Introduction To Hierarchical Bayesian Modeling

For Ecological Data E Parent download

https://ebookbell.com/product/introduction-to-hierarchical-
bayesian-modeling-for-ecological-data-e-parent-4347918

Explore and download more ebooks at ebookbell.com

Here are some recommended products that we believe you will be
interested in. You can click the link to download.

Introduction To Modern Analysis 2nd Edition 2nd Kantorovitz

https://ebookbell.com/product/introduction-to-modern-analysis-2nd-
edition-2nd-kantorovitz-44870612

Introduction To The Speechmaking Process 15th Edition Diana K Leonard

Raymond S Ross

https://ebookbell.com/product/introduction-to-the-speechmaking-
process-15th-edition-diana-k-leonard-raymond-s-ross-44874488

Introduction To Construction Management 2nd Edition 2nd Fred Sherratt

https://ebookbell.com/product/introduction-to-construction-
management-2nd-edition-2nd-fred-sherratt-44899008

Introduction To Analysis With Complex Numbers Irena Swanson

https://ebookbell.com/product/introduction-to-analysis-with-complex-
numbers-irena-swanson-44912170
Introduction To Quantitative Methods In Business With Applications
Using Microsoft Office Excel 1st Edition Bharat Kolluri

https://ebookbell.com/product/introduction-to-quantitative-methods-in-
business-with-applications-using-microsoft-office-excel-1st-edition-
bharat-kolluri-44915766

Introduction To Biostatistical Applications In Health Research With

Microsoft Office Excel And R 2nd Edition Robert P Hirsch

https://ebookbell.com/product/introduction-to-biostatistical-
applications-in-health-research-with-microsoft-office-excel-and-r-2nd-
edition-robert-p-hirsch-44915830

Introduction To Strategies For Organic Synthesis 2nd Edition Laurie S

Starkey

https://ebookbell.com/product/introduction-to-strategies-for-organic-
synthesis-2nd-edition-laurie-s-starkey-44915846

Introduction To Hydrogen Technology 2nd Edition K S V Santhanam

https://ebookbell.com/product/introduction-to-hydrogen-technology-2nd-
edition-k-s-v-santhanam-44916142

Introduction To Behavioral Research Methods 7th Edition 7th Mark R

Leary

https://ebookbell.com/product/introduction-to-behavioral-research-
methods-7th-edition-7th-mark-r-leary-44975468
 Statistics

Chapman & Hall/CRC Chapman & Hall/CRC

Bayesian Modeling for Ecological Data

Applied Environmental Statistics Applied Environmental Statistics

Introduction to

Introduction to Hierarchical
Making statistical modeling and inference more accessible to
ecologists and related scientists, Introduction to Hierarchical
Bayesian Modeling for Ecological Data gives readers a flexible and
effective framework to learn about complex ecological processes
Hierarchical
from various sources of data. It also helps readers get started on
building their own statistical models. Bayesian Modeling
for Ecological Data
The text begins with simple models that progressively become
more complex and realistic through explanatory covariates and
intermediate hidden states variables. When fitting the models to
data, the authors gradually present the concepts and techniques
of the Bayesian paradigm from a practical point of view using real
case studies. They emphasize how hierarchical Bayesian modeling
supports multidimensional models involving complex interactions
between parameters and latent variables. Data sets, exercises, and
R and WinBUGS codes are available on the authors’ website.
This book shows how Bayesian statistical modeling provides an
intuitive way to organize data, test ideas, investigate competing
hypotheses, and assess degrees of confidence of predictions. It
also illustrates how conditional reasoning can dismantle a complex
reality into more understandable pieces. As conditional reasoning
is intimately linked with Bayesian thinking, considering hierarchical

Parent • Rivot
models within the Bayesian setting offers a unified and coherent
framework for modeling, estimation, and prediction.

Éric Parent
C9195
Étienne Rivot

C9195_Cover.indd 1 7/11/12 2:42 PM

 IntroductIon to
HIerarcHIcal
BayesIan ModelIng
for ecologIcal data

C9195_FM.indd 1 7/19/12 3:41 AM

 CHAPMAN & HALL/CRC
APPLIED ENVIRONMENTAL STATISTICS
University of North Carolina
TATISTICS
Series Editor
Richard Smith
University of North Carolina
U.S.A.

Published Titles

Michael E. Ginevan and Douglas E. Splitstone, Statistical Tools for

Environmental Quality
Timothy G. Gregoire and Harry T. Valentine, Sampling Strategies for
Natural Resources and the Environment
Daniel Mandallaz, Sampling Techniques for Forest Inventory
Bryan F. J. Manly, Statistics for Environmental Science and
Management, Second Edition
Steven P. Millard and Nagaraj K. Neerchal, Environmental Statistics
with S Plus
Song S. Qian, Environmental and Ecological Statistics with R
Wayne L. Myers and Ganapati P. Patil, Statistical Geoinformatics for
Human Environment Interface
Éric Parent and Étienne Rivot, Introduction to Hierarchical Bayesian
Modeling for Ecological Data
 Chapman & Hall/CRC
Applied Environmental Statistics

introduction to
HierarcHical
Bayesian Modeling
for ecological data

Éric Parent
Étienne rivot
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2013 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Version Date: 20120531

International Standard Book Number-13: 978-1-58488-920-5 (eBook - PDF)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts
have been made to publish reliable data and information, but the author and publisher cannot assume
responsibility for the validity of all materials or the consequences of their use. The authors and publishers
have attempted to trace the copyright holders of all material reproduced in this publication and apologize to
copyright holders if permission to publish in this form has not been obtained. If any copyright material has
not been acknowledged please write and let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmit-
ted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented,
including photocopying, microfilming, and recording, or in any information storage or retrieval system,
without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.
com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood
Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and
registration for a variety of users. For organizations that have been granted a photocopy license by the CCC,
a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used
only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com
Contents

List of Figures xiii

List of Tables xvii

Foreword xix

I Basic blocks of Bayesian modeling 1

1 Bayesian hierarchical models in statistical ecology 3
1.1 Challenges for statistical ecology . . . . . . . . . . . . 4
1.1.1 The three steps of quantitative modeling . . . 4
1.1.2 Complexity, uncertainty and observations . . . 7
1.1.2.1 Increasing model realism . . . . . . 7
1.1.2.2 Different sources of uncertainty . . . 8
1.1.3 Ecological models and statistics . . . . . . . . 9
1.1.4 Motivating examples . . . . . . . . . . . . . . . 10
1.2 Conditional reasoning, graphs and hierarchical models 17
1.2.1 Think conditionally . . . . . . . . . . . . . . . 17
1.2.2 Graphical models . . . . . . . . . . . . . . . . 18
1.2.3 Conditional reasoning and complexity . . . . . 20
1.2.4 Observation processes . . . . . . . . . . . . . . 23
1.2.5 Hierarchical modeling . . . . . . . . . . . . . . 24
1.3 Bayesian inferences on hierarchical models . . . . . . . 27
1.3.1 The intuition: What if a node is observed? . . 27
1.3.2 Getting the Bayesian posterior distribution . . 28
1.3.3 Capturing the posterior distribution in practice 33
1.4 What can be found in this book? . . . . . . . . . . . . 34

