Ethics and Uncertainty Wendy Parker, Department of Philosophy, Durham University, wendy.parker@durham.ac.uk Nancy Tuana, Department of Philosophy and Rock Ethics Institute, Penn State ntuana@psu.edu, http://scrimhub.org With inputs from Klaus Keller, Lauren Mayer, Rob Lempert, Bryan Cwik, and many others. All errors and opinions are (unless cited) ours. ASP Summer Colloquium August 1, 2014 Contains some privileged materials, please do not cite or distribute 1 Ethical Analysis How should we act? Responsible selection of research topics Science to Society impacts Epistemic Analysis What can we know? Coupled Ethical-Epistemic Analysis How do we support responsible action with what we know? Uncertainty quantification Model selection Values that inform epistemic decisions Epistemic decisions that have ethical import Ethical values Justice Sustainability Utility Human Security Epistemic values Robustness of evidence Predictive power Convergence of evidence Completeness 2 IPCC Treatment of Uncertainty: The Role of Values • IPCC Guidance Note – The characterizations of uncertainties is a deliberative process • “However, the extent that this procedure can be considered reliable…is contingent on agreements based on value judgements, for instance on spatiotemporal resolution, parameterization of models, or significance level for accuracy in empirical studies.” Adler and Hadorn 2014 3 Example of coupled ethical-epistemic questions: (1) What do we know about climate sensitivity? (2) Which decision-criterion is relevant? (3) How do we estimate climate sensitivity? 1 − Cumulative Frequency [dimensionless] 10 0 Olson et al. (2012) Libardoni and Forest (2013) Aldrin et al. (2012) Urban et. al. (2014) CMIP High Resolution Models Missing Tails 10-1 1 in 50 ● 10-2 10-3 10 1 in 10,000 -4 0 2 4 6 Climate Sensitivity [K] 8 10 Keller et al (in prep) 4 Epistemic/Evidentiary Epistemic strengths and weaknesses Positive and negative impacts of ethical features on future gathering of evidence Model / Variable Choice Ethically desirable or undesirable consequences of epistemic features Ethically desirable and undesirable features Ethical 5 Valles et al. forthcoming Wendy: Coupled ethicalepistemic choices in modeling Nancy: Coupled ethicalepistemic issues and decision support Coupled epistemic-ethical choices in modeling • choices in model construction & evaluation – which quantities to prioritize for accurate simulation – choice of uncertain parameter value(s) / parameterizations • how to estimate and communicate uncertainty • which methods/studies to pursue Coupled epistemic-ethical choices in modeling • choices in model construction & evaluation – which quantities to prioritize for accurate simulation – choice of uncertain parameter value(s) / parameterizations • how to estimate and communicate uncertainty • which methods/studies to pursue Ethical values should never influence how we build and evaluate mathematical models in the study of climate change. (e.g. which variables we focus on, which values we assign to parameters) 1. True 2. False 3. I’m really not sure 58% 32% 11% 1. 2. 3. It is sometimes appropriate for ethical values to influence choices in model construction and evaluation. What we care about (e.g. avoiding harm to humans) What we can hope to provide information about, given today’s science choice of priority variables/outcomes sea level rise loss of life in floods economic losses model improvement activities, choice of performance metrics It is sometimes appropriate for ethical values to influence choices in model construction and evaluation. Underestimating X would be worse than overestimating it. The best value for parameter μ could plausibly be anywhere lower μ lower X in this range… (The lower the value of μ, the greater the risk that the model will underestimate variable/outcome X.) sea level rise choice of numerical value for uncertain parameter μ The flip side is that even when ethical values are not directly influencing choices in model development, those choices can have consequences with ethical import! loss of life in floods economic losses larger or smaller risk of overestimating/underestimating X Coupled epistemic-ethical choices in modeling • choices in model construction & evaluation – which quantities to prioritize for accurate simulation – choice of uncertain parameter value(s) / parameterizations • how to estimate and communicate uncertainty • which methods/studies to pursue A probability density function (pdf) is always an appropriate way to represent uncertainty. 1. True 2. False 3. I have no idea. 95% 5% 1. 0% 2. 