Methods for use with discovery GWAS

This vignette demonstrates how our R package, winnerscurse, can be used to obtain new adjusted association estimates for each SNP when only the summary statistics from a single genetic association study, namely the discovery GWAS, are available. In order to do so, we create a toy data set using the function sim_stats, which is included in the package, and subsequently, illustrate how a user could engage with each of the package’s Winner’s Curse correction functions using this data set.

The methods currently available for implementation are:

  1. Conditional Likelihood methods - Ghosh et al. (2008)
  2. Empirical Bayes method - Ferguson et al. (2013)
  3. FDR Inverse Quantile Transformation method - Bigdeli et al. (2016)
  4. Bootstrap method - Forde et al. (2023)

An important feature which distinguishes these four methods is detailed in the table below:

Adjusts \(\beta\) estimate of only SNPs with \(p\)-values less than specified threshold, \(\alpha\) Adjusts \(\beta\) estimate of all SNPs
  • Conditional Likelihood methods
  • Empirical Bayes method
  • FDR Inverse Quantile Transformation method
  • Bootstrap method

Creating a toy data set

  • The function sim_stats provides users with an opportunity to construct various simulated data sets of GWAS summary statistics which the correction methods can then be applied to. This allows users to perhaps gain a better understanding of each method before applying them to their own real data sets.

  • sim_stats constructs a large set of independent SNPs, in which there exists a polygenic background of a specified portion of these SNPs for the trait of interest while all other SNPs demonstrate no effect. When considering obtaining a simulated set of summary statistics for solely the discovery GWAS, sim_stats requires the user to specify four arguments, namely nsnp, h2, prop_effect and nid:

    1. nsnp: the total number of SNPs
    2. h2: the heritability of the trait, i.e. the variance explained in the trait by all SNPs
    3. prop_effect: the proportion of the total number of SNPs truly associated with the trait
    4. nid: the number of samples in the study

  • With the above specifications, a set of rather realistic summary statistics can be obtained as shown below. For convenience purposes, the following default values have been supplied for the function’s arguments: nsnp=10^6, h2=0.4, prop_effect=0.01, nid=50000. These values simulate summary statistics for 1,000,000 independent SNPs in which 1% have a true effect on this trait with moderate heritability and a total of 50,000 individuals have been sampled in the study.

  • Note: For further information regarding simulating GWAS summary statistics, refer to the text supplement of Forde et al. (2023). This function has been designed based on the details described under the heading ‘Quantitative trait with independence assumed’. However, note that for step 3) of these details, \(\beta_i\) is in fact sampled from the distribution \(\beta_i \sim N(0,1)\) instead of \(\beta_i \sim N(0, 2 \text{maf}_i(1-\text{maf}_i)\).

  • sim_stats outputs a list with three components. When only the discovery GWAS is considered, the third element in this list is merely NULL, and so we are just concerned with the first two, true and disc. true is a data frame with two columns, rsid which contains identification numbers for each SNP and true_beta which is each SNP’s simulated true association value. This data frame is most useful if we wish to check how well our Winner’s Curse correction methods are performing. disc has three columns representing the summary statistics one would obtain in a discovery GWAS. For each SNP, this data frame contains its ID number, its estimated effect size and associated standard error. We note that this function outputs the set of summary statistics, disc, in a suitable way so that the Winner’s Curse correction methods can be directly applied, i.e. in the form of a data frame with three columns rsid, beta and se.

set.seed(1998)

sim_dataset <- sim_stats(nsnp=10^6,h2=0.4,prop_effect=0.01,nid=50000)

## simulated GWAS summary statistics
summary_stats <- sim_dataset$disc
head(summary_stats)
#>   rsid         beta          se
#> 1    1 -0.011423267 0.019222517
#> 2    2  0.002883483 0.009187098
#> 3    3 -0.043813790 0.016459133
#> 4    4 -0.018164924 0.019560463
#> 5    5  0.021349091 0.006335237
#> 6    6 -0.001559945 0.007697227

Method 1: Conditional Likelihood

  • The function conditional_likelihood requires a data frame, in the form described above, and a value for alpha, the significance threshold used in the GWAS. Note that all columns of the data frame must contain numerical values and each row must represent a unique SNP, identified by rsid.