2 The Beta-Binomial model 43

2.1 From a scientific question to a Bayesian analysis . . . 44
2.2 What is modeling? . . . . . . . . . . . . . . . . . . . . 46
2.2.1 Model making . . . . . . . . . . . . . . . . . . 46
2.2.2 Discussing hypotheses . . . . . . . . . . . . . . 47
2.2.3 Likelihood . . . . . . . . . . . . . . . . . . . . 47

v
vi

2.3 Think conditionally and make a graphical representation 49

2.4 Inference is the reverse way of thinking . . . . . . . . 50
2.5 Expertise matters . . . . . . . . . . . . . . . . . . . . . 52
2.6 Encoding prior knowledge . . . . . . . . . . . . . . . . 53
2.7 The conjugate Beta pdf . . . . . . . . . . . . . . . . . 54
2.8 Bayesian inference as statistical learning . . . . . . . . 56
2.9 Bayesian inference as a statistical tool for prediction . 57
2.10 Asymptotic behavior of the Beta-Binomial model . . . 59
2.11 The Beta-Binomial model with WinBUGS . . . . . . . 60
2.12 Further references . . . . . . . . . . . . . . . . . . . . 64

3 The basic Normal model 65

3.1 Salmon farm’s pollutants and juvenile growth . . . . . 65
3.2 A Normal model for the fish length . . . . . . . . . . . 66
3.3 Normal-Gamma as conjugate models to encode expertise 69
3.3.1 Encoding prior knowledge on µ . . . . . . . . . 69
3.3.2 Encoding prior knowledge on σ . . . . . . . . . 70
3.3.3 Bayesian updating of the mean . . . . . . . . . 71
3.3.4 Joint prior for (µu , µf , µd , σ −2 ) . . . . . . . . . 74
3.4 Inference by recourse to conjugate property . . . . . . 75
3.4.1 Bayesian updating in a closed-form . . . . . . 75
3.4.2 Does the fish farm perturb salmons’ growth? . 76
3.5 Bibliographical notes . . . . . . . . . . . . . . . . . . . 79
3.6 Further material . . . . . . . . . . . . . . . . . . . . . 80

4 Working with more than one Beta-Binomial element 83

4.1 Capture-mark-recapture analysis . . . . . . . . . . . . 84
4.1.1 Motivating example . . . . . . . . . . . . . . . 84
4.1.2 Sampling distributions and likelihood . . . . . 87
4.1.3 Prior distributions . . . . . . . . . . . . . . . . 89
4.1.4 Getting the posterior distribution . . . . . . . 90
4.2 Successive removal analysis . . . . . . . . . . . . . . . 91
4.2.1 Motivating example . . . . . . . . . . . . . . . 91
4.2.2 Sampling distributions . . . . . . . . . . . . . 92
4.2.3 Prior distributions . . . . . . . . . . . . . . . . 94
4.2.4 Full joint posterior . . . . . . . . . . . . . . . . 96
4.3 Testing a new tag for tuna . . . . . . . . . . . . . . . . 96
4.3.1 Motivating example . . . . . . . . . . . . . . . 96
4.3.2 From the question to a Bayesian problem . . . 97
4.3.2.1 Setting a prior pdf for π . . . . . . . 97
4.3.2.2 Posterior for π . . . . . . . . . . . . 98
4.3.3 Hypotheses testing via Bayes Factors . . . . . 100
4.3.3.1 Bayes Factors . . . . . . . . . . . . . 101
 vii

4.3.3.2 Back to the tuna experiment . . . . 102

4.4 Further references . . . . . . . . . . . . . . . . . . . . 103

5 Combining various sources of information 105

5.1 Motivating example . . . . . . . . . . . . . . . . . . . 106
5.1.1 The dangerous run of spawners . . . . . . . . . 107
5.1.2 Observed variables . . . . . . . . . . . . . . . . 108
5.2 Stochastic model for salmon behavior . . . . . . . . . 109
5.2.1 Bernoulli trials and population features . . . . 109
5.2.2 Parameters are assumed stationary . . . . . . 110
5.2.3 Prior expertise is available for the parameters 110
5.2.4 Prior elicitation . . . . . . . . . . . . . . . . . 111
5.2.5 Introducing latent variables . . . . . . . . . . . 113
5.2.6 Stochastic model as a directed acyclic graph . 114
5.2.7 The interannual model . . . . . . . . . . . . . 115
5.3 Inference with WinBUGS . . . . . . . . . . . . . . . . 116
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.4.1 Estimation for the year 1995 . . . . . . . . . . 118
5.4.2 Inference when taking into account 5 years of
data . . . . . . . . . . . . . . . . . . . . . . . . 120
5.5 Discussion and conclusions . . . . . . . . . . . . . . . 121

6 The Normal linear model 125

6.1 The decrease of Thiof abundance in Senegal . . . . . . 125
6.2 Linear model theory . . . . . . . . . . . . . . . . . . . 128
6.2.1 Basics and notations . . . . . . . . . . . . . . . 128
6.2.2 Bayesian updating of the Normal linear model 129
6.2.3 Defining appropriate Zellner’s prior for θ . . . 130
6.2.3.1 Contrasts . . . . . . . . . . . . . . . 130
6.2.3.2 Zellner’s priors for qualitative effects 132
6.3 A linear model for Thiof abundance . . . . . . . . . . 134
6.3.1 A first look at the data . . . . . . . . . . . . . 134
6.3.2 A first simple linear regression . . . . . . . . . 136
6.3.3 Explaining Thiof abundance . . . . . . . . . . 136
6.3.3.1 Selecting relevant covariates . . . . 137
6.3.3.2 Bayesian Inference . . . . . . . . . . 137
6.3.3.3 Model selection . . . . . . . . . . . . 138
6.4 Further reading . . . . . . . . . . . . . . . . . . . . . . 140
6.4.1 Linear modeling . . . . . . . . . . . . . . . . . 140
6.4.2 Model selection . . . . . . . . . . . . . . . . . 141
6.4.3 Variable selection . . . . . . . . . . . . . . . . 142
6.4.4 Bayesian model averaging . . . . . . . . . . . . 143
viii

