Skip to content
Six Sigma | Part 6 of 10

Hypothesis Tests and Regression in Plain Language

How a t-test separates real differences from chance, how regression puts a number on a relationship, and how to explain both to a manager.

Statistics often fails in the meeting, not in the calculation. A green belt reports “p equals 3.4 times ten to the minus seven” and the room goes quiet. The analysis was correct and the message was lost.

From many ideas to one hypothesis

In the bore example, the team collected 23 possible causes in a cause and effect diagram. Then it split the baseline data by every context variable it had recorded: shift, setter, tool count, lot and hours since the start of the shift.

One split stood out. Oversize bores clustered in the first hour after every start. That turned 23 ideas into one testable hypothesis: the bore is larger while the machine warms up.

The test

The team compared 30 parts from the first hour with 60 parts from the rest of the shift. Means: 24.019 mm against 24.010 mm, a difference of 0.009 mm. A two-sample t-test gave a p-value below 0.001.

The plain sentence for the champion: if there were no real difference, a gap this large would appear by chance in fewer than 1 of 1,000 comparisons. We treat it as real. It also matters in practice: 0.009 mm is 30 percent of the half tolerance.

Choose the test by the question

Question Test
Do two groups differ in their mean? two-sample t-test
Do several groups differ? analysis of variance
Do two groups differ in spread? F-test or Levene test
Are defect shares different? test of proportions, chi-square
Is there a relationship? correlation and regression

Regression puts a number on it

The warm-up hypothesis pointed to temperature. The team recorded spindle temperature and bore diameter for 30 parts. Regression gave a slope of 0.0012 mm per degree Celsius with an R squared of 0.68: temperature explained about two thirds of the variation in this sample.

Between a cold start at about 22 °C and a warm spindle at 36 °C, the bore grows by about 0.017 mm. That is more than half of the upper half tolerance.

Four rules

  • Write the hypothesis first. Patterns found after looking at all the data are often chance.
  • Separate statistical from practical significance. With enough data, tiny and useless differences become significant.
  • Stay inside the measured range. The regression is valid between 21.5 and 35.8 °C, not beyond.
  • Correlation is not cause. Temperature rises with time since the start, and tool wear runs alongside it.

Verify at the process

The last step is not statistical. The team switched the suspected cause on and off: with a warm-up cycle before the first part, the first-hour shift disappeared. Without it, the shift came back. Only then was the warm-up accepted as a root cause.

Tests and regression narrow the search. Switching the cause on and off at the process confirms it.

Newsletter

One Lean idea.
Every two weeks.