Common Support for Batch-Adjusted Comparisons

Why Common Support Matters

A batch-adjusted condition comparison is grounded most directly in batches that contain samples from both conditions. Within such a batch, the two groups share the recorded batch label, so their observed difference can be formed without comparing one batch's treated samples with another batch's control samples. This is the practical meaning of common support for a discrete batch variable.

The restriction has a cost: it changes which samples, batches, and potentially which biological settings the comparison describes. If some batches contain only one condition, their samples cannot contribute a direct within-batch contrast. A result estimated from the remaining batches is not automatically an estimate for every sample originally collected. The key tasks are therefore to identify support before fitting or interpreting a comparison, and to state the target population that remains.

Finding Supported Batches and Samples

Start from a cross-tabulation of biological samples by batch and condition. A batch with a positive sample count in both condition columns supports a within-batch comparison. A batch with a zero in either column does not. Count biological samples, not cells, when deciding whether each condition is represented: thousands of cells from one donor or specimen do not create thousands of independent biological replicates.

This support check is distinct from asking whether the comparison is precise. A batch with one sample in each condition is structurally supported, but its contrast may be highly uncertain and sensitive to those two samples. Likewise, a large number of cells cannot repair a lack of biological replication. Common support answers whether both groups are present; it does not establish adequate sample size, balance, or reliable measurement.

Defining the Comparison on Common Support

Let \(Y_i\) denote a sample-level outcome for a specified gene or other prespecified quantity, \(b_i\) the batch for sample \(i\), and \(c_i\) its condition. A sample-level outcome could be a summary derived from that sample's cells; the inferential unit should still reflect biological replication. For each supported batch \(b\), define the condition contrast as the treated mean minus the control mean within that batch:

A summary across supported batches requires a choice of weights. For prespecified weights \(w_b\) satisfying \(w_b\geq 0\) and \(\sum_{b\in S}w_b=1\), where \(S\) is the common-support set, the supported contrast is

Equal weights give each supported batch the same influence. Other weights may reflect a defined target distribution of batches or samples, but must be stated: weighting by sample counts, for example, targets a different mixture from equal-batch weighting. Weights cannot create information in unsupported batches.

Why Within-Batch Cancellation Requires an Assumption

An additive model helps explain what a within-batch comparison removes. For a sample in batch \(b\) and condition \(c\), write

Here \(\mu\) is a baseline, \(\gamma_b\) is the batch contribution, \(\tau_c\) is the condition contribution, and \(\varepsilon_{ibc}\) is the residual. To derive the conditional-mean contrast from this model, assume explicitly that the residual has conditional mean zero in each observed batch-condition cell: \(E(\varepsilon\mid b,c)=0\). The model also assumes a common condition effect across batches; if effects differ by batch, a single \(\tau_c\) does not represent every within-batch difference.

The zero conditional-mean assumption is what permits each conditional expectation to omit its residual term. If instead residual means differ between conditions within a batch, the difference includes that residual-mean difference. Common support makes the within-batch contrast available, but does not by itself guarantee the model assumptions or eliminate all bias.

Diagnostics, Alternatives, and Limitations

Before interpreting a batch-adjusted comparison, inspect the batch-by-condition table and report both the supported batches and the excluded samples. Then examine the number of independent samples per cell, not just the binary presence of a group. Large imbalances, singleton cells, or a contrast driven by one batch are warning signs. Compare batch-specific contrasts as a diagnostic: substantial differences may indicate effect heterogeneity, residual imbalance, or unstable estimation, and should not be concealed by a single average.

For a discrete batch variable, restricting to common support is transparent and avoids pretending that a direct within-batch contrast exists in unsupported batches. A model-based analysis may retain more observations by imposing assumptions that permit comparisons across batches, but those comparisons rely on extrapolation where a condition is absent. Neither strategy is universally preferable: the choice depends on the target population and the credibility of the assumptions. Matching or reweighting can also alter the represented sample mixture, but cannot supply an observed comparison in a batch-condition cell with no samples.

Common support is limited in several ways. First, it is only support with respect to the recorded batch labels; other imbalanced characteristics may remain. Second, sparse overlap can yield noisy contrasts even when support technically exists. Third, an average over supported batches may not transport to unsupported batches, especially if treatment effects vary with batch or biological context. Finally, support is determined by the actual experimental design: adding cells to existing samples does not resolve a missing condition at the biological-sample level.