Tartu University, Estonia
Krista Fischer has a PhD in mathematical statistics (University of Tartu, 1999). One of her main research areas between 1994-2010 was causal inference in clinical trials, mainly structural mean models for analyzing the effect of non-compliance. She worked as a post-doctoral researcher at the University of Ghent, Belgium (1999-2001), Associate Professor in Biostatistics at the Medical Faculty of the University of Tartu (2001-2007), Investigator Scientist at the MRC Biostatistics Unit, Cambridge, UK (2007-2010) and has been working as Senior Researcher at the Estonian Genome Center, University of Tartu since 2010. Her current research interests include different areas of statistical modeling of -omics data, including the aspects of causal modeling. In addition, she is working on the development of polygenic risk scores for personalized risk prediction and on models involving metabolomics data.
Causal association structures in -omics data: how far can we get with statistical modelling?
This talk mainly concentrates on the setting where association of one genotype marker (typically SNP) with two correlated phenotypes is studied. In so-called “Mendelian Randomization” studies the main parameter of interest corresponds to a causal effect of one phenotypic trait on another trait, whereas a genetic marker is used as an instrument. Despite of the increasing number of publications using this methodological approach, the underlying assumptions are often overlooked. Therefore, many of the published effect estimates may actually be biased and misleading. One of the main untestable assumptions is the “no pleiotropy” assumption – the genotype has a direct causal effect on one phenotype only, whereas the effect on the second phenotype is fully mediated by the first one.
When this is not fulfilled, the genotype is said to have a pleiotropic effect on both phenotypes, whereas another class of models is been designed to estimate such effects. However, we will show that mathematically one cannot distinguish between the two models: the model underlying the Mendelian Randomization scenario and the model for pleiotropic effect. We will discuss whether some sensitivity analysis methods may help to draw a correct conclusion here.
In addition, we discuss another assumption underlying the Mendelian Randomization idea: the “no-treatment-effect heterogeneity” assumption. Here a parallel can be drawn with randomized clinical trials, where this assumption is crucial to allow for active treatment on the control arm. Using also simulation results, the effect of deviations from this assumption is studied.