x ˉ$ is a sample mean. Its standard deviation is $σ / n$ , period. The formula $Z = (x - μ) / σ$ applies to a single observation $x$ . For a sample mean $x ˉ$ , you always divide by $σ / n$ . Several students asked this — it is a common trap.

Pitfalls for CLT computation problems. 1. Using $σ$ instead of $σ / n$ in the Z-score denominator. This is the #1 exam mistake. If the question mentions "sample mean" or $x ˉ$ , the denominator is always $σ / n$ . 2. Forgetting to check $n \geq 30$ . If $n = 10$ and the population is not stated to be normal, you cannot use the Z-table. The CLT does not apply. 3. Misreading the probability direction. "Between a and b" → subtract two CDF values. "Greater than c" → $1 - F (Z_{c})$ . "Less than d" → $F (Z_{d})$ . Draw the bell curve and shade the area before computing. 4. Rounding intermediate values too early. Keep at least 3 decimal places in Z-scores before table lookup. Rounding 2.667 to 2.7 changes the probability.

Recap. Every CLT problem follows the same pattern: check $n \geq 30$ , compute $σ_{x ˉ} = σ / n$ , convert to Z-scores, look up the table, combine areas. The only thing that varies is the story (vehicles, shoppers, customers) and the specific boundary values. Next: what changes when the population size $N$ is known and finite?

10.7 Finite Population Correction Factor

10.7.1 When the Population Size Is Known

Hook. You work at a company with exactly 350 employees. You sample 45 of them. The standard error formula $σ / n$ assumes the population is infinite — that you could keep sampling forever and never run out. But with only 350 people, after you have sampled 45, you have seen over 12% of everyone. The remaining uncertainty is smaller than the infinite-population formula suggests. The finite correction factor accounts for this.

Intuition. Imagine drawing marbles from a jar without replacement. If the jar has 1,000,000 marbles, drawing 50 changes almost nothing — the probability of drawing a red marble barely shifts. That is the infinite-population case. If the jar has only 100 marbles and you draw 50, you have seen half the jar — there is much less mystery about what remains. The finite correction factor shrinks the standard error to reflect this reduced uncertainty.

The factor $(N - n) / (N - 1)$ is always between 0 and 1. When $n ≪ N$ (sample is tiny relative to population), the factor is close to 1 — the correction is negligible. When $n$ approaches $N$ , the factor approaches 0 — the standard error nearly vanishes because you have almost measured everyone.

Finite population correction. For an infinite (or very large) population:

σ_{x ˉ} = \frac{σ}{n}

For a finite population of known size $N$ , sampled without replacement:

σ_{x ˉ} = \frac{σ}{n} \times \frac{N - n}{N - 1}

The term $\frac{N - n}{N - 1}$ is the finite population correction factor (FPC). It shrinks the standard error because sampling without replacement from a finite group reduces variability — once you have drawn most units, the remaining ones are nearly determined.

When to use it: The FPC matters when the sampling fraction $n / N > 0.05$ (you are sampling more than 5% of the population). Below that, the correction is negligible and the infinite-population formula is fine.

10.7.2 Worked Example — Employee Ages

Problem: A production company has $N = 350$ hourly employees. The average age is $μ = 37.6$ years with $σ = 8.3$ years. A random sample of $n = 45$ employees is taken. Find $P (x ˉ < 40)$ .

Step 1 — Standard error with FPC. $σ_{x ˉ} = \frac{8.3}{45} \times \frac{350 - 45}{350 - 1}$

First, compute each piece: $\frac{8.3}{45} = \frac{8.3}{6.708} \approx 1.237$ $\frac{305}{349} = 0.8739 \approx 0.9348$ $σ_{x ˉ} = 1.237 \times 0.9348 \approx 1.157$

Without FPC, $σ_{x ˉ}$ would be 1.237. The correction shrinks it by about 6.5%.

Step 2 — Z-score. $Z = \frac{40 - 37.6}{1.157} = \frac{2.4}{1.157} \approx 2.07$

Step 3 — Probability. $P (x ˉ < 40) = P (Z < 2.07) = F (2.07)$ From the standard normal table, $F (2.07) \approx 0.9808$ .

Answer: 0.9808 (about 98.1%). Sense-check: 40 is about 2.07 standard errors above the mean — nearly 98% of samples should fall below it. Without the FPC, Z would be about $2.4/1.237 \approx 1.94$ and the probability would be about 97.4%. The FPC tightened the interval slightly, as expected.

Scope. The FPC applies only when: - The population size $N$ is known and finite. - Sampling is done without replacement (the typical case in surveys). - The sampling fraction $n / N > 0.05$ — otherwise the correction can be ignored. - The CLT still applies ( $n \geq 30$ for the normal approximation), using the corrected standard error.

Visual intuition. Picture the standard error as the width of a confidence band. The infinite-population formula gives a band of width $2 \times σ / n$ . The FPC squeezes that band narrower — the more you sample relative to $N$ , the tighter the squeeze. When $n = N$ (you sample everyone), the band width collapses to zero because $x ˉ = μ$ exactly.

