Bayesian change point quantile regression approach to enhance the understanding of shifting phytoplankton-dimethyl sulfide relationships in aquatic ecosystems
Graphical abstract
Introduction
Dimethyl sulfide (DMS) was recognized as an anti-greenhouse gas because its oxidized products acted as cloud condensation nuclei, which reflected solar irradiation and thereby contributed to the reduction of earth temperature (i.e., the CLAW hypothesis, Charlson et al., 1987). Despite recent debates or rejection of the CLAW hypothesis (Cropp, Gabric, van Tran, Jones, Swan, Butler, 2018, Quinn, Bates, 2011), the hypothesis is likely relevant in some regions like the Southern and Arctic Oceans (Krüger, Graßl, 2011, Levasseur, 2013). DMS also plays multiple essential roles in aquatic ecosystems, such as serving as an antioxidant for phytoplankton (Sunda et al., 2002) and facilitating a tritrophic mutualism between primary producers and top predators (Savoca and Nevitt, 2014). In addition, DMS is important to the global sulfur cycle (Eyice et al., 2015), accounting for about 80 of global biogenic sulfur emissions to the atmosphere (Kettle and Andreae, 2000). Phytoplankton, indicated by chlorophyll a (CHL) (Bates, Kiene, Wolfe, Matrai, Chavez, Buck, Blomquist, Cuhel, 1994, Zhang, Yang, Zhu, 2008), is the major producer of DMS (Charlson, Lovelock, Andreae, Warren, 1987, Gondwe, Krol, Gieskes, Klaassen, de Baar, 2003). Therefore, understanding the CHL-DMS relationship is critical for estimating regional or global DMS emissions from aquatic ecosystems (Anderson, Spall, Yool, Cipollini, Challenor, Fasham, 2001, Galí, Devred, Levasseur, Royer, Babin, 2015, Simó, Dachs, 2002).
Correlation analysis and ordinary linear regression have been the most widely used methods to explore the CHL-DMS relationship. Most studies revealed a positive effect of CHL on DMS (e.g., a significantly positive correlation coefficient or regression slope) (Gao, Yang, Zhang, Liu, 2017, Iverson, Nearhoof, Andreae, 1989, Lana, Simó, Vallina, Dachs, 2011, Law, Smith, Harvey, Bell, Cravigan, Elliott, Lawson, Lizotte, Marriner, McGregor, Ristovski, Safi, Saltzman, Vaattovaara, Walker, 2017, Lizotte, Levasseur, Galindo, Gourdal, Gosselin, Tremblay, Blais, Charette, Hussherr, 2020, Tan, Wu, Liu, Yang, 2017, Tortell, Guguen, Long, Payne, Lee, Ditullio, 2011, Walker, Harvey, Bury, Chang, 2000, Yang, 1999, Yang, 2000, Yang, Tsunogai, 2005, Yang, Zhang, Su, Zhou, 2009, Yang, Zhang, Zhou, Yang, 2011, Zhang, Yang, Zhang, Yang, 2014), while several studies reported a negative relationship (Froelichd et al., 1985) or no relationship at all (Nemcek, Ianson, Tortell, 2008, Watanabe, Yamamoto, Tsunogai, 1995). We note that maximum CHL concentrations in studies deducing positive CHL-DMS relationships were always much lower than those deducing negative or no relationships. For example, CHL concentrations in a series of studies on Chinese seas (Yang, 1999, Yang, 2000, Yang, Tsunogai, 2005, Yang, Zhang, Su, Zhou, 2009, Yang, Zhang, Zhou, Yang, 2011) were all lower than 4 g/L. In contrast, the CHL concentration can reach approximate 60 g/L in Froelichd et al. (1985).
