The accuracy of LD Score regression as an estimator of confounding and genetic correlations in genome‐wide association studies

Lee, James J.; McGue, Matt; Iacono, William G.; Chow, Carson C.

doi:10.1002/gepi.22161

Cited by 58 publications

(57 citation statements)

References 47 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The level of bias in h 2 SNP was higher in the LDAK simulations than for the GCTA simulations (Appendix 4.1). Our finding of inadequate adjustment for confounding is concordant with the results of two recent analyses of stratified populations (DeVlaming et al, 2017;Berg et al, 2018), but not with Lee et al (2018) who considered confounding by parental genotype. This is because parental average genotype generates an additive genetic effect, which inflates the slope but not the intercept of the summary statistic regression.…”

Section: Confounding (Cor(x Z) = 0)supporting

confidence: 89%

“…Furthermore, if the level of overlap is known, then the inflation in the expected statistic can be predicted in advance. In contrast, the population structure in our simulations implies some relatedness among all individuals, leading to a relationship between confounding and LD that cannot be modelled using A or C. The example of non-trivial intercepts in Lee et al (2018) appropriately correcting for confounding is similar to sample overlap, as it was based on twins, so can be viewed as two dependent samples, each consisting of unrelated individuals. q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qqq LDAK model q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qqq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q Figure 6 (a) shows that use of a well-chosen reference panel (which is the case for our simulations) has minimal impact on parameter estimates compared with computing LD coefficients from the GWAS genotypes (In-sample estimates).…”

Section: Sample Overlap and Mis-specification Of The Heritability Modelmentioning

confidence: 94%

See 1 more Smart Citation

Summary statistic analyses do not correct confounding bias

Holmes

Speed

Balding

2019

Preprint

View full text Add to dashboard Cite

LD SCore regression (LDSC) has become a popular approach to estimate confounding bias, heritability and genetic correlation using only genome wide association study (GWAS) test statistics. SumHer is a newly-introduced alternative with similar aims. We show using theory and simulations that both approaches fail to adequately account for confounding bias, even when the assumed heritability model is correct. Consequently, these methods may estimate heritability poorly if there was inadequate adjustment for confounding in the original GWAS analysis. We also show that choice of summary statistic for use in LDSC or SumHer can have a large impact on resulting inferences. Further, covariate adjustments in the original GWAS can alter the target of heritability estimation, which can be problematic for test statistics from a meta-analysis of GWAS with different covariate adjustments.

show abstract

Section: Confounding (Cor(x Z) = 0)supporting

confidence: 89%

Section: Sample Overlap and Mis-specification Of The Heritability Modelmentioning

confidence: 94%

Summary statistic analyses do not correct confounding bias

Holmes

Speed

Balding

2019

Preprint

View full text Add to dashboard Cite

show abstract

“…This extension therefore predicts an inflation of the bivariate LDSC intercept if relatives span both studies, which is more general than the restriction to actual sample overlap. Another contribution by Lee et al (2018) 8 is also worth mentioning here as it provides a rigorous mathematical framework that not only refines our understanding of the LD score regression methodology but also helps clarifying its interpretation. These two examples, among many others, both illustrate the effervescence of researches driven by the LD score regression methodology.…”

Section: Discussionmentioning

confidence: 98%

“…Equation (8) shows a strong similarity with equation (1) from Bulik-Sullivan et al (2015) but one can already notice the inclusion of the second term on the right side of the equation, representing the contribution of population stratification within and between cohorts. We now further simplify equation (8).…”

Section: Notationsmentioning

confidence: 99%

See 1 more Smart Citation

Expectation of the intercept from bivariate LD score regression in the presence of population stratification

