Decision Support Systems for Just War Deliberations

Unless one is a pure pacifist, the general assumption is that some wars are justified. For centuries a body of literature called just war theory has developed concerning what distinguishes a just from an unjust war.  The criteria come under several headings, like (1) just cause, (2) right intention, (3) last resort, (4) legal authority, (5) probability of success, and (6) that the war not produce greater harms than it intends to solve.

If these criteria, which conform to common sense and moral philosophy alike, were applied scrupulously, most wars would be avoided. The problem comes in practice:  governments, if they consider these criteria at all, typically pay mere lip service to them. For example, to satisfy the just cause criteria, threats posed by foreign powers are greatly exaggerated; and the predicted costs of a war, both economically and in terms of human life and suffering, are greatly minimized. Further, as happened in the case of the 2001 Afghan War and the 2003 Iraq War, intellectuals spend more time arguing tedious fine points about the precise technical meanings of just war criteria than in applying them in a practical and sensible way.

Considering this, it struck me how there is a close similarity between the decision to make war and a medical decision to perform some drastic and risky procedure — say, a dangerous operation. In the latter case, because of the complexity of the choices involved and the fallibility of human decision-makers, expert systems and artificial intelligence have been used as decision support tools. In fact, I’ve developed one or two such systems myself.

Computerized medical decision-support systems offer several benefits. First, they can help a physician decide how to treat a particular patient. For example, based on such variables as the patient’s age, health, genes, and physiology, the system might supply the physician with the estimated probabilities of success for several treatment options (e.g., surgery, medication, naturalistic treatment, or perhaps no treatment at all). The physician isn’t required to follow the recommendation — but he or she can take it into account. Usually it is found that, in the long run, incorporating such a system into medical practice reduces the number of unnecessary procedures and improves practice overall.

Second, and perhaps more importantly, the process of developing of a medical decision support system is itself very valuable. It requires physicians and medical scientists to focus attention on how actual treatment decisions are made. Ordinarily, diagnosis and treatment selection can be a very subjective and ad hoc thing — something physicians do based on habit, wrong practices, or anecdotal evidence. Developing an expert system forces physicians to explicitly state how and why they make various decisions — and this process not infrequently reveals procedural errors and forces people to re-think and improve their practices.

Both of these advantages might accrue were we to similarly develop a computerized support system to decide whether a war is just. From the technical standpoint, it would not be difficult to do this; a functional prototype could easily be developed in, say, 6 weeks or less. Off-the-shelf software packages enable the rapid development of such a system.

Another advantage of such systems is that they do not produce yes/no results, but rather a probability of success. That is, they are inherently probabilistic in nature. All inputs — for example, whether a foreign power has weapons of mass destruction — would be supplied as probabilities, not definite facts. Probabilities can be estimated based on mathematical models, or expert consensus (e.g., the Delphi method).

A decision support system helps one see how uncertainties accumulate in a complex chain of inferences. For example, if the success of choice C depends on facts A and B both being true, and if A and B are only known as probabilities, then a system accordingly takes uncertainty concerning A and B into account in estimating the probability of C’s success. In a medical decision based on a dozen or more variables, none known with complete certainty, the net uncertainty concerning success or failure of a particular treatment option can be considerable. In that case, a physician may elect not to perform a risky procedure for a particular patient. The same principle would apply for a just war decision support system.

Such, then, is my proposal. From experience, I’ve learned that it is better to start with a simpler decision support system, and then to gradually increase its complexity. Accordingly, I suggest that we could begin with a system to model only one part of just war theory — say, just cause, or ‘no greater harms produced.’ I further propose that we could take the decision to invade Iraq in 2003 as guiding example. My guess is that were such a model produced, it would show that the likelihood of success, the immediate necessity, and the range of possible harms were all so uncertain in 2003 that we should have not intervened as we did.

A final advantage of such a system is that it would connect moral philosophy with science. Science is cumulative: one scientific or mathematical advance builds on another. The same is not true of moral philosophy. Philosophers can go back and forth for centuries, even millennia, rehashing the same issues over and over, and never making progress.

Perhaps this is a project I should pursue myself. Or it might be an excellent opportunity for a young researcher. In any case,  I’m throwing it out into cyber-space for general consideration. If anyone reads this and finds it interesting, please let me know.

Incidentally, military analysts have developed many such computerized systems to aid combat decisions.  (When working at the RAND Corporation, I worked on a system to help US forces avoid accidentally shooting at their own aircraft — something called fratricide.) Since it is clearly in the interests of the military to avoid pursuing unwinnable wars, possibly it is they who could take a lead in developing the line of research proposed here.  US Naval War College and West Point — are you listening?

Swine Flu, Vaccines, and Mathematical Models

The recent swine flu outbreak in Mexico reminds me that, although lately I’ve been working on other things, I should also continue my work in health policy research and related areas.

Here we consider the problem:  in a flu pandemic, what strategies can we use to conserve scarce vaccine?

