Tutorial 4

Economics and Genetics (FEB13089)

Erasmus School of Economics – Bachelor (Block 2)

Learning goals
• Being able to perform power calculations in R.
• Being able to run a Genome-wide Association study in PLINK.
• Being able to visualize Genome-wide Association study results in R.

Power calculations in R
As discussed in Lecture 4, it is essential to well-powered from a statistical point of view to find
reliable and replicable associations between SNPs and phenotypes. We perform the power
calculation in R (see Tutorial 1 for how to get R running on your computer). Launch R by starting
RStudio. To perform power calculations, we make use of the R package “pwr”. This add-on package
needs to be installed first. Go to Tools → Install packages… Type pwr, select the package, and click
“install”. Alternatively, install the pwr package from the command line using the following
command:
install.packages("pwr")

To be able to work with add-on R packages, you should first load them using the “library” command:
library(pwr)

For our interest, the function “pwr.r.test” is most important. This function takes the following five
arguments as input:

n Number of observations (individuals)

r Linear correlation coefficient (in our case, squareroot of the explained variance R2)
sig.level Significance level (Type I error probability, α)
power Power of test (1 minus Type II error probability, π)
alternative a character string specifying the alternative hypothesis, must be one of "two.sided"
(default), "greater" or "less"

Exactly one of the input arguments “r”, “n”, “power” and “sig.level” must be passed as NULL, and
that parameter is then determined from the others. Through three examples, we practice how
power calculations help to determine the conditions under which a well-powered GWAS can be
conducted. Remember, in a GWAS the significance threshold α is set to 5x10-8, and 80% power is
usually considered to be adequate.

Example 1: How many observations are needed to find a SNP that explains 0.02% of
the phenotypic variance at the genome-wide significance level of 5x10-8 with 80%
power?
The “pwr.r.test” works with standardized regression coefficients as input, which is in our case the
square root of the variance explained. Therefore, we use r = sqrt(0.0002) as input parameter. We

1
perform a two-sided test, because here we are not interested in the sign of the association but in
the variance explained.

pwr.r.test(n = NULL, r = sqrt(0.0002), sig.level = 5e-8, power = 0.80, alt

ernative = "two.sided")

approximate correlation power calculation (arctangh transformation)

n = 197984.4
r = 0.01414214
sig.level = 5e-08
power = 0.8
alternative = two.sided

Hence, 197984 observations are needed in this setting.

Example 2: How much power do we have for a in a sample of 100,000 individuals for a
SNP that explains 0.02% of the phenotypic variance, if we impose a significance level
of 5x10-8?
pwr.r.test(n = 100000, r = sqrt(0.0002), sig.level = 5e-8, power = NULL, a
lternative = "two.sided")

approximate correlation power calculation (arctangh transformation)

n = 1e+05
r = 0.01414214
sig.level = 5e-08
power = 0.1637795
alternative = two.sided

The results show that in this setting, we only have 16% power to detect the SNP. You’ll see that large
samples are needed in a GWAS.
Example 3: What is the minimum variance explained you could detect in a sample of
100,000 individuals at the genome-wide significance level of 5x10-8 with 80% power?
pwr.r.test(n = 100000, r = NULL, sig.level = 5e-8, power = 0.80, alternati
ve = "two.sided")

approximate correlation power calculation (arctangh transformation)

n = 1e+05
r = 0.01989582
sig.level = 5e-08
power = 0.8
alternative = two.sided

As output we get a standardized regression coefficient of 0.01989582. We have to square this

number, and multiply the result with 100% to obtain what we want:
0.01989582*0.01989582*100
[1] 0.03958437

Hence, the minimum variance explained we can detect with 80% power in this setting is 0.04%.
Thus, SNPs with a smaller effect are unlikely to be detected in this setting.

2
Running a GWAS in PLINK
Now that we can determine the conditions for a sufficiently powered GWAS, let’s practice with
running a GWAS. For this purpose, we switch to PLINK (see Tutorial 2 for how to run PLINK from the
command line). In this tutorial, we make again use of the same example data as in Tutorial 2 and
Tutorial 3. The files Example.bed, Example.bim, and Example.fam are probably still in your working
directory (“C:\Users\Dilnoza\Documents\EUR\E&G\Tutorials”) If not, please put these files there
again. These data files include information about 23,825 SNPS for 2,000 individuals (have a look at
the BIM and FAM file to check this). Also, make sure that the file “Example_height.pheno” (Tutorial
2) is available in this folder.