7 Nonlinear models for stock-recruitment analysis 145

7.1 Stock-recruitment motivating example . . . . . . . . . 146
7.2 Searching for a SR model . . . . . . . . . . . . . . . . 148
7.3 Which parameters? . . . . . . . . . . . . . . . . . . . . 151
7.3.1 Natural parameters . . . . . . . . . . . . . . . 151
7.3.2 Working with management parameters . . . . 153
7.4 Changing the error term from logNormal to Gamma . 157
7.5 From Ricker to Beverton-Holt . . . . . . . . . . . . . . 160
7.6 Model choice with informative prior . . . . . . . . . . 162
7.7 Conclusions and perspectives . . . . . . . . . . . . . . 165

8 Getting beyond regression models 169

8.1 Logistic and Probit regressions . . . . . . . . . . . . . 169
8.1.1 Changes in salmon age at smoltification . . . . 169
8.1.2 Formalization: The Binomial-logit model . . . 172
8.1.3 Bayesian inference . . . . . . . . . . . . . . . . 176
8.1.3.1 Noninformative Zellner priors . . . . 176
8.1.3.2 Posterior distributions . . . . . . . . 177
8.1.4 Model comparison . . . . . . . . . . . . . . . . 178
8.1.5 Discussion . . . . . . . . . . . . . . . . . . . . 180
8.1.5.1 Ecological implications of the results 180
8.1.5.2 About GLM . . . . . . . . . . . . . 182
8.2 Ordered Probit model . . . . . . . . . . . . . . . . . . 184
8.2.1 Skate species repartition . . . . . . . . . . . . 184
8.2.2 Formalization . . . . . . . . . . . . . . . . . . 186
8.2.3 Inferring the ordered multinomial Probit model 190
8.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 191

II More elaborate hierarchical structures 193

9 HBM I: Borrowing strength from similar units 195
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 196
9.2 HBM for capture-mark-recapture data . . . . . . . . . 198
9.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . 198
9.2.2 First phase observation submodels . . . . . . . 199
9.2.3 Second phase observation submodels . . . . . . 200
9.2.4 Latent layers . . . . . . . . . . . . . . . . . . . 201
9.2.5 Results . . . . . . . . . . . . . . . . . . . . . . 204
9.3 Hierarchical stock-recruitment analysis . . . . . . . . . 206
9.3.1 Data . . . . . . . . . . . . . . . . . . . . . . . 207
9.3.2 Model assuming independence . . . . . . . . . 209
9.3.2.1 Likelihood . . . . . . . . . . . . . . 209
9.3.2.2 Prior . . . . . . . . . . . . . . . . . 209
 ix

9.3.2.3 Inference assuming independence . . 211

9.3.3 Hierarchical model with partial exchangeability 212
9.3.4 Results from the hierarchical approach . . . . 215
9.3.5 Prediction of S ∗ and h∗ given the latitude . . 216
9.4 Further Bayesian comments on exchangeability . . . . 218

10 HBM II: Piling up simple layers 221

10.1 HBM for successive removal data with habitat and year 222
10.1.1 Why sampling freshwater juveniles? . . . . . . 222
10.1.2 Study site, sampling design and data . . . . . 223
10.1.3 General model formulation . . . . . . . . . . . 225
10.1.3.1 Observation model . . . . . . . . . . 225
10.1.3.2 Role of the sampling area . . . . . . 226
10.1.3.3 A Normal HBM . . . . . . . . . . . 226
10.1.4 Making ecological sense . . . . . . . . . . . . . 227
10.1.4.1 Correlation between δ and π . . . . 227
10.1.4.2 Effects of year and habitat . . . . . 227
10.1.4.3 Model comparison . . . . . . . . . . 228
10.1.5 DAG and prior specification . . . . . . . . . . 228
10.1.6 Extrapolation to the whole river . . . . . . . . 230
10.1.7 Linking the 0+ population size to the smolts run 231
10.1.8 Bayesian computations . . . . . . . . . . . . . 233
10.1.8.1 Missing data . . . . . . . . . . . . . 233
10.1.8.2 Controlling the flow of information . 234
10.1.8.3 MCMC sampling . . . . . . . . . . . 234
10.1.9 Results of the model comparison . . . . . . . . 235
10.1.10 Estimation with model M0 . . . . . . . . . . . 235
10.1.11 Posterior checking . . . . . . . . . . . . . . . . 237
10.1.12 Estimation of the total number of 0+ juveniles 239
10.1.13 Linking 0+ juveniles to smolts run . . . . . . . 240
10.1.14 Discussion . . . . . . . . . . . . . . . . . . . . 242
10.1.14.1 Changing prior assumptions? . . . . 242
10.1.14.2 Possible improvements . . . . . . . . 244
10.2 Electrofishing with successive removals . . . . . . . . . 245
10.2.1 HBM for electrofishing with successive removals 245
10.2.2 A rapid sampling technique . . . . . . . . . . . 249
10.2.3 Predictive efficiency of the 5-mn protocol . . . 254

11 HBM III: State-space modeling 257

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 258
11.2 State-space modeling of a Biomass Production Model 259
11.2.1 The Namibian hake fishery . . . . . . . . . . . 259
11.2.2 State-space modeling . . . . . . . . . . . . . . 261
x

11.2.2.1 Process equation . . . . . . . . . . . 262

11.2.2.2 Observation equation . . . . . . . . 264
11.2.3 Deriving inferences and predictions . . . . . . 264
11.2.4 The Namibian hake stock dynamics . . . . . . 266
11.2.4.1 Priors and computational details . . 266
11.2.4.2 Inference (Schaefer function) . . . . 268
11.2.4.3 Schaefer versus Fox . . . . . . . . . 272
11.3 State-space modeling of A. salmon life cycle model . . 274
11.3.1 Dynamics process equations . . . . . . . . . . 275
11.3.2 Priors on vital rates . . . . . . . . . . . . . . . 280
11.3.3 Data and observation equations . . . . . . . . 281
11.3.3.1 Updating stage stocks . . . . . . . . 282
11.3.3.2 Prior on the trapping efficiencies . . 285
11.3.3.3 Updating the demographic structure 286
11.3.3.4 River age in smolt run . . . . . . . . 286
11.3.3.5 Sea and River-age in spawner run . 286
11.3.3.6 Sex ratio in spawner run . . . . . . 287
11.3.4 Results . . . . . . . . . . . . . . . . . . . . . . 287
11.3.4.1 MCMC simulations . . . . . . . . . 287
11.3.4.2 Spawners and smolt production . . 287
11.3.4.3 Egg and juvenile production . . . . 290
11.3.4.4 Smolt production . . . . . . . . . . 292
11.3.4.5 Smolt to spawner transitions . . . . 292
11.3.4.6 Forecasting ability . . . . . . . . . . 294
11.3.5 Discussion . . . . . . . . . . . . . . . . . . . . 295
11.3.5.1 A powerful data synthesis method . 295
11.3.5.2 Data and realism . . . . . . . . . . . 295
11.4 A tool of choice for the Ecological Detective . . . . . 297
11.4.1 Linking population dynamics and field data . . 297
11.4.2 Statistics for diagnosis and management . . . . 298

