When comparing models to observations, it is important to have tools that explore the parameter space, quantify how well the model represents observations, describe the probability distributions of the model parameters and help choose which model is most likely to be the true one. As well as making use of existing tools I contribute to efforts to make new ones.
I am developing ampere, which aims to make it easy for astronomers to fit heterogeneous datasets with their model of choice, even when the model is unable to reproduce all the features of the data. The key element to ampere’s approach is to realise that an incomplete model will produce structured residuals, but structured residuals in the model-data comparison are indistinguishable from correlated noise in the data itself. By introducing a noise model for the data which optimises the amount of correlated noise during the fit, we can in effect down-weight those features that the model is unable to reproduce. This allows us to exploit the features of the data which the model represents well, without getting distracted by the ones that it doesn’t. This makes it easier to model very heterogeneous data sets where different elements have very different information content. Ampere is still in development, but we hope it will be ready for usage soon!
I also recently released PyBayCor (“Pie Baker”). Although this is only an early version of the package, it is intended to help determine Bayesian correlation coefficients for multi-dimensional datasets. It is capable of robust inference on data both with and without measurement uncertainty.
Peter Scicluna 