You are on page 1of 7

European Journal of Operational Research 226 (2013) 132138

Contents lists available at SciVerse ScienceDirect

European Journal of Operational Research


journal homepage: www.elsevier.com/locate/ejor

Interfaces with Other Disciplines

Moral mid-level principles in modeling


Sven Diekmann
Eindhoven University of Technology, Department of Philosophy and Ethics, Fac. of Innovation Sciences, IPO 1.14, P.O. Box 513, 5600 MB Eindhoven, The Netherlands

a r t i c l e

i n f o

a b s t r a c t
Modelers, especially in operational research, are becoming increasingly aware that their role in decisionmaking raises moral problems. This paper discusses two questions: How do moral issues in modeling arise? How can these moral issues be addressed? I propose a framework that (1) provides tools for discovering moral issues raised by models, and (2) provides practical guidance for solving moral problems in modeling. As regards (1), I discuss three moral perspectives on modeling: a perspective that focuses on the benecial or harmful consequences of using a model; a perspective that focuses on the intentions of using a model; and a perspective that focuses on whether a model promotes virtuous behavior. In order to achieve practical action guidance, (2), four moral mid-level principles are introduced: (i) The principle of transparency expresses the obligations to explain the structure, assumptions and further properties of the model; (ii) the principle of integrity demands for the application of professional standards; (iii) the principle of comprehensiveness stresses that all moral concerns should be acknowledged thoroughly; and nally, (iv) the principle of efcacy states that models should evaluate moral issues explicitly. 2012 Elsevier B.V. All rights reserved.

Article history: Received 3 October 2011 Accepted 18 September 2012 Available online 13 October 2012 Keywords: Decision analysis Decision-making Ethical guidance Ethics Moral mid-level principles Modeling

1. Introduction Models, especially operational research models, often play an important role in decision-making. Through this role, the construction and use of models can raise explicit or implicit moral issues. Operational researchers are becoming increasingly aware of this problem.1 Several authors have raised a number of moral issues that arise in modeling. Wenstp and Magnus (2001) describe a model for decision-support in the Norwegian HIV policy. Rabins and Harris (1997) discuss modeling failures in engineering and conclude that every modeling process should respect moral standards such as the Golden Rule, informed consent, and utility. Kleijnen (2011) approaches the topic by raising several morally problematic modeling cases. However, almost all existing approaches are explorative. Hence, aside from recent approaches that use discourse ethics (Le Menestrel and Van Wassenhove, 2004; White and Bourne, 2007; Kunsch et al., 2009; Drake et al., 2011; Mingers, 2011; De Brucker et al., forthcoming), there are no systematic analyses of how modelers should tackle moral problems. In this article, I attempt to provide a basis for more thorough ethical discussions. I develop a framework for discovering and dealing with practical moral problems in modeling, addressing two fundamental questions: (1) How and when do

Tel.: +31 40 247 4987; fax: +31 40 244 4602.


E-mail address: s.diekmann@tue.nl Two milestones of this debate are Wallace (1994) and the special issue Ethical Issues in OR/MS in Omega, see Le Menestrel and Van Wassenhove (2009). Wenstp (2010) provides comprehensive overview of the eld of operations research and ethics.
1

moral issues arise in modeling? (2) Which moral norms should guide modelers in resolving these moral issues? The current literature neglects the rst question, leaving it to the modelers intuition to recognize moral problems. Some authors even seem to assume that moral problems are external to the modeling profession. For example, Walker (2009) states that the best way to deal with moral problems is to follow the best practices of ones profession. In the following, I show that this assumption is incorrect. By considering several moral perspectives, one can see that modeling practice itself raises moral issues. Indeed, moral issues in modeling arise under a variety of circumstances for diverse stakeholders and interest groups. In order to cope with this variety, we need a broad ethical approach that can accommodate and address most concerns. (Van de Poel, 2001, p. 430) For recognizing the full breadth of moral issues, I employ the three most prominent moral perspectives: consequentialism, deontic ethics, and virtue ethics. These three perspectives lead to specic questions regarding morality in modeling: (1) Can the consequences of using a model be morally good or bad? (2) Are the intentions of employing the model good or bad? (3) Does the model discuss matters related to the human way of living? Following a pluralistic approach, this article proposes moral mid-level principles to provide systematic ethical guidance. The principles are called mid-level principles, because they bridge between fundamental ethical principles and specic moral problems. They are denoted as principles, because one should abide by them in general. However, their mid-level status allows for some exibility; different principles may need to be balanced out against each other if circumstances lead to their mutual conict. The rst

0377-2217/$ - see front matter 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.ejor.2012.09.027

S. Diekmann / European Journal of Operational Research 226 (2013) 132138

133

mid-level principle is the principle of transparency. This principle refers to the obligations to explain, describe, and document the structure, assumptions and further relevant properties of the model. The second is the principle of integrity. This mid-level principle expresses the responsibility of modelers to apply professional and scientic standards in their work. The third is the principle of comprehensiveness. It stresses that modelers should acknowledge the moral concerns of all stakeholders thoroughly and extensively. Fourth and nally is the principle of efcacy. It states that models should evaluate all moral issues with respect to their consequences and implications for stakeholders. When arguing for these moral mid-level principles, I refer to moral theory, on the one hand, and practical experiences and solutions of professionals, on the other. Moral mid-level principles thereby support a more sophisticated and theory-based ethical debate in operational research without losing the connection to practice. The mid-level principles can thus also be used to structure an ethical discussion about modeling in operational research. In Sections 5 and 6, I explain in detail how moral mid-level principles are employed.