12 Decision and planning 301

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 301
12.2 The Sée-Sélune River network . . . . . . . . . . . . . . 305
12.3 Salmon life cycle dynamics . . . . . . . . . . . . . . . 306
12.3.1 Production of 0+ juveniles . . . . . . . . . . . 307
12.3.2 Smolts . . . . . . . . . . . . . . . . . . . . . . 308
12.3.3 Marine phase . . . . . . . . . . . . . . . . . . . 309
12.3.4 A bivariate decision for angling . . . . . . . . . 309
12.3.5 Assessing parameter values . . . . . . . . . . . 310
12.4 Long-term behavior: Collapse or equilibrium? . . . . . 311
12.5 Management reference points . . . . . . . . . . . . . . 313
12.6 Management rules and implementation error . . . . . 316
 xi

12.7 Economic model . . . . . . . . . . . . . . . . . . . . . 321

12.7.1 Cumulative discounted revenue . . . . . . . . . 321
12.7.2 Revenue per time period . . . . . . . . . . . . 321
12.7.3 Uncertainties . . . . . . . . . . . . . . . . . . . 322
12.7.4 Other indicators of performance . . . . . . . . 323
12.8 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 323
12.8.1 Extinction probabilities . . . . . . . . . . . . . 323
12.8.2 Indicators of performance . . . . . . . . . . . . 324
12.8.3 Cumulated benefits . . . . . . . . . . . . . . . 325
12.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 326
12.9.1 Optimum solution? . . . . . . . . . . . . . . . 327
12.9.2 Expected utility . . . . . . . . . . . . . . . . . 328
12.9.2.1 Decisionmakers’ behavior . . . . . . 328
12.9.2.2 Robustness . . . . . . . . . . . . . . 329
12.9.3 Viability theory . . . . . . . . . . . . . . . . . 329
12.9.4 More than one player in the field . . . . . . . . 330
12.9.4.1 The tragedy of commons . . . . . . 330
12.9.4.2 Cooperation . . . . . . . . . . . . . 330

A The Normal and Linear Normal model 333

A.1 Student distributions . . . . . . . . . . . . . . . . . . . 333
A.1.1 Student distributions as a ratio . . . . . . . . . 333
A.1.2 Student distributions as a mixture . . . . . . . 334
A.1.3 Multivariate Student distributions . . . . . . . 335
A.2 The linear Normal regression model . . . . . . . . . . 336
A.2.1 Likelihood . . . . . . . . . . . . . . . . . . . . 336
A.2.2 Conjugate prior . . . . . . . . . . . . . . . . . 337
A.2.3 Zellner’s G-prior . . . . . . . . . . . . . . . . . 340
A.3 The Zellner Student as a prior . . . . . . . . . . . . . 341

B Computing marginal likelihoods and DIC 343

B.1 Computing predictive distribution . . . . . . . . . . . 343
B.1.1 Method of Marin and Robert . . . . . . . . . . 343
B.1.2 Extension of the method of Raftery . . . . . . 344
B.2 The Deviance Information Criterion . . . . . . . . . . 345

C More on Ricker stock-recruitment 347

C.1 A closer look at the Ricker model . . . . . . . . . . . . 347
C.1.1 Origin . . . . . . . . . . . . . . . . . . . . . . . 347
C.1.2 A deterministic controlled state equation . . . 348
C.1.3 Attractors and chaos . . . . . . . . . . . . . . 349
C.2 Optimal harvesting with Ricker deterministic behavior 350
C.2.1 Standard optimization of a dynamic system . . 350
xii

C.2.2 Hamiltonian in the discrete case . . . . . . . . 351

C.2.3 Searching for a stationary decision rule . . . . 352
C.2.4 Special cases . . . . . . . . . . . . . . . . . . . 352
C.3 Stochastic Ricker model . . . . . . . . . . . . . . . . . 353
C.3.1 A stochastic controlled state equation . . . . . 353
C.3.2 Dynamic stochastic programming . . . . . . . 353
C.3.3 Optimal escapement level . . . . . . . . . . . . 354
C.3.4 Stochastic case: Theory and practice . . . . . . 354
C.4 Equilibrium distribution . . . . . . . . . . . . . . . . . 355
C.4.1 Recalling the Normal characteristic function . 356
C.4.2 SR transition . . . . . . . . . . . . . . . . . . . 356

D Some predictive and conditional pdfs 359

D.1 Predictive for the Logistic model . . . . . . . . . . . . 359
D.2 Predictive for the LogPoisson model . . . . . . . . . . 360
D.3 Full conditional for the categorial probit model . . . . 360

E The baseball players’ historical example 363

E.1 The baseball players’ example . . . . . . . . . . . . . . 363
E.2 Borrowing strength from neighbors . . . . . . . . . . . 364
E.2.1 What about Williams’ skill? . . . . . . . . . . 364
E.2.2 Winning bets . . . . . . . . . . . . . . . . . . . 366
E.3 Fully exchangeable model . . . . . . . . . . . . . . . . 366
E.3.1 Prior distributions . . . . . . . . . . . . . . . . 368
E.3.1.1 Prior on population parameters . . . 368
E.3.1.2 Exchangeable prior latent structure 369
E.3.2 Posterior distributions and borrowing strength 369
E.3.2.1 Posterior distributions . . . . . . . . 369
E.3.2.2 Borrowing strength . . . . . . . . . 371
E.4 Shrinkage effect in the exchangeable structure . . . . . 372

