Dependence of Bayesian Model Selection Criteria and Fisher Information Matrix on Sample Size

Geostatistical analyses require an estimation of the covariance structure of a random field and its parameters jointly from noisy data. Whereas in some cases (as in that of a Matérn variogram) a range of structural models can be captured with one or a few parameters, in many other cases it is necessary to consider a discrete set of structural model alternatives, such as drifts and variograms. Ranking these alternatives and identifying the best among them has traditionally been done with the aid of information theoretic or Bayesian model selection criteria. There is an ongoing debate in the literature about the relative merits of these various criteria. We contribute to this discussion by using synthetic data to compare the abilities of two common Bayesian criteria, BIC and KIC, to discriminate between alternative models of drift as a function of sample size when drift and variogram parameters are unknown. Adopting the results of Markov Chain Monte Carlo simulations as reference we confirm that KIC reduces asymptotically to BIC and provides consistently more reliable indications of model quality than does BIC for samples of all sizes. Practical considerations often cause analysts to replace the observed Fisher information matrix entering into KIC with its expected value. Our results show that this causes the performance of KIC to deteriorate with diminishing sample size. These results are equally valid for one and multiple realizations of uncertain data entering into our analysis. Bayesian theory indicates that, in the case of statistically independent and identically distributed data, posterior model probabilities become asymptotically insensitive to prior probabilities as sample size increases. We do not find this to be the case when working with samples taken from an autocorrelated random field.