In Tutorial 2, we used the --assoc command to perform a basic association analysis. The --linear flag
(for continuously distributed traits) comes with more flexibility. Let’s compare the two output files;
let’s run the --assoc command first:
plink --bfile Example --assoc --pheno Example_height.pheno --out
GWAS_height_assoc

The output file “GWAS_height_assoc.qassoc” looks as follows (we already discussed the contents of
this file in Tutorial 2):

Let’s now run the GWAS with the --linear command:

plink --bfile Example --linear --pheno Example_height.pheno --out
GWAS_height_linear

The resulting file “GWAS_height_linear.assoc.linear” looks as follows:

The columns in this file represent: Chromosome, SNP identifier, Basepair position, Reference allele,
Type of model (here: additive), Number of non-missing individuals for this analysis, Regression
coefficient, Standard error of the coefficient, t-statistic for regression of phenotype on allele count
(number of reference alleles; 0, 1, or 2), Asymptotic significance value for coefficient. You can see

3
that --assoc and --linear give precisely the same results, and that many columns in the output files
are identical.

Principal components
As discussed in Lecture 4, an important advantage of GWAS analysis is the possibility to control for
population stratification with principal components. Let’s construct these principal components first,
and then add them to the GWAS model as control variables. We can construct the principal
components with the --pca command. Let’s construct the first four principal components (because
for many samples, analysts use four principal components to control subtle population
stratification):
plink --bfile Example --pca 4 --out Example_pcs

The command above creates two files in your working directory: “Example_pcs.eigenval” and
“Example_pcs.eigenvec”. The first file contains the eigenvalues of the genetic relationship matrix.
The second file contains the first four principal components:

Although a bit difficult to see, there are six columns in this file (separated by spaces): The family ID,
the individual ID, and the first four principal components.

Controlling for principal components in a GWAS

In Tutorial 3, we excluded cryptically related individuals from the GREML analysis. In a GWAS, we use
principal components to control for cryptic genetic relatedness. A nice feature of the --linear
command is the possibility to include covariates in the model with the --covar option. Let’s use the
constructed principal components as control variables in our GWAS:
plink --bfile Example --linear --pheno Example_height.pheno --out
GWAS_height_linear_pcs --covar Example_pcs.eigenvec

4
For every SNP, PLINK reports the regression results for the control variables on a separate row. In
practice, this is quite inconvenient, because you are often only interested in the regression results
for the SNPs. If you add --hide-covar, only the regression results for the SNPs are included in the
output file:
plink --bfile Example --linear --pheno Example_height.pheno --out
GWAS_height_linear_pcs2 --covar Example_pcs.eigenvec --hide-covar

• Exercise: Check that the SNP with the lowest p-value is rs1953405.

Visualizing GWAS results in R

We are now going to visualize the GWAS results using QQ-plots and Manhattan plots. For this
purpose, we switch back to R. Launch R by starting RStudio. First, we need to read in our GWAS
results in R. Therefore, we first change R’s working directory to our folder with data and results:
setwd("C:/Users/Dilnoza/Documents/EUR/E&G/Tutorials")

Subsequently, we read in the GWAS results (the model with PCs) using the “read.table” function. We
store the results in the object “results”.
results<-read.table("GWAS_height_linear_pcs2.assoc.linear",header=TRUE)

In the “Environment” you now see the object “results” under “Data”. By clicking on the object, you
can inspect it contents. Importantly, the p-value for each SNP is available in the column “P”.

We will create both a Manhattan plot and a QQ-plot to visualize the GWAS results. In both plots, it is
common work with logarithmically transformed p-values, -log10(p-value), rather than with the raw
p-values. Let’s do this transformation, and store the results in the column “observed_log_pvalues”:
results$observed_log_pvalues <- (-log10(results$P))

5
Manhattan plot
Based on the –log10(p-values) of the GWAS results, we can create a Manhattan plot. In a Manhattan
plot, you plot the (transformed) p-values according to their position in the genome. In the GWAS
results, the chromosome number and basepair position of the SNP is available. The numbering of
basepairs restarts every chromosome, so we first need to determine the plotting position of a SNP as
a number reflecting its chromosome and its BP position on the chromosome.

