A Cautionary Note on the Effects of Population Stratification Under an Extreme Phenotype Sampling Design

Abstract

Extreme phenotype sampling (EPS) is a popular study design used to reduce genotyping or sequencing costs. Assuming continuous phenotype data are available on a large cohort, EPS involves genotyping or sequencing only those individuals with extreme phenotypic values. Although this design has been shown to have high power to detect genetic effects even at smaller sample sizes, little attention has been paid to the effects of confounding variables, and in particular population stratification. Using extensive simulations, we demonstrate that the false positive rate under the EPS design is greatly inflated relative to a random sample of equal size or a “case-control”-like design where the cases are from one phenotypic extreme and the controls randomly sampled. The inflated false positive rate is observed even with allele frequency and phenotype mean differences taken from European population data. We show that the effects of confounding are not reduced by increasing the sample size. We also show that including the top principal components in a logistic regression model is sufficient for controlling the type 1 error rate using data simulated with a population genetics model and using 1,000 Genomes genotype data. Our results suggest that when an EPS study is conducted, it is crucial to adjust for all confounding variables. For genetic association studies this requires genotyping a sufficient number of markers to allow for ancestry estimation. Unfortunately, this could increase the costs of a study if sequencing or genotyping was only planned for candidate genes or pathways; the available genetic data would not be suitable for ancestry correction as many of the variants could have a true association with the trait.

Comparison of mixed model based approaches for correcting for population substructure with application to extreme phenotype sampling.

“…For both approaches, the estimated type 1 error for all the 295 sample sizes ranges between 0.042 − 0.049, which is only slightly lower than the nominal level of 0.05. The type 1 error rate for the PCA-approach is also close to the nominal value, though possibly slightly elevated; similar results for the PCA approach were found in [28]. However, CARAT shows higher type 1 error rates than the nominal level of 0.05.…”

“…The total cohort size, N , was set to 5000, 10, 000 or 20, 000. The F st value between the two populations -a measure of genetic population differentiation -was set to 0.01; this value is higher than would be expected between typical European populations but it ensures substantial substructure [28]. Genetic data was simulated using the Balding-Nichols method [37,14] as previously described [28].…”

“…For each individual, we simulated a total of p = 5000 SNPs. Though true genome-wide data would consist of much larger numbers of SNPs, our previous work with data simulated using this model has shown that this number of SNPs is sufficient to correct for population stratification [28]. For each SNP, the generating allele frequency, p, was sampled from a uniform [0.1, 0.9] distribution.…”

“…Therefore, for EPS designs it is very important to include correction for population stratification. We have shown that including the top principal components in a logistic regression model adequately limits the type 1 error rate when the candi-65 date variant was common; however, there was a slight inflation when the candidate variant was rare [28]. The mixed model-based approaches for correcting for population substructure were developed assuming binary traits from case-control type studies.…”

Effect of population stratification on SNP‐by‐environment interaction

“…For Figure , we assumed that and were assumed to be 0.1 and 0.3, respectively. We selected the values for and , as they seem to be real choices based on our experience in applications and similar parameter choices in other studies (Panarella & Burkett, ). Different values for both can be utilized but the similar patterns are expected as long as is same.…”