🧬 genomics

A Bias Correction Method for Substructure Estimation with Small Reference Sample Sizes

This paper demonstrates that small reference sample sizes introduce systematic bias in population substructure estimates using Summix2, identifies a threshold of 150 individuals per group for accuracy, and proposes two effective bias correction methods tailored to different reference sample size scenarios.

Original authors: Tifour, S., Price, A., Marker, K., Hendricks, A.

Published 2026-11-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tifour, S., Price, A., Marker, K., Hendricks, A.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

To understand the genetic makeup of a population, scientists often look at the frequency of specific genetic variants, much like counting the colors of marbles in a jar to guess the mix of flavors inside. When researchers study groups of people with mixed ancestry, such as African Americans or Latinos, they need to know exactly how much of their DNA comes from different ancestral roots, like West Africa, Europe, or Indigenous America. This breakdown, known as population substructure, is crucial for ensuring that medical studies and genetic predictions are accurate and fair. However, a major hurdle exists: to make these calculations, scientists need a reference library of genetic data from each ancestral group. For some groups, like those of European descent, these libraries are vast and detailed. For others, particularly Indigenous American populations, the available data is often sparse, containing only a handful of individuals. When the reference library is too small, the resulting calculations can be skewed, leading to inaccurate maps of a person's ancestry and potentially flawed medical conclusions.

A team of researchers set out to solve this specific problem of small reference libraries. They focused on a widely used computer method called Summix2, which estimates ancestry proportions using summary genetic data. The team wanted to know exactly how much error creeps into these estimates when the reference groups are small, and whether they could build a mathematical fix to correct the mistake. They did not have real-world ground truth to test against, so they created thousands of simulated populations with known ancestry mixes. They then fed these simulations into the Summix2 program using reference libraries of varying sizes, ranging from just five individuals to one thousand. By comparing the program's output against the known truth of their simulations, they could measure exactly how much the small sample sizes distorted the results.

The researchers discovered that small reference samples introduce a systematic bias, meaning the errors are not random but consistently push the estimates in the wrong direction. When they reduced the size of a reference group, the error did not just stay isolated to that group; it spilled over and distorted the estimates for the other groups as well. For instance, when they shrunk the reference sample for Indigenous Americans in a simulation of a Latino-like population, the accuracy of the European ancestry estimate dropped significantly, even though the European reference data remained large and robust. This finding revealed that the quality of the entire ancestry map depends on the weakest link in the chain, not just the size of the specific group being measured.

Through their simulations, the team identified a clear threshold for reliability. They found that once a reference group reaches a sample size of 150 individuals, the estimation error drops below a level that is generally considered acceptable for most scientific purposes. Below this number, the errors become substantial. However, the researchers also realized that in the real world, waiting for large datasets for every population is not always possible, especially for Indigenous groups where data collection is limited by ethical considerations and community governance. To address this, they developed two new correction methods to adjust the results when small sample sizes are unavoidable.

The first method is a simple adjustment that assumes the error is inversely related to the sample size, essentially adding or subtracting a small amount based on how few people were in the reference group. The second method is more complex and flexible, using a power-law relationship that allows the error to change at a different rate as the sample size grows. The researchers tested both methods against their simulations and found that the best choice depends on the situation. If all the reference groups in an analysis are small, the flexible power-law correction works best, providing the most accurate results. However, if only one reference group is small while the others are large, the simpler correction method performs better.

To ensure these findings held up in the real world, the team applied their methods to actual genetic data from African American populations. They treated the estimates derived from large reference samples as a benchmark for stability and then tested how the estimates changed when they artificially reduced the sample sizes. The real-world data confirmed their simulation results: reducing the sample size of one group increased the error across the board, and applying the appropriate correction significantly reduced this deviation. The study concludes that while larger reference samples are always preferable, researchers can now use these specific correction tools to salvage accurate ancestry estimates even when working with limited data, provided they choose the right method for their specific mix of reference groups. This work offers a practical path forward for genetic studies involving underrepresented populations, ensuring that the lack of large reference datasets does not compromise the accuracy of scientific findings.

Technical Summary: A Bias Correction Method for Substructure Estimation with Small Reference Sample Sizes

Problem Statement
Accurate estimation of population substructure from genetic summary data (e.g., allele frequencies and GWAS summary statistics) is critical for downstream analyses such as polygenic score development and meta-analysis. Methods like Summix2 estimate mixing proportions by minimizing the least-squares difference between observed allele frequencies and a weighted combination of reference group allele frequencies. However, these methods rely on reference allele frequencies derived from finite reference samples. In practice, reference datasets are unevenly distributed; some populations (e.g., Indigenous American) have significantly smaller sample sizes than others (e.g., European or African). The authors demonstrate that small reference sample sizes introduce systematic bias into substructure proportion estimates. Furthermore, this bias is not isolated; a small sample size in one reference group can degrade the accuracy of estimates for other reference groups, even when those other groups have large sample sizes.