First, we need to determine per chromosome the minimum and maximum Base Pair (BP) position:
for(i in 1:22) {
results$MinBP[results$CHR==i]<-min(results$BP[results$CHR==i])
results$MaxBP[results$CHR==i]<-max(results$BP[results$CHR==i])
}

Within a chromosome, we make sure that the plotting position start at 1. Therefore, we need to
subtract the minimum BP in a chromosome from the SNP's BP position, and add 1 to it:
results$Plot_position<-results$BP - results$MinBP + 1

The only thing left to do now (to get a continuous plotting position), is to add the maximum plotting
position from a SNP on the preceding chromosome to the plotting position of a SNP within a
chromosome. For this purpose we could use a “loop”, but for simplicity I spell out the commands for
each chromosome here. We skip chromosome 1, because there is no preceding chromosome.
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==1])
results$Plot_position[results$CHR==2]<-results$Plot_position[results$CHR==2]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==2])
results$Plot_position[results$CHR==3]<-results$Plot_position[results$CHR==3]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==3])
results$Plot_position[results$CHR==4]<-results$Plot_position[results$CHR==4]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==4])
results$Plot_position[results$CHR==5]<-results$Plot_position[results$CHR==5]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==5])
results$Plot_position[results$CHR==6]<-results$Plot_position[results$CHR==6]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==6])
results$Plot_position[results$CHR==7]<-results$Plot_position[results$CHR==7]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==7])
results$Plot_position[results$CHR==8]<-results$Plot_position[results$CHR==8]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==8])
results$Plot_position[results$CHR==9]<-results$Plot_position[results$CHR==9]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==9])
results$Plot_position[results$CHR==10]<-results$Plot_position[results$CHR==10]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==10])
results$Plot_position[results$CHR==11]<-results$Plot_position[results$CHR==11]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==11])
results$Plot_position[results$CHR==12]<-results$Plot_position[results$CHR==12]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==12])
results$Plot_position[results$CHR==13]<-results$Plot_position[results$CHR==13]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==13])
results$Plot_position[results$CHR==14]<-results$Plot_position[results$CHR==14]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==14])
results$Plot_position[results$CHR==15]<-results$Plot_position[results$CHR==15]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==15])
results$Plot_position[results$CHR==16]<-results$Plot_position[results$CHR==16]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==16])
results$Plot_position[results$CHR==17]<-results$Plot_position[results$CHR==17]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==17])
results$Plot_position[results$CHR==18]<-results$Plot_position[results$CHR==18]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==18])
results$Plot_position[results$CHR==19]<-results$Plot_position[results$CHR==19]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==19])
results$Plot_position[results$CHR==20]<-results$Plot_position[results$CHR==20]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==20])
results$Plot_position[results$CHR==21]<-results$Plot_position[results$CHR==21]+MAX_position_preceding_chromosome
MAX_position_preceding_chromosome<-max(results$Plot_position[results$CHR==21])
results$Plot_position[results$CHR==22]<-results$Plot_position[results$CHR==22]+MAX_position_preceding_chromosome

With the transformed p-values and plotting positions calculated, we can create the Manhattan plot.1
In doing, we are going to give every chromosome a different color:

1
Sometimes, you get the error message "Error in plot.new() : figure margins too large". This error occurs when
your "Plots" panel is too small on your screen. You can easily solve this issue by making your "Plots" panel

6
plot(results$Plot_position,
results$observed_log_pvalues,col=results$CHR,xaxt="n",ylab=expression(paste(Observed~~-log[10],"(p-
value)")),main="Manhattan plot",ylim=c(0,8),xlab="Chromosomes")

In the resulting plot, all chromosomes have a different color. Let’s also put the chromosome numbers on the x-
axis.

for(i in 1:22) {
text(mean(results$Plot_position[results$CHR==i]),-0.50-0.30*((i %% 2) ==
0), adj=0.5,label=i,cex=0.7, xpd=TRUE)
}

The genome-wide significance level is 5×10-8 (Note that –log10(5×10-8)=7.3). Let’s add this line to the
Manhattan plot:
abline(h=-log10(5e-8),lwd="2",lty="dashed")

large, see e.g., https://support.rstudio.com/hc/en-us/articles/200488548-Problem-with-Plots-or-Graphics-

Device.

7
The resulting Manhattan plot nicely presents the results of your GWAS. Unfortunately, you can see
that there is no single SNP genome-wide significant in our analysis.