3. Sometimes precise probabilities aren’t appropriate epistemically or ethically • Sometimes uncertainty is deeper than a pdf would imply. The science is insufficient to assign precise probabilities to outcomes. – The probability of more than two Cat 5 hurricanes making landfall in the U.S during the 2070s under RCP 6.0. • In these cases, representing uncertainty with precise probabilities (or a full pdf) would be inaccurate, implying that we know more than we really do. • It would also be misleading. If decisions with real consequences will be informed by these uncertainty estimates, offering precise probabilities may be inappropriate from an ethical point of view; we can expect bad consequences. Being epistemically responsible “…striving for epistemic excellence captures well what is fundamental to…epistemic responsibility. It is for one to do the best she can with what she has available to her, epistemically speaking. It is to be circumspect in seeking truth and avoiding error.” (Corlett 2008, p.180) Epistemic responsibility & uncertainty Estimating uncertainty about future climate change in an epistemically responsible way would seem to require: – striving to take account of all available evidence – striving to take account of all recognized sources of uncertainty – trying to factor in the possibility of “unknown unknowns” – avoiding a one-size-fits-all mentality about the nature of one’s uncertainty (e.g. probability-like vs. deeper) – offering a depiction of uncertainty that does not misrepresent one’s epistemic state (indicating that one knows more / less / different than one in fact does) -- ownership “likely range” = we are roughly 66-100% confident that the change would be in this range How to interpret these results? “UKCP09 offers projections of the future climate that [are] based on the current understanding of the climate system – there may be scientific unknowns that would affect the information provided. Hence UKCP09 should be seen as providing possible projections rather than absolute predictions or forecasts of future climate.” (http://ukclimateprojections.defra.gov.uk/content/view/633/531/) “Probabilistic projections, although they are designed to quantify uncertainty, … are themselves uncertain. …However, as a general guideline we suggest that users should be able to use the distribution from the 10 to the 90% probability levels, but not outside this range...” (Murphy et al. 2009, Section 4.1) Coupled epistemic-ethical choices in modeling • choices in model construction/selection & evaluation – which quantities to prioritize for accurate simulation – choice of uncertain parameter value(s) / parameterizations • how to estimate and communicate uncertainty • which methods/studies to pursue If a modeling study is on scientifically shaky ground, and we know that decision makers will be tempted to use the quantitative results in decisions with real consequences, should we not do the study? 70% 1. I think we probably should not do the study. 2. It depends… 3. As a scientist, it’s not my job to worry about that. 4. I have no idea. 15% 10% 1. 5% 2. 3. 4. Attribution of extreme events Douglas on the moral responsibilities of scientists “With full awareness of science's efficacy and power, scientists must think carefully about the possible impacts and potential implications of their work. Although there is no qualitative difference between this responsibility and the responsibility of automobile drivers to proceed with due care and caution, the quantitative burden is much greater. The ability to do harm (and good) is much greater for a scientist, and the terrain almost always unfamiliar. The level of reflection such responsibility requires may slow down science, but such is the price we all pay for responsible behavior. The driver may need to take more time getting to his destination; similarly the scientist may need to take more time in developing her research and determining how to present the results.” (Douglas 2003, p.66) Coupled ethical-epistemic issues and decision support 1. Why are coupled ethical-epistemic issues important to decision support? 2. What are ethically and epistemically responsible approaches to decision support? 3. A challenge… 26 Why are coupled ethical-epistemic issues important to decision support? 27 Overconfidence can result in downwards biased risk estimates. (Slide from K Keller) Legally acceptable flooding probabilities often range between 1/100 and 1/10 000. Revised estimate with high reliability Overconfident projections can lead to downwards biased risk estimates of tail area events and downwards biased (constrained) optimal risk management strategies. This example is about adaptation. Does this also apply to mitigation? 28 Do the analysts and the users share the same understanding? • How are instances of overconfidence and/or uncertainties communicated? “The gaps between the authors’ intentions and the readers’ understanding of the probabilistic communications are large andBudescu systematic et al 2012” 29 Rosenzweig et al 2011 Are Value Choices Transparent? Risks from climate change, by reason for concern—2001 compared with updated data. Smith J B et al. PNAS 2009;106:4133-4137 ©2009 by National Academy of Sciences 30 Aggregating impacts risks obscuring distributive justice concerns Smith J B et al. PNAS 2009;106:4133-4137 ©2009 by National Academy of Sciences 31 How do we balance epistemic robustness with ethical salience in modeling? 