DNA Evidence: What the Match Can and Cannot Support

DNA carries a reputation no other forensic discipline enjoys: the gold standard, the test that cannot be wrong. Part of that reputation is earned. For a complete profile from a good single-source sample, comparison is objective and the random-match probability is genuinely astronomically small. The President's Council of Advisors on Science and Technology, reviewing the field in 2016, found single-source and simple two-person mixtures foundationally valid. That is the strong core, and an honest witness does not give it away.

The difficulty is what the strength attaches to. Forensic scientists work with a hierarchy of propositions: sub-source (whose DNA is this?), source (what body fluid or tissue is it, and whose?), activity (how and when did it get there?), and offence (did this person commit the crime?). A random-match statistic answers only the bottom rung. It is a statement about the origin of a profile, not about an act.

So the honest frame for the whole discipline is this. DNA is often the strongest single piece of evidence in a case, and the sub-source claim, whose profile this is, deserves the confidence it usually gets. But that confidence belongs to one rung of a ladder, and it does not transfer upward on its own. The examiner who lets an overwhelming sub-source match be heard as proof of what the defendant did, or who offers a confident profile from a few picograms of degraded, mixed material, has climbed past what the evidence supports. The sections that follow are the places where that climb happens: counting contributors, the software, the reproducibility, the statistic, the transfer, the contamination, and the wording.

Real casework is not a clean reference swab. A large share of crime samples are mixtures of two or more people, and the first hard question is the most basic: how many contributed? Often the honest answer is that you cannot be sure. The common counting method, taking the locus with the most alleles and dividing, systematically undercounts. Hinda Haned and colleagues showed that for five-person mixtures this maximum-allele-count method is wrong the overwhelming majority of the time, correct in as few as 2 to 5% of cases even with a full set of markers. The method can only miss people; it cannot over-count.

The more capable tools do better but hit a floor. Testing the probabilistic method NOCIt against the alternatives, Lauren Alfonse and colleagues found it needed about 0.07 nanograms from a contributor, the DNA of roughly ten cells, to recover the true number reliably. Below about 0.01 nanograms of minor DNA, no method exceeded 40% accuracy. At the low template levels where these cases are actually contested, the number of contributors is frequently unknowable.

None of this makes mixture interpretation worthless. A clean two-person mixture with a clear major and minor contributor and ample template is routinely and reliably resolved; the counting methods agree and the deconvolution is sound. The Association's own guidelines require each laboratory to establish and validate its thresholds precisely so the choices are documented and defensible. The honest position is narrower than the reputation. On a good two-person mixture the interpretation is strong. On a low-template, three-plus-person, high-drop-out mixture, the examiner cannot always say how many people are present, cannot be sure which peaks are real, and is working with a profile whose shape depended on a cut-off the laboratory chose.

When a mixture is too complex to interpret by hand, laboratories now feed it to probabilistic genotyping software: STRmix, TrueAllele, EuroForMix and others. The programs build a statistical model of the mixture and return a likelihood ratio. This is genuine progress over the old subjective methods, and on validation data the leading programs perform well. But two facts about how they work matter in court, and neither is obvious from the single number on the report.

The first is that the number is not fixed. These systems sample randomly as they search, seeding from the computer's clock, so, as the STRmix developers themselves state, one should expect a degree of variation between runs of the same profile on the same data. And the answer moves with the operator's choices. Change the population-substructure value a single input, and one program's likelihood ratio shifts by more than two orders of magnitude on identical data. The figure in the report is one draw, under one set of settings, not a measurement.

The developers' answer is not unreasonable, and the section states it fairly: reliability, they argue, is shown by extensive validation on real and simulated data, not by reading code, and the source code is a legitimate trade secret. Validation does establish that the programs work well within the conditions tested. But validation and code review answer different questions. Validation shows the program gave the right answer on the samples the makers chose; it cannot show the code correctly implements the method on the next sample, which is precisely how the Forensic Statistical Tool's hidden function survived years of validation. A New Jersey appellate court, in State v. Pickett, held that a defendant who shows a particularised need is entitled to the source code under a protective order. For most cases, no one outside the developer has ever read the code that produced the number, and the honest witness says so.

