Evolution of the marginal likelihood estimation for the four tested models with sample size. The known bias of the importance sampling estimator used (Perrakis et al. 2014) is evident, but seems to become negligible for sample sizes larger than around 2000. The inset shows an enlargement of the region with sample sizes between 1000 and 10 000, where a slight preference for the squared exponential models is seen. A small random noise was added in the x-direction to facilitate viewing.