The Weighting is the Hardest Part: On the Behavior of the Likelihood Ratio Test and the Score Test Under a Data-Driven Weighting Scheme in Sequenced Samples

Abstract

Sequence-based association studies are at a critical inflexion point with the increasing availability of exome-sequencing data. A popular test of association is the sequence kernel association test (SKAT). Weights are embedded within SKAT to reflect the hypothesized contribution of the variants to the trait variance. Because the true weights are generally unknown, and so are subject to misspecification, we examined the efficiency of a data-driven weighting scheme. We propose the use of a set of theoretically defensible weighting schemes, of which, we assume, the one that gives the largest test statistic is likely to capture best the allele frequency-functional effect relationship. We show that the use of alternative weights obviates the need to impose arbitrary frequency thresholds in sequence data association analyses. As both the score test and the likelihood ratio test (LRT) may be used in this context, and may differ in power, we characterize the behavior of both tests. We found that the two tests have equal power if the set of weights resembled the correct ones. However, if the weights are badly specified, the LRT shows superior power (due to its robustness to misspecification). With this data-driven weighting procedure the LRT detected significant signal in genes located in regions already confirmed as associated with schizophrenia – the PRRC2A (P=1.020E-06) and the VARS2 (P=2.383E-06) – in the Swedish schizophrenia case-control cohort of 11,040 individuals with exome-sequencing data. The score test is currently preferred for its computational efficiency and power. Indeed, assuming correct specification, in some circumstances the score test is the most powerful. However, LRT has the advantageous properties of being generally more robust and more powerful under weight misspecification. This is an important result given that, arguably, misspecified models are likely to be the rule rather than the exception in weighting-based approaches.

Epigenetic signatures relating to disease-associated genotypic burden in familial risk of bipolar disorder

“…VARS2 encodes a mitochondrial aminoacyl-tRNA synthetase involved in mitochondrial protein synthesis. Mitochondrial abnormalities are evident in BD [ 76 ] and other psychiatric illnesses [ 77 ], and loss-of-function mutations in VARS2 have been previously associated with mitochondrial encephalopathies [ 78 ], epilepsy [ 79 ], and schizophrenia [ 80 ]. Furthermore, an EWAS of depressive symptoms in 724 monozygotic Danish twins identified a differentially methylated region in a putative active enhancer of VARS2 , which spanned 9-probes including cg00933603 [ 81 ].…”

Recent advances and challenges of rare variant association analysis in the biobank sequencing era

“…Many such annotations may also be cell-type specific, requiring consideration for the phenotype under study. Appropriate consideration for variant filtering and weighting may substantially improve statistical power for RV association discovery ( Byrnes et al, 2013 ); conversely, misspecification of variant weights could lead to loss of power by inadvertently removing and/or down-weighting key disease-related functional RVs ( Minica et al, 2017 ).…”

Convex combination sequence kernel association test for rare‐variant studies

“…Statistical power of variant set tests can be improved by weighting SNVs by their hypothesized effect on the trait or disease risk as a function of available annotation, for example, a function of MAF giving more weight to rarer SNVs. Fixed SNV weights, however, may misspecify the contribution of the variants and lower the power of the variant set test (Minica et al, 2017). To reduce weight misspecification, adaptive tests have been proposed that compute many different statistics and select the test with the smallest p value, such as the multi‐kernel sequence kernel association test (MK‐SKAT; Urrutia et al, 2016; Wu et al, 2013) or the omnibus test statistic (OMNI; Barnett et al, 2017).…”