Population Databases for Forensic Statistics

Three requirements define a valid reference database: the samples must come from the right population, the database must be large enough to estimate frequencies with acceptable uncertainty, and the collection and typing process must be reproducible. The right population is the one that a trier of fact would consider a plausible source of the crime scene sample, given what is known about the suspect and the circumstances of the offence. In practice, forensic databases are built around ethnic or geographic categories because those are the groupings for which biological frequency data can be systematically collected. The resulting mismatch between biological reality (which is continuous) and database categories (which are discrete) is one of the persistent tensions in forensic population genetics.

Sample size requirements depend on the rarity of the feature being measured and the number of subpopulations to be covered. The National Research Council's 1996 report on DNA evidence (NRC II) recommended databases of at least 100 unrelated individuals per population group for each locus. At this size, a 95% confidence interval around an estimated frequency of 0.01 spans roughly 0.003 to 0.025, which is adequate for most forensic reporting. The SWGDAM 2016 guidelines for the US and the European Network of Forensic Science Institutes (ENFSI) guidance for EU member states both set minimum sample sizes in the same range, while noting that larger databases reduce uncertainty and are preferred.

Reproducibility requires that the typing method, quality control criteria, and data-entry procedures are documented and consistently applied. A database built partly from one kit chemistry and partly from another is of limited value unless the cross-kit concordance has been validated. This issue arose concretely when STR profiling replaced RFLP profiling in the 1990s and when expanded CODIS loci were introduced in 2017: existing databases had to be retyped or re-validated before the new locus frequencies could be trusted.

Human populations are not randomly mating globally. Geographic barriers, cultural practices, and shared ancestry mean that allele frequencies in a Punjabi population differ from those in a Tamil population, which differ from those in a Bengali population. The same is true within nominally single ethnic groups in the United States (African American allele frequencies differ between northern and southern states), in Australia (Aboriginal Australian allele frequencies differ substantially between regional groups), and in every other country where systematic studies have been done. This internal variation is called population substructure or genetic stratification.

Substructure matters for forensic statistics because the product rule, which multiplies per-locus match probabilities to obtain an overall profile frequency, assumes that the loci are independent and that the sampled population is in Hardy-Weinberg equilibrium. When substructure is present, members of a subpopulation share alleles more often than the product rule predicts, because they share recent common ancestors. The result is that the product rule overstates how rare a profile is within the suspect's subpopulation, making the evidence appear stronger than it is.

The theta correction, formalised by Balding and Nichols in 1994, adjusts for this effect. For a homozygous genotype at a single locus, the uncorrected frequency is p squared (where p is the allele frequency in the database). The theta-corrected frequency is [2theta + (1-theta)p][3theta + (1-theta)p] divided by (1+theta)(1+2theta). In practice, most forensic laboratories apply theta values between 0.01 and 0.03 for within-country subpopulations and up to 0.05 for indigenous or highly isolated groups. The correction is conservative: it makes the match probability higher (less incriminating) than the uncorrected figure.

The UK Forensic Science Service adopted theta = 0.03 as a standard correction for DNA evidence presented in English and Welsh courts. The FBI applies theta = 0.01 as a default for CODIS loci, with higher values available for specific groups. Indian forensic practice under the DNA Technology (Use and Application) Regulation Bill 2019 (pending enactment as of 2026) proposes separate databases for major ethnic and tribal groups to reduce reliance on a single correction factor.

The largest and most extensively validated forensic reference databases are for DNA STR profiles. The FBI's CODIS population data, published by Steffen et al. in 2017, covers four US population groups: European American, African American, US Hispanic, and Asian American. Each group contains 361 to 1036 individuals typed at 29 loci including the 20 core CODIS loci. The UK National DNA Database (NDNAD) uses population frequency tables maintained by the Forensic Science Regulator. Germany, the Netherlands, France, and other EU member states maintain national STR frequency databases published in peer-reviewed literature, with the Caucasian European population best represented.

In South Asia, the forensic DNA database infrastructure is less developed. Published STR frequency data exists for several Indian regional populations, including studies on Punjabi, Tamil, Bengali, and Rajasthani groups, but no single authoritative national database exists for forensic use. The Central Forensic Science Laboratory (CFSL) in India uses internally compiled frequency tables that are not uniformly published. This creates a transparency problem: a defence lawyer challenging a DNA statistic in an Indian court cannot easily access the underlying database to verify the calculation. The DNA Technology Bill, if enacted, would require the National DNA Data Bank to maintain and publish population data, moving India closer to the transparency standard already operating in the UK and US.

