In this thesis, we address the identification of biomarkers in high-dimensional omics data. The identification of valid biomarkers is especially relevant for personalized medicine that depends on accurate prediction rules. Moreover, biomarkers elucidate the provenance of disease, or molecular changes related to disease. From a statistical point of view the identification of biomarkers is best cast as variable selection. In particular, we refer to variables as the molecular attributes under investigation, e.g. genes, genetic variation, or metabolites; and we refer to observations as the specific samples whose attributes we investigate, e.g. patients and controls. Variable selection in high-dimensional omics data is a complicated challenge due to the characteristic structure of omics data. For one, omics data is high-dimensional, comprising cellular information in unprecedented details. Moreover, there is an intricate correlation structure among the variables due to e.g internal cellular regulation, or external, latent factors. Variable selection for uncorrelated data is well established. In contrast, there is no consensus on how to approach variable selection under correlation. Here, we introduce a multivariate framework for variable selection that explicitly accounts for the correlation among markers. In particular, we present two novel quantities for variable importance: the correlation-adjusted t (CAT) score for classification, and the correlation-adjusted (marginal) correlation (CAR) score for regression. The CAT score is defined as the Mahalanobis-decorrelated t-score vector, and the CAR score as the Mahalanobis-decorrelated correlation between the predictor variables and the outcome. We derive the CAT and CAR score from a predictive point of view in linear discriminant analysis and regression; both quantities assess the weight of a decorrelated and standardized variable on the prediction rule. Furthermore, we discuss properties of both scores and relations to established quantities. Above all, the CAT score decomposes Hotelling’s T 2 and the CAR score the proportion of variance explained. Notably, the decomposition of total variance into explained and unexplained variance in the linear model can be rewritten in terms of CAR scores. To render our approach applicable on high-dimensional omics data we devise an efficient algorithm for shrinkage estimates of the CAT and CAR score. Subsequently, we conduct extensive simulation studies to investigate the performance of our novel approaches in ranking and prediction under correlation. Here, CAT and CAR scores consistently improve over marginal approaches in terms of more true positives selected and a lower model error. Finally, we illustrate the application of CAT and CAR score on real omics data. In particular, we analyze genomics, transcriptomics, and metabolomics data. We ascertain that CAT and CAR score are competitive or outperform state of the art techniques in terms of true positives detected and prediction error.

Um Daten am effektivsten zu nutzen, muss die moderne Ökonometrie ihren Werkzeugkasten an Modellen erweitern und neu denken. Das Feld, in dem diese Transformation am besten beobachtet werden kann, ist die kausale Inferenz. Diese Dissertation verfolgt die Absicht Probleme zu untersuchen, Lösungen zu präsentieren und neue Methoden zu entwickeln Machine Learning zu benutzen, um kausale Parameter zu schätzen. Dafür werden in der Dissertation zuerst verschiedene neuartige Methoden, welche als Ziel haben heterogene Treatment Effekte zu messen, eingeordnet. Im zweiten Schritt werden, basierend auf diesen Methoden, Richtlinien für ihre Anwendung in der Praxis aufgestellt. Der Parameter von Interesse ist der „conditional average treatment effect“ (CATE). Es kann gezeigt werden, dass ein Vergleich mehrerer Methoden gegenüber der Verwendung einer einzelnen Methode vorzuziehen ist. Ein spezieller Fokus liegt dabei auf dem Aufteilen und Gewichten der Stichprobe, um den Verlust in Effizienz wettzumachen. Ein unzulängliches Kontrollieren für die Variation durch verschiedene Teilstichproben führt zu großen Unterschieden in der Präzision der geschätzten Parameter. Wird der CATE durch Bilden von Quantilen in Gruppen unterteilt, führt dies zu robusteren Ergebnissen in Bezug auf die Varianz. Diese Dissertation entwickelt und untersucht nicht nur Methoden für die Schätzung der Heterogenität in Treatment Effekten, sondern auch für das Identifizieren von richtigen Störvariablen. Hierzu schlägt diese Dissertation sowohl die „outcome-adaptive random forest“ Methode vor, welche automatisiert Variablen klassifiziert, als auch „supervised randomization“ für eine kosteneffiziente Selektion der Zielgruppe. Einblicke in wichtige Variablen und solche, welche keine Störung verursachen, ist besonders in der Evaluierung