"We identified mean age at assessment, measure of educational achievement, and sample ancestry as significant moderators in individual meta-regression models (Tables S16-S23). When added simultaneously to the meta-regression model, the association pooled estimate increased to ρ = 0.50, p < 0.001 (95% CI from 0.39 to 0.61) (Fig. S5). Two moderators were significant in the model: Mean age of assessment (β = 0.014, p < 0.001, 95% CI from 0.007 to 0.021), with polygenic score predictions being stronger in adolescents than in younger children, and the measure of educational achievement (β = − 0.116, p = 0.025, 95% CI from − 0.206 to − 0.026), with predictions being higher for school grades than for standardized test scores. Sample ancestry was no longer significant (β = − 0.008, p = 0.884, 95% CI from − 0.120 to − 0.104). The Q statistic reduced substantially (i.e., by 65.64%) but remained significant (Q = 152.583, p < 0.001), occurring mainly at the within-study level (I2Level 2 = 89.91%). Yet, ANOVA model comparisons showed significant heterogeneity at both levels 2 and 3 (σ2.1 = 0.003 p < 0.001, σ2.2 = 0.001 p < 0.001)."
Kinda unclear what this means in real units, but if we assume their beta is the slope, one age older samples had 0.014 higher correlation with polygenic scores. So the difference between age 5 and age 30 should be about 25*0.014=0.35 r. Seems hard to believe though, so I assume this is not a slope.
Thanks for this. That Age x PGS prediction interaction should have been pretty expected (wilson effect, smaller children gaps &c), but is there any reason why this happened?
“Sample ancestry was no longer significant (β = − 0.008, p = 0.884, 95% CI from − 0.120 to − 0.104)”
Haven’t read it yet, so unsure if this is method-related or a byproduct of a homogeneous sample or whatever.
Good effort. By the way, it is not true that direct genetic measures don't show age interaction. In the just published meta-analysis, they found:
https://link.springer.com/article/10.1007/s10648-024-09928-4
"We identified mean age at assessment, measure of educational achievement, and sample ancestry as significant moderators in individual meta-regression models (Tables S16-S23). When added simultaneously to the meta-regression model, the association pooled estimate increased to ρ = 0.50, p < 0.001 (95% CI from 0.39 to 0.61) (Fig. S5). Two moderators were significant in the model: Mean age of assessment (β = 0.014, p < 0.001, 95% CI from 0.007 to 0.021), with polygenic score predictions being stronger in adolescents than in younger children, and the measure of educational achievement (β = − 0.116, p = 0.025, 95% CI from − 0.206 to − 0.026), with predictions being higher for school grades than for standardized test scores. Sample ancestry was no longer significant (β = − 0.008, p = 0.884, 95% CI from − 0.120 to − 0.104). The Q statistic reduced substantially (i.e., by 65.64%) but remained significant (Q = 152.583, p < 0.001), occurring mainly at the within-study level (I2Level 2 = 89.91%). Yet, ANOVA model comparisons showed significant heterogeneity at both levels 2 and 3 (σ2.1 = 0.003 p < 0.001, σ2.2 = 0.001 p < 0.001)."
Kinda unclear what this means in real units, but if we assume their beta is the slope, one age older samples had 0.014 higher correlation with polygenic scores. So the difference between age 5 and age 30 should be about 25*0.014=0.35 r. Seems hard to believe though, so I assume this is not a slope.
Thanks for this. That Age x PGS prediction interaction should have been pretty expected (wilson effect, smaller children gaps &c), but is there any reason why this happened?
“Sample ancestry was no longer significant (β = − 0.008, p = 0.884, 95% CI from − 0.120 to − 0.104)”
Haven’t read it yet, so unsure if this is method-related or a byproduct of a homogeneous sample or whatever.