What is Model thinking, and how can models and paters be useful in resolving problems?
Joshua M. Epstein∗
Based on the author’s 2008 Bastille Day keynote address to the Second World Congress on Social Simulation, George Mason University, and earlier addresses at the Institute of Medicine, the University of Michigan, and the Santa Fe Institute.
The modeling enterprise extends as far back as Archimedes; and so does its misunderstanding. I have been invited to share my thoughts on some enduring misconceptions about modeling. I hope that by doing so, I will give heart to aspiring modelers, and give pause to misguided critics.
The first question that arises frequently–sometimes innocently and sometimes not–is simply, “Why model?” Imagining a rhetorical (non-innocent) inquisitor, my favorite retort is, “You are a modeler.” Anyone who ventures a projection, or imagines how a social dynamic–an epidemic, war, or migration–would unfold is running some model.
But typically, it is an implicit model in which the assumptions are hidden, their internal consistency is untested, their logical consequences are unknown, and their relation to data is unknown. But, when you close your eyes and imagine an epidemic spreading, or any other social dynamic, you are running some model or other. It is just an implicit model that you haven’t written down.
This being the case, I am always amused when these same people challenge me with the question, “Can you validate your model?” The appropriate retort, of course, is, “Can you validate yours?” At least I can write mine down so that it can, in principle, be calibrated to data, if that is what you mean by “validate,” a term I assiduously avoid (good Popperian that I am). The choice, then, is not whether to build models; it’s whether to build explicit ones. In explicit models, assumptions are laid out in detail, so we can study exactly what the entail. On these assumptions, this sort of thing happens. When you alter the assumptions that is what happens. By writing explicit models, you let others replicate your results. You can in fact calibrate to historical cases if there are data, and can test against current data to the extent that exists. And, importantly, you can incorporate the best domain (e.g.,biomedical, ethnographic) expertise in a rigorous way. Indeed, models can be the focal points of teams involving experts from many disciplines.
∗ Senior Fellow in Economic Studies and Director of the Center on Social and Economic Dynamics, the Bookings Institution, and External Professor, The Santa Fe Institute. I thank Ross A. Hammond for insightful comments.
Another advantage of explicit models is the feasibility of sensitivity analysis. One can sweep a huge range of parameters over a vast range of possible scenarios to identify the most salient uncertainties, regions of robustness, and important thresholds. I don’t see how to do that with an implicit mental model. It is important to note that in the policy sphere (if not in particle physics) models do not obviate the need for judgment. However, by revealing tradeoffs, uncertainties, and sensitivities, models can discipline the dialogue about options and make unavoidable judgments more considered.
Can You Predict?
No sooner are these points granted than the next question inevitably arises: “But can you predict?” For some reason, the moment you posit a model, prediction–as in a crystal ball that can tell the future–is reflexively presumed to be your goal. Of course, prediction might be a goal, and it might well be feasible, particularly if one admits statistical prediction in which stationary distributions (of wealth or epidemic sizes, for instance) are the regularities of interest. I’m sure that before Newton, people would have said “the orbits of the planets will never be predicted.” I don’t see how macroscopic prediction–pacem Heisenberg–can be definitively and eternally precluded.
Sixteen Reasons Other Than Prediction to Build Models
But, more to the point, I can quickly think of 16 reasons other than prediction (at least in this bald sense) to build a model. In the space afforded, I cannot discuss all of these, and some have been treated en pass-ant above. But, off the top of my head, and in no particular order, such modeling goals include:
1. Explain (very distinct from predict)
2. Guide data collection
3. Illuminate core dynamics
4. Suggest dynamical analogies
5. Discover new questions
6. Promote a scientific habit of mind
7. Bound (bracket) outcomes to plausible ranges
8. Illuminate core uncertainties.
9. Offer crisis options in near-real time
10. Demonstrate tradeoffs / suggest efficiencies
11. Challenge the robustness of prevailing theory through perturbations
12. Expose prevailing wisdom as incompatible with available data
13. Train practitioners
14. Discipline the policy dialogue
15. Educate the general public
16. Reveal the apparently simple (complex) to be complex (simple)
Explanation Does Not Imply Prediction
One crucial distinction is between explain and predict. Plate tectonics surely explains earthquakes, but does not permit us to predict the time and place of their occurrence. Electrostatics explains lightning, but we cannot predict when or where the next bolt will strike. In all but certain (regrettably consequential) quarters, evolution is accepted as explaining speciation, but we cannot even predict next year’s flu strain. In the social sciences, I have tried to articulate and to demonstrate an approach I call generative explanation, in which macroscopic explanation–large scale regularities such as wealth distributions, spatial settlement patterns, or epidemic dynamics–emerge in populations of heterogeneous software individuals (agents) interacting locally under plausible behavioral rules1. For example, the computational reconstruction of an ancient civilization (the Anasazi) has been accomplished by this agent-based approach2. I consider this model to be explanatory, but I would not insist that it is predictive on that account. This work was data-driven. But I don’t think that is necessary.
To Guide Data Collection
On this point, many non-modelers, and indeed many modelers, harbor a naive inductive that might be paraphrased as follows: ‘Science proceeds from observation, and then models are constructed to ‘account for’ the data.’ The social science rendition–with which I am most familiar–would be that one first collects lots of data and then runs regressions on it. This can be very productive, but it is not the rule in science, where theory often precedes data collection. Maxwell’s electromagnetic theory is a prime example. From his equations the existence of radio waves was deduced. Only then were they sought…and found! General relativity predicted the deflection of light by gravity, which was only later confirmed by experiment. Without models, in other words, it is not always clear what data to collect!
illuminate Core Dynamics: All the Best Models are Wrong!
Simple models can be invaluable without being “right,” in an engineering sense. Indeed, by such lights, all the best models are wrong. But they are fruitfully wrong. They are illuminating abstractions. I think it was Picasso who said, “Art is a lie that helps us see the truth.” So it is with many simple beautiful models: the Lotka-Volterra ecosystem model, Hooke’s Law, or the Kermack-McKendrick epidemic equations. They continue to form the conceptual foundations of their respective fields. They are universally taught: mature practitioners, knowing full-well the models’ approximate nature, nonetheless entrust to them the formation of the student’s most basic intuitions. And this because they capture qualitative behaviors of overarching interest, such as predator-prey cycles, or the nonlinear threshold nature of epidemics and the notion of herd immunity. Again, the issue isn’t idealization–all models are idealizations. The issue is whether the model offers a fertile idealization. As George Box famously put it, “All models are wrong, but some are useful.”
1See Joshua M. Epstein, 2006. GenerativeSocialScience:StudiesinAgent-Based Computational
Modeling (Princeton University Press) and the review: Philip Ball, “Social Science Goes Virtual”
2 See Axtell, RL, JM Epstein, JS Dean, GJ Gumerman, AC Swedlund, JHarberger, S Chakravarty,
RHammond, JParker, and M Parker, “Population Growth and Collapse in a Multi-Agent Model of the
Kayenta Anasazi in Long House Valley). Proceedings oftheNationalAcademyof Sciences,Colloquium
99(3): 7275-7279, and the review: Jared M. Diamond, “Life with the Artificial Anasazi,” Nature419: 567-
From Ignorant Militancy to Militant Ignorance
To me, however, the most important contribution of the modeling enterprise–as distinct from any particular model, or modeling technique–is that it enforces a scientific habit of mind, which I would characterize as one of militant ignorance–an iron commitment to “I don’t know.” That all scientific knowledge is uncertain, contingent, subject to revision, and falsifiable in principle. (This, of course, does not mean readily falsified. It means that one can in principle specify observations that, if made, would falsify it). One does not base beliefs on authority, but ultimately on evidence. This, of course, is a very dangerous idea. It levels the playing field, and permits the lowliest peasant to challenge the most exalted ruler–obviously an intolerable risk. This is why science, as a mode of inquiry, is fundamentally antithetical to all monolithic intellectual systems. In a beautiful essay5, Feynman talks about the hard-won “freedom to doubt.” It was born of a long and brutal struggle, and is essential to a functioning democracy. Intellectuals have a solemn duty to doubt, and to teach doubt. Education, in its truest sense, is not about “a saleable skill set.” It’s about freedom, from inherited prejudice and argument by authority. This is the deepest contribution of the modeling enterprise. It enforces habits of mind essential to freedom.
5 Richard P. Feynman, “The Value of Science.” In Feynman, R. P. 1999. The Pleasure of Finding Things
Out. Perseus Publishing