• We are sorry, but NCBI web applications do not support your browser and may not function properly. More information
Logo of ajhgLink to Publisher's site
Am J Hum Genet. Feb 8, 2008; 82(2): 524–526.
Published online Feb 1, 2008. doi:  10.1016/j.ajhg.2007.10.014
PMCID: PMC2427240

Comment on a Simple and Improved Correction for Population Stratification

Main Text

To the Editor: In the May 2007 issue of the American Journal of Human Genetics, Epstein, Allen, and Satten1 (hereafter referred to as EAS) introduced a new method for controlling population stratification in case-control association studies. The method computes a stratification score by performing partial least-squares regression (PLS) of phenotypes (case-control status) on a matrix of genotypes at markers used to correct for ancestry. The quantitative stratification score is then used to divide subjects into a number of strata, so that a stratified test of case-control association may be performed at any test locus not in linkage disequilibrium with the ancestry-informative markers. The stratification and testing procedure are implemented in the program StratScore, available as SAS code from the authors.

EAS described a retrospective case-control model involving the latent true stratification variable and provided practical recommendations for dividing the estimated stratification score into a number of strata. The PLS procedure, however, was presented in less detail, although it is key to the performance of the overall approach. A primary motivation was the claim that stratified analysis based on principal components2 or genomic control3 cannot fully control for population ancestry. The authors cited an example and provided simulations in which stratification resulted in inflated type I errors when using these methods for 100 ancestry-informative markers. An immediate concern is whether these results reflect current practice—in a modern whole-genome scan, hundreds of thousands of markers are available for ancestry control. The results of Price et al.2 suggest that, with the availability of thousands of markers, principal components do provide effective ancestry control, and indeed a large number of markers may be necessary for correcting stratification within continental-level populations.5 Moreover, the use of principal components does not require predefined ancestry-informative markers and thus may potentially control for unanticipated strata, including technical phenomena unrelated to ancestry.2 In terms of statistical power, the principal-components-based approach appeared to fare quite well in EAS.1

To better understand the issues and how the EAS approach might be best applied, we examined the PLS procedure more closely. Here, PLS finds linear combinations T of the matrix of ancestry-informative markers X such that the covariance between phenotypes Y and T is maximized (see 4 for details on partial least-squares regression). Predictions of case status from a logistic-regression model (Y on T) are then used as the stratification score. A risk of PLS is the potential for finding spurious relationships, although EAS employed a variable selection technique to control the number of T variables used. If spurious apparent stratification arises from PLS, it has the potential to greatly reduce statistical power because the stratification variable could account for phenotype variation caused by a true disease gene. Moreover, although the inclusion of a large number of ancestry-informative markers should be desirable for ancestry prediction, the resulting increased flexibility in the PLS factors might produce even stronger spurious stratification, thereby resulting in decreased power as the number of such markers increases.

To further investigate the utility of StratScore and to test our predictions about the method, we performed simulations under no stratification, for random unlinked markers with minor allele frequencies (MAF) ranging uniformly from 0.1 to 0.5. Table 1 shows the results from representative simulations analyzed by StratScore, with m = 100, 200, 500, and 800 markers used to infer ancestry. Note that the stratification score has a very high correlation with case-control status, although no true correlation exists between the markers and phenotype because no stratification exists. As the number of markers increases, the spurious correlation increases, and the case-control numbers for many of the strata become highly imbalanced. Such strata cannot meaningfully contribute to detection of case-control association.

Table 1
Illustrative Simulations of Case-Control Status versus StratScore Inferred Strata

We further performed simulations of case-control association data, by following the conditions and terminology described in EAS. For each setup, 5000 simulations were performed for 500 cases and 500 controls, with three underlying populations of equal size. We simulated substantial stratification by sampling cases in the proportions 0.45, 0.33, and 0.22 from subpopulations 1, 2, and 3. Moderate stratification was achieved by sampling in the proportions 0.40, 0.33, and 0.27. The alternative hypothesis was simulated with odds of disease increasing by a factor 1.4 for each copy of the risk allele for the test locus, which had Fst values of 0.03 and 0.15 in various simulation setups. EAS simulated ancestry markers on the basis of Fst selection criteria applied to SNPs from a real data set. To reproduce their results and to better control the simulation conditions, we simulated marker SNPs following the method in Price et al.1 For each of MAF values 0.1, 0.25, and 0.4, sets of random marker SNPs were simulated with Fst = 0.03, and highly ancestry-informative markers with Fst values were drawn uniformly from 0.5 to 0.8. Although EAS reported results for sets of m = 100 ancestry markers,1 we also performed simulations for sets of m = 200, 500, and 800 markers.

