Towards a comprehensive uncertainty assessment in environmental research and decision support

Scientific integrity principles and ethical concerns require scientists to be open about and proactively communicate the uncertainty in their predictions. In environmental decision support, predictions of the consequences of decision alternatives have to be assessed for the fulfilment of societal goals. This adds additional uncertainty to the decision support process, as elicited societal preferences are uncertain due to uncertainty of and temporal changes in the preferences of individuals, different perceptions and values of different individuals in the society, and uncertainties induced by the parameterization of quantified preferences and the elicitation process.

Over the past few decades, many promising concepts and methodologies have been developed to address these uncertainties. Nevertheless, most environmental decision support processes address uncertainty only incompletely. To stimulate more comprehensive uncertainty analyses in the future, it is the goal of this paper to review and discuss techniques that can contribute to better considering uncertainty in environmental research and decision support and to outline a decision support procedure that more comprehensively addresses uncertainty. In particular, the goal is to emphasize how the uncertainty of preferences and the ambiguity in specifying prior distributions can be included in a decision support process as these uncertainties, despite their high relevance, are often neglected in environmental decision support.

Many reviews on modelling environmental systems and on the modelling process exist. Most of these reviews emphasize the importance of quantifying uncertainty (e.g. Clark 2005; Refsgaard et al. 2007; Schuwirth et al. 2019). Other studies focus on expert or stakeholder involvement in model building (e.g. Voinov & Bousquet 2010; Krueger et al. 2012; Voinov et al. 2016) which also relates to extending structural diversity and considering uncertainty. This article builds on this literature, but focuses on useful methodologies for uncertainty quantification rather than on guidelines for the model building or decision support process.

In a modelling process that primarily focuses on increasing our understanding of the investigated system, we need:

• (a)

  a model that attempts to describe the underlying mechanisms of the observed behaviour of a system and makes it possible to test these model formulations (‘hypotheses’) with observed data.

To support environmental decisions, we need:

• (b1) a description of societal preferences about what should be achieved, ideally in quantitative terms expressed as functions of observable system attributes; and

• (b2) a model that is based on the current state of scientific knowledge and predicts the consequences of decision alternatives on output variables that are relevant for assessing the fulfilment of the societal goals (the attributes mentioned in b1).

Note that the requirements for models of category (a) and models of category (b2) are somewhat different, so that despite many synergies, a model designed for (a) is not always the best model for (b2). In particular, a model for decision support (b2) has to predict the output variables (attributes) used to quantify the preferences (b1) and it has to describe the dependence of these on input variables that distinguish the decision alternatives (Schuwirth et al. 2019).

In the following sections, conceptual aspects of uncertainty quantification and model building are discussed before moving on to models for predicting the consequences of decision alternatives and models for quantifying preferences. Then a decision support framework with comprehensive uncertainty quantification is described. This is followed by a short section on numerical approaches and software to implement the suggested techniques. Finally, conclusions are drawn about the transfer of the suggested techniques to research and practice.

The most important conceptual decision is on how to describe uncertain knowledge. Bayesian (epistemic) probabilities are the most straightforward choice as they provide a consistent framework for conditional beliefs and iterative learning and they are compatible with adopting randomness (aleatory probabilities) as part of the uncertain knowledge about future outcomes (Reichert et al. 2015).

As environmental decisions should be based on the best available current state of scientific knowledge, prior knowledge must be carefully elicited as intersubjective knowledge (Gillies 1991; Gillies 2000; Reichert et al. 2015). This is implemented in practice by using well-designed eliciting techniques (Morgan & Henrion 1990; Meyer & Booker 2001; O'Hagan et al. 2006; Rinderknecht et al. 2011) to elicit priors, combining the assessments from multiple experts, and by carefully documenting the use of prior information from the literature to get justifiable priors.

Here, θ are model parameters; y_obs represent potentially observed states; p(y_obs|θ,x) is the probabilistic model for observations conditional on the input, x, and the parameter values (called likelihood function if viewed as a function of the parameters with actual observed values substituted for y_obs); p(θ) is the probability distribution that quantifies the prior knowledge about the model parameters; and p(θ|y_obs,x) is the probability distribution representing the updated, posterior knowledge about the parameters given the observations. Although Equation (1) is correct in the right context (learning about model parameters), its interpretation (‘prior times likelihood’) obscures:

• (i)

  that in nearly all practical applications, the main part of the prior knowledge is formulated by the probabilistic model (likelihood), which is needed to define the meaning of the parameters (or may not have parameters);

• (ii)

  that one is usually interested in predicting future or conditional states, y_new, (conditional on changes in influence factors or potential measures suggested as decision alternatives), p(y_new|y_obs,x), rather than in model parameters (which are just an auxiliary tool for an intermediate step); and

• (iii)

  that, from a statistical perspective, there is often no fundamental difference between model parameters and unobserved states and Bayesian inference can be applied to condition the joint distribution of observed and unobserved model variables on those observed to get an update of those unobserved.

