Abstract: In principal component analysis (PCA), the first few principal components possibly reveal interesting systematic patterns in the data, whereas the last may reflect random noise. The researcher may wonder how many principal components are statistically significant. Many methods have been proposed for determining how many principal components to retain in the model, but most of these assume non-standardized data. In agricultural, biological and environmental applications, however, standardization is often required. This article proposes parametric bootstrap methods for hypothesis testing of principal components when variables are standardized. Unlike previously proposed methods, the proposed parametric bootstrap methods do not rely on any asymptotic results requiring large dimensions. In a simulation study, the proposed parametric bootstrap methods for standardized data were compared with parallel analysis for PCA and methods using the Tracy–Widom distribution. Parallel analysis performed well when testing the first principal component, but was much too conservative when testing higher-order principal components not reflecting random noise. When variables are standardized, the Tracy–Widom distribution may not approximate the distribution of the largest eigenvalue. The proposed parametric bootstrap methods maintained the level of significance approximately and were up to twice as powerful as the methods using the Tracy–Widom distribution. SAS and R computer code is provided for the recommended methods.
Introduction
Principal Component Analysis (PCA) is a fundamental tool for dimensionality reduction, allowing researchers to separate systematic "signal" from random "noise." In many scientific fields, such as agriculture and biology, variables are measured in different units (e.g., mass, length, concentration), necessitating standardization—scaling data to zero mean and unit variance—before analysis.
However, a significant gap exists in statistical methodology: most classical tests for determining the number of significant components (such as the Tracy–Widom distribution or the Kaiser rule) were developed for non-standardized data. When applied to standardized data, these tests often fail to accurately identify meaningful patterns.
The Impact of Standardization
Standardization fundamentally alters the distribution of eigenvalues, which represent the variance explained by each component. Traditional reference distributions, like Tracy-Widom, do not account for the constraints imposed by scaling every variable to a variance of one.
As shown in Figure 1, the distribution of the largest eigenvalue shifts depending on the scaling method. Using a standard reference distribution on scaled data is mathematically inappropriate and leads to "conservative" results, where genuine scientific signals are mistakenly discarded as noise.
Methodology: The Parametric Bootstrap
To address this, the authors propose a Full Parametric Bootstrap approach. This method is specifically designed to handle the nuances of standardized data and remains accurate even for small-sample sizes typical in field trials. The process involves:
- Null Hypothesis Modeling: Assuming the data consists of a fixed systematic structure plus random normal noise.
- Simulation: Generating thousands of synthetic datasets that mimic the original data's structure, applying the exact same standardization process to each.
- Sequential Testing: Evaluating components one by one ($K=0, 1, 2...$) until a component is found that is not statistically significant.
Results and Discussion
The proposed method was validated using three real-world datasets (Peanut, Sparrow, and Fish) and extensive simulations. A key finding was that Parallel Analysis, a common industry standard, is often too conservative for higher-order components. It frequently misses real signals because it fails to account for how the variance captured by the first component affects the remaining data.
Case Study: The Peanut Dataset
In the analysis of the Peanut dataset—measuring yields of 10 genotypes across 15 environments—the new bootstrap method revealed three significant principal components. Existing methods might have suggested only one or two.
As visualized in Figure 4, these additional components provide critical insights into genotype-by-environment interactions. Identifying the third component changed the scientific conclusion regarding which peanut genotypes performed best in specific environmental conditions.
Conclusion
The parametric bootstrap method provides a more powerful and reliable alternative for researchers working with standardized data. It maintains strict Type I error control and is up to twice as powerful as Tracy-Widom-based methods. To facilitate adoption, the authors have provided R and SAS code, ensuring these improved diagnostic tools are accessible to the broader scientific community.
Summary updated