## Introduction

Genome-wide association studies (GWAS) and candidate gene studies have highlighted regions of the genome containing variants affecting disease susceptibility. The next stage is fine-mapping of these regions to identify the variants most likely to be causal. This task is confounded by high correlation between variants in a small chromosomal region. The effects of this correlation as well as sampling variation mean that in tests of association the variant with the largest likelihood or smallest *p*-value will not necessarily be the causal variant. Several statistical methods for analysing fine-mapped data have now been published but guidelines are needed to determine which of these will give the highest true positive rates (TPRs) and lowest false positive rates (FPRs) and in which scenarios.

Methods for analysing fine-mapped data include those that analyse multiple variants in a region simultaneously, for example, penalised and nonpenalised regression methods and Markov chain Monte Carlo routines. Some such methods are given in reviews by Ayers & Cordell (2010) and Abraham et al. (2013), including the popular HyperLasso (Hoggart et al., 2008). There are also fully Bayesian methods implemented in the software pi-MASS (Guan & Stephens, 2011). Also, some recent methods attempt to include external data such as functional annotation, for example, *p*-value weighting (Saccone et al., 2008) and a Bayesian latent variable model (BLVM, Fridley et al., 2011). However, we have chosen to compare a subset of statistical analyses which should work well when a single causal variant is present in the chromosomal region of interest. In these methods, each single nucleotide polymorphism (SNP) is analysed separately and they are then ranked in some way based on the likelihood or *p*-value from a logistic model or based on linkage disequilibrium (LD) with or proximity to the top hit SNP in the region. The methods we consider do not make use of any available functional data. To our knowledge this set of methods has not previously been compared in a thorough simulation study such as this.

All of the statistics that this report examines could be used as filters to remove noncausal variants from the set of all candidate causal variants. The variants considered in this work are SNPs but the methods and results discussed can be applied directly to any other variants which can be modelled via a logistic regression model. Successful filters will reduce the initial set of SNPs down to a much smaller group in which it is highly probable that the true causal variant remains. Other techniques, such as the biological analysis of pathways in cell lines, can then be used to identify the causal variant. These methods are expensive, so reducing the number of variants to take forward is of paramount importance.

The first methods we examine are based on *p*-values and likelihoods. It is common in GWAS to rank SNPs by *p*-values either from Cochran–Armitage trend tests or from Wald tests and both of these methods have now also been used in the context of fine-mapping (Miki et al., 2010; Adrianto et al., 2012). An alternative to using *p*-values is to use the likelihood (or equivalently log-likelihood) from fitted regression models. Several studies (including Easton et al., 2007; Udler et al., 2009, 2010a; French et al., 2013), rank SNPs based on likelihoods and the usual practice is to retain the set of SNPs with likelihoods within a prespecified ratio of the highest likelihood. This method leads to variable numbers of SNPs being retained. We examine this relative likelihood (RL) filter as well as the alternative of retaining a prespecified proportion of all SNPs based on ranking by likelihood. These statistics are attractive for filtering because they are easily obtained from standard analyses.

The remaining methods relate to LD structure. Within a small chromosomal region, LD can be high between SNPs. When the top hits from GWAS are found, these are not assumed to be the causal SNPs, but it is often postulated that the causal SNP lies within the same gene or LD block as the tagSNP. Alternatively, a handful of candidates may be suggested based on high LD with the tagSNP (, for example). We formalise three filtering methods based on these ideas: ranking by genetic map distance, *r*^{2} and with the top hit (the SNP with the largest likelihood). The final method (Zhu et al., 2012) we examine is also LD-based, but takes into account the LD between each SNP and the top hit compared to the LD between the SNP and tagSNPs in the region. Although we use the analyses set out by Zhu et al. (2012), we use it in a slightly different setting, as it is designed for use with tagSNPs from a GWAS. As far as we are aware the application of these LD- and distance-based methods to fine-mapped genotype data and their comparison with standard univariate statistical methods is novel.

We found that percentile filtering based on ranked likelihoods was the most efficacious method in all the scenarios we investigated. To explore the utility of this approach, this study considers the impact of effect size, sample size, minor allele frequency (MAF), mode of inheritance and filter threshold on the effectiveness of the filter proposed. We also consider whether these results apply to filtering in regions of the genome with strikingly different LD structures. A range of plausible odds ratios (ORs) were used in our simulations, as well as relatively large sample sizes consistent with numbers being used in the era of disease-specific consortia.