Let’s assume, for example, that during the first 3 months of a flu pandemic, a country has 1 million doses of flu vaccine.  How can this quantity, which is not sufficient to immunize the entire at-risk population, be used as effectively as possible?

First we need to decide what “as effectively as possible” means.  Is the objective to minimize total mortality, to minimize mortality and morbidity, to maximize what are called QALY’s (quality-adjusted years of life), or to reduce negative economic impact?  All of these are defensible criteria.  This requires some careful analytical modeling and work.

As just one example related to this, should scarce vaccines be direct more towards children, young adults, or older adults?  Older adults are a likely target, as they have the highest mortality rates in a flu pandemic.  However they are, unfortunately, least likely to exhibit a positive immune response to flu vaccines.

Conversely, children respond well to the vaccines; and by potentially saving a child’s life, one theoretically gains  many years of productive life. Further, while this may require further epidemiological study, children, who attend school along with dozens or hundreds of other children, are probably disproportionately both at risk for flu and involved in transmission once they catch it.   However school-age children also tend to have fewer complications and lower mortality rates with flu.

In the end, an optimal allocation of flu vaccine may require a fairly complex analysis and/or computer simulation.  Various parameters that feed these analyses would need to be quantified beforehand.  For this we would have two choices:  (1) either estimate the parameters based on a combination of guesswork and literature review, or (2) to conduct small experimental studies aimed to supply more realistic values.

The choice between (1) and (2) could itself be made by performing mathematical sensitivity analyses within the simulation models; highly sensitive parameters — those for which small differences have a large effect on results — would be worth investing more money to quanity precisely.

In general, it should be noted that everything discussed here — simulations, literature reviews, mathematical analyses, etc. — are extremely inexpensive compared to the costs of large-scale population immunizations.  Half a million dollars, say, buys an immense amount of mathematical research.  And it could easily save tens or even hundreds of millions of dollars by preventing disease or streamlining immunization efforts.

Predicting Individual Response to Vaccine

Another productive area of mathematical modeling here would be to try to predict individual  response to vaccines.  For a given flu vaccine, only a certain proportion of people develop the intended antibodies.  For a particular population and vaccine, for example, this rate may be only 70%.  It would be worthwhile to know in advance whether a given person is among the 70% that respond to a vaccine or the 30% that do not.  If someone won’t probably won’t respond, spare the vaccine dose and give it to someone who will.

Such analyses can be performed using routine predictive statistical methods, like logistic regression, or perhaps more modern techniques.  Possible predictor variables might include:  subject age, sex, immunization history, flu history, ethnicity, overall health, weight.

Other predictive variables might be measured via blood tests or even DNA testing.  The choice concerning how heroically to collect predictive variables would depend on factors unique to the pandemic, such as the virulence of the strain, and the amount of existing vaccine.  In theory, if a flu strain is dangerous enough, and if vaccine is scarce enough, literally every available dose must be directed to someone it can potentially benefit.  In that case even as expensive (currently) a procedure as micro-array DNA screening could be utilized.

Other benefits from mathematical modeling and prediction in a pandemic might come by analyzing cross-reactivity of previously-developed vaccines for the current flu strain.  In the past vaccines have been developed for perhaps dozens of flu strains.  In theory, each of these vaccines is unique.  The usual assumption is that a vaccine for one flu strain offers little or no protection for a new strain.
However, that is not always the case.

The only way to be sure would be to test old vaccines against the new flu strain.  In theory, this could be done using human subjects in only a few days, at the outset of a pandemic.  All that is required is to administer an old flu vaccine to a subject, wait a few days, and then see if their blood contains antibodies effective against the new strain.

Perhaps this is a long-shot, but we might get lucky, and would lose nothing by trying.

An even more elaborate strategy would involve trying to predict cross-reactivity of previous flu vaccines to the new strain in a particular patient.  That is, by considering demographic, biological, or genetic variables of a given subject, we might identify those will exhibit favorable crossreactivity.

In addition, we could probably make some good guesses about crossreactivity simply by comparing the genetic composition of the new strain to previous ones, and applying mathematical or artificial intelligence models.

More broadly, there’s a lot more we can do at the behavioral level to prevent or limit a flu pandemic.  Public information aimed at teaching people how to prevent spread of flu is effective and cost-effective.  The pharmaceutical company GSK, for example,  has produced some excellent web-based presentations that teach people about flu prevention.  People need to learn, for example how to wash their hands correctly (30 seconds; warm water; wash both sides and between fingers).

Personally, I would like to see studies done on the potential preventive effects of wearing surgical masks on airplanes or subways.  Or perhaps, in the case of airlines, does anybody know what’s going on with the air recirculation system?  Is it filtered, and, if so, can the filters trap virus-bearing dust particles?  Airlines might be reluctant to address this issue.  Pictures of mask-wearing passengers isn’t exactly good advertising.  But on the other hand, people now are already avoiding air travel because of flu fears.  If the airlines could show that masks significantly reduce risk of contagion it might actually be good for them.