  • As the conditional likelihood methods have been designed to be only used for SNPs which are deemed significant, the data frame returned contains only SNPs that have \(p\)-values below the inputted genome-wide significance threshold value, alpha.

  • Note: If the \(z\)-statistic of a particular SNP is greater than 100, then merely the original naive estimate will be outputted for the second form of the adjusted estimate, namely beta.cl2, for that SNP.

  • If no SNPs are detected as significant, the function returns a warning message: WARNING: There are no significant SNPs at this threshold.

  • The returned data frame has SNPs ordered based on their \(p\)-values from smallest to largest, or equivalently, in descending order of \(\mid z \mid\) in which \(z\) is the estimated value for \(\beta\) divided by its standard error.

  • We implement the function on our toy data set as follows, choosing a significance threshold of 5e-8:

out_CL <- conditional_likelihood(summary_data = summary_stats, alpha = 5e-8) 
head(out_CL)
#>   rsid        beta          se    beta.cl1    beta.cl2    beta.cl3
#> 1 7815  0.04771727 0.006568299  0.04709173  0.04575084  0.04642129
#> 2 3965  0.04609999 0.006346232  0.04549476  0.04419814  0.04484645
#> 3 4998 -0.04527792 0.006458823 -0.04421716 -0.04238468 -0.04330092
#> 4 9917  0.04164616 0.006334188  0.03909824  0.03598923  0.03754373
#> 5 7261  0.04162686 0.006343244  0.03900942  0.03584977  0.03742959
#> 6 6510  0.04254046 0.006500057  0.03975867  0.03645266  0.03810567

  • Three columns have been added to the right of the inputted data frame: each represents a correction method suggested in Ghosh et al. (2008). The first, beta.cl1, is referred to as the conditional MLE while the second, beta.cl2, is the mean of the normalized conditional likelihood. The third, beta.cl3 is merely the average of the first two, known as the compromise estimator.

  • Ghosh et al. (2008) noted that for instances in which the true \(\beta\) value is close to zero, beta.cl2 often has greater mean squared error than beta.cl1 but for true \(\beta\) values further from zero, beta.cl2 performs better. Thus, it was suggested that beta.cl3 be used to combine the strengths of these two estimators. However, this function, conditional_likelihood, outputs values for all three estimators in order to allow the user to make their own decision on which they believe to be the most appropriate.

Method 2: Empirical Bayes

  • The function empirical_bayes implements the empirical Bayes method for correcting for Winner’s Curse, detailed in Ferguson et al. (2013), with a slight modification.

  • The function has two arguments with the first being summary_data, the data frame containing rsid, beta and se columns. Again, all columns of the data frame must contain numerical values and each row must represent a unique SNP, identified by rsid. In addition, the function requires that the data frame must contain at least 50 rows corresponding to 50 unique SNPs. This requirement is included as the empirical Bayes method performs poorly when only data of a small number of SNPs is available. The second argument, method, is a string parameter which allows the user to choose what modelling approach to take for the purpose of estimating the log density function. The default setting is method="AIC", which is a method very similar to that described by Ferguson et al. (2013). Other options include method="fix_df", method="scam", method="gam_nb" and method="gam_po". If method="fix_df", the degrees of freedom is set to 7. The other three options all enforce additional constraints on the shape of the estimated log density function by i) fitting a generalized additive model (GAM), with integrated smoothness estimation, (method="gam_nb" and method="gam_po") and ii) applying two shape constrained additive models (SCAMs) (method="scam"). These additional variations of the empirical Bayes approach are detailed in the ‘Materials and methods’ section of Forde et al. (2023).

  • empirical_Bayes returns the above described data frame but with a fourth column, beta_EB in which the adjusted estimates of this procedure have been added.

  • As the empirical Bayes method makes adjustments to all SNPs, not only those that have been considered significant, this data frame contains all SNPs.

