Statistics, causality and dynamic path modelling
Odd O. Aalen
One of the exciting new developments in the field of statistics is a renewal of interest in the causality concept. Causality, however is a complex and difficult concept with an ancient history in philosophy and science. Traditionally, the statistical community has shown a very cautious attitude towards talking about causal connections. Pearl (2000) has strong objections to the extreme caution towards discussing causality that has been common in statistics, stating for instance: "This position of caution and avoidance has paralyzed many fields that look to statistics for guidance, especially economics and social science." There is a contradiction between the cautious attitude of the statistical community and the actual practice of statistics. For instance, any statistician working in medical research, which is arguably the largest single area of application of statistics, will have noticed that the interest in statistics in this field is to a large extent due to the belief that statistics can help in proving causality. In the first issue of New England Journal of Medicine in 2000, there was an editorial about the eleven most important developments in the last millennium (Editors, 2000). Rather surprisingly, one of these was "Application of statistics to medicine". The examples given were typically of a nature where statistics had actually contributed to proving causality. Recently, there have been made very interesting efforts to establish a serious relationship between statistics and causal thinking, in particular the counterfactual and graphical modelling approaches. Cox & Wermuth (2004) also stresses the provisional nature of much causal explanation, realizing for example that there would often be a more detailed explanation on a deeper level which is presently not available. Methods in statistics for analyzing causal connections typically focus on the interplay of variables. Causality is seen as a relation between variables, e.g. being a smoker, having high blood pressure, having heart disease. Often time is hidden in these considerations, but it is quite clear that causality is intimately related to time. It would be natural to view all components in the analysis as stochastic processes and see how they influence one another at a local (that is, infinitesimal time) level.

Model Assessment and Sensitivity Analysis when Data are Incomplete
Geert Molenberghs
Every statistical analysis should ideally start with data exploration, proceed with a model selection strategy, and assess goodness of fit. Graphical tools are an important component of such a strategy. In the context of longitudinal data, oftentimes analyzed using mixed models, these steps are not necessarily straightforward, and the issues are compounded when data are incomplete. Indeed, some familiar results from complete (balanced) data, such as the desire for observed and expected curves to be close to each other, or the well known equivalence between OLS and normal-based regression, do not hold as soon as data are incomplete, unless in very specific cases. Such facts challenge the statistician’s intuition and great care may be needed when exploring, building, and checking a longitudinal or multivariate model for incomplete data. It is also a fertile basis to understand there is a need for sensitivity analysis, some principles of which will be discussed.