A recent study examining the CHL-DMS relationship across a broad range of CHL concentrations implemented a change point model to capture the ascending and descending limbs of this relationship (Deng et al., 2020). The change point model aims to determine one or more unknown change points at which the stressor-response relationship changes. In Deng et al. (2020), the authors used 246 paired observations of CHL and DMS from 100 Chinese lakes and collected 426 paired observations from global oceans. They applied a piecewise linear regression model (Muggeo, 2003) to detect thresholds of CHL concentration, at which CHL-DMS relationships significantly changed. Benefiting from the novel application of piecewise regression, the authors revealed hump-shaped CHL-DMS relationships in both lakes and seas, which were expected to increase the estimation accuracy of global DMS emissions from aquatic ecosystems (Deng et al., 2020). The hump-shaped relationship also seemed to resolve the contradiction of the sign of the CHL-DMS relationship in previous studies, whose deductions might have been constrained by a relatively smaller sample size, a narrow range of sampled CHL concentration, or an application of a overly simplified linear regression model.
Although many informative studies have investigated the CHL-DMS relationship, we note that those studies mainly used mean regression methods (e.g., the ordinary linear regression or piecewise regression), by which the relationship between CHL and the mean of DMS distribution was estimated. A practically important alternative to classical mean regression methods is quantile regression (QR) (Koenker and Bassett, 1978). To the best of our knowledge, QR has not been used to examine CHL-DMS relationships.
QR explores the effect of one or more predictors on any quantile of the response variable distribution (Das, Krzywinski, Altman, 2019, Koenker, Bassett, 1978). Compared with mean regression methods, QR can provide a more complete view of possible causal relationships and can reveal useful predictive relationships at some parts of the response variable distribution, even when there is a weak or no predictive relationship between the predictor(s) and the mean of the response variable distribution (Cade and Noon, 2003). In addition, QR appears more robust to outliers of the response variable (Scharf et al., 1998) and is not constrained by the equal variance assumption (Cade, Noon, 2003, Das, Krzywinski, Altman, 2019). QR has been successfully applied to environmental and ecological studies. QR has been used to 1) illustrate a relatively complete view of stressor-response relationships at multiple regression quantiles (Cade, Terrell, Porath, 2008, Liang, Xu, Qiu, Liu, Lu, Wagner, 2021, Muller, Cade, Schwarzkopf, 2018, Niinemets, Valladares, 2006, Simkin, Allen, Bowman, Clark, Belnap, Brooks, Cade, Collins, Geiser, Gilliam, Jovan, Pardo, Schulz, Stevens, Suding, Throop, Waller, 2016, Xu, Schroth, Isles, Rizzo, 2015), 2) obtain reliable prediction intervals of the response variable (Heiskary, Bouchard, 2015, Kampichler, Sierdsema, 2018), and 3) reveal the limiting effect of the stressor on the response variable via the upper boundary of the stressor-response relationship (Fornaroli, Cabrini, Zaupa, Bettinetti, Ciampittiello, Boggero, 2016, Keeley, Macleod, Forrest, 2012, Youngflesh, Jenouvrier, Li, Ji, Ainley, Ballard, Barbraud, Delord, Dugger, Emmerson, Fraser, Hinke, Lyver, Olmastroni, Southwell, Trivelpiece, Trivelpiece, Lynch, 2017). The upper boundary of a stressor-response relationship illustrates the behavior of response variable when the stressor is the limiting factor (Cade, Terrell, Schroeder, 1999, Sankaran, Hanan, Scholes, Ratnam, Augustine, Cade, Gignoux, Higgins, Roux, Ludwig, Ardo, Banyikwa, Bronn, Bucini, Caylor, Coughenour, Diouf, Ekaya, Feral, February, Frost, Hiernaux, Hrabar, Metzger, Prins, Ringrose, Sea, Tews, Worden, Zambatis, 2005).
Because the CHL-DMS relationship represents a stressor-response relationship (McDowell et al., 2018), QR seems applicable and helpful to enhance the understanding of the CHL-DMS relationship. Considering the recent finding on the shifting nature of CHL-DMS relationships (Deng et al., 2020), a simple linear QR might not be adequate. A QR method with the ability to detect a change point is required but has rarely been explored (an exploration of this approach could be found in Zhou et al. (2015) who proposed a sequential change point detection method for linear QR).
