Hypothesis Testing

11.1 What Is Hypothesis Testing?

11.1.1 Definition and Intuition

A hypothesis is a statement or claim about a population — not about a sample, but about the entire group you care about. For example: "The average life of this product is two years." "Smoking causes cancer." "This two-wheeler gives 55 km per litre." These are all hypotheses — general statements that claim something about a whole population.

Hypothesis testing (or testing of hypothesis) is the statistical procedure for validating such claims. You cannot test every unit in a population — that would mean testing every product ever made, or every person who ever smoked. Instead, you take a sample, analyse it, and then infer back to the population. This is why hypothesis testing sits under inferential statistics — you infer about the population from a sample.

11.1.2 Population, Sample, and the Core Logic

The population (size $N$) is the whole group. The sample (size $n$) is the part you actually observe. You use the sample to make a decision about the population-level claim.

11.1.3 Visual Intuition

Picture a courtroom. The defendant stands accused. The starting assumption — the "null" position — is that the defendant is innocent. The prosecution brings evidence (sample data). The jury asks: "If the defendant were truly innocent, how likely is it that we would see evidence this damning?" If the answer is "very unlikely," the jury rejects the presumption of innocence and convicts. If the evidence is weak, the jury does not reject innocence — which is not the same as declaring the defendant innocent, only that the evidence did not meet the burden of proof.

This courtroom logic is exactly the logic of hypothesis testing. The null hypothesis is "innocent until proven guilty." The sample data is the evidence. The significance level $\alpha$ is the standard of proof.

11.1.4 Pitfalls

11.1.5 Recap and Bridge

11.1.6 Student Questions and Answers

11.1.7 Real-World & Domain Connection

Hypothesis testing is the backbone of evidence-based decision-making in every quantitative field. Pharmaceutical companies use it to prove a new drug works better than a placebo before the FDA will approve it. Manufacturing plants use it to decide whether a production batch meets quality specifications. Political pollsters use it to determine whether a candidate truly leads or the gap is just sampling noise. In machine learning, A/B testing — comparing two model versions on live traffic — is hypothesis testing in disguise. Mastering this framework means you can separate signal from noise in any data-driven claim you encounter.

11.2 Null and Alternative Hypotheses

11.2.1 Definition

Every hypothesis test involves two competing statements:

• Null hypothesis ($H_0$): The statement you set out to test. It represents the status quo, the existing belief, or the claim being validated. It must contain an equality sign ($=$, $\le$, or $\ge$). The word "null" means "no effect, no difference, no change" — it is the boring, default position.
• Alternative hypothesis ($H_1$ or $H_a$): The complement of $H_0$. If the data forces you to reject $H_0$, then $H_1$ is what you conclude instead. This is often what the researcher wants to prove — the "exciting" result. $H_0$ and $H_1$ are mutually exclusive — they cannot both be true.

For example, if you are testing whether the average life of a product equals two years:

11.2.2 The Equal-to-Sign Rule — Why $H_0$ Must Contain Equality

This is a critical practical rule. When you compute a test statistic like $Z=\frac{\bar{X}-\mu }{\sigma /\sqrt{n}}$, you substitute $\mu$ with the value from $H_0$. This substitution is only justified if $H_0$ contains an equality sign.

The term ${\mu }_0$ (called the null value) must be a specific number. If $H_0$ were strictly $\mu <750$ (no equality), you would have no specific number to plug in for $\mu$. The test statistic formula simply would not work.

Therefore, always ensure $H_0$ is written with $=$, $\le$, or $\ge$. This sometimes means swapping the natural reading of the problem statement. The worked example on light bulb life (Section 11.10) demonstrates this swap in detail.

11.2.3 Three Standard Forms for Comparing Two Means

When comparing two population means (${\mu }_1$ and ${\mu }_2$), there are three standard hypothesis pairs:

You do not pick arbitrarily. The problem statement dictates which pair applies. The alternative hypothesis $H_1$ determines whether the test is one-tailed (left or right) or two-tailed.

11.2.4 Visual Intuition

Imagine a balance scale. $H_0$ says the scale is perfectly balanced — the two sides weigh the same (${\mu }_1={\mu }_2$). Your sample data is a small weight you place on the scale. A one-tailed test asks: "Did the scale tip in this specific direction far enough that it cannot be chance?" A two-tailed test asks: "Did the scale tip in either direction far enough?" The rejection region is how far the scale must tip before you declare it genuinely unbalanced.

