What Is Hypothesis Testing? Definition and Uses

Hypothesis testing is the formal procedure for deciding whether a difference you observe in data reflects a genuine change in the process or merely the ordinary noise every process produces. It begins by assuming nothing has changed — the null hypothesis — and then asks whether the evidence is strong enough to overturn that assumption. In Six Sigma, it is the referee that separates real improvement from wishful thinking.

How it works

Every hypothesis test has the same skeleton. You state a null hypothesis, which says there is no difference — the new method performs like the old one, the two machines produce the same average, the training changed nothing. You state an alternative hypothesis, which says a difference exists. You choose a significance level before collecting data, conventionally 5 percent, which caps how often you are willing to declare improvement when nothing actually changed. Then you gather the data, compute a test statistic, and obtain a p-value: the probability of seeing results this extreme if the null hypothesis were true. A small p-value means the data are hard to explain as noise, and you reject the null.

Inside DMAIC, hypothesis tests carry two loads. In Analyze, they verify suspected root causes — does the defect rate really differ by shift, by supplier, by machine? In Improve, they confirm that a piloted fix produced a real change rather than a lucky sample. Teams that skip the test tend to install fixes that quietly evaporate within a quarter.

A worked example

Imagine a regional insurer piloting a redesigned claims intake form. On a sample of claims processed the old way, average cycle time is 12.4 days; on a pilot sample using the new form, it is 11.1 days. The gap looks encouraging, but samples bounce around. The team runs a two-sample test: the null hypothesis says the forms perform identically and the 1.3-day gap is chance. Given the variation and sample sizes involved, the test returns a small p-value — a gap this large would be rare if nothing had changed — so the team rejects the null and rolls out the form with evidence, not enthusiasm, behind it. The numbers here are illustrative; the discipline is not.

• Set the significance level before you see the data, not after — moving the goalposts invalidates the test.
• Failing to reject the null is not proof of no difference; it may simply mean your sample was too small to detect one.
• Statistical significance is not practical significance — a real difference can still be too small to be worth acting on.
• Test one clearly stated question; running dozens of comparisons and reporting the winner manufactures false discoveries.
• Check that your sample actually represents the process — a biased sample defeats any test run on it.

A hypothesis test is the moment a team stops trading opinions and starts weighing evidence.

Hypothesis testing anchors the Analyze phase of our Green Belt program — $299 for 35 hours of training and a 100-question proctored exam, with a 70 percent passing score and one free retake. Black Belt training extends the same logic into ANOVA, regression, and designed experiments.