Pitfalls. 1. Applying the FPC when the population is effectively infinite. If $N$ is in the millions and $n = 100$ , the correction factor is $\approx 0.99995$ — pointless to compute. 2. Forgetting the FPC when it matters. If $n = 50$ from $N = 200$ (25% sampled), the correction factor is $150/199 \approx 0.868$ . Ignoring it overstates the standard error by about 15%. 3. Using $N$ instead of $N - 1$ in the denominator. The formula is $(N - n) / (N - 1)$ , not $(N - n) / N$ . The difference is tiny for large $N$ but matters for small populations.

Recap. The finite population correction shrinks the standard error when you sample a non-trivial fraction of a known, finite population. Use it when $n / N > 0.05$ ; ignore it otherwise. Everything else — Z-scores, table lookups, probability calculations — stays the same. Next: proportions, where the story is similar but the standard error formula changes.

Real-world & domain connection. The FPC is standard in survey sampling. When a polling firm surveys 1,000 people from a town of 10,000 (10% sampled), they must apply the FPC to their margin of error. Without it, they would overstate uncertainty and report a wider confidence interval than the data actually supports. When inspecting a batch of 500 manufactured parts by testing 50 (10%), the FPC tightens the acceptance bounds. You can be more certain about the batch quality than the infinite-population formulas would suggest.

10.8 Population Proportion

10.8.1 Definition

Hook. "60% of voters support the bill." That number — 60% — came from a poll of maybe 1,000 people. The true proportion across all millions of voters is unknown. How far off could that 60% be? The sampling distribution of the proportion answers exactly that.

Intuition. A proportion is just a mean in disguise. Suppose you code "supports the bill" as 1 and "opposes" as 0. The sample proportion $p^$ is the average of those 0s and 1s — it is a sample mean of binary data. So everything you learned about the sampling distribution of $x ˉ$ applies to $p^$ as well, with one twist: the variance of a binary (Bernoulli) variable is $P (1 - P)$ , not $σ^{2}$ . That single change gives us the standard error for proportions.

10.8.2 Sampling Distribution of the Proportion

Let the population proportion be $P$ (the true fraction with the characteristic). Let $Q = 1 - P$ . For a random sample of size $n$ , define:

p^= \frac{x}{n}

where $x$ is the count of "successes" in the sample. Then, for large $n$ :

p^\sim N (P, \frac{P Q}{n})

Equivalently: $μ_{p^} = P, σ_{p^} = \frac{P Q}{n}$

The normality condition: $n P \geq 5$ and $n (1 - P) \geq 5$ . These ensure enough successes and failures for the normal approximation to be reasonable.

The Z-score for a sample proportion: $Z = \frac{p ^ - P}{\frac{P Q}{n}}$

Quick check. A sample of 100 factory workers contains 30 union members. Then $p^= 30/100 = 0.30$ . If the true population proportion of union membership is claimed to be $P = 0.25$ , the standard error is: $σ_{p^} = \frac{0.25 \times 0.75}{100} = 0.001875 \approx 0.0433$ So $p^= 0.30$ is about $(0.30 - 0.25) /0.0433 \approx 1.15$ standard errors above the claimed $P$ — not unusual.

10.8.3 Worked Example — Cancer Treatment

Problem: 60% of cancer patients are cured by a new drug ( $P = 0.60$ ). A random sample of $n = 120$ patients is taken. Find $P (p^\geq 0.50)$ — the probability that at least 50% of the sample is cured.

Step 1 — Check normality: $n P = 120 \times 0.6 = 72 \geq 5, n (1 - P) = 120 \times 0.4 = 48 \geq 5 ✓$

Step 2 — Standard error: $σ_{p^} = \frac{0.60 \times 0.40}{120} = \frac{0.24}{120} = 0.002 \approx 0.0447$

Step 3 — Z-score: $Z = \frac{0.50 - 0.60}{0.0447} = \frac{- 0.10}{0.0447} \approx - 2.24$

Step 4 — Probability: $P (p^\geq 0.50) = P (Z \geq - 2.24) = 1 - F (- 2.24)$ From the standard normal table, $F (- 2.24) \approx 0.0125$ . $P (p^\geq 0.50) = 1 - 0.0125 = 0.9875$

Answer: 0.9875 (about 98.8%). Sense-check: the true cure rate is 60%. Asking for "at least 50%" means $p^$ can be as low as 2.24 standard errors below $P$ . That left tail contains only about 1.25% of the distribution — so nearly 99% of samples will show 50%+ cured. This makes sense: with 120 patients, the sample proportion should be close to 60%.

Scope: when the normal approximation for proportions holds. - Requires $n P \geq 5$ AND $n (1 - P) \geq 5$ . If either fails, the binomial distribution is too skewed for the normal approximation. Use exact binomial methods instead. - The sample must be random. A convenience sample of 120 patients from one hospital does not represent all cancer patients. - For finite populations, apply the FPC to $σ_{p^}$ as well.

