Hypothesis Testing: Null, Alternative, and How to Set Them Up

Hypothesis testing is the formal procedure researchers use to decide whether the patterns in their data are real or just the product of chance. You start with two competing statements about the world: a null hypothesis that says nothing is going on, and an alternative hypothesis that says something is. You then collect data and ask a single question. If the null hypothesis were true, how surprising would this data be? When the answer is "very surprising," you reject the null hypothesis in favor of the alternative. When the answer is "not surprising at all," you don't.

This guide covers what the null and alternative hypotheses are, how to write each one correctly, the difference between one-tailed and two-tailed tests, the step-by-step logic of a hypothesis test, a fully worked example using real published research, and the mistakes that most often trip up graduate writers. Get the setup right and everything downstream, from your p-value to your conclusion, follows cleanly. Get it wrong and the rest of the analysis rests on a cracked foundation.

What Is a Hypothesis in Statistics?

In everyday speech, a hypothesis is an educated guess. In statistics, the word is more precise. A statistical hypothesis is a specific, testable claim about a population parameter: a mean, a proportion, a difference between two groups, or a correlation. The claim has to be specific enough that you can collect data and check it.

Here's the key idea. You can almost never measure an entire population. You can't survey every doctoral student, test every patient, or interview every voter. So you work with a sample, and you use what you find in the sample to make an inference about the population. Hypothesis testing is the formal machinery for making that leap responsibly, with a built-in account of how much you might be wrong by chance.

Every hypothesis test pits two statements against each other. They have to be mutually exclusive, which means only one of them can be true, and they have to cover every possibility between them. Those two statements are the null hypothesis and the alternative hypothesis.

The Null Hypothesis (H0)

The null hypothesis, written H0 and read "H-naught," is the statement of no effect. It says the thing you're investigating isn't happening. There's no difference between the groups. There's no relationship between the variables. The treatment does nothing. The coin is fair.

The null hypothesis is your starting assumption. You treat it as true the way a court treats a defendant as innocent: not because you believe it, but because the burden of proof sits on the other side. Your data has to make a strong enough case before you're willing to give the null up. If the evidence is weak, the null stands, the same way a defendant goes free when the case isn't proven.

Here are null hypotheses written correctly for three common research situations.

• Comparing two groups. There is no difference in mean financial risk tolerance between men and women.
• Testing a relationship. There is no correlation between hours of sleep and exam performance.
• Testing a treatment. The new teaching method produces no change in average test scores compared to the standard method.

Notice the pattern. Each null hypothesis is a statement of zero: zero difference, zero correlation, zero change. That's what "null" means. It's also why hypothesis testing has a clean mathematical structure. A statement of exactly zero gives you a precise point to build a probability calculation around.

The Alternative Hypothesis (H1 or Ha)

The alternative hypothesis, written H1 or Ha, is the statement that something is going on. There is a difference. There is a relationship. The treatment works. The alternative is almost always what the researcher actually believes and set out to investigate. The whole study is usually built to find evidence for it.

For each null hypothesis above, here's the matching alternative.

• Comparing two groups. There is a difference in mean financial risk tolerance between men and women.
• Testing a relationship. There is a correlation between hours of sleep and exam performance.
• Testing a treatment. The new teaching method changes average test scores compared to the standard method.

Together, the null and the alternative cover every outcome. Either the means are equal or they aren't. Either there's a correlation or there isn't. There's no third option, which is exactly the structure a hypothesis test needs.

One point that confuses many writers: you never "prove" the alternative hypothesis, and you never "accept" the null. You either reject the null or you fail to reject it. That careful wording isn't academic fussiness. It reflects what the test can actually tell you, which is a point we return to below.

One-Tailed vs. Two-Tailed Hypotheses

When you write the alternative hypothesis, you're making a choice about direction, and that choice changes the test. There are two options.

Two-tailed (non-directional)

A two-tailed alternative says only that the groups differ, without saying which one is larger. "There is a difference in mean risk tolerance between men and women" is two-tailed. It allows for the possibility that men score higher or that women score higher. You're testing for a difference in either direction.

Most research uses two-tailed tests, and most reviewers expect them. A two-tailed test is the more conservative, more honest default because it doesn't assume you already know which way the effect runs.

One-tailed (directional)

A one-tailed alternative commits to a direction. "Men have higher mean risk tolerance than women" is one-tailed. You're only testing for an effect in that one direction, and you'd treat a strong effect in the opposite direction as a failure to reject the null.

One-tailed tests have more statistical power to detect an effect in the predicted direction, which sounds appealing. The catch is that you have to justify the direction in advance, on theoretical or prior-evidence grounds, not because the data turned out that way. Switching to a one-tailed test after seeing the results is a form of data dredging that reviewers and statisticians treat as a serious error. When in doubt, use a two-tailed test.

How a Hypothesis Test Works, Step by Step

Every hypothesis test, no matter how complex the statistics underneath, follows the same five-step logic. Here it is in order.

1. State the hypotheses. Write the null (H0) and the alternative (H1) before you look at the data. Decide whether the test is one-tailed or two-tailed at this stage, not later.
2. Set the significance level. Choose your threshold for "surprising enough," called alpha. The standard choice is 0.05, meaning you're willing to accept a 5% chance of wrongly rejecting a true null. Some fields use 0.01 for a stricter bar.
3. Collect data and calculate a test statistic. Run your study, then compute a single number (a t-value, an F-value, a chi-square value, and so on) that summarizes how far your sample result sits from what the null hypothesis predicts.
4. Find the p-value. The p-value is the probability of getting a result at least as extreme as yours if the null hypothesis were true. A small p-value means your data would be very unlikely under the null.
5. Make a decision. If the p-value is below alpha, reject the null hypothesis. If it's at or above alpha, fail to reject it. Then state what that means in plain terms for your research question.