2. Perspectives on morality in modeling Moral issues in modeling arise at a large scale (e.g. governmental decisions), as well as in more localized situations (e.g. the design of a particular artifact). Moreover, moral problems may affect only a few individuals or a variety of stakeholders. To identify moral issues in the midst of such heterogeneity, I take a pluralistic perspective on moral theories: I assume that all commonly acknowledged moral theories are able to point out (different) morally problematic issues. Since consequentialism, deontic ethics, and virtue ethics subsume most moral theories, I shall focus on these three. As mentioned in the introduction, my analysis focuses on practical moral problems, as they arise in the work of operational researchers. First consider the consequentialist approach to moral theory. The consequentialist credo is that an action should be performed, if and only if no other action leads to more desirable overall consequences. Thus, consequentialists focus on the consequences of using a model. Take crash simulations of cars as an example. These simulations involve models that help to decrease human harm in an efcient and cheap way. Since this provides a benet with only minimal cost, the use of these models has good overall consequences. That said, use of models can also have bad consequences. This was shown by an unfortunate event during the rst Gulf War, which resulted in 28 deaths. On 25th February 1991, a Patriot defense system failed to intercept a Scud rocket, due to misusage of the system and its underlying model (Blair et al., 1992). Whether such usage is an intended use or a misuse is less relevant from a consequentialist perspective. Neither the model itself, nor the intentions with which it was used are considered. Indeed, in consequentialism the model itself can be treated as a black box: use of the very same model may be morally right in some cases (where the consequences are most desirable) and morally wrong in other cases (where the consequences are undesirable). The second perspective is deontic ethics, which focuses on the action of use as such. I use Kantian ethics as a paradigmatic example of a deontic moral theory. Central to the Kantian theory is the categorical imperative of which Kant offers two main formulations2: (1) [A]ct only in accordance with that maxim through which you can at the same time will that it become a universal law (Kant, 1785/1998, p. 31); and (2) act that you use humanity, whether in your own person or in the person of any other, always at the same
2

Kant claims that the two formulations are equivalent.

time as an end, never merely as a means (Kant, 1785/1998, p. 38). Formulation (1) expresses a principle of universality: Good actions should always be performed regardless of the situation or the agent. For example, deceiving is morally wrong, because deceit on a regular basis (as a universal law) would interfere with the core function of communication of transmitting reliable information. Formulation (2) tells us that humans have a dignity, which we should respect in all our actions. For instance, employers should never see their employees as pure resources, but always acknowledge that they are also persons who have further, e.g. social, needs. In Kantian ethics, the intentions behind an action determine its moral evaluation. When applied to modeling, both formulations restrict the desirable intentions of employing or developing models. For example, models should never be used for deceit. The reason for employing models is that models are elucidating instruments. If models were intended to deceive in general, they would lose their elucidating function and there would be no reason for using them anymore. Moreover, models should never support exploiting people, because this would ignore that people should (also) be treated as ends. Note that maximized benecial consequences do not affect the Kantian moral judgment, because the elucidating nature would still be undermined or the people would still be exploited. It seems likely that most models are developed with good intentions. However, there are also counter examples. Fienberg (1994) discusses the problem of undercounting minorities in the 1990s US decennial census. In order to solve this problem, new statistical procedures were introduced. As a supplement to the statistical model of the US Bureau of the Census, statistical tools were developed by an independent group of mathematicians. By rening the statistical methods, the additional tools reduced undercounting. Fienberg claims that unofcial political interest of a few powerful white Americans caused the additional tools to be neglected in the analysis. If such a claim is true, the intention of using the model without additional tools was personal benet. Personal benet is inconsistent with the scientic, impartial idea behind modeling. Thus, this intention cannot be generalized and is morally wrong. Virtue ethicists focus on the good life and the pursuit of it. Virtuous behavior, managed by practical reasoning, is seen as the means to reach such good life. Aristotle sees virtues as those character traits that lead to appropriate behavior under various circumstances. Appropriate behavior in the Aristotelian sense lies always between extremes (not necessarily in the middle): A courageous man is neither cowardly nor brash; a virtuously generous person is neither stingy nor spendthrift (Aristotle, 350BC/1980 [1107b1 1109b26]). Specication of the appropriate middle ground is dependent on the situation at hand and must be judged by practical reasoning (Aristotle, 350BC/1980 [1138b201139a2]). Practical reasoning is not just a matter of intuition; rather it involves moral decisions and judgment informed by theoretical knowledge (Aristotle, 350BC/1980 [1144b]). Here models can support virtuous behavior by informing practical reasoning. For example, many people see it as a part of good life to consume meat on a daily basis. However, models which forecast the environmental impact of such lifestyle predict substantial negative effects if all humans will consume meat in such amounts (Stehfest et al., 2009). Thus, meat consumption on a non-daily basis is more virtuous, and models help us to understand this. The climate debate is a further example. Almost all our knowledge about desirable sustainable behavior is based on models. Correct practical reasoning is a necessary part of virtuous behavior, because acting appropriately by chance cannot be called virtuous (Aristotle, 350BC/1980 [1144b11145a6]). For this reason, any virtuous person will seek to improve his or her practical reasoning. Following this line of thought, a morally good model sheds light on the desirable and undesirable implications of human

