Injustice - Developers Among Us (SciFiDevCon 2024)
Jobim2011 o gimenez
1. Modèles à structure cachée pour la
dynamique des populations
Olivier Gimenez
Centre d’Ecologie Fonctionnelle et Evolutive - Montpellier
2. Population dynamics in the wild
• Conservation: what are the reasons of a population decline
and how to stop it?
• Harvesting: how many individuals can be harvested in a
sustainable way?
• Pest control: what is the most efficient (cheapest) way to
get rid of an alien species?
• Evolutionary ecology: to understand the evolution of life
histories.
3. Population dynamics in the wild
Investigating process in natural populations
Long-term individual monitoring datasets
Methodological issues when moving from lab
to natural conditions
Issue 1: detectability < 1
Issue 2: individual heterogeneity (IH)
4. Population dynamics in the wild
Investigating process in the wild
Long-term monitoring
Methodological issues when moving from lab
to natural conditions
Issue 1: detectability < 1
Issue 2: individual heterogeneity
5. Issue of detectability < 1
How to reliably estimate demographic
parameters in the wild?
Individuals may be seen or not
If they’re not... Are they breeding? Are they
on the study site? Are they dead?
Individually mark and monitor individuals:
capture-recapture (CR) data
7. Incomplete registration
Laplace, P.S., 1786. Sur les naissances les mariages et les morts. Histoire de
l’Académie Royale des Sciences. Année 1783, p. 693
"POPULATION dans le Royaume, compris l’île de Corse,
suivant l'ordre des généralités, pendant l'année 1783."
8. Why bother with p < 1?
Capture-
recapture
approach
Naïve
approach
with p = 1
9. Why bother with p < 1?
1.0
0.8
Survival Capture-
0.6
recapture
approach
0.4
Naïve
0.2
2 4 6 8 10 12 14 approach
Age with p = 1
Bias in survival and rate of senescence
(Gimenez et al. 2008 Am. Nat.)
10. Why bother with p < 1?
1.0
0.8
0.6
Capture-
recapture
Survival
0.4
approach
0.2
Naïve
0.0
-4 -2 0 2 4
approach
with p = 1
Body mass
Bias in shape of selection
(Gimenez et al. 2008 Am. Nat.)
11. Investigating evolution in the wild
Investigating evolution in the wild (Grant,
Reznick, ...)
Long-term monitoring
Methodological issues when moving from lab
to natural conditions
Issue 1: detectability < 1
Issue 2: individual heterogeneity (IH)
12. Issue of individual heterogeneity
Simple CR models assume homogeneity
From a statistical point of view, IH can cause
bias in parameter estimates – see later on
13. Issue of individual heterogeneity
Standard CR models assume homogeneity
From a statistical point of view, IH can cause
bias in parameter estimates
From a biological point of view, IH is of
interest – individual quality
14. What is individual quality?
Quality varies among individuals within a
population
High quality individuals have greater fitness
than low quality ones
⇒ Among-individual heterogeneity that
is positively correlated to fitness
Wilson & Nussey 2009 TREE
15. What is individual quality?
Quality varies among individuals within a population
High quality individuals have greater fitness than low
quality ones
⇒ Among-individual heterogeneity that is
positively correlated to fitness (Wilson & Nussey 2009)
Why is it so important?
Natural selection can occur if individuals
vary in phenotype and fitness
A response to selection depends on this
variation having a genetic basis
IH may lead to flawed inference
16. Accounting for individual heterogeneity
CR models do not cope that well with quality
Accounting for individual heterogeneity
If you’re a biologist, rely on empirical
measures (mass, gender, age, experience, etc.)
How to incorporate this information?
If you’re a statistician, intrinsic property of
individuals
How to filter out the signal from noisy observations?
17. Capture-recapture models
Introduction: CR data
How to account for individual variation
Case study 1: estimating abundance
Case study 2: detecting trade-offs
Can quality have a genetic basis or is it a
consequence of environmental effects?
Case study 3: quantifying heritability
Perspectives
18. Capture-recapture models
Introduction: CR data
How to account for individual variation
Case study 1: estimating abundance
Case study 2: detecting trade-offs
Can quality have a genetic basis or is it a
consequence of environmental effects?