Bibliography 375

Index 403
List of Figures

1.1 The three steps of quantitative modeling . . . . . . . . 4

1.2 Biomass dynamics of an exploited fish stock. . . . . . . 10
1.3 Biomass production model . . . . . . . . . . . . . . . . 13
1.4 Simplified Atlantic salmon life cycle model . . . . . . . 14
1.5 Some acyclic conditional links between three variables. 19
1.6 A simple example of DAG . . . . . . . . . . . . . . . . 21
1.7 Step-by-step construction of a complex model . . . . . 22
1.8 DAG of a prototype Hierarchical Model . . . . . . . . . 24
1.9 DAG of the A. salmon life cycle . . . . . . . . . . . . . 26
1.10 Information retropropagation . . . . . . . . . . . . . . . 28
1.11 The Bayes rule . . . . . . . . . . . . . . . . . . . . . . . 29
1.12 Hierarchical modeling strategy . . . . . . . . . . . . . . 35
1.13 DAG with random effects . . . . . . . . . . . . . . . . . 38
1.14 DAG for a state-space model . . . . . . . . . . . . . . . 39

2.1 Trapping device on the Scorff river . . . . . . . . . . . . 44

2.2 10000 random draws from a Binomial distribution . . . 49
2.3 DAG for a Beta-Binomial model . . . . . . . . . . . . . 50
2.4 Likelihood as a function of π . . . . . . . . . . . . . . . 51
2.5 Priors matter . . . . . . . . . . . . . . . . . . . . . . . . 53
2.6 Beta(a,b) pdf with parameters a=5 and b=10 . . . . . . 55
2.7 The Bayesian crank . . . . . . . . . . . . . . . . . . . . 57
2.8 Prior and posterior pdfs for the Binomial parameter . . 61
2.9 Joint and marginal posterior pdfs for π and y new . . . 62
2.10 Sensitivity of the Beta posterior pdf to prior and data . 63

3.1 A. salmon 0+ juveniles . . . . . . . . . . . . . . . . . . 66

3.2 Histogram of the lengths of juvenile salmon . . . . . . . 68
3.3 Prior knowledge about salmon mean length . . . . . . . 70
3.4 Prior knowledge about the variability of salmon length 72
3.5 Marginal posterior of the difference in lengths . . . . . 79

4.1 The salmon section of the Oir River . . . . . . . . . . . 84

4.2 Marking an anesthetized smolt . . . . . . . . . . . . . . 85

xiii
xiv List of Figures

4.3 The Cerisel trapping facility . . . . . . . . . . . . . . . 86

4.4 DAG for a capture-mark-recapture model . . . . . . . . 87
4.5 Posterior pdf of (π, ν) for capture-mark-recapture . . . 88
4.6 An electrofishing team. . . . . . . . . . . . . . . . . . . 91
4.7 DAG for a successive removal model . . . . . . . . . . . 93
4.8 Posterior pdf of (π, ν) for successive removals . . . . . . 94
4.9 A skipjack tuna. . . . . . . . . . . . . . . . . . . . . . . 97
4.10 Probability that the two tags are a posteriori identical 100

5.1 A spawner . . . . . . . . . . . . . . . . . . . . . . . . . 106

5.2 What the future holds in store . . . . . . . . . . . . . . 108
5.3 Informative Gamma Prior for the stock size κ . . . . . 112
5.4 Informative Beta pdfs for (θ, α, β, τ, δ, π) . . . . . . . . 114
5.5 DAG for the life of a spawner in the Scorff River . . . . 116
5.6 DAG for the interannual model . . . . . . . . . . . . . . 117
5.7 Posterior for basic parameters given year 1995 . . . . . 118
5.8 Posterior distributions of (θ, κ) . . . . . . . . . . . . . . 120
5.9 Posterior of κ based on 5 years . . . . . . . . . . . . . . 121
5.10 Posterior for probability parameters given 5 years . . . 122

6.1 M’bour traditional fishery harbor . . . . . . . . . . . . 126

6.2 Thiof fishing zones in Senegal . . . . . . . . . . . . . . 127
6.3 Residuals versus fitted values of CPUEs . . . . . . . . . 135
6.4 Log-transformed catch as a function of years . . . . . . 136
6.5 Simplified DAG of the Tioff model . . . . . . . . . . . . 139
6.6 Posterior predictive pdf of the log-CPUE . . . . . . . . 140

7.1 Margaree River area. . . . . . . . . . . . . . . . . . . . 146

7.2 Margaree River stock and recruitment data . . . . . . . 148
7.3 Ricker versus Beverton-Holt stock-recruitment curves . 149
7.4 DAG for the Ricker model . . . . . . . . . . . . . . . . 152
7.5 Ricker management-related parameters . . . . . . . . . 154
7.6 Management-related posterior pdf . . . . . . . . . . . . 157
7.7 Ricker model fitted with logNormal noise . . . . . . . . 158
7.8 Ricker model fitted with Gamma noise . . . . . . . . . 160
7.9 LogNormal Beverton-Holt SR model . . . . . . . . . . . 162
7.10 Beverton-Holt posterior pdf . . . . . . . . . . . . . . . . 163
7.11 Checking Importance sampling weight distribution . . . 165

8.1 Three geographical units in Brittany . . . . . . . . . . . 171

8.2 An archived scale from the database . . . . . . . . . . . 172
8.3 Smolt proportions per years . . . . . . . . . . . . . . . 173
8.4 Smolt proportions per regions . . . . . . . . . . . . . . 174
8.5 Smolt sample sizes . . . . . . . . . . . . . . . . . . . . . 175
 List of Figures xv

8.6 DAG for the smolt ratio model . . . . . . . . . . . . . . 177

8.7 Posterior predictive for sea ages . . . . . . . . . . . . . 179
8.8 Posterior predictive for smolt proportions . . . . . . . . 180
8.9 Checking importance sampling weight distribution . . . 181
8.10 Winter, thorny and smooth stakes . . . . . . . . . . . . 185
8.11 Empirical frequencies for the three skate species . . . . 186
8.12 Random mechanism for an ordered categorical response 187
8.13 Posterior joint pdf of the ordered Probit model . . . . . 189
8.14 How well does the model fit? . . . . . . . . . . . . . . . 190

9.1 Marking a spawner . . . . . . . . . . . . . . . . . . . . 199

9.2 DAG for the hierarchical CMR analysis . . . . . . . . . 201
9.3 Posterior mean of the capture probabilities . . . . . . . 202
9.4 Posterior mean of the recapture probabilities . . . . . . 203
9.5 Posterior distribution of the number of spawners . . . . 205
9.6 Posterior predictive of the trapping efficiency. . . . . . . 206
9.7 Location of the 13 index rivers . . . . . . . . . . . . . . 208
9.8 Fitted SR relationships for the 13 monitored rivers . . . 210
9.9 Marginal posterior distributions of log(S ∗ ) . . . . . . . 212
9.10 Marginal posterior distributions of h∗ . . . . . . . . . . 213
9.11 DAG for the hierarchical SR analysis . . . . . . . . . . 214
9.12 Posterior pdfs for α, β, δ, κ, σ and τ . . . . . . . . . . 216
9.13 Posterior pdfs of log(S ∗ ) with latitude as a covariate . 217
9.14 Posterior pdfs of h∗ with latitude as a covariate . . . . 218

