Genotype imputation estimates variants your DNA test never measured. It compares your measured markers with reference genomes and relies on the fact that nearby variants are inherited together. It works well for common variants in well-represented ancestries.
Imputation is much less reliable for rare variants and under-represented populations. Imputed calls are statistical estimates, so never treat them as measured results for medical decisions.
You upload the same raw DNA file to two services. One says you carry a variant and the other says you do not, or does not mention it at all. Both used your data, so how can they disagree? Very often the answer is imputation. At least one of those reports is describing a position your test never measured.
What imputation actually does
Most consumer DNA tests use genotyping arrays that measure a preset list of hundreds of thousands of positions. Human genetic variation is far larger than that. The NHLBI TOPMed program catalogued more than 400 million variants across 53,831 sequenced genomes, and 97% of them had frequencies below 1%. An array samples a small slice of that variation.
Imputation fills in the gaps statistically. Nearby variants tend to be inherited together in blocks, a pattern called linkage disequilibrium. Software first estimates which of your measured alleles sit together on each chromosome copy (phasing). It then compares those stretches with a reference panel of fully sequenced genomes. If your measured markers match a block that, in the reference genomes, usually carries a particular allele at an unmeasured position, the software infers that you probably carry it too.
The output is a probability, not an observation. Imputation tools report quality measures such as an r² or an INFO score that estimate how reliable each imputed variant is. Those scores are estimates, too. They are calculated without knowing your true genotype.
Think autocomplete, not microscope. Imputation predicts what is most likely there based on people with similar DNA. It is very good at common patterns and much weaker at the unusual.
Why two reports can disagree
- Measured vs imputed. One service reports only positions on your array. Another reports imputed positions alongside them, sometimes without marking which is which.
- Different reference panels. Panels differ in size and in which populations they include, so the same file can produce different imputed genotypes.
- Different quality cutoffs. One service may drop low-confidence calls, while another reports them.
- Different interpretation databases. Even when the genotype matches, two services can attach different meanings to it.
Accuracy depends on frequency and ancestry
Rare variants are hard to impute
Imputation works by matching you to shared haplotypes, so rare variants, which few reference genomes carry, are the hardest to call. In a 2021 comparison of phasing and imputation tools in Frontiers in Genetics, concordance started to decline once a variant’s minor allele frequency fell below 5% and dropped drastically below 1%. That study used European data only.
That matters because many medically significant variants are rare. Even direct measurement on arrays struggles there. A 2021 BMJ study of 49,908 UK Biobank participants found that only 16% of array calls for very rare variants (below 0.001% frequency) were confirmed by sequencing.
Reference panels do not represent everyone equally
Imputation is most accurate when your ancestry is well represented in the reference panel. A 2019 PLoS Genetics study noted that the widely used Haplotype Reference Consortium panel is predominantly European. It found that using the larger, more diverse TOPMed panel improved imputation quality in African American and Hispanic/Latino samples, especially for rare variants. The authors also cautioned that the gains likely do not extend to populations poorly represented in TOPMed, such as South Asians.
A 2021 Genome Research study gives a sense of the gap. It imputed a standard array against the 1000 Genomes reference. For common variants, the array reached an average r² of about 0.91 in European-ancestry samples and 0.83 in African-ancestry samples. The study was small (120 individuals), and its authors work for a company that sells low-pass sequencing software, but the direction of the gap matches the wider literature.
| Factor | Tends to impute well | Tends to impute poorly |
|---|---|---|
| Variant frequency | Common (above ~5%) | Rare (below ~1%), and private to a family |
| Ancestry vs reference panel | Well represented in the panel | Under-represented or highly admixed |
| Nearby measured markers | Dense, in strong linkage disequilibrium | Sparse or weakly linked |
How upload services use imputation
Some third-party services impute your uploaded array file before generating reports, and at least one advertises turning a file of roughly 750,000 markers into more than 200 million variants with a single headline accuracy percentage. Treat a figure like that with care. An average across all variants is dominated by common ones, which impute well, and says little about the rare variants that matter most for health.