Case study 3: quantifying heritability
Perspectives
19. Common marking methods
• Ear tags for mammals / leg bands for birds.
• Passive integrated transponder (PIT) tags.
21. Marking by noninvasive genetic sampling
• Individuals are uniquely identified using
microsatellite profiling on hair, dung, … samples
bear (hair) bat (droppings)
wolf (dung)
orang-utan (hair) elephant (dung)
27. Modelling CR data
An encounter history: hi = (1 0 1)
φ φ
1 0 1 Pr (hi ) = φ (1 − p ) φ p
1− p p
Survival probability φ
Detection probability p
28. Modelling CR data
A probabilistic framework
Pr (hi ) = φ (1 − p ) φ p
Central role of likelihood (frequentist / bayesian)
L = ∏ Pr (hi )
i
How to account for IH in
φi
29. Capture-recapture models
Introduction: CR data
How to account for individual variation
Case study 1: estimating abundance
Case study 2: detecting trade-offs
Can quality have a genetic basis or is it a
consequence of environmental effects?
Case study 3: quantifying heritability
Perspectives
30. Cas 1: estimation de l’abondance
L’exemple de la recolonisation du loup (Canis lupus)
dans les Alpes
Sarah Cubaynes & Lucile Marescot
- Cubaynes et al. (2010). Importance of accounting for detection heterogeneity when
estimating abundance: the case of French wolves. Conservation Biology. 24:621-626.
- Marescot et al. (2011). Capture-recapture population growth rate as a robust tool against
detection heterogeneity for population management. Ecological Applications.
31. Cas 1: estimation de l’abondance
L’exemple de la recolonisation du loup (Canis lupus)
dans les Alpes
32. Cas 1: estimation de l’abondance
Echantillonnage (ONCFS) Séquençage ADN (LECA)
Des données génétiques de capture-recapture
33. Cas 1: estimation de l’abondance
Echantillonnage (ONCFS) Séquençage ADN (LECA)
Des données génétiques de capture-recapture
Les individus dominants sont plus facilement « capturés » (marquage
du territoire)
La population est un mélange de 2 classes d’individus :
« facilement capturables » et « difficilement capturables »
34. L’information à laquelle on aimerait accéder
Les états :
• Vivant et facilement capturable (L)
• Vivant et difficilement capturable (H)
• Mort (†)
35. Les informations dont on dispose…
Les états :
• Vivant et facilement capturable (L)
• Vivant et difficilement capturable (H)
• Mort (†)
Les données : Présence (1) / Absence (0)
36. Modèle de Markov caché (Pradel 2005 Bcs)
États initiaux : L H †
Π = (1 − π π 0)
Transition entre états (survie) :
L H †
L φ 0 1−φ
Φ=
H 0 φ 1−φ
† 0 0 1
37. L’hétérogénéité de capture
Lien entre observations et états :
Pas
détecté Détecté
L 1 − pL pL
B = H 1 − pH pH
† 1 0
pL : probabilité de capture des individus facilement capturables
pH : probabilité de capture des individus difficilement capturables
38. Probability of a capture history
Under homogeneity, the capture
history ‘101’ has probability
Pr(101 = φ ⋅ (1− p) ⋅φ ⋅ p
)
φ is survival
p is detection for all individuals
39. Probability of a capture history
Under heterogeneity:
( ) ( )
Pr(101 = π ⋅φ ⋅ 1− pL ⋅φ ⋅ pL + (1− π ) ⋅φ ⋅ 1− pH ⋅φ ⋅ pH
)
π is the probability that the individual
belongs to state L
pL is the detection for lowly detectable
individuals
pH is the detection for highly detectable
individuals
40. In brief: 2-step analysis for estimating abundance
Model selection procedure (AIC)
1) Patterns of temporal variation in survival and detection
Constant, seasonal or annual?
