Sensitivity of non-conditional climatic variables to climate-change deep uncertainty using Markov Chain Monte Carlo simulation

As stated above this study aims to shed light on the deep uncertainties that are associated with the climate change phenomenon. The seasonally-averaged surface air temperature, hereafter simply referred to as temperature, was selected as the non-conditional climatic variable to be monitored within the Karkheh River Basin, Iran, during the baseline period (1975–2005). The CORDEX datasets (RCP 8.5) were employed to make climate-change projections.

The first step in the proposed framework is to identify the most suitable theoretical distribution function to represent the stochastic behavior patterns of both historical and climate change data sets. Such identification considered the following theoretical distributions: normal, lognormal, exponential, Weibull, 3-parameter Weibull, extreme value, gamma, logistic, and loglogistic. It is important to note here that the primary strategy in this study is to analyze the data from a numeric standpoint without any presumption about the stochastic structure of the data⁴⁴. As such, the study would opt for any distribution that is deemed fittest to describe the data. A summary of the fitted distributions to represent the prior distributions and likelihood functions is found in Tables S1 through S4 (see the Appendix). Furthermore, the climate-change period was divided into three mutually exclusive time frames which are short-term (2010–2039), mid-term (2040–2069), and long-term (2070–2099) future to gain a better understanding of the evoluton of future temperature changes.

With Bayes’ theorem in mind, a Markov Chain Monte Carlo (MCMC) method was then applied to merge the prior distributions and the likelihood functions and to generate a sample set from the posterior distribution set. After a series of trials-and-errors, the sample size for the MCMC algorithm was set to be 1000 (n = 1000). These generated sample sets were then used to specify the most suitable theoretical pdfs to represent the posterior distribution functions. Figure 3, for instance, illustrates the most appropriate theoretical distribution that could represent the posterior distribution for the Seimareh sub-basin during spring under the short-term period.

Figure 4 demonstrates the frequency with which each specific theoretical distribution functions was deemed the most suitable to characterize the prior, likelihood, and posterior distributions. Analyzing the fitted pdfs in Fig. 4 reveals an important point about the nature of RCMs’ raw projections. Specifically, the most frequently chosen distribution function for prior and posterior distributions is the 3-parameter Weibull. As for the likelihood function, however, it was the normal distribution that outperformed other available alternatives. Furthermore, the type of selected theoretical distribution for prior and posterior pdfs seems far more diverse compared to those from the likelihood functions. In fact, the likelihood functions were only limited to three types of distributions, most of which are normal distributions. Keep in mind that these functions are the most suitable pdfs that were fitted to the RCMs projected results. The cause behind this notion might be traced back to the nature of RCMs’ projections. RCMs operate at a finer horizontal resolution than GCMs, and thus they provide localized and high-resolution detailed climatic information that can be of importance for many management purposes, especially in regions with complex topography. However, the analyzed data revealed that among the distributions fitted for the likelihood function the normal distribution was found to be the best distribution to describe the data 70% of the time. This could be interpreted as signaling that employing RCMs’ raw projections, especially for regions that have considerable volatility in their climatic variables, should be used with caution, and further adjustment to the raw projected data may be required in some cases. Note that from a statistical standpoint, the normal distribution is not heavy-tailed, and as such, may not be the best way to portray this data. The fact that, in most cases, it has been selected as the best way to portray the stochastic nature of the likelihood function (i.e., RCM’s projections) means that innate characteristics of these data might prevent them to truly represent these types of variables on their own.

Figure 5 provides additional information regarding the frequency in which each individual theoretical distributions were deemed suitable to represent the posterior pdfs. While posterior distribution sets are, indeed, the most diverse in terms of the number of different types of distributions, a significant proportion of fitted pdfs (approximately 52%), however, are fitted by the 3-parameter Weibull distribution. Further information regarding the fitted distributions to represent the posterior pdfs is found in Tables S5 to S7 (Appendix).

The computed posterior distribution functions can be interpreted as modified representations of the stochastic behavior of temperature variable concerning the short-term, mid-term, and long-term climate change projections. In that spirit, employing the confidence interval of 95%, the average temperature of the entire basin is depicted in Fig. 6 associated with historical and climate change conditions. Two sets of behavior patterns are observed. The first one is a broad trend in summer. The second pattern describes the rest of the seasons. In summer (Fig. 6b) the presence of a mild, yet, steady positive trend (upward) is detected. Here, one can expect the average temperature of the basin to increase steadily with the passage of time. As for the rest of the seasons, while it seems that the average temperature of the basin would experience a mild drop in the short-term, the temperature would begin to rise with a steady trend with time. In spring (Fig. 6a) and autumn (Fig. 6c) time series, it is projected that the expected average temperature in the basin would eventually surpass those that had been experienced in the baseline condition in the mid-term and long-term future. Concerning winter temperature it is seen in Fig. 6d that it is projected to increase over time. Yet, it has been estimated that it might not reach the observed average temperature of the basin in neither of the expressed time frames. Of course, given the upward trend in the data, this temperature would indeed be reached in a longer timeframe. It is worth noting that these patterns are in line with the idea that the earlier impacts of climate change are to amplify the historical patterns in climatic variables. That is why the data show a slight drop in colder seasons and an uptick in the warmer ones. That is, of course, until eventually, a new climatic equilibrium is reached on a global scale. At this point, the temperature as shown here would start to increase gradually.

