Arne Bathke: How do you build a statistical model for real data – and draw valid conclusions?
Building a good model requires keeping a balance between correctly representing reality and simplifying for practically feasibility.
We will talk about some of the enchanting beauty, as well as the pitfalls and the difficulties of statistical modeling.
Hereby, we plan to use examples where regression and inference methods are applied, and to touch on parametric vs. nonparametric approaches.
Richard Bradley: Earthquakes, Tsunamis and Hurricances: Making decisions with catastrophe models.
Both emergency response and mitigation planning for catastrophes such as earthquakes, Tsunamis and hurricanes draw on scientific models to predict the frequency and severity of these events. In this paper I look at the various uncertainties associated with these predictions – both those quantified by scientific models and those that are not – and consider how planning should be conducted in the light of this uncertainty.
Nancy Cartwright & Sarah Wieten : What’s wrong with pragmatic trials?
‘[P]ragmatic trials … ask “we now know it can work, but how well does it work in real world clinical practice?’* This presentation will argue that in principle nothing is wrong with them – so long as we don’t overbid our cards. The trouble comes when we suppose they have some special advantage when it comes to ‘external validity’. A well-conducted trial can give us an estimate of the average treatment effect in the trial population. It cannot establish that the treatment ‘works’ in general nor that it will work anywhere else (except where just the same causal structure obtains) nor can it tell us what factors help or hinder the treatment. Similarly, a pragmatic trial can establish that the treatment worked in some particular ‘real world’ setting, not that ‘it work[s] in real world clinical practice’. Nor can they tell us what is causally significant about those practices, information without which we cannot generalise.
* Williams, Burden-Teh & Nunn, Journal of Investigative Dermatology (2015) 135, e33
Andrea Fischer: Iconic glacier changes: The role of theory in creating evidence and phenology of climate change
The visible and intuitive evidence of climate change in time series of retreating glaciers made the cryosphere to icons of climate change. Therefore, glaciers are attributed a major role in creating public climate change awareness as premise of a societal transformation. But how much theory is involved in this – at the first glance – purely empirical image analysis, and what are the uncertainties?
On a very basic level, the comparison of time series of glacier changes involves media (photographs), or the theoretical interpretation of the genesis of landscape forms (moraines). For both, perception and anticipation of processes and their cause play a major role.
On a scientific level, we face the problems of scaling and equifinality at the interface between model and ‘real world’. This rises the question if we could expect a match between modelled and measured parameters at all. Changes of mountain glaciers are governed by atmospheric conditions showing high gradients and feedback mechanisms involving thresholds as the snowfall limit or the melting point. The basic problem in tackling the relation between glaciers and climate change involves the scale problem between short time small scale atmospheric processes in the field and modelled spatial and temporal climate means in a model topography. The conditions (elevation, snow cover, temperature…) at the location of the real – world meteorological measurement (observation data) is not necessarily corresponding to the conditions at any model gridpoint, and in most cases statistics is used to scale the model to the observation (or) vice versa. Common assumptions are the validity of the statistical relation also out the range of calibration values, and the representativity of measured values for the total sample. The variability of observables within a gridpoint, where we would expect the gridpoint value to be representative for, in the scientific practice often is fairly unknown. For example, one stake measurement of ablation on the glacier surface is extrapolated to highly variable surroundings. Therefore, a most careful analysis of variability and classical measurement accuracies is at least as important as recording data itself. The high number of involved parameters implicate the problem of equifinality. In the scientific praxis, many parameters act as fit parameters, and validation data in geoscience is limited. Only very few of the world’s glaciers are monitored with high accuracy. The data recorded at this so-called benchmark glaciers is often extrapolated to other glaciers in the region, assuming some representativeness, which can not be rigidly proven. For example, densities of measured glaciers are used to convert volume changes from remote sensing to mass losses. The densities are based on very few observations, and there is no empirical way to validate them. As typically for geoscience, values with higher occurrence rates have a more sound empirical basis than extreme values. The modelling of past glacier changes at the scale of a photographic comparison of a glacier tongue at different times so far is still extremely challenging. Therefore, we can state that there is still a gap of knowledge between the intuitive understanding of climate change looking at glacier photographs and a rigid scientific understanding of all processes behind the curtains. In any case, the perception that there is any change ongoing might be a more important basis for decision making than a rigid scientific understanding of the details.
The uncertainties in glaciology differ from uncertainties in quantum theory in the way that the measurement of the state of the cryosphere is not theoretically, but only practically impossible. Nevertheless, as the practical outcome seems to be similar, framing theoretical uncertainty concepts might be step to build a bridge between model world and observed reality.
Roman Frigg: Expert Judgment for Climate Change Adaptation
Robert Junker: Evidence and uncertainty in predicting plant responses to climate warming
The year 2016 was globally the warmest since records started and thus represents the (provisional) peak of the general trend of global warming, which has severe negative effects on biodiversity and ecosystem processes. In order to understand the causes and predict or even weaken consequences of these impacts, it is mandatory to quantify ecological responses of communities to increasing temperatures. In experimental studies and in approaches using elevational gradients as proxy for climate warming, it has been shown that plant species respond with altered phenotypes and shifted distribution ranges and plant communities with alterations in species composition and diversity. However, mostly it remains unclear whether responses result from phenotypic plasticity or represent adaptations to new environments, which adds substantial uncertainty to the evaluation of the vulnerability of ecosystems. Using the concept fundamental and realized niches, I will discuss the strengths and limits of different approaches to investigate plant responses to environmental changes and of predictions regarding effects on ecosystem processes and services.
Andreas Lang: Climate impact on Earth surface systems – uncertainties unfolding
The Earth’ surface results from the interplay of tectonic and climatic processes, and increasingly also from human action. Understanding how this surface operates and evolves is of direct societal relevance but predicting system behavior has proven notoriously difficult. Quantifying how climate change will trigger a response of the Earth surface is limited due to uncertainties on different levels: analytical uncertainties limit the precision and accuracy of observations and propagation of measurement uncertainty leads to vagueness even in well understood relationships. Often, system response to past climate changes is reconstructed and used as an analog for the future. This approach is limited because complex interplays of multiple processes result in a range of possible outcomes. Evolutionary path dependence of a system’s sensitivity, for example, will trigger different system responses to impacts of similar magnitude occurring at different times.
In the presentation sources of uncertainty will be showcased and approaches highlighted that are commonly used in Earth Sciences to deal with uncertainty.
Charlotte Werndl: Initial Conditions Dependence and Initial Conditions Uncertainty in Climate Science
This talk examines initial conditions dependence and initial conditions uncertainty for climate projections. Climate projections are often described as experiments that do not depend on the initial conditions and that estimate the forced response of the system. Although a prominent claim, it is hardly ever scrutinized, and this talk aims to fill this gap. The conclusion will be that evidence does not support the independence of projections on initial conditions and that thus the forced response of a system is ill-defined. Concerning initial conditions uncertainty, the main contribution will be to identify three kinds of initial conditions uncertainty. The first kind (the one usually discussed) is the uncertainty associated with the spread of the ensemble simulations. The second kind of initial conditions uncertainty arises because the theoretical initial ensemble (relative to which a projection is defined) cannot be used in calculations and has to be approximated by finitely many initial states. The third kind of initial conditions uncertainty arises because of uncertainties in the construction process of the possible initial conditions. To my knowledge, the second and third kinds of uncertainty have hardly been discussed in the philosophy of climate science before.
Lena Zuchowski: Scaling Mountains: Modelling Multi-Scale Geological Processes with Scale-Specific Models