Large biobanks release genetic data in waves. Each new release adds participants, and analysts rerun the same pipeline each time to ask which parts of the genome carry more than their share of a trait's heritability. A typical question is whether intron or regulatory annotations are "enriched." Every rerun is another chance to declare enrichment by luck. Many of the traits involved, such as diabetes, heart disease, smoking and alcohol use, rest partly on what participants say about themselves rather than on measurements or medical records. Clinical meta-analysis has a tool for the first problem. Trial sequential analysis (TSA) treats a growing pile of trials as one trial with repeated interim looks. It asks whether enough total evidence has accumulated before a result counts as definitive, and it sets significance thresholds that start strict and relax as information grows. No comparable tool exists for genomic enrichment, and none accounts for the extra noise or bias that self-reported traits introduce.
This note proposes to build that tool and test it. The tool computes, for each functional annotation, how much biobank sample is required before an enrichment claim is conclusive. That requirement is inflated by measured self-report error. Each data release is then checked against boundaries that control false positives across all looks and all annotations. The central observation is this. When people misreport in a way unrelated to their genes, the error should leave the enrichment ratio roughly unchanged and only make its estimate noisier, so the right correction is a larger required sample. When misreporting is itself heritable, for example tied to education or personality, it can manufacture enrichment, and no amount of extra data fixes that. The method must tell these two cases apart and report them differently.
Core Claim
Sequential monitoring can be transferred from trial meta-analysis to stratified heritability enrichment. Stratified LD score regression (S-LDSC) is the method that splits a trait's heritability across genome annotations using only association summary statistics. Monitoring means each annotation gets its own information-size target and boundaries that tighten as data accumulate.
For self-reported binary traits, the information-size target should be scaled by the observed reliability of the phenotype. The standard attenuation argument gives the scaling. If self-report has sensitivity $Se$ and specificity $Sp$ against a reference diagnosis, and errors are independent of genotype, then genetic signal is diluted by a factor of roughly $(Se+Sp-1)$. The noncentrality of association tests, which grows with the square of the effect, falls by its square. The adjusted target is therefore
$$ N_{\text{req}}^{\text{adj}} = \frac{N_{\text{req}}}{(Se+Sp-1)^2}, $$
where $N_{\text{req}}$ is the required information size computed for an error-free phenotype. Because dilution scales every annotation's heritability by the same factor, the enrichment ratio (an annotation's share of heritability divided by its share of SNPs) is unchanged in expectation. Only the time to conclusiveness changes.
The second half of the claim covers misreporting that is itself heritable. That case breaks the proportional scaling and can shift enrichment toward annotations tied to the misreporting behavior. It must therefore be flagged as bias, not absorbed into sample size.
What would be new is three things together. The first is a sequential, family-wise error-controlled decision rule for enrichment across data releases. The second is an information-size target tied to measured phenotype reliability. The third is an explicit diagnostic that separates dilution from directional self-report bias.
How This Sits Relative to Prior Work
The TSA papers in the file supply the method, and they also show its limits. The probiotic colorectal surgery meta-analysis (Exploring the clinical value of probiotics...) shows a cumulative Z-curve crossing sequential boundaries before the required information size (1,602 of 2,897 patients). It is the template for an "early but conclusive" verdict. The denosumab analysis (Efficacy and safety of denosumab...) shows the opposite case: rare-event safety outcomes far from their required size remain inconclusive. That is exactly the regime we expect for small annotations under noisy self-report.
Both papers, and the critical-care applications such as Continuous vs. intermittent meropenem infusion..., monitor one clinical outcome over chronologically ordered trials. Existing TSA software has no notion of many correlated hypotheses or of genomic summary statistics. We keep the alpha-spending machinery but replace "trials ordered by year" with "nested biobank releases." We also replace one endpoint with a family of annotations.
The enrichment papers constrain the design. Assumptions about frequency-dependent architectures of complex traits bias measures of functional enrichment shows that the S-LDSC allele-frequency parameter (default $\alpha=1$, versus an empirical average near 0.38) systematically shrinks enrichment estimates across 47 traits and 96 annotations. A sequential rule built on a biased estimator would converge confidently to the wrong answer. We therefore run monitoring under both settings and treat disagreement as a reason to withhold a "conclusive" verdict.