11.2.5 Pitfalls

11.2.6 Recap and Bridge

11.2.7 Student Questions and Answers

11.2.8 Real-World & Domain Connection

The null/alternative framework underpins every regulatory decision you have ever benefited from. The FDA requires drug trials to test $H_0$: "new drug = placebo" against $H_1$: "new drug > placebo." Only when $H_0$ is rejected at $\alpha =0.05$ (or stricter) does a drug reach the market. Environmental agencies test $H_0$: "emissions ≤ legal limit" against $H_1$: "emissions > legal limit." The burden of proof always sits on the alternative — you must prove something is wrong, not assume it. This conservative design protects consumers, patients, and the environment from false alarms and from reckless claims alike.

11.3 Parametric vs Non-Parametric Tests

11.3.1 Definition

Tests fall into two broad categories:

• Parametric tests: The hypothesis involves a parameter of the population — typically the mean ($\mu$) or proportion ($P$). You also know (or assume) the population follows a specific distribution, usually the normal distribution. The name comes from the fact that you are testing a parameter of a known distributional form. Examples: Z-test, T-test, ANOVA (F-test).
• Non-parametric tests: The hypothesis is a general statement, not tied to a specific parameter like the mean. Or you have no knowledge of the population's distribution. These tests make fewer assumptions and are sometimes called "distribution-free" tests. Examples: Chi-square test, Mann-Whitney U test, Kruskal-Wallis test.

11.3.2 Tests Covered

11.3.3 Visual Intuition

Draw two distributions side by side. On the left, a clean bell-shaped normal curve — this is the world where parametric tests thrive. Every parametric formula (Z, T, F) was derived assuming data looks roughly like this. On the right, a jagged, skewed, multi-modal mess — real-world data often looks more like this. Non-parametric tests work here because they do not assume any particular shape. They rely on ranks, counts, and order statistics instead of means and variances.

11.3.4 Pitfalls

11.3.5 Recap and Bridge

11.3.6 Real-World & Domain Connection

In clinical trials, the choice between parametric and non-parametric can affect whether a drug gets approved. A t-test might show significance where a Mann-Whitney test does not, or vice versa. Regulatory agencies like the FDA often require both as a sensitivity check. In data science, non-parametric methods like bootstrap confidence intervals have become popular precisely because they do not assume normality — they let the data speak for itself. The rise of computing power has made non-parametric methods practical for datasets where, fifty years ago, you would have been forced to assume normality just to get a closed-form answer.

11.4 One-Tailed vs Two-Tailed Tests

11.4.1 Definition

The alternative hypothesis $H_1$ determines the tail structure:

• Two-tailed test: $H_1$ uses $\ne$. The rejection region is split equally across both tails. Example: $H_1:\mu \ne 3$. If the sample mean is far above 3 OR far below 3, you reject $H_0$.
• One-tailed test: $H_1$ uses $<$ or $>$. The entire rejection region sits in one tail.
  • Right-tailed: $H_1$ uses $>$. Rejection region in the right tail. You only reject if the sample statistic is far above the null value.
  • Left-tailed: $H_1$ uses $<$. Rejection region in the left tail. You only reject if the sample statistic is far below the null value.

11.4.2 Visual Interpretation

For a right-tailed test at $\alpha =0.05$, the rightmost 5% of the normal curve is the rejection region. The remaining 95% is the acceptance region. The boundary Z-value $Z=1.645$ separates them.

For a two-tailed test at $\alpha =0.05$, each tail gets $\alpha /2=0.025$. The boundaries are $Z=\pm 1.96$.

11.4.3 Common Z Critical Values

These are the most-used values — worth memorising:

11.4.4 Pitfalls

11.4.5 Recap and Bridge

11.4.6 Real-World & Domain Connection

One-tailed tests are common in quality control and regulatory compliance: "Is the impurity level above the legal limit?" — you only care about one direction. Two-tailed tests dominate academic research: "Does this intervention have any effect?" — you must be open to surprise in either direction. The choice has real consequences: in 2020, several COVID-19 treatment trials used one-tailed tests to maximise the chance of detecting a benefit, prompting debate about whether potential harm was adequately monitored. The statistical choice encodes a value judgment about which errors you are willing to tolerate.

