Biological age test accuracy: reliability, noise and repeat results

Oct 20, 2026

Prepared by: Research Desk

Prepared by: Research Desk. Sources are linked for verification.

Data point first: A 2022 analysis of epigenetic clock reliability measured the same DNA samples more than once. On some widely used clocks, the age estimates differed by up to about 9 years between these technical replicates, from technical noise alone [5]. A clock can be scientifically sound and still be noisy for a single person's single sample. This explainer covers where that noise comes from, which clock designs reduce it, and what to look for in a test's published accuracy data.

For a primer on how clocks are built and what each one predicts, start with our guide to how epigenetic clocks work. This page focuses on one question: how much trust does a single number deserve, and how much does a change between two numbers deserve?

Three different meanings of "accuracy"

When a company says its biological age test is "accurate", it can mean at least three different things. They are not interchangeable.

Accuracy against chronological age. First-generation clocks were trained to predict calendar age. Horvath's 2013 pan-tissue clock was built from DNA methylation data across many human tissues and cell types. It was evaluated by how closely its estimate tracked the known age of each sample donor [1]. A low error here shows the clock captures age-related methylation. On its own, it does not show that a gap between estimated and calendar age is meaningful for an individual.
Association with outcomes. Later clocks were trained on different targets. PhenoAge was trained on a composite of blood-chemistry markers linked to mortality risk [2]. GrimAge combined methylation-based surrogates for plasma proteins and smoking history, and was reported to predict lifespan and healthspan measures in human cohorts [3]. DunedinPACE was trained on the rate of change in many physiological measures, tracked over years in a single birth cohort [4]. For these clocks, "accuracy" usually means how strongly the score is associated with outcomes across a population.
Reliability (precision). This asks whether the same sample, or two samples taken close together, give the same answer. For anyone comparing a result over time, this is the property that matters most. It is also the one most often missing from marketing pages.

A clock can score well on the first two and still be weak on the third. Population-level associations average out random error across thousands of people. A single retest does not.

Where the noise comes from

Unreliable individual CpG sites

Clocks combine methylation levels at hundreds of CpG sites, each weighted by a coefficient. Higgins-Chen and colleagues examined technical replicates, meaning the same biological sample measured more than once. They found that many individual CpG sites used in clocks show poor reliability [5]. When a clock places heavy weight on noisy sites, small measurement fluctuations add up to large swings in the final age estimate. That mechanism produced the multi-year replicate differences reported in the study.

Stochastic variation that is real but hard to measure

Not all variation is measurement error. Methylation also changes in partly random ways with age, a phenomenon studied as "stochastic epigenetic mutations". A 2024 human study described methods aimed at detecting these changes reliably and reported associations with cardiovascular aging measures [6]. For measurement, the lesson is that separating genuine biological drift from technical noise is a methods problem in its own right. The two can look alike in a single array readout.

Tissue type

Methylation patterns differ between tissues. Horvath's clock was designed to work across tissues [1]. Many other clocks were trained on blood. CheekAge, by contrast, was developed specifically for buccal (cheek swab) samples and reported associations with lifestyle and health variables in its human data [8]. If a lab applies a blood-trained clock to saliva or cheek cells, or switches sample type between retests, some of the difference between results may reflect tissue rather than aging.

Clock designs that address reliability

Principal-component (PC) clocks

The 2022 reliability study proposed a computational fix. Instead of training on individual CpG sites, the authors first compressed methylation data into principal components. These summarise shared patterns across many sites, so random noise at any single site carries less weight. They then retrained versions of existing clocks on those components [5]. The authors reported that the resulting PC clocks agreed much more closely between technical replicates, and noted that this matters directly for clinical trials and longitudinal tracking [5]. The same principal-component approach was later used to build a brain-tissue clock from multi-region methylation data, studied in the context of Alzheimer's disease [9].

Selecting reliable probes from the start

DunedinPACE took a related route. During development, the authors restricted the clock to probes with acceptable test-retest reliability, and they reported the clock's reliability as part of its validation [4]. Because it reports a pace (years of biological change per calendar year) rather than an age, its scale and error look different from age-based clocks. Comparing a DunedinPACE value directly with a Horvath or PhenoAge "age" is not a like-for-like comparison.

Reading change between two results

This is where accuracy questions become practical. Suppose a first test reads 44 and a second, months later, reads 41. Is that a real change?

A 2024 preprint by Borrus and colleagues addressed this question in the context of aging interventions. It argued that changes in epigenetic clock estimates can be false positives driven by noise, sample handling, batch effects or the choice of clock, and it proposed checks for when a change can be trusted [7]. This is a computational and methods paper posted as a bioRxiv preprint, so it had not completed peer review when it was indexed. Its central point agrees with the replicate data in the 2022 study [5]: a single before-and-after pair on a noisy clock cannot separate signal from error.

Points that published methods work raises about interpreting two results:

Compare like with like: same lab, same sample type, same clock version, and ideally similar collection timing and handling.
Know the clock's replicate error. A change smaller than the documented technical variation is consistent with no real change.
Watch for batch effects. Samples processed in different array batches can shift systematically. Labs that re-run or normalise earlier samples alongside new ones reduce this problem.
Prefer clocks built for longitudinal use. PC-based clocks and reliability-filtered clocks were designed with repeat measurement in mind [4][5].
Read multiple clocks moving in different directions as a caution flag, not a mixed result to cherry-pick from.

What to compare before ordering a test

When choosing a biological age test, these neutral criteria separate well-documented products from thin ones:

Criterion	What to look for
Clocks reported	Named, published clocks (e.g. DunedinPACE, PC versions of established clocks) rather than an unnamed proprietary "score"
Replicate data	Published technical-replicate or test-retest figures for the exact clock and sample type offered
Sample type match	The clock was trained or validated on the tissue collected (blood, saliva or cheek swab)
Lab standards	The lab states its CLIA certification; CLIA sets federal quality standards for laboratory testing [12]
Regulatory clarity	Plain statement that the test is a laboratory-developed test, not an FDA-cleared device [10][11]
Price transparency	Clear cost for retests, since accuracy questions only resolve with repeat measurement
Data privacy	Explicit policy on storage, deletion and any research use of methylation data

Lab certification and regulatory status are a separate layer from clock reliability. Our explainer on whether biological age tests are FDA approved covers how laboratory-developed tests are regulated after the 2025 rule change [10][11], and what CLIA does and does not cover.

Frequently asked questions

Why can biological age change between two tests?

Part of any change can come from technical noise. Replicate studies show some clocks vary by several years on the same sample [5]. Differences in sample type, lab batch or clock version can also shift results, so a change only becomes interpretable once it is larger than the clock's documented replicate error.

Which clocks are most reliable for repeat testing?

Published work points to clocks designed with reliability in mind. These include principal-component versions of established clocks [5] and DunedinPACE, which was built from reliability-filtered probes [4]. Reliability still depends on the lab's handling and processing, so the provider's own replicate data matter.

Is a saliva or cheek swab test less accurate than blood?

Not inherently. The concern is a mismatch between the sample and the clock's training tissue. A clock developed for buccal samples, such as CheekAge [8], is a different case from a blood-trained clock applied to saliva, so the validated tissue matters more than the sample format alone.

Related on this site

Not medical advice. This page reports published research and regulatory records for information only.

Research-only notice: Content on this site is provided for research and educational purposes only and is not intended for human use. It is not medical advice.

Translation notice: Some content on this site may have been machine-translated using Google Translate and may contain errors.

Sources

‹ GrimAge explained: how the mortality-trained epigenetic clock works