10.1 Electrofishing in the Oir River . . . . . . . . . . . . . . 223

10.2 DAG of the baseline model M0 . . . . . . . . . . . . . . 229
10.3 DAG of the model combining the two blocks . . . . . . 232
10.4 Posterior pdfs of the αδi ’s under model M0 . . . . . . . 236
10.5 Posterior pdfs of the βδh ’s under model M0 . . . . . . . 237
10.6 Correlation between δ and π . . . . . . . . . . . . . . . 238
10.7 Posterior predictive checks of χ2 -discrepancies . . . . . 240
10.8 Prediction of the 0+ juvenile population size . . . . . . 241
10.9 Estimates of the smolts run size by cohort . . . . . . . 242
10.10 Survival rate . . . . . . . . . . . . . . . . . . . . . . . . 243
10.11 Electrofishing for abundance indices . . . . . . . . . . . 245
10.12 A hierarchical model for successive removals. . . . . . . 247
10.13 Posterior pdf for successive removal HBM . . . . . . . . 248
10.14 Salmon densities posterior distributions . . . . . . . . . 249
10.15 Model considering independence between sites . . . . . 250
10.16 Model with sites sharing information . . . . . . . . . . 250
10.17 Salmon juvenile estimates versus Abundance Indices . . 251
10.18 DAG for successive removals with AI . . . . . . . . . . 252
xvi List of Figures

10.19 Posterior pdfs for the full AI+ successive removal model 253
10.20 Predictive pdf of Salmon juvenile density versus AI . . 254

11.1 Location of the ICSEAF fishery areas . . . . . . . . . . 260

11.2 DAG for biomass production state-space model . . . . . 266
11.3 Posterior pdfs for the biomass production model . . . . 268
11.4 Updating M SY for the Schaefer model . . . . . . . . . 269
11.5 Quality of fit for the Schaefer model . . . . . . . . . . . 270
11.6 Biomass posterior pdf for the Schaefer model . . . . . . 271
11.7 Biomass : Schaefer versus Fox models . . . . . . . . . . 272
11.8 M SY : Schaefer versus Fox models . . . . . . . . . . . 273
11.9 Atlantic salmon life cycle model . . . . . . . . . . . . . 275
11.10 DAG of the Atlantic salmon dynamics . . . . . . . . . . 276
11.11 Posterior estimates of spawners and smolts . . . . . . . 288
11.12 Posterior estimates of trapping efficiences . . . . . . . . 289
11.13 From eggs to juveniles . . . . . . . . . . . . . . . . . . . 290
11.14 Posterior pdfs for key demographic parameters . . . . . 291
11.15 Bayesian estimates of survival rates . . . . . . . . . . . 293
11.16 One step ahead Bayesian forecasting . . . . . . . . . . . 294

12.1 Fly fishing . . . . . . . . . . . . . . . . . . . . . . . . . 303

12.2 The Sée-Sélune River network . . . . . . . . . . . . . . 306
12.3 Atlantic Salmon controlled Markov chain . . . . . . . . 307
12.4 Probability of extinction . . . . . . . . . . . . . . . . . 312
12.5 Cumulative return rates. . . . . . . . . . . . . . . . . . 317
12.6 Coefficient of variation of Salmon returns . . . . . . . . 325
12.7 Captures versus Salmon Returns . . . . . . . . . . . . . 326

C.1 Ricker stable equilibrium point . . . . . . . . . . . . . . 349

C.2 Ricker limit cycle . . . . . . . . . . . . . . . . . . . . . 349
C.3 Ricker period 4 limit cycle . . . . . . . . . . . . . . . . 349
C.4 Ricker chaotic behavior . . . . . . . . . . . . . . . . . . 349

E.1 Bayesian approach to assess Williams’ ability at bat . . 365

E.2 HBM for the 1970 baseball players. . . . . . . . . . . . 368
E.3 Boxplot of 1970 baseball players’ posterior . . . . . . . 373
List of Tables

3.1 Length of juvenile salmons . . . . . . . . . . . . . . . . 67

5.1 Data to estimate spawning run in The Scorff River . . . 109

5.2 Informative Beta pdfs for (θ, α, β, τ, δ, π) . . . . . . . . 113
5.3 Marginal posterior statistics based on year 1995 . . . . 119
5.4 Posterior correlation between parameters . . . . . . . . 119
5.5 Posterior statistics of parameters based on 5 years . . . 123

6.1 Senegalese artisanal fishery data . . . . . . . . . . . . . 128

6.2 Posterior estimates of the parameters . . . . . . . . . . 138

7.1 Margaree River S-R data . . . . . . . . . . . . . . . . . 147

7.2 Ricker marginal posterior parameters . . . . . . . . . . 153
7.3 Management-related parameter prior . . . . . . . . . . 156
7.4 Posterior for management-related parameters . . . . . . 156
7.5 Ricker Bayesian inference with informative prior . . . . 159
7.6 Posterior statistics with a Beverton-Holt-type model . . 161
7.7 Three competing models with informative priors . . . . 164

8.1 Spending either one or two years at sea . . . . . . . . . 170

8.2 Posterior statistics for θ in model M0 . . . . . . . . . . 178
8.3 Marginal likelihoods for the competing models . . . . . 182
8.4 Bayes Factors for the competing models . . . . . . . . . 182
8.5 Ordered multinomial Probit posterior statistics . . . . . 188

9.1 Capture-mark-recapture data of the Oir River . . . . . 200

9.2 SR features for the 13 monitored A. salmon rivers . . . 207

10.1 Surface area (m2 ) of the sampling sites . . . . . . . . . 224

10.2 Captures at the first and second passes . . . . . . . . . 225
10.3 Prior distributions of parameters . . . . . . . . . . . . . 230
10.4 Surface area by habitat type . . . . . . . . . . . . . . . 230
10.5 Structures of competing models . . . . . . . . . . . . . 235
10.6 Posterior estimates of models M0 and M1 parameters . 239
10.7 Three passes during successive removal electrofishing . 246

xvii
xviii List of Tables

10.8 Abundance Indices on the Nivelle River . . . . . . . . . 250

11.1 Namibian Hake catches . . . . . . . . . . . . . . . . . . 261

11.2 Prior pdfs for the biomass production state-space model 267
11.3 Posterior parameters for the Schaefer model . . . . . . 269
11.4 Prior distribution on parameters. . . . . . . . . . . . . 277
11.5 CMR and sampling data for smolts . . . . . . . . . . . 282
11.6 CMR and sampling data for 1SW spawners . . . . . . . 283
11.7 CMR and sampling data for 2SW spawners . . . . . . . 284
11.8 Key demographic parameter estimates . . . . . . . . . . 292