2) Patterns of individual variation in detection
Homogeneity vs Heterogeneity
Plug in parameters of
the model best
supported by the data
Abundance estimation with heterogeneity
ˆ ≈ m + π × u + (1 − π )× u
N
ˆ ˆ
pH
ˆ pL
ˆ pH
ˆ
43. Homogeneity vs. Heterogeneity
detection probability
1.0
1.0
Homogeneity Heterogeneity
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.0
0.0
Summer Autumn Winter Spring Summer Autumn Winter Spring
44. Homogeneity vs. Heterogeneity
detection probability
1.0
1.0
• Detectability strongly Heterogeneity
differs between classes
0.8
0.8
H individual = dominant
L individual = young +
0.6
0.6
subordinate
0.4
0.4
0.2
0.2
0.0
0.0
Summer Autumn Winter Spring Summer Autumn Winter Spring
45. Seasonal variation in abundance
N
111
25
64
10 29
0
Overall growth of the population
Marked seasonal variations
46. Ignoring detection heterogeneity leads
to strong bias in abundance estimation
Heterog.
Homog.
N
64 [29 ; 111]
33 [17 ; 54]
Underestimation by 50% of abundance
47. Ignoring detection heterogeneity leads
to bias in survival estimation
0.83 (0.06)
0.9
0.68 (0.08)
φ
0.8
survie annuelle
Homogeneity vs.
heterogeneity in
0.7
detection
0.6
CJS H
underestimation by12% for survival
48. Capture-recapture models
Introduction: CR data
How to account for individual variation
Case study 1: estimating abundance
Case study 2: detecting trade-offs
Can quality have a genetic basis or is it a
consequence of environmental effects?
Case study 3: quantifying heritability
Perspectives
49. M. Buoro (thèse, co-dir. E. Prévost1)
1 UMR INRA/UPPA Ecobiop, Saint Pée s/ Nivelle, France
Photo: Paul Nicklen (National Geographic)
50. M. Buoro (thèse, co-dir. E. Prévost1)
1 UMR INRA/UPPA Ecobiop, Saint Pée s/ Nivelle, France
Photo: Paul Nicklen (National Geographic)
51. Assessing life-history tradeoffs in the wild
• Natural selection favours individuals that
maximize their fitness
• Limited resource acquisition: strategy of
resource allocation
• Trade-off between traits related to
fitness
• Issue of detectability, again
52. State-space modelling of CR data
(Gimenez et al. 2007)
Dynamic process model
Hidden states
Xi,t-1
Xi,t
54. State-space modelling of CR data
Dynamic process model
Hidden states
Xi,t-1
φi,t
survival
Xi,t
State equation
55. State-space modelling of CR data
Dynamic process model Observation
Hidden states Observations
Xi,t-1 Yi,t-1
φi,t
survival
Xi,t Yi,t
56. State-space modelling of CR data
Dynamic process model Observation
Hidden states
Xi,t-1 Yi,t-1
detection
Xi,t Yi,t Pt
Observation equation
57. State-space modelling of CR data
Dynamic process model Observation
Hidden states Observations
Xi,t-1 Yi,t-1
φi,t detection
survival
Xi,t Yi,t Pt
State equation Observation equation
58. Atlantic salmon life cycle
Reproduction Development of
juveniles
Freshwater
Migration to stream Migration to sea
Growth at sea
Sea
60. Juveniles Autumn
1st year of life
Migrants Spring
Freshwater
Sea
61. Juveniles Autumn
1st year of life
Migrants Residents Spring
Freshwater
Sea
62. Juveniles Autumn
1st year of life
Migrants Residents Spring
Freshwater
Sexual maturation Autumn
(males)
2nd year of life
Migrants Spring
Sea
63. Juveniles Autumn
1st year of life
Migrants Residents Spring
Freshwater
Sexual maturation Autumn
(males)
2nd year of life
Migrants Spring
Adults
Sea
64. Juveniles Autumn
Winter survival 1st year
Résidents + 1
Migrants
an
Spring
Freshwater
Is there a tradeoff between
migration and winter survival?