Probabilistic inference of the genetic architecture underlying functional enrichment of complex traits analyzes 382,466 UK Biobank and 32,500 Estonian Biobank participants for height, BMI, coronary artery disease and type 2 diabetes. It reports introns explaining 32–44% of genetic variance and proximal regulatory regions at most 10%. These fixed-sample estimates are the endpoints our sequential curves should approach, and they set plausible anticipated effect sizes for the required-size calculation.
Two papers address the many-correlated-hypotheses issue, and we borrow from them selectively. pareg fuses estimates of overlapping pathways. gtCLIP clusters redundant gene sets by embedding them; it notes that over 80% of related pathway pairs share no genes. We use annotation correlation, estimated from the S-LDSC jackknife, to spend alpha across the annotation family rather than Bonferroni-splitting it. We do not adopt either method's estimator. Finally, funkea provides a unified, Spark-backed S-LDSC implementation that makes rerunning enrichment at many looks feasible.
The research file contains no papers on self-report bias itself. It also contains no prior work on sequential monitoring of enrichment; its section on closest existing work was empty or truncated. The novelty claim therefore rests on the absence of such work in the file, not on a verified literature search.
Research Program
The first component is the monitoring statistic. At each release $t$, S-LDSC run on the cumulative summary statistics gives an enrichment estimate and a block-jackknife standard error for each annotation. These are converted to a Z-score for enrichment minus one. Because each release's sample contains the previous one, the cumulative Z-scores should behave approximately like a process with independent increments indexed by information fraction. That approximation is what licenses Lan-DeMets alpha spending (a rule for distributing the false-positive budget across looks). We verify this empirically rather than assuming it, by checking the correlation of successive Z-scores against the square root of the ratio of their information fractions.
The second component is the required information size per annotation. It is computed from four inputs:
- an anticipated enrichment taken from the earliest release or from external estimates such as BayesRR-RC;
- the annotation's size and its LD-score variance;
- a heterogeneity term for between-cohort differences, analogous to TSA's diversity adjustment;
- the reliability factor from the core claim.
Reliability is measured, not assumed. Where a biobank holds both self-reported and record-linked diagnoses for the same people, as UK Biobank does for diabetes and heart disease, $Se$ and $Sp$ are estimated directly against hospital records.
The third component is the bias diagnostic. For each self-reported trait we also run enrichment on the record-confirmed version in the subset that has it. We then test whether the vector of enrichment ratios differs beyond what dilution predicts. Dilution predicts equal ratios with larger variance. A directional bias predicts annotation-specific departures. As a second check, we compute the genetic correlation between "self-report minus record" discordance and the trait, and between the discordance and education-related summary statistics. Annotations whose enrichment is carried by the discordance component are reported as "bias-suspect" instead of being monitored to conclusion.
Baselines
The naive reference point is cumulative S-LDSC with a nominal per-look threshold. At every release, each annotation is declared enriched if its two-sided p-value is below 0.05, or below the Benjamini–Hochberg threshold across annotations. This is how enrichment is actually reported release after release. It is the procedure whose false-positive inflation we expect to expose, and it is run on exactly the same cumulative summary statistics and annotations as our method.
The strongest existing method is fixed-sample S-LDSC on the final, largest release, with the allele-frequency parameter set to the better-supported value (0.38) per the frequency-dependence study. Annotations are corrected by Bonferroni or FDR. That study reports qualitative shifts rather than a single error-rate figure. The BayesRR-RC study supplies variance shares (introns 32–44%, distal regulatory 22–28%, coding 12–25%) as reference values. For the metric that matters here, false-positive rate across repeated looks or time-to-correct-conclusion, no reported result was found on Lacuna for any enrichment method.
Fair comparison means both procedures see the same data and annotation set and have the same total alpha. The fixed-sample baseline is allowed one look at the final release. Our method looks at every release. Beating it means two things. First, our method matches its final-release power within a few points. Second, it reaches correct verdicts on strongly enriched annotations at earlier releases while keeping family-wise error at or below 5% across all looks.