12.1 Estimates for Salmon life parameters . . . . . . . . . . 311

12.2 Reference points for the Salmon cycle . . . . . . . . . . 315
12.3 10 competing management strategies . . . . . . . . . . 318
12.4 Extinction probabilities under various policies . . . . . 324
12.5 Cumulated benefits under sustainable policies . . . . . 327

E.1 First 45 at bats of the 1970 season . . . . . . . . . . . . 364

E.2 True batting averages and their Bayesian estimates . . 367
Foreword

Statistics? Not for us! Surprisingly, this is still what can sometimes be
heard when discussing statistics with scientists from applied disciplines.
How is it that such people disregard the science of doubt? No doubt that
practitioners possess a deep understanding of phenomenological behav-
ior of the system under study. They also have an acute feeling of its
variability and its complexity. They have elaborated a mental image of
the phenomenon, with operational shortcuts and permanent reformula-
tions. Dare we say: the seed for a model is already set in their minds.
At the same time, they never blindly trust model outputs and agree
that statistics is necessary to depict things that happened in the past
and to make projections about what may occur in the future. However,
they also often complain that the basic statistical procedures are too
simple and not flexible enough to address real problems they have to
struggle with, and one has to acknowledge that instruction manuals for
more advanced statistical tools are rarely funny readings! The present
book might not be exactly a funny reading, but we hope that it could
contribute to make the first steps of statistical modeling and inference
more accessible, and tear-free, to a wide community of practitioners.
In Ecology, interestingly, complex issues have often been the hotbed
of development of innovative statistical methodology. Statistical Ecol-
ogy, a branch of ecological sciences, has proved incredibly productive
in the last thirty years. However, temptation remains to apply ready-
to-use statistical recipes, and frustration can be heavy when the data
and the problem at hand do not fit exactly into the mold of the clas-
sical statistical toolboxes. But, to sacrifice the model realism for the
sake of statistical tractability, or to defer to the sole mathematician the
inference of the more complex model formulation, is that the unavoid-
able dilemma? In both cases, the analyst is deprived from the exciting
intellectual adventures of model making!
This book was written because we believe that Hierarchical Bayesian
Modeling is like a golden key to be given to scientists in Ecology. This
avenue of thought should highly contribute to free their creativity in
designing statistical models of their own.

xix
xx Foreword

In this book, we first recall that the principles of Bayesian statistics

are not difficult to learn and can be intuitively understood. Bayesian
statistical inference is merely learning from data by updating prior prob-
abilistic judgments into posterior beliefs. Bayesian statistical modeling
offers an intuitive avenue to put structures on data, to test ideas, to
investigate several competing hypotheses and to assess degrees of confi-
dence of predictions.
We also illustrate how conditional reasoning is a way to dismantle
a complex reality into more understandable pieces. The construction
of complex models is achieved by assembling basic elements linked via
conditional probabilities that form the core of the hierarchical model
structure. As conditional reasoning is intimately linked with Bayesian
thinking, considering hierarchical models within the Bayesian setting
offers a unified coherent framework for modeling, estimation and predic-
tion.
Key ideas of this book have emerged through discussions with col-
leagues and PhD students. Étienne Prévost spent a whole research career
working on salmons and Bayes, and many chapters are inspired by our
collaborations. We are grateful that he allowed us to present in this book
a lot of examples and data from his work. For the Scorff databases, field
data collection has been carried out under his supervision by the per-
sonnel from the Moulin des Princes experimental station. Many thanks
to the team of Fisheries and Oceans Canada in Moncton, where it all
began. In 2005, we set up there a two-week Bayesian school with our
friend Étienne Prévost. Gerald Chaput, Hugues Benoit and their col-
leagues suggested we write this book. We heard again the same sugges-
tions during the 2006 and 2007 doctoral sessions that we organized for
the Marine Exploited Ecosystems Research Unit (UMR EME) in Sète,
France, to promote Bayesian thinking for modeling ecological data. We
are particularly grateful to Daniel Gaertner, Nicolas Bez, Jean-Marc Fro-
mentin and Frédéric Ménard who encouraged us on the way. We would
like to thank Jean-Luc Baglinière (INRA, UMR ESE, Rennes) without
whom the long series of data on A. salmon populations in the Oir and
Scorff Rivers would not exist, and all the staff of the Unité Expérimentale
d’Écologie et d’Écotoxicologie (INRA, U3E, Rennes), in particular Nico-
las Jeannot and Frédéric Marchand for their invaluable work in the field
collecting the salmon trapping data in the Scorff and Oir Rivers.
There are many people who also deserve our thanks, although they
did not know they were so actively influencing us and contributing to
this book, such as our families, our mentors, our PhD students and also
the many collaborators from our research institutions, which fostered a
nurturing environment for the whole lifetime of this book project. We
also had the great fortune of counting on our long-standing friendship
 Foreword xxi

with Jacques Bernier and Lucien Duckstein for useful discussions and
careful proofreading.
Jérôme Guitton patiently helped us to master enough html program-
ming language to implement our website hbm-for-ecology.org. This web-
site gives many of the datasets, R and WinBUGS or OpenBUGS codes
that we used to derive inference and figures in the book.

Éric Parent and Étienne Rivot

Paris, France
This page intentionally left blank
 Part I

Basic blocks of Bayesian

modeling and inference
for ecological data

1
This page intentionally left blank
Chapter 1
Bayesian hierarchical models in
statistical ecology

Summary
The Salmon life cycle and the biomass production model exemplify
the key idea of this book: Hierarchical Bayesian Modeling (HBM) is a
Directed Acyclic Graph (DAG) modeling technique with the capacity to
cope with high-dimensional complex models typically needed for ecolog-
ical inferences and predictions ([64]; [65]; [67]; [75]; [313]).
HBM works through conditional decomposition of high-dimension
problems into a series of probabilistically linked simpler substructures.
HBM enables to exploit diverse sources of information to derive infer-
ences from large numbers of latent variables and parameters that de-
scribe complex relationships while keeping as close as possible to the
basic phenomena.
Based on these two motivating case studies from fisheries sciences,
we detail the three basic layers of hierarchical statistical models ([75];
[76]; [312]; [313]):

1. A data level that specifies the probability distribution of the ob-

servables at hand given the parameters and the underlying pro-
cesses;
2. A latent process level depicting the various hidden ecological mech-
anisms that make sense of the data;
3. A parameter level identifying the fixed quantities that would be
sufficient, were they known, to mimic the behavior of the system
and to produce new data statistically similar to the ones already
collected.