65. State-space model Juveniles
Migrants Residents
Dynamic process model Observation
Juveniles
marked in
autumn
Migrants
recapture in
spring
66. State-space model Juveniles
Migrants Residents
Dynamic process model Observation
Juveniles
Migration Migration
probability choice
marked in
autumn
Migrants
recapture in
spring
67. State-space model Juveniles
Migrants Residents
Probabilistic reaction norm
Dynamic process model Observation
Juveniles
Migration Migration
Size
probability choice
marked in
autumn
Migrants
recapture in
spring
68. State-space model Juveniles
Migrants Residents
Dynamic process model Observation
Juveniles
Migration Migration
Size
probability choice
marked in
autumn
Migrants
Survival Migrants
probability Survivor
recapture in
spring
69. State-space model Juveniles
Migrants Residents
Observation
Selective mortality
Juveniles
Migration Migration
Size
probability choice
marked in
autumn
Migrants
Random Survival Migrants
probability Survivor
recapture in
effect
spring
70. State-space model Juveniles
Migrants Residents
Dynamic process model Observation
Juveniles
Migration Migration
Size
probability choice
marked in
autumn
Detection
probability
Migrants
Random Survival Migrants
probability Survivor
recapture in
effect
spring
71. State-space model Juveniles
Migrants Residents
Dynamic process model Observation
Juveniles
Migration Migration
Size
probability choice
marked in
autumn
Detection
probability
Migrants
Random Survival Migrants
probability Survivor
recapture in
effect
spring
72. α1 T0
β1
κ
α2 T0.1 S0.1
β2
Φ1 Lf
Φ2 T1 S1
T1.1 pL1
Ψmâle
S1.1 S1.2
T1.2.1 T1.2.2
S1.1.1 S1.1.2 S1.2.1 S1.2.2 IV1
T2.2 Pr.det T1.2.2det
pP1
A1
T1.2capt T2.2capt T1.3.1 T1.3.2a T1.3.2b T1.3.3
pA1
pC1
S3 S2.1 S2.2 S2.3
δ2
Φ3 Φ3
δ1
Processus d’observation
Processus d’observation des
des smolts 2+ (marqués
pL2 smolts 2+ (marqués au stade
au stade tacon 1+) au
tacon 0+) au printemps 2007
printemps 2007 i in 1:1851
j in 1:286
IV2 pP2
73. Results (1)
Probabilistic reaction norm
1.0
Migration Probability
0.2 0.4 0.6 0.8
Size-dependent
probabilistic reaction
norm for age at
0.0
50 55 60 65 70 75 80 85 90 95 100 105 110 115 120 125 130 migration
Size (mm)
Buoro, M., Prévost, E. and O. Gimenez (2010). Investigating evolutionary
trade-offs in wild populations of Atlantic salmon (Salmo salar): incorporating
detection probabilities and individual heterogeneity. Evolution. 64: 2629-2642
74. Results (1)
Probabilistic reaction norm
1.0
Migration Probability
0.2 0.4 0.6 0.8
Size-dependent
probabilistic reaction
norm for age at
0.0
50 55 60 65 70 75 80 85 90 95 100 105 110 115 120 125 130 migration
Size (mm)
Juveniles longer than 100 mm in autumn has a probability to migrate close to 1.
75. Results (1)
Probabilistic reaction norm
1.0
Migration Probability
0.2 0.4 0.6 0.8
Size-dependent
probabilistic reaction
norm for age at
0.0
50 55 60 65 70 75 80 85 90 95 100 105 110 115 120 125 130 migration
Size (mm)
Juveniles longer than 100 mm in autumn has a probability to migrate close to 1.
A juvenile of 90 mm has 50% of chance of migrating to the sea at 1year of age.
76. Results (1)
Probabilistic reaction norm
1.0
Migration Probability
0.2 0.4 0.6 0.8
Size-dependent
probabilistic reaction
norm for age at
0.0
50 55 60 65 70 75 80 85 90 95 100 105 110 115 120 125 130 migration
Size (mm)
Juveniles longer than 100 mm in autumn has a probability to migrate close to 1.
A juvenile of 90 mm has 50% of chance of migrating to the sea at 1year of age.
Juveniles shorter than 60 mm in autumn has a probability to migrate almost null.
