This paper on the evolution of overconfidence came out a couple of weeks ago in Nature. The authors claim that their simple model offers an explanation as to why humans are typically overconfident when faced with novel or uncertain situations, such as climate change or the financial crisis. In general, the I like the author – Dominic Johnstone’s – work. He wrote a great paper on the logic of cooperation a few years ago and did some interesting stuff using evolutionary thinking to study adaptation rates for the US Army and insurgents….the structural and incentive differences mean insurgents modify their behaviour far more rapidly than the Army can keep up with.
Unfortunately, I really dislike this latest paper. The only thing novel about it is the overconfidence spin – but sadly, their model is not really about the evolution of overconfidence. The novelty they introduce to a standard animal conflict model is a parameter, k, which they have defined as overconfidence; but that is a dangerously loaded term to give such a general parameter. Essentially, all the model shows is that when the cost of gaining a resource goes down, an individual is more likely to go for it – a basic result known since the 1970s. The model results also rest upon some unrealistic assumptions – such as that if the chance of an individual winning the resource is <0.5 then they won’t go for it – I don’t think this is realistic, as it would leave resources often unclaimed…what if a resource is extremely valuable, but there is only a 10% chance of winning it? Natural selection will often favour high-risk high-reward strategies.
What would a model need to show in order to convince me that overconfidence was an adaptive trait? I think you’d need two parameters – one giving your confidence that you can get a resource, and another reflecting how likely you can get the resource based upon an estimate of your personal quality (with an error included so we do not assume perfect information)…The ‘quality estimate’ trait value would vary with partner quality / resource value, but the ‘confidence’ trait value would be independent of resource / partner quality – consistently positive for overconfident, negative for under confident. We could then examine the evolution of the ‘confidence’ trait value (i.e. how over / under confident and individual is favoured to be) – my guess is that natural selection would favour this trait to go to 0 i.e. to have no consistent over / under confidence bias…but I’d love to see someone model it.
It’s a pity this paper made it through the review process, but the publications of papers of this quality happens so frequently, it is not surprising any more.