An anchored comparison keeps the nuisance sample on both sides, which is how you ask whether a PAIR is special rather than whether a SAMPLE is unusual
If you want to know whether two samples from one patient are unusually alike, compare them to pairs that include one of those same samples. Otherwise you are also measuring whether that patient's samples were unusual in some other way.
The same thing, for a scientist
Comparing a target pair's overlap statistic to a cohort distribution conflates two questions: is this pair special, and are its members atypical samples (in depth, quality, batch, or biology)? Anchoring answers only the first: for a pair (A,B), compare `s(A,B)` to `{s(A,X)}` and `{s(B,X)}` over comparators X, so that A or B appears on both sides of every comparison and its sample-level nuisances cancel. Applied here, the unanchored comparison put the same-patient pair at the 37th percentile of 27 pairs and the anchored, institution-matched comparison put it below all six matched comparators in all six analysis cells - and the difference between those two answers was entirely samples from a different hospital pairing badly with everything.
The worked contrast
Unanchored, on raw Jaccard, the same-patient pair BEATS 8 of 12 anchored comparisons across all six KMT2A::AFF1 comparators - which looks like support for the same-patient hypothesis. On excess over the matched null it beats 4 of 12. Split by institution:
Texas Children's comparators (excess) 1.532 1.560 1.469 1.406 1.397 1.361
St Jude comparators (excess) 1.189 1.165 1.200 1.160 1.333 1.353
the same-patient pair 1.329
The raw-Jaccard 'advantage' was the St Jude samples. Anchoring alone does not fix that; anchoring plus matching on the nuisance variable does.
The costs, which are real
- Anchored comparisons are not independent. They share a sample by construction, so a sign test over them overstates the evidence and its nominal p-value should be read as a direction, not a significance.
- Matching shrinks n. Going from 27 unrelated pairs to 6 institution-matched anchored comparisons puts the sign test's floor at 0.031, so no anchored analysis at this scale can be strongly significant however clean the direction.
- The matching variable must be chosen on independent evidence. Here the institution effect was measured on unrelated pairs only (p = 0.0016) before being used to define the matched set; choosing it because it produced the preferred answer would be the standard post-hoc failure, and this analysis is labelled post-hoc regardless.
- Anchored designs are what make a single-patient claim checkable at all. Without one, this corpus's same-patient conclusion would have rested on a comparator distribution that contained a measurable batch structure.
- Is there a permutation scheme that gets anchoring's confound control AND the full cohort's n - for instance permuting patient labels within institution and recomputing the whole comparator distribution?
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.review
Batch effects become a serious problem precisely when they are correlated with the outcome of interest, which is the general statement of why an unanchored, unmatched cohort comparison fails for a same-patient pair.
Tackling the widespread and critical impact of batch effects in high-throughput data · Leek JT, Scharpf RB, Bravo HC, Simcha D, Langmead B, Johnson WE, Geman D, Baggerly K, Irizarry RA · Nature Reviews Genetics 11:733-739 2010What this evidence cannot show
A review of the general problem in high-throughput data; the anchoring remedy described here is not from that paper.
Evidence that argues against this (1)
- Formal batch-correction methods (the review's own recommendation) would be an alternative to matching, and could recover more comparators - but correcting a batch effect estimated from 27 samples across 4 sites risks removing the signal along with the batch. Leek JT et al. 2010
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Measured on GSE252378: unrelated same-institution pairs have a higher excess over the matched null than cross-institution pairs (medians 1.456 versus 1.372, n = 83 versus 236, Mann-Whitney one-sided p = 0.0016), and a same-patient pair is same-institution by construction.
GSE252378: AutoCUT&RUN of KMT2A-rearranged primary leukemia samples (KMT2A N- and C-terminal antibodies, ENL, DOT1L, H3K4me3, IgG) · Janssens DH, et al. · NCBI Gene Expression Omnibus 2024What this evidence cannot show
Institution read from a free-text GEO description field; the effect could be a referral-population effect rather than a processing effect, which changes its interpretation but not the need to match on it.
Evidence that argues against this (1)
- The effect is modest (1.06x on the excess scale) and could in principle be a depth effect wearing an institution label, since the St Jude libraries are the shallowest in this deposit. Janssens DH et al. 2024
the design the post-hoc analysis uses
the nuisance variable that had to be matched
the comparator half of the same problem