The logic running through all five steps is a kind of proof by contradiction. You assume the null is true, work out what the data should look like under that assumption, and then check whether your actual data fits. When the data is wildly inconsistent with the null, you conclude the null was probably the wrong assumption. The test statistic in step three is typically standardized, which is why a working knowledge of z-scores and how standardized values work makes the whole procedure easier to follow.

A Worked Example: Risk Tolerance by Gender

Abstract steps are easier to grasp with real numbers, so let's walk through a complete test using a genuine research question. Fisher and Yao (2017) studied gender differences in financial risk tolerance, a well-documented finding in consumer economics. We'll build a simplified two-group hypothesis test in that spirit, using realistic figures to show how the five steps play out.

Suppose you survey graduate students and measure financial risk tolerance on a standardized scale where higher scores mean greater willingness to take financial risk. You want to know whether men and women in your sample differ on average.

Step 1: State the hypotheses

Because you have no strong basis to predict the direction in advance, you choose a two-tailed test.

• H0: The mean risk tolerance score is the same for men and women. In symbols, the two population means are equal.
• H1: The mean risk tolerance score differs between men and women. The two population means are not equal.

Step 2: Set the significance level

You set alpha at 0.05, the standard choice in the social sciences. You're accepting a 5% risk of concluding there's a gender difference when there really isn't one.

Step 3: Collect data and calculate a test statistic

Say your sample gives these results.

The observed difference in sample means is 0.5 points, with men scoring higher. To test whether a gap that size is more than chance variation, you run an independent-samples t-test, which produces a t-value. With these numbers, the test statistic comes out to roughly t = 3.32. The exact arithmetic belongs in a dedicated t-test guide. What matters here is that the t-value measures how many standard errors separate the two group means. A t-value near 0 means the groups look alike. A large t-value, positive or negative, means they look different.

Step 4: Find the p-value

A t-value of 3.32 with this sample size corresponds to a p-value of about 0.001. Read that carefully. It means that if there were truly no difference between men and women in the population, you'd see a sample gap this large or larger only about once in a thousand studies. That's a genuinely surprising result under the null.

Step 5: Make a decision

Your p-value of 0.001 is well below your alpha of 0.05, so you reject the null hypothesis. In plain language: the data provides strong evidence of a real difference in mean financial risk tolerance between men and women in this population, with men scoring higher on average. You did not prove it with certainty. You showed that the no-difference explanation fits the data poorly enough to set aside.

Note one thing the test doesn't tell you. A p-value of 0.001 says the effect is unlikely to be chance. It says nothing about whether a 0.5-point gap is large enough to matter in practice. Statistical significance and practical importance are different questions, and treating them as the same is one of the most common mistakes in applied research.

Common Mistakes in Setting Up Hypotheses

Most hypothesis-testing errors happen at the setup stage, before any calculation. Here are the ones that show up most often in graduate work, with the fix for each.

Putting the effect in the null hypothesis

The null is always the statement of no effect. If you find yourself writing "H0: the treatment improves outcomes," you've reversed them. The improvement belongs in the alternative. The null says the treatment does nothing.

Writing vague hypotheses

"There is a relationship between the variables" is too loose. Specify which variables, and specify the parameter you're testing. A reader should be able to tell exactly what would count as evidence for and against your claim.

Choosing a one-tailed test to chase significance

Switching from two-tailed to one-tailed after the results come in, because it pushes a borderline p-value under 0.05, is data dredging. Decide the direction before you see the data, and only when theory genuinely justifies it.

Saying you "accept" or "prove" the null

Failing to reject the null doesn't mean the null is true. It means you didn't find enough evidence against it, which can happen because there's no effect or because your study was too small to detect one. Write "failed to reject," never "accepted" or "proved."

Confusing statistical and practical significance

A tiny, unimportant effect can be statistically significant in a large sample, and a meaningful effect can miss significance in a small one. Always report and interpret the size of the effect alongside the p-value, not the p-value alone.

Where Hypothesis Testing Fits in Your Statistics

Hypothesis testing is the engine of inferential statistics. It's the framework underneath the specific tests you'll actually run: t-tests for comparing two groups, ANOVA for comparing three or more, chi-square tests for categorical data, and correlation and regression for relationships. Each of those is a different way of generating a test statistic and a p-value, but the five-step logic above is identical across all of them.

The reason the procedure works mathematically is that test statistics follow known probability distributions when the null is true. Many of those distributions are built on, or approximate, the normal distribution and its properties. That's what lets you translate a test statistic into a precise p-value. Once you're comfortable with the setup here, the individual tests become variations on a theme you already understand.

If any of those gives you pause, it's worth a second look before submission. The statistical writing in a dissertation or manuscript is exactly where a subject-matter editor earns their keep, because a reviewer who spots a reversed hypothesis or an unjustified one-tailed test will question everything that follows. Editor World's editors hold advanced degrees and read this kind of analysis daily across dissertation editing and journal article editing. You choose your own editor by field, so the person reviewing your statistics actually knows the conventions of your discipline.

Frequently Asked Questions

Page last reviewed: June 2026. Content reviewed and edited by the Editor World editorial team. Editor World, founded in 2010 by Patti Fisher, PhD, provides professional human-only editing, proofreading, and writing services for graduate students, academics, and researchers worldwide. 100% human editing, no AI at any stage. BBB A+ accredited since 2010 with 5.0 / 5 Google Reviews and 5.0 / 5 Facebook Reviews. More than 100 million words edited for over 8,000 clients in 65+ countries. Recommended by the Boston University Economics Department.