Null and Alternate Hypotheses

You need to know this because it gives you a reliable way to decide if what you see in data is actually meaningful or just a random coincidence. Imagine you launch a small change on your website, like moving a button from red to blue, and sales go up by 2%. Without this concept, you cannot tell if that increase was caused by the color change or if it would have happened anyway by chance. By using these two opposing statements, you gain the practical ability to separate signal from noise, allowing you to make confident business decisions rather than guessing based on lucky breaks or bad luck in your data.

Suppose you run a coffee shop and you're thinking of switching to a new espresso blend. You don't know if the new blend is better, worse, or the same as what you use now. To find out, you set up two competing stories about reality. The first story assumes nothing has changed: the customer satisfaction stays exactly the same. This is the null hypothesis—it is the boring, safe default assumption that your change did nothing. The second story assumes something has happened: the new blend has an effect, meaning customers will either prefer it or dislike it more than the old one. This is the alternative hypothesis.

In any test, you start by pretending the null hypothesis is true. You gather data to see how well this 'no change' story holds up. If the results you get are extremely unlikely to happen if nothing had actually changed, you reject the null hypothesis and accept the alternative. It is like a courtroom: the null hypothesis is 'innocent until proven guilty.' The defendant (the idea that nothing changed) is assumed innocent until you produce strong evidence against them. You do not need to prove the new blend is perfect; you just need to show that it is statistically impossible for the observed improvement to be just random luck. If the data does not meet that high bar, you fail to reject the null, meaning you stick with the assumption that nothing meaningful changed.

Imagine you go to a doctor with a vague symptom and they run a diagnostic test. The null hypothesis is that you are healthy. The alternative hypothesis is that you have the specific disease being tested for. The medical test is designed to look for signs of that disease. If the test comes back positive, you might reject the assumption of health and conclude you have the disease (though there is always a small chance the test gave a false alarm). If the test comes back negative, you fail to reject the assumption of health—you remain assumed healthy, but this is not proof that you are perfectly disease-free, just that the test did not find evidence of illness. This analogy works because it shows how we treat the 'no problem' state as the default and only change our mind when the evidence is strong enough to overcome that bias.

Caveat: Unlike a medical diagnosis where treatment might follow immediately, in statistics we rarely 'accept' the alternative as absolute truth; we merely say the data supports it more than the null does.

The null hypothesis is the default assumption that there is no effect or difference, while the alternative hypothesis is the claim that an effect or difference exists, which you support only if the data provides strong enough evidence to reject the default.

This concept sits next to probability, which measures how likely different outcomes are, giving you the tools to judge whether evidence is 'strong' or weak. It also connects to statistical significance, which is the specific threshold of evidence required before you reject the null hypothesis and declare your result meaningful.