The covariation method of estimation Add_my_pet Dina Lika Dept of Biology Texel, 15/4/2013 UNIVERSITY OF CRETE Contents • The covariation method for parameter estimation – – – – – – – DEB parameters Auxiliary theory Real & pseudo data Zero & variate data Estimation criteria Numerical implementation Evaluation of the estimation The standard DEB model variables 1 food type, 1 reserve, 1 structure, isomorph Extended: V1-morphic early juvenile stage • structure, reserve, maturity, density of damage inducing compounds, and density of damage compounds parameters • Core parameters – Control changes of the state variables – Linked to the concepts on which the model is based on • Auxiliary parameters – Convert measurement (e.g. from dry to wet mass, length to volume etc.) – Quantify effects of temperature on rates and time • Primary parameters – Connected to a single underlying process • Compound parameters – Depend on several underlying processes Core parameters assimilation {pAm} feeding {Fm} digestion κX product formation κXP mobilisation v allocation reproduction R turnover,activity [pM] heating,osmosis {pT} development kJ Growth [EG] life cycle EHb life cycle EHj life cycle EHp aging ha aging sG z max surface-specific assim rate Lm ( 22.5z J cm-2 d-1) surface- specific searching rate (6.5 l d-1 cm-2) digestion efficiency (0.8) defecation efficiency (0.1) energy conductance (0.02 cm d-1) allocation fraction to soma (0.8) reproduction efficiency (0.95) volume-specific somatic maint. costs ( 18 d-1cm-3) surface-specific somatic maint. costs (0 d-1cm-2) maturity maintenance rate coefficient (0.002 d-1) specific growth for structure (2800 J cm-3) maturity at birth (0.275z3 J) maturity at metamorphosis ( z3 J) maturity at puberty (166z3 J) Weibul aging acceleration (10-6z d-2) Gompertz stress coefficient (0.01) zoom factor z= Lm / Lmref, with Lmref =1 maximum length Lm = {pAm} / [pM] Auxiliary parameters Conversion parameters δM dO =(dX, dV, dE, dP) μO =(μX, μV, μE, μP) μM =(μC, μH, μO, μN) nO =(nX, nV, nE, nP) nM =(nC, nH, nO, nN) wO=(12 1 16 14) nO shape coefficient (-) specific densities (g/cm3) chemical potentials (J/mol) chemical potentials (J/mol) chemical indices (-) chemical indices (-) molecular weights (-) Temperature parameters Tref TA TL, TH TAL, TAH reference temperature (273 K) Arrhenius temperature (8000 K) temperature tolerance range (277 K, 318 K) Arrhenius temperatures for transitions to inert state (20 kK, 190kK) Assumptions of auxiliary theory • A well-chosen physical length (volumetric) structural length for isomorphs – Physical length Lw is the actual length of a body, defined for a particular shape – Structural length L is the volumetric length of structure, where the individual is assumed to consist of structure, reserve and the reproduction buffer. δM = L/ Lw • Volume, wet/dry weight have contributions from structure, reserve, reproduction buffer • Constant specific mass & volume of structure, reserve, reproduction buffer • Constant chemical composition of juvenile growing at constant food Data • Real-data Empirical observations of physiological process – zero-variate – uni-variate • Pseudo-data Prior knowledge of a selection of parameter values – zero-variate Zero-variate data Life history events: hatching, birth, metamorphosis, puberty, death Real data: age, length, dry-, wet-weight at life history events max rates: reproduction, respiration, feeding, growth Modified by food, temperature Pseudo-data Typical parameter values of the generalized animal Species specific parameters should not be included as pseudo-data (e.g., z, δM, EHb, EHp) Growth efficiency κG vary less than the specific cost for structure [EG], and should be preferred for pseudo-data [EG] = μV [MV] / κG with [MV] =dV /wV Typical values for the ash-free-dry-weight over wet-weight ratio. Scyphomedusa 0.04 Ctenophora 0.04 Ascidia 0.06 Ectoprocta 0.07 Priapulida 0.07 Cheatognata 0.07 Actinaria 0.08 Bivalvia 0.09 Echinodermata 0.09 Porifera 0.11 Sipuncula 0.11 Gastropoda 0.15 Polychaeta 0.16 Crustacea 0.17 Cephalopoda 0.21 Pisces 0.22 Turbellaria 0.25 Aves 0.28 Reptilia 0.30 Mammalia 0.30 Uni-variate data • length, weight, reproduction, respiration, feeding as functions of time, temperature, food • incubation time, juvenile period, life span as functions of time, temperature, food • weight as function of length • egg number as function of weight/length Completeness of Real-data 0 maximum length and body weight; weight as function of length 1 age, length and weight at birth and puberty for one food level; mean life span (due to ageing) 2 growth (curve) at one food level: length and weight as function of age at constant (or abundant) food level 3 reproduction and feeding as function of age, length and/or weight at one food level 4 growth (curve) at several (>1) food levels; age, length and weight at birth and puberty at several food levels 5 reproduction and feeding as function of age, length and/or weight at several (>1) food levels 6 respiration as function of length or weight and life span at several (>1) food levels 7 elemental composition at one food level, survival due to ageing as function of age 8 elemental composition at several (>1) food levels, including composition of food 9 elemental balances for C, H, O and