If interpretation were objective, the same profile would yield the same conclusion in any competent laboratory. It does not. Itiel Dror and Greg Hampikian took the mixture evidence from a real Georgia gang-rape case, where the original laboratory had reported the defendant could not be excluded, and sent it, with no case context, to 17 experienced analysts in an accredited laboratory. One agreed with the original conclusion. The other sixteen did not: most excluded the defendant outright, the rest called it inconclusive. Same electropherograms, same guidelines, one analyst in seventeen reaching the answer that helped convict.

The fair counter is that the variation is not irreducible, and it concentrates on hard samples. When Lourdes Prieto and colleagues gave 18 laboratories shared software, common training and agreed propositions, the likelihood ratios converged to within less than an order of magnitude. Agreement on clean, high-template mixtures is high; the disagreement bites on low-level, high-sharing, complex ones, which is exactly the category that reaches a contested trial. The safeguard the field itself endorses is sequential unmasking: interpret and record the evidence profile before looking at the suspect's reference, so the answer cannot be steered by the target. It is a procedure a laboratory either followed or did not, and "I kept an open mind" is not the same thing.

The one-in-a-billion figure is the most persuasive number a jury will hear, and it rests on assumptions worth naming. The random-match probability is built by multiplying allele frequencies together, which assumes the population mates at random. It does not, so a correction for population substructure is applied, the theta or coancestry term, set at 0.01 for most populations and 0.03 for small isolated ones. The correction is deliberately conservative, made in the defendant's favour, and cross-examining it usually backfires. The number is not junk. But it answers one narrow question: how often an unrelated stranger would coincidentally share this profile.

Two things are commonly done to it that the number cannot bear. The first is the prosecutor's fallacy, transposing the conditional. "The chance this came from someone other than the accused is one in a billion" is not what the figure means; it is the chance an innocent, unrelated person would match, which is a different quantity. In R v Deen a conviction was quashed over exactly this slip. The second is the database hit. When a profile is found by trawling a database of millions rather than by testing a named suspect, the relevant statistic changes, and matches become expected: a nine-locus match was found in a database of 65,000 profiles where the chance of some matching pair was about 94%, precisely as the birthday problem predicts.

The fair statement holds both halves. For a good single-source profile the random-match probability genuinely is astronomically small, the theta corrections are conservative rather than defence-friendly, and a well-run laboratory has a low measured error rate. The error-rate argument does not nullify a clean result; it bounds its weight. But the bound matters, because the number the jury remembers is the billion, and the number that actually governs the chance of a false match is the one the report usually omits. The honest witness gives the match probability for what it is, the coincidence figure, and concedes that the probability of a laboratory or handling error, which is larger and rarely quoted, is the one that limits how far the evidence can be trusted.

A match tells you whose DNA it is. It does not tell you how or when the DNA arrived, and DNA moves. It transfers from person to person, from person to object, and onward again, and it can end up on someone or something with no direct contact at all. Georgina Meakin and Allan Jamieson, reviewing the transfer literature, reach the conclusion the whole section turns on: neither the quantity of DNA recovered nor the quality of the profile can tell you the mechanism by which it was deposited. A strong, clean profile does not mean a hand was there.

The experiments are vivid. In Cynthia Cale's study, pairs shook hands for two minutes and then each handled a knife; in 85% of samples the other person's DNA transferred to the knife, and in a quarter of them the person who had never touched the knife was the only or the major contributor on it. A full, single-source, astronomically discriminating profile on a weapon can belong to someone who never held it.

