Upload
benedict-simon
View
213
Download
0
Tags:
Embed Size (px)
Citation preview
Mechanistic models for macroecolgy:
moving beyond correlationNicholas J. Gotelli
Department of BiologyUniversity of VermontBurlington, VT 05405
??What causes geographic variation
in species richness??
Understanding species richness patterns
• Data sources
• A critique of current methods
• Range cohesion and the mid-domain effect
• Mechanistic models for species richness
• Model selection
• Summary
Nicholas Gotelli, University of Vermont
Gary EntsmingerAcquired Intelligence
Rob ColwellUniversity of Connecticut
Gary GravesSmithsonian
Carsten RahbekUniversity of Copenhagen
Thiago Rangel
Federal University of Goiás
Understanding species richness patterns
• Data sources
• A critique of current methods
• Range cohesion and the mid-domain effect
• Mechanistic models for species richness
• Model selection
• Summary
Data sources
• Gridded map of domain
Avifauna of South America
“There can be no question, I think, that South America is the most peculiar of all the primary regions of the globe as to its ornithology.” P.L. Sclater (1858)
South American Avifauna
• 2891 breeding species
• 2248 species endemic to South America and associated land-bridge islands
Minimum:
18 species
Minimum:
18 species
Maximum:
846 species
Data sources
• Gridded map of domain
• Species occurrence records within grid cells
Geographic Ranges For Individual
Species
Myiodoorus cardonai
Phalacrocorax brasilianus
Anas puna
Geographic Ranges
Species Richness
Geographic Ranges
Species Richness
Data sources
• Gridded map of domain
• Species occurrence records within grid cells
• Quantitative measures of potential predictor variables within grid cells (NPP, temperature, habitat diversity)
Climate, Habitat Variables Measured at Grid Cell Scale
Understanding species richness patterns
• Data sources
• A critique of current methods
• Range cohesion and the mid-domain effect
• Mechanistic models for species richness
• Model selection
• Summary
How are these macroecological data typically analyzed?
Net Primary Productivity (Tonnes/hectare)
Ob
serv
ed
Sp
eci
es
Ric
hn
ess
0 2 4 6 8 10 12 14
02
00
40
06
00
80
0
Net Primary Productivity (Tonnes/hectare)
Ob
serv
ed
Sp
eci
es
Ric
hn
ess
0 2 4 6 8 10 12 14
02
00
40
06
00
80
0
OLS
Net Primary Productivity (Tonnes/hectare)
Ob
serv
ed
Sp
eci
es
Ric
hn
ess
0 2 4 6 8 10 12 14
02
00
40
06
00
80
0
OLS
LOESS
Poisson
How are these macroecological data typically analyzed?
Curve-fitting!
Criticisms of Curve-Fitting
• “Correlation does not equal causation”
Criticisms of Curve-Fitting
• “Correlation does not equal causation”Common to all of macroecology!
Criticisms of Curve-Fitting
• “Correlation does not equal causation”Common to all of macroecology!
• Non-linearity & non-normal, spatially correlated errors
Criticisms of Curve-Fitting
• “Correlation does not equal causation”Common to all of macroecology!
• Non-linearity & non-normal, spatially correlated errorsLOESS, Poisson, Spatial Regression (SAM)
Criticisms of Curve-Fitting
• “Correlation does not equal causation”Common to all of macroecology!
• Non-linearity & non-normal, spatially correlated errorsLOESS, Poisson, Spatial Regression (SAM)
• Choosing among correlated predictor variables
Criticisms of Curve-Fitting
• “Correlation does not equal causation”Common to all of macroecology!
• Non-linearity & non-normal, spatially correlated errorsLOESS, Poisson, Spatial Regression (SAM)
• Choosing among correlated predictor variablesModel selection strategies, stepwise regression, AIC
Criticisms of Curve-Fitting
• “Correlation does not equal causation”Common to all of macroecology!
• Non-linearity & non-normal, spatially correlated errorsLOESS, Poisson, Spatial Regression (SAM)
• Choosing among correlated predictor variablesModel selection strategies, stepwise regression, AIC
• Sensitivity to spatial scale, taxonomic resolution, geographic range size
Criticisms of Curve-Fitting
• “Correlation does not equal causation”Common to all of macroecology!
