Parameter Redundancy and Identifiability in Ecological Models

Slides:



Advertisements
Similar presentations
Pattern Recognition and Machine Learning
Advertisements

Modeling of Data. Basic Bayes theorem Bayes theorem relates the conditional probabilities of two events A, and B: A might be a hypothesis and B might.
Uncertainty and confidence intervals Statistical estimation methods, Finse Friday , 12.45–14.05 Andreas Lindén.
Workshop on Parameter Redundancy Part II Diana Cole.
The Problem with Parameter Redundancy Diana Cole, University of Kent.
Estimation  Samples are collected to estimate characteristics of the population of particular interest. Parameter – numerical characteristic of the population.
EARS1160 – Numerical Methods notes by G. Houseman
Detecting Parameter Redundancy in Ecological State-Space Models Diana Cole and Rachel McCrea National Centre for Statistical Ecology, University of Kent.
Parameter Redundancy and Identifiability in Ecological Models Diana Cole, University of Kent.
Parameter Redundancy and Identifiability Diana Cole and Byron Morgan University of Kent Initial work supported by an EPSRC grant to the National Centre.
Detecting Parameter Redundancy in Complex Ecological Models Diana Cole and Byron Morgan University of Kent.
Maximum likelihood estimates What are they and why do we care? Relationship to AIC and other model selection criteria.
Resampling techniques Why resampling? Jacknife Cross-validation Bootstrap Examples of application of bootstrap.
The Islamic University of Gaza Faculty of Engineering Civil Engineering Department Numerical Analysis ECIV 3306 Chapter 3 Approximations and Errors.
Parameter Redundancy in Ecological Models Diana Cole, University of Kent Byron Morgan, University of Kent Rachel McCrea, University of Kent Ben Hubbard,
Determining Parameter Redundancy of Multi-state Mark- Recapture Models for Sea Birds Diana Cole University of Kent.
Linear and generalised linear models
Linear and generalised linear models Purpose of linear models Least-squares solution for linear models Analysis of diagnostics Exponential family and generalised.
Classification and Prediction: Regression Analysis
Likelihood probability of observing the data given a model with certain parameters Maximum Likelihood Estimation (MLE) –find the parameter combination.
Statistical approaches to analyse interval-censored data in a confirmatory trial Margareta Puu, AstraZeneca Mölndal 26 April 2006.
Adding individual random effects results in models that are no longer parameter redundant Diana Cole, University of Kent Rémi Choquet, Centre d'Ecologie.
7-1 Introduction The field of statistical inference consists of those methods used to make decisions or to draw conclusions about a population. These.
Detecting Parameter Redundancy in Integrated Population Models Diana Cole and Rachel McCrea National Centre for Statistical Ecology, School of Mathematics,
Various topics Petter Mostad Overview Epidemiology Study types / data types Econometrics Time series data More about sampling –Estimation.
A Hybrid Symbolic-Numerical Method for Determining Model Structure Diana Cole, NCSE, University of Kent Rémi Choquet, Centre d'Ecologie Fonctionnelle et.
MECN 3500 Inter - Bayamon Lecture 3 Numerical Methods for Engineering MECN 3500 Professor: Dr. Omar E. Meza Castillo
Empirical Efficiency Maximization: Locally Efficient Covariate Adjustment in Randomized Experiments Daniel B. Rubin Joint work with Mark J. van der Laan.
Maximum Likelihood Estimation Methods of Economic Investigation Lecture 17.
Integrating archival tag data into stock assessment models.
Estimation of Number of PARAFAC Components
2005MEE Software Engineering Lecture 11 – Optimisation Techniques.
BRIEF INTRODUCTION TO ROBUST DESIGN CAPTURE-RECAPTURE.
Estimating age-specific survival rates from historical ring-recovery data Diana J. Cole and Stephen N. Freeman Mallard Dawn Balmer (BTO) Sandwich Tern.
Chapter 5 Sampling Distributions. The Concept of Sampling Distributions Parameter – numerical descriptive measure of a population. It is usually unknown.
Class 5 Multiple Regression CERAM February-March-April 2008 Lionel Nesta Observatoire Français des Conjonctures Economiques
Parameter Redundancy in Mark-Recapture and Ring-Recovery Models with Missing Data Diana Cole University of Kent.
Using distributions of likelihoods to diagnose parameter misspecification of integrated stock assessment models Jiangfeng Zhu * Shanghai Ocean University,
Hypothesis Testing. Statistical Inference – dealing with parameter and model uncertainty  Confidence Intervals (credible intervals)  Hypothesis Tests.
 Multi-state Occupancy. Multiple Occupancy States Rather than just presence/absence of the species at a sampling unit, ‘occupancy’ could be categorized.
Review. Common probability distributions Discrete: binomial, Poisson, negative binomial, multinomial Continuous: normal, lognormal, beta, gamma, (negative.
 Occupancy Model Extensions. Number of Patches or Sample Units Unknown, Single Season So far have assumed the number of sampling units in the population.
Presentation : “ Maximum Likelihood Estimation” Presented By : Jesu Kiran Spurgen Date :
Multi-state Occupancy. Multiple Occupancy States Rather than just presence/absence of the species at a sampling unit, ‘occupancy’ could be categorized.
STA248 week 121 Bootstrap Test for Pairs of Means of a Non-Normal Population – small samples Suppose X 1, …, X n are iid from some distribution independent.
Part 3: Estimation of Parameters. Estimation of Parameters Most of the time, we have random samples but not the densities given. If the parametric form.
Hierarchical Models. Conceptual: What are we talking about? – What makes a statistical model hierarchical? – How does that fit into population analysis?
Pharmacometrics research group Department of Pharmaceutical Biosciences Uppsala University Approximations of the population Fisher information matrix-
Stats Methods at IC Lecture 3: Regression.
Estimating standard error using bootstrap
Chapter 7. Classification and Prediction
Probability Theory and Parameter Estimation I
7-1 Introduction The field of statistical inference consists of those methods used to make decisions or to draw conclusions about a population. These.
CJT 765: Structural Equation Modeling
Chapter 6 Floating Point
Data Mining (and machine learning)
School of Mathematical Sciences, University of Nottingham.
Extension to the Hybrid Symbolic-Numeric Method for Investigating Identifiability Diana Cole, University of Kent, UK Rémi Choquet, CEFE, CNRS, France.
Estimating Population Size
More about Posterior Distributions
Modelling data and curve fitting
Geology Geomath Chapter 7 - Statistics tom.h.wilson
Where did we stop? The Bayes decision rule guarantees an optimal classification… … But it requires the knowledge of P(ci|x) (or p(x|ci) and P(ci)) We.
Multistate models Lecture 10.
PSY 626: Bayesian Statistics for Psychological Science
Chapter 4, Regression Diagnostics Detection of Model Violation
Econometrics Chengyuan Yin School of Mathematics.
Principles of the Global Positioning System Lecture 11
Maths for Signals and Systems Linear Algebra in Engineering Lecture 6, Friday 21st October 2016 DR TANIA STATHAKI READER (ASSOCIATE PROFFESOR) IN SIGNAL.
Learning From Observed Data
Presentation transcript:

Parameter Redundancy and Identifiability in Ecological Models Diana Cole, University of Kent

x Introduction Occupancy Model example Parameters: 𝜓 −site is occupied, 𝑝 – species is detected. Can only estimate 𝜓𝑝 rather than 𝜓 and 𝑝. Model is parameter redundant or parameters are non-identifiable. (Solution: MacKenzie et al, 2002, robust design; visit sites several times in one season). x Species present and detected Species present but not detected Species absent Prob =𝜓𝑝 Prob =𝜓 1−𝑝 Prob =1−𝜓 Prob not detected =𝜓 1−𝑝 +1−𝜓 =1−𝜓𝑝 Prob detected =𝜓𝑝

Parameter Redundancy and Identifiability Suppose we have a model 𝑀(𝜃) with parameters 𝜃. A model is globally (locally) identifiable if 𝑀 𝜃 1 =𝑀( 𝜃 2 ) implies that 𝜃 1 = 𝜃 2 (for a neighbourhood of 𝜃). A model is parameter redundant model if it can be reparameterised in terms of a smaller set of parameters. A parameter redundant model is non-identifiable. There are several different methods for detecting parameter redundancy, including: numerical methods (e.g. Viallefont et al, 1998), symbolic methods (e.g. Cole et al, 2010), hybrid symbolic-numeric method (Choquet and Cole, 2012). Generally involves calculating the rank of a matrix, which gives the number of parameters that can be estimated.

