Likelihood Ratio: Definition, Use, and Fallacies in Forensics

A likelihood ratio in forensic evidence is a single number that compares two probabilities: how likely the evidence would be if the prosecution’s explanation is correct, divided by how likely the same evidence would be if the defense’s explanation is correct. The result tells a court how strongly the evidence favors one side’s account over the other. It does not tell the court the probability that the defendant is guilty, and treating it as if it did has produced some of the most notorious miscarriages of justice in modern forensic history.

What the Number Actually Compares

The numerator is the probability of observing the evidence if the prosecution hypothesis is true. The denominator is the probability of observing that same evidence if the defense hypothesis is true. One divided by the other gives the ratio.¹

A ratio of exactly 1 means the evidence is equally probable under either explanation, so it carries no weight. Anything above 1 favors the prosecution’s account, and the higher the number, the stronger the support. Anything below 1 favors the defense’s account. When the ratio comes out below 1, analysts usually report its reciprocal and describe the strength of support for the defense hypothesis instead.¹

That is the entire mechanism. The complexity sits in choosing the two hypotheses and estimating the two probabilities, not in the arithmetic.

Why the Two Hypotheses Matter

Every likelihood ratio rests on a pair of competing propositions. The prosecution hypothesis, often labeled Hp, typically proposes that the suspect is the source of the evidence. The defense hypothesis, Hd, proposes that someone else left it.¹

How Hd gets framed changes the answer. If the alternative contributor is assumed to be an unrelated stranger, the denominator relies on general population frequencies. If the alternative is a close relative, the denominator shifts considerably because relatives share more genetic material. A 2024 primer from the National Institute of Standards and Technology is blunt about this point: there is no single “correct value” for a likelihood ratio, and the result depends on “personal or organizational choices” about which hypotheses to test.²

When case facts suggest more than one plausible alternative, analysts may calculate several ratios, pairing Hp against each defense hypothesis in turn.²

The denominator also depends on the reference population used to estimate how common a profile or feature is. In DNA cases, laboratories draw allele frequencies from databases that cover multiple population groups.³ Because people do not mate randomly across an entire country, certain alleles cluster more heavily in some subgroups than population-wide averages suggest. If an analyst ignores that substructure, the matching profile can look rarer than it really is, inflating the ratio against a defendant from a smaller subgroup. Analysts apply a correction factor called theta to make the estimate more conservative.⁴

Reading the Number

A ratio of 50,000 does not communicate much on its own. Forensic reporting guidelines pair numerical ranges with plain-language labels so non-statisticians can grasp the order of magnitude. The scale published by the European Network of Forensic Science Institutes is the most widely adopted:⁵

LR of 1: no support for either explanation
LR of 2 to 10: weak support
LR of 10 to 100: moderate support
LR of 100 to 1,000: moderately strong support
LR of 1,000 to 10,000: strong support
LR of 10,000 to 1,000,000: very strong support
LR above 1,000,000: extremely strong support

These labels work as a continuum, not as hard thresholds. A ratio of 999 and a ratio of 1,001 do not meaningfully leap categories. The point is to communicate whether the evidence nudges the scale slightly or shoves it hard.⁵

Why a Large Ratio Is Not a Verdict

The reason a likelihood ratio cannot tell you whether a defendant is guilty comes from Bayes’ theorem. In its odds form:

Posterior odds = Prior odds × Likelihood ratio

Prior odds reflect everything the jury already knows before the forensic evidence enters: eyewitness testimony, alibis, motive, opportunity. The likelihood ratio is the multiplier that updates those prior odds. The posterior odds are the overall strength of the case with the forensic evidence folded in.

That multiplication structure matters. A DNA likelihood ratio of 1,000,000 against a suspect already linked to the crime by independent evidence makes the case overwhelming. The same ratio against a suspect pulled at random from a database of ten million people, with no other connection to the crime, does not produce certainty. The ratio is the evidence’s contribution. It is never the whole answer.

The Two Fallacies That Keep Showing Up

Two logical errors have specific names in the forensic literature because they recur in trial after trial.

The Prosecutor’s Fallacy

This fallacy treats the likelihood ratio as though it were the probability of innocence. An analyst testifies that the DNA evidence is one million times more probable if the defendant left the sample than if a random person did. A prosecutor then tells the jury there is a “one-in-a-million chance” the defendant is innocent. The two statements are not equivalent. The first describes how the evidence behaves under two scenarios; the second claims to know the probability of guilt, which the evidence alone cannot establish.

The most well-known example is Sally Clark, a British mother convicted in 1999 after an expert witness told the jury the odds of two children in the same family dying of sudden infant death syndrome were 1 in 73 million. The figure treated the two deaths as statistically independent, ignoring possible shared genetic or environmental causes, and was then presented as though it were the probability of Clark’s innocence. Her conviction was overturned on appeal.

The Defense Attorney’s Fallacy

The mirror-image error runs the other way. A defense attorney argues that if a DNA profile matches 1 in every 20,000 people, the defendant is just one of thousands who could have left the sample, so the match proves nothing. The fallacy asks the jury to evaluate the forensic match as if it were the only evidence in the case, ignoring eyewitness identification, physical presence at the scene, motive, and opportunity. When this reasoning lands, strong forensic evidence gets treated as near-worthless.

Both fallacies share one root cause: confusing what a likelihood ratio measures with what the jury is actually supposed to decide. The ratio tells you how the evidence behaves. Guilt is a separate question, answered by weighing all the evidence together.