Hardy-Weinberg equilibrium testing and linkage disequilibrium testing are the primary validation tools for STR databases. If a database passes HWE tests at each locus and shows no significant linkage disequilibrium between loci, the product rule can be applied without adjustment. The Steffen et al. CODIS dataset and the major European national databases pass these tests at the locus level, supporting the standard reporting approach. Databases with HWE departures at multiple loci are an indicator of sample quality problems or significant unaccounted substructure and should not be used without further investigation.

The same principles that govern DNA databases apply to reference collections for other forensic evidence types, but the databases are generally smaller, less rigorously validated, and more jurisdiction-specific. Glass refractive index databases are a well-studied example. The refractive index (RI) of a glass fragment tells a forensic scientist the optical density of the glass, which varies systematically with glass type and manufacturer. A reference database of RI values from known glass sources allows the scientist to calculate how common the measured RI is in the target population of glass likely to appear in that jurisdiction.

The UK Forensic Science Service built a glass RI database from windscreen and building glass samples collected over many years, which underpins the likelihood ratio calculations reported in UK glass evidence. The database reflects UK glass manufacturers and imports prevalent at the time of collection: its relevance to a case involving glass from a different country or a different time period requires separate justification. Similar databases exist in Australia (maintained by the Australian Federal Police forensic laboratory) and in several European countries, but they are not directly interchangeable because glass composition varies with manufacturer and national building standards.

Fingerprint frequency databases are a more contested area. Traditional fingerprint examination uses categorical classification (arches, loops, whorls) and point-counting, neither of which directly produces a probabilistic frequency statement from a database. Statistical approaches to fingerprint frequency, including the models developed by Srihari et al. and by Neumann et al., use large databases of known prints to estimate the probability of observing a given minutiae configuration. These databases are maintained by institutions such as the FBI and the Netherlands Forensic Institute. Courts in the US and UK have been cautious about full probabilistic reporting of fingerprint evidence, partly because the underlying databases and models have not been as thoroughly peer-reviewed as DNA databases.

Using a database that does not represent the suspect's population can produce a match probability that is either too low or too high, depending on the direction of the frequency difference. If the suspect belongs to a population in which the relevant alleles are more common than in the database, the calculated match probability understates the true probability of a coincidental match, making the evidence appear more incriminating than the data supports. If the suspect belongs to a population in which the alleles are rarer, the database will produce an overly conservative probability, slightly disadvantaging the prosecution.

The typical magnitude of the error depends on the genetic distance between the database population and the suspect's actual population. For closely related groups, such as two Western European populations, the difference in reported match probabilities is usually less than one order of magnitude. For more distant groups, such as comparing a Western European database to an indigenous Amazonian population, differences of several orders of magnitude have been documented. In practice, the worst-case mismatches arise when a database built from one continental population is used to evaluate evidence from a suspect drawn from a different continent.

Courts in England and Wales addressed this issue in R v Doheny and Adams (1997), which established that the prosecution must specify which database was used and must justify its choice. The US Supreme Court's decisions related to DNA evidence have not directly mandated database choice, but SWGDAM guidelines require laboratories to document and justify their population database selection. Under the Bharatiya Sakshya Adhiniyam 2023 (which replaced the Indian Evidence Act 1872), expert opinion evidence is admissible when the expert is qualified and the basis for the opinion is disclosed: this framework supports, though does not yet specifically require, disclosure of population database choices in DNA cases.

Validation of a forensic reference database involves three categories of checks. First, internal consistency checks confirm that the database's own data is coherent: HWE tests at each locus, linkage disequilibrium tests between loci, and quality control checks on individual sample typing. Second, external comparison tests confirm that the frequency estimates agree with published data from independent studies of the same population. If a newly built Japanese STR database produces allele frequencies that differ substantially from previously published Japanese studies, that is a signal of either sampling bias or typing error. Third, proficiency testing assesses whether analysts using the database produce consistent results: the same evidence profile typed in two different laboratories against the same database should produce the same match probability.

Databases require maintenance because populations change over time. Migration, admixture, and the passage of generations gradually shift allele frequencies. A database built from samples collected in 1995 may not accurately represent the frequency distribution in 2025, particularly in countries with high immigration rates. Most forensic science authorities require periodic review of database currency, though specific re-collection intervals are not universally mandated. The US CODIS database was last substantially updated with the expanded 20-locus panel validated on samples collected in the 2000s and published in 2017.

Expansion of forensic profiling systems requires parallel database expansion. When the UK Forensic Science Service introduced the SGM Plus 10-locus system, new population frequency data had to be collected for the additional loci. When the US moved from 13 to 20 CODIS loci in 2017, the Steffen et al. study provided the new frequency estimates. In both cases, the transition required laboratories to manage a period in which some samples had been typed at the old locus set and others at the new set, and the statistical reporting had to be clearly labelled with which loci were included. The same transition challenge will arise in any jurisdiction that adds new loci to its national profiling system.