Next weekend I’m heading to Minneapolis to present at the Second Biennial Conference of the Society for Philosophy of Science in Practice (SPSP)
Increasing Returns in Science
A Formal Model of Consensus and Dissensus
Abstract: We construct a formal model describing the dynamics of science under increasing returns. The model restates the problem of the division of labor in science as an attempt to bring in increasing returns to the dynamics of science. By assuming increasing returns our model differs significantly from Kitcher (1990) and is closer to a series of models originally developed by Brian Arthur (1994). While Arthur’s models describes the economics of technology choice, we demonstrate that a similar model can be used to replicate the dynamics of science.
We build an abstract computational agent-based model. In our model there is a population of heterogeneous scientists. Their main activity is to produce evidence. By producing evidence (e.g. writing a paper) each scientist employs the methods of a particular ’school of though’, ‘paradigm’, or cluster. The decisons at the micro-level produce a particular pattern at the macro-level. Several ’schools of thought’, or clusters exist side by side (diversity), or one cluster gets dominant (specialization) with several smaller clusters relegated to the fringes.
The individual scientist does not directly react to a objective world but to the available evidence produced by his fellow scientists. He relies on his colleagues’ testimony. His decision – specialize or diversify – is based on his own preferences and the available evidence produced by his fellows. This introduces a kind of herd behavior where, under certain conditions, uniformity of opinions emerges as a result of positive feedback effects. In Arthur’s model the corresponding situation would be a lock-in: all producers adopting the same (potentially ineffective) technology. In the dynamics of science, however, we rarely find uniformity of opinions. There are always some sceptics out there, opposing conventional wisdom. We take this into account and tune our model so that complete lock-in is only a special case. In the more general case of the model we see a dominant cluster besides several small ones. The process in which one cluster gets dominant is path-dependent and nonergodic. Random events are not averaged away over as time passes, and small fluctuations matter for the selection of the dominant cluster. Although we cannot predict which cluster will get dominant, we know that one cluster will get dominant for sure, hence the process is predictable.
By modeling the scientist’s choice as a nonlinear Polya process we take into account increasing returns. The strength of the increasing returns effect depend on available evidence as well as on the strength of clusters. Within stronger clusters scientists are more likely to stick to the accepted methods and specialize. In addition to the effect of a cluster’s strength and evidence there are the scientist’s preferences. By making a contribution to a cluster a scientist invests time and money. These sunk costs lead to a change in the scientist’s preference, making the agent more likely to contribute to the same cluster again. Or more pithily: higher sunk costs make it more likely that scientists specialize.
By calibrating our model we are able to explain both, the formation of consensus and the dissolution of consensus. We explore the parameter space of the model and look how institutional factors and policy influence the dynamics of the model. In particular, we look at policies that can effectively reduce the CO-IR discrepancy. Using some examples from the history of economic thought we link our abstract model to particular periods in the history of science.