Visual intuition. The sampling distribution of $p^$ is a bell curve centered at $P$ . Its spread is $P Q / n$ . The spread is widest when $P = 0.5$ (maximum uncertainty — the population is evenly split) and narrows as $P$ approaches 0 or 1 (when almost everyone or almost no one has the characteristic). For fixed $P$ , increasing $n$ narrows the curve — larger samples give more precise proportion estimates.

Pitfalls. 1. Forgetting the normality check. If $n P < 5$ or $n (1 - P) < 5$ , the Z-table gives wrong probabilities. Always check both conditions. 2. Using $p^$ instead of $P$ in the standard error formula. The formula is $P Q / n$ , not $p^(1 - p^) / n$ (that version is for confidence intervals, where $P$ is unknown). In CLT problems, $P$ is given — use it. 3. Confusing the count $x$ with the proportion $p^$ . The Z-score uses $p^= x / n$ , not $x$ directly. If the problem gives a count, divide by $n$ first.

Recap. A proportion is a mean of 0/1 data. Its sampling distribution is about normal when both $n P$ and $n (1 - P)$ exceed 5, with standard error $P Q / n$ . The Z-score formula and table lookups work exactly as they did for $x ˉ$ . Next: we turn the logic around — instead of "given $P$ , what is $P (p^)$ ?" we ask "given $p^$ , where is $P$ ?" That is estimation.

Real-world & domain connection. Every election poll, every clinical trial success rate, every manufacturing defect rate uses the sampling distribution of the proportion. When you read "45% ± 3% (19 times out of 20)," the ±3% is $1.96 \times p^(1 - p^) / n$ , and the "19 times out of 20" is the 95% confidence level. The sample size $n$ is chosen to make that margin small enough to be useful — typically around 1,000 for national polls.

10.9 Estimation — From Sample Back to Population

10.9.1 The Goal of Estimation

Hook. You have measured 85 phone bills. The average is 510 minutes. What is the average for all customers? You cannot say "510" with certainty — another sample of 85 bills would give a different average. So you give a range: "Probably between 500 and 520." How wide should that range be? How confident can you be? This is estimation.

Intuition. A point estimate is like guessing someone's age by looking at them — you give one number and hope you are close. A confidence interval is like saying "I am 95% sure they are between 25 and 35." The interval is wider but more honest — it admits you do not know exactly. In statistics, interval estimates are almost always what you should report, because they quantify your uncertainty.

Point vs. interval — the dartboard analogy. A point estimate is a single dart thrown at a dartboard — you hope it hits the bullseye ( $μ$ ), but you have no idea how close it landed. A confidence interval is like throwing a hoop — you can say "I am 95% confident the hoop encloses the bullseye." The hoop has a known radius and a known success rate.

10.9.2 Confidence Level and Level of Significance

Confidence level = $1 - α$ . Common choices: 90% (α = 0.10), 95% (α = 0.05), 99% (α = 0.01).
Level of significance = $α$ . The probability that the interval does not contain the true parameter.

Interpretation: "I am 95% confident that $μ$ lies between $L$ and $U$ " means: if you repeated the sampling-and-estimation procedure 100 times, about 95 of those intervals would contain the true $μ$ . Any single interval either contains $μ$ or it does not — the "95%" describes the procedure, not the particular interval.

The confidence-precision trade-off. Higher confidence → wider interval. A 99% interval is wider than a 95% interval for the same data. A shopkeeper offering a "99% guarantee" is making a bolder claim than one offering "60%" — and in statistics, that bolder claim costs you precision.

10.9.3 The $Z_{α /2}$ Values

For a normal sampling distribution, the confidence interval is centered at $x ˉ$ and extends $Z_{α /2}$ standard errors in each direction. The tails each contain $α /2$ of the probability.

Memorize these three values:

Confidence	$α$	$α /2$	$Z_{α /2}$
90%	0.10	0.05	1.645
95%	0.05	0.025	1.96
99%	0.01	0.005	2.58

How to find $Z_{α /2}$ from the table (reverse lookup). For 95% confidence, $α /2 = 0.025$ . The area to the left of $Z_{α /2}$ is $0.95 + 0.025 = 0.975$ . Search inside the standard normal table for 0.975 — you will find it at row 1.9, column 0.06, giving $Z = 1.96$ . This is the reverse of what you did in CLT problems: instead of Z → probability, you go probability → Z.

