Choosing the Right Hypothesis Test

1. The Logic of Choosing a Statistical Test

Selecting a statistical test isn't about memorizing a long list of options. It's a structured process that starts with asking a few fundamental questions about your data and your goal.

To begin, let's watch a video that frames this selection process as a logical, step-by-step procedure, much like following a clinical guideline or a process flowchart.

As the video highlighted, the selection process boils down to three core questions, which we will now explore in the context of a manufacturing improvement project.

The 3 Key Questions for Test Selection:

1. What is the type of your data? (Specifically, the data for your response variable or CTQ).
2. What is the purpose of your analysis? (e.g., comparing groups, looking for relationships).
3. How many groups/samples are you analyzing, and what is their relationship? (e.g., two independent machines, three shifts, paired measurements before and after a change).

2. A Tool for the Job: The Hypothesis Test Decision Tree

Engineers love a good flowchart, and for good reason—it makes complex decisions simple and repeatable. Selecting a hypothesis test is no different. The image below is a decision tree that is an indispensable tool for any Green Belt.

We will use this "Hypothesis Test Decision Tree" as our guide. Let's break down the key branching points.

Decision Point 1: Data Type (Continuous vs. Discrete)

This is the first and most important split.

• Continuous Data (Interval/Ratio): Measurements that can take any value within a range. Examples: part diameter, cycle time, temperature, pressure. These tests often focus on comparing averages (means) or variation.
• Discrete/Attribute Data (Nominal/Ordinal): Data that is counted or categorized. Examples: pass/fail counts, defect type categories (e.g., scratch, dent, crack), shift number (1, 2, 3). These tests often focus on comparing proportions or counts in categories.

Decision Point 2: Analysis Purpose & Number of Groups

Your graphical analysis from the previous lesson will guide you here.

• Comparing one group to a standard: "Is our average fill volume still calibrated to 500ml?" (e.g., 1-Sample t-test).
• Comparing two groups: "Is Machine A faster than Machine B?" (e.g., 2-Sample t-test).
• Comparing more than two groups: "Is there a difference in defect rates across our three suppliers?" (e.g., ANOVA or Chi-Square).
• Looking for a relationship (Correlation/Regression): "Does furnace temperature affect material hardness?" We will cover these tests in detail in Module 5.

Decision Point 3: Parametric vs. Non-Parametric Tests

Notice how the "Continuous" path on our decision tree splits based on whether the data is "Normal." This leads to the choice between parametric and non-parametric tests.

To summarize the key points:

• Parametric Tests (like t-tests and ANOVA) are powerful but have assumptions, the most important being that your data is approximately normally distributed. You already learned how to check for normality using probability plots in Module 3.
• Non-Parametric Tests (like Mann-Whitney or Kruskal-Wallis) are the robust alternative when your data is not normal, or when you are working with ordinal (ranked) data. They are the "safer bet" when assumptions are violated.

3. Practical Application: Using the Decision Tree

Let's solidify this by walking through some common manufacturing scenarios. Refer to the "Hypothesis Test Decision Tree" image as we go.

Scenario A: Calibrating a Filling Machine

• Problem: A filling machine is supposed to dispense 2.0 Liters. You take a sample of 35 bottles and measure their volume to see if the machine's average is off-target.
• Applying the Tree:
  1. Data Type? Volume is Continuous.
  2. Objective? Comparing the process Mean to a target (2.0 L). This is a "1 Sample" problem.
  3. Distribution? We run a normality test and find the data is Normal.
  • Conclusion: The path leads us to the 1-Sample t-Test.

Scenario B: Comparing Two Suppliers

• Problem: You suspect that Supplier A provides raw material with a higher average purity than Supplier B. You measure the purity from 15 batches from each supplier.
• Applying the Tree:
  1. Data Type? Purity level is Continuous.
  2. Objective? Comparing the Means of 2 Samples (Supplier A vs. B).
  3. Distribution? Let's assume the data from both groups is Normal.
  • Conclusion: The path leads to the 2-Sample t-Test. (If the data were not normal, the path would lead to the Mann-Whitney Test).

Scenario C: Defect Analysis by Production Line

• Problem: You're analyzing final inspection data. You have defect counts categorized by three production lines and want to know if the proportion of defective units is the same across all three lines.
• Applying the Tree:
  1. Data Type? The data is Discrete (defective or not defective).
  2. Objective? We are comparing defect counts/proportions across more than two groups. This is a test for association between the categorical variable "Production Line" and the categorical variable "Defect Status".
  • Conclusion: The path leads to the Chi-Square Test.

4. A Final Detail: One-Tailed vs. Two-Tailed Tests

Once you've selected the test (e.g., a 2-Sample t-test), there is one final choice to make: the direction of your hypothesis. This determines whether you use a one-tailed or two-tailed test.

Here's a simple breakdown for manufacturing contexts:

• Use a One-Tailed Test when you only care about a change in one specific direction.

  • Question: "Does the new process reduce cycle time?"
  • Hypothesis: You are testing if New Time < Old Time. You don't care if it's simply different; you only care if it's an improvement.
  • This is common in improvement projects.

• Use a Two-Tailed Test when a difference in either direction is important.

  • Question: "Are the diameters of parts from Mold A different from Mold B?"
  • Hypothesis: You are testing if Diameter A ≠ Diameter B. A difference in either direction (larger or smaller) indicates a lack of consistency that may need to be addressed.
  • This is common in problem-solving or validation scenarios where you're comparing a process to a standard or two processes to each other.

Conclusion

In this lesson, we established a formal process for moving from visual clues to statistical validation. You now have a framework for selecting the correct hypothesis test, which is a cornerstone of the Analyze phase of any Six Sigma project.

Key Takeaways:

• Selecting the right hypothesis test is a logical process based on your data type, analysis objective, and the number/nature of your samples.
• The Hypothesis Test Decision Tree is your roadmap for navigating this selection process.
• The choice between parametric (e.g., t-test, ANOVA) and non-parametric (e.g., Mann-Whitney) tests depends primarily on whether your continuous data meets the assumption of normality.
• The choice between a one-tailed and two-tailed test depends on whether you are testing for a change in a specific direction (improvement) or for any difference from a standard.

Preview of the Next Lesson:

Now that you know how to choose the right tool, it's time to learn how to use it. In our next lesson, we will perform our first formal statistical analyses. We will focus on the versatile family of t-tests, learning how to perform and interpret t-tests (1-sample, 2-sample, and paired) to compare process means.