PCA score regression: the art of losing power

Technical Summary: PCA Score Regression: The Art of Losing Power

Problem Statement
The paper addresses a pervasive methodological issue in the analysis of functional data: the regression of subject-specific principal component scores (RPCS) on scalar covariates. This two-step approach—first performing Principal Component Analysis (PCA) or Singular Value Decomposition (SVD) to extract scores, then regressing these scores on covariates—is widely used in fields ranging from genomics and neuroimaging to physical activity monitoring. The authors argue that while RPCS is computationally convenient, it is statistically equivalent to fitting a misspecified model. This misspecification leads to a severe, often unrecognized loss of statistical power, inflated Type I error rates when multiplicity is ignored, and invalid inference regarding the true functional effects.

Methodology
The authors establish a rigorous theoretical framework comparing two approaches:

1. Function on Scalar Regression (FoSR): A correctly specified model where the functional outcome $W_i(s)$ is modeled directly as a function of scalar covariates $X_i$ and random effects:
   $$W_i(s) = \beta_0(s) + \sum_{q=1}^Q X_{iq}\beta_q(s) + \sum_{l=1}^{L_{true}} \xi_{il}\phi_l(s) + \epsilon_i(s)$$
   Here, $\beta_q(s)$ represents the fixed-effect functions of interest, and inference is conducted directly on these functions.

2. RPCS (Misspecified Model): The standard two-step procedure where PCA is first applied to the data matrix $W$ ignoring covariates to extract scores $\xi^M_{il}$ based on estimated eigenfunctions $\phi^M_l(s)$. These scores are then regressed on covariates:
   $$\xi^M_{il} = b_{0l} + \sum_{q=1}^Q b_{ql}X_{iq} + e_{il}$$

The core theoretical contribution is Theorem 1, which demonstrates that the regression coefficient $b_{ql}$ in the RPCS framework is not the true effect $\beta_q(s)$, but rather the inner product (projection) of the true effect onto the estimated eigenfunctions:
$$b_{ql} = \int_S \beta_q(s)\phi^M_l(s) ds$$

Key Contributions and Theoretical Findings
The paper provides five primary insights into the limitations of RPCS:

• Power Loss via Orthogonality: The power of RPCS to detect an association depends entirely on the correlation between the true effect $\beta_q(s)$ and the estimated eigenfunctions $\phi^M_l(s)$. If the true effect is orthogonal to the leading principal components (i.e., the integral above is zero), RPCS has zero power to detect the effect, regardless of the effect's magnitude.
• Signal Dilution: Even when not perfectly orthogonal, RPCS replaces a single test for the true function $\beta_q(s) = 0$ with $L_{miss}$ tests for smaller projected signals ($b_{ql}$). This fragmentation of the signal inherently reduces statistical power compared to the direct modeling in FoSR.
• Multiplicity and Inflation: Testing multiple coefficients $b_{ql}$ introduces a multiplicity problem. If not corrected, this leads to inflated $\alpha$-levels. If corrected (e.g., via Bonferroni), power is further reduced.
• Interpretability Issues: The coefficients $b_{ql}$ are difficult to interpret in the original data space. Reconstructing the functional effect $\hat{\beta}_q(s)$ from significant PCs alone can yield fundamentally different shapes depending on which PCs are selected, leading to biased or ambiguous biological interpretations.
• Inference Validity: Current RPCS methods do not provide valid inference for the true effect $\beta_q(s)$ because the hypothesis being tested ($b_{ql}=0$) is distinct from the scientific hypothesis of interest ($\beta_q(s)=0$).

Empirical Results
The authors validate these theoretical findings through extensive simulations and a real-data application:

• Simulations: In scenarios where the true effect is orthogonal or only partially correlated with the principal components, RPCS power drops to the level of the test size (no detection), while FoSR maintains high power. The simulations confirm that power loss is a direct function of the correlation between the true effect and the PCs.
• NHANES Application: Using minute-level accelerometry data from the National Health and Nutrition Examination Survey (2011–2014), the authors analyzed the effect of age on physical activity.
  • RPCS Results: Failed to detect a statistically significant age effect across the first four principal components (p-values > 0.05), leading to the conclusion that age has no effect on activity patterns in the 56–62 age group.
  • FoSR Results: Detected a significant negative effect of age on activity specifically during early morning hours (4 AM – 7 AM) and evening periods, with p-values < 0.001.
  • Reconstruction Failure: The authors demonstrated that attempting to reconstruct the age effect from the significant RPCS scores yielded a result that did not match the FoSR estimate, highlighting the inability of RPCS to recover the true functional shape.

Significance and Claims
The paper claims that the "PCA-then-regression" pipeline, while popular for its simplicity, is fundamentally flawed for hypothesis testing and effect estimation in functional data settings. The authors assert that:

1. RPCS is statistically equivalent to fitting a misspecified model, leading to a systematic loss of power that is often overlooked.
2. The method can produce false negatives, particularly when the true signal is small-to-moderate or not perfectly aligned with the directions of maximal variance in the data.
3. The problem is not limited to functional data but extends to any high-dimensional outcome analysis (e.g., genomics, imaging) where eigenvector scores are regressed on covariates.
4. Function on Scalar Regression (FoSR) offers a robust alternative that avoids these pitfalls by modeling the conditional mean of the functional outcome directly, providing valid inference and higher power.

The authors conclude that the convenience of RPCS comes at a significant cost to statistical power, signal reconstruction, and the validity of scientific inference, urging a shift toward direct functional modeling approaches like FoSR.