10.9.4 The Confidence Interval Formula

Derivation. Start from the probability statement:

P (- Z_{α /2} \leq \frac{x ˉ - μ}{σ / n} \leq Z_{α /2}) = 1 - α

Multiply all three parts by $σ / n$ :

P (- Z_{α /2} \cdot \frac{σ}{n} \leq x ˉ - μ \leq Z_{α /2} \cdot \frac{σ}{n}) = 1 - α

Subtract $x ˉ$ from all parts (which flips the inequalities):

P (- x ˉ - Z_{α /2} \cdot \frac{σ}{n} \leq - μ \leq - x ˉ + Z_{α /2} \cdot \frac{σ}{n}) = 1 - α

Multiply through by $- 1$ (flipping the inequalities back):

P (x ˉ - Z_{α /2} \cdot \frac{σ}{n} \leq μ \leq x ˉ + Z_{α /2} \cdot \frac{σ}{n}) = 1 - α

So, the $(1 - α) \times 100%$ confidence interval for $μ$ (with $σ$ known) is:

x ˉ \pm Z_{α /2} \cdot \frac{σ}{n}

The ingredients: - $x ˉ$ — sample mean (your data) - $Z_{α /2}$ — critical value from the table (your chosen confidence level) - $σ$ — population standard deviation (assumed known) - $n$ — sample size - $σ / n$ — standard error of the mean

10.9.5 Worked Example — Cell Phone Minutes

Problem: A cellular company wants to estimate the population mean monthly call minutes. A sample of $n = 85$ bills yields $x ˉ = 510$ minutes. Historical data gives $σ = 46$ minutes. Find the 95% confidence interval for $μ$ .

Given: $n = 85$ , $x ˉ = 510$ , $σ = 46$ , confidence = 95% $\to Z_{α /2} = 1.96$ .

Step 1 — Standard error: $\frac{σ}{n} = \frac{46}{85} = \frac{46}{9.220} \approx 4.989$

Step 2 — Margin of error: $Z_{α /2} \cdot \frac{σ}{n} = 1.96 \times 4.989 \approx 9.78$

Step 3 — Confidence interval: $x ˉ \pm margin = 510 \pm 9.78$ $Lower limit = 510 - 9.78 = 500.22$ $Upper limit = 510 + 9.78 = 519.78$

Answer: [500.2, 519.8] minutes. We are 95% confident the true population mean monthly call minutes lies between 500.2 and 519.8.

Sense-check: The margin of error is about 9.78 minutes, which is roughly 1.9% of $x ˉ$ . With $n = 85$ and $σ = 46$ , a ±10 minute band feels reasonable — not too tight (which would overstate precision) and not too wide (which would be useless).

Same data, different confidence levels. - 90% CI: $Z_{0.05} = 1.645$ . Margin = $1.645 \times 4.989 \approx 8.21$ . CI = [501.8, 518.2] — narrower. - 99% CI: $Z_{0.005} = 2.58$ . Margin = $2.58 \times 4.989 \approx 12.87$ . CI = [497.1, 522.9] — wider.

Notice: more confidence costs you width. The 99% interval spans 25.8 minutes; the 90% interval spans 16.4 minutes. There is no free lunch.

Scope: assumptions behind this confidence interval. 1. $σ$ is known. This is rare in practice. When $σ$ is unknown (the usual case), you use the $t$ -distribution instead of $Z$ — covered in later sessions. 2. The sample is random. If data were collected haphazardly, the interval is meaningless. 3. The sampling distribution of $x ˉ$ is about normal. This holds by CLT if $n \geq 30$ , or exactly if the population is normal. 4. Independent observations. No repeated measurements, no time-series dependence.

Visual intuition. Picture the normal curve centered at $x ˉ = 510$ with spread $σ / n \approx 5$ . The 95% confidence interval marks the boundaries at $x ˉ \pm 1.96 \times 5$ . The area between these boundaries captures 95% of the sampling distribution — but the interval is about $μ$ , not about $x ˉ$ . The correct mental picture: imagine many different samples, each producing its own interval $x ˉ \pm 1.96 σ / n$ . About 95% of those intervals will overlap the true $μ$ . Our single interval is one draw from this process.

Pitfalls. 1. Saying "the probability that $μ$ is in the interval is 95%." Wrong. $μ$ is a fixed number — it either is in the interval or it is not. The 95% describes the procedure, not the specific interval. Correct: "We are 95% confident that the interval contains $μ$ ." 2. Using the wrong $Z_{α /2}$ . For a 95% CI, $Z = 1.96$ , not 1.645 (that is for 90%). Mixing these up is a common exam error. 3. Confusing the standard error with the margin of error. The standard error is $σ / n$ . The margin of error is $Z_{α /2} \times σ / n$ . You need both. 4. Reporting a CI without stating the confidence level. "The mean is between 500 and 520" is incomplete — is that 90%, 95%, or 99%?

Recap. A confidence interval is $x ˉ \pm Z_{α /2} \cdot σ / n$ . It gives a range of plausible values for $μ$ with a stated confidence level. Higher confidence → wider interval. The formula needs $σ$ known and normality of $x ˉ$ (via CLT or population normality). Next: a preview of hypothesis testing — what do you do with the interval once you have built it?