Yengo¹,

Yang

Visscher

2018

Preprint

View full text Add to dashboard Cite

Linkage disequilibrium (LD) score regression is an increasingly popular method used to quantify the level of confounding in genome-wide association studies (GWAS) or to estimate heritability and genetic correlation between traits. When applied to a pair of GWAS, the LD score regression (LDSC) methodology produces a statistic, referred to as the bivariate LDSC intercept, which deviation from 0 is classically interpreted as an indication of sample overlap between the two GWAS. Here we propose an extension of the theory underlying the bivariate LDSC methodology, which accounts for population stratification within and between GWAS. Our extended theory predicts an inflation of the bivariate LDSC intercept when sample sizes and heritability are large, even in the absence of sample overlap. We illustrate our theoretical results with simulations based on actual SNP genotypes and we propose a re-interpretation of previously published results in the light of our extended theory.Initially introduced in Bulik-Sullivan et al. (2014) 1 , the LD score regression (LDSC) methodology relies on a derivation for a particular SNP j of the expectation of its association χ 2 -statistic (χ 2 j ) as function of the LD score ( j ), that is the sum of pairwise squared correlation between minor allele counts at SNP j versus all SNPs or versus SNPs in its vicinity, the heritability (h 2 ) of the trait plus a term indicating the level of confounding in the GWAS attributable to population stratification. The derivation proposed in Bulik-Sullivan et al. (2014) predicts that the latter term (a.k.a the LDSC intercept) increases linearly with the sample size (N ) of the GWAS and the heritability of the trait, which raises a number of challenges in its interpretation as an indication of confounding when N is very large (N ∼400,000 for example). This problematic behavior of the LDSC intercept has been underlined in a recent publication by Loh et al. (2017) 2 , which recommends the use of an alternative statistic to quantify the influence of population stratification on GWAS results.The LDSC methodology was later extended in Bulik-Sullivan et al. (2015) 3 to analyse pairs of GWAS in order to quantify the genetic correlation between focus traits of each GWAS. The theory proposed in Bulik-Sullivan et al. (2015) introduced the bivariate LD score intercept, obtained from the regression of the product of association statistics (defined further below) onto LD scores, as a measure of sample overlap between the two GWAS. Compared to Bulik-Sullivan et al. (2014), the model underlying the bivariate LDSC methodology does not consider the presence of population stratification within and between studies, neither discusses how this might contribute to the expectation of the bivariate LDSC intercept. We therefore address this question here and propose an extension of the initial theory.

show abstract

Genetic meta‐analysis of obsessive–compulsive disorder and self‐report compulsive symptoms

Smit

Cath

Zilhão

et al. 2019

American J of Med Genetics Pt B

View full text Add to dashboard Cite

We investigated whether obsessive-compulsive (OC) symptoms from a populationbased sample could be analyzed to detect genetic variants influencing obsessivecompulsive disorder (OCD). We performed a genome-wide association studies (GWAS) on the obsession (rumination and impulsions) and compulsion (checking, washing, and ordering/precision) subscales of an abbreviated version of the Padua Inventory (N = 8,267 with genome-wide genotyping and phenotyping). The compulsion subscale showed a substantial and significant positive genetic correlation with an OCD case-control GWAS (r G = 0.61, p = .017) previously published by the Psychiatric Genomics Consortium (PGC-OCD). The obsession subscale and the total Padua score showed no significant genetic correlations (r G = −0.02 and r G = 0.42, respectively). A meta-analysis of the compulsive symptoms GWAS with the PGC-OCD revealed no genome-wide significant Single-Nucleotide Polymorphisms (SNPs combined N = 17,992, indicating that the power is still low for individual SNP effects). A genebased association analysis, however, yielded two novel genes (WDR7 and ADCK1).The top 250 genes in the gene-based test also showed a significant increase in enrichment for psychiatric and brain-expressed genes. S-Predixcan testing showed that for genes expressed in hippocampus, amygdala, and caudate nucleus significance increased in the meta-analysis with compulsive symptoms compared to the original PGC-OCD GWAS. Thus, the inclusion of dimensional symptom data in genome-wide association on clinical case-control GWAS of OCD may be useful to find genes for OCD if the data are based on quantitative indices of compulsive behavior. SNP-level power increases were limited, but aggregate, gene-level analyses showed increased enrichment for brain-expressed genes related to psychiatric disorders, and increased association with gene expression in brain tissues with known emotional, reward processing, memory, and fear-formation functions.

show abstract

The accuracy of LD Score regression as an estimator of confounding and genetic correlations in genome‐wide association studies

Cited by 58 publications

References 47 publications

Summary statistic analyses do not correct confounding bias

Summary statistic analyses do not correct confounding bias

Expectation of the intercept from bivariate LD score regression in the presence of population stratification

Genetic meta‐analysis of obsessive–compulsive disorder and self‐report compulsive symptoms

Contact Info

Product

Resources

About