134

S. Diekmann / European Journal of Operational Research 226 (2013) 132138

Table 1 The table presents the analysis of the four cases in summary. Aspects that raise moral concerns are bold. No. 1 2 3 4 Case Patriot failure 1990s US decennial census Forecasting the effects of human meat consumption on climate change The harmonic oscillator Consequences of use Failure to prevent death of 28 soldiers Probably disadvantageous for minorities Description of the expected climate effects of different diets Understanding of oscillations Intentions behind use Prediction of missile trajectories for intercepting them Partial benet of an interest group Explore the consequences of diet changes Description and understanding Questions elucidated Trajectories of Scud rockets Number and type of US inhabitants Effects of diets on climate change Behavior of linearly swinging masses

behavior. Hence, whenever a model touches on norms, actual behavior, and its implications, it raises moral issues from a virtue ethics perspective. In summary, the three moral perspectives reveal moral issues that stand in fundamentally different relations to models and modeling. The consequentialist perspective focuses on the use of a model and its consequences. The deontic perspective focuses on modelers and model-users and their respective intentions in developing and using a model. Finally, the virtue ethics perspective focuses on whether a model promotes virtuous behavior.

3. Modeling and examples of moral issues Philosophers of science have made a number of attempts to clarify what scientic models are, how they relate to scientic theories and to the phenomena being studied. For analysis of application-oriented sciences, such as operational research, the account of modeling of Morgan and Morrison (1999) provides a suitable description of a modeling process. Their account treats phenomena as similarly important as scientic theory. In short they claim that models are autonomous instruments that mediate between theory and phenomena, such that most models embody parts of the theory as well as parts of the phenomena. Even though the usage of, the intentions of using, and the question discussed by a model are eventually up to the model-user, modelers can inuence them in the construction of models. First, Morgan and Morrison formulate different kinds of representation3: [A] representation is seen as a kind of rendering a partial representation that either abstracts from, or translates into another form, the real nature of the system or theory, or one that is capable of embodying only a portion of a system. (Morgan and Morrison, 1999, p. 27, my emphasis) Second, models can fulll several functions, viz. they can be used in theory construction and exploration; in measurement for structuring data or as measurement devices; and they can be used in design and intervention. (Morgan and Morrison, 1999, pp. 1825) These different kinds of representation and functions represent degrees of freedom in constructing models for particular uses, with specic intentions, or for varying questions, without harming the scientic character of models. Having introduced these moral perspectives, I now turn to showing how these can be used to systematically analyze real cases. Four examples are considered relative to the core question of each mid-level principal, namely: (1) Are the consequences of using the model morally good or bad? (2) Are the intentions behind employing the model right or wrong? (3) Does the model elucidate matters related to the good way of living? The analysis of the cases is summarized in Table 1.
Some modelers claim that their models, in particular agent-based models, do not represent anything. There are also other kinds of models, such as the Ford Model T and fashion models. Both sorts of models can be neglected, since relevance for decisions, i.e. moral relevance, presumes some kind of representation.
3

Example 1. The Patriot failure during the rst Gulf War is an example of the use of a model with morally bad consequences, namely 28 deaths. The Patriot defense system was employed to intercept Iraqi Scud missiles. In order to do this, the patriot software modeled the trajectory of the missiles by using velocity and time data. Time was represented by real values with 24 bits. Time was modeled with increasing numbers, starting from 0 when the software booted. Problematically, the higher the values of time the more bits are required to represent the digits before the comma of the time values. Consequently fewer bits are available for representing decimal places after the comma. Thus, longer periods of use decrease the time decimal places represented and thereby the precision of the model. This poses no problem for the originally intended short-term use of Patriots. However, during the rst gulf war, patriots had often been running for several consecutive days, while they became useless after 20 h.4 (Blair et al., 1992) From a consequentialist view, it was morally wrong to use this model for several consecutive days, because the Patriot failed to intercept a Scud missile and 28 people died in consequence. Instead, it would have been morally right to use the Patriot and their model over short time periods or to employ a model that allows for long consecutive use. In such case, the soldiers lives could have been saved.5 Example 2. The case of the 1990s decennial census (Fienberg, 1994), as discussed in Section 2, points on morally problematic intentions behind model deployment. Statistical tools that would have reduced the undercounting of socially disadvantaged groups were rejected on the basis of the economic interests of a powerful minority. Since the intention of benetting personal interests contradicts the accuracy that census models aim for, Fienberg describes an intention which one cannot want to be the general case. Following Kantian ethics, this case of model employment for personal benets is a moral wrong. Example 3. Stehfest et al. (2009) employ the Integral Model to Assess the Global Environment (IMAGE) to forecast the impacts of human diets on climate change. Even though their study provides only scientic claims, it touches the moral question of what virtuous eating habits are. They compare ve diet scenarios to explore how changes in the use of animal products affect landuse and the emission of greenhouse gases. The reference scenario is an extrapolation of the current income-driven increase in per capita meat consumption without any counter-acting policies. They predict that even a change to the healthiest diet of meat consumption has a signicant positive impact on land-use and greenhouse gases emissions.6
4 The discussed patriot system worked for 100 consecutive hours at the moment of failure. 5 Its a tragic fact that an improved model arrived at the failing patriot system the next day (Blair et al., 1992). 6 For the healthiest diet Stehfest et al. refer to Willet (2001).

