How Much Missing Heritability Is a Snapshot Artifact?

Twin and family studies say most variation in traits like blood pressure, body mass, or brain volume is inherited. Genome-wide methods that add up the effects of all measured DNA variants recover only part of that. This gap is the "missing heritability" problem. Nearly every estimate on both sides of the gap uses a phenotype measured once, at whatever age the person happened to visit the clinic. A single measurement mixes three things: a person's stable set point, how far they have drifted from it by that age, and day-to-day fluctuation plus measurement error. Each piece can have a different genetic architecture, and the transient piece adds noise that is not genetic. A snapshot can therefore understate heritability, or make it look like it depends on the age mix of the cohort, for reasons that have nothing to do with the genome.

This note proposes measuring how large that artifact is. We use cohorts with repeated visits to split each trait into a stable level, a rate of change, and transient variance. We estimate genetic variance for level and rate jointly, while accounting for how precisely each person's rate is known. We then ask what share of the twin/pedigree-versus-SNP gap closes once the transient piece is removed from the denominator and the change piece is given its own term. The answer matters in both directions. If the share is large, much of "missing heritability" is a phenotyping problem rather than a genotyping one. If it is near zero, the field can stop attributing the gap to cross-sectional design.

Core Claim

Write a person's measurement at age $a$ as $y(a) = L + aS + T + \varepsilon$. Here $L$ is the person's stable level, $S$ is their rate of change, $T$ is a transient state that does not persist between visits, and $\varepsilon$ is measurement error. If $g_L$, $g_S$ and $g_{LS}$ are the genetic variances of level and slope and their genetic covariance, and $\sigma^2_L, \sigma^2_S, \sigma_{LS}, \sigma^2_T, \sigma^2_\varepsilon$ are the corresponding total variances, then the heritability a single snapshot would report at age $a$ is

$$h^2_{\text{snap}}(a) = \frac{g_L + a^2 g_S + 2a\, g_{LS}}{\sigma^2_L + a^2\sigma^2_S + 2a\,\sigma_{LS} + \sigma^2_T + \sigma^2_\varepsilon}.$$

This expression makes three testable predictions. First, snapshot heritability is diluted by $\sigma^2_T + \sigma^2_\varepsilon$, which repeated visits can estimate and remove. Second, it shifts with the cohort's age distribution whenever $g_S$ is not negligible. Third, a large slope variance in the denominator paired with a small $g_S$ can drive a residualized static trait toward zero heritability even when the level is heritable.

The third case is a candidate explanation for a puzzle in the file. Mapping the Genetic Variation of Regional Brain Volumes as Explained by All Common SNPs from the ADNI Study used whole-genome mixed models on one baseline MRI per person. It found that common variants explain 85% of intracranial volume variance (p = 0.04), yet grey and white matter volume adjusted for head size have estimates near zero (p = 0.5). The new contribution is not a new association test. It is an accounting of the gap by component, with uncertainty, in cohorts where both pedigree and SNP estimates can be computed on the same people.

Relation to Prior Work

Two papers motivate the proposal by showing that genetic effects on traits change over time. The Function-on-Scalar LASSO with Applications to Longitudinal GWAS maps genotypes to whole systolic blood pressure curves in Framingham (1,924 subjects, up to seven exams, about 4.3 million variants). It finds a ZBTB20 variant whose protective effect fades with age. Bayesian group Lasso for nonparametric varying-coefficient models with application to functional genome-wide association studies does the same for BMI trajectories using polynomial effect curves on about 33,000 preselected variants. Both are variant-selection methods. They tell us which loci have time-varying effects, not how much total heritable variance a snapshot misses. Barber et al. also assume unrelated subjects and report shrinkage bias, both of which make them unsuitable as heritability estimators. We reuse their datasets and their evidence that effects vary with age, but our estimand is aggregate variance.

