InsightFab
Knowledge Base/ANOVA: Is There Really a Difference Among 4 Machines—It's Not Judged by Sight
6 Sigma8 min read

ANOVA: Is There Really a Difference Among 4 Machines—It's Not Judged by Sight

In a factory with 4 injection molding machines, M2's average dimension was 0.3mm higher than M1. When questioned by a supervisor about the randomness of this difference, ANOVA provides the statistical framework to let the data speak for itself.

Scenario

In the factory, there are 4 injection molding machines. You measured 15 product dimensions from each machine, created a table, and found that the average value of M2 was 0.3mm higher than M1. You went to your supervisor and said, "M2 seems to have a problem."

The supervisor asked, "How do you know that difference isn't random?" You fell silent.

Plain Language Explanation

The question ANOVA (Analysis of Variance) addresses is:

Are the differences between these groups real, or are they caused by random fluctuations?

It breaks down "total variance" into two parts:

  • Between-group variance: The differences between the means of each group (if there's a difference between machines, this will be large).
  • Within-group variance: The dispersion of individual measurements within the same group (random noise).

F-value = Between-group variance ÷ Within-group variance. If the F-value is large enough, it indicates that the differences between machines are far greater than random noise, and then the P-value will be less than 0.05.

Practical Application

ANOVA only tells you "if there is a difference"; it doesn't tell you "which groups differ from each other." To know which machine has a problem, you need to perform a post-hoc test, with Tukey HSD being commonly used.

Assumptions for ANOVA:

  1. Data in each group follows a normal distribution (tested using Anderson-Darling test).
  2. Variances of each group are equal (tested using Levene's Test).
  3. Observations are independent.

If the assumptions are not met (e.g., severely skewed data), switch to the non-parametric Kruskal-Wallis Test.

How InsightFab Does It

In Minitab, you need to set "response variable" and "factors," then review the F-value and P-value tables, and separately run a Tukey comparison. After uploading grouped data to InsightFab, it directly provides a visual report, showing which groups have significant differences, labeled using the connecting letters method (A, B, AB), making it clear at a glance.

Golden Quote

"Comparing averages is not engineering analysis; it's gambling on intuition. ANOVA is the tool that lets the data speak for itself."

Want to try it yourself?

Every tool mentioned in this article is available on InsightFab — just upload a CSV to analyze.

Go to Tools

More in this category

P-value: It's Not a "Pass/Fail" Switch

P = 0.043, and you declare it "effective," updating the SOP with new parameters. Three months later, the issue re-emerges, and you find the original sample size was only 12. P-value conveys significantly less information than commonly assumed.

Pareto Principle: The Problem You're Solving Might Not Be the Most Important One

In a QA meeting, individual efforts to improve specific areas led to only a 3% reduction in defect rate after three months. By contrast, using a Pareto chart to focus all resources on the "solder bridge" issue resulted in a 22% decrease in total defects over the same period.

Linear Regression: Finding the Process's "Control Lever"

When curing oven temperature is adjusted, product strength changes, yet the relationship remains undefined, leading to adjustments based solely on expert intuition. Linear regression provides a formula to predict product strength based on temperature.