S. Diekmann / European Journal of Operational Research 226 (2013) 132138

135

Example 4. The model of the linear oscillator is a central model in physics. It describes the behavior of a mass point on a swing that is in linear movement. I include this example as a representative of models that are generally not perceived as raising moral problems. The linear oscillator is usually developed and used with epistemic intentions; it discusses no human habits, but rather a physical object (a swinging mass); and although a use with morally wrong consequences is surely imaginable, the vast majority of its uses in physics do not directly touch on benets or harms.

4. Moral mid-level principles in modeling In this section, I propose moral mid-level principles as a pluralistic method for dealing with diverse moral problems in modeling. The notion of moral mid-level principles goes back to Beauchamp and Childress (2008), who develop a moral framework for the biomedical context, consisting of four principles: Respect for Autonomy, Benecence, Nonmalecence and Justice. Respect for Autonomy demands that professionals acknowledge the moral authority and judgment of humans under treatment; Benecence demands the promotion of wellbeing; Nonmalecence demands the avoidance of harm; and Justice demands that all humans are treated fairly with impartial concern. Since these guidelines cease being principles outside biomedical contexts (e.g. Nonmalecence cannot be applied to modeling directly) they are not general principles, but mid-level principles. Similarly, moral mid-level principles in modeling are designed to solve moral problems in modeling contexts. To enable this, moral mid-level principles differ from moral theories. Moral theories claim to be generally applicable and ignore any specic norms or codes of conduct of a professional eld. Mid-level principles however relate to such eld-specic norms and codes as well as they are connected to moral theories. 7 In this section, I relate each principle to moral theories and to established norms. It is difcult, if not impossible, to derive the uniquely best action for a given situation from moral mid-level principles. Rather the moral mid-level principles should demarcate a realm of morally appropriate actions. It might be that there are more mid-level principles than the four presented here. Such additional mid-levels principles would add precision to my approach. Moreover, I do not claim that the mid-level principles are unique. Rather, each principle subsumes a variety of moral considerations. Which principle aggregates which considerations or the number of principles is not relevant for the overall coverage of moral considerations. Finally, many moral mid-level principles are partly epistemically justied. Despite this, the mid-level principles are still moral principles, because epistemic motives such as accuracy, simplicity, or theoretical consistency do not sufce to support the mid-level principles presented here. 4.1. Transparency The principle of transparency refers to the open communication of the specic model properties. It is grounded in the fact that no model result can be understood without some clarication of how the model and its result relate to the world. For example, there are many energy policy models that explore the question of to what extent the use of different energy sources such as coal, gas, nuclear, and renewable energies should be prioritized. Several
7 Mid-level principles stand in between unconditional principles and contextspecic obligations. Unconditional principles, e.g. impartiality, are supposed to hold under all circumstances in any case. Obligations (e.g. One should be quiet during a conference talk.) hold only in particular situations under specic circumstances. The obligations change when the situation or the circumstances change (e.g. One should be talkative and raise questions after a conference talk.).

