State-Space Models for Biological Monitoring Data Devin S

State-Space Models for Biological Monitoring Data Devin S. Johnson University of Alaska Fairbanks and Jennifer A. Hoeting Colorado State University

The work reported here was developed under the STAR Research Assistance Agreement CR-829095 awarded by the U. S. Environmental Protection Agency (EPA) to Colorado State University. This presentation has not been formally reviewed by EPA. The views expressed here are solely those of presenter and the STARMAP, the Program he represents. EPA does not endorse any products or commercial services mentioned in this presentation. # CR - 829095

Outline n Biological monitoring data n Previous methods n Bayesian hierarchical models n previous models n Multiple trait models n Continuous trait models n Analysis of fish functional traits

Biological Monitoring Data n n n Organisms are sampled at several sites. “Individuals” are classified according to a set F of traits Example: Longevity Trophic guild 1. 6 years 1. herbivore 2. > 6 years 2. omivore 3. invertivore 4. picsivore Individual response vector:

Environmental Conditions n A set, , of site specific environmental measurements (covariates) are also typically recorded. Example: stream order, watershed area, elevation n Let Denote the vector of environmental covariates for a single sampling site

Functional Trait vs. Species Analysis n Distributions of functional traits are often more interesting Species are geographically constrained n Limited ecological inference n n Analysis of functional traits is portable n Functional traits allow inference of the biological root of species distribution and environmental adaptation.

Previous Functional Trait Analysis Methods n Ordination methods n Canonical Correspondence Analysis (CCA) (ter Braak, 1985) n Ordinate traits along a set of environmental axes n Product moment correlations n “Solution to the 4 th Corner Problem” (Legendre et al. 1997) n Estimate correlation measure between trait counts and environmental covariates

Shortcomings of previous methods n Measure marginal association between environmental variables and traits n Conditional relationships give a more detailed measure of association n Interaction between traits can give a different view n No predictive ability n Cannot predict community structure at a site using remotely sensed covariates (GIS)

State-space models for a single trait Billhiemer and Guttorp (1997) Csi = number of individuals belonging to category i at site s = 1, …, S n xs = site specific environmental covariate n Parameter estimation using a Gibbs sampler n

Extending the Billheimer-Guttorp model n Generalize the BG model to explicitly allow for multiple trait inference n Allow for a range of trait interaction n Parameterize to allow parsimonious modeling n BG model based on random effect categorical data models n Use graphical model structure with random effects n Allow inference for trait interactions

Multiple trait analysis Notation: i I Psi Csi f a Realization of Y (cell) Sample space of Y (not necessarily I 1×…×I|F|) Probability density of Y at site s = 1, …, S (cell probability) number of individuals of type i at site s (cell count) Single trait (f F) Subset of traits ( a F )

Bayesian hierarchical model n Data model: where n Parameter model:

Interaction parameterization n n and traits in a measure interaction between the For model identifiability choose reference cell i* and set:

Conditional independence statements n If I = I 1×…×I|F|, then if n For certain model specifications: if

Interaction example n Data: F = {1, 2}; Y = {Y 1, Y 2}; no covariates n Saturated model: n Conditional independence model: implies Y 1 Y 2 | e (and in this case Y 1 Y 2 )

Continuous traits n In addition, for each individual, a set, G, of continuous traits are measured Example: Shape = Body length / Body depth (How hydrodynamic is the individual? ) n Individual response vector: YG R|G| is a vector of interval valued traits n (i, y) represents a realization of Y n

Conditional Gaussian distribution n The conditional Gaussian distribution (Lauritzen, 1996, Graphical Models) n η(a) and Ψ(a) measure interactions between discrete and continuous traits n Homogeneous CG: = 0 for a

Random effects CG Regression where, n Reference cell identifiability constraints imposed n Conditional independence inferred from zero-valued parameters and random effects

RECG Hierarchical model However, note the simplification

Parameter estimation n A Gibbs sampling approach is used for parameter estimation Analyze (b, e, T) and (η, Ψ, d, K) with 2 separate Gibbs chains n The CG to Mult×N formulation of the likelihood n Independent priors for (b, e, T) and (η, Ψ, d, K) Problem: n Rich random effects structure can lead to poor convergence n Solution: Hierarchical centering

Hierarchical centering n n b(a), Ta and Ka have closed form full conditional distributions l and η need to be updated with a Metropolis step in the Gibbs sampler.

Fish species trait richness n n 119 stream sites visited in an EPA EMAP study Discrete traits: Longevity 1. 6 years 2. > 6 years n n Trophic guild 1. herbivo re 2. omivore 3. invertivor e 4. picsivore Continuous trait: Shape factor = Body length / Body depth i* = ( 6 years, Herbivore) = (1, 1)

Stream covariates Environmental covariates: values were measured at each site for the following covariates 1. 2. 3. 4. 5. Stream order Minimum watershed elevation Watershed area % area impacted by human use Areal % fish cover

Fish trait richness model n Interaction models: n Random effects:

Environment effects on Longevity Table 1. Comparison of “null” model to model including specified covariate for Longevity. Values presented are ≈2 ln(BF). Covariate Stream order Elevation Area Use Fish cover > 6 years -0. 588 ( ) 4. 139 3. 022 7. 319 5. 032

Environmental effects on Trophic Guild Table 2. Comparison of “null” model to model including specified covariate for trophic guild. Values presented are ≈2 ln(BF). Trophic Guild Covariate Omnivore Invertivore Piscivore Stream order 5. 550 5. 393 3. 538 Elevation -0. 824 ( ) 2. 166 6. 368 Area 4. 863 7. 142 6. 292 Use 5. 498 7. 241 0. 704 ( ) Fish cover 7. 031 5. 487 6. 351

Environmental effects on Trophic Guild Table 3. Comparison of “null” model to model including specified covariate for Shape. Values presented are ≈2 ln(BF). Covariate Stream order Elevation Area % impacted Fish cover Shape 8. 116 8. 228 8. 225 8. 249 7. 636

Trait interaction Table 4. Comparison of “null” model to model including interaction parameters. Interaction 2 ln(BF) {L, T} -18. 492 {L, S} -52. 183 {T, S} -104. 348 {L, T, S} -126. 604

Comments / Conclusions n Multiple traits can be analyzed with specified interaction n Continuous traits can also be included n Markov Random Field interpretation for trait interactions n BG model obtainable for multi-way traits n Allow full interaction and correlated R. E. n MVN random effects imply that the cell probabilities have a “constrained” LN distribution