Hypothesis Testing

Hypothesis testing turns a vague worry like "is this machine off-target?" into a number you can compute and compare against a cutoff. You start from a default claim, the null hypothesis $H_0$, and ask: if $H_0$ were true, would data this extreme be unusual enough to doubt it? The method never proves $H_0$ true or false; it only measures how inconsistent the sample looks with the null model.

The Formula And Its Symbols

For a one-sample mean test with known population standard deviation, the test statistic is

where $\bar{x}$ is the sample mean, $\mu_0$ is the value claimed by $H_0$, $\sigma$ is the population standard deviation, and $n$ is the sample size. The denominator $\sigma/\sqrt{n}$ is the standard error, the typical spread of sample means. The test statistic itself depends on the situation: a $z$-test, $t$-test, chi square test, and many others are all hypothesis tests, so there is no single formula for all of hypothesis testing.

Why The Formula Works

The statistic answers a simple question: how many standard errors is the sample mean away from the null value? The numerator $\bar{x} - \mu_0$ is the raw gap between what you saw and what $H_0$ predicted. Dividing by the standard error rescales that gap into a universal unit, so a $z$ of $-2$ always means "two standard errors below the claim," whatever the original units were. Because sample means cluster tightly around the true mean (the standard error shrinks as $n$ grows), a large $|z|$ is hard to produce by chance alone if $H_0$ holds, and that is exactly what makes it evidence.

This also explains the surrounding machinery. Every test pits two statements against each other:

1. The null hypothesis $H_0$, the default claim being tested.
2. The alternative $H_1$ or $H_a$, what you support if the data argues strongly enough against $H_0$.

You fix a significance level $\alpha$, often $0.05$, before looking at the result. It is the amount of evidence you demand before rejecting $H_0$. Two outcomes follow: reject $H_0$ when the data is sufficiently inconsistent with the null model, or fail to reject $H_0$ when it is not strong enough to rule the null model out. "Fail to reject" is not "accept as true"; it only means the sample did not provide strong enough evidence against $H_0$.

Worked Example, Step By Step

A filling machine is supposed to average $500$ mL per bottle. A quality-control team samples $36$ bottles and gets a sample mean of $496$ mL. Assume the population standard deviation is known, $\sigma = 12$ mL, and the conditions justify a one-sample $z$-test.

Set up the hypotheses:

This is a left-tailed test, since the concern is underfilling. Compute the standard error:

Then the test statistic:

For $\alpha = 0.05$ on a left-tailed $z$-test, the critical value is about $-1.645$. Because $-2 < -1.645$, the result falls in the rejection region, so reject $H_0$ at the $5\%$ level. In context, the sample provides evidence that the machine is underfilling on average. That conclusion still depends on the test assumptions; weak assumptions can make it unreliable even when the arithmetic is correct.

Try It Yourself

Take the same bottle-filling setup but change the sample mean to $498$ mL. Recompute the standard error (it is unchanged), then the test statistic, and decide at $\alpha = 0.05$ whether the decision flips. You should find the evidence weakens as the sample mean slides toward $500$, which shows how the gap in the numerator drives the whole result.

Calculation Pitfalls

• Misreading the tail. A left-tailed test compares $z$ against a negative critical value; flipping the sign sends you to the wrong region.
• Wrong denominator. The standard error is $\sigma/\sqrt{n}$, not $\sigma$ alone. Forgetting the $\sqrt{n}$ inflates the standard error and shrinks $|z|$.
• Misreading the $p$-value. A $p$-value is the probability, assuming $H_0$ is true and the assumptions hold, of a result at least as extreme as observed. It is not the probability that $H_0$ is false, nor a vague "happened by chance," nor the size of the effect.
• Type I vs Type II error. A Type I error rejects a true $H_0$ (its probability is controlled by $\alpha$); a Type II error fails to reject a false $H_0$ (probability $\beta$). Lowering $\alpha$ cuts false alarms but can make true effects harder to detect, which is why sample size matters.
• Significance is not importance. A tiny effect can be statistically significant in a very large sample, and a clean-looking $p$-value never rescues a test whose independence, distribution, variance, or data-type assumptions are wrong.

When Hypothesis Testing Is Used

It appears in science, manufacturing, medicine, surveys, A/B testing, and policy analysis, always to decide whether a sample gives enough evidence to question a default claim. Good testing is not only the calculation: it needs a sensible null hypothesis, a defensible design, and an interpretation that matches what the test can actually say.