• Non-linearity & non-normal, spatially correlated errorsLOESS, Poisson, Spatial Regression (SAM)
• Choosing among correlated predictor variablesModel selection strategies, stepwise regression, AIC
• Sensitivity to spatial scale, taxonomic resolution, geographic range sizeStratify analysis
Conceptual Weakness of Curve-Fitting Paradigm
Predicted Species Richness
(S / grid cell)
Potential Predictor Variables
(tonnes/ha, C°)
Observed Species Richness
(S / grid cell)
Conceptual Weakness of Curve-Fitting Paradigm
Predicted Species Richness
(S / grid cell)
Potential Predictor Variables
(tonnes/ha, C°)
Observed Species Richness
(S / grid cell)
minimizeresiduals
Conceptual Weakness of Curve-Fitting Paradigm
Predicted Species Richness
(S / grid cell)
Potential Predictor Variables
(tonnes/ha, C°)
Observed Species Richness
(S / grid cell)
??MECHANISM
??
minimizeresiduals
ExplicitSimulation
Model
Alternative Strategy:Mechanistic Simulation Models
Predicted Species Richness
(S / grid cell)
Potential Predictor Variables
(tonnes/ha, C°)
Observed Species Richness
(S / grid cell)
ExplicitSimulation
Model
Alternative Strategy:Mechanistic Simulation Models
Predicted Species Richness
(S / grid cell)
Potential Predictor Variables
(tonnes/ha, C°)
Observed Species Richness
(S / grid cell)
mechanism
How can we build explicit simulation models for
macroecology?
Understanding species richness patterns
• Data sources
• A critique of current methods
• Range cohesion and the mid-domain effect
• Mechanistic models for species richness
• Model selection
• Summary
One-dimensional geographic domain
One-dimensional geographic domain
Species geographic ranges randomly placed line segments within domain
One-dimensional geographic domain
Species geographic ranges randomly placed line segments within domain
Peak of species richness in geographic center of domain
One-dimensional geographic domain
Species geographic ranges randomly placed line segments within domain
Peak of species richness in geographic center of domain
Species
Number
domain
domain
geographic range
der PfankuchenGuild
Pancakus spp.
Reduced species richnessat margins of the domain
Mid-domainpeak of species richnessin the center of the domain
2-dimensional MDE Model
• Random point of originationwithin continent (speciation)
• Random spread of geographicrange into contiguousunoccupied cells
• Spreading dye model (Jetz & Rahbek 2001) predicts peak richness incenter of continent (r2 = 0.17)
Assumptions of MDE models
• Placement of ranges within domain is random with respect to environmental gradients– Controversial, but logical for a null model for
climatic effects
Assumptions of MDE models
• Placement of ranges within domain is random with respect to environmental gradients– Controversial, but logical for a null model for
climatic effects
• Geographic ranges are cohesive within the domain– Rarely discussed, but important as the basis
for a mechanistic model of species richness
Range Cohesion Range Scatter
At the 1º x 1º scale, > 95% of species of South American birds have contiguous
geographic ranges
Causes of Range Cohesion
• Extrinsic Causes
Causes of Range Cohesion
• Extrinsic Causes– Coarse Spatial Scale– Spatial Autocorrelation in Environments
Causes of Range Cohesion
• Extrinsic Causes– Coarse Spatial Scale– Spatial Autocorrelation in Environments
• Intrinsic Causes
Causes of Range Cohesion
• Extrinsic Causes– Coarse Spatial Scale– Spatial Autocorrelation in Environments
• Intrinsic Causes– Limited Dispersal– Philopatry & Site Fidelity– Metapopulation & Source/Sink Structure– Fine-scale Genetic Structure & Local Adaptation– Spatially Mediated Species Interactions
Strict Range Cohesion Stepping Stone
* The mid-domain effect does not require strict range cohesion. A mid-domain peak in species richness will also arise from stepping stone models with limited dispersal and from neutral model dynamics (Rangel & Diniz-Filho 2005)
Homogenous Environment
HeterogeneousEnvironment
Almost all MDE models have assumed a homogeneous environment: grid cells are equiprobable
Enforced
Relaxed
Homogeneous Heterogeneous
RANGECOHESION
ENVIRONMENT
Enforced
Relaxed
Homogeneous Heterogeneous
RANGECOHESION
ENVIRONMENT
Classic MDE
Statistical Null(slope = 0)
Enforced
Relaxed
Homogeneous Heterogeneous
RANGECOHESION
ENVIRONMENT
Classic MDE
Statistical Null(slope = 0)
Enforced
Relaxed
Homogeneous Heterogeneous
RANGECOHESION
ENVIRONMENT
Classic MDE
Statistical Null(slope = 0)
Range ScatterModels
Range Cohesion Models
Enforced
Relaxed
Homogeneous Heterogeneous
RANGECOHESION
ENVIRONMENT
Classic MDE
Statistical Null(slope = 0)
Range ScatterModels
Range Cohesion Models
Range Cohesion Models are a hybrid that describes a stochastic MDE model in a more realistic heterogeneous environment.