Real-world & domain connection. Confidence intervals are the standard output of every clinical trial. When a drug trial reports "the treatment lowered blood pressure by 8.2 mmHg (95% CI: 5.1–11.3)," the 8.2 is the point estimate $x ˉ_{treatment} - x ˉ_{control}$ , and [5.1, 11.3] is the confidence interval. If the interval does not include 0, the effect is "statistically significant." Regulatory agencies like the FDA require confidence intervals, not just point estimates, because the interval reveals how much uncertainty remains.}

10.10 Hypothesis Testing — A Preview

Hook. You have built a confidence interval for $μ$ . Now someone claims " $μ = 100$ ." Your interval is [94, 98]. Their claim falls outside. Do you believe them? This is the jump from estimation to decision — the domain of hypothesis testing.

Intuition — the sweet shop analogy. You walk into a sweet shop. The shopkeeper claims: "This new sweet is excellent." You don't blindly accept his claim. You ask for a tiny piece — a sample. You taste it. If it tastes good, you accept the claim. If it is terrible, you reject it. This is hypothesis testing in miniature: a claim (hypothesis), a sample, an analysis, and a decision.

10.10.1 The Sweet Shop Analogy — Formalized

Map the sweet shop to statistical terms:

Null hypothesis ( $H_{0}$ ): The shopkeeper's claim — "the sweet is excellent." In statistics, $H_{0}$ is the default position: nothing unusual is happening.
Alternative hypothesis ( $H_{a}$ ): The opposite — "the sweet is not excellent."
Sample: The tiny piece you taste.
Test statistic: Your taste judgment — sweet? bitter? salty?
Decision rule: If the taste is bad enough, reject $H_{0}$ and refuse the sweet.

The five-step logic of every hypothesis test:

A claim ( $H_{0}$ ) is made about a population parameter.
You draw a random sample.
You compute a test statistic from the sample.
You compare the statistic to a threshold (critical value).
You reject $H_{0}$ if the statistic falls in the rejection region; otherwise, you fail to reject $H_{0}$ .

Notice "fail to reject" — you never "accept" $H_{0}$ as true. The sweet tasting good does not prove it is excellent; it only means you lacked evidence to reject it. This asymmetry is fundamental.

Scope. Hypothesis testing is the subject of the next lecture. The formal framework — Type I error (rejecting a true $H_{0}$ ), Type II error (failing to reject a false $H_{0}$ ), p-values, and the t-test — will be developed in full then. For now, understand the conceptual flow: claim → sample → test → decision.

Pitfall — jumping ahead. Do not confuse a confidence interval with a hypothesis test, even though they are two sides of the same coin. A 95% CI for $μ$ that excludes 100 is equivalent to rejecting $H_{0} : μ = 100$ at α = 0.05. But the formal machinery — p-values, power, one-tailed vs. two-tailed — comes later.

Recap. Hypothesis testing is decision-making under uncertainty. You start with a claim, gather evidence (a sample), and decide whether the evidence contradicts the claim strongly enough to reject it. The sweet shop analogy captures the entire logic. Next lecture: the full mathematical framework.

Real-world & domain connection. Every A/B test you will ever run is a hypothesis test. $H_{0}$ : "The new button color has no effect on click rate." You collect data (sample), compute the difference in click rates (test statistic), and decide whether the difference is large enough to reject $H_{0}$ and ship the change. The entire tech industry runs on this loop.

10.11 Making Sampling Effective

Hook. You have all the formulas now — CLT, standard errors, confidence intervals. But none of it matters if your sample is bad. A biased sample with $n = 10, 000$ is worse than a random sample with $n = 30$ . The three conditions below are your quality checklist before you compute anything.

Three conditions for effective sampling.

1. Representativeness. The sample must reflect the population. If the population is not homogeneous (e.g., class imbalance), use stratified sampling. A representative sample is one where every subgroup appears in roughly the same proportion as in the population. Stir the pot before you taste.

2. Adequate sample size. Small samples amplify chance. A sample of 10 from a population of 1,000 is too small — the standard error is huge, your estimates are noisy, and the CLT may not apply. The rule of thumb: $n \geq 30$ for the CLT, and at least 20-30% of the population for small finite populations. The sample should not be "small-big" — disproportionate to what you are trying to measure.

3. The right measure. If you want to understand academic performance, you measure exam scores — not heights, not shoe sizes. Every analytical question demands its own appropriate statistic. Using the wrong measure gives a precise answer to the wrong question.

The vegetable-buying analogy (from the professor). When you buy tomatoes, you inspect them by looking and feeling. When you buy ladyfingers (okra), you break the tip to test freshness — a different technique entirely. If your child learns only the tomato technique and applies it to okra, it will fail. Each vegetable needs its own quality test. Each statistical question needs its own appropriate measure. You cannot use a proportion test where a mean test is needed, and you cannot use random sampling where stratification is needed.

Pitfalls. 1. Confusing sample size with sample quality. A big biased sample is still biased. Internet polls with millions of voluntary responses are worse than a carefully designed random sample of 500. 2. Using the wrong statistic for the question. Asking "what proportion of students pass?" and reporting the mean score is a category error. Match your statistic to your research question. 3. Ignoring population heterogeneity. If your population has distinct subgroups (age groups, regions, device types) and you do not stratify, your sample may misrepresent all of them.