Before you rely on an upload service’s report, check its documentation for:
- whether it imputes at all, and against which reference panel;
- whether each reported genotype is marked as measured or imputed;
- whether it shows a per-variant quality score; and
- what it does with your file afterward. Our privacy checklist covers storage, deletion and sharing.
Unsure what your download actually contains? The file inspector identifies consumer TXT/CSV files, VCFs and sequencing files locally, without uploading anything.
Why imputed calls are not medical results
An imputed genotype is a statistical guess about a position nobody measured. It can be useful for research and for population-level questions such as ancestry. It is not evidence that you carry a disease-causing variant. If a report built on imputed data flags something that could affect screening, treatment or family planning, treat it as a question to take to a clinician or genetic counselor, not as an answer. A clinical lab can test the actual position, as we explain in how confirmatory testing works.
Do not act on an imputed result alone. No medication, screening or surgical decision should rest on a genotype your test inferred rather than measured.
How whole genome sequencing differs
Whole genome sequencing reads DNA across the genome instead of checking preset positions. At 30× depth, each position is covered by about 30 reads on average, so rare and even family-specific variants can be observed directly rather than inferred. Our genotyping vs sequencing explainer and the test comparison cover the trade-offs.
Two honest caveats. First, not all sequencing avoids imputation. Low-pass sequencing, at an average depth below 1×, is designed to be combined with imputation, as the Genome Research study above describes (see low-pass vs 30× sequencing). Second, deep sequencing is not flawless. Some repetitive regions are hard to read with standard short reads, and a consumer genome is still not a clinical test. Medically important findings still need clinical confirmation.
Commercial option · affiliate link
Want the whole genome and every raw file?
Dante Labs sequences the whole genome at 30× and lets you download the raw BAM, FASTQ and VCF files.
Check Dante’s current offer →Our code GENOME may apply; codes may not combine with sitewide sales. Why this is an affiliate link.
If your goal is ancestry or general curiosity, an array plus imputation is often enough. If you want to know what is actually at a rare position, imputation cannot give you that. Measuring it can.
Common questions
Is imputed DNA data the same as measured data?
No. An imputed genotype is a statistical estimate based on reference genomes. A measured genotype was actually read from your sample.
Why is imputation less accurate for some ancestries?
Accuracy depends on how well your ancestry is represented in the reference panel. Many widely used panels have been predominantly European, and studies show quality improves when panels are larger and more diverse.
Can imputation find rare disease variants?
Poorly. Imputation accuracy falls off sharply for rare variants, which are often the medically important ones, so any such finding needs clinical confirmation.
Does whole genome sequencing use imputation?
Standard 30× sequencing measures positions directly. Low-pass sequencing, below 1× average depth, is designed to be combined with imputation.
How can I tell if a report used imputed data?
Check the service’s documentation or report for a measured vs imputed flag and per-variant quality scores. If it does not say, ask before relying on the result.
Keep going
Sources
Checked October 2026: the Stahl et al. 2021 Frontiers in Genetics article (imputation method and frequency thresholds), Kowalski et al. 2019 PLoS Genetics (TOPMed reference panel and ancestry), Li et al. 2021 Genome Research (array vs low-pass imputation r² by ancestry; authors disclose a commercial interest), Taliun et al. 2021 Nature (TOPMed variant counts), Weedon et al. 2021 BMJ (array accuracy for rare variants), and one upload service’s published imputation claim. We did not name individual services’ accuracy figures or cite a fold-improvement number from Kowalski et al., because we could not confirm which metric it applied to.
- Stahl et al., Frontiers in Genetics (2021): assessment of imputation quality
- Kowalski et al., PLoS Genetics (2019): TOPMed improves imputation in admixed populations
- Li et al., Genome Research (2021): low-pass sequencing vs genotyping arrays
- Taliun et al., Nature (2021): sequencing of 53,831 diverse genomes (TOPMed)
- Weedon et al., BMJ (2021): SNP chips and rare pathogenic variants
Educational content, not medical advice. A genetic result is information, not a diagnosis; for anything medically important, talk to a qualified clinician or genetic counselor. Some outbound links are affiliate links; see the disclosure.