Methodology
The study utilized Summix2 to estimate mixing proportions across five continental reference groups: African (AFR), East Asian (EAS), European (EUR), Indigenous American (IAM), and South Asian (SAS). The authors employed a dual approach involving simulation studies and real-data evaluation:

Simulation Study:
- Data Generation: Reference allele frequencies were simulated for sample sizes ranging from $N=5$ to $N=1000$ . Observed allele frequencies were generated for three admixed populations (Latino-like, African American-like, and Peruvian-like) with known true mixing proportions.
- Scenarios: Three scenarios were tested: (1) Joint variation of all reference group sample sizes; (2) Variation of one reference group's sample size while others were fixed at $N=500$ ; and (3) Variation of one group while others were fixed at $N=150$ .
- Bias Correction Development: Two post-estimation correction methods were developed to adjust the Summix2 mixing proportion estimates ( $\hat{\pi}_k$ $\overset{π}{^}_{k}$ ) based on the reference sample size ( $N_k$ $N_{k}$ ):
  - Simple Inverse Correction ( $1/N$ ): Assumes bias magnitude is inversely proportional to sample size. The correction adds or subtracts $1/N_k$ based on whether the estimated proportion is above or below the uniform expectation ( $1/K$ ).
  - Flexible Power-Law Correction ( $a/N^b$ ): Models bias using a power-law relationship to account for varying rates of decay in bias as sample size increases. Parameters $a$ and $b$ were estimated via nonlinear least squares on simulation data.
- Constraints: Both methods included a post-correction step where negative estimates were truncated to zero and the vector was renormalized to sum to one.
Real-Data Evaluation:
- Observed allele frequencies were taken from gnomAD v3.1.2 (African/African American data).
- Reference allele frequencies were derived from 1000 Genomes and HGDP, subsampled to various sizes ( $N=5$ to $N=500$ ).
- Since true mixing proportions are unknown in real data, estimates derived from the largest available reference samples served as a stability benchmark.

Key Results

Inverse Relationship: Total absolute estimation error decreases approximately inversely with reference sample size. A practical threshold of $N=150$ individuals per reference group was identified, below which mean estimation error exceeds 0.05.
Cross-Group Contamination: Limiting the sample size of a single reference group (e.g., IAM) increased estimation error for other groups (e.g., EAS), even when those other groups maintained large sample sizes ( $N=500$ ). This suggests that the joint estimation constraints in Summix2 propagate errors across groups.
Correction Performance:
- Scenario 1 (All groups small): The power-law correction provided the greatest reduction in mean total absolute estimation error when all reference groups had small sample sizes.
- Scenarios 2 & 3 (One group small): The simple $1/N$ correction performed better when only a single reference group had a small sample size while others were large.
- Diminishing Returns: For sample sizes above $N=150$ , both corrections offered minimal additional benefit over uncorrected estimates.
Real-Data Validation: The real-data evaluation confirmed that small sample sizes increase deviation from the benchmark. However, the benefit of correction was most pronounced at the smallest sample sizes ( $N < 45$ ); at larger sizes, corrected estimates sometimes deviated more from the benchmark than uncorrected ones, indicating that correction is most critical when data is scarce.

Significance and Claims
The paper claims to provide practical guidance for researchers using Summix2 or similar summary-data-based substructure estimation tools when reference data is limited or imbalanced. The primary contributions are:

Identification of a Threshold: Establishing $N=150$ as a practical threshold for accurate estimation (error < 0.05).
Demonstration of Interdependence: Showing that the adequacy of a reference sample cannot be evaluated in isolation; small samples in one group compromise estimates for the entire set of reference groups.
Context-Specific Correction Strategy: Offering a decision framework (summarized in Table 1) where users should apply the power-law correction if all reference groups are small, but the simple $1/N$ correction if only one group is small.
Methodological Limitation: The authors modestly note that these are post-estimation corrections and do not directly account for uncertainty in the reference allele frequencies during the optimization process. They suggest future work could integrate this logic directly into the estimation procedure.

The study concludes that while reference sample size is a known factor, its specific impact on summary-data substructure estimation—and the specific bias patterns it creates—requires these targeted adjustments to ensure reliable results, particularly for underrepresented populations.

More like this