TG15 – Statistical problems in fitting models to heterogeneous data

TG Leaders: David Cameron / Michael Dietze

Description: This task group will focus on generating a concepts paper detailing the challenges involved with constraining models with multiple data types. Using multiple sources of data to calibrate models has numerous advantages in terms of better constraining parameters and broadening the applicability of models. However, biases in observations or model structure can lead to inconsistent parameter estimates for different data types. These problems can be greatly exacerbated when different data types have very unequal sample sizes, where one data type will dominate parameter estimates. Conventional ad hoc solutions, such as weighting different parts of the likelihood or subsampling/time-averaging the data, affects parameter estimates and leads to subjective uncertainty estimates based on the choice of weight or sampling. Approaches accounting for autocorrelation or information content are promising, but can lead to unexpected paradoxes. This group will review these challenges and possible solutions, using models fit to simulated data to demonstrate common pitfalls.


Participation: If you are interested in participating in the task group, please contact the task group coordinator directly. If you are interested in the working group in general, please contact the working group leaders.