Increasingly, genetic studies contribute to our understanding of the pathogenesis of diabetes and its complications at the molecular level [1, 2]. With the advent of powerful new high-throughput analytical methods [3], there is every reason to expect further insights. Virtually all such genetic studies include some test of the association between genome sequence variation and the phenotype of interest, be that diabetes itself, the presence of a given complication, or measures of some diabetes-related intermediate trait. Association studies of this type have scored several recent successes in relation to both type 1 (CTLA4, INS and PTPN22 genes) [2, 4, 5] and type 2 diabetes (PPARG, KCNJ11) [6–8], identifying variants that are unequivocally associated with disease. Two further association studies, examining variation in the genes encoding the Krüppel-like transcription factor family and the receptors for adiponectin, are published in this issue of Diabetologia [9, 10].

Despite these advances, there are major concerns regarding the overall performance and robustness of genetic association studies. All too often, initial positive (or negative) association findings fail the test of replication, leaving the literature littered with the detritus of uncertain and poorly reproducible associations. As set out in a number of excellent review articles [11–15], the origins of such inconsistency lie in a series of common methodological failings. These include (but are not limited to) the use of sample sizes that are inadequately powered for the task in hand, incomplete assessments of sequence variation within the locus of interest, technical errors in genotyping, and the inappropriate interpretation of data when large number of statistical tests have been performed. Furthermore, discrepancies introduced by these failings are compounded by the rather low prior odds that any given variant contributes to susceptibility to a given trait and by the understandable bias of reviewers and journal editors towards the publication of novel, superficially interesting, positive associations (whilst otherwise well-performed association studies reporting no association are often dismissed as ‘negative’). In fact, where power is low and liberal thresholds for declaring significance are used, the vast majority of such ‘positive’ associations will be erroneous [16].

Of course, it is not all bad news. Susceptibility variants for diabetes and its complications do exist. However, with the exception of HLA in type 1 diabetes, the effect sizes are usually modest. Such effects can be reliably and reproducibly detected provided the studies are adequately powered and feature appropriately constructed sample sets, careful genotyping and appropriate analysis [2, 4–8]. For example, the variants P12A in PPARG and E23K in KCNJ11 have been detected repeatedly within large type 2 diabetes case-control samples, such that the overall evidence for association now exceeds any reasonable correction for genome-wide significance [17]. Other type 2 diabetes-susceptibility effects—at CAPN10 for example—look increasingly convincing as larger sample sizes have been assembled, though the combined evidence is not, as yet, incontrovertible [18].

As these successes show, and as connoisseurs of the association game will know, single studies rarely, if ever, provide conclusive evidence that a given variant is (or is not) associated with disease. Instead, evidence typically accumulates over multiple studies. Crucially, replication not only provides access to increasingly large cumulative sample sizes but also insures against the inevitable biases, confounders and technical problems (latent population stratification and genotyping errors, for example) that can unavoidably afflict any individual study [19]. Only when extensive replication studies have been completed does a meaningful estimate of the true effect size emerge.

All of this leaves journals such as Diabetologia in something of a quandary. On the one hand, no journal wants to publish papers that have a depressingly high chance of being ‘wrong’ [16]. On the other hand, to insist that each individual study should reach some standard of unequivocal proof would paralyse the field. It is hard to think of any genetics paper published in Diabetologia that could be considered to have cleared the latter hurdle.

Journal editors with limited page numbers also struggle with the question of bias referred to above. Manuscripts reporting positive associations are undoubtedly more ‘newsworthy’ than those that find no association (not least because the large number of variants within the genome means that only a small percentage of variants will ever be truly associated with a given trait). Yet, in the interests of establishing a true, unbiased picture of the overall strength of the evidence for (or against) association at a given variant, it is important that all well-performed studies stand an equitable chance of publication, whatever their outcome. To facilitate more complete capture of association data, there have been calls for electronic registration [14, 20], and some provisional efforts have been made to establish such registries (e.g. http://geneticassociationdb.nih.gov, last accessed in April 2005). However, for the time being, scientific probity requires that well-performed studies that fail to detect associations (especially those that convincingly fail to replicate published positive results) should be regarded as competitive for publication.