This is not an argument that a match proves nothing. It is an argument that whose and how are separate questions, and the second one has to be worked, not assumed. The field has a formal framework for it: transfer, persistence, prevalence and recovery, evaluated against activity-level propositions, the subject of the International Society for Forensic Genetics guidance. With proper background and transfer data and paired propositions, a large quantity of fresh cellular material and a plausible contact can support a strong direct-transfer inference. What is not defensible is to report an overwhelming match and let the jury supply the act. The honest witness states whose DNA it is, concedes that the profile alone cannot say how or when it was deposited, and says whether any activity-level analysis was actually done.

The clearest puncture of the infallibility reputation is the record of cases where DNA got it wrong. Contamination is not an exotic mishap; it is, as Norah Rudin and colleagues write, inherent in modern sensitive DNA analysis, something that can be detected and minimised but not eliminated. The measured rates are real. An Austrian audit found police contamination in about 0.75% of crime-scene samples, rising to 1.33% as detection improved, with almost one in ten officers having contaminated evidence at least once, and the authors estimating the true rate at 1.5 to 2%. The Netherlands Forensic Institute, one of the few laboratories that publishes its numbers, logged three false matches reaching reports in five years.

And there is a phrase worth dismantling. When a report says no contamination was associated with a case, that does not mean none occurred; it means nothing unexpected was seen. A contaminant that happens to match the suspect looks exactly like a genuine result. A clean negative control does not prove the evidence sample was uncontaminated either; it shows the reagents were clean, and contamination at the scene or on the bench sails straight past it, which was the error at the heart of the Amanda Knox case.

The low-template end has its own cautionary case. In the Omagh bombing trial, R v Hoey, the judge rejected low copy number DNA because it had not been validated by the wider scientific community and, worse, because the laboratory's own method proved unreproducible on the very case: a third test overturned the consensus of the first two. The fair counter is that these failures drove reform rather than abandonment. Controls, elimination databases and accreditation catch most contamination before a report issues; the review that followed Hoey concluded the underlying science was sound; the exoneration statistics are selected for exonerations and understate nothing so much as the routine cases that go right. Both things are true. A mature quality system catches most errors, and contamination is inherent, not always caught, and has convicted innocent people. The honest witness does not claim the process is error-free, and does not let a clean control stand in for proof that this sample was clean.

How a DNA result is worded decides how a jury weighs it, and the recurring failures have names. Jonathan Koehler catalogued them from the trial record: the source-probability error, treating the match frequency as the chance the defendant is not the source; the ultimate-issue error, treating it as the chance he is not guilty; and the impossibility claim, that a false positive cannot happen. Tellingly, these slips are committed by defence and prosecution alike, which tells you they are driven by misunderstanding, not advocacy. Underneath sits the fact that a reported match is not certainly a true match: false positives run at roughly one to four per cent of match reports in open proficiency tests.

The misunderstanding does not stop at the lawyers. When William Thompson and Eryn Newman tested jury-eligible adults, nearly two-thirds endorsed a source-probability-error reading of the evidence as correct, while rating their own understanding highly. No presentation format was clean, and the likelihood-ratio wording was the most misread of all. And the choice of words can flip the signal: hand the same mixed profile to nineteen qualified experts and, depending only on how each wrote it up, the same trace was rated incriminating by one and strongly exculpatory by another.

The through-line of the whole reading is a single distinction the witness has to hold: the difference between a strong sub-source match and everything the courtroom wants to read into it. DNA earns its reputation at the bottom of the hierarchy, on a good single-source profile, where the coincidence probability is genuinely tiny. It loses that footing rung by rung, and the sections here are the rungs: the mixture whose contributors cannot be counted, the software whose number will not repeat, the profile two laboratories read differently, the error rate that dwarfs the match probability, the DNA that arrived by transfer, the contaminant that matched by chance. The examiner who states what the match supports, under stated propositions, with the error rate acknowledged, and who declines to convert it into an account of the act, is on ground the evidence can hold. Beyond that, the reputation is doing the work, not the science.