Statistics › Section I · Hypothesis testing with one sample › Topic I.1

Null and alternative hypotheses

Every hypothesis test starts with two competing statements about a population. Writing them correctly, before any data are examined, decides everything the test can later conclude.

Innocent until evidence

Statistics has a courtroom. The accused is a claim: "the machine fills accurately", "the drug does nothing", "support is unchanged". The trial's rules mirror justice - the claim is presumed true, the data are the evidence, and only evidence strong enough overturns the presumption. The claim under trial is the null hypothesis; the prosecution's case is the alternative. Writing the two statements carefully, before any trial begins, is this topic's whole job.

The null hypothesis

The null hypothesis \(H_0\) is the default claim: no effect, no change, nothing unusual. It is written as a statement about a population parameter with a specific value: \(H_0: \mu = 500\) grams (the machine fills accurately), or \(H_0: p = 0.5\) (the coin is fair). By convention the null always carries the equality: whatever direction the investigation leans, \(H_0\) states the value being tested. It is presumed true at the start - not because anyone believes it, but because overturning a presumption requires evidence.

The alternative hypothesis

The alternative hypothesis \(H_a\) is what the study seeks evidence for: the claim that something is going on. It comes in three forms. \(H_a: \mu < 500\) (one-sided, left): underfilling. \(H_a: \mu > 500\) (one-sided, right): exceeding. \(H_a: \mu \ne 500\) (two-sided): differing in either direction. \(H_0\) and \(H_a\) are contradictory: they never overlap, so at most one is true. With a one-sided \(H_a\), the null written \(\mu = 500\) stands for \(\mu \ge 500\) (or \(\le\)); many textbooks write that form outright, and either way the boundary value 500 is the one tested. The form of \(H_a\) is decided by the question, before any data are collected.

Parameters, never statistics

Hypotheses are about population parameters - \(\mu\), \(p\), \(\sigma\) - never about sample statistics. \(H_0: \mu = 500\) is a hypothesis; \(H_0: \bar{x} = 500\) is a category error, because \(\bar{x}\) is already known from the data and needs no testing. The whole logic of testing is that the sample tells us about the population: the statistic (known) speaks about the parameter (unknown). A sample mean quoted in a word problem belongs in the calculation - never in the hypothesis.

Reading the wording into the form

The question's phrasing writes \(H_a\). "Is the machine underfilling?" looks one way: \(\mu < 500\). "Does the new battery outlast the old?" looks one way: \(\mu > \mu_0\). "Has the average changed?" looks both ways: \(\mu \ne 500\). "Different from", "no longer" and "either direction" signal two-sided; "more than", "exceeds", "less than" signal one-sided. Read twice, write once: the form chosen here decides how the p-value will be computed several steps later.

Worked example: three claims, six statements

Write the hypotheses for each. a) A consumer group suspects a cereal brand underfills its 500 g boxes: \(H_0: \mu = 500\), \(H_a: \mu < 500\) - one-sided left, because only underfilling is a complaint. b) A lab tests whether a drug's effect differs from the standard 10 mmHg: \(H_0: \mu = 10\), \(H_a: \mu \ne 10\) - two-sided, because "differs" names no direction. c) A battery maker claims its cells last more than 20 hours and a sceptic doubts it: the boast being tested is the alternative, so \(H_0: \mu = 20\), \(H_a: \mu > 20\) - one-sided right. Check: each pair contradicts, and each names a parameter with a value.

Worked example: debugging bad hypotheses

Spot the errors. a) "\(H_0: \bar{x} = 496\)" - the sample mean in the hypothesis: \(\bar{x}\) is known data, not a claim; it should read \(H_0: \mu = 500\). b) "\(H_a: \mu \ne 500\)" for an underfilling complaint - two-sided where one-sided belongs: the complaint names a direction, so \(H_a: \mu < 500\). c) "\(H_0: \mu < 500\)" - a strict inequality in the null: the null must contain the equality, so \(H_0: \mu = 500\) (or \(\mu \ge 500\)). d) "\(H_0\): the sample of 36 boxes averages under 500" - a description of data, not a hypothesis about the population. Check: every corrected pair follows the two rules - a parameter with a value in \(H_0\), a direction in \(H_a\).

Why the null gets the presumption

The presumption is what makes a test a fair trial. Presume \(H_0\) true, compute what the data would look like under it, and measure how surprising the observed result is - that surprise is the p-value, the machinery of the coming topics. Flip the presumption - assume the alternative, demand proof of nothing - and every claim would pass, however weak its evidence. The courtroom's logic is the statistician's: the burden of proof sits on the claim of change, and "not enough evidence" is a verdict, not a proof of innocence.

Common mistakes to avoid

First, writing sample statistics into the hypotheses: \(\bar{x}\) and \(\hat{p}\) are data, not claims. Second, choosing the direction after seeing which way the data lean - the question names the direction before measurement. Third, putting a strict inequality in \(H_0\): the null always contains the equality. Fourth, using a two-sided alternative for a one-sided question, or the reverse - the form changes the p-value's computation. Fifth, "accepting" \(H_0\) when a test fails to reject it - a preview of the verdict vocabulary. Sixth, testing a claim about one individual rather than a parameter.

Key takeaways

A hypothesis test tries a claim in a courtroom. The null \(H_0\) is the default - no effect, a specific parameter value, carrying the equality - presumed true unless the evidence argues. The alternative \(H_a\) is the claim being sought: \(\mu < \mu_0\), \(\mu > \mu_0\), or \(\mu \ne \mu_0\), one-sided or two-sided, fixed by the question's wording before the data arrive. Hypotheses are always about parameters (\(\mu\), \(p\)), never about statistics (\(\bar{x}\), \(\hat{p}\)). The pair must contradict - and the presumption of \(H_0\) is exactly what the coming topics' evidence must overcome.

Flashcards

Flashcard

H0

The null hypothesis: the default claim

No effect; carries the equality.

Test your understanding

Question 1 of 5

Which is a valid pair of hypotheses for testing a claimed mean of 500 g?

Question 2 of 5

A complaint says the boxes are underfilled (claim: 500 g). Which alternative?

Question 3 of 5

A test asks whether a machine's fills differ from 500 g, in either direction. Which alternative?

Question 4 of 5

Hypotheses in a test are always statements about:

Question 5 of 5

At the start of a hypothesis test, the null hypothesis is:

← Previous topicNext topic →