Limited overlap between genetic effects on disease susceptibility and disease survival motivates the progression side and also constrains it. Across more than 120,000 patients in seven biobanks, susceptibility polygenic scores reached hazard ratios up to 1.90 for diagnosis but were near 1.0 for post-diagnosis death. Under matched sample sizes, susceptibility scans found 33 loci and mortality scans found zero. The authors attribute this partly to mortality being a blunt endpoint and call for continuous progression measures. Our slope component is such a measure. Their result is also a warning: it is plausible that $g_S$ is genuinely small, and the design must be able to report that cleanly.

A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study supplies a genetic predictor we reuse: module-level risk scores weighted by large Alzheimer's GWAS. Its imaging and diagnostic measures are cross-sectional stage labels, which we replace with per-person atrophy rates.

Discovering non-additive heritability using additive GWAS summary statistics shows that local gene-by-gene interactions tagged by additive statistics are detectable in 24 of 25 UK Biobank traits. It is the main competing explanation for the gap. We do not model interactions. We treat that work as the alternative our decomposition must be compared against, not absorbed into.

Longitudinal Flow Matching for Trajectory Modeling generates continuous MRI trajectories from sparse ADNI scans and names genetic conditioning as future work. We deliberately do not build a genetics-conditioned generator. A forecaster's accuracy would not say how much heritable variance a snapshot hides, which is the question here.

Research Program

The estimator is a joint random-regression genetic mixed model. Person-level intercept and slope each receive a genetic random effect, with covariance given by the genomic relationship matrix. In Framingham, a parallel version uses the pedigree relationship matrix. The model also includes a person-level non-genetic intercept and slope and a visit-level residual. With at least three visits per person, the visit-level residual is identifiable separately from slope variance. Fitting on the visit-level data rather than on extracted slopes lets people with two widely spaced visits contribute less slope information than people with seven. This is the only methodological departure from standard practice, and it is modest. To keep the comparison honest, age, sex, ancestry principal components, medication (antihypertensive use for blood pressure, as an offset sensitivity analysis), scanner and field strength (for ADNI) enter as fixed effects at the visit level.

The work has three stages. The first is a simulation study that uses real Framingham and ADNI genotype matrices. We plant known values of $g_L$, $g_S$, $g_{LS}$ and $\sigma^2_T$ and impose the cohorts' actual visit schedules and dropout patterns. This checks whether each estimator recovers the planted components and whether the snapshot formula above predicts the bias of single-visit estimates. We include a scenario where dropout depends on the unobserved slope, as happens in Alzheimer's cohorts, to measure how much rates are biased under missing-not-at-random follow-up. The second stage is the Framingham decomposition for systolic blood pressure and BMI. Pedigree and SNP estimates are computed per component, so the gap can be reported separately for snapshot, level, and slope. The third stage applies the same decomposition to ADNI regional volumes, alongside a better-powered polygenic-score test described below. We expect ADNI's sample size to make direct heritability estimates of slopes too imprecise to be decisive.

Baselines

The naive reference point is single-visit genomic-relationship heritability, the design of Mapping the Genetic Variation of Regional Brain Volumes as Explained by All Common SNPs from the ADNI Study. We run it once on the baseline visit and once on a randomly chosen visit per person, using the same samples, covariates and relationship matrix as every other method. The reported figure to reproduce in ADNI is 85% of intracranial volume variance (p = 0.04), with near-zero estimates for head-size-adjusted grey and white matter (p = 0.5). Reproducing these figures within their uncertainty is a sanity check before any comparison.

The second baseline is the repeatability approach: heritability of each person's mean over visits. It removes some transient variance from the denominator but folds slope into level. Comparing it with the joint model isolates how much the change component contributes beyond noise reduction.

The strongest competitor is the two-stage slope approach. We fit a per-person least-squares or empirical-Bayes slope, then run standard heritability estimation on it. This is what most applied groups would do, and it is cheap. It is run on identical visits and covariates. The joint model beats it only if, under the real irregular schedules, it recovers planted $g_S$ with materially lower bias or interval width in simulation. Our pre-specified bar is that bias reduction of at least a third or narrower intervals are required to claim an advantage. Without that, the two-stage method is the recommended tool and our contribution reduces to the decomposition itself.

