Genetic studies of complex traits usually ask how much of the variation in a trait's level comes from genetic differences. Many diseases, though, unfold in stages: tissue changes first, and symptoms follow after a delay. Two people with similar genetic risk can show similar brain shrinkage but begin losing memory years apart. We propose to treat that delay as a trait in its own right. For each person we estimate how long the decline in an upstream measurement (hippocampal volume on MRI) takes to show up in a downstream one (cognitive test scores). We then ask whether this person-specific lag is shaped by genotype. If it is, models that link the two measurements at a fixed delay, or at the same moment, average across people whose stages are out of step. That averaging would hide genetic signal that lives in timing rather than in level.
The tool comes from lagged correlation analysis, which finds the time shift at which two series line up best. The cross-correlation function, the correlation between $X(t)$ and $Y(t+\tau)$ plotted across shifts $\tau$, is standard in epidemiology and climate science. Prison transfers led COVID-19 incidence by 3 to 5 weeks in Brinkley-Rubinstein et al., and a 3-month phase error in simulated river discharge produced peak-discharge errors of 18% to 117% in the CESM1 Evaluation. The missing heritability problem is the gap between the large genetic share of trait variation estimated from twins and the small share recovered from DNA markers. It has not yet been approached through between-stage delays. The concrete deliverable is a mixed model in which the lag carries a genetic random effect. It will be tested on simulated phenotypes built on real genotypes and on the Alzheimer's Disease Neuroimaging Initiative (ADNI).
What Exists and What Is New
Three lines of longitudinal genetics come closest, and each keeps everyone on one shared clock. Functional MINQUE estimates genetic covariance functions for a single trajectory. Applied to lung function in the Childhood Asthma Management Program, it attributes about 46% of variation to common variants. FS-LASSO regresses blood-pressure curves on millions of variants and finds a ZBTB20 variant whose effect fades with age, but it evaluates every effect at the person's concurrent age. PFLR is the most direct precedent. It links baseline hippocampal surface shape plus selected variants in ADNI to cognitive scores measured exactly 12 months later. It reports that hippocampal shape explains 1% to 4.6% of future cognitive variation.
These papers motivate the proposal, because each finds time-varying genetic structure. None lets the delay between two measurements differ across people, and none asks whether that delay is genetic. MPRS-NetWAS shows that amygdala change statistically carries part of the path from Alzheimer's genetic risk to diagnosis, but its analysis is cross-sectional. It also supplies a reusable method: module-level polygenic scores weighted by the IGAP and Jansen et al. meta-analyses, which we adopt as genetic predictors.
The ADNI GCTA study estimates the genetic share of variation from all common markers at once, the quantity called SNP heritability. It reports 84.5% for intracranial volume with $p = 0.04$ in 747 people. That result constrains the proposal. A $p$ value that marginal for an estimate that large means whole-genome variance estimates at ADNI scale carry standard errors comparable to the estimate itself. A lag heritability estimated in a few hundred people will be similarly imprecise, so our primary genetic test must be a polygenic-score association rather than a variance estimate.
Core Claim and Model
The claim is that between-stage delays carry genotype-dependent variation that both concurrent and fixed-lag models discard. We further expect that a model letting genotype shift the delay will predict future cognition better than one that shifts only a single person-level disease clock.
For person $i$ with cognitive score $C_i(t)$ and smoothed hippocampal volume curve $H_i(t)$, we fit
$$C_i(t) = \alpha + a_i + \beta\, H_i(t - \tau_i) + \varepsilon_i(t), \qquad \tau_i = \bar\tau + \gamma^\top s_i + u_i + e_i, \qquad u \sim N(0, \sigma_g^2 K).$$
The terms are defined as follows. The random intercept $a_i$ has deliberately limited variance. The delay $\tau_i$ is the person's atrophy-to-decline lag, with population mean $\bar\tau$. The vector $s_i$ holds polygenic and module-level scores with weights $\gamma$, and $K$ is the genetic relatedness matrix computed from markers. Genetic and residual parts of the lag are $u_i$ and $e_i$. The lag heritability is $h^2_\tau = \sigma_g^2/(\sigma_g^2+\sigma_e^2)$. Fitting runs by profile likelihood over a grid of $\tau$ values, initialized at each person's cross-correlation peak computed on first-differenced, prewhitened series.
There is a critical identification point. If both trajectories were straight lines, shifting $H$ by $\tau_i$ would be indistinguishable from changing the intercept $a_i$. The lag is therefore identified only by curvature, meaning the timing of acceleration in decline. This forces two choices. We restrict the analysis to people who enter with mild cognitive impairment or Alzheimer's disease and have at least four visits. We also constrain lags to the observed follow-up span, so that $H_i$ is never extrapolated. Sample size will fall well below ADNI's 747 genotyped participants. We accept that cost rather than report lags that are artifacts of a linear fit.