  • Note: The conditional likelihood methods adjust each statistic one at a time, and so summary statistics of one individual SNP can easily be inputted into the conditional_likelihood function. However, the empirical Bayes method requires many SNPs as the more information given to the function, the more accurate it will be at modelling the distribution and thus, making adjustments. Thus, as mentioned above, summary_data must contain information related to more than 50 SNPs.

  • Below is a demonstration of using empirical_bayes with our toy data set with the default setting for method:

out_EB <- empirical_bayes(summary_data = summary_stats)
head(out_EB)
#>   rsid        beta          se     beta_EB
#> 1 7815  0.04771727 0.006568299  0.04599530
#> 2 3965  0.04609999 0.006346232  0.04443623
#> 3 4998 -0.04527792 0.006458823 -0.03999539
#> 4 9917  0.04164616 0.006334188  0.03998556
#> 5 7261  0.04162686 0.006343244  0.03996389
#> 6 6510  0.04254046 0.006500057  0.04083638

  • Unfortunately, the Empirical Bayes estimator originally described by Ferguson et al. (2013) is known to perform poorly in the extreme tails of the distribution. Therefore, it is quite possible that a lack of adjustment for the \(\beta\) values of SNPs with the most extreme \(z\)-statistics could be witnessed with this method. A somewhat ad hoc approach to overcome this issue of combining the Empirical Bayes and the conditional likelihood methods was suggested by Ferguson et al. (2013).

  • In order to ensure that shrinkage does occur for the \(\beta\) values of these extreme SNPs, it was decided that the following modification would be added to the empirical_bayes function with method="AIC". The basis function of the natural cubic spline has been altered so that the boundary knots are no longer the most extreme \(z\)-values. Instead, the lower boundary knot is defined as the \(10^{\text{th}}\) \(z\)-statistic when the \(z\)-statistics lie in increasing order while the upper boundary knot is the \(10^{\text{th}}\) \(z\)-statistic when the \(z\)-statistics have been arranged in decreasing order. The natural cubic spline is then merely linear beyond these boundary knots. This assists in fixing the ‘tails’ issue discussed above. We see that the reduction of the estimated \(\beta\) values for all of the top 6 most significant SNPs has occurred.

Method 3: FDR Inverse Quantile Transformation

  • The function FDR_IQT implements the Winner’s Curse adjustment method detailed in Bigdeli et al. (2016).

  • Similar to the approaches above, the function requires a data frame with three columns rsid, beta and se, all columns of the data frame must contain numerical values and each row must represent a unique SNP, identified by rsid.

  • It also has an argument min_pval, for which the default setting is 1e-300. This is included in order to avoid zero \(p\)-values which can prove problematic for the function when evaluating qnorm(). Furthermore, due to the nature of Winner’s Curse, it is in fact rare that observations with \(\mid z \mid > 37\) are biased.

  • The function outputs a data frame similar to that inputted with an additional column containing the adjusted estimate, beta_FIQT, and again, the SNPs have been reordered according to their individual \(z\)-statistics.

out_FIQT <- FDR_IQT(summary_data = summary_stats)
head(out_FIQT)
#>   rsid        beta          se   beta_FIQT
#> 1 7815  0.04771727 0.006568299  0.03422825
#> 2 3965  0.04609999 0.006346232  0.03307103
#> 3 4998 -0.04527792 0.006458823 -0.03188772
#> 4 9917  0.04164616 0.006334188  0.02823824
#> 5 7261  0.04162686 0.006343244  0.02827861
#> 6 6510  0.04254046 0.006500057  0.02897770

Method 4: Bootstrap

  • Inspired by the bootstrap resampling method first detailed in Faye et al. (2011), the function BR_ss implements a similar approach which can be easily applied to published sets of GWAS summary statistics without the requirement of access to original individual-level data. The major advantage of our alternative suggested method is the considerable improvement in computational efficiency, with a reasonably low level of conceptual difficulty being maintained. A comprehensive description of this new bootstrap method can be found in the ‘Materials and methods’ section of Forde et al. (2023).