11.5 Type I and Type II Errors

11.5.1 The Four Possibilities

When you make a decision about $H_0$, there are four possible outcomes:

11.5.2 Which Error Is More Critical?

Type I error (rejecting a true $H_0$) is generally considered more serious in the hypothesis testing framework. The professor gave two powerful analogies:

The grading analogy: A student writes a correct answer (so $H_0$: "answer is correct" is true). The faculty member marks it wrong and gives zero marks (rejects $H_0$). This is a Type I error. The consequence: the student may fail, lose a job offer, or delay graduation. Compare this with the reverse — a wrong answer marked correct (Type II error). That reflects poorly on the faculty member but has less devastating consequences for the student. Type I hurts the student more.

The medical diagnosis analogy: A patient has chest pain that is actually a heart attack ($H_0$: "patient has heart disease" is true). The doctor dismisses it as gastric trouble (rejects $H_0$). This is a Type I error — the true condition is missed. The patient goes home, ignores the symptoms, and the next episode is critical. Compare with the reverse: the patient's chest pain is NOT a heart attack, but the doctor says it is and recommends surgery (Type II error). Both are bad, but the first can be fatal.

11.5.3 Why Focus on Type I Error?

In hypothesis testing, the primary focus is on controlling Type I error. The probability of Type I error is $\alpha$, the level of significance — and you get to choose it. You set the bar for how much false-positive risk you are willing to tolerate before you declare a finding "significant."

Type II error (probability $\beta$) is also important — it tells you how sensitive your test is — but it is not the primary decision-making tool in introductory hypothesis testing. The reason is practical: $\alpha$ is a single number you control directly, while $\beta$ depends on the unknown true effect size. You can always reduce $\beta$ by increasing the sample size, but you must fix $\alpha$ before the test.

11.5.4 Visual Intuition

Draw two overlapping normal curves side by side. The left curve is the sampling distribution of the test statistic when $H_0$ is true (centred at the null value). The right curve is the distribution when $H_1$ is true (centred at some true effect). The vertical line is your critical value — the decision boundary.

The area under the $H_0$ curve to the right of the critical value is $\alpha$ (Type I error region). The area under the $H_1$ curve to the left of the critical value is $\beta$ (Type II error region). As you slide the critical value to the right, $\alpha$ shrinks but $\beta$ grows — this is the fundamental trade-off. You cannot reduce both simultaneously without increasing the sample size, which narrows both curves.

11.5.5 Pitfalls

11.5.6 Recap and Bridge

11.5.7 Real-World & Domain Connection

The Type I / Type II trade-off shapes legal systems, not just statistics. "Innocent until proven guilty" means the legal system sets $\alpha$ very low — it would rather let guilty people go free (Type II) than convict the innocent (Type I). In contrast, airport security screening sets $\beta$ very low — it would rather pat down innocent passengers (Type I) than let a weapon through (Type II). Every decision system that operates under uncertainty implicitly chooses its tolerance for these two errors. Hypothesis testing makes that choice explicit and quantifiable.

11.6 Level of Significance and Confidence Level

11.6.1 Definition

• Level of significance ($\alpha$): The probability of making a Type I error — rejecting $H_0$ when it is actually true. It is your self-imposed tolerance for false positives. You choose $\alpha$ before seeing the data.
• Confidence level ($1-\alpha$): The probability of NOT making a Type I error when $H_0$ is true. It quantifies how sure you are that your rejection is correct.

11.6.2 Where $\alpha$ Comes From

The problem statement typically gives $\alpha$ explicitly — "test at 5% level of significance" means $\alpha =0.05$. If not specified, common practice uses $\alpha =0.05$.

11.6.3 Why 0.05?

The convention of $\alpha =0.05$ traces back to R.A. Fisher in the 1920s, who suggested that a 1-in-20 chance was a reasonable threshold for "significant" evidence. It is not a mathematical truth — it is a cultural convention that stuck. Some fields (physics, genomics) now use much smaller thresholds ($\alpha =0.001$ or even $5\times 10^{-8}$ for genome-wide significance) because they run millions of tests and need to control the overall false-positive rate. The key insight: $\alpha$ is a choice, not a law of nature. Pick the value that matches the cost of being wrong in your domain.