Research Program
The work has three stages. The first is a simulation study built on ADNI's real genotypes. We generate sigmoidal atrophy and cognition curves with known $h^2_\tau$ ranging from 0 to 0.6, and sample them at ADNI's actual visit schedules, so that visit density, missing visits, and dropout match the real data. On these simulations we measure two things. One is how much genetic variance the concurrent and fixed-lag approaches recover at each age. The other is whether our model recovers $\bar\tau$, $\gamma$, and $h^2_\tau$ without bias. The second stage fits the model to ADNI, pairing hippocampal volume with the Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS11) and the Mini-Mental State Examination (MMSE), the outcomes PFLR used. The third stage swaps in the cross-tissue gene-expression risk scores of Zhao et al. as the vector $s_i$. This tests whether genes acting through brain structure predict delay, not just level.
Baselines
The naive reference fits a random-intercept, random-slope model of cognition on concurrent hippocampal volume, a lag of zero for everyone. It then estimates the SNP heritability of each person's cognitive slope with GCTA-style restricted maximum likelihood. Any lag model must at least beat this in held-out prediction.
The fixed-lag baseline re-implements PFLR at its 12-month lag and at the best single population lag chosen by cross-validation. Its reported result is that hippocampal surface explains 1% to 4.6% of future cognitive variation, and genetics explains 62.69% to 78.01%. Those genetic figures come from selecting variants with an $\ell_0$ penalty within a sample of 606 people. We therefore rerun PFLR inside the same held-out cross-validation folds rather than compare against the published numbers.
The concurrent functional baselines are FS-LASSO, refit without penalty after its selection step, and F-MINQUE, applied to cognition curves. F-MINQUE's only reported figure is about 46% for lung function in CAMP. We found no reported result on Lacuna for either method on ADNI cognition.
The key ablation is a shared-clock model, a disease-progression model in which each person has one time shift applied to both measurements. That forces the atrophy-to-decline lag to be identical across people. We found no reported result on Lacuna for such a model on ADNI.
All models use the same participants, visits, covariates, polygenic scores, and five-fold cross-validation split by person. Each gets an equal hyperparameter search budget. Beating the baselines means two things. The first is a gain in held-out $R^2$ for the last observed cognitive score that is significant by a paired bootstrap across folds. The second is a polygenic-score effect on $\tau_i$, the coefficient $\gamma$, that is nonzero at a pre-registered threshold and survives the controls described under Decisive Checks.
Why not just…?
Why not just run a larger cross-sectional GWAS? Larger studies find more variants associated with trait level, but they cannot detect a genetic effect on timing that only appears when two trajectories are compared. The FS-LASSO ZBTB20 result shows that age-dependent effects are invisible in a single snapshot. The naive baseline tests this alternative directly. If the slope-heritability model matches ours, timing adds nothing.
Why not just use random slopes or a static polygenic score on decline rate? This is a serious alternative. A variant that speeds decline also changes when decline appears, so slope and lag are correlated. We will settle it with a likelihood comparison of nested models. We also fit simulations in which the slope is genetic but the lag is not. If the lag term absorbs genetic signal in those simulations, the lag claim is not identified and we will report that.
Why not just fit a shared latent-time disease-progression model? This is the strongest competitor, and it is our ablation. It captures the possibility that some people are simply further along. It cannot capture a gap between stages that differs across people. If the shared-clock model predicts as well and $\gamma$ is null, we conclude that apparent lags are disease stage. We would then drop the lag-heritability claim.
Why not just compute each person's cross-correlation peak and run GCTA on it? With four to eight visits per person, the peak is very noisy, and that noise attenuates any heritability estimate toward zero. The simulations quantify this attenuation against the joint model, in which genetic shrinkage of $\tau_i$ is meant to reduce it.
Decisive Checks
The decisive checks are four. First, with $h^2_\tau = 0$ in simulation, the false-positive rate for $\gamma$ must stay at the nominal level. Second, permuting genotypes across people must remove the effect. Third, reversing the direction of the lag, so that cognition leads atrophy, must give a null or negative result. This mirrors the reverse-causality check in Brinkley-Rubinstein et al.. Fourth, the result must survive prewhitening, which removes shared trends that can create spurious alignment between two declining series. A positive outcome requires held-out prediction gains over the shared-clock ablation together with an association between Alzheimer's polygenic or module scores and lag. Under the GCTA power limits, $h^2_\tau$ itself will be reported only as an estimate with a wide interval.
References
- The association between intersystem prison transfers and COVID-19 incidence in a state prison system
- Evaluation of hydroclimatic biases in the Community Earth System Model (CESM1) within the Mississippi River basin
- Estimating Variance Components in Functional Linear Models With Applications to Genetic Heritability
- The function-on-scalar LASSO with applications to longitudinal GWAS
- A partially functional linear regression framework for integrating genetic, imaging, and clinical data
- A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study
- Mapping the Genetic Variation of Regional Brain Volumes as Explained by All Common SNPs from the ADNI Study
- Transcriptome-wide association analysis of brain structures yields insights into pleiotropy with complex neuropsychiatric traits