32 Where is the problem located in the joint knowledge / values space? Keller et al (in prep.) 33 What are ethically and epistemically responsible approaches to decision support? 34 Whose Values? The other uncertainty: values 35 Figure from K. Keller What is the underlying value? • Equity? – Equity based on what measure? • Attention to the least well off? • Sustainability? – Sustainability based on what measure? • ……… 36 Values Uncertainty • Experts know what’s best… • Stakeholders know what’s best… Whose values? What knowledge do we need given the values? 37 “Deliberation with Analysis” Offers an Effective Decision Support Process for Decisionmaking Under Uncertainty Deliberate: • Participants to decision define objections, options, and other parameters Analysis: • Participants work with experts to generate and interpret decisionrelevant information How do we include values in this process? • Just ask them? • How self-transparent are values? • What if our stakeholder group isn’t representative? • What if it is, but it still misses important values? What are the best methods for collecting inputs for Ethically Informed Robust Decision Making? • Can we ensure values comprehensiveness and objectivity? – Rawl’s Veil of Ignorance insure impartiality? • How do we include “absent” stakeholders? • How do we adjust for value uncertainties? • What if the analysts values are different than the stakeholder values? 39 A simplified flow diagram of ethically informed RDM Values Informed Mental Models Veil of Ignorance deliberations Ethically informed robust strategies 40 But what if the analysts values inform the models/scenarios? Values Informed Mental Models Veil of Ignorance deliberations Ethically informed robust strategies 41 GR SCRIM Decision analysts talk statistically significantly more often about consequentialist values than decisionmakers • SCRiM Project Leaders – prevalence of nonconsequentialist values + 20 30 40 50 60 70 80 Frequency [%] – prevalence of consequentialist values 0.06 0.05 0.04 Density • West Michigan Climate Resiliency Consortium + 95% CI 0.03 0.02 0.01 0 10 20 30 Difference in mean w elfare response frequency (SCRiM − GR) [%] 40 42 H1: Decision analysts are mostly consequentialists; decisionmakers are mostly not. H2: This difference is decision-relevant 43 Challenge • When you do your RDM exercise, consider how many coupled ethical-epistemic issues are involved. Potential discussion questions 1. 2. 3. 4. 5. 6. 7. 8. What do we need to know to contribute to responsible decision-making? What are some epistemic or ethical pitfalls to avoid if we are working in ‘climate services’? How can we identify the most decision-relevant uncertainties? What is a useful characterization of decision-relevant uncertainties? How do we deal with the incompleteness of our models and their potential overconfidence? How do we find strategies that perform reasonably well in the face of deep and dynamic uncertainties and across a wide range of objectives / ethical frameworks? How can we measure and improve the usefulness of decision-support tools / models? What (if any) advice do we give to (i) people who need to pour concrete for coastal infrastructures now and (ii) a grad student who wants to study the most important part of the problem? 45 RESERVE MATERIALS 46 References • • • • • Adler, C. E. and G. Hirsch Hadorn. (2014) The IPCC and treatment of uncertainties: topics and sources of dissensus WIREs Clim Change 2014. doi: 10.1002/wcc.297. Budescu, David V., Han-Hui Por & Stephen B. Broomell (2012) Effective communication of uncertainty in the IPCC reports Climatic Change 113:181–200. Mastrandrea MC, Field CB, Stocker TF, Edenhofer O, Ebi KL, Frane DJ, Held H, Kriegler E, Mach KJ, Matschoss PR, et al. Guidance Note for Lead Authors of the IPCC Fifth Assessment Report on Consistent Treatment of Uncertainties. Geneva, Switzerland: Intergovernmental Panel on Climate Change (IPCC); 2010. Available at: http://www.ipcc.ch. Rosenzweig, C., W. D., Solecki, R. Blake, M. Bowman, C. Faris, V. Gornitz, R. Horton, K. Jacob, A. LeBlanc, R. Leichenko. M. Linkin, D. Major, M. O’Grady, L. Patrick, E. Sussman, G. Yohe, R. Zimmerman. (2011). Developing coastal adaptation to climate change in the New York City infrastructure-shed: process, approach, tools, and strategies. Climatic Change (2011) 106:93– 127. Smith, Joel B., Stephen H. Schneider, Michael Oppenheimer, Gary W. Yohe, William Hare, Michael D. Mastrandrea, Anand Patwardhan, et al. 2009. Assessing dangerous climate change through an update of the Intergovernmental Panel on Climate Change (IPCC) ‘reasons for concern.’ Proceedings of the National Academy of Sciences of the United States of America 106(11) (March 17): 4133-37. 47 Risks from climate change, by reason for concern—AR5 48