11.6.4 Pitfalls

11.6.5 Recap and Bridge

11.6.6 Real-World & Domain Connection

Different industries set $\alpha$ differently because their error costs differ. Pharmaceutical trials typically use $\alpha =0.05$ for primary endpoints. Manufacturing quality control often uses $\alpha =0.01$ (or even Six Sigma's 3.4 defects per million, which corresponds to a very tiny $\alpha$). Algorithmic trading systems may use $\alpha =0.001$ to avoid acting on spurious patterns that would lose money. The choice of $\alpha$ is not a statistics question — it is a business decision about risk tolerance, dressed in mathematical clothing.

11.7 The Hypothesis Testing Procedure

11.7.1 Purpose

The six-step procedure standardises every hypothesis test into a repeatable workflow. It separates the conceptual work (steps 1–4: what are we testing and what is our rule?) from the computational work (step 5: plug in numbers) from the interpretive work (step 6: what does it mean?). Following these steps in order prevents the most common errors: wrong hypothesis direction, wrong critical value, wrong conclusion language.

11.7.2 Inputs & Outputs

11.7.3 The Six Steps

11.7.4 Trace — Sodium Example (Preview)

Apply the six steps to the sodium content problem (fully worked in Section 11.9):

Step 1: $H_0:\mu \le 920$ (restaurant's claim: sodium ≤ 920 mg)
Step 2: $H_1:\mu >920$ (sodium is more than claimed) → right-tailed
Step 3: $\alpha =0.05$
Step 4: Reject $H_0$ if $Z>1.645$
Step 5: $Z=\frac{925-920}{18/\sqrt{44}}=1.842$
Step 6: $1.842>1.645$ → Reject $H_0$. The restaurant's claim is not supported.

11.7.5 Decision Criteria / Critical Region in Detail

The decision criteria (or critical region) is the rule that tells you when to reject $H_0$. It depends on three things:

• Whether the test is one-tailed or two-tailed (from $H_1$)
• The level of significance $\alpha$
• Whether you are using Z or T

For example: at $\alpha =0.05$, right-tailed Z-test, the decision criteria is: reject $H_0$ if $Z>1.645$. This is because the rightmost 5% of the standard normal curve begins at $Z=1.645$.

The entire normal curve together with this boundary forms the decision framework. If your computed Z falls in the shaded tail, you reject. If it falls in the unshaded body, you do not reject.

11.7.6 The "Do Not Reject" Language

Pure statisticians do not say "accept $H_0$." They say "do not reject $H_0$" or "there is not enough evidence to reject $H_0$." This is a precise phrasing that reflects an important philosophical point: failing to find evidence against $H_0$ is not the same as proving $H_0$ true.

11.7.7 Pitfalls

11.7.8 Recap and Bridge

11.7.9 Real-World & Domain Connection

This six-step procedure is not just an academic exercise — it is the template used by regulatory agencies worldwide. When the FDA reviews a new drug application, the statistical analysis plan must pre-specify $H_0$, $H_1$, $\alpha$, the test statistic, and the critical region before any data is unblinded. Deviating from this order is considered scientific misconduct. The procedure exists to prevent analysts from subconsciously (or consciously) tuning their method to get a desired result — a phenomenon called "p-hacking" that has contributed to the replication crisis in psychology, medicine, and economics.

11.8 The Complete Decision Tree for Choosing a Test

11.8.1 Purpose

The decision tree answers the single most common confusion in hypothesis testing: "Which formula do I use?" Instead of memorising a dozen disconnected formulas, you learn one branching logic. Given a problem statement, the tree tells you the correct test, the correct formula, and whether to use Z or T — all from the words in the problem.

11.8.2 The Master Flow Chart

Testing of Hypothesis
├── MEAN (discussion involves average/mean/μ)
│   ├── One Mean (one population)
│   │   ├── Large sample (n ≥ 30) → Z-test
│   │   └── Small sample (n < 30) → T-test
│   ├── Two Means (comparing two populations)
│   │   ├── Large sample (n₁ + n₂ - 2 ≥ 30) → Z-test
│   │   └── Small sample (n₁ + n₂ - 2 < 30) → T-test
│   └── Several Means (3+ populations) → ANOVA (F-test)
│
└── PROPORTION (discussion involves percentages, fractions, rates — no mean)
    ├── One Proportion
    │   ├── Large sample → Z-test
    │   └── Small sample → T-test (rare in practice)
    ├── Two Proportions → Z-test (large sample formula)
    └── Several Proportions → Chi-square (χ²) test

11.8.3 How to Trace the Tree

11.8.4 The Two-Means Sample Size Rule

For two means, do NOT judge each sample individually. Instead, compute:

If $\text{df}\ge 30$, treat as large sample (Z-test). If $\text{df}<30$, treat as small sample (T-test).