  • BR_ss takes a data frame summary_data with three columns: rsid, beta, se, and returns the inputted data frame along with the adjusted estimate for \(\beta\), beta_BR_ss. Note that all columns of the data frame, summary data must contain numerical values, each row must represent a unique SNP, identified by rsid and the function requires that there must be at least 5 unique SNPs so that the smoothing spline can be implemented successfully. The function also has a logical parameter seed_opt which can be set to TRUE by the user to permit reproducibility. Its default setting is seed_opt=FALSE. As this method involves bootstrap implementation, the use of a seed can allow the user to obtain the exact same results for beta_BR_ss when running the function many times on the same dataset. The parameter seed provides the user with an option to set their desired value for the seed when using seed_opt=TRUE. The default for seed is the arbitrary value of 1998.

\(~\)

Brief method description:

  • This procedure begins with arranging all \(N\) SNPs according to their original \(z\)-statistics, \(z_i = \frac{\hat\beta_i}{\hat{\text{se}(\hat\beta_i)}}\), in descending order. A randomized estimate of the extent of ranking bias for the \(k^{th}\) largest \(z\)-statistic is calculated by means of the parametric bootstrap as follows:

    1. A value \(\hat\beta_i^{\text{b}}\) is simulated for SNP \(i, i=1,...,N\), independently, from a Gaussian distribution with mean \(\hat\beta_i\) and standard deviation \(\hat{\text{se}(\hat\beta_i)}\), i.e. \[\hat\beta_i^{\text{b}} \sim N(\hat\beta_i, \hat{\text{se}(\hat\beta_i)}).\]

    2. Upon obtaining \(\hat\beta_i^{\text{b}}\) for \(i, i=1,...,N\), the SNPs are subjected to an ordering in which they are arranged based on their \(z_i^{\text{b}}\)-statistics, in descending order. The \(z_i^{\text{b}}\)-statistic of SNP \(i\) is defined as \[z_i^{\text{b}} = \frac{\hat\beta_i^{\text{b}}}{\hat{\text{se}(\hat\beta_i)}}.\]Due to this ordering, each SNP is assigned a rank \(A(k)\). Equivalently, we define \(A(k)\) as the index corresponding to the \(k^{th}\) largest entry in the vector: \([z_1^{\text{b}}, ..., z_N^{\text{b}}] = [\frac{\hat\beta_1^{\text{b}}}{\hat{\text{se}(\hat\beta_1)}}, ..., \frac{\hat\beta_N^{\text{b}}}{\hat{\text{se}(\hat\beta_N)}}]\).

    3. Then, the estimated bias of SNP \(k\), the SNP with the \(k^{th}\) largest original \(z\)-statistic, takes the following form: \[\text{bias}_k = \frac{\hat\beta_{A(k)}^{\text{b}}- \hat\beta_{A(k)}^{\text{oob}}}{\hat{\text{se}(\hat\beta_{A(k)})}} = \frac{\hat\beta_{A(k)}^{\text{b}}- \hat\beta_{A(k)}}{\hat{\text{se}(\hat\beta_{A(k)})}},\]in which \(\hat\beta_{A(k)}^{\text{b}}\) is the bootstrap value of the SNP ranked in position \(k\) in the ordering of \(z_i^{\text{b}}\)-statistics, \(\hat\beta_{A(k)}^{\text{oob}} = \hat\beta_{A(k)}\) is that same SNP’s original \(\beta\) estimate and \(\hat{\text{se}(\hat\beta_{A(k)})}\) its standard error.

  • In the next step of the process, a cubic smoothing spline is fitted to the data in which the \(z\)-statistics are considered as the inputs and \(\text{bias}_k\), their corresponding outputs. The predicted values from this model fitting provides new estimates for the bias correction, \(\text{bias}_k^*\) for each SNP. This additional stage in which \(\text{bias}_k^*\) is obtained eliminates the need for several bootstrap iterations for each SNP in order to ensure competitive performance of the method. Thus, this results in an approach that is faster as only one bootstrap iteration is required per SNP, while using \(\text{bias}_k^*\) instead of \(\text{bias}_k\) ensures that the method has greater accuracy.

