Batch / Plate Confounding Detection

Batch confounding occurs when experimental group and batch assignment are correlated: all control samples on Plate 1, all treated samples on Plate 2. Any plate-level variation (temperature, reagent lot, technician, day) becomes indistinguishable from the treatment effect. The apparent "treatment effect" may be entirely due to batch, and there is no statistical correction that can disentangle perfectly confounded designs.

STEP 1 —The Pitfall

In practice, batching by group is the path of least resistance: run all controls together, run all treatments together. The problem is invisible in the data —the values look like a clean effect. It only becomes apparent when someone checks whether the batch assignment was randomized with respect to group.

Common scenarios: all samples from one condition processed on Monday, all from another on Friday; different technicians running different groups; reagent lots changing between groups. Each creates a potential confounder that is undetectable if not recorded —and uncorrectable if perfectly confounded.

STEP 2 —Journal Requirement

ARRIVE 2.0 requires reporting of randomization procedures including batch allocation. RNA-seq and metabolomics guidelines universally require batch correction documentation and verification that batch is not confounded with group. Nature Protocols and Methods Enzymology guidelines require that batch information be recorded and reported, and that confounding be addressed in the analysis.

STEP 3 — What Licklider Currently Provides

What the current product does

• The batchConfounding gate summarizes structural batch / plate risk from contract columns (batch, plate, run-order) and row distribution across groups.
• Confounding severity is classified as complete, partial, or none based on whether group assignment is concentrated within batch levels.
• Card 36 presents the user with disclosure options: disclose the batch/plate structure and continue, keep batch shape/color/facet visible, or fall back to descriptive-only output.
• The user's choice affects export-facing disclosure and can influence whether claim-bearing output is treated as ready.

What the product does not do

• The product does not run a chi-squared or Fisher's exact test on the group-by-batch crosstable.
• The product does not prove or disprove a causal confounding relationship.
• The product does not offer ComBat, limma removeBatchEffect, or other batch-correction execution.
• A dedicated confounding heatmap panel is not part of the current product UI.

Known limitations

Batch confounding detection is based on structural hints (contract columns and row distribution), not on statistical testing of the group-by-batch association. The product surfaces risk and records disclosure, but the researcher must independently verify whether confounding is present and decide on appropriate correction methods.

For the detailed quality-checks documentation, see Batch and Plate Confounding Detection in the quality-checks docs.

STEP 4 — Draft Output (Draft / Needs review)

The following is an example of how a researcher might write a batch confounding disclosure, not output that Licklider generates automatically.

Samples were processed across two experimental batches. Batch assignment
was partially correlated with treatment group. Batch was therefore
included as a covariate in the primary analysis. Sensitivity analysis
without batch correction is provided in Supplementary Table 4.