Planning the next generation general population assessment model Mark Maunder (IATTC) and Simon...
-
Upload
sophia-dennis -
Category
Documents
-
view
215 -
download
3
Transcript of Planning the next generation general population assessment model Mark Maunder (IATTC) and Simon...
Planning the next generation general population assessment
model
Mark Maunder (IATTC)
and
Simon Hoyle (SPC)
Outline• Why we need a new general model
• Advantages of a general model
• Existing general models
• Important features of the next generation general model
• Features required for protected species
• Issues with developing a general model
• Summary
Recent advances
• Improved computer performance
• Parallel processing and distributed computing
• Automatic differentiation and MCMC
• Convergence of approaches towards integrated population dynamics modeling
Why we need a new general model
• Too many populations to assesses
• Not enough qualified analysts
• Common language
• Current models are reaching their limitations
• Fit to data
Common language
• Facilitates discussions
• Easier to review– use of SS2 in west coast STAR panel process
and Pacific cod assessment
• Comprehensive analysis and testing to develop best practices
• Focuses development
• Reduces duplication
Advantages of a general model
• Less development time
• Tested code
• Familiarity
• Diagnostics and output
Existing general models• Stock assessment
– Coleraine– MULTIFAN-CL– SS1/SS2– CASAL– Gadget– Xsurvivers– ADAPT
Table of model comparisonsA-SCALA MFCL SS2 CASAL
Approach ADMB AUTODIFF ADMB BETADIFF
Normal approx Yes Yes Yes Yes
Automatic profile likelihood Yes No No Yes
Bayesian MCMC No MCMC MCMC
MCMC practical for tuna No NA No No
Model uncertainty MCMC No No No No
Bootstrapping No No Automatic Automatic
Review Dual programming, comparisons with MFCL, publication review
Publication review, comparison with A-SCALA (no spatial or tagging)
Independent expert review, intensive reviews of applications, comparisons with other models, simulation tests
Comparison with Coleraine and other models, applications reviewed by independent experts
Assessments IATTC Assessments (YFT, BET, SKJ) and comparisons with WCPO
WCPO YFT BET, ALB, SKJ, BUM, SOW, Blue shark, Lobster
Atlantic BET ALB
15 west coast and Alaska groundfish assessments, SEPO swordfish
From 10 to 20 stocks in NZ and CCAMLR, fin fish and Shellfish
Max parameters estimated in application
2000 (RE) 3000 (RE) 200 200
Time required for tuna app 4 hrs 4 hrs 40 min4 Not evaluated
Model StructureStructure A-SCALA MFCL SS2 CASAL
Spatial No Yes Yes Yes
Fishing mortality Effort devs Effort devs Pope’s/catch Eq Pope’s
Seasons Restricted General General General
Modeling of discards No No Yes No
Sex structured No Under development Optional Optional
Growth morphs No No Yes Yes
multi-species No Under development/no predator prey
No Yes/no predator prey
Selectivity Smoothness penalties Functional forms, smoothness penalties, splines
Functional forms and nonparametric
Functional forms, smoothness penalties
Selectivity basis Age, length penalty Age or Length Age, length, and sex Age, length, partition
Time varying parameters Catchability Catchability All parameters Limited
Environment R and q R All parameters R (untested)
Stock-recruitment relationship
B-H B-H B-H, Ricker B-H, Ricker
M Full age-structure Full age-structure with smoothness
2 breakpoints Full age-structure with smoothness
Movement NA Transfer rates with implicit time steps
Transfer rates Transfer rates, density dependent
Aging error No No Yes Yes
Variable length bin size No No Yes Yes
Additional model structureQuestion A-SCALA MFCL SS2 CASAL
Recruitment deviates
Penalized likelihood
Penalized likelihood
Penalized likelihood/MCMC
Penalized likelihood/MCMC
Uncertainty Normal approximation but MCMC and profile likelihood possible but impractical
Normal approximation profile likelihood by hand and limited in practice
Normal approximation, MCMC, profile likelihood, bootstrap
Normal approximation, MCMC, profile likelihood, bootstrap
Covariate approach
Fit to index or as relationship
Undetermined Relationship Relationship
Projections Point estimates or likelihood based with normal approximation