  • Finally, the new estimate for the true effect size of SNP k, the SNP with the \(k^{\text{th}}\) largest original \(z\)-statistic, is defined as \(\hat\beta_k^* = \hat\beta_k - \hat{\text{se}(\hat\beta_k)} \cdot \text{bias}_k^*\).

out_BR_ss <- BR_ss(summary_data = summary_stats, seed_opt=TRUE, seed=2020)
head(out_BR_ss)
#>   rsid        beta          se  beta_BR_ss
#> 1 7815  0.04771727 0.006568299  0.03628244
#> 2 3965  0.04609999 0.006346232  0.03505042
#> 3 4998 -0.04527792 0.006458823 -0.02990042
#> 4 9917  0.04164616 0.006334188  0.02915571
#> 5 7261  0.04162686 0.006343244  0.02909271
#> 6 6510  0.04254046 0.006500057  0.02965867

Comparing Results

As we have simulated our data set, we have the ability to quantify the amount of bias present in the original naive \(\beta\) estimates and also, in our adjusted estimates for each method. Thus, we can briefly examine how each correction method performed with respect to our simulated data set. Similar to Forde et al. (2023), assessment can take place using measures such as the estimated root mean square error of significant SNPs (rmse) and the estimated average bias over all significant SNPs (bias). These two evaluation metrics can be be mathematically defined as follows, in which there are \(N_{\text{sig}}\) significant SNPs:

  • rmse \(= \sqrt{\frac{1}{N_{\text{sig}}}\sum_{i=1}^{N_{\text{sig}}} (\hat\beta_{\text{adj},i} - \beta_i)^2}\)
  • bias \(= \frac{1}{N_{\text{sig}}} \sum_{i=1}^{N_{\text{sig}}} ( \hat\beta_{\text{adj},i}- \beta_i)\)

\(\star\) Note: This average bias will be computed separately for significant SNPs with positive and negative true effects to avoid the bias in either direction simply cancelling each other out.

\(\star\) Note: We compare our methods based on their performance with respect to SNPs that pass a certain threshold as these are the SNPs that have association estimates affected by Winner’s Curse and most often, it is the effect sizes of these SNPs that researchers are most interested in.

We consider a significance threshold of \(\alpha = 5 \times 10^{-8}\) and subset the outputted data frames as required.

## Simulated true effect sizes:
true_beta <- sim_dataset$true$true_beta 

## Subset required data sets:
out_EB_sig <- out_EB[2*pnorm(abs(out_EB$beta/out_EB$se), lower.tail = FALSE) < 5e-8,]
out_FIQT_sig <- out_FIQT[2*pnorm(abs(out_FIQT$beta/out_FIQT$se), lower.tail=FALSE) < 5e-8,]
out_BR_ss_sig <- out_BR_ss[2*pnorm(abs(out_BR_ss$beta/out_BR_ss$se), lower.tail=FALSE) < 5e-8,]

The estimated root mean square error of significant SNPs for each method is computed below. It can be seen that all methods demonstrate improvements over merely using the naive unadjusted association estimates as the rmse values for each method are lower than that of naive. The bootstrap method provides the lowest value for rmse.

## rmse
rmse <- data.frame(naive = sqrt(mean((out_CL$beta - true_beta[out_CL$rsid])^2)), CL1 = sqrt(mean((out_CL$beta.cl1 - true_beta[out_CL$rsid])^2)), CL2 = sqrt(mean((out_CL$beta.cl2 - true_beta[out_CL$rsid])^2)), CL3= sqrt(mean((out_CL$beta.cl3 - true_beta[out_CL$rsid])^2)),EB = sqrt(mean((out_EB_sig$beta_EB - true_beta[out_EB_sig$rsid])^2)), FIQT = sqrt(mean((out_FIQT_sig$beta_FIQT - true_beta[out_FIQT_sig$rsid])^2)), boot= sqrt(mean((out_BR_ss_sig$beta_BR_ss - true_beta[out_BR_ss_sig$rsid])^2)))
rmse
#>        naive       CL1         CL2        CL3         EB        FIQT
#> 1 0.01528707 0.0165505 0.009738775 0.01285767 0.01217377 0.008195889
#>          boot
#> 1 0.008379167

The next metric, the average bias over all significant SNPs with positive association estimates, is included below. Here, we wish to obtain values closer to zero than the naive approach and it can be seen that all methods achieve this.

