Sunday, 19 January 2014

Parsimony in the social sciences

Parsimony - in the form of Occam's razor or Solomonoff induction - is an important foundation of science.

However, in the social sciences, simplicity is often seen as a bad thing. Often models are criticized as being overly simplistic.

I haven't written about this before - since Boyd and Richerson's Simple Models of Complex Phenomena says practically everything I would say - and much more in far more detail than I would bother with.

However, Jonathan Haight's critique of "the pursuit of parsimony" (see also 6:10 here) makes the subject topical - so now seems like a good moment for my two cents.

Complex models have a number of drawbacks. They are difficult to construct, difficult to use, difficult to maintain, difficult to publish, and difficult to teach. In short, complexity costs.

The argument for complex models is that they are more accurate. This is often not true, but even when it is, complex models are often not worth the associated costs.

Simple models are ubiquitous in science. People still use Newtonian mechanics - even though it is less accurate than general relativity. In population genetics, one often assumes that the population size is infinite, or that gene fitnesses combine linearly to produce organism fitnesses, or that mutations occur at random. If other scientists took the social science approach, no work would get done.

Why are social scientists so obsessed with making their models perfect? I think it's because they are such sharp critics of each others' work. Anything that falls short of total realism gets shot down in flames as a naive over-simplification. Inaccurate social science models aren't just scientific mistakes, they can led to capitalism, euthanasia or Social Darwinism.

Sometimes the charge of oversimplification is genuinely useful. Perfectionism can sometimes be a virtue - but often it is a destructive and paralysing obsession. This has certainly been a common result in the social sciences. These have yet to embrace Darwinism in many cases. Much social science has been stuck in a pre-Darwinian backwater for 150 years. If you reject simplified models, you don't have basic models that can be improved on. Whatever methodology social scientists have been pursuing, it hasn't been working very well.

Anyway, no one should ever write:

Let's retire the pursuit of parsimony from the social sciences. Parsimony is beautiful when we find it, but the pursuit of parsimony is sometimes an obstacle to the pursuit of truth.


  1. To illustrate your misunderstanding of Solomonoff induction's application to social science, let's look at a linear correlation between two social variables, A and B. You have a population of people and for each you measure property A and measure property B. These produce a scatter of people on a graph. If they line up exactly (which they never do unless you have only 2 people) you have a correlation coefficient of 1.0 or -1.0. But since they never do, you have errors and a correlatin of between 1.0 and -1.0. As with all such linear regressions, the line has a slope and a y intercept (say A = slope * B + y_intercept) that are chosen so that the ERROR is minimized. Solomonoff induction forces you to add, to the size of your model, a commensurable measurement of the total error of your line. Since the size of your model is in terms of bits of information required to encode your linear equation, the number of bits required to precisely describe its errors are added to the model's length to produce the measurement of how parsimonious your model actually is. This is how Solomonoff induction avoids oversimplification and its constant companion, confirmation bias.

    Indeed, the time is ripe to retire the social sciences as we have known them and leverage Moore's Law by encourging social scientists to construct comprehensive data encoding social measures across disciplines, and compare comprehensive social models against each other according to which can produce the shortest executable archive of said data. This will revitalize intersectional disciplines like political economy but it will do far more than that: It will bring social science out of the realm of quasi theology and sophistry into honest public discourse.

    1. I mentioned Solomonoff induction and Occam's razor to indicate the scientific preference for brevity in explanation.

      It seems a bit patronizing to conclude that I am under the delusion that these ideas don't offer a way of including complexity when it is required.

      If I was really that ignorant, I would probably have been incapable of writing the rest of this article.