iiinova — where your digital self keeps evolving

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

I will be presenting this work in a poster about this work at PEQG in June, and I would love to discuss it and learn from your feedback! (15/15)

110 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

We've already switched to block jackknife in our own work: the forthcoming new version of PGSUS (www.biorxiv.org/content/10.1...), where accurate sib-GWAS SEs matter a lot. (14/15)

biorxiv.org

120 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

Block jackknife can be straightforwardly applied with current family GWAS software — PLINK and snipar. We also provide a simple formula for choosing the number of resampling replicates, which balances precise estimates and computational cost. (13/15)

100 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

It assumes neither homoskedastic residuals (unlike OLS) nor a null effect (unlike permutation). It's approximately unbiased under all scenarios we tested. (11/15)

100 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

Specifically, we recommend a block jackknife SE estimator. It estimates uncertainty by repeatedly resampling the data — dropping random blocks of families and re-estimating the effect each time. The variance across resamples approximates the true uncertainty. (10/15)

210 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

Enter our hero: more of the same! (a.k.a. resampling.) (9/15)

100 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

Biases in both methods get worse for rare alleles and small samples — the settings where accurate estimates of uncertainty matter most. (8/15)

100 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

In UKB data, we see that the stronger the evidence for a real allelic effect (smaller p-value), the more the permutation overestimates uncertainty – exactly as predicted by theory. We see the same pattern in the OLS uncertainty estimates. (7/15)

100 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

But the permutation method assumes the true effect is zero. Turns out it's most biased for the SNPs we care about most: the ones with non-zero effects. It overestimates uncertainty, making real signals look noisier than they are. (6/15)

100 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

The permutation method fixes the heteroskedasticity problem. It breaks the genotype-phenotype association within families and builds a null distribution of effect estimates. The sample variance of the null distribution approximates the uncertainty. No homoskedasticity assumption needed. (5/15)

120 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

In reality, families differ in genetic backgrounds and environments. This means residual variance isn't equal across families, violating OLS's core assumption. The result is biased SEs whenever allelic effects vary across families or background noise is heteroskedastic. (4/15)

100 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

The ordinary least squares (OLS) method estimates uncertainty from how scattered the data are around the regression line. OLS is simple and fast, but it assumes all families contribute equally to estimation noise. (3/15)

100 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

Family studies like sib-GWAS are the gold standard for estimating direct genetic effects. But uncertainty in these estimates feeds everything downstream: polygenic scores, Mendelian randomization, and measuring assortative mating. Getting the uncertainty (e.g., SE) wrong can propagate errors. (2/15)

110 0

xinyimiao.bsky.social @xinyimiao.bsky.social · 15/05/2026

I am excited to share my first paper, our new preprint: www.biorxiv.org/content/10.6... with @arbelharpak.bsky.social and @docedge.bsky.social. Family GWAS provides direct genetic effect estimates and is increasingly important. But how much can we trust these estimates? (1/15)

biorxiv.org

22514 0