Where Likelihood Ratios Are Used

DNA analysis is the most developed application. Laboratories examine short tandem repeat markers across the genome and compare the resulting profile against population frequency databases.³ For a clean single-source sample, the ratios often exceed billions to one. The framework becomes essential for mixed samples, where crime scene evidence contains DNA from several people and a simple yes-or-no match conclusion would be scientifically indefensible. Probabilistic genotyping software models the possible combinations of contributors and produces a ratio that accounts for the ambiguity.

Trace evidence disciplines use the same framework. In glass fragment analysis, scientists compare the refractive index and elemental composition of recovered fragments against the broken source, then estimate how common those properties are in reference databases of manufactured glass. The Organization of Scientific Area Committees standard for trace evidence recommends calibrated likelihood ratios when the underlying data support them, treating values above 1,000 as very strong to extremely strong support for a same-source conclusion.⁶

Firearms examination has historically relied on subjective visual comparison of markings, but researchers at NIST developed a Congruent Matching Cells method that automates comparison of breech face impressions and produces a quantitative likelihood ratio.⁷ Fingerprint comparison has been slower to quantify, though a tool integrating a likelihood ratio model into automated fingerprint identification systems was still undergoing validation as of 2025 and had not been widely deployed in casework.⁸

What Can Go Wrong

Likelihood ratios are the best tool forensic science has for quantifying the weight of evidence. They are not infallible, and three limitations are worth knowing before you accept any ratio at face value.

The first is subjectivity in hypothesis selection. The choice of competing propositions, the reference population, and the statistical model all involve judgment. Two competent analysts examining the same evidence can produce different ratios by making different, equally defensible assumptions.² This does not make the number meaningless, but attorneys and judges should ask what went into it.

The second is the foundational validity of methods that depend on human visual comparison. A 2016 report by the President’s Council of Advisors on Science and Technology raised pointed concerns about disciplines like bitemark analysis, hair microscopy, and traditional firearms examination. PCAST concluded that foundational validity could only be established through rigorous “black-box” studies in which multiple examiners evaluate many independent tests against known ground truth. Without those studies, the report warned, an examiner’s claim that two samples match, including any likelihood ratio derived from subjective comparison, is “scientifically meaningless” and has “no probative value, and considerable potential for prejudicial impact.” The report also rejected the idea that experience or certifications could substitute for empirical validation, pointing out that casework rarely gives you the right answer and so cannot be used to estimate error rates.⁹

The third is reference database limitations. Even in DNA analysis, using broad population-average allele frequencies when the actual contributor comes from a smaller, genetically distinct subgroup can make a matching profile look rarer than it is. The theta correction helps, but only if the analyst recognizes the issue and applies the right value.¹⁰

Getting a Likelihood Ratio In Front of a Jury

Before a jury hears a likelihood ratio, the judge decides whether the testimony is admissible. Federal Rule of Evidence 702, amended in December 2023, requires the proponent to show that it is more likely than not that the expert’s opinion rests on sufficient facts, reliable methods, and a reliable application of those methods to the case.¹¹

In most federal courts and many state courts, the admissibility test comes from Daubert v. Merrell Dow Pharmaceuticals, which directs judges to consider whether the method has been tested, whether it has been peer-reviewed, whether it has a known error rate, and whether it is generally accepted in the relevant scientific community.¹² A smaller number of states still follow the older Frye standard, which asks only whether the method is generally accepted by experts in the field.¹³

DNA-based likelihood ratios rarely fail these tests, because the underlying science has decades of validation behind it. The real admissibility fights happen with newer or more subjective applications: probabilistic genotyping software for complex mixtures, automated firearms comparison tools, and any discipline PCAST flagged. In those areas, challenges typically probe the analyst’s choice of hypotheses, the size and relevance of the reference database, and whether the verbal scale used to describe the ratio overstates what the number actually supports.¹²

The practical consequences of exclusion cut both ways. The prosecution may lose its only quantitative link between a suspect and the scene. The defense may lose its ability to show that the forensic evidence is weaker than it looks. Either way, the admissibility hearing is where undocumented assumptions tend to collapse, and where the question of what a likelihood ratio really measures gets its most serious test.

1
National Institute of Justice. Population Genetics and Statistics for Forensic Analysts – Likelihood Ratio
2
National Institute of Standards and Technology. Forensic Science: Statistics Related to Results – Primer: Probability and Likelihood Ratios
3
National Institute of Justice. Population Data on the Expanded CODIS Core STR Loci for Eleven Populations of Significance for Forensic DNA Analyses in the United States
4
National Institute of Justice. Population Genetics and Statistics for Forensic Analysts – Theta Correction
5
European Network of Forensic Science Institutes. ENFSI Guideline for Evaluative Reporting in Forensic Science
6
National Institute of Standards and Technology. OSAC 2022-S-0029 Standard Guide for Interpretation and Reporting of Forensic Comparisons of Trace Evidence
7
National Institute of Standards and Technology. Estimating Likelihood Ratio for Firearm Evidence Identifications in Forensic Science
8
National Institute of Standards and Technology. Likelihood Ratio Model in AFIS – Testing a New Tool to Measure the Weight of Latent Evidence
9
The White House. Forensic Science in Criminal Courts: Ensuring Scientific Validity of Feature-Comparison Methods
10
National Center for Biotechnology Information. The Evaluation of Forensic DNA Evidence: Population Genetics
11
Legal Information Institute. Federal Rules of Evidence Rule 702 – Testimony by Expert Witnesses
12
Legal Information Institute. Daubert v. Merrell Dow Pharmaceuticals, Inc.
13
Legal Information Institute. Frye Standard