Problems with Parameter Redundancy There will be a flat ridge in the likelihood of a parameter redundant model (Catchpole and Morgan, 1997), resulting in more than one set of maximum likelihood estimates. Numerical methods to find the MLE will not pick up the flat ridge, although it might be picked up by trying multiple starting values and looking at profile log-likelihoods. If parameter redundant model is fitted will get biased parameter estimates. The Fisher information matrix will be singular (Rothenberg, 1971) and therefore the standard errors will be undefined. The exact Fisher information matrix is rarely known. Standard errors are typically approximated using a Hessian matrix obtained numerically. Can get explicit (but wrong) estimates of standard errors in some cases. Model selection based on identifiable models. (e.g. AIC needs number of parameters can estimate). Recommended you check identifiability of model before fitting (or at least as part of model fitting process).

Symbolic Method (Occupancy Example) Catchpole and Morgan (1997), Cole et al (2010) Exhaustive summary: 𝜿= 𝜓𝑝 1−𝜓𝑝 Parameters: 𝜽=[𝜓 𝑝] Derivative matrix: 𝑫= 𝜕𝜿 𝜕𝜽 = 𝑝 −𝑝 𝜓 −𝜓 Rank of D is 1. If rank less than number of parameters, model is parameter redundant. Solve set of PDEs to find whether we can estimate 𝜓𝑝 (estimable parameter combinations). Also able to generalise results (e.g. to multiple surveys). Requires symbolic algebra package, e.g. Maple, in most models to find rank. Can run out of memory finding rank in complex models. Extended in Cole et al (2010) to deal with complex models.

Numeric Methods (Occupancy Example) Hessian Method (Viallefont et al, 1998) Hessian matrix should be singular and therefore have at least one Eigenvalue that is 0. Found numerically the Hessian is not necessarily singular. Hessian: 𝒉= 180.25 179.99 179.99 179.74 Standardised Eigenvalues: 1, 0.000003 Eigenvalue close to zero, model is parameter redundant. Inaccurate - can give wrong result (e.g. Cole and Morgan, 2010) It is not able to find estimable parameter combinations. It is not able to find general results. Can be added to a software package (e.g. Mark). Other numerical methods include simulation, likelihood profile, data cloning.

Hybrid Symbolic-Numeric Method Choquet and Cole (2012) Find derivative matrix, D, symbolically (or using automatic differentiation). Find the rank of D at 5 random points, and the model rank is the maximum of all 5 ranks. Occupancy model: rank 1 in all 5 cases, therefore parameter redundant. Can also show if any of the original parameters can be estimated, but not the estimable parameter combinations. Can be added to a software package (e.g. E-Surge and M-Surge)

Identifiable Parameters Estimable Parameter Combinations Comparison of Methods Method Accurate Identifiable Parameters Estimable Parameter Combinations General Results Complex Models Automatic Numeric (eg Viallefont et al, 1998) ×  Symbolic (Catchpole and Morgan, 1997) Extended Symbolic (Cole et al, 2010) Hybrid Symbolic-Numeric (Choquet and Cole, 2012) Extended Hybrid ?

Occupancy Model - Likelihood Profile Flat profile = Parameter Redundant

Occupancy Model - Likelihood Profile Line is 𝑝= 𝛽 𝜓 = 0.5 𝜓 ⇒0.5=𝜓𝑝 Profile gives the estimable parameter combinations

Extended Hybrid Method – Subset profiling Eisenberg and Hayashi (2014) use subset profiling to identify estimable parameter combinations manually in compartment modelling. Combine hybrid method (Choquet and Cole, 2012) with subset profiling (Eisenberg and Hayashi, 2014) and extend to automatically detect relationships between confounded parameters. Step 1: Check whether model is parameter redundant using hybrid method and remove any parameters that are individually identifiable. Step 2: Find subsets of parameters that are confounded, using subset profiling. Looking for subsets of parameters which are nearly full rank. Step 3: For each subset, fix all the parameters not in the subset. Then for each pair of parameters, in the subset, produce a likelihood profile. Fix one parameter, 𝜃 1,𝑖 , and store the MLE of the other parameter, 𝜃 2,𝑖 (𝑖=1,…,𝑛).

