The war on error.

Risk is theworld’s biggest industry and permeates almost every aspect of social and economic life (Adams, 1995). For the search term ‘risk’ Google will return 893 million hits, as compared with only 491 million for ‘disease’, but 1670 million for ‘sex’. Risk and risk determining factors (risk factors) are the domain of clinical epidemiology. The term ‘clinical epidemiology’ was coined by Alvan Feinstein in 1968 (Feinstein, 1968). He described it as: ‘The clinicostatistical study of diseased populations; the intellectual activities of clinical epidemiology include the following: the occurrence rates and geographic distribution of disease; the patterns of natural and post-therapeutic events that constitute varying clinical courses in the diverse spectrumof disease; and the clinical appraisal of therapy’. Clinical epidemiology aims to describe what the burden of disease is in a community and to provide the methods to help understand causes, risk factors and contributing or predisposing conditions that influence disease (Morrow, 2015). Risk factors in essence are however simple fractionsmade up of a numerator and a denominator. Fractions are all they are, and risk is what they represent, a statistic, or a predictor atbest.But even ‘the claim thatpast behavior is thebestpredictor of future behavior does not mean that past behavior causes future behavior’ (Wikstrom, 2007). This quotation should be on every clinician’s office wall. Clinical epidemiology has developed progressively more sophisticated tools to estimate and express risk, frequently in so-called ‘adjusted’ odds ratios. But the more sophisticated these tools became the less critical researchers, readers, journalists and policy makers appear to have become. Details—and especially limitations—of new research findings yielded to over-simplistic (and biologically implausible) headlines in the popular press: ‘IVF causes skin cancer!’ And now we have entered the world of bioinformatics, with its mammoth datasets, subjected to unfathomably complex high-level statistical software packages. Information technology has proliferated dramatically and the ability to store, retrieve and analyze large amounts of data has increased concomitantly. There are two major forms of clinical databases: administrative databases (often maintained by health insurance companies or health care policy makers) that focus on incidence rates and costs; and clinical registries (by doctors or hospitals) that focus on quality issues and treatment outcome (Cook and Collins, 2015). Both may enable research as a secondary activity. A third novel form of big database, as a primary research activity, is found in genome research. GenomeWide Association Studies (GWAS) examine the associationbetween genetic variants, usually single-nucleotide polymorphisms (SNPs), and a trait, usually a disease, by comparing theDNAof cases (with the disease) and controls (without disease). Immense numbers of genetic variants become available—typically hundreds of thousands to one million or more of tested SNPs—with fold changes as the critical yardstick. Poorly defined cases, insufficient sample size and absence of adjustment for multiple testing are but a few of the problems encountered (Pearson and Manolio, 2008). And also here, the finding that a certain genetic variant co-segregates with a certain cancer does not mean that that particular SNP causes the cancer. All database research has in common that the data are not always being used for the purpose they were designated for. Another problem is that the large number of statistical tests performed presents an unprecedented potential for falsepositive results (Pearson andManolio, 2008). Themore comparisons are made, the greater is the chance that something will turn out statistically significant, even when no relationship exists. Checking for twenty different, normally distributed risk factors will—by definition—produce one significant finding if a P-value of 0.05 is set as the limit of significance. Checking for one million SNPs demands even more statistical restraint. P-values of 0.00000005 or less are the rule in this field of research. In the current issue of the journal we present such a big data study (Reigstad et al., 2015). The authors studied, with impeccable epidemiological techniques, a population-based cohort consisting of all women