77. Results (2)
Probabilistic reaction norm
1.0
Migration Probability
0.2 0.4 0.6 0.8
Size-dependent
probabilistic reaction
norm for age at
0.0
50 55 60 65 70 75 80 85 90 95 100 105 110 115 120 125 130 migration
Size (mm)
Selective mortality
1.0
Migrants
0.8
Survival cost in
Survival Probability
deciding to stay an
0.6
Residents extra year in
0.4
freshwater
0.2
0.0
78. Capture-recapture models
Introduction: CR data
How to account for individual variation
Case study 1: estimating abundance
Case study 2: detecting trade-offs
Can quality have a genetic basis or is it a
consequence of environmental effects?
Case study 3: quantifying heritability
Perspectives
79. Heritability in the wild
Animal models: mixed models incorporating
genetic, environmental and other factors.
Capture-recapture models: assess demographic
parameters with p < 1 and individual variability.
The idea of combining Animal and Capture-
recapture models is in the air (O’Hara et al. 2008;
Cam 2009).
81. Heritability in the wild
Animal models: mixed models incorporating
genetic, environmental and other factors.
Capture-recapture models: assess demographic
parameters with p < 1 and individual variability.
The idea of combining Animal and Capture-
recapture models is in the air (O’Hara et al. 2008;
Cam 2009).
For (demographic) parameters strongly related to
fitness, we expect low heritability. But, predictions
not so clear in wild populations
82. Heritability in the wild
Animal models: mixed models incorporating
genetic, environmental and other factors.
Capture-recapture models: assess demographic
parameters with p < 1 and individual variability.
The idea of combining Animal and Capture-
recapture models is in the air (O’Hara et al. 2008;
Cam 2009).
For (demographic) parameters strongly related to
fitness, we expect low heritability. But, predictions
not so clear in wild populations
M2R BioStat J. Papaïx & S. Cubaynes’s PhD (co-dir.
C. Lavergne)
83. State-space modelling of CR data
Dynamic process model Observation
Hidden states Observations
Xi,t-1 Yi,t-1
φi,t detection
survival
Xi,t Yi,t Pt
State equation Observation equation
84. Introducing the threshold model
• Survival is related to a continuous underlying latent
variable li,t which, given Xi,t = 1, satisfies
1 if li,t > κ
X i,t +1 =
0 if li,t ≤ κ
• where κ is a threshold value, and li,t is usually
referred to as the liability
86. Introducing the threshold model
• Survival is related to a continuous underlying latent
variable li,t which, given Xi,t = 1, satisfies
1 if li,t > κ
X i,t +1 =
0 if li,t ≤ κ
• where κ is a threshold value, and li,t is usually
referred to as the liability with
(
li,t ~ N µi ,t , σ ε2 )
• For identifiability reasons, κ = 0 and σε = 1 without
loss of generality
87. Plug in the animal model
It can be shown that φi,t = F µi ,t ( ) where F is
the c.d.f. of a N(0,1)
Then, we write:
F (φi,t ) = probit (φi,t ) = µi ,t = η + bt + ei + ai
−1
88. Plug in the animal model
It can be shown that φi,t = F µi ,t ( ) where F is
the c.d.f. of a N(0,1)
Then, we write:
F (φi,t ) = probit (φi,t ) = µi ,t = η + bt + ei + ai
−1
mean survival
89. Plug in the animal model
It can be shown that φi,t = F µi ,t ( ) where F is
the c.d.f. of a N(0,1)
Then, we write:
F (φi,t ) = probit (φi,t ) = µi ,t = η + bt + ei + ai
−1
mean survival
yearly effect
(
bt ~ N 0, σ t
2
)
90. Plug in the animal model
It can be shown that φi,t = F µi ,t ( ) where F is
the c.d.f. of a N(0,1)
Then, we write:
F (φi,t ) = probit (φi,t ) = µi ,t = η + bt + ei + ai
−1
mean survival
non-genetic effect
yearly effect
(
ei ~ N 0, σ e2 )
(
bt ~ N 0, σ t
2
)
91. Plug in the animal model
It can be shown that φi,t = F µi ,t ( ) where F is
the c.d.f. of a N(0,1)
Then, we write:
F (φi,t ) = probit (φi,t ) = µi ,t = η + bt + ei + ai
−1
mean survival
non-genetic effect
yearly effect (
ei ~ N 0, σ e2 )
(
bt ~ N 0, σ t2 ) additive genetic effect
(a1,K, aN ) ~ MN (0,σ 2
a )
A
92. Calculating heritability
- Capture-recapture animal models (CRAMs)
- Decomposing components of variance in survival
- Heritability = contribution of genetic variance to the total
σ a2
h2 = 2 2 2
σ t + σ e + σ a +1
93. Case study on blue tits in Corsica
Mark-recapture data Social pedigree
• Blue tits – Study site in
Corsica. 654 individuals,
• 1979 – 2007 ⇒ 29 years of 218 fathers (sires),
monitoring)! 215 mothers (dams),
12 generations.