QQ-plot
As explained in Lecture 4, a Manhattan plot visualizes specific SNP associations. A QQ-plot is used to
inspect the distribution of p-values. Under the null hypothesis of no association, the p-values in a
GWAS are randomly distributed on the interval [0,1]. That is, they are uniformly distributed on the
interval [0,1]. Let’s “draw” values from this distribution for exactly as many SNPs we have in our
results (length(results$P)), and transform them to the logarithmic scale. For this purpose, we create
a series of numbers from 1 to the number of SNPs using “1:length(results$P)”. We divide these
numbers by length(results$P) to get to our distribution on the interval [0,1]:
null_log_pvalues <- -log10(1:length(results$P)/length(results$P))

To plot the observed and null distributions of p-values against each other, we first need to sort the
p-values:
observed_log_pvalues_sorted <- sort(results$observed_log_pvalues, decreasing=TRUE)

null_log_pvalues_sorted <- sort(null_log_pvalues, decreasing=TRUE)

For the sorted p-values under the null distribution, we also construct the 95% confidence intervals
because we want to add these to the QQ-plot. Fortunately, some researchers have derived that the
j-th order statistic from a uniform (0,1) sample follows a beta(j,n-j+1) distribution. Hence, we can use
the quantiles of this distribution to determine the upper (97.5% quantile) and lower bound (2.5%
quantile) of the 95% confidence interval:
c975 <- rep(0, length(results$P))
c025 <- rep(0, length(results$P))
for(i in 1: length(results$P)){
c975[i] <- qbeta(0.975,i, length(results$P)-i+1)
c025[i] <- qbeta(0.025,i, length(results$P)-i+1)
}

In a QQ-plot, you plot the expected distribution of p-values against the observed distribution of p-
values. We add some labels to the axes and figure itself as well. Moreover, we make sure to have a
squared plotting area.
plot(null_log_pvalues_sorted, observed_log_pvalues_sorted,xlab=expression(paste(Expected~~-
log[10],"(p-value)")),ylab=expression(paste(Observed~~-log[10],"(p-value)")),main="QQ-
plot",xlim=c(0,max(observed_log_pvalues_sorted)),ylim=c(0,max(observed_log_pvalues_sorted)))

8
The resulting plot looks like:

With the following lines of code, we add a diagonal line to the plot. The diagonal represent the null
hypothesis of no association. The command “par(new=T)” makes sure that the diagonal is added to
the earlier created QQ-plot.
par(new=T)
abline(0,1,lwd="2",lty="dashed")

Your plot should like look this now:

In the plot, you can see that there is some deviation from the diagonal. However, the deviation is
small.

Finally, we add the confidence intervals to the QQ-plot:

polygon(x=c(null_log_pvalues_sorted,rev(null_log_pvalues_sorted)),y=c(-
log(c975,10),rev(null_log_pvalues_sorted)),col= rgb(0.5, 0.5, 0.5, 0.5),border=NA)

polygon(x=c(null_log_pvalues_sorted,rev(null_log_pvalues_sorted)),y=c(-
log(c025,10),rev(null_log_pvalues_sorted)),col= rgb(0.5, 0.5, 0.5, 0.5),border=NA)

Your final result should look like:

9
We see that all p-values are more or less within the 95% confidence interval (a few SNPs outside the
95% confidence interval is not surprising given the large number of SNPs we analyze).2 Moreover,
there is no SNP with a p-value below the genome-wide significance level of 5×10-8 (i.e., above ~7.3
on the y-axis). Hence, we conclude that the distribution of p-values does not provide evidence for
association between SNPs and height in our data.

Individual assignment3
• Use power calculations to assess under what statistical conditions (e.g., the expected R2 of
SNP) a GWAS on personal income is feasible. Use reasonable assumptions (e.g., based on
earlier genetic associations reported in scientific papers) in your calculations, and use at
least 1 scientifically formatted table to report the results.
• Run two times a GWAS on personal income, one for females and one for males, using the
data from Tutorial 3. Use at least 2 scientifically formatted figures (QQ-plots and Manhattan
plots) to report the results. Given your power calculations, discuss whether you are
surprised by the GWAS results.

2
In rare cases, the part of the confidence interval above the diagonal is not correctly plotted. This happens as a
result of a bug in some older R versions. If this happens to you, please update your R version or create the QQ-
plots using RStudio Cloud (rstudio.cloud).
3
Don’t forget to include your analysis code in the Appendix of your report.

10