studies investigate this issue, employing different models with varying assumptions, based on different data. Sundqvist et al. (2004) compare models about estimations of external costs for a variety of energy sources. The assumptions of the models used differ in (1) focus on energy technologies, (2) local characteristics of the researched sites, (3) scope of researched external cost, (4) implicit trade-offs among environmental effects, and (5) different methods to deal with complexity or uncertainty. (Sundqvist et al., 2004, pp. 234235) Nevertheless, all models are only valid relative to these characteristics. For instance, many models are developed for Europe. Employing them for tropical regions certainly leads to inappropriate results. (Sundqvist et al., 2004, p. 236) Fullling transparency means stating assumptions clearly, clarifying the scope of problems to which it can be applied, and being clear as to the restrictions that the model has; the goal of transparency is to make the design of a model understandable to modelusers.8 In fact, transparency prevents two kinds of problems. First, an epistemic problem: inappropriate models can generate pseudoknowledge. Second, a moral problem: misunderstood models can intentionally be used to deceive. 4.2. Integrity Having integrity means that a modeler applies professional and scientic standards, even in situations where it is compelling to follow personal interests or to neglect rules. The case described in Fienberg (1994) (example 2), about the 1990s US decennial census is an example of lacking integrity. Here a political decision favored a less accurate model in order to give benets to a particular group of people. Professional standards, e.g. objectivity, or scientic standards, e.g. maximized accuracy, would have led to the usage of the right model. Integrity implies that a modeler regards professional and scientic standards as role-based virtues of modelers. Furthermore, regarding professional standards as role-virtues strengthens morality in conict situations. First, standards give orientation about what is right or wrong. Second, standards provide a well-accepted argument for maintaining minimal moral standards; namely the role-based obligations of professionals. Also deontic ethics support integrity, because it emphasizes the value of following generalizable standards (maxims). Finally, integrity has also a legitimizing function. On the one hand, professional and scientic standards guarantee scientic rigor and scientically justied results. On the other hand, professional codes formulate a minimum of moral behavior. This minimum does not imply morally right actions, but applying ethical codes prevents morally reckless modeling. 4.3. Comprehensiveness The principle of comprehensiveness summarizes the obligations of a full-edged consideration of all moral issues that are connected to the decision alternatives under scrutiny. Even though comprehensiveness is partly an epistemic concern it has a moral dimension, as the study on meat consumption (Stehfest et al., 2009) shows. Environmental and health issues are connected to meat consumption. Many people regard animal products as a necessary part of a healthy diet. Problematically, the production of meat is environmentally invasive. Comparing different degrees of meat consumption, Stehfest et al. support the claim that there are healthy diets with acceptable environmental impact. Hence,
8 I do not advocate that model-users should have the same knowledge about the models as the modelers. Rather, transparency wants the modeler to communicate actively, and to be open to doubts and questions of model-users, such that the modelusers do not treat the models as black-boxes (Ross and Harris, 1994).

136

S. Diekmann / European Journal of Operational Research 226 (2013) 132138

the comprehensive design of their discussion resolves a virtue conict between health and sustainability. As the example suggests, comprehensiveness is an issue of impartiality and practical reason. It aims at broad coverage of moral concerns. If only a few concerns are incorporated, one-sided decisions are more likely. Accordingly, consequentialism can be taken as a motivation for comprehensiveness. From a virtue ethical perspective comprehensiveness is praiseworthy, because of its support of practical reason. If we want to understand what virtuous behavior is, we need to know what kind of problems can arise from current and potential behavior. Even more strongly, moral theories that see the acknowledgment of all stakeholders as a moral ideal such as discourse ethics in Drake et al. (2011) or Kunsch et al. (2009) go as far as making comprehensiveness a necessary condition. 4.4. Efcacy The principle of efcacy denotes the detailed explication of consequences or effects concerning each moral issue. Taebi and Kadak (2010, p. 1358) give a good example of modeling that respects efcacy. They compare different kinds of nuclear energy production by providing a score card, summarizing issues of environmental friendliness, public safety, security concerns, resource durability, economic viability, and technological applicability. Hence, Taebi and Kadak provide a list of consequence details for each kind of nuclear energy plant such that decision-makers have a concise overview of advantages and disadvantages. The arguments for efcacy relate to the impact of moral issues for assessing their actual importance. If a moral concern is unlikely to have noteworthy consequences, it might be reasonable to cope with other concerns rst. Additionally, some moral concerns exist in theory, but do not become problematic in practice. The consequentialist argument for efcacy is therefore straightforward: Following efcacy means to assess moral concerns for their benets and disadvantages, to estimate the consequences of an action, and to compare the moral desirability of actions. Finally, efcacy is important for practical reason in virtue ethics, since it assesses the effects of human behavior in detail. 4.5. Mid-level principles and established norms Transparency as a professional norm is widely accepted. For example, the contributors to the book Ethics in Modeling agree on some aspects of ethical conduct, such as . . . to make clear to the model-user what the model can do and what its limitations are (Wallace, 1994, p. 8). Ross and Harris (1994) see the pro-active education of the model-user as a major obligation: A model builder has a professional responsibility to explain not only the strengths of a particular model but its inadequacies . . . (Ross and Harris, 1994, p. 161). Drake et al. (2011, p. 10) claim that . . . the way in which models are created, what problems they address, and how they are implemented should be made available to stakeholders of the research activity. Valuing reproducibility, Kijowski et al. state that modelers need to describe the underlying theories, assumptions, and limitations of their models (Kijowski et al., forthcoming, p. 9). Although integrity is not mentioned explicitly in the literature, a number of claims can be summarized under it. First of all, ethical codes of different organizations formulate professional standards. Gass (2009) gives a good overview: Professional standards include norms of how to perform ones work, as e.g. Military OR professionals must strive to be . . . [t]ruthful, complete and accurate in what they say and write. Professional standards also include norms about social obligations, e.g. ACM members will contribute to society and human well-being. Finally, they include