For effect-curve variant selection, the published Framingham results of The Function-on-Scalar LASSO with Applications to Longitudinal GWAS are the reference. That paper reports a locus, not a heritability, so no reported result was found on Lacuna for trajectory heritability of systolic blood pressure in Framingham. The same holds for pedigree heritability of any Framingham trait. These quantities are new measurements, not comparisons against a stated number.

Two ablations isolate the key ingredients. The first drops the transient term, forcing all within-person scatter into slope or residual. The second drops per-person precision weighting by fitting the joint model to visit-thinned data in which everyone has exactly two visits. Every method gets the same restricted maximum-likelihood fitting routine and the same bootstrap over individuals for intervals.

Why not just…?

Why not just run standard heritability estimation on per-person slopes? This is the real threat, and it may well be enough when people have many evenly spaced visits. The weakness is that slopes from two or three closely spaced visits carry very unequal noise, which attenuates slope heritability. Treating them as exact is the same error that makes snapshots understate heritability. The visit-thinning experiment in Framingham settles it. We estimate slope heritability from all seven exams, then from random subsets of two or three, and see which estimator stays close to the full-data value. If the two-stage method does, we adopt it.

Why not just average repeated measurements? Averaging is the cheapest way to shrink transient variance and will recover part of the gap. It cannot say whether the remaining gap sits in change, and in an aging cohort the mean is itself a mixture of level and accumulated drift. The comparison between repeatability and joint-model level heritability measures exactly what averaging leaves on the table. If that difference is negligible, averaging is the practical recommendation.

Why not just include age and age squared as covariates in a large cross-sectional GWAS? Age covariates remove a population-average age curve. They do not remove person-specific slopes or transient states from the residual variance, so they cannot change the denominator of the snapshot formula. The Yang et al. result, where age-adjusted susceptibility scores predicted onset (hazard ratio up to 1.90) but not post-diagnosis survival (near 1.0), shows that cross-sectional architectures do not carry over to within-person change. Our simulations show directly how much bias remains after age adjustment.

Why not just attribute the gap to interactions, or wait for bigger GWAS? Discovering non-additive heritability using additive GWAS summary statistics makes interactions a genuine competing explanation. Our decomposition does not rule it out. The two explanations are complementary accountings of the same gap, and we report what fraction ours closes rather than claiming it closes the whole gap. Larger cross-sectional samples shrink error bars around a snapshot estimate but leave its expected value unchanged under the formula above. More snapshots cannot remove transient variance from a single measurement.

Proposed Experiments

The first decisive check is in Framingham, the only cohort in the file with both multigenerational pedigrees and dense genotyping. For systolic blood pressure and BMI, we compute the pedigree-minus-SNP gap three ways: for a single-visit phenotype, for the level component, and for the slope component. The snapshot-artifact hypothesis predicts a smaller gap for level than for single visits, with the difference tracking the estimated transient share. If the gaps are statistically indistinguishable, the hypothesis is falsified for these traits.

The second check tests the age-dependence prediction. We fit single-visit heritability within age bands and compare the observed trend against $h^2_{\text{snap}}(a)$ computed from the jointly estimated components, with no refitting. The reported ZBTB20 pattern, an effect that fades with age, gives an expected direction for blood pressure.

The third check addresses the Yang et al. tension in ADNI, where direct slope heritability will be underpowered. We build an Alzheimer's polygenic score and the module-level scores of the NetWAS paper, both with and without the APOE region. We test whether they predict per-person hippocampal and grey-matter atrophy rates after conditioning on baseline volume. Atrophy rates come from the same mixed model, with a sensitivity analysis for slope-dependent dropout. If susceptibility scores predict atrophy rate while Yang et al. found them near-null for mortality, the endpoint was masking progression genetics. If they are null for atrophy too, that supports genuinely different genetic architectures for onset and progression. Either outcome answers the question the survival paper left open.

The simulation study is the gate for everything else. If no estimator recovers planted slope variance under ADNI's real visit schedule, the ADNI heritability analyses are reported as descriptive only, and the claims rest on Framingham.