Recap. Before you apply any formula from this lecture, verify three things: is your sample representative? Is it large enough? Are you measuring the right thing? If any answer is "no," fix the sampling plan before you compute. The best math cannot rescue bad data.

Real-world & domain connection. In survey method, these three conditions are formalized as the Total Survey Error framework: sampling error (sample size), coverage error (representativeness), and measurement error (right measure). Professional polling firms spend more effort on sampling design than on analysis, because they know that a $10M analysis of a $10 sample is worthless.

Exam Guidance Summary

Exam note: what to expect and how to prepare.

High-probability exam topics

CLT numerical problems. Given $μ$ , $σ$ , and $n$ (with $n \geq 30$ ), find $P (a \leq x ˉ \leq b)$ using Z-scores and the standard normal table. This is the most likely computational question. Practice the five-step recipe from Section 10.6 until it is automatic.
Confidence interval computation. Given $x ˉ$ , $σ$ , $n$ , and a confidence level, compute the interval for $μ$ and interpret it. Memorize the three $Z_{α /2}$ values: 1.645 (90%), 1.96 (95%), 2.58 (99%). Do not confuse them.
Finite vs. infinite population. When $N$ is given, use the finite correction factor $(N - n) / (N - 1)$ in the standard error. When $N$ is not given, use $σ / n$ .
Proportion problems. $Z = (p^- P) / P Q / n$ . Always check the normality conditions: $n P \geq 5$ and $n (1 - P) \geq 5$ .
Stratified sampling. Understand why it is needed (class imbalance makes random sampling unreliable) and how stratify=y works in train_test_split. Expect a conceptual or short-answer question.

Exam pattern

Problems are straightforward — one concept per sub-question. Numerical computation with interpretation. Show your work: write the formula, substitute values, compute the Z-score, look up the table, state the final probability or interval, and add a one-sentence interpretation.

Grading notes

Relative grading is used. The distribution of marks should follow a normal curve. Do not expect to answer every question perfectly — maximum marks often cluster around 60%.
After evaluation, papers and answer keys are shared. Genuine concerns can be raised through the revaluation process.

Key Industry Applications and Real-World Connections

The concepts in this lecture are not just exam material — they are the foundation of data-driven decision-making across every industry that uses statistics.

Machine Learning

Train-test split in every ML pipeline is sampling. Understanding representativeness and stratification prevents biased models. A model that performs well on an unstratified test set of imbalanced data may fail catastrophically in production.
K-fold cross-validation is repeated sampling for strong model evaluation. Each fold is a sample; the CV score is an aggregate across $K$ different train-test splits.
Mini-batch gradient descent uses a different random sample (mini-batch) in each iteration — an application of repeated random sampling at the core of deep learning optimization.
Stratified sampling is essential in medical AI (imbalanced disease datasets), fraud detection (rare events), and any classification problem with skewed class distributions.

A/B Testing and Experimentation

Confidence intervals quantify uncertainty around estimated treatment effects. When a tech company reports "the new feature increased conversion by 2.3% (95% CI: 1.1%–3.5%)," the interval comes directly from $x ˉ \pm Z_{α /2} \cdot σ / n$ .
The CLT justifies using Z-tests and t-tests on experiment metrics (click rates, revenue per user, time on page) even when the raw metric distributions are heavily skewed.

Clinical Trials and Biostatistics

Confidence intervals are required by regulatory agencies (FDA, EMA) for reporting treatment effects. A drug is approved only when the CI for its benefit excludes zero.
Sample size planning uses the standard error formula $σ / n$ in reverse: choose $n$ large enough so the margin of error is clinically meaningful.

Quality Control and Manufacturing

Acceptance sampling uses the sampling distribution of the proportion to decide whether a batch of products meets specifications. A random sample of $n$ items is tested; if too many fail, the batch is rejected.
Control charts track the sample mean $x ˉ$ over time, with control limits set at $μ \pm 3 σ / n$ — another direct application of the sampling distribution.

Finance and Risk

Value at Risk (VaR) estimates use the CLT to approximate the distribution of portfolio returns. The sampling distribution of the mean return underlies many risk models.
Credit scoring models are validated on stratified samples to ensure they perform across different demographic segments.

Survey method and Polling

Every published poll margin of error is $Z_{α /2} \times p^(1 - p^) / n$ . The sample size $n$ is chosen to achieve a target margin (typically ±3% for national polls, requiring $n \approx 1, 067$ ).
Finite population correction is applied when polling small populations (e.g., employee satisfaction in a company of 500).

ISM Lecture 10 notes · Sampling, Sampling Distributions, and Estimation

Introduction to Statistical Methods· postgraduate· 2026-07-06

Sections Breakdown

110.1 Population and Sample — The Core Idea

Defines population vs. sample, the foundation of inferential statistics, and the train-test split as a sampling operation.