## bias positive
pos_sig <- out_CL$rsid[out_CL$beta > 0]
pos_sigA <- which(out_CL$rsid %in% pos_sig)

bias_up <- data.frame(naive = mean(out_CL$beta[pos_sigA] - true_beta[pos_sig]), CL1 = mean(out_CL$beta.cl1[pos_sigA] - true_beta[pos_sig]), CL2 = mean(out_CL$beta.cl2[pos_sigA] - true_beta[pos_sig]), CL3= mean(out_CL$beta.cl3[pos_sigA] - true_beta[pos_sig]),EB = mean(out_EB_sig$beta_EB[pos_sigA] - true_beta[pos_sig]), FIQT = mean(out_FIQT_sig$beta_FIQT[pos_sigA] - true_beta[pos_sig]), boot= mean(out_BR_ss_sig$beta_BR_ss[pos_sigA] - true_beta[pos_sig]))
bias_up
#>        naive          CL1          CL2          CL3         EB         FIQT
#> 1 0.01268454 -0.003919891 -0.000829573 -0.002374732 0.01037972 -0.002254763
#>           boot
#> 1 -0.001818847

In a similar manner, the average bias over all significant SNPs with negative association estimates, is computed. Again, we see that the application of all methods result in, on average, less biased association estimates for these SNPs, as desired.

## bias negative
neg_sig <- out_CL$rsid[out_CL$beta < 0]
neg_sigA <- which(out_CL$rsid %in% neg_sig)

bias_down <- data.frame(naive = mean(out_CL$beta[neg_sigA] - true_beta[neg_sig]), CL1 = mean(out_CL$beta.cl1[neg_sigA] - true_beta[neg_sig]), CL2 = mean(out_CL$beta.cl2[neg_sigA] - true_beta[neg_sig]), CL3= mean(out_CL$beta.cl3[neg_sigA] - true_beta[neg_sig]),EB = mean(out_EB_sig$beta_EB[neg_sigA] - true_beta[neg_sig]), FIQT = mean(out_FIQT_sig$beta_FIQT[neg_sigA] - true_beta[neg_sig]), boot= mean(out_BR_ss_sig$beta_BR_ss[neg_sigA] - true_beta[neg_sig]))
bias_down
#>         naive         CL1         CL2         CL3           EB        FIQT
#> 1 -0.01250495 0.005887153 0.002532757 0.004209955 -0.006742937 0.002466867
#>          boot
#> 1 0.003934355

\(\star\) Note: In the above results, it is seen that the empirical Bayes method doesn’t seem to be performing as well as the other methods, i.e. the reductions in rmse and bias are small in comparison. The reason for this is perhaps that it is the original form of the empirical Bayes method that has been used here. It is possible that using another variation of the method, for example setting method="gam_nb" or method="scam", would result in improvements, as shown in Forde et al. (2023).

To complement the above, we provide an illustration of boxplots obtained for the root mean square error of significant SNPs (RMSE of sig. SNPs at \(5 \times 10^{-8}\)), plotted for each Winner’s Curse correction method (Method).

library(RColorBrewer)
library(ggplot2)
col <- brewer.pal(8,"Dark2")

all_results <- data.frame(rsid = c(rep(out_CL$rsid,7)), beta = c(rep(out_CL$beta,7)), se = c(rep(out_CL$se,7)), adj_beta = c(out_CL$beta,out_CL$beta.cl1,out_CL$beta.cl2,out_CL$beta.cl3,out_EB_sig$beta_EB,out_FIQT_sig$beta_FIQT,out_BR_ss_sig$beta_BR_ss), method = c(rep("naive",33),rep("CL1",33),rep("CL2",33),rep("CL3",33),rep("EB",33),rep("FIQT",33),rep("boot",33)))
all_results$rmse <- sqrt((all_results$adj_beta - true_beta[all_results$rsid])^2)
all_results$method <- factor(all_results$method, levels=c("naive","CL1", "CL2", "CL3", "EB", "boot", "FIQT"))