Two ablations isolate the ingredients. The first removes the reliability factor, which is standard TSA boundaries applied to enrichment Z-scores. It tests whether the self-report adjustment matters. The second removes the bias diagnostic and monitors self-reported traits to conclusion regardless. It tests whether the diagnostic catches cases where boundaries alone would certify a spurious enrichment.
Why not just…?
Why not just analyze only the final release with FDR control? For a single paper this is correct and cheaper. But biobank results are in fact published and acted on at each release, and the question is what to conclude now. The fixed-sample baseline cannot say whether an interim result is premature. The experiments measure how often interim nominal hits fail to survive later releases. If that rate is low, the case for sequential monitoring weakens to a convenience. We will report it either way.
Why not just use record-confirmed phenotypes and drop self-report? This is the most serious alternative. If record-confirmed subsets retain most of the power, the reliability adjustment is moot. Record linkage shrinks case counts and is itself selective, since hospitalized cases skew severe. We compare time-to-conclusiveness and final enrichment estimates for record-only, self-report-only and combined phenotypes on diabetes and heart disease. If record-only matches combined power at every release, we will recommend it and present the reliability factor only as a planning tool.
Why not just use liability-scale conversion or LDSC's intercept to handle misclassification? Liability conversion rescales total heritability but leaves the standard error of an enrichment ratio unchanged. It therefore does not alter how much data is needed. The LDSC intercept absorbs confounding that is uniform across the genome, not heritable misreporting that loads on particular annotations. The bias diagnostic exists for exactly that case. Its value is shown if, in the planted-bias simulation below, intercept-based correction leaves the spurious enrichment in place while the diagnostic flags it.
Why not just apply standard TSA unchanged? That is the first ablation. With nondifferential error it should be anticonservative about when conclusions are reached: its information targets are too small, so it declares small annotations conclusive too early. If simulations show no meaningful difference at realistic reliability levels, the self-report component reduces to a footnote.
Decisive Experiments
All experiments use UK Biobank and Estonian Biobank summary statistics as the data source, following the BayesRR-RC study. They are run through funkea's S-LDSC on the 96-annotation baseline model used in the frequency-dependence study. Releases are emulated as nested random fractions (10% through 100%) of the cohort, and also ordered by recruitment centre to mimic non-random accrual. Height and measured BMI serve as low-error controls, and diabetes and heart disease as self-report-affected traits.
The type-I error check uses null annotations matched on size and LD-score distribution but randomly placed. It asks whether family-wise error across all looks stays at or below 5% under our boundaries, and how far the naive per-look procedure inflates it. The power check measures, for annotations BayesRR-RC identifies as strongly enriched, such as introns, the release at which each method reaches a verdict and whether that verdict agrees with the full-cohort estimate.
The self-report check replaces record-confirmed status with self-report in increasing proportions. It verifies that enrichment ratios stay stable while time-to-conclusiveness lengthens by roughly $1/(Se+Sp-1)^2$, using the measured $Se$ and $Sp$.
The planted-bias check creates synthetic phenotypes in which misreporting probability depends on a polygenic score built from brain-expressed annotations. It asks whether boundaries alone certify spurious enrichment of those annotations, and whether the diagnostic flags them. The proposal succeeds if four things hold: family-wise error is controlled where the naive procedure is not; strong enrichments reach correct verdicts earlier than the fixed-sample baseline would permit; the reliability factor predicts observed slowdowns; and planted bias is caught. It fails if record-only phenotypes make the adjustment unnecessary, or if successive enrichment Z-scores violate the independent-increments approximation badly enough that the boundaries do not hold their nominal error rate.
References
- Exploring the clinical value of probiotics in the perioperative management of colorectal cancer: Meta‑analysis with trial sequential analysis
- Efficacy and safety of denosumab compared with conventional medications in the treatment of osteoporosis: A meta-analysis with trial sequential analysis
- Continuous vs. intermittent meropenem infusion in critically ill patients with sepsis: a systematic review and meta-analysis of randomized controlled trials with trial sequential analysis
- Assumptions about frequency-dependent architectures of complex traits bias measures of functional enrichment
- Probabilistic inference of the genetic architecture underlying functional enrichment of complex traits
- Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression
- Contrastive Learning for Gene Set Enrichment Analysis Post-Processing
- funkea: Functional Enrichment Analysis in Python