We are delighted to announce that the esteemed speakers have graciously accepted our invitation to deliver keynote speeches at the Applied Statistics 2025 conference.
Professor of Statistics
Ghent University, Gent, Belgium
Homepage
Abstract
Causal inference has taken off over the past decades. New and ever more powerful statistical methods have been proposed with well developed properties under well-defined but sometimes complex and typically untestable assumptions [1]. There is now a wealth of promising methods (with software) to choose from. Few are easy to justify and implement in observational settings, however. Fortunately, the new emphasis on the estimand framework helps steer statistical research towards more direct real world relevance and transparency. Challenges arise, for instance, when estimating treatment effects on repeated outcome measures in a mortal population [2]. Policy oriented estimands are then designed to involve continued outcome measures after non-terminal intercurrent events occur. Popular alternative estimands yield different answers for a different purpose [3]. We consider limitations of the (time-varying) survivor average causal effect and of alternative principal stratum like estimands. Given their restrictions and those of other known estimands, such as the hypothetical estimand often presented following mixed models, we present the two-dimensional outcome (survival time and disease outcome while alive) as a basis for causal estimands with prime interpretation and relevance. For estimation, we compare “double inverse probability weighting” with “outcome regression followed by adapted standardization” which handle censoring and death in a distinct manner. As an important secondary problem we discuss the often joint occurrence of missing data and intercurrent events which may lead to non-positivity and/or missingness dependent on underlying values. On the more technical side, we point to naïve combinations of Inverse Probability Weighting and outcome regression which do not generally lead to double robust estimation, especially for right censored survival outcomes [3]. With a case study on the evaluation of Patient Reported Outcomes in late stage oncology, we will highlight sensible estimands, solutions currently available for estimation and suggestions for further research.
Emeritus Professor of Statistics
Eötvös Loránd University, Budapest, Hungary
Homepage
Abstract
In many sequential experiments whether or not a further observation is made depends on the outcome of a previous observation. For example, one may wish to estimate the efficacy of a vaccine based on data collected from an experiment in which those not reacting to the first vaccination receive a booster shot after some time, and to those who still do not show a reaction, a second booster shot is administered. In such a design, the total number of observations is not known in advance and this fact needs to be taken into account in the analysis of the observed data. There are several statistical problems with this structure, including test-retest problems, data about offsprings, and event history (survival) analysis. Such designs have a tree structure and the resulting data, in the discrete case, may be represented in an incomplete contingency table. The talk discusses the correct distributional assumptions, the maximum likelihood estimates, and their properties. The data generating process implies a multiplicative structure of the parameters which generalizes log-linear models. The sampling probabilities are as in multinomial sampling, but the total number of observations is random. This introduces an adjustment factor in the maximum likelihood estimates and also in their covariances. The results are illustrated on several data sets. This is joint work with Anna Klimova.