11.8.5 Formula Reference for Z-Tests

11.8.6 T-Distribution vs Z-Distribution

The T-distribution is also bell-shaped and symmetric, but slightly flatter (more spread out) than the normal distribution. It accounts for the extra uncertainty when $n$ is small and you must estimate $\sigma$ with $s$.

11.8.7 Pitfalls

11.8.8 Recap and Bridge

11.8.9 Real-World & Domain Connection

Decision trees like this one are not just for classrooms. Statistical software (R, Python's scipy.stats, SAS) implements exactly this branching logic internally when you call a function like ttest_ind or ztest. Knowing the tree means you understand what the software is doing under the hood. More importantly, you can catch it when it does the wrong thing because you gave it the wrong data type. In industry, being able to look at a business question and say "that is a two-sample Z-test for proportions" is worth more than computing the formula by hand.

11.9 Worked Example 1 — Sodium Content (One Mean, Right-Tailed Z-Test)

11.9.1 Problem Statement

A food restaurant estimates that the mean sodium content in one of its breakfast sandwiches is not more than 920 milligrams. A random sample of 44 breakfast sandwiches has a mean of 925 mg with a standard deviation of 18 mg. Test at $\alpha =0.05$.

11.9.2 Tracing the Decision Tree

• Involves "mean sodium content" → MEAN branch
• One product, one population → One mean
• $n=44\ge 30$ → Large sample → Z-test

11.9.3 Setting Up Hypotheses

The claim to test: mean sodium ≤ 920 mg.

• $H_0:\mu \le 920$ (the restaurant's claim — contains $\le$, works in the formula with ${\mu }_0=920$)
• $H_1:\mu >920$ (sodium is more than claimed)

Since $H_1$ uses $>$, this is a right-tailed test.

11.9.4 Decision Criteria

At $\alpha =0.05$, right-tailed Z-test:

• Reject $H_0$ if $Z>1.645$

The Z value 1.645 comes from the standard normal table: the point where the right-tail area is 0.05, meaning the body area (to the left) is 0.95.

11.9.5 Computing the Test Statistic

Given: $\bar{X}=925$, ${\mu }_0=920$, $s=18$, $n=44$.

Population standard deviation $\sigma$ is not available, so use $s$:

11.9.6 Conclusion

Computed $Z=1.842$ is greater than the critical value $1.645$. It falls in the rejection region.

Reject $H_0$. The restaurant's claim that sodium is ≤ 920 mg is not supported. The evidence suggests the mean sodium content is greater than 920 mg.

11.9.7 Real-World & Domain Connection

This is exactly how food regulatory bodies enforce labelling laws. The FDA and similar agencies periodically sample products from shelves, run hypothesis tests on nutritional content, and issue violations when $H_0$ (the label claim) is rejected. The $\alpha =0.05$ standard means they accept a 5% chance of wrongly accusing a compliant manufacturer — a trade-off between consumer protection and business fairness.

11.10 Worked Example 2 — Light Bulb Life (One Mean, Left-Tailed Z-Test, with Hypothesis Swap)

11.10.1 Problem Statement

A light bulb manufacturer guarantees that the mean life of a certain type of light bulb is less than 750 hours. A sample of 36 bulbs has a mean life of 745 hours with a standard deviation of 60 hours. Test at $\alpha =0.01$.

11.10.2 The Hypothesis Swap — Critical Discussion

11.10.3 Setting Up Hypotheses (Final Form)

• $H_0:\mu \ge 750$
• $H_1:\mu <750$

Since $H_1$ uses $<$, this is a left-tailed test.

11.10.4 Decision Criteria

At $\alpha =0.01$, left-tailed Z-test:

• Reject $H_0$ if $Z<-2.33$

The critical value $-2.33$ is from the standard normal table. For a left-tailed test at $\alpha =0.01$, the left-tail area is 0.01, corresponding to $Z=-2.33$ (symmetric to the right-tail value $+2.33$).

11.10.5 Computing the Test Statistic

Given: $\bar{X}=745$, ${\mu }_0=750$, $s=60$, $n=36$.