In this study, we propose a novel Bayesian change point quantile regression (BCPQR) approach to investigate the CHL-DMS relationship in aquatic ecosystems. The BCPQR model integrates two well-developed Bayesian models: a Bayesian change point (BCP) model (Barry, Hartigan, 1993, Erdman, Emerson, 2007) and a Bayesian quantile regression (BQR) model (Benoit, Van den Poel, 2017, Yu, Moyeed, 2001). Both the BCP model (Beckage, Joseph, Belisle, Wolfson, Platt, 2007, Liang, Qian, Wu, Chen, Liu, Yu, Yi, 2019, Thomson, Kimmerer, Brown, Newman, Nally, Bennett, Feyrer, Fleishman, 2010) and the BQR model (Barneche, Kulbicki, Floeter, Friedlander, Allen, 2016, Uranchimeg, Kim, Kim, Kwon, Lee, 2018, Yu, Zou, Wang, 2019, Zou, Shi, 2020) have been recently introduced and applied to develop a stressor-response relationship in environmental and ecological fields. However, to our knowledge, this is the first proposal of BCPQR model in environmental and ecological studies.
There are several features of the BCPQR model that makes it desirable for ecological investigations. First, the BCPQR model inherits advantages of the BQR model and the BCP model. It is expected to be able to provide a complete view on the stressor-response relationship (Muller, Cade, Schwarzkopf, 2018, Xu, Schroth, Isles, Rizzo, 2015). The detection of any change point in the regression intercept, slopes, and/or variance of residuals is possible (Beckage, Joseph, Belisle, Wolfson, Platt, 2007, Liang, Qian, Wu, Chen, Liu, Yu, Yi, 2019). Second, the Bayesian framework would provide the convenience for parameters estimation. We can straightforwardly incorporate the change point into the BQR model structure. Parameter estimation of BCPQR model could then be achieved using Markov-chain Monte Carlo (MCMC) methods (Qian et al., 2003). Moreover, the parameter estimation framework would allow for the calculation of probability densities representing the uncertainty of parameters (including the change point and the other model parameters) (Ellison, 2004, Gende, Hendrix, Harris, Eichenlaub, Nielsen, Pyare, 2011, Underwood, Rizzo, Schroth, Dewoolkar, 2017). In addition, based on posterior distributions of parameters, comparing parameters is straightforward (Alameddine, Qian, Reckhow, 2011, Qian, Craig, Baustian, Rabalais, 2009). Finally, prior information – if available – could be used during model development (Ellison, 1996, Ellison, 2004).
Our objective was to examine whether or not the BCPQR approach can enhance the understanding of CHL-DMS relationships in lakes and seas. We applied the proposed BCPQR model to reevaluate shifting CHL-DMS relationships revealed in Deng et al. (2020). We separately fitted BCPQR models at five regression quantiles. To avoid over confidence in the change point model, a common practice is comparing the change point model with a model without any change point (Cahill, Rahmstorf, Parnell, 2015, Liang, Qian, Wu, Chen, Liu, Yu, Yi, 2019). Therefore, we also fitted a BQR model at each regression quantile and compared performances of the two models as a means to select the best model for characterizing the CHL-DMS relationship.
Section snippets
Data source
Observations of CHL, DMS, and pH in lakes were directly obtained from Deng et al. (2020), in which the authors sampled 246 sites from 100 shallow lakes in China. Locations of these lakes range from 111E to 122E in longitude and from 28N to 39N in latitude. CHL concentrations varied widely, ranging from 0.55 g/L to 58 g/L, with an average of 11.87 g/L and a standard deviation of 10.95 g/L. The average DMS concentration was 175 ng/L, with a standard deviation of 189 ng/L.
In seas,
Model selection
DIC values for each pair of BQR and BCPQR models at each regression quantile in lakes or seas are summarized in Table 1. The DIC difference was calculated by subtracting the DIC value of the BQR model from the DIC value of BCPQR model. For all model pairs, the BCPQR model had a much smaller DIC value compared with the BQR model (DIC differences 47 for all comparisons; Table 1). According to Ribatet (2020), a DIC difference larger than 10 indicates that the model with a larger DIC value has no
Relationships at multiple regression quantiles
While a mean regression method, e.g., the piecewise linear regression used in Deng et al. (2020), focuses on the mean of DMS distribution, the BCPQR model revealed a more complete view of the CHL-DMS relationships at multiple regression quantiles (Fig. 2), thus allowing for a more thorough understanding of the response of DMS to changes in CHL across the DMS distribution.
