Veranstaltungsübersicht -- Sommersemester 2012 (Archiv)

Hospital-acquired (nosocomial) infections (NIs) constitute a major medical problem in that they increase patients’ morbidity and mortality. They also lead to an increase of costs due to additional medical care and prolongation of length of hospital stay. Some of these infections are thought to be preventable; identification of patients at high risk is therefore of central importance, as is the assessment of their sequelae. Incidence of nosocomial infections varies according to type, size and other characteristics of hospitals and wards and is usually highest in intensive care units.

The vast majority of evidence on determinants and consequences of NIs comes from observational studies that besides being prone to various, well-known biases often suffer from inadequate statistical analyses. Two major flaws are prominent in the infection literature: The first is caused by ignoring the temporal dynamics, i. e. that NIs are time-dependent exposures. Implicitly assuming that they are already known at time of admission leads to so-called time-dependent bias and usually to an exaggerated estimate of their consequences. This issue is frequently related to design aspects and sometimes further complicated by delayed study entry that leads to left truncated data. The second flaw is caused by ignoring competing risks so that patients are censored at the time a competing event occurs. Since the cumulative incidence function depends on the hazard of the event of primary interest as well as on the hazard of all competing events ignoring them may lead to a completely misleading assessment.

To address these issues, we advocate a multistate model where the temporal dynamics as well as competing risks are incorporated, quantities of interest can be estimated and inference can be based on with R packages being available. The model and the resulting strategies for the statistical analysis are explained and illustrated by using data from studies on nosocomial pneumonia in intensive care units and on in-hospital bacteraemia caused by Staphylococcus aureus in a Scottish hospital. Furthermore, we also will explain the key issues made using results from a study on hospital-acquired bacteraemia in African children, a study on meticillin-resistant Staphylococcus aureus colonisation and infection in surgical patients and - outside the realm of NIs - from a popular investigation on the mortality of Oscar nominees.

In the last part of the presentation we will discuss problems related with defining and estimating the attributable mortality due to NIs. It is shown that this concept can be nicely embedded into the multistate model framework. However, interpretational problems remain since the evidence that can be derived from observational studies is limited. So not only an appropriate statistical treatment of data on nosocomial infections, but also more randomized studies investigating the impact of preventive measures are urgently needed!

Dieser Vortrag findet im Rahmen des "Mini-symposium on Statistical approaches for integrating high-dimensional molecular data from different sources" statt.

Abstract:

The Lasso is a popular tool for high-dimensional model building. First I will review recent computational advances that enable the Lasso to be applied to large datasets. Then I will describe very recent work on fitting interaction models. Statisticians commonly demand that an interaction only be included in a model if both variables are marginally important. We study the problem of identifying hierarchical two-way interaction models from the viewpoint of the Lasso (i.e., L1-penalized regression). We show that by adding a set of convex constraints to the Lasso problem, we can produce sparse interaction models that honor the hierarchy restriction. In contrast to stepwise procedures that are most commonly used for building interaction models, our formulation is convex, and its solution is completely characterized by a set of optimality conditions. This makes it easier to study as a statistical estimator. We argue that restricting to hierarchical interactions can be advantageous both statistically and computationally. We study its properties, give examples and present an efficient computational algorithm.

This is based on the PhD thesis work of my student Jacob Bien and is also joint with Jonathan Taylor.

Ecological data are a mess: environmental states are difficult to measure, extremely variable, governed by processes at various spatial and temporal scales and describing highly adaptive systems. Ecologists are rarely trained well enough in statistics to even recognise the problems they are facing. At the same time, environmental questions are high on the political agenda and ecologists desire to support policy with their knowledge. A typical example is the attempt to predict the ``whereabouts'' of species under climate change. Large data bases are currently being filled with geographical locations of where species currently are, analysed statistically and the predicted to climate change scenarios. In this talk I will present some statistical challenges that our discipline is facing and the strategies it has developed. Specifically, I will touch on spatial autocorrelation, multicollinearity and typical modelling approaches. I would like to dwell a bit on prediction uncertainty and on the unrelatedness of two fundamental developments in the trade, Bayesian statistics (focussing on embracing detection probabilities) and machine learning (focussing on flexible relationships between predictors and the response). In the end I hope to have given the audience an overview of the many challenges ecological statistics are stubbornly trying to address.