N at several body sizes and several food levels 10 energy balance at several body sizes and several food levels (including heat) Each level includes all lower levels Abstract World Auxiliary Parameters Core Primary Parameters [pM] [EG] v f {pAm} ... δM dV yEV ... Mapping Functions estimation prediction Lm = {pAm}/[pM] [Em] = {pAm}/v = ref rB = 1/(3/ [pM]/[EG] + 3 * f * Lm/ v) Wm = Lm3dV(1+fyEV [Em]/[EG]) LWm = Lm/δM Lw (t)= Lwm - (Lwm - Lwb) exp(-rBt) Wm maximum dry mass (g) LWm maximum body length (cm) rb von Bertalanffy growth rate (1/day) ... Zero-variate Observations Real World [pM]ref vref [EG]ref ref Zero-variate Pseudo-data t (time, days) LW (body lenght,cm) t1 t2 t3 ... LW(t1) LW(t2) LW(t3) Uni-variate Observations Lika et al., 2011 J. Sea Research 22:270-277 The covariation method Estimates all parameters simultaneously using all data: single-step-procedure Independently normally distributed error with constant variation coefficient Estimation criteria • Weighted Least Square (WLS) • Maximum Likelihood (ML) WLS criterion Minimization of a weighted sum of squared deviations between observations yij and predictions fij w ij j i ( yij f ij ) 2 yij2 The weight coefficients : wij / yij2 account for differences in units of the various data The dimensionless weight factor wij account for the certainty of the individual data point ML criterion For independently normally distributed dependent variables, the ln-likelihood function is n 1 ( ,c ) ln( 2 ) n ln c ln f ( xi ; ) 2 2 2c i 2 ( y / f ( x ; ) 1 ) i i i The ML estimator for the squared variation coeff ˆc2 1 2 ( y / f ( x ; ) 1 ) i i n i The ML estimates ˆ minimize 1 ln ˆc ln f ( xi ; ) n i Numerical implementation Nelder-Mead method A simplex method for finding a local minimum of a function of several variables For 2 variables, a simplex is a triangle The function is evaluated at the vertices of the triangle. The worst vertex xh , where f is largest, is rejected and replaced with a new vertex xC obtained via a sequence of transformations (reflect, expand or contract) or shrink the triangle towards the best. Does not require any derivative info Reflection Contraction outside Shrinking Expansion Contraction inside Numerical implementation Nelder-Mead simplex method debtool/lib/regr/nmregr (WLS) debtool/lib/regr/nmvcregr (ML) Numerical implementation Newton-Raphson A method for finding successively the roots of an equation f(x)=0. The iteration scheme: xn1 f ( xn ) xn f ( xn ) debtool/lib/regr/nrregr (WLS) debtool/lib/regr/nrvcregr (ML) Source wikipedia Evaluation of the estimation • Effects of pseudo-data – Elasticity coefficients ˆ1 ˆ0 e ˆ0 θ a core parameter to be estimated ˆ0 estimate of θ given the pseudo data θ0 α percentage increase in pseudo-value ˆ1 estimate of θ given the pseudo data θ0(1+α) Evaluation of the estimation • Goodness of fit – Mean relative error for the real data estimation criterion MRE function WLS n 1 n i 1 exp i 1 obs i ML 2 debtool/lib/regr/mre FIT =10 (1-MRE) n 1 n i 1 obs i 1 exp i 2 debtool/lib/regr/mrevc Parameter identifiability κ data on growth and reproduction and size at birth and puberty are required simultaneously z, δM zero-variate data and growth data, while additional uni-variate data reduce the standard deviation of the estimate. feeding data reproduction at several food levels mean life span survival as a function of age κΧ, {Fm} kJ, EHp , κR ha sG Kooijman et al. 2008 Biol. Rev., 83:533-552. Lika et al., 2011 J. Sea Research 22:278-288 Properties of the covariation method estimation of parameter κ The effect of the pseudo-value κ is reduced only when there is information for both growth and reproduction estimation of parameter the effect of the pseudo-value is reduced only when information on age at birth and puberty is given estimation of parameter [pM] the effects of the pseudo-value [pM] are reduced as information on real data increases the least effect is obtained when information on respiration is included the estimation of [EG] the effects of the pseudo-data κG are reduced as information on real data increases estimation of the parameter kJ the pseudo-value for kJ does not play significant role The covariation method for parameter estimation • Estimation of all parameters of the standard DEB model simultaneously • Real-data and pseudo-data, exploiting the rules for the covariation of parameter values among species implied by the standard DEB model • The least required information is the maximum size, but the pseudo-data fully control the result • Increasing the number of type of data decreases the role of pseudo data Add_my_pet collection 2011 : ~ 60 species 2013 : 240 species 1 0.8 0.6 10 0.4 8 FIT mark 0.2 6 0 0 1 2 3 4 COMPLETE mark 5 6 1 4 0.8 2 0.6 0 1 2 3 COMPLETE mark 4 5 0.4 0.2 0 0 2 4 6 FIT mark 8 10 Max specific assimilation rate -2 log {pAm}, J cm d 10 -1 4 3 2 1 0 -2 -1 0 log L, cm 10 1 2 Before acceleration After acceleration Kooijman, 2013 Oikos 122:348-357 Maturity levels 10 8 5 log Ej 10 0 10 log Eb H H 5 2 -1 -5 -4 0 , cm L log 10 -7 2 1 -1 -2 10 8 6 H -1 log Ep -2 10 -10 4 2 0 -2 -4 -6 -2 -1 0 log L , cm 10 1 2 0 , cm L log 10 1 2 Energy conductance 2 10 log v, cm d-1 1 0 -1 -2 -3 -4 Before acceleration After acceleration -2 -1 0 log L, cm 10 1 2 Thank you for your attention