11.10.6 Conclusion

Computed $Z=-0.5$ is greater than $-2.33$ (it falls in the acceptance region — to the right of the critical value).

Do not reject $H_0$. That means $\mu \ge 750$ stands. Translating back to the original claim: the manufacturer's guarantee that mean life is less than 750 hours is rejected. The data does not support the claim.

11.10.7 Real-World & Domain Connection

This type of test — where a manufacturer's claim becomes $H_1$ via the swap — is standard in false-advertising litigation. A company says "our product lasts less than X" (which sounds bad, but maybe they mean it degrades safely). Regulators test $H_0:\mu \ge X$ and only act if they can prove the opposite. The burden of proof is on the regulator to show the claim is false, not on the manufacturer to prove it true. The swap encodes a legal principle: the accuser must prove guilt.

11.11 Worked Example 3 — Package Delivery (One Mean, Two-Tailed Z-Test)

11.11.1 Problem Statement

A company claims the average delivery time for its packages is 3 days. A sample of 50 packages has a mean delivery time of 3.4 days with a standard deviation of 1.2 days. Test at $\alpha =0.05$.

11.11.2 Tracing the Decision Tree

• Involves "average delivery time" → MEAN branch
• One company, one population → One mean
• $n=50\ge 30$ → Large sample → Z-test

11.11.3 Setting Up Hypotheses

The claim to test: mean delivery time = 3 days. The question is simply: is this true or not?

• $H_0:\mu =3$
• $H_1:\mu \ne 3$

Since $H_1$ uses $\ne$, this is a two-tailed test. You would reject if delivery is significantly faster OR significantly slower than 3 days.

11.11.4 Decision Criteria

At $\alpha =0.05$, two-tailed Z-test:

• Reject $H_0$ if $Z<-1.96$ or $Z>+1.96$

The total $\alpha =0.05$ is split: $\alpha /2=0.025$ in each tail. The critical values $\pm 1.96$ bound the middle 95% of the standard normal.

11.11.5 Computing the Test Statistic

Given: $\bar{X}=3.4$, ${\mu }_0=3$, $s=1.2$, $n=50$.

11.11.6 Conclusion

Computed $Z=2.36$ is greater than $+1.96$. It falls in the right-tail rejection region.

Reject $H_0$. The company's claim that average delivery time is 3 days is not supported. The evidence suggests the mean differs from 3 days (specifically, it appears to be greater — deliveries are taking longer than promised).

11.11.7 Real-World & Domain Connection

Delivery time testing is a standard operations-analytics use case. E-commerce companies continuously run hypothesis tests on delivery promises. If $H_0:\mu =\text{promised days}$ is rejected, either the promise is updated or the logistics chain is investigated. Two-tailed tests are appropriate here because being consistently faster than promised can also be a problem — it might mean you are over-spending on premium shipping that customers did not pay for.

11.12 Worked Example 4 — Tire Brands (Two Means, Two-Tailed Z-Test)

11.12.1 Problem Statement

Is there a significant difference between two brands of tires? Sample from Brand 1: $n_1=30$, ${\bar{X}}_1=38,600$ km, $s_1=3,200$ km. Sample from Brand 2: $n_2=40$, ${\bar{X}}_2=40,100$ km, $s_2=2,800$ km. Test at $\alpha =0.05$.

11.12.2 Tracing the Decision Tree

• Involves "mean" (tread life in km) and comparison between two brands → Two means
• $\text{df}=n_1+n_2-2=30+40-2=68\ge 30$ → Large sample → Z-test

11.12.3 Setting Up Hypotheses

"Significant difference" means testing whether the means are equal or not equal:

• $H_0:{\mu }_1={\mu }_2$ (no difference between brands)
• $H_1:{\mu }_1\ne {\mu }_2$ (there IS a difference)

Two-tailed test — you are open to Brand 1 being either better or worse.

11.12.4 Decision Criteria

At $\alpha =0.05$, two-tailed: reject $H_0$ if $|Z|>1.96$.

11.12.5 Computing the Test Statistic

Use the two-means formula with sample standard deviations (population variances unknown):

Under $H_0:{\mu }_1={\mu }_2$, the term $({\mu }_1-{\mu }_2)=0$ and drops out.

11.12.6 Conclusion

$|Z|=2.05>1.96$ → Reject $H_0$. There is a statistically significant difference between the two tire brands. Brand 2 appears to have a higher mean tread life.