responsibilities of modelers, e.g. IEEE members agree to accept responsibility in making engineering decisions consistent with safety, health, and welfare of the public. (Gass, 2009, pp. 1046 1047) An attempt to install such a code among operational researchers has been made by Brans (2002), called the Oath of Prometheus. The Oath is addressed to different elds of operational research and includes professional, scientic, and moral norms. In addition to Gass and Brans, Walker (1994, 2009) also promotes professional standards. He states that modelers will be acting in an ethical manner if they apply the general accepted best practice of their profession (Walker, 2009, p. 1051). Comprehensiveness is also an established norm in modeling. Kunsch et al. (2009) see a moral need for thorough information assessment: The main issues are not primarily the technicalities or heuristics. The denition of the human context, the identication of stakeholders and their moral values, the systematic analysis of all connections and entanglement with society impose in-depth recurrent analysis. (Kunsch et al., 2009, p. 1101) Drake et al. claim that every morally appropriate model should consider at least nancial, environmental, and social outcomes. (Drake et al., 2011, pp. 1617) Wenstp (2005) argues that rational decision-making should also take emotions into account. Additionally, in his discussion he already assumes that values and the different moral perspectives consequentialism, deontic ethics, and virtue ethics are part of morally sensitive decision-making. Gallo claims that the comprehensive inclusion of all interests is a responsibility principle: Applied to our eld this principle suggests, for example, taking into account not only the point of view of the client, . . . but also the point of view of all stakeholders (Gallo, 2004, p. 469). And recently Kijowski et al. put forward that modelers must also take into account trade-offs between the computational cost, parallel processing capability, storage cost and ease of implementation . . . and their relative importance to the applications. (Kijowski et al., forthcoming, p. 4) Finally, efcacy has also become an essential part of any thorough decision process. The multi-criteria approach, cost-benet analysis, and the score card are assessments that follow this principle. Since the multi-criteria approach can incorporate conicting objectives, Brans (2004) advocates this method for moral-laden modeling problems. For the example of the Norwegian HIV policy, Wenstp and Magnus (2001) describe how multi-criteria decision analyses can incorporate moral concerns into rational decisionmaking, such as the extent of future HIV cases, privacy, general and personal anxiety, life quality reduction, and stigmatization of individuals and groups (Wenstp and Magnus, 2001, pp. 6162). De Brucker et al. (forthcoming) acknowledge social cost-benet analysis and multi-criteria analysis both as feasible tools for addressing issues of sustainable development and discuss their advantages and disadvantages extensively. In addition, Gallo (2004) discusses, what he calls the sharing and cooperation principle, meaning that modelers should share openly the insights they gain. Mason sees it as an obligation of the modeler to insure that actions the client takes based on the model have the desired effect. (Mason, 1994, p. 184)

5. Putting the principles to work The meaning of transparency, integrity, comprehensiveness, and efcacy are clear in general. However, in particular contexts, these principles have to be further specied. As an example of such further specication, imagine a simplied optimization of a surgery schedule in a European hospital. Since I take a pluralistic approach, I do not claim that there is only one right unique specication of moral mid-level principles.

S. Diekmann / European Journal of Operational Research 226 (2013) 132138

137

Every hospital has several surgery rooms, e.g. four such rooms. Due to increasing cost pressure, hospitals need to use these rooms in a cost-optimal way. Besides this aim, three secondary aims are pursued: (1) The rooms should be used continuously throughout the working day. (2) The rooms should not be used longer than the regular working day lasts. (3) No surgeries should be canceled. Intuitively we feel that modeling surgery schedules has a moral dimension. The tools developed in the rst part of the article can help to recognize why: The consequentialist perspective reveals that particular schedules might lead to disadvantages for some patients. As a worst consequence, it seems imaginable that patients die because of blocked rooms. The other criteria however do not reveal moral problems. The intentions of developing the model are benecent: The model should help increasing the efciency of hospital facilities.9 The model is not concerned about questions of a good life, because the behavior of all stakeholders (surgeons, nurses, patients, etc.) is neither questioned nor affected. In light of the consequentialist moral issues, I show how the principles can be specied. (Specications are written in italics.) Comprehensiveness: The modeler should acknowledge all stakeholders and assess potential moral issues in detail. The stakeholders of the surgery schedule are nurses, physicians, patients, and society. On the one hand, physicians and nurses benet from higher salaries during overtime. On the other hand, there is probably a higher benet from a schedule that reliably stays within the usual working hours. Society is a stakeholder, because European health costs are renanced via social contracts.10 The major interest on the societal level is cost effectiveness, because all saved money can be used for other purposes. Patients are the nal group of stakeholders. We can distinguish between patients with scheduled surgeries and emergency patients with unscheduled surgeries. Since one of the secondary aims is decreasing surgery cancelations, patients with scheduled surgeries can expect a benet. However, those patients could also face longer waiting times, due to higher time variability in the new schedule. Emergency patients can be affected positively or negatively. Dependent on the changes in the scheduling policy, it is imaginable that surgery rooms could be more easily available or that they are blocked for longer periods of time. Efcacy: Nurses, physicians, and society at large can expect benets from an optimal scheduling system. The modeler should assess how big these benets are dependent on the scheduling scheme. Scheduled patients are less likely to have their appointments cancelled, but might face longer waiting time within a day. The modeler should analyze the actual advantage or disadvantage of scheduled patients. For emergency patients the assessment is more complex. If emergency patients arrive while all surgery rooms are blocked, they might face chronic damages or death, since surgery cannot usually be avoided or delayed. Therefore, the modeler should explicitly evaluate the disadvantages connected to the needs of emerging patients.