210.2 Parameters and Statistics — Two Vocabularies

Distinguishes population parameters (Greek letters) from sample statistics (Roman letters), and explains Bessel's correction (n-1) for sample variance.

310.3 Sampling Methods

Covers random sampling, stratified sampling for imbalanced data, and K-fold cross-validation as repeated sampling.

410.4 Sampling Distribution

Introduces the sampling distribution of the mean, its properties, and why multiple samples matter.

510.5 The Central Limit Theorem

The CLT states that sample means are normally distributed for n >= 30, regardless of the population shape.

610.6 Worked Examples — Applying the CLT

Three fully worked CLT problems demonstrating the five-step recipe for computing probabilities.

710.7 Finite Population Correction Factor

Adjusts the standard error when sampling a non-trivial fraction of a known finite population.

810.8 Population Proportion

The sampling distribution of the proportion with standard error sqrt(PQ/n).

910.9 Estimation — From Sample Back to Population

Constructs confidence intervals for the population mean. Covers confidence levels, Z-values, and interpretation.

1010.10 Hypothesis Testing — A Preview

Introduces the logic of hypothesis testing: claim, sample, test statistic, decision.

1110.11 Making Sampling Effective

Three conditions for effective sampling: representativeness, adequate sample size, and the right measure.

Postgraduate students in statistics, data science, and machine learning

Exam Revision Notes

Below is the distilled, exam-ready core of this lecture. Every entry is built from the full textbook notes above. Use this section for rapid review — but if something doesn't make sense, go back to the full explanation in the main content.

Exam Revision Notes

Population and Sample

Must-know: A population is the complete set (size $N$ ), a sample is a subset (size $n$ ). A parameter (Greek letter) describes the population; a statistic (Roman letter) describes the sample. The sample mean $x ˉ$ estimates the population mean $μ$ .

$x ˉ = \frac{1}{n} i = 1 \sum n x_{i}, μ = \frac{1}{N} i = 1 \sum N x_{i}$

Top pitfall: Confusing sample with population. A number computed from data is a statistic; the true unknown value is a parameter. Never say "the mean is 72" without specifying whether it's $x ˉ$ or $μ$ .

Self-check: Why does the sample variance use $n - 1$ instead of $n$ in the denominator?

Connects to: Sampling distribution (10.4), estimation (10.9)

Sampling Methods

Must-know: Random sampling gives every item an equal chance. Stratified sampling divides the population into homogeneous subgroups (strata) and samples proportionally from each — always use stratify=y for imbalanced classification. K-fold cross-validation reduces variance by averaging across multiple train-test splits.

$n_{stratum k} = \frac{n _{k}}{N} \times n_{total}$

Top pitfall: Using random sampling on imbalanced data. A model that says "not fraud" every time gets 99.9% accuracy when only 0.1% of transactions are fraud. Always stratify when classes are skewed.

Self-check: When should you use stratified sampling instead of random sampling?

Connects to: Population and sample (10.1), estimation (10.9)

The Central Limit Theorem (CLT)

Must-know: For any population with finite variance, the sampling distribution of $x ˉ$ approaches a normal distribution as $n \to \infty$ . At $n \geq 30$ , treat $x ˉ$ as about normal regardless of the population shape. The mean of the sampling distribution equals $μ$ , and its standard deviation is $σ / n$ (the standard error).

$X ˉ_{n} \sim N (μ, \frac{σ ^{2}}{n}), Z = \frac{x ˉ - μ}{σ / n}$

Top pitfall: Using $σ$ instead of $σ / n$ in the Z-score denominator when the question is about a sample mean. The denominator is always $σ / n$ for sample means.

Self-check: A population has $μ = 100$ and $σ = 20$ . You take a sample of $n = 25$ . What is the standard error of the mean?

Connects to: Sampling distribution (10.4), confidence intervals (10.9), hypothesis testing (10.10)

Finite Population Correction Factor (FPC)

Must-know: When sampling without replacement from a finite population of known size $N$ , multiply the standard error by $(N - n) / (N - 1)$ . Apply when the sampling fraction $n / N > 0.05$ . The FPC shrinks the standard error because sampling without replacement reduces variability.

$σ_{x ˉ} = \frac{σ}{n} \times \frac{N - n}{N - 1}$

Top pitfall: Forgetting the FPC when $n$ is a large fraction of $N$ , or applying it when the population is effectively infinite (FPC ≈ 1 when $n ≪ N$ ).

Self-check: Why does the FPC always produce a value between 0 and 1?

Connects to: Sampling distribution (10.4), CLT (10.5)

Population Proportion

Must-know: A proportion is a mean of binary (0/1) data. The sample proportion $p^= x / n$ is about normal when $n P \geq 5$ and $n (1 - P) \geq 5$ , with standard error $P Q / n$ where $Q = 1 - P$ .

