#Reference frequency data

Every match probability is conditional on a reference population, and that population arrives as a FrequencyDb: per locus, the allele frequencies and n, the number of individuals typed (2N alleles). The struct fields are in the Cheatsheet.

#The 5/(2N) floor

A reference sample of N people has seen 2N alleles at a locus, so an allele rarer than about 1/(2N) is often absent from the table. NRC II floors every allele frequency at

5 / (2N)

for that locus's own N: 0.0025 at N = 1000, 0.025 at N = 100. The floor is applied on read, so the stored table keeps its published numbers:

forensicgenetics.minFrequency($db, "TH01");              # 5 / (2N)
forensicgenetics.alleleFrequency($db, "TH01", "9.3");    # max(published, floor), capped at 1

An unlisted allele is scored at the floor; a listed allele below the floor is raised to it. A locus with n = 0 has no floor, and alleleFrequency throws for an allele it does not list. MIN_ALLELE_COUNT is the 5.

#Building a table

def db as forensicgenetics.FrequencyDb init forensicgenetics.db("Example population");
$db = forensicgenetics.withLocus($db, "TH01", 1000,
    {"6": 0.25, "7": 0.15, "8": 0.30, "9.3": 0.30});
$db = forensicgenetics.withLocus($db, "D3S1358", 1000,
    {"15": 0.25, "16": 0.20, "17": 0.30, "18": 0.25});

withLocus returns a copy and throws on an empty locus name, a negative n, a frequency outside (0, 1] (leave an unobserved allele out; the floor covers it), or frequencies summing past MAX_FREQUENCY_SUM (1.05). A sum below 1 is accepted, since a table may list only part of a locus.

Each call copies the whole database; see Limitations for the cost at hundreds of loci.

#Published tables

The core reads no file format. A reader deck imports the core and produces a FrequencyDb:

import "@deckyard/strider/" as strider;

def db as forensicgenetics.FrequencyDb init strider.load($xml, "STRidER AT");

The deck bundles no reference data. The table in examples/lib/exampledb.j is synthetic. Real autosomal STR frequencies for European populations come from STRidER, the ENFSI reference database, whose formulae page is the specification these estimators follow.

#Which dataset

STRidER publishes per-country, per-continent and whole-database tables, and they give different answers. Name the dataset when you build the table; the label travels as FrequencyDb.name.

forensicgenetics.loci($db);                  # locus names, insertion order
forensicgenetics.hasLocus($db, "TH01");      # bool
forensicgenetics.sampleSize($db, "TH01");    # N

A locus the table lacks throws everywhere, with a message naming the locus and the table.

#When the population is uncertain

For "the donor was German or Austrian", use the law of total probability:

forensicgenetics.randomMatchProbabilityAcross([$de, $at], [0.5, 0.5], $suspect,
    forensicgenetics.THETA_GENERAL);
forensicgenetics.matchLikelihoodRatioAcross([$de, $at], [0.5, 0.5], $suspect,
    forensicgenetics.THETA_GENERAL);

It computes the match probability within each population and averages the results, weighted by the prior that the donor came from each. The weights must be non-negative and sum to 1 (within 0.001).

combineFrequencies pools tables instead:

forensicgenetics.combineFrequencies([$de, $at], "DE+AT");

It sums observed allele counts (p * 2N) over the loci every input carries, and describes one merged population. Pooling averages allele frequencies where the Across form averages genotype probabilities. A merged population whose parts differ has fewer heterozygotes than Hardy-Weinberg predicts (the Wahlund effect), so the RMP from a pooled table is biased downwards and overstates the evidence.

Two populations, one locus, equal sample sizes:

For a 6,6 homozygote at theta 0:

The pooled answer overstates this match by a third at one locus, and the gap compounds across a profile. Use combineFrequencies only when the merged group is itself the reference population.

A conservative alternative is to compute in each candidate population and report the largest RMP.