Tyranny of Models

Models pervade R&D: physiology models, toxicology models, statistical models, disease progression models, dynamic models, algorithmic devices, clinical models, structural models, and even the instruments we use to run our laboratory experiments. With models you input data representing the current state, turn the crank, and the model outputs a prediction of the future state. Some models are simple, for example, a tick bite that turns into a bull’s eye rash is a good predictor of someone contracting Lyme disease. Complex models typically take a series of inputs and combine them using algorithms to calculate an output. These are most useful in engineering or mechanical models where the interactions between inputs can be unambiguously calculated.

the life sciences, models blind us to the most important pieces of evidence – causality. For models to work you must select and hard code a specific formulation of the cause and effect relationship, and you must hard code your own generalization+ of patients, diseases, surrogate measures, markets, etc. The tyranny comes from the fact that models force you down a path of causal reasoning and generalization without asking your assent to the underlying rationale – a rationale that has been cleansed of the messy particulars.  Some models give access to the underlying rationale; most don’t. Frequently the underlying rationale is so complex that no one has the time to unravel it. A fundamental flaw of the modeling mindset is that causal relationships should be constantly updated and challenged in the human mind; causality should be part of our daily introspections.

If from a conceptual standpoint models are problematic, from a behavioral standpoint they’re a disaster. Modeling is merely the equivalent of building a gigantic, complex hypothesis, hiding all the assumptions behind the hypothesis, and then forcing it on our colleagues. We enslave our people and ourselves to models. Models entrap future generations with assumptions built in the past. Individuals not involved in the modeling exercise feel no ownership in the output. Being forced to rely on the model, I excuse whatever I do by running to the model. I abdicate responsibility for my children.

Models exclude those not involved in their development. They’re oligarchical. Even if someone had the time to dissect them they are disadvantaged by not having been part of the development team for the model. Many of the implicit assumptions built into the model are inaccessible: many times they cannot even be communicated outside the development team in intelligible terms. Try to argue with someone about their model and they’ll bury you in a welter of detail which they alone can have mastered: after all they built the model.

Models enslave the modelers. The overall model building exercise is complex and difficult, and the modelers fool themselves into thinking the complexity of the model increases its validity. Agreement with past results, no matter how arbitrary the timeframe, are taken as proof of validity. Laboratory results are conflated into the real world. The mental models of the developers take hold. Over time they become fully blinded to disconfirming evidence+. They go through every possible gyration to preserve their mental creations. Modelers spend far too much time polishing the model and not enough time interacting with peers, impoverishing interpersonal relationships.

Models can be useful as far as their ability to suggest future or unexpected avenues to be empirically researched. Simple models are often not problematic: the evidence can often be unambiguous. Complex models are typically problematic and require orders of magnitude more validation, often spanning generations of scientists. It’s too easy to make these models fit an arbitrary set of historical data, ignoring many important variables that may have been quiescent or counterbalancing during the selected timeframe. The complexity of the model allows greater freedom for the modeler to make the data fit. Importantly, even a perfect fit does not imply causality: it doesn’t tell us how changing the input changes the output, a sine qua non for drug discovery.