PCA Diagnostics for Batch Effects Before and After Correction

Why Diagnose Batch Effects With PCA?

Technical batches can create systematic differences between samples. If those differences are large relative to other sources of variation, samples from the same batch may cluster in a principal-component (PC) score plot. Such a pattern can help identify batch as a prominent feature of the data, and comparing plots before and after correction can show whether the pattern changes. PCA is a diagnostic, however, not a test that proves a batch effect is present or that correction has succeeded.

A PCA diagnostic is useful only when the data and analysis are defined clearly. The rows should represent the same samples before and after correction, and the columns should represent the same features. For a meaningful comparison, use a consistent feature set and documented transformation, centering, scaling, and block-weighting choices. The correction may change the values, but changing other analysis choices at the same time makes it harder to tell what caused a changed plot.

For example, a PC1-versus-PC2 score plot shows each sample's coordinates on the first two principal components. Colors can identify batch, while a second visual encoding, such as point shape, can identify a biological group. The plot is most informative when batch labels and relevant biological covariates are available and displayed together.

What PCA Variance Summaries Mean

Let \(Z\) be the centered data matrix, after any selected feature scaling, and let \(S=Z^\mathsf{T}Z/(n-1)\) be its sample covariance matrix. The eigenvalues of \(S\), written \(\lambda_1 \geq \lambda_2 \geq \cdots \geq 0\), are the variances along the corresponding principal components. The proportion of total variance explained by PC \(j\) is its eigenvalue divided by the sum of all eigenvalues.

The justification follows directly from the definition of a principal component. For a unit vector \(v\), the projected scores are \(t=Zv\), and their sample variance is \(t^\mathsf{T}t/(n-1)=v^\mathsf{T}Sv\). The first principal-component direction maximizes this quantity over unit vectors; subsequent directions maximize it subject to orthogonality to earlier directions. The eigenvalues are these successive variances, and their sum is the total variance because the trace of \(S\) is the sum of its eigenvalues.

A scree plot displays eigenvalues or explained-variance proportions by component. It helps show whether a small number of directions account for much of the observed variation. It does not identify the source of that variation: a high PC1 proportion might reflect batch, biology, an outlier, or a combination.

Assessing Batch Alignment in Score Plots

Start by coloring samples by batch and looking for separation, gradients, or clusters across the leading PCs. Then check whether those patterns also align with biological covariates, such as phenotype or collection time. Batch-aligned clustering is a warning that technical variation may be prominent in the displayed directions; overlap between batches is not proof that batch has no effect, because an effect may lie in later PCs or be obscured by other variation.

A descriptive summary for a single PC score is the fraction of its total sum of squares explained by batch. For scores \(t_i\), let \(\bar{t}\) be their mean, and let \(\bar{t}_g\) be the mean in batch \(g\). The between-batch sum of squares divided by total sum of squares gives the usual one-factor \(R^2\). It ranges from 0 to 1 when the total sum of squares is positive. This is a descriptive measure of score separation; it is not the proportion of all feature-level variation caused by batch.

This summary is most interpretable when the batch labels are accurate and the samples provide relevant comparisons across batches. If biological groups are imbalanced across batches, report that imbalance and inspect the biological labels as well. A batch-associated score pattern cannot, on its own, distinguish technical effects from biology that is aligned with batch.

Assumptions, Confounding, and Comparisons

PCA describes linear directions of variation in the matrix supplied to it. Its diagnostics depend on the input: centering, feature scaling, filtering, and block weights affect which directions dominate. The usual covariance-based calculation also requires numerical values for the analyzed entries; missing data need an explicitly chosen handling strategy before applying it. Extreme samples can strongly influence the leading components, so inspect the score plot for isolated points and check whether conclusions depend on them.

The main interpretive requirement is adequate overlap between batches and biological groups. If every case is in one batch and every control in another, batch and phenotype are confounded. No PCA plot can identify which factor explains their shared pattern. The same caution applies if only a subset of groups appears in some batches. When groups are represented across batches, color and shape can help assess whether a batch pattern remains after considering the biological labels, but PCA still does not isolate causal sources of variation.

PCA diagnostics complement, rather than replace, other checks. A score plot summarizes sample-level structure in a few directions; it can miss batch patterns outside the plotted PCs. A batch association summary for a PC adds a numerical description but remains limited to that score and does not adjust away confounding. Feature-level comparisons can reveal changes not apparent in sample-level projections, while multivariable adjustment methods use a specified model and assumptions. Neither a plot nor a single summary should be treated as a substitute for understanding the design and the correction model.

Comparing Before and After Correction Carefully

Use the same samples and selected features where possible, and keep the transformation, scaling, and weighting choices consistent. Fit PCA to each matrix and examine the score plots, scree summaries, and batch association in several leading PCs, not only PC1. Because PCA is refitted on each matrix, its axes can rotate, change sign, or exchange order. Thus, a score shift on a separately fitted PC1 is not automatically a directly comparable change in a fixed direction.

For a clearer comparison, describe what is stable and what changes: for example, whether batch-colored clusters are less distinct in the displayed projections, whether batch-associated \(R^2\) values change, and whether biological labels still show structure. If correction is assessed using a different set of features or newly chosen scaling in each analysis, say so; those choices may contribute to the apparent difference.

Common Mistakes and Limitations

• Calling less visible separation proof of successful correction. A batch pattern may remain in later PCs, while biology may also have been reduced. Inspect multiple components and biological labels; PCA alone cannot establish preservation of relevant biological effects.
• Comparing PC1 coordinates as if they were on a fixed axis. Separate PCA fits can produce different directions and signs. Compare the patterns and summaries in context, not the raw coordinates as though the axes were identical.
• Ignoring scaling and block weights. A block with greater numerical variance or influence can dominate the leading components. Interpret the PCA relative to the actual preprocessing choices.
• Attributing a confounded pattern to batch. If batch and phenotype coincide, the observed separation cannot distinguish them. The study design, not a more detailed plot, must provide the missing overlap.
• Relying only on PC1 versus PC2. Important structure may occur in later components or may not be captured by linear PCA. Examine variance summaries and appropriate additional diagnostics.

The practical value of these diagnostics is that they expose prominent structure and make assumptions visible. Their limitation is equally important: PCA is a projection of the observed data, not a causal attribution procedure or a guarantee that the correction preserves the signal of interest.