Range Scatter Models also incorporate environmental heterogeneity, but do not place any constraints on species geographic ranges.
ExplicitSimulation
Model
Alternative Strategy:Mechanistic Simulation Models
Predicted Species Richness
(S / grid cell)
Potential Predictor Variables
(tonnes/ha, C°)
Observed Species Richness
(S / grid cell)
mechanism
Understanding species richness patterns
• Data sources
• A critique of current methods
• Range cohesion and the mid-domain effect
• Mechanistic models for species richness
• Model selection
• Summary
Modeling Strategy
• Establish simple algorithms that describe P(occupancy) based on environmental variables
Modeling Strategy
• Establish simple algorithms that describe P(occupancy) based on environmental variables
• Simulate origin and placement of each species geographic range in heterogeneous landscape (with or without range cohesion)
Modeling Strategy
• Establish simple algorithms that describe P(occupancy) based on environmental variables
• Simulate origin and placement of each species geographic range in heterogeneous landscape (with or without range cohesion)
• Repeat simulation to estimate predicted species richness per grid cell
Geographic Ranges
Species Richness
What determines P(cell occurrence)?
• Simple environmental modelsP(occurrence) Measured Environmental
Variable (NPP, Temperature, etc.)
What determines P(cell occurrence)?
• Simple environmental modelsP(occurrence) Measured Environmental
Variable (NPP, Temperature, etc.)
• Formal analytical models
What determines P(cell occurrence)?
• Simple environmental modelsP(occurrence) Measured Environmental
Variable (NPP, Temperature, etc.)
• Formal analytical models– Species-Energy Model (Currie et al. 2004)– Temperature Kinetics (Brown et al. 2004)
What determines P(cell occurrence)?
• Simple environmental modelsP(occurrence) Measured Environmental
Variable (NPP, Temperature, etc.)
• Formal analytical models– Species-Energy Model (Currie et al. 2004)
P(occurrence) (NPP)(Grid-cell Area)– Temperature Kinetics (Brown et al. 2004)
P(occurrence) e-E/kT
Understanding species richness patterns
• Data sources
• A critique of current methods
• Range cohesion and the mid-domain effect
• Mechanistic models for species richness
• Model selection
• Summary
Model-Selection in Curve-Fitting Analyses
• Simple tests against the null hypothesis that b=0
• No consideration of what expected slope should be with a specific mechanism
• Least-square and AIC criteria to try and select a subset of variables that best account for variation in S
Net Primary Productivity (Tonnes/hectare)
Ob
serv
ed
Sp
eci
es
Ric
hn
ess
0 2 4 6 8 10 12 14
02
00
40
06
00
80
0
OLS
H0: b = 0
Model Selection with Mechanistic Simulation Models
• Models make quantitative predictions of expected species richness
• Test slope of observed richness versus predicted richness
• Hypothesis of an acceptable fit H1: b = 1.0
• Rank acceptable models according to slope, intercept, and r2
• AIC criteria not appropriate
Predicted S
Observed S
Theoretical b = 1.0
Observed b
Understanding species richness patterns
• Data sources
• A critique of current methods
• Range cohesion and the mid-domain effect
• Mechanistic models for species richness
• Model selection
• Summary
Summary
• Curve-fitting framework does not incorporate explicit mechanisms
Summary
• Curve-fitting framework does not incorporate explicit mechanisms
• Use mechanistic simulations to define the placement of geographic ranges in a gridded domain
Summary
• Curve-fitting framework does not incorporate explicit mechanisms
• Use mechanistic simulations to define the placement of geographic ranges in a gridded domain
• Specify rules for P(occurrence)= f(environmental variables)
Summary
• Curve-fitting framework does not incorporate explicit mechanisms
• Use mechanistic simulations to define the placement of geographic ranges in a gridded domain
• Specify rules for P(occurrence)= f(environmental variables)
• Test model fit against expected slope = 1.0
Criticisms & Rejoinders
Criticisms & Rejoinders
• “Each species has a unique and distinctive response to different environmental variables. Species ranges should be modeled independently, not with a single function for all species.”
Criticisms & Rejoinders
• “Each species has a unique and distinctive response to different environmental variables. Species ranges should be modeled independently, not with a single function for all species.”
If this is true, why are there widespread repeatable patterns of species richness (e.g., latitude, elevation, area, productivity)?
Criticisms & Rejoinders
• “Each species has a unique and distinctive response to different environmental variables. Species ranges should be modeled independently, not with a single function for all species.”
If this is true, why are there widespread repeatable patterns of species richness (e.g., latitude, elevation, area, productivity)?
Often not enough data to model each species individually. We need a simple framework for analysing entire floras and faunas at a biogeographic scale.