Two papers in this issue of Diabetologia illustrate the issues discussed above rather well. Kanazawa et al. [9] describe association studies in a series of Japanese case-control samples, which were designed to evaluate a family of biological candidates (the Krüppel-like transcription factors). One of the single nucleotide polymorphisms (SNPs) typed (in KLF7) shows an apparently impressive association (p=0.000057, odds ratio=1.59) with type 2 diabetes. There is evidence of replication across the two sample sets and some functional data to support the candidacy of the gene concerned. Although these results are interesting, the authors are right to be circumspect in their interpretation of the data given the chequered history of association findings. At the very least, this study is likely to have overestimated the effect size at the KLF7 SNP (the so-called ‘winner’s curse’ [20]). At worst, future replication studies (in Japanese and other populations) may fail to confirm the association at all, and the candidacy of KLF7 will founder on the rocks that have claimed so many other promising associations. These results are therefore far from conclusive. However, publication of this well-performed study enables other groups to test the robustness of this association in additional data sets.

In the second paper, Hara et al. [10] have undertaken a detailed assessment of the relationship between the variation in two excellent candidate genes (encoding the receptors for the adipocytokine, adiponectin) and type 2 diabetes, considering both the discrete disease phenotype and diabetes-related intermediate traits. A dense inventory of common variants in these genes was examined and, to the extent allowed by the sample sizes deployed, no significant association effects were detected. Once again, the authors have been appropriately cautious in their interpretation, pointing out that the variants typed do not represent a complete inventory of sequence variation in and around the genes, and that larger sample sizes would have provided greater power to detect (and/or exclude) more modest effects. However, given the biological credentials of the genes concerned as candidates for susceptibility to type 2 diabetes, these remain interesting data which, despite these limitations, add significantly to our knowledge.

Editorial judgements about which papers should be published in a journal such as Diabetologia are inevitably somewhat subjective. The vagaries of the peer-review process are well known to all who participate. In an attempt to foster consistency in the treatment of such manuscripts and to encourage transparency in the decisions reached, Diabetologia has developed a set of guidelines for genetic association studies. These are posted on the journal’s website (http://www.diabetologia-journal.org/genetics%20guidelines.htm).

The term ‘guidelines’ is used advisedly. These are very definitely not meant to be thresholds that must be cleared if a paper is to be accepted for publication in Diabetologia. Very few association studies are entirely blameless, and the authors of those that are might reasonably be seeking publication in the very highest impact journals. What these guidelines do reflect are the criteria that we expect those participating in the peer-review process for Diabetologia to use when evaluating association studies. As well as encouraging consistency and transparency, we also hope that these guidelines will lead to a greater awareness of these important methodological issues by those undertaking association studies. In addition, by asking authors to ensure that submissions include information essential for their evaluation (e.g. rs numbers of variants, information on genotyping accuracy), the guidelines may even accelerate the review process.

Finally, we hope that these guidelines will contribute to a more mature dialogue between authors and the readership of the journal, one marked less by emphasis on the ‘headline’ p value (which all too often induces authors to construct some post-hoc narrative around the single nominally significant result that emerged from the forest of statistical output) and more by methodological and technical probity, adequate sample size, and appropriate and considered interpretation of the findings (including a realistic assessment of the limitations of the study).

The perceived unreliability of association studies has been a weight around the neck of genetic researchers over the past decade. In years to come, genome-wide association studies will generate data on a previously unimaginable scale (billions of genotypes). These studies should provide a comprehensive view of the association landscape of type 2 diabetes, but only if they are well performed and correctly interpreted. Greater emphasis on the application of robust methodologies for association studies is therefore particularly timely.