11.12.7 Student Q&A

11.12.8 Pitfalls

11.12.9 Real-World & Domain Connection

Tire manufacturers and automotive magazines run exactly these tests. Consumer Reports, for example, buys tires off the shelf, mounts them on test vehicles, measures tread wear under controlled conditions, and publishes comparisons. A two-sample Z-test (or more commonly, a two-sample T-test) determines whether the observed differences are "real" or within the margin of sampling error. The result directly affects purchasing recommendations read by millions of consumers.

11.13 Worked Example 5 — Academic Performance (Two Means, Right-Tailed Z-Test, with Hypothesis Swap)

11.13.1 Problem Statement

Students in Group A performed better on average than students in Group B. Sample: Group A — $n_1=40$, ${\bar{X}}_1=78$, $s_1=8$. Group B — $n_2=35$, ${\bar{X}}_2=74$, $s_2=6$. Test at $\alpha =0.05$.

11.13.2 Tracing the Decision Tree

• Means involved, comparing two groups → Two means
• $\text{df}=40+35-2=73\ge 30$ → Large sample → Z-test

11.13.3 Setting Up Hypotheses

The claim: Group A > Group B → ${\mu }_1>{\mu }_2$ (strict inequality, no equality). Following the equal-to-sign rule — swap so $H_0$ gets the equality:

• $H_0:{\mu }_1\le {\mu }_2$ (Group A is not better — contains $\le$, so equality sign present)
• $H_1:{\mu }_1>{\mu }_2$ (Group A IS better — the original claim becomes $H_1$)

Right-tailed test. As with the light bulb example, you will interpret the result backward: if you reject $H_0$, you support the claim that Group A is better.

11.13.4 Decision Criteria

At $\alpha =0.05$, right-tailed: reject $H_0$ if $Z>1.645$.

11.13.5 Computing the Test Statistic

11.13.6 Conclusion

$Z=2.47>1.645$ → Reject $H_0$. Accept $H_1$: Group A performs significantly better than Group B.

11.13.7 Real-World & Domain Connection

This is the standard setup for educational research: compare test scores between a treatment group (new teaching method) and a control group (traditional method). The hypothesis swap is standard here because the researcher's claim ("the new method is better") is a strict inequality that becomes $H_1$. A significant result means the new method is adopted; a non-significant result means sticking with the status quo. The burden of proof is on the innovation, not on tradition.

11.14 Worked Example 6 — Disease Survival (One Proportion, Right-Tailed Z-Test)

11.14.1 Problem Statement

20 people were attacked by a disease and only 18 survived. Test whether the survival rate is 85% or more at $\alpha =0.05$. Use a large sample test.

11.14.2 Tracing the Decision Tree

• No mean discussed. The problem talks about "survival rate" — a percentage/fraction → PROPORTION branch
• One disease, one population → One proportion
• $n=20$ would normally be a small sample, but the problem explicitly says "use a large sample test" → Z-test

11.14.3 Setting Up Hypotheses

The problem asks: is the survival rate 85% or is it more than 85%?

• $H_0:P=0.85$ (contains equality — works in the formula)
• $H_1:P>0.85$ (the question: is it MORE than 85%?)

Right-tailed test.

11.14.4 Decision Criteria

At $\alpha =0.05$, right-tailed: reject $H_0$ if $Z>1.645$.

11.14.5 Computing the Test Statistic

Sample proportion: $\hat{p}=\frac{18}{20}=0.9$

Population proportion under $H_0$: $P=0.85$, $Q=1-P=0.15$

11.14.6 Conclusion

$Z=0.63<1.645$ → Do not reject $H_0$. The hypothesis that the survival rate is 85% stands. There is not enough evidence to say the survival rate is more than 85%.

11.14.7 Real-World & Domain Connection

Survival rate testing is standard in medical research. A Phase II clinical trial might test $H_0:P=0.85$ (the current standard-of-care survival rate) against $H_1:P>0.85$ (the new treatment improves survival). With only 20 patients, the trial is underpowered — even a 5-percentage-point improvement cannot be detected. This is why medical trials typically require hundreds or thousands of patients: to have enough power to detect clinically meaningful improvements. Sample size planning (power analysis) is the tool that determines how many patients you need BEFORE running the trial.

