InsightFab
Knowledge Base/Analyze Phase: Selecting Statistical Tools for Root Cause Analysis
DMAIC6 min read

Analyze Phase: Selecting Statistical Tools for Root Cause Analysis

This article elucidates the "Analyze" phase through a real-world case study where a semiconductor factory's Cpk value plummeted to an alarm level. It demonstrates how to systematically apply statistical tools to sift through data, identify the true root cause, and resolve issues, rather than relying on intuition.

That Day, When the CPK Report Came Out, The Whole Room Fell Silent for Three Seconds

I remember years ago, our machine's Cpk suddenly dropped to 1.08. You must understand, this is an alarm-level figure in a semiconductor factory. In the meeting room, RD, Process, Equipment personnel looked at each other, the atmosphere so heavy it felt like the air had frozen. The boss's face was ashen, asking only one question: "Where's the problem?" At that moment, all we could do was pull out every statistical tool at hand, trying to find the "culprit" from a massive amount of data. The Analyze phase, simply put, is "detective work" – finding the root cause.

Where's the Problem?

Frankly, the core of the Analyze phase is to find the "true culprit" from a pile of seemingly related and unrelated data. It's not about wild guessing; it requires you to use scientific methods to list all possible suspects (i.e., "potential factors"), then eliminate them one by one until you find the "root cause" that truly led to the problem. In this phase, the biggest fear is drawing conclusions based on intuition; that's not solving the problem, but merely "trying your luck."

So, the key is to use statistical tools for "screening" and "validation." When your product yield drops from 99.99% to 99.38% (translating to DPMO skyrocketing from 100 to 6210), it's definitely not a "random event"; there must be an underlying pattern. At this point, you'll need basic tools like "histograms," "Pareto charts," and "scatter plots"; they can help you quickly "visualize" data and identify some initial "clues."

How It's Done in Practice

When Cpk drops to 1.08, our first step is usually "Pareto analysis." This tool is very simple: it arranges all defect modes that could cause Cpk to drop, from highest to lowest, based on their frequency of occurrence or impact. For example, we found that "film thickness non-uniformity" accounted for 70% of defects, followed by "particles" at 20%, with the remaining minor issues adding up to just 10%. At this point, you know your main effort should focus on resolving the "film thickness non-uniformity" issue.

Next, we use "scatter plots" to check if "film thickness non-uniformity" is related to certain process parameters. For instance, if we find a clear negative correlation between film thickness non-uniformity and "chamber pressure" (higher pressure leads to worse non-uniformity), then this is a strong clue. If you want to go further and understand the simultaneous impact of multiple factors on the result, or even if there are interactions between them, then "regression analysis" or "Analysis of Variance (ANOVA)" will be your good helpers.

In other words, these tools are not meant for an all-in bet; they're like a combination punch. First, use a Pareto chart to identify the main direction, then use scatter plots or histograms to deeply investigate individual factors, and finally, use regression or ANOVA to validate your conclusions, ensuring you haven't caught a scapegoat.

Most Common Pitfalls

Frankly, the biggest pitfall I've encountered is "drawing conclusions with insufficient data." Once, we saw a slight correlation between a certain parameter and yield, and we rushed to make changes. What was the result? After the change, the yield didn't improve; instead, we messed up another seemingly unrelated parameter, causing a bigger disaster. Only later did we realize that the initial sample size collected was too small, and the data noise was too high, leading us to misinterpret the correlation.

Another common issue is "only looking at the mean, not the variation." Sometimes, the average yield of two processes might be similar, but the yield fluctuation of one process is very large, sometimes very high, sometimes very low. If you only look at the mean, you'll miss this critical problem of "high variation." Therefore, when looking at data, you shouldn't just look at "central tendency" (mean, median), but also at "dispersion" (standard deviation, range).

Simply put, you cannot just look at the surface of the data; you must deeply understand the "story" behind the data to avoid these elementary mistakes.

One Thing You Can Do Today

Open Excel and plot a Pareto chart of your latest yield data!

Want to try it yourself?

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

Go to Tools