94. Résultats
Julien Papaïx
Papaïx, J., S. Cubaynes, M. Buoro, A. Charmantier, P. Perret and O. Gimenez
(2010). Combining capture–recapture data and pedigree information to assess
heritability of demographic parameters in the wild. Journal of Evolutionary
Biology. 23: 2176-2184
98. Capture-recapture models
Introduction: CR data
How to account for individual variation
Case study 1: estimating abundance
Case study 2: detecting trade-offs
Can quality have a genetic basis or is it a
consequence of environmental effects?
Case study 3: quantifying heritability
Perspectives
99. Conclusions
CR methodology is catching up with ‘p=1’ world
(medicine)
IH needs to be accounted for, otherwise
Bias in abundance estimation (PhD S. Cubaynes)
Obscur life-history tradeoffs (PhD M. Buoro)
Recent statistical methods: hidden-Markov model
and state-space models - cope with IH when p < 1
If possible, biological view – measure quality
100. Perspectives - Methods
Continue efforts in developing methods to properly
account for individual heterogeneity
Estimation des états cachés – Viterbi ou autre
Semi-chaîne de Markov pour modéliser la durée de
séjour dans un état (R. Choquet, collab. Y. Guédon)
Fit and compare models (PhD S. Cubaynes)
Is heritability important in blue tits (model selection)?
Shall we go for discrete or continuous heterogeneity?
Speed up estimation (algorithms; random effects)?
Software implementation
101. Implementation issues: software
Program E-SURGE
Hidden Markov models, individual
covariates, mixtures and individual
random effects
102. Implementation issues: software
Program E-SURGE
Hidden Markov models, individual
covariates, mixtures and individual
random effects
R. Choquet & E. Nogué
103. Perspectives - Biology
Consider other demographic parameters (dispersal
and breeding probabilities e.g.);
→ A. Charmantier, B. Doligez, E. Cam, B. Sheldon
Fixed vs. dynamic individual heterogeneity:
→ E. Cam and S. Tuljapurkar
From individuals to species
→ E. Papadatou’s post-doc & S. Cubaynes’s PhD
→ Museum for community ecology aspects
Integrating evolutionary and demography views:
→ S. Servanty’s post-doc and M. Gamelon’s PhD
105. Estimating abundance in open populations
Standard capture-recapture models provide
estimates of survival and detection probabilities
An estimate of abundance N is obtained as:
ˆ=n
N
ˆ
p
106. Estimating abundance in open population
Standard capture-recapture models provide
estimates of survival and detection probabilities
An estimate of abundance N is obtained as:
Number of
individuals
ˆ=n
N
detected
ˆ
p
107. Estimating abundance in open population
Standard capture-recapture models provide
estimates of survival and detection probabilities
An estimate of abundance N is obtained as:
ˆ=n
N
ˆ
p
Estimated
detection
probability
108. What if heterogeneity in detection?
The number of counted individuals can be split
into two quantities
Newly marked (u) and previously marked (m)
n=u+m
109. What if heterogeneity in detection?
The number of counted individuals can be split
into two quantities
Newly marked (u) and previously marked (m)
n=u+m
Number of previously marked individuals
with probability pH
110. What if heterogeneity in detection?
The number of counted individuals can be split
into two quantities
Newly marked (u) and previously marked (m)
n=u+m
Number of unmarked individuals, made of:
• π×u individuals with low capturability pL and
• (1-π)×u individuals with high capturability pH
111. Abundance with detection heterogeneity
An estimate of abundance N accounting for
heterogeneity is obtained as:
ˆ ≈ m + π × u + (1 − π ) × u
N
ˆ ˆ
ˆH
p pˆL pˆH