11.15 Worked Example 7 — Two Judges (Two Proportions, Z-Test)

11.15.1 Problem Identification

This problem involves comparing proportions between two judges — two populations, proportions, large sample → two-proportion Z-test. The professor presented this as a quick identification exercise to confirm understanding of the decision tree.

11.15.2 Real-World & Domain Connection

Two-proportion tests are used extensively in legal analytics: comparing conviction rates between judges, between jurisdictions, or before and after a policy change. They also appear in A/B testing for conversion rates — does landing page A convert more visitors than landing page B? The pooled proportion formula accounts for the fact that, under $H_0$, both groups are drawn from the same underlying conversion rate.

Exam Guidance Summary

What to prioritise

• The master flow chart (Section 11.8) is your single most important reference. If you understand how to trace any problem through that tree, you can solve virtually any hypothesis testing question. Print it out. Memorise the branching logic.
• Common question types (roughly in order of exam frequency):
  1. One mean, large sample — Z-test (most common)
  2. One mean, small sample — T-test
  3. Two means — Z-test (large) or T-test (small)
  4. One proportion — Z-test
  5. ANOVA and Chi-square — separate topics, covered next session

High-risk concepts

• The equal-to-sign rule for $H_0$ is the most frequent source of confusion. When the natural claim is a strict inequality ($<$ or $>$), swap the pair so $H_0$ gets the equality ($\le$ or $\ge$). Then interpret back at the end. Practice the light bulb example (Section 11.10) until the swap feels automatic.
• The two-means sample size rule uses $n_1+n_2-2$, not individual sample sizes. Exam questions may deliberately give you one small and one large sample to test whether you apply the rule correctly.

What to memorise

These six numbers appear in nearly every problem. Know them cold.

Language conventions

Statistical language uses "do not reject $H_0$" rather than "accept $H_0$." For problem-solving, treat them as equivalent — the professor accepts both phrasings on exams. But in your conclusion sentence, always answer the original question in plain English: "The evidence suggests the mean sodium content exceeds 920 mg," not just "Reject $H_0$."

Looking ahead

• Chi-square and ANOVA will be covered in the next session. These handle "several proportions" and "several means" respectively — the bottom branches of the decision tree.
• The next session will include 10–15 more numerical problems to solidify concepts. The worked examples in Sections 11.9–11.15 cover every major problem type that will appear.

Key Industry Applications

Hypothesis testing is not just an academic exercise — it is the decision engine behind quality assurance, medical approvals, and policy evaluation across every quantitative industry.

Quality Control & Manufacturing

Testing whether mean sodium content, product life, or fill volume meets specifications. Every production line in the world uses some form of hypothesis testing — often automated — to decide whether to stop the line, adjust a process, or release a batch. The Z-test and T-test are the mathematical backbone of Statistical Process Control (SPC).

Medical Research & Pharmaceuticals

Testing survival rates and treatment efficacy, where Type I vs Type II error trade-offs have life-or-death implications. The FDA's drug approval process is, at its core, a hypothesis test: $H_0$: new drug = placebo vs $H_1$: new drug > placebo, tested at $\alpha =0.05$ (or stricter). False positives mean dangerous drugs reach the market. False negatives mean effective drugs never do.

Education & Social Science

Comparing academic performance between groups (teaching methods, curricula, demographics). Every "statistically significant" finding in an education journal traces back to the same six-step procedure covered in this lecture. The hypothesis swap (Section 11.10) is particularly common here because research hypotheses are almost always directional ("the new method is better").

Legal & Judicial Analytics

Evaluating agreement between judges using two-proportion tests. Are two judges ruling consistently, or does one have a significantly different conviction rate? These tests have been used in appellate cases to demonstrate judicial bias. The two-proportion Z-test is the standard tool.

Marketing & E-Commerce

Testing whether a campaign improved response rates (proportion tests) or whether a new website layout increases average order value (mean tests). A/B testing — the dominant experimental method in tech — is hypothesis testing with a different name. Every time Google, Amazon, or Netflix shows you a variant and measures your response, a Z-test or T-test is running behind the scenes.

Food Industry & Regulatory Compliance

Validating nutritional label claims against sample testing. When a cereal box says "10g protein per serving," that number came from lab tests, and regulatory agencies verify it using the exact procedure from Section 11.9. The $\alpha$ level is chosen to balance consumer protection against the cost of false accusations — a policy decision dressed in statistical clothing.