ggplot(all_results,aes(x=method,y=rmse,fill=method, color=method)) + geom_boxplot(size=0.7,aes(fill=method, color=method, alpha=0.2)) + scale_fill_manual(values=c(col[1],col[2],col[3],col[4],col[5],col[6],col[7])) + scale_color_manual(values=c(col[1],col[2],col[3],col[4],col[5],col[6],col[7]))  + xlab("Method") +
  ylab(expression(paste("RMSE of sig. SNPs at  ", 5%*%10^-8))) + theme_bw() + geom_hline(yintercept=0, colour="black") +  theme(axis.text.x=element_text(angle=90, vjust=0.5, hjust=1),text = element_text(size=12),legend.position = "none", strip.text = element_text(face="italic"))

In addition, we can gain a visual insight into the adjustments made by these functions by constructing a density plot with the adjusted absolute values along with the naive estimates and the true absolute \(\beta\) values of significant SNPs, as follows:

true <- data.frame(rsid = c(out_CL$rsid), beta=c(out_CL$beta), se=c(out_CL$se),adj_beta=true_beta[out_CL$rsid],method=c(rep("true",33)))
all_resultsA <- rbind(all_results[,1:5],true)

ggplot(all_resultsA, aes(x=abs(adj_beta),colour=method,fill=method)) + geom_density(alpha=0.05,size=1) + scale_color_brewer(palette="Dark2") + scale_fill_brewer(palette="Dark2") + xlim(-0.01,0.08) +
  ylab("Density") + xlab(expression(paste("|", beta, "| " , "of sig. SNPs at  ", 5%*%10^-8))) + theme_bw() +  theme(text = element_text(size=12)) + labs(fill = "Method",colour="Method")

\(~\) \(~\)

Application of methods to a real data set

In this final section, we will provide a short demonstration of applying these Winner’s Curse correction methods to a real data set.

Briefly, the data set we will use is a subset of the first BMI data set used in Forde et al. (2023), which was obtained from the UK Biobank. In order to obtain a smaller data set suitable for demonstration purposes in this vignette, LD clumping was performed on the data set used in the aforementioned paper using the ld_clump() function from the ieugwasr R package. The LD (r-squared) threshold for clumping was set to 0.01 while the physical distance threshold used was 100 kb. This procedure provided a data set of 272,604 SNPs with corresponding association estimates (beta) and standard errors (se) as well as association estimates obtained in an independent replication GWAS of a similar size (beta_rep). The first six rows of this data set are outputted below.

BMI_UKBB <- read.csv("https://raw.githubusercontent.com/amandaforde/winnerscurse_realdata/main/data/summary_stats_BMI.txt",quote = "", row.names = NULL, col.names = c("chr","pos","rsid","beta","se","beta_rep"), stringsAsFactors = FALSE) 

head(BMI_UKBB)
#>   chr    pos          rsid        beta        se    beta_rep
#> 1   1 801536  "rs79373928"  0.00691695 0.0672237 -0.00845073
#> 2   1 809876  "rs57181708"  0.08085120 0.0271055  0.05181780
#> 3   1 815656 "rs557970564"  0.16602900 0.0887715 -0.13927300
#> 4   1 834430 "rs533316052" -0.06663160 0.0685312 -0.05314070
#> 5   1 837657 "rs149737509" -0.00116905 0.0623581 -0.06175060
#> 6   1 852758   "rs4970462" -0.02576140 0.0200347 -0.02956890

We apply the correction methods to this data set as follows and obtain adjusted association estimates for significant SNPs at the genome-wide significance threshold of \(5 \times 10^{-8}\).