Criticisms & Rejoinders
• “1:1 scaling of environmental variables with P(occurrence) is unrealistic and arbitrary.”
Criticisms & Rejoinders
• “1:1 scaling of environmental variables with P(occurrence) is unrealistic and arbitrary.”
Perhaps, but this is a parsimonious mechanistic model that relates environmental variables to geographic range placement.
Criticisms & Rejoinders
• “1:1 scaling of environmental variables with P(occurrence) is unrealistic and arbitrary.”
Perhaps, but this is a parsimonious mechanistic model that relates environmental variables to geographic range placement.
Linearity in P(occurrence) is not unreasonable over the empirical ranges of environmental variables measured in South America. (Linearity of P(occurrence) ≠ Linearity of (Species Richness))
Criticisms & Rejoinders
• “1:1 scaling of environmental variables with P(occurrence) is unrealistic and arbitrary”
Perhaps, but this is a parsimonious mechanistic model that relates environmental variables to geographic range placement.
Linearity in P(occurrence) is not unreasonable over the empirical ranges of environmental variables measured in South America. (Linearity of P(occurrence) ≠ Linearity of (Species Richness))
Mechanistic models are scarce in this literature (n = 2)! We have to begin somewhere!
Criticisms & Rejoinders
• “Many environmental variables, but especially NPP, show non-linear relationships with peaks in richness at intermediate levels. This is not captured by linear models.”
Criticisms & Rejoinders
• “Many environmental variables, but especially NPP, show non-linear relationships with peaks in richness at intermediate levels. This is not captured by linear models.”
At least at this spatial scale, no evidence for a diversity hump of avian species richness when plotted with NPP or other variables
Criticisms & Rejoinders
• “Using slopes comparisons will not successfully distinguish between models with intercorrelated predictor variables.”
Criticisms & Rejoinders
• “Using slopes comparisons will not successfully distinguish between models with intercorrelated predictor variables.”
Not a problem for these analyses. From an initial set of ~ 100 candidate models (10 variables x 2 algorithms x 5 range size quartiles), we reduced the set down to only 4 or 5 possible contenders.
Criticisms & Rejoinders
• “The model is not truly mechanistic because it does not model the sizes of the geographic ranges, only their placement.”
Criticisms & Rejoinders
• “The model is not truly mechanistic because it does not model the sizes of the geographic ranges, only their placement.”
True! Our model takes range sizes as a given and then uses algorithms to place them in a heterogeneous domain. A more realistic model would describe the processes of speciation, dispersal, and extinction of an evolving fauna.
Criticisms & Rejoinders
• “The model is not truly mechanistic because it does not model the sizes of the geographic ranges, only their placement.”
True! Our model takes range sizes as a given and then uses algorithms to place them in a heterogeneous domain. A more realistic model would describe the processes of speciation, dispersal, and extinction of an evolving fauna.
But how can the parameters of such a model (e.g. speciation and dispersal rates) ever be measured in the real world? Same problems have plagued most empirical evaluations of the neutral model.
Criticisms & Rejoinders
• “The model is not truly mechanistic because it does not model the sizes of the geographic ranges, only their placement.”
True! Our model takes range sizes as a given and then uses algorithms to place them in a heterogeneous domain. A more realistic model would describe the processes of speciation, dispersal, and extinction of an evolving fauna.
But how can the parameters of such a model (e.g. speciation and dispersal rates) ever be measured in the real world? Same problems have plagued most empirical evaluations of the neutral model.
Our models are designed to analyze the data that macroecologists typically have: gridded maps of environmental variables and species geographic ranges.
Criticisms & Rejoinders
• “The range cohesion and range scatter models don’t’ seem like they would give predictions that are any different from just a regression with the underlying variables themselves. What is the added value of these simulation models?”
Criticisms & Rejoinders
• “The range cohesion and range scatter models don’t’ seem like they would give predictions that are any different from just a regression with the underlying variables themselves. What is the added value of these simulation models?”
The predictions are not the same. For species with large geographic ranges, the range cohesion models always fit the data better than the range scatter models, regardless of which environmental variable is considered.
Key Differences
Curve-Fitting Mechanistic Models
Unit of Study Species Richness Underlying geographic ranges
Predicted values Minimization of residuals (data dependent)
Algorithms for origin and spread of geographic ranges (data independent)
Model Selection Criteria Smallest number of variables that reduce residual sum of squares
Quantitative fit to model predictions
Statistical Tests H0: (b = 0) tests for any effect that is larger than 0
H0: (b = 1.0) tests for quantitative match between observed and predicted S
To Be Continued…
Carsten Rahbek. Perception of Species Richness Patterns:
The Role of Range Sizes