von Politikmaßnahmen aber auch im medizinischen Sektor wichtig, insbesondere dann, wenn kein randomisiertes Experiment möglich ist. / To use data effectively, modern econometricians need to expand and rethink their toolbox. One field where such a transformation has already started is causal inference. This thesis aims to explore further issues, provide solutions, and develop new methods on how machine learning can be used to estimate causal parameters. I categorize novel methods to estimate heterogeneous treatment effects and provide a practitioner’s guide for implementation. The parameter of interest is the conditional average treatment effect (CATE). It can be shown that an ensemble of methods is preferable to relying on one method. A special focus, with respect to the CATE, is set on the comparison of such methods and the role of sample splitting and cross-fitting to restore efficiency. Huge differences in the estimated parameter accuracy can occur if the sampling uncertainty is not correctly accounted for. One feature of the CATE is a coarser representation through quantiles. Estimating groups of the CATE leads to more robust estimates with respect to the sampling uncertainty and the resulting high variance. This thesis not only develops and explores methods to estimate treatment effect heterogeneity but also to identify confounding variables as well as observations that should receive treatment. For these two tasks, this thesis proposes the outcome-adaptive random forest for automatic variable selection, as well as supervised randomization for a cost-efficient selection of the target group. Insights into important variables and those that are not true confounders are very helpful for policy evaluation and in the medical sector when randomized control trials are not possible.

In this thesis, we address the identification of biomarkers in high-dimensional omics data. The identification of valid biomarkers is especially relevant for personalized medicine that depends on accurate prediction rules. Moreover, biomarkers elucidate the provenance of disease, or molecular changes related to disease. From a statistical point of view the identification of biomarkers is best cast as variable selection. In particular, we refer to variables as the molecular attributes under investigation, e.g. genes, genetic variation, or metabolites; and we refer to observations as the specific samples whose attributes we investigate, e.g. patients and controls. Variable selection in high-dimensional omics data is a complicated challenge due to the characteristic structure of omics data. For one, omics data is high-dimensional, comprising cellular information in unprecedented details. Moreover, there is an intricate correlation structure among the variables due to e.g internal cellular regulation, or external, latent factors. Variable selection for uncorrelated data is well established. In contrast, there is no consensus on how to approach variable selection under correlation. Here, we introduce a multivariate framework for variable selection that explicitly accounts for the correlation among markers. In particular, we present two novel quantities for variable importance: the correlation-adjusted t (CAT) score for classification, and the correlation-adjusted (marginal) correlation (CAR) score for regression. The CAT score is defined as the Mahalanobis-decorrelated t-score vector, and the CAR score as the Mahalanobis-decorrelated correlation between the predictor variables and the outcome. We derive the CAT and CAR score from a predictive point of view in linear discriminant analysis and regression; both quantities assess the weight of a decorrelated and standardized variable on the prediction rule. Furthermore, we discuss properties of both scores and relations to established quantities. Above all, the CAT score decomposes Hotelling’s T 2 and the CAR score the proportion of variance explained. Notably, the decomposition of total variance into explained and unexplained variance in the linear model can be rewritten in terms of CAR scores. To render our approach applicable on high-dimensional omics data we devise an efficient algorithm for shrinkage estimates of the CAT and CAR score. Subsequently, we conduct extensive simulation studies to investigate the performance of our novel approaches in ranking and prediction under correlation. Here, CAT and CAR scores consistently improve over marginal approaches in terms of more true positives selected and a lower model error. Finally, we illustrate the application of CAT and CAR score on real omics data. In particular, we analyze genomics, transcriptomics, and metabolomics data. We ascertain that CAT and CAR score are competitive or outperform state of the art techniques in terms of true positives detected and prediction error.