BMI_data <- BMI_UKBB[,3:5]

out_CL_BMI <- conditional_likelihood(summary_data = BMI_data, alpha = 5e-8)
out_EB_BMI <- empirical_bayes(summary_data = BMI_data)
out_FIQT_BMI <- FDR_IQT(summary_data = BMI_data)
out_boot_BMI <- BR_ss(summary_data = BMI_data, seed=2020)

## Subset required data sets:
out_EB_BMI_sig <- out_EB_BMI[2*pnorm(abs(out_EB_BMI$beta/out_EB_BMI$se), lower.tail = FALSE) < 5e-8,]
out_FIQT_BMI_sig <- out_FIQT_BMI[2*pnorm(abs(out_FIQT_BMI$beta/out_FIQT_BMI$se), lower.tail=FALSE) < 5e-8,]
out_boot_BMI_sig <- out_boot_BMI[2*pnorm(abs(out_boot_BMI$beta/out_boot_BMI$se), lower.tail=FALSE) < 5e-8,]

\(\star\) Note: In the case of real data, the true association values are unknown. Therefore, in order to evaluate the performance of methods, we will make use of beta_rep, the association estimates obtained in an independent replication GWAS of a similar size. These association estimates are considered to be unbiased, and thus can be used to obtain an expression for the estimated mean square error (MSE) of significant SNPs. This expression is given in Eq [14] of Forde et al. (2023) and can be computed for each method as shown below. It can be seen that each method provides an estimated MSE value lower than that of the naive approach of using the unadjusted association estimates from the discovery GWAS.

## mse
beta_rep <- BMI_UKBB$beta_rep[BMI_UKBB$rsid %in% out_CL_BMI$rsid]

mse <- data.frame(naive = mean((beta_rep - out_CL_BMI$beta)^2) - mean(out_CL_BMI$se^2), CL1 = mean((beta_rep - out_CL_BMI$beta.cl1)^2) - mean(out_CL_BMI$se^2), CL2 = mean((beta_rep - out_CL_BMI$beta.cl2)^2) - mean(out_CL_BMI$se^2), CL3 = mean((beta_rep - out_CL_BMI$beta.cl3)^2) - mean(out_CL_BMI$se^2), EB = mean((beta_rep - out_EB_BMI_sig$beta_EB)^2) - mean(out_EB_BMI_sig$se^2), boot = mean((beta_rep - out_boot_BMI_sig$beta_BR_ss)^2) - mean(out_boot_BMI_sig$se^2), FIQT = mean((beta_rep - out_FIQT_BMI_sig$beta_FIQT)^2) - mean(out_FIQT_BMI_sig$se^2))

mse
#>        naive        CL1        CL2        CL3         EB       boot       FIQT
#> 1 0.03608036 0.02946674 0.02870324 0.02897353 0.03029724 0.02874184 0.02859518

For illustration purposes, we again simply produce a density plot with the adjusted absolute values along with the naive estimates and the true absolute \(\beta\) values of significant SNPs.

all_results <- data.frame(rsid = c(rep(out_CL_BMI$rsid,7)), beta = c(rep(out_CL_BMI$beta,7)), se = c(rep(out_CL_BMI$se,7)), adj_beta = c(out_CL_BMI$beta,out_CL_BMI$beta.cl1,out_CL_BMI$beta.cl2,out_CL_BMI$beta.cl3,out_EB_BMI_sig$beta_EB,out_FIQT_BMI_sig$beta_FIQT,out_boot_BMI_sig$beta_BR_ss), method = c(rep("naive",214),rep("CL1",214),rep("CL2",214),rep("CL3",214),rep("EB",214),rep("FIQT",214),rep("boot",214)))
all_results$method <- factor(all_results$method, levels=c("naive","CL1", "CL2", "CL3", "EB", "boot", "FIQT"))
rep <- data.frame(rsid = c(out_CL_BMI$rsid), beta=c(out_CL_BMI$beta), se=c(out_CL_BMI$se),adj_beta=beta_rep,method=c(rep("rep",214)))
all_resultsA <- rbind(all_results[,1:5],rep)

ggplot(all_resultsA, aes(x=abs(adj_beta),colour=method,fill=method)) + geom_density(alpha=0.05,size=1) + scale_color_brewer(palette="Dark2") + scale_fill_brewer(palette="Dark2") + xlim(-0.1,0.4) +
  ylab("Density") + xlab(expression(paste("|", beta, "| " , "of sig. SNPs at  ", 5%*%10^-8))) + theme_bw() +  theme(text = element_text(size=12)) + labs(fill = "Method",colour="Method")