HBM stands out as an approach that can accommodate complex sys-

tems in a fully consistent framework and can represent a much broader
class of models than the classical statistical methods from ready-to-use

3
4 Introduction to Hierarchical Bayesian Modeling for Ecological Data

toolboxes. Finally, we use graphical modeling techniques as a guiding

thread to announce the contents of the book’s chapters, starting with
simple stochastic structures (elementary blocks) in Part I that will pro-
gressively be assembled to lead to the more elaborate hierarchical models
of Part II.

1.1 Challenges for statistical ecology

1.1.1 The three steps of quantitative modeling
Quantitative modeling is typically motivated by knowledge and ac-
tion: i) modeling seeks to improve our understanding of a phenomenon
(cognitive objective); ii) many applied ecological questions require mod-
els as a tool to derive predictions (applied objectives) (see Fig. 1.1).

FIGURE 1.1: The three steps of quantitative modeling. Steps 1 and 2

correspond to the learning phase. The little character on the left side is
inspired from the Ecological Detective in the book of R. Hilborn and M.
Mangel ([136]).

1. Step 1. Propose a tentative model for the process under consider-

ation.
 Bayesian hierarchical models in statistical ecology 5

Models are used to synthesize knowledge and speculate about the

system of interest. But modeling is also governed by the objec-
tives of the study and some subjectivity from the analyst ([236]).
For instance, plants or animals in ecosystems can be represented
one by one by dynamically interacting rule-based agents (well-
documented examples of individual-based models for ecology can
be found in [125] and [122]). In such models, local heterogeneity
among entities is built in for studying, by simulation, the global
scale consequences of a particular trait of ecological interest, such
as body size or some behavioral characteristics such as an alter-
ation in reproductive strategy. On the contrary, in a population
model ([164]), like the biomass production model depicted in Fig.
1.2, the characteristics of the population are averaged together and
the model attempts to depict the evolution of the population as
a whole. Knowledge is generally understood as structural quanti-
tative knowledge with mathematical equations that formalize hy-
potheses and quantify interactions between variables. Models rely
on i) key variables of interest with respect to the objectives of the
modeling enterprise, and ii) the main cause-to-effects interactions
between these variables which are built from a deductive reasoning.
Interactions mimic either deterministic or stochastic cause-to-effect
relationships. Consider as an example, a π = 40% survival rate be-
tween two stages A and B of some population with size NA and NB
like the juvenile to smolt transition in the salmon life cycle of Fig.
1.4; a deterministic model would state that the population at stage
B is 40% of the one at stage A while a stochastic model would say
that population size NB is some random variable, distributed as
a Binomial with order NA and probability 0.4 (mathematical lay-
outs for the Binomial distribution are given in Chapters 2 and 4).
Because natural systems are often poorly understood, there may
be alternative competing hypotheses. Equations generally involve
unknown quantities (e.g., unknown parameters or hidden system
states) that must be estimated: for instance the survival rate π
from state A to state B in the previous example might be left un-
known and one would expect this quantity to be estimated from
the available data.
Knowledge also involves qualitative knowledge or expertise ([223]).
From past experience when dealing with a similar situation in an-
other location, use of proxies or a long life understanding of the
biological phenomenon, scientists may sometimes specify valuable
best guesses and credible intervals for unknown model quantities.
For instance, referring to the previous example, some probabilistic
judgment could be made a priori for the unknown survival rate π
6 Introduction to Hierarchical Bayesian Modeling for Ecological Data

such as, “considering the number of potential predators and the

length of juveniles, the probability of survival π is most likely be-
tween 0.25 and 0.5, unlikely below 0.20, but extremely rarely above
0.8.”
2. Step 2. Learning from observations.
In the second step, one seeks to learn from the data at hand, i.e.,
to use the data to update the prior knowledge that has been coded
in Step 1. This is often referred to as confronting the model to
the data or more simply as model fitting or inference. Objectives
are i) to compare the credibility of alternative competing hypothe-
ses with respect to the observations; ii) to estimate the unknown
quantities of the model and the associated uncertainty condition-
ally upon the data at hand. This step relies heavily on statistical
science to infer pattern and causality from the data. By comparison
with the deductive way of thinking that works with cause-to-effect
relationships, inference relies on inductive reasoning and promotes
a reverse way of thinking, working on effect-to-cause relationships.
Depending on the field of application, it is sometimes termed in-
verse modeling (engineering), data assimilation (geosciences), sta-
tistical learning (neural networks), among others.
3. Step 3. Use the model in a deductive way as a decision tool.
In this third step, the model can be used in a deductive way as
a tool for prediction. This generally takes the form of simulations
designed to explore the response of the system under different sce-
narios. Typically, models for renewable resource assessment are
used to establish a diagnostic about past evolution and the present
state of the resource, to estimate key management reference points
such as the sustainable level of exploitation, and to predict the po-
tential evolution of the resource under several scenarios for future
exploitation or evolution of the environment. As an example of
choice, the 2000 special report of the International Panel on Cli-
mate Change of Working Group III includes a summary for policy
makers that describes long-term greenhouse gas emissions depend-
ing on different future storylines with contrasted demographic, so-
cioeconomic, technological and environmental developments.
However, as models become more complex, they require more pa-
rameters and they inevitably become more sensitive to parameter
values. Before using the model in a predictive mode, a critical issue
for the Ecological Detective ([136]) is to have a precise and rigorous
quantification of the uncertainties in the different components of
the model ([133]; [184]), not least because the general adoption of
 Bayesian hierarchical models in statistical ecology 7

the precautionary principles and precautionary approaches to man-

agement of ecological systems ([67]; [106]) requires an assessment
of the risks that will occur as a result of the different scenarios.
Risk can be defined as the probability distribution of the conse-
quences of an undesirable event, (e.g., irreversible environmental
damage). As such, risks are inevitable consequences of uncertainty.
If there is no uncertainty, the concept of risk is irrelevant because
the probability of any event degenerates to either 0 or 1.

Here, the first two steps have been isolated for the purpose of clarity.
But this is somewhat artificial as they are practically embedded within
an iterative process in which the confrontation of the proposed model to
the available data (Step 2) may lead to reconsider the hypotheses that
were made in Step 1.
The relