Agnostic studies in epidemiology

Abstract

In epidemiology there has been an consistent effort to construct refined methods aiming towards the ability to draw casual inference between an exposure and an outcome. This thesis, partly inspired by the genome-wide association studies, has on the contrary strived towards exploration of data. The fundamental idea has been to decipher associations between one or many exposures with one or many outcomes.

Using population-based register data from Sweden, this thesis explored the association between ABO blood group and RhD status in 1,217 disease categories, the occurrence of transfusion-transmitted disease examining 1,155 disease categories, the spectrum of adverse events in tyrosine kinase-inhibitor treated patients with chronic phase chronic myeloid leukemia in 670 disease categories, and lastly the occurrence of familial aggregation of 60 cancerous disease. To account for multiple testing we use previously employed methods of adjustment.

In Paper I, using the large-scale donation-transfusion database, SCANDAT-3S with 8 million individuals, we identified 49 associations with ABO blood group and disease. Many associations were previously known but we identified a novel association of a protective role for blood group B, as compared to O, in suffering kidney stones. For RhD status, we identified only one disease after adjustment for multiple testing, namely pregnancy-induced hypertension.

In Paper II, which used the same database and the unique connection between blood donor, the blood product and the recipients of blood, we identified 15 disease categories that seemed to be transfusion-transmitted. Among them there were strong signals suggesting transmission in hepatitis virus and HIV. For most other findings, the effect sizes were small. A general conclusion was that the current practice in Sweden, regarding transfusion safety, seems acceptable in terms of the risk of transfusion-transmission of disease.

In Paper III, we used a database covering the full Swedish chronic myeloid leukemia-population diagnosed since 2002. In this study, also consisting of a matched control cohort, we identified 142 disease categories with increased incidence as compared to the control cohort. We also found 41 associations between tyrosine kinase inhibitors, used for the treatment of chronic myeloid leukemia, and a disease category. No unknown severe adverse events were found. In Paper IV, we created 3.5 million pedigrees using the Multi-generation Register and explored cancerous disease clustering within pedigrees. We identified multiple cancer syndromes, e.g. BRCA1/2 and hereditary colon cancer.

The approach, agnostic in that we strive towards testing all possible hypotheses without prejudice, has the advantage of removing or at least reducing the researcher bias – where the research hypothesis is constrained by the environment of the individual researcher. The approach is mainly limited by the problems of misclassification of exposures and outcomes, the inability to optimally construct modelling for a large set of hypotheses, and by false discoveries. We have proposed some solutions to overcome these issues. A main aspect is that the method should not be used to draw inference but rather to generate hypothesis for future refined studies in a world with increasing amounts of high-resolution data.