Design Experiments From Reported Results

 Quick Analysis Plan Copied

 Long Analysis Plan

Experimental Design: Ancestry-Stress-Tested APOE-Exclusive PRS

Hypothesis under test (falsifiable): A PRS trained on non-European GWAS summary statistics, excluding the APOE region, will show diminished fold-risk extremes and reduced AAO separation within APOE ε4 strata unless ancestry-specific LD and allele frequencies are modeled. Failure of transferability (miscalibration) would disprove naive trans-ancestral portability. Baseline motivation: APOE ε4 dosage effects on Aβ burden attenuated sharply between a discovery cohort (n=626; ε4 β=27.75/41.93 for 1/2 alleles) and a diverse replication cohort (n=726; β=11.9/24.46), demonstrating that genetic effect magnitudes are composition-sensitive Multi-cohort imaging-genetics effect attenuation.

Note: These are reported cohort-level β values, not PRS fold-risks; they are used as sourced evidence that effect magnitudes shrink across populations — the same mechanism the stress test probes.

Analysis Plan

• Data: Obtain non-European AD GWAS summary statistics (target ancestries: African, East Asian, Hispanic/Latino, admixed American) plus a held-out target cohort with age-at-onset, APOE genotype, and genotypes. HABS-HD is a useful template: diverse replication sample with stratified subgroups and documented ancestry composition Diverse replication cohort composition.
• PRS construction (excl. APOE): Lift all SNPs to a common build; exclude chr19 APOE region (e.g., hg38: 44.9–46.5 Mb). Train with ancestry-matched LD reference panels using clumping+thresholding (PLINK/PRSice-2) and a penalized/LDpred2-style continuous-shrink model as the ancestry-aware arm. Use snakemake to version two pipelines: (a) naive-EUR-trained, (b) ancestry-retrained.
• Calibration metrics within APOE ε4 strata (0/1/2 alleles): decile-based observed/expected fold-risk ratio with 95% CI; slope of PRS on AAO via Cox/AFT models; variance explained (R² on the liability scale); calibration-in-the-large and calibration slope; stratified by ancestry.
• Pre-registered decision rule: Transferability holds if top-vs-bottom decile fold-risk ratio ≥ 1.5× the null expectation and AAO hazard separation remains significant (p<0.05, FDR-corrected) within each ε4 stratum and ancestry. Attenuation of fold-risk extremes in the naive arm that is rescued by ancestry-specific LD/AF modeling confirms the prediction; failure to rescue falsifies transferability of the naive score.

Confounds and Blind Spots

Sample size in non-European AD GWAS is often limiting, shrinking predictive power independent of LD/AF mismatch (power confound — report per-ancestry N and SE). Cross-sectional designs limit causal inference Cross-sectional causal-inference limitation. Rare-variant evidence such as protective FN1 missense burden was found only in small European-ancestry APOE ε4 homozygote samples (SKAT-O p=0.008; 3/212 cases vs 5/61 controls), underscoring that even within-ancestry strata analyses suffer small-n instability Small-n ancestry-restricted stratum evidence. Pre-register all thresholds before unblinding to avoid HARKing.