With a significance threshold of α = 0.05 and a test locus with Fst = 0.03, we found approximately correct type I error control by using the StratScore approach for all choices of m markers (Table 2, effectively an expanded version of Tables 2 and 3 in EAS). However, when the test locus had Fst = 0.15, we found type I errors ranging from 0.02 to 0.098 (Table 3), depending on the ancestry marker setup and degree of stratification. EAS had reported correct StratScore error control for some of these same setups (see Table 4 in EAS). We are unsure of the reason for the discrepancy, although minor variation in generalized PLS1 versus the standard PLS implemented in StratScore is a possibility. To investigate whether the results might be specific to our use of the simulation approach of Price et al.1 (beta sampling of minor allele frequencies, followed by rejection sampling of Fst values), we also employed a deterministic approach. We set allele frequencies for the three populations (order determined randomly) as p/a, p, and pa, where a and p were determined to achieve specified Fst and MAF values. Our conclusions under this scheme were unchanged. Although our main focus is on the power of StratScore, these results suggest a lack of robustness that may be problematic in StratScore error control and deserves further inquiry.

Table 2
Type I Error under Substantial and Moderate Stratification
Table 3
Type I Error, Test Locus Fst = 0.15
Table 4
Power under Substantial and Moderate Stratification

We next investigated power for StratScore as the number of markers increases. Table 4 presents the power under the alternative hypothesis for Cochran Mantel Haenszel (CMH) tests under moderate and substantial true stratification. Here, the best-case scenario of known strata is compared to the StratScore approach for various numbers of ancestry markers. As predicted, the power drops dramatically as the number of ancestry markers increases, thereby restricting the number of markers that can be used. Note that this restriction depends in an essential way on the case-control sample size. Studies in which the true stratification is subtle may require a larger number of markers for ancestry control and therefore limit the utility of StratScore.

Another aspect of EAS that was unclear was the degree of correspondence between the stratification score and the true subpopulations. For the alternative-hypothesis simulation setups, we computed average ANOVA R2 values for the stratification score versus the three true subpopulations. For m = 100 markers and substantial stratification, R2 was ~0.19 when highly ancestry-informative markers were used, regardless of MAF, and 0.12 for random markers with Fst = 0.03. Under moderate stratification, the R2 values were 0.07 for highly ancestry-informative markers, and 0.04 for random markers. As m increased, the R2 values dropped even further. These relatively low values were apparently enough to provide error-control correction for the simulations reported in EAS, and other measures of correspondence than R2 might be preferred. Nonetheless, these results further call into question the robustness of the PLS procedure, in which the stratification score does not strongly reflect the true stratification.

In summary, we conclude that aspects of the EAS method may be worthy of further exploration and development. However, in its present form, we have concerns about the routine use of StratScore, especially in the context of genome-wide scans. At the very least, the genomics community should be aware of the potential for power loss and sensitivity to the number of ancestry-informative markers employed. Additional, larger simulations in the context of whole-genome scans are necessary to provide convincing comparisons of the major approaches for controlling spurious association in case-control association studies.


The authors are supported in part by NIH grant R01 GM074175 and EPA RD-83272001. We thank the editors and reviewers for their comments.


1. Epstein M.P., Allen A.S., Satten G.A. A simple and improved correction for population stratification in case-control studies. Am. J. Hum. Genet. 2007;80:921–930. [PMC free article] [PubMed]
2. Price A.L., Patterson N.J., Plenge R.M., Weinblatt M.E., Shadick N.A., Reich D. Principal components analysis corrects for stratification in genome-wide association studies. Nat. Genet. 2006;38:904–909. [PubMed]
3. Devlin B., Roeder K. Genomic control for association studies. Biometrics. 1999;55:997–1004. [PubMed]
4. Abdi H. Partial least squares regression (PLS-regression) In: Lewis-Beck M., Bryman A., Futing T., editors. Encyclopedia for Research Methods for the Social Sciences. Sage; Thousand Oaks, CA: 2003. pp. 792–795.
5. Freedman M.L., Reich D., Penney K.L., McDonald G.J., Mignault A.A., Patterson N., Gabriel S.B., Topol E.J., Smoller J.W., Pato C.N. Assessing the impact of population stratification on genetic association studies. Nat. Genet. 2004;36:388–393. [PubMed]

Articles from American Journal of Human Genetics are provided here courtesy of American Society of Human Genetics
PubReader format: click here to try


Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...


  • PubMed
    PubMed citations for these articles

Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...