In the context above, it would be better to state: ‘our joint prior knowledge of parameters and observed states is equal to the likelihood times the prior of its parameters; conditioning this joint probability on the observations leads to the posterior of the parameters’.

Model construction based on a graphical model as shown in Figure 1 and formalized by Equation (2) became prominent as ‘Bayesian belief network modelling’ (see e.g. Borsuk et al. 2004). However the underlying concept applies to all probabilistic models and is also known as ‘hierarchical modelling’ (with a slight variation in focus, see e.g. Clark 2005). Note that also ‘structural equations modelling’ (see e.g. Kline 2011) is based on very similar concepts (although initially limited to normal distributions). In the context of time-series models, again very similar model structures are known as ‘state-space models’ (see e.g. Künsch 2001), ‘hidden Markov models’ (see e.g. Künsch 2001) or ‘dynamic Bayesian belief network models’ (see e.g. Murphy 2002). Sometimes, those different terminologies for essentially the same underlying concepts can be confusing.

This concept is illustrated in Figure 2. The left panel illustrates how shapes of non-normalized probability densities (green) are bounded by non-normalized upper (u, red) and lower (l, blue) non-normalized densities. The right panel illustrates the non-normalized density (green, dashed) with the highest probability in the interval [θ₁, θ₂].

This class of probability distributions was chosen because of its invariance under Bayesian updating, under marginalization, and under propagation through deterministic functions (see Rinderknecht et al. 2014) which makes sequential learning possible within the same framework.

Data science methodologies provide new opportunities for exploring data (see e.g. LeCun et al. 2015; or more specifically regarding application in ecology and water resources research, Peters et al. 2014; Shen 2018). Even as the main purpose of these methods is exploration and prediction without explicitly considering mechanisms, these methods can contribute to a primarily mechanistically oriented modelling process by:

• developing ‘prediction models’ for aspects for which understanding is not as important as for other aspects (e.g. image recognition from remote sensing data or plankton species identification to compile input data for mechanistic models);

• identifying patterns to stimulate mechanistic model development (e.g. finding clusters of organisms that have similar properties or behave similarly);

• developing ‘black box’ models that serve as benchmarks to analyse the potential for the improvement of more sparsely parameterized mechanistic models (Kratzert et al. 2019a);

• trying to interpret the functioning of the model developed by applying machine learning techniques in the sense of ‘interpretable data science’ (e.g. Papernot & McDaniel 2018; Gilpin et al. 2019; Kratzert et al. 2019b);

• constraining the model or the learning algorithm to consider physical characteristics of the system to facilitate interpretation and learning or developing ‘hybrid models’ that bridge between mechanistic and data-based models (‘theory-guided data science’, e.g. Karpatne et al. 2017).

Wherever possible, models that are intended to predict beyond their calibration range should be designed to represent underlying mechanisms. Hardly any environmental system behaves deterministically at the level at which we observe its behaviour. Reasons for this can be true stochasticity resulting from quantum-mechanical processes, demographic stochasticity of birth and death processes, genetic stochasticity, or apparent non-deterministic behaviour caused by the limited temporal and spatial resolution of the initial state of the modelled system and of external driving forces. For these reasons, to get an appropriate description of system behaviour and of its uncertainty, we need stochastic, mechanistic models. Environmental stochasticity can be considered by making inputs and/or model parameters stochastic processes in time (see e.g. Reichert & Mieleitner 2009). Typically, considering stochasticity leads to hierarchical models that complement unobserved parameters with unobserved states as illustrated in Equation (2) and Figure 1. As shown by Chou & Greenman (2016) for a density-dependent, age-structured population model, it can be inconsistent if one tries to formulate deterministic models directly: they may not necessarily represent the development of the mean of a (more realistic) stochastic model. Ideally, stochastic modelling is combined with a multi-model approach to account for the consequences of structural uncertainty on predictions even better.

