Disclosing the truth: Are models better than observations?

Abstract

The aphorism, ‘All models are wrong, but some models are useful’, originally referred to statistical models, but is now used for scientific models in general. When presenting results from a marine simulation model, this statement effectively stops discussions about the quality of the model, as there is always another observation to mismatch, and thereby another confirmation why the model cannot be trusted. It is common that observations are less challenged and are often viewed as a ‘gold standard’ for judging models, whereas proper interpretations and the true value of models are often overlooked. Models are not perfect, and there are many examples where models are used improperly to provide misleading answers with great confidence, but to what extent does an observation represent the truth? The precision of the observational gear may be high, but what about representativeness? The interpretation of observations is simply another model, but this time not coded in a computer language but rather formed by the individual observer. We submit that it would be more productive to initiate a process where the norm is that models and observations are joined to strengthen both. In the end, neither method is the goal, but both are useful tools for disclosing the truth. Biased views on either observational or modeling approaches would limit us from achieving this goal.

Exploring the impact of climate change on the global distribution of non‐spinose planktonic foraminifera using a trait‐based ecosystem model

“…Model‐data biases may arise from a lack of plasticity in the model calcification trade‐offs, spatial and temporal sparsity in data, especially for highly undersampled parts of the Pacific, South Atlantic, and Indian Ocean, and very likely, because of the absence of fine‐scale ocean dynamics and inter‐annual variability. Due to sparse physiological and distribution data in temporal and spatial scales, it is unknown to us to what degree the model overestimates and the empirical observations underestimate the biomass stock and calcium carbonate flux at this point (Skogen et al, 2021). Additional empirical observations would aid the development of a large‐scale zooplankton database which includes less abundant groups, such as foraminifera, enhance the model‐data comparison, identify and help to overcome model limitations (Jonkers et al, 2021).…”

Building bridges between natural and social science disciplines: a standardized methodology to combine data on ecosystem quality trends

“…Typical examples of interactions between the different types of natural science data include the validation of remote sensing through ground truthing surveys via in situ sampling, assimilation of data from either sampling or remote sensing into models [49] and interactions between time series and models for validation or development of new paradigms [50]. These projects strengthen the argument that different types of data make different contributions, and that their combination provides important added value [51].…”

Making spatial-temporal marine ecosystem modelling better – A perspective

“…Promising approaches include adopting a modular design to model construction that allows for bypassing internal computations with the advice from dedicated expert models ( Christensen et al, 2014 ; Coll et al, 2020 ; Steenbeek et al, 2016 ). Climate models quantify IIV uncertainty by starting models at different times with different realizations (e.g., Nadiga et al, 2019 ), which is hard to achieve for marine ecosystem models that have much more complex starting states to represent the living components in the system (e.g., Skogen et al, 2021 ). This difficulty is exacerbated by the much sparser nature of ecological data, which has yet to achieve the precision and coverage of physiochemical ocean properties.…”

“…For more complex models, adaptive parameter sensitivity screening ( Pantus, 2007 ) may serve to greatly reduce the number of needed parametric uncertainty iterations. Uncertainties in observations ( Skogen et al, 2021 ) and driver data ( Coll et al, 2020 ) can be considered as parametric uncertainty too, and should be treated as such in uncertainty assessments. Scenario uncertainty relates to the inability to accurately define future contexts within which the dynamic conditions captured within a MEM play out.…”