Extended Hybrid Method – Subset profiling Step 4: Identify the relationships. Find 𝛽 1 , 𝛽 2 , 𝛽 3 , 𝛽 4 that minimise 𝑖=1 𝑛 𝜃 2,𝑖 − 𝛽 1 + 𝛽 2 𝜃 1,𝑖 𝛽 3 + 𝛽 4 𝜃 1,𝑖 2 to identify most common relationships. (Should equal 0). Occupancy Example: 𝜃 1 =𝜓, 𝜃 2 =𝑝, 𝛽 1 =0.5, 𝛽 2 =0, 𝛽 3 =0, 𝛽 4 =1 𝑝− 0.5+0×𝜓 0+1×𝜓 =𝑝− 0.5 𝜓 =0⇒𝑝𝜓=0.5 Step 5: Turn relationships between pairs of parameters to form estimable parameter combinations. Steps 1-4 can be programed to be automatic, step 5 not yet.

Mark-Recovery Example Animals marked and recovered dead. Probability animals marked in year 𝑖 and recovered in year 𝑗: 𝑃= 1− 𝜙 1,1 𝜆 1 𝜙 1,1 1− 𝜙 2 𝜆 2 𝜙 1,1 𝜙 2 1− 𝜙 3 𝜆 3 1− 𝜙 1,2 𝜆 1 𝜙 1,2 1− 𝜙 2 𝜆 2 1− 𝜙 1,3 𝜆 1 𝜙 1,𝑡 probability animal survives 1st year at time t. 𝜙 𝑖 probability an animal age 𝑖 survives their 𝑖th year. 𝜆 𝑖 probability animal aged 𝑖 is recovered dead. Model is parameter redundant with estimable parameter combinations 𝜙 1,1 , 𝜙 1,2 , 𝜙 1,3 , 𝜆 1 , 1− 𝜙 2 𝜆 2 , 𝜙 2 1− 𝜙 3 𝜆 3 (result found symbolically in Cole et al, 2012). 𝜙 1,1 , 𝜙 1,2 , 𝜙 1,3 , 𝜆 1 – individually identifiable (can still be estimated).

Mark-Recovery Example Model has parameters 𝜽= 𝜙 1,1 , 𝜙 1,2 , 𝜙 1,3 , 𝜙 2 , 𝜙 3 , 𝜆 1 , 𝜆 2 , 𝜆 3 Step 1: Hybrid method can be used to show model is parameter redundant, 𝜙 1,1 , 𝜙 1,2 , 𝜙 1,3 , 𝜆 1 are individually identifiable. Non-identifiable parameters are 𝜙 2 , 𝜙 3 , 𝜆 2 , 𝜆 3 . Step 2: Nearly full rank subsets are: { 𝜙 3 , 𝜆 3 },{ 𝜙 2 , 𝜙 3 , 𝜆 2 }, { 𝜙 2 , 𝜆 2 , 𝜆 3 },

Mark-Recovery Example Step 3: x profiles for subset { 𝜙 3 , 𝜆 3 }: Step 4: fitted relationship: 𝜆 3 − 0.0736−0.0000 𝜙 3 1.2267−1.2268 𝜙 3 =0⇒ 1− 𝜙 3 𝜆 3 =0.06

Mark-Recovery Example Step 3: x profiles for subset { 𝜙 2 , 𝜙 3 , 𝜆 2 } Step 4: fitted relationships: 𝜙 3 − −0.1800+0.9999 𝜙 2 0.0000+0.9999 𝜙 2 =0⇒ 𝜙 2 1− 𝜙 3 =0.18 𝜆 2 − 0.1200+0.0000 𝜙 2 0.9999−0.9999 𝜙 2 =0⇒ 1− 𝜙 2 𝜆 2 =0.12 𝜆 2 − 0.0013−0.0013 𝜙 3 0.0088−0.0106 𝜙 3 =0⇒ 𝜆 2 0.83− 𝜙 3 1− 𝜙 3 =0.12 (complex)

Mark-Recovery Example Step 3: x profiles for subset { 𝜙 2 , 𝜆 2 , 𝜆 3 } Step 4: fitted relationships: 𝜆 2 − 0.1200+0.0000 𝜙 2 0.9999−0.9999 𝜙 2 =0⇒ 1− 𝜙 2 𝜆 2 =0.12 𝜆 3 − 0.1472−0.0000 𝜙 2 0.0000+1.2267 𝜙 2 =0⇒ 𝜙 2 𝜆 3 =0.12 𝜆 3 − −0.0000+0.0110 𝜆 2 −0.0110+0.0913 𝜆 2 =0⇒ 𝜆 3 𝜆 2 0.12− 𝜆 2 =−0.12 (complex)