The multi-criteria decision analysis (MCDA) methodology (Keeney & Raiffa 1976; Keeney 1992; Eisenführ et al. 2010) provides an ideal framework for modelling preferences for environmental decision support (see e.g. Reichert et al. 2015).

Similarly to the case of the outcome prediction model (Equation (2) and Figure 1) this decomposition of the overall value function into value functions of lowest-level sub-objectives and aggregation functions at higher levels facilitates the construction of the overall value function by requiring less complex functions at each decomposition step.

This framework leads to a unique ranking that considers the uncertain utilities. As outlined in the next section, combining this framework with imprecise probabilities makes it possible to assess the ambiguity of this ranking resulting from the ambiguity about the probability distributions used to quantify uncertainty of outcome predictions as well as values and utilities; this adds an essential new element to the analysis.

By combining the ‘standard steps’ of decision support, such as problem definition, stakeholder analysis, elicitation of an objectives hierarchy, etc (see e.g. Reichert et al. 2015) with the techniques outlined above, a comprehensive uncertainty analysis can be performed by modifying the following three steps.

When using parameterized value functions and a parameterized conversion function to utilities, parameter estimation can be done based on elicited discrete choice selections (Hoyos 2010) or indifference replies (Haag et al. 2019a) using Bayesian inference with density ratio class priors. This leads to a posterior density ratio class of the value/utility parameters (Rinderknecht et al. 2014) that jointly describes uncertainty and the ambiguity about its quantification. If a multi-model approach was performed, either the best fitting value model can be selected or multiple models considered for further analysis.

For predicting the outcomes of the decision alternatives, either density ratio class priors (for their elicitation, see e.g. Rinderknecht et al. 2011) can be used directly, or they can be updated by Bayesian inference if data is available. In the latter case, this also leads to a (posterior) density ratio class of the parameters (Rinderknecht et al. 2014). Again, a multiple model approach is useful for considering structural uncertainty.

Evaluating the expected expected utilities in Equation (6) for imprecise probability distributions of and leads to an interval of expected expected utilities EEU(a), for each alternative, a. These intervals for all alternatives lead to an incomplete ordering of the decision alternatives that reflects the ambiguity in addition to the uncertainty. However, this seems to be the best representation of knowledge and uncertainty and thus provides the best information to be communicated to decision makers.

Efficient numerical algorithms, mainly based on Markov Chain Monte Carlo (MCMC) techniques, such as Metropolis-Hastings sampling (see e.g. Gelman et al. 2013), have been developed to sample from posteriors in Bayesian inference for cases in which the likelihood function can easily be evaluated. This is no longer true for hierarchical models that require computationally much more demanding techniques, such as Gibbs sampling (see e.g. Gelman et al. 2013), Particle Markov Chain Monte Carlo (PMCMC; see e.g. Andrieu et al. 2010), Hamiltonian Monte Carlo (HMC; see e.g. Betancourt 2017), or Approximate Bayesian Computation (ABC; see e.g. Beaumont 2010; Albert et al. 2015). Krapu & Borsuk (2019) provide an overview of software designed for this purpose.

On the other hand, the extension to imprecise probabilities based on the density ratio class does not add a lot of additional computational effort, as a (unweighted) sample of a single distribution from the class can easily be turned into a weighted sample of any other distribution from the class without having to do inference again (Rinderknecht et al. 2014). Still calculating the EEU-intervals requires the evaluation of expected expected utilities for a large sample of elements from the class.

An ideal framework for environmental decision support would be to combine mechanistic, stochastic models for prediction with uncertain utilities for quantifying preferences and to use imprecise probabilities for robustness analysis. We outlined how this can be done without significantly increasing the computational burden. It would be a very interesting next step to apply this procedure to an actual decision problem and to analyse the benefits and challenges of considering all sources of uncertainty. In particular, it would be interesting to explore whether decision makers are willing to accept the outcome in the form of an incomplete ranking based on intervals of expected expected utility rather than a unique ranking. Easy-to-apply software could significantly contribute to facilitating the application of the suggested approach in practice.

The development of this paper has benefited from discussions with many scientists over the last few years. I would like to mention in particular Nele Schuwirth, Simon Lukas Rinderknecht, Johanna Mieleitner, Carlo Albert, Simone Langhans, Fridolin Haag, and Judit Lienert. This paper is dedicated to my former PhD students Johanna Mieleitner (1975–2019) and Simon Lukas Rinderknecht (1979–2019), who contributed significantly to the development of some of the techniques discussed in this paper.