$p^\sim N (P, \frac{P Q}{n}), Z = \frac{p ^ - P}{P Q / n}$

Top pitfall: Forgetting the normality check ( $n P \geq 5$ and $n (1 - P) \geq 5$ ). If these conditions fail, the normal approximation is invalid. Also: using $p^$ instead of $P$ in the standard error formula for CLT problems.

Self-check: True cure rate is $P = 0.60$ . A sample of $n = 120$ is taken. Is the normal approximation valid?

Connects to: CLT (10.5), estimation (10.9)

Confidence Intervals for the Mean

Must-know: A $(1 - α) \times 100%$ confidence interval for $μ$ (when $σ$ is known) is $x ˉ \pm Z_{α /2} \cdot σ / n$ . Memorize the three critical values: 1.645 (90%), 1.96 (95%), 2.58 (99%). Higher confidence gives a wider interval.

$x ˉ \pm Z_{α /2} \cdot \frac{σ}{n}$

Top pitfall: Saying "the probability that $μ$ is in the interval is 95%." Wrong. $μ$ is a fixed number — it either is or isn't in the interval. The 95% describes the procedure: if you repeated the sampling many times, 95% of the intervals would contain $μ$ .

Self-check: A 95% CI is [500.2, 519.8]. What is the margin of error? What is the point estimate?

Connects to: CLT (10.5), hypothesis testing (10.10)

Hypothesis Testing (Preview)

Must-know: Hypothesis testing is a formal decision procedure: start with a null hypothesis $H_{0}$ (the default claim), draw a sample, compute a test statistic, and reject $H_{0}$ if the evidence is strong enough. You never "accept" $H_{0}$ — you only "fail to reject." A 95% CI that excludes a value of $μ$ is equivalent to rejecting $H_{0} : μ = that value$ at $α = 0.05$ .

$H_{0} : μ = μ_{0} vs. H_{a} : μ \neq = μ_{0}$

Top pitfall: Confusing a confidence interval with a hypothesis test. They are two sides of the same coin — a CI that excludes $μ_{0}$ implies rejection of $H_{0} : μ = μ_{0}$ at the corresponding $α$ .

Self-check: What is the difference between "failing to reject $H_{0}$ " and "accepting $H_{0}$ "?

Connects to: Confidence intervals (10.9), CLT (10.5)

Previous LectureLecture 09 · Descriptive Statistics, Probability, and Random Variables

Next LectureLecture 11 · 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.

Prerequisite Knowledge

Previously Covered in This Subject

Sampling, Sampling Distributions, and Estimation

10.1 Population and Sample — The Core Idea

10.1.1 Definition and Intuition

10.1.2 The 80-20 Split as Sampling

10.2 Parameters and Statistics — Two Vocabularies

10.2.1 Definition

10.2.2 The N vs. N−1 Question

10.3 Sampling Methods

10.3.1 Random Sampling

10.3.2 Stratified Sampling — Fixing the Class Imbalance Problem

10.3.3 K-Fold Cross-Validation as Sampling

10.4 Sampling Distribution

10.4.1 Why Multiple Samples?

10.4.2 Key Results

10.5 The Central Limit Theorem

10.5.1 The Big Idea

10.5.2 Why This Matters

10.5.3 The Two-Part Summary

10.5.4 Visual Intuition

10.6 Worked Examples — Applying the CLT

10.6.1 Example 1 — Vehicle Age

10.6.2 Example 2 — Customer Expenditure

10.6.3 Example 3 — Shopping Hours

10.7 Finite Population Correction Factor

10.7.1 When the Population Size Is Known

10.7.2 Worked Example — Employee Ages

10.8 Population Proportion

10.8.1 Definition

10.8.2 Sampling Distribution of the Proportion

10.8.3 Worked Example — Cancer Treatment

10.9 Estimation — From Sample Back to Population

10.9.1 The Goal of Estimation

10.9.2 Confidence Level and Level of Significance

10.9.3 The Zα/2​ Values

10.9.4 The Confidence Interval Formula

10.9.5 Worked Example — Cell Phone Minutes

10.10 Hypothesis Testing — A Preview

10.10.1 The Sweet Shop Analogy — Formalized

10.11 Making Sampling Effective

Exam Guidance Summary

High-probability exam topics

Exam pattern

Grading notes

Key Industry Applications and Real-World Connections

Machine Learning

A/B Testing and Experimentation

Clinical Trials and Biostatistics

Quality Control and Manufacturing

Finance and Risk

Survey method and Polling

ISM Lecture 10 notes · Sampling, Sampling Distributions, and Estimation

Sections Breakdown

Exam Revision Notes

Exam Revision Notes

Population and Sample

Sampling Methods

The Central Limit Theorem (CLT)

Finite Population Correction Factor (FPC)

Population Proportion

Confidence Intervals for the Mean

Hypothesis Testing (Preview)

Was this lecture useful?

BitsNotes AI Assistant

Configure AI Chat

Sign in to use AI Chat

10.9.3 The $Z_{α /2}$ Values