Small References, Big Biases: How Sample Sizes Distort Ancient Indian Admixture Models

Modern population genetics relies heavily on reference panels—collections of individual genomes used as baseline templates to untangle complex ancestral ancestries. When computational tools model an individual or a population as a mix of distinct historical components, the accuracy of those proportions depends fundamentally on how well those reference panels capture baseline variation.

In South Asian paleogenomics, building these reference sets remains a structural hurdle. Ancient high-coverage genomes from the Indian subcontinent remain scarce compared to those available for Western Eurasia. As a result, statistical models frequently substitute small, geographically distant, or modern proxy panels to estimate ancient ancestral components.

A new methodological preprint published on bioRxiv, titled "A Bias Correction Method for Substructure Estimation with Small Reference Sample Sizes", systematically evaluates what happens when statistical mixture models operate on small reference sample sizes. The study reveals that sample size deficits do not merely introduce random noise; they generate systematic, predictable biases in ancestry proportion estimates.

The Mechanics of Sample-Size Bias

The study focuses on Summix2, an established mixture modeling framework designed to estimate and adjust for genetic substructure using aggregate summary statistics. Summix2 calculates ancestry proportions by matching allele frequencies (variant frequency patterns across the genome) in a target sample against reference groups representing major global continental ancestries: African, East Asian, European, Indigenous American, and South Asian.

When reference sample sizes fall below critical thresholds, allele frequency estimates derived from those panels become noisy. The authors demonstrated that this sampling noise forces the underlying optimization algorithms to systematically misallocate ancestry proportions across the reference categories.

To quantify this effect, the researchers identified a sample size threshold of 150 individuals per reference group required to maintain stable, reliable proportion estimates with a mean estimation error below 0.05. Below this threshold, proportion estimates skew systematically. To resolve this, the authors developed two statistical corrections: an inverse sample size correction for scenarios where a single reference group is small, and a flexible power-law correction that mitigates estimation error when all reference groups suffer from small sample sizes.

Implications for South Asian Paleogenomics

While the bioRxiv paper tested continental panels and clinical genetic cohorts, its mathematical findings address a longstanding challenge in South Asian archaeogenetics.

Reconstructing South Asian population history relies on modeling admixture among deep ancestral streams. In genetic literature, these streams are typically conceptualized as Ancestral South Indian (ASI)—itself modeled as a mix between Ancient Ancestral South Indian (AASI) lineage and ancient Iranian agriculturalist-related ancestry—and Ancestral North Indian (ANI), which incorporates additional Steppe-related pastoralist ancestry.

Because direct high-coverage ancient DNA samples from many ancient South Asian time periods remain sparse, models frequently rely on small ancient reference sets or modern endogamous proxy populations.

When reference panels for specific ancestral components are small, mathematical models can over- or under-estimate regional proportions. If a reference panel representing a deep local ancestral lineage consists of only a handful of individuals, its internal allele frequency variance increases. Under mixture models like Summix2, this variance can cause the algorithm to falsely assign genuine local alleles to larger, better-sampled reference panels—such as Eurasian or Western Steppe panels—simply because larger panels offer more statistically stable allele frequency baselines.

⚖ Scholarly Debate: Interpreting Admixture Proportions

The mathematical sensitivity of admixture modeling highlights an ongoing debate among geneticists, archaeologists, and anthropologists regarding how ancient migration and continuity should be quantified.

• Migration and Admixture Models: Proponents of primary migration models argue that despite limitations in sample size, formal statistics (such as f-statistics and qpAdm modeling) consistently detect distinct pulses of gene flow into South Asia across different epochs. They maintain that while exact numerical percentages fluctuate based on reference choice, the overall directional sequence of population movements remains statistically robust.

• Local Continuity and Sampling Skepticism: Conversely, scholars emphasizing indigenous continuity and localized demographic development caution against taking precise admixture percentages at face value. They argue that when local reference panels are sparse or geographically restricted, statistical algorithms over-simplify complex, continuous isolation-by-distance patterns into discrete "pulses" of outside admixture. From this perspective, uncorrected sampling biases risk creating artificial genetic shifts where gradual, localized micro-evolution occurred.

Moving Beyond Uncorrected Models

The findings in the bioRxiv paper underscore that statistical precision cannot be separated from sampling scale. As paleogenomics expands across South Asia, applying analytical adjustments—such as the power-law and inverse sample size corrections proposed for Summix2—will be essential for ensuring that ancient ancestry estimates reflect demographic history rather than mathematical artifacts of small sample sizes.

For researchers interpreting ancient genomes, the lesson is straightforward: before drawing sweeping historical conclusions from genetic admixture fractions, one must first audit the size and stability of the reference panels that generated them.