Integrity: The modeler should accurately incorporate all these moral evaluations into the model and set them in relation to the cost results. Comprehensiveness: The modeler should make the moral problems explicit to the hospital management. Transparency: The modeler should highlight the origins of the problems and that the goal-criteria set in the beginning neglect these problems. Integrity: The modeler and the management should look for a solution according to the managements wishes and the moral problems of scheduling surgeries. Let us assume in this example that the new scheduling scheme jointly blocks all rooms for some periods of time. We can distinguish between two cases: Scenario A the hospital lies in an urban area; and scenario B the hospital lies in a rural area. Scenario A: During the conversation with the management it becomes clear that the potential problem fortunately does not exist. In emergencies, the ambulances check hospitals for free capacities. Due to the comparably high number of hospitals in an urban area, the expected waiting times for emergencies are always comparably low. Scenario B: During the conversation, the management understands the problem for emergency patients. Since further hospitals are far away, a solution needs to be found. Transparency: The modeler should educate the management about the possible solutions to the problem. Efcacy: The modeler should suggest incorporating an evaluation of the consequences of the trade-off between risk for waiting emergency patients and cost-efciency. Integrity: The modeler should present the rened results to the management and discuss the trade-off and its consequences in detail. Based on the rened results, it could be decided to restrict the expected time of jointly blocked rooms to 30 minutes. 6. Conicts of moral mid-level principles The different principles sometimes suggest mutually conicting obligations. In those cases the principles can be balanced against each other. Taket (1994) describes a case where transparency is in conict with integrity: Out of a desire to make my report for [the clients] as accessible as possible, I did not include a discussion of the methods used (cognitive mapping). This decision was based on my judgment (grounded in my own experience) that understanding cognitive mapping requires some exposure to the use of the technique, beyond a simple written explanation. . . . however, as I have not produced any other [technical] report, I have not followed the ORSA guidelines . . . (Taket, 1994, p. 124). Taket aimed at actual transparency for her clients, and thus she left out some information. Thereby, she acted against the professional integrity as proposed in the ORSA guidelines which demanded that a full-edged technical description of the model should be provided. In this case she weighed the ORSA guidelines as less important and decided to override them. Although balancing is necessary for applying moral mid-level principles, it poses a threat: it is imaginable that modelers might wrongly balance according to their intuitions rather than moral norms and principles. Therefore Beauchamp and Childress state six criteria of balancing (Beauchamp and Childress, 2008, p. 23)11:
11

9 I take these intentions as not raising any moral problems by themselves. Rather the consequences of one-sided optimization are problematic. 10 In many countries, health insurance functions via insurance companies. Nevertheless I speak of a social contract for Europe, because even in those cases the decision on whether one wants to be insured is hardly a personal matter. Furthermore, there are strong political inuences on the insurance companies and their policies.

I leave it to the reader to show why Takets decision fullls these criteria.