The other notable implication that can be understood by analyzing Fig. 6 is the variation in the width of the confidence intervals under baseline and climate change conditions. In comparison to the baseline condition, the length of the 95% confidence intervals would dramatically decrease under climate change conditions. This shrinking indicates that the generated results are more densely surrounding the central tendency measure herein chosen as the mean (μ) of the data. This notion is still in line with the idea that RCM’s projections mostly resemble the stochastic characteristic of normal distributions. A normal distribution is by nature not a heavily-tailed distribution, meaning that it rarely generates tail values. Even though the MCMC framework has mitigated this effect to some extent, they inevitably inherit this stochastic property from the likelihood functions.

Again, to truly understand the obtained results here, one must first acknowledge how Bayesian models work. The main idea behind a Bayesian-based framework is to adjust the prior assumptions about a stochastic phenomenon through observed samples. In this case, the prior information represents the historical data, and the likelihood function (i.e., the samples) is obtained from RCM projects. As can be seen here, while RCMs’ projections might be perfectly capable of portraying the normal behavior of a variable under climate change conditions, which is usually sufficient for most lumped evaluation of climate change impact assessments, they might not be suitable to study extreme hydro-climatic events. The main problem with the raw RCM projections is that they follow a normal distribution, which is a symmetric distribution. Figure 4 suggests that while the MCMC framework here is mitigating this impact the final projections inherit this property from the likelihood functions. This simply means that while any RCM-based projection is perfectly suitable to understand the general outline of the climate change impacts, they are not the best option to study extreme events because even by modifying their pdfs, they rarely generate truly extreme values. The average temperatures in all sub-basins under baseline and climate change conditions are summarized in Table 1.

As for the impact of climate change, it is clear that these data are associated with deep uncertainty; that is, the parameters used to describe the stochastic behavior of a variable may be subjected to some degree of uncertainty. These parameters, μ for one, may also be represented by a pdf of their own. This study focuses on highlighting this type of deep uncertainty that might interfere with the central tendency measure μ.

The deep uncertainty in this instance dictates that the recorded parameters for each posterior distribution are not deterministic values. While for a given prior distribution and likelihood function the MCMC would lead to a specific type of posterior pdf, the parameters that are used to define this pdf (e.g., μ), could vary each time the algorithm is used. If this variation is mild, there is more certainty about the nature of the variable’s stochastic behavior pattern (i.e., the posterior distribution function). If it is determined that the parameters are experiencing severe variations then the deep uncertain environment would leave the decision-makers unsure about the variables’ stochastic behavior pattern.

With that idea in mind the combination of prior distributions and likelihood functions was executed for 100 times, and in each iteration the mean of 1000 samples was recorded. A theoretical distribution function was then fitted to the recorded values. Naturally, if the recorded values are generally close to one another numerically, the parameters of the computed posterior pdfs are less subjected to deep uncertainty. If, however, these values show significant fluctuation then the deep uncertainty of climate change would impede predictions of the stochastic behavior pattern of temperature. Figure 7, for example, portrays the uncertainty of the computed μ parameter for Seimareh sub-basin in spring under short-term future condition.

Figure 8 demonstrates the number of times each theoretical distribution was chosen to portray the stochastic behavior of the μ parameter. As can be seen here, the normal and lognormal distributions are the most common pdfs used to describe the variation in the μ parameter. One should also note the fact that about 65% of the distributions used to describe the future condition are normal distributions. The list of fitted pdfs is summarized in Tables S8 to S10.

Table 2 summarizes the variation in the computed μ parameter in each given sub-basin. It is seen in Table 2 the 95% confidence interval of the μ parameter in all cases ranges between ± 0.1 and ± 0.3 °C. In 55% of the cases, this interval was found not to be more than ± 0.1 °C, and, furthermore, in 97% of them the interval was less than ± 0.2 °C. Needless to say, a widened confidence interval for the μ parameter can only signal that the deep uncertainty has a more pronounced impact on the temperature’s stochastic behavior. As for the case of the spring data set of the Seimareh sub-basin under the short-term condition, or the case of the Gharesou sub-basin’s winter data series under short-term period, the confidence interval for the μ parameter is estimated to be ± 0.3 °C wide. This indicates that compared to other projected posterior pdfs there is less certainty about the predicted stochastic behavior pattern of temperature variable for these particular cases. As shown in Table 2 in some cases, the variation in the projected μ temperature’ posterior pdfs is decreasing over time (for a given season over different timeframes). As discussed earlier, this was interpreted as the deep uncertainties of the climate change projections, meaning that lower volatility in this measure indicates that the said variable is less affected by the deep uncertainty of the climate-change phenomenon. This observation is in line with the general belief that, in the near future, the climate change phenomenon is most likely to intensify the historical patterns in climatic variable, but gradually we expect to see an upward trend in temperature in the longer run⁴⁵. In this case, there is more volatility in the earlier time frames, but as time progresses, this volatility seems to decrease in some cases. This means that the obtained projections are showing less uncertainty about the outline behavior of the parameter for the long-term future as the models that are simulating the climatic behavior under climate change conditions have already reached a new equilibrium by that point.

Source: Ecology - nature.com