How Do Cluster Analysis and Factor Analysis Improve Quality Management?

In quality engineering, understanding why defects occur often means finding hidden patterns in large datasets. Two powerful statistical tools help you do exactly that: cluster analysis and factor analysis. While both reduce complexity, they answer different questions. This article explains what they are, how they work, and how to apply them—using only standard, published methods.

---

What They Are

• Cluster analysis groups observations (e.g., products, batches, customers) so that items within a group are similar to each other and different from items in other groups. It is an unsupervised technique—you do not pre-label the groups. Common methods include k-means (partitioning) and hierarchical clustering (building a tree of nested groups). It answers: "Which units behave alike"

• Factor analysis reduces a large set of correlated variables (e.g., many measurement probes) into a smaller set of underlying latent factors. It answers: "Which variables move together, and what hidden dimension explains them" A standard rule for deciding how many factors to keep is the Kaiser criterion: retain only factors with an eigenvalue > 1 (Everitt, An Introduction to Applied Multivariate Analysis).

---

How They Work

### Cluster Analysis (k-means, simplified steps)

1. Choose the number of clusters, k (e.g., via the elbow method or prior knowledge).
2. Randomly assign k initial cluster centers.
3. Assign each observation to the nearest center (using Euclidean distance).
4. Recalculate each center as the mean of its assigned points.
5. Repeat steps 3–4 until assignments no longer change.

Hierarchical clustering instead builds a dendrogram by successively merging the two closest clusters (agglomerative approach), letting you cut the tree at any height to get the desired number of groups.

### Factor Analysis (with Kaiser criterion)

1. Compute the correlation matrix of your variables.
2. Extract initial factors (typically via principal component analysis).
3. Calculate eigenvalues—each represents the variance explained by one factor.
4. Kaiser rule: keep only factors with eigenvalue > 1. This ensures each retained factor explains more variance than a single standardized variable.
5. Apply rotation (e.g., varimax) to make factor loadings more interpretable—each variable loads strongly on one factor and weakly on others.

---

A Worked Illustrative Example

Example data (illustrative only) — not from any real factory.

Suppose you measure 5 quality parameters (hardness, thickness, gloss, density, and surface roughness) on 200 machined parts.

• Factor analysis yields 5 eigenvalues: 3.2, 1.1, 0.6, 0.5, 0.4. By the Kaiser criterion, you keep only the first two factors (eigenvalues 3.2 and 1.1 > 1). After varimax rotation, you find that hardness, density, and thickness load highly on Factor 1 (interpreted as "material strength"), while gloss and roughness load on Factor 2 ("surface finish"). You have reduced 5 variables to 2 meaningful dimensions.

• Cluster analysis on the same parts (using the two factor scores) with k-means (k=3) reveals three natural groups: Group A (high strength, poor finish), Group B (moderate both), Group C (low strength, excellent finish). Investigating further, you discover Group A corresponds to one production shift—pointing to a root cause you had not previously considered.

---

Common Pitfalls

• Kaiser criterion is a guideline, not a law. For borderline eigenvalues (e.g., 1.05), also check the scree plot or interpretability.
• Cluster results depend on scaling. Always standardize variables (mean 0, variance 1) before clustering, or variables with large units will dominate distances.
• Do not confuse the two tools. Cluster analysis groups rows (observations); factor analysis groups columns (variables).
• Rotation does not change model fit—it only changes the coordinate system for easier interpretation.

---

Try It Free

Ready to uncover hidden patterns in your own quality data? Use our free, browser-based tool for cluster and factor analysis—no coding required. It implements the Kaiser eigenvalue rule and common rotation methods described above.

? Launch the Cluster & Factor Analysis Tool