138

S. Diekmann / European Journal of Operational Research 226 (2013) 132138 Brans, J.P., 2002. OR, ethics and decisions: the OATH of PROMETHEUS. European Journal of Operational Research 140, 191196. Brans, J.P., 2004. The management of the future. Ethics in OR: respect, multi-criteria management, happiness. European Journal of Operational Research 153, 466 467. De Brucker, K., Macharis, C., Verbeke, A., forthcoming. Multi-criteria analysis and the resolution of sustainable development dilemmas: a stakeholder approach, European Journal of Operational Research. Drake, M.J., Gerde, V.W., Wasieleski, D.M., 2011. Socially responsible modeling: a stakeholder approach to the implementation of ethical modeling in operations research. OR Spectrum 33 (1), 126. Fienberg, S.E., 1994. Ethical modeling considerations in correcting the results of the 1990 decennial census. In: Wallace, W.A. (Ed.), Ethics in Modeling. Elsevier/ Pergamon, Oxford, pp. 103144. Gallo, G., 2004. Operations research and ethics: responsibility, sharing and cooperation. European Journal of Operational Research 153, 468476. Gass, S.I., 2009. Ethical guidelines and codes in operations research. Omega 37, 10441050. Kant, E., 1998. Groundwork of the Metaphysics of Morals (M. Gregor, Trans.). Cambridge University Press, New York (Original work published 1785). Kijowski, D.J., Dankowicz, H., Loui, M.C., forthcoming. Observations on the Responsible Development and Use of Computational Models and Simulations. Science and Engineering Ethics. Kleijnen, J.P.C., 2011. Ethical issues in engineering models: an operations researchers reection. Science and Engineering Ethics 17 (3), 539552. Kunsch, P.I., Kavathatzopoulos, I., Rauschmayer, F., 2009. Modelling complex ethical decision problems with operations research. Omega 37, 11001108. Mason, R.O., 1994. Morality and models. In: Wallace, W.A. (Ed.), Ethics in Modeling. Elsevier/Pergamon, Oxford, pp. 183194. Le Menestrel, M., Van Wassenhove, L., 2004. Ethics outside, within, or beyond OR models? European Journal of Operational Research 153, 477484. Le Menestrel, M., Van Wassenhove, L., 2009. Ethics in operations research and management sciences: a never-ending effort to combine rigor and passion. Omega 37, 10391043. Mingers, J., 2011. Ethics and OR: operationalising discourse ethics. European Journal of Operational Research 210, 114124. Morgan, M.S., Morrison, M., 1999. Models as Mediators: Perspectives on Natural and Social Science. Cambridge University Press, Cambridge. Rabins, M.J., Harris Jr., C.E., 1997. The ethics of modeling. Control Engineering Practice 5 (4), 519526. Ross, N.P., Harris, S., 1994. From model building to risk management: evolving standards of professional responsibility. In: Wallace, W.A. (Ed.), Ethics in modeling. Elsevier/Pergamon, Oxford, pp. 161166. Stehfest, E., Bouwman, L., Van Vuuren, D.P., Den Elzen, M.G.J., Eickhout, B., Kabat, P., 2009. Climate benets of changing diet. Climatic Change 95, 83102. Sundqvist, T., Sderholm, P., Stirling, A., 2004. Electric power generation: valuation of environmental costs. In: Cleveland, C.J. (Ed.), Encyclopedia of Energy. Academic Press, San Diego, pp. 229243. Taket, A., 1994. Undercover agency? Ethics, responsibility and the practice of OR. The Journal of the Operational Research Society 45 (2), 123132. Taebi, B., Kadak, A.C., 2010. Intergenerational considerations affecting the future of nuclear power: equity as a framework for assessing fuel cycles. Risk Analysis 30 (9), 13411362. Van de Poel, I., 2001. Investigating ethical issues in engineering design. Science and Engineering Ethics 7, 429446. Walker, W.E., 1994. Responsible policy making. In: Wallace, W.A. (Ed.), Ethics in Modeling. Elsevier/Pergamon, Oxford, pp. 226241. Walker, W.E., 2009. Does the best practice or rational-style model-based policy analysis already include ethical considerations? Omega 37, 10511062. Wallace, W.A., 1994. Ethics in Modeling. Elsevier/Pergamon, Oxford. Wenstp, F., Magnus, P., 2001. Value focused rationality in AIDS policy. Health Policy 57, 5772. Wenstp, F., 2005. Mindsets rationality and emotion in multi-criteria decision analysis. Journal of Multi-Criteria Decision Analysis 13, 161172. Wenstp, F., 2010. Operations research and ethics: development trends 19662009. International Transactions in Operational Research 17, 413426. White, L., Bourne, H., 2007. Voices and values: linking values with participation in OR/MS in public policy making. Omega 35, 588603. Willet, W.C., 2001. Eat, drink, and be healthy: the Harvard Medical School Guide to healthy eating. Simon & Simon, New York.

1. Good reasons can be offered to act on the overriding norm rather than on the infringed norm. 2. The moral objective justifying the infringement has a realistic prospect of achievement. 3. No morally preferable alternative actions are available. 4. The lowest level of infringement, commensurate with achieving the primary goal of the action, has been selected. 5. Any negative effects of the infringement have been minimized. 6. All affected parties have been treated impartially. All balancing should be done with care, because no mid-level principle can be judged beforehand to be more important than any other one. 7. Concluding remarks In the previous sections I have developed a pluralistic framework that helps to recognize a variety of moral issues by taking different moral perspectives. Moreover, the framework supports dealing with diverse moral issues by following the pluralistic ethical approach of moral mid-level principles. Moral mid-level principles are loosely connected to moral theories. They do not claim that moral theories in general are wrong or misleading; rather they explicate the local moral obligations of a specic eld, in this case modeling. Moral mid-level principles build established norms into a concise framework, backed up by moral theories. Consequently, moral mid-level principles offer a structured approach for dealing with and discussing ethical issues in modeling. Nevertheless, due to changes within the eld of operational research and changes in morality, the moral mid-level principles might also change over time. Acknowledgments Several people have given me substantial feedback on this paper. Among them are my colleagues from the philosophy department at the University of Technology Eindhoven, colleagues from OZSE conference in 2010, participants of the conference Ethics, Energy and the Future in Delft, 2010, commentators on earlier versions of this paper, and the reviewers of the manuscript. Special thanks go to Martin Peterson, Sjoerd Zwart, Kees van Overveld, Valentina Moskalenko, Petr Kulikov, and Stefan E. Mendritzki for reviewing and helping to improve the claims in this article. This research has been made possible by the nancial support of the 3TU Center for Ethics and Technology. References
Aristotle, S., 1998. The Nicomachean Ethics (D. Ross, Trans.), third ed. Oxford University Press, New York (Original work published 350 BC). Beauchamp, T.L., Childress, J.F., 2008. Principles of Biomedical Ethics, sixth ed. Oxford University Press, New York. Blair, M., Obenski, S., & Bridickas, P., 1992. GAO/IMTEC-92-26 Patriot Missile Software Problem. <http://www.fas.org/spp/starwars/gao/im92026.htm> (last accessed 14.09.11).