Mark-Recovery Example Step 5: There could be more than one way of parameterising the estimable parameter combinations. Looking for the simplest. From step 1 also can deduce we need to find 2 combinations of parameters. From step 2, using subsets that are not nearly full rank, we can deduce that 𝜙 2 , 𝜙 3 , 𝜆 3 are related (or 𝜙 3 , 𝜆 2 , 𝜆 3 ) Relationships that involve 𝜙 2 , 𝜙 3 , 𝜆 3 : 𝜙 2 1− 𝜙 3 =0.18, 𝜙 2 𝜆 3 =0.12, 1− 𝜙 3 𝜆 3 =0.06 are consistent with: 𝜙 2 1− 𝜙 3 𝜆 3 1− 𝜙 2 𝜆 2 =0.12 is consistent with: 1− 𝜙 2 𝜆 2 Estimable parameter combinations: 𝜙 2 1− 𝜙 3 𝜆 3 & 1− 𝜙 2 𝜆 2 The other (complex) relationships can be shown (with a bit of algebra) to be consistent with estimable parameter combinations.

General Results Symbolic method: general results found using the extension theorem (Catchpole and Morgan, 1997, Cole et al, 2010). This can be extended to the Hybrid method. Mark Recovery Example: Reparameterise in terms of estimable parameters, 𝜙 1,1 , 𝜙 1,2 , 𝜙 1,3 , 𝜆 1 , 𝑏 1 = 1− 𝜙 2 𝜆 2 and 𝑏 2 = 𝜙 2 1− 𝜙 3 𝜆 3 . Hybrid Method shows can estimate 𝜙 1,1 , 𝜙 1,2 , 𝜙 1,3 , 𝜆 1 , 𝛽 1 , 𝛽 2 . Add an extra year of ringing and recovery adds extra parameters 𝜙 1,4 and 𝑏 3 = 𝜙 2 𝜙 3 1− 𝜙 4 𝜆 4 . Hybrid method shows can also estimate 𝜙 1,4 , 𝑏 3 . By the extension theorem, for n years of ringing and recovery, model is always parameter redundant, estimable parameter combinations will be 𝜙 1,1 , 𝜙 1,2 ,…, 𝜙 1,𝑛 , 𝜆 1 , 1− 𝜙 2 𝜆 2 , 𝜙 2 1− 𝜙 3 𝜆 3 ,…, 𝑖=2 𝑛−1 𝜙 𝑖 1− 𝜙 𝑛 𝜆 𝑛

Identifiable Parameters Estimable Parameter Combinations Conclusion Method Accurate Identifiable Parameters Estimable Parameter Combinations General Results Complex Models Automatic Numeric (eg Viallefont et al, 1998) ×  Symbolic (Catchpole and Morgan, 1997) Extended Symbolic (Cole et al, 2010) Hybrid Symbolic-Numeric (Choquet and Cole, 2012) Extended Hybrid Semi

References Catchpole, E. A. and Morgan, B. J. T. (1997) Detecting parameter redundancy Biometrika, 84, 187-196 Choquet, R. and Cole, D.J. (2012) A Hybrid Symbolic-Numerical Method for Determining Model Structure. Mathematical Biosciences, 236, p117. Cole, D. J. and Morgan, B. J. T. (2010) A note on determining parameter redundancy in age-dependent tag return models for estimating fishing mortality, natural mortality and selectivity, JABES, 15, 431-434. Cole, D. J., Morgan, B. J. T., Catchpole, E. A. and Hubbard, B.A. (2012) Parameter redundancy in mark-recovery models, Biometrical Journal, 54, 507-523. Eisenberg, M. C. and Hayashi, M. A. L (2014) Determining identifiable parameter combinations using subset profiling, Mathematical Biosciences, 256, 116–126. MacKenzie, D. I., et al (2002) Estimating site occupancy rates when detection probabilities are less than one. Ecology, 83, 2248-2255. Viallefont, A., et al (1998) Parameter identifiability and model selection in capture-recapture models: A numerical approach. Biometrical Journal, 40, 313-325.