Likelihood based with normal approximation
Likelihood based with normal approximation, MCMC
MCMC, point estimates, parametric or nonparametric recruitment
Weighting data sets
Estimate process error component
Spatial structure Only in fisheries In fisheries and population dynamics, uses tagging data
In fisheries and population dynamics, does not use tagging data
In fisheries and population dynamics, uses tagging data
Data typesData A-SCALA MFCL SS2 CASAL
Catch-effort Effort dev Effort dev Index Index
Catch-at-age √ √ √
Catch-at-length √ √ √ √
Abundance index √ √ √
Tagging √ √
Catch-at-weight √
Age-length √ √ √
Average weight √
Discard (fit) √
Proportions mature
√
Proportions migrating
√
Age at maturity √
Existing general models• Multi-species/Ecosystem
– Ecopath/Ecosim• Mark recapture
– MARK– M-SURGE– Barker’s Mother of All Models
• Wildlife– St Andrews state-space framework
• PVA– ALEX– RAMAS– VORTEX– GAPPS– INMAT
Existing general models• Multi-species
– Similar to integrated models
• Ecosystem– Simple structure and data use
• Mark recapture– Generally limited to mark-recapture data
• Wildlife– Only a framework, not a general model
• PVA– Not fit to data
State-space models• Models processes as probability
distributions• Not all SS models need to be integrated*
or Bayesian• Not all integrated* or Bayesian models
have to be SS• Most process variation is due to the
environment not demographic processes– Random effects
*Integrated in this context means use multiple data types
FLR (Fisheries Library in R)
• Collection of R tools that facilitate the construction of models representing fisheries and ecological systems.
• Focuses on evaluating fisheries management strategies• Includes several models for stock assessment and
simulation• Some components are written entirely in R, while others
use C++ or Fortran to accommodate existing programs or to recode programs for greater efficiency.
• (http://flr-project.org/doku.php, Kell et al. 2007)
Important features to consider for the next generation general model
• Integrated multiple data types• Priors• Include process error• spatial structure• Sub-population structure (as well as spatial structure)• Covariates• Age, length, stage, sex
• Multi-species• Meta analysis• Genetics
• Estimate uncertainty• Model selection and averaging
• Simulate data for model testing and MSE• Ability to include user defined functions• Ability to run each component of the model separately• MSE
• Abundance– Absolute or relative
• Composition– Age, length, stage, sex, weight, otolith size
• Aggregated
• Mark-recapture
• Archival tags
• Mortality/catch
• Future types of data
Data
Features required for protected species• Alternative stock-recruitment curves (density
dependence) mate pairing, widowing, skip breeding• Density dependence in other processes
– Survival– Movement
• Stage structure• Small population sizes
– Random variation in population processes• Mark-recapture data• Occupancy data• Minimum counts• Habitat data• Individual characteristics
Management strategy evaluation
• Data to collect
• Method to analyze data
• Management rule
• Evaluation criteria
• Operating models
Output
• Management quantities– MSY– Extinction risk– Projections
• Impact plots
• Diagnostics– Not well developed for integrated models
Some issues with developing a general model
• Tradeoff between generality and computational efficiency
• Using the model incorrectly
• Weighting of data sets
• Missing data in covariates
How to get it done• Open source and Free
– Create a community for development, testing, training, and assistance
• Collaboration– Expertise scattered among countries, organizations,
and disciplines – Efficient algorithms: statisticians and mathematicians– Efficient code: computer scientist– Appropriate statistical framework (e.g. likelihood
functions): statisticians– Population dynamics: ecologists and biologists
• Funding– Who will pay– Who will get paid
• Some experts do not have their salaries covered
Summary
• A general model is needed to fulfill management’s increasing needs, and to focus and accelerate research
• It will take a well planned collaboration from diverse disciplines
• Organizations are willing to fund it