In practice, it is difficult to obtain the true upper boundary of the relationship due to the lack of data at the tail ends
Conclusions
Integrating the BCP model and the BQR model, we proposed a novel BCPQR model that was able to detect a change point in the QR. We employed the proposed approach to investigate the CHL-DMS relationship in aquatic ecosystems. We revealed new findings in the CHL-DMS relationship modeling, relationship differences between lakes and seas, and factors impacting the CHL-DMS relationship. We thereby concluded that the BCPQR model could indeed enhance the understanding of shifting CHL-DMS relationship.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
We would like to thank the editor and reviewers for their insightful and detailed comments and suggestions. We appreciate the kind help on the Bayesian quantile regression model from Dr. Qingrong ZOU (Beijing Information Science and Technology University), Dr. Shanshan WANG (Beihang University), and Dr. Yang YU (Beihang University). We are grateful to helpful suggestions in pre-reviews by Dr. Wei GAO (Guangdong University of Technology) and Dr. Christopher Thomas Filstrup (University of
References (105)
- et al.
A Bayesian changepoint–threshold model to examine the effect of TMDL implementation on the flow–nitrogen concentration relationship in the neuse river basin
Water Res.
(2011) - et al.
Global fields of sea surface dimethylsulfide predicted from chlorophyll, nutrients and light
J. Mar. Syst.
(2001) Quantile regression neural networks: implementation in r and application to precipitation downscaling
Comput. Geosci.
(2011)- et al.
An MCMC approach to classical estimation
J. Econom.
(2003) - et al.
Quantile regression analysis as a predictive tool for lake macroinvertebrate biodiversity
Ecol. Indic.
(2016) - et al.
A remote sensing algorithm for planktonic dimethylsulfoniopropionate (DMSP) and an analysis of global patterns
Remote Sens. Environ.
(2015) - et al.
Reduced sulfur gases in saline lakes of the Vestfold Hills, Antarctica
Palaeogeogr Palaeoclimatol Palaeoecol
(1991) - et al.
On the usefulness of prediction intervals for local species distribution model forecasts
Ecol. Inform.
(2018) - et al.
Combining best professional judgement and quantile regression splines to improve characterisation of macrofaunal responses to enrichment
Ecol. Indic.
(2012) - et al.
Using Bayesian change point model to enhance understanding of the shifting nutrients-phytoplankton relationship
Ecol. Model.
(2019)
A framework to develop joint nutrient criteria for lake eutrophication management in eutrophic lakes
J. Hydrol. (Amst)
Global distribution of dissolved organic matter along the aquatic continuum: across rivers, lakes and oceans
Sci. Total Environ.
On Monte Carlo methods for Bayesian inference
Ecol. Modell.
The case against climate regulation via oceanic phytoplankton sulphur emissions
Nature
Bayesian change point analysis of abundance trends for pelagic fishes in the upper San Francisco Estuary
Ecol. Appl.
Evaluating spatial variability in sediment and phosphorus concentration-discharge relationships using Bayesian inference and self-organizing maps
Water Resour. Res.
Modeling spatially varying landscape change points in species occurrence thresholds
Ecosphere
Precipitation gradient determines the tradeoff between soil moisture and soil organic carbon, total nitrogen, and species richness in the loess plateau, china
Sci. Total Environ.
Modeling short-term post-offering price-volume relationships using Bayesian change-point panel quantile regression
Appl. Stoch. Model. Bus. Ind.
Quantile regression improves models of lake eutrophication with implications for ecosystem-specific management
Freshw. Biol.
Spatial distributions of dimethylsulfide in the South China Sea
Deep Sea Res. Part I
Biogeochemistry of dimethylsulfide (DMS) and dimethylsulfoniopropionate (DMSP) in the surface microlayer of the western North Pacific
Deep Sea Res. Part I
Biogenic emission of dimethylsulfide (DMS) from the North Yellow Sea, China and its contribution to sulfate in aerosol during summer
Atmos. Environ.
Temporal and spatial variations of dimethylsulfide (DMS) and dimethylsulfoniopropionate (DMSP) in the east china sea and the yellow sea
Cont. Shelf Res.
Circumpolar analysis of the adélie penguin reveals the importance of environmental variability in phenological mismatch
Ecology
Brq: an r package for Bayesian quantile regression
Working Paper
Breaking the sticks: a hierarchical change-point model for estimating ontogenetic shifts with stable isotope data
Methods Ecol. Evol.
Energetic and ecological constraints on population density of reef fishes
Proceed. Roy. Soc. B
A Bayesian analysis for change point problems
J. Am. Stat. Assoc.
The cycling of sulfur in surface seawater of the northeast pacific
J. Geophys. Res.
Intercomparison of homogenization techniques for precipitation data continued: comparison of two recent Bayesian change point models
Water Resour. Res.
Bayesian change-point analyses in ecology
New Phytol.
BayesQR: a Bayesian approach to quantile regression
J. Stat. Softw.
Predicting the frequency of water quality standard violations: a probabilistic approach for TMDL development
Environ. Sci. Technol.
General methods for monitoring convergence of iterative simulations
J. Comput. Graph. Stat.
A gentle introduction to quantile regression for ecologists
Front. Ecol. Environ.
Estimating fish body condition with quantile regression
North Am. J. Fish. Manage.
Estimating effects of limiting factors with regression quantiles
Ecology
Change points of global temperature
Environ. Res. Lett.
Oceanic phytoplankton, atmospheric sulphur, cloud albedo and climate
Nature
Coral reef aerosol emissions in response to irradiance stress in the Great Barrier Reef, Australia
Ambio
Dimethylsulfoniopropionate biosynthesis in marine bacteria and identification of the key gene in this process
Nat. Microbiol.
Quantile regression
Nat. Methods
Eco-chemical mechanisms govern phytoplankton emissions of dimethylsulfide in global surface waters
Natl. Sci. Rev.
An introduction to Bayesian inference for ecological research and environmental decision-making
Ecol. Appl.
Bayesian inference in ecology
Ecol. Lett.
bcp: an r package for performing a Bayesian analysis of change point problems
J. Stat. Softw.
SIP Metagenomics identifies uncultivated methylophilaceae as dimethylsulphide degrading bacteria in soil and lake sediment
ISME J.
Evidence for regional nitrogen stress on chlorophyll a in lakes across large landscape and climate gradients
Limnol. Oceanogr.
Arsenic, barium, germanium, tin, dimethylsulfide and nutrient biogeochemistry in charlotte harbor, florida, a phosphorus-enriched estuary
Deep Sea Res. Part B
Cited by (8)
Eutrophication-driven infochemical dimethylsulfide accelerates carbon transfer in freshwater food chain
2024, Journal of Environmental ManagementTracking changes in opioid prescriptions dispensed following the enactment of a prescription drug monitoring program use mandate
2023, Research in Social and Administrative PharmacyA statistical framework to track temporal dependence of chlorophyll–nutrient relationships with implications for lake eutrophication management
2021, Journal of HydrologyCitation Excerpt :Considering the lack of statistical metrics to determine trend convergence of accumulative CNRs in our framework, a reliable indicator to better assess the trend convergence is an area of future research. The method applied in the framework can also be extended to other methods, such as a Bayesian QR approach (Liang et al., 2021a) and QR neural network (Cannon, 2011). The contribution of this study is the proposed statistical framework which can reveal the temporal dependence of CNR.
Approaching the upper boundary of driver-response relationships: identifying factors using a novel framework integrating quantile regression with interpretable machine learning
2023, Frontiers of Environmental Science and EngineeringAssociation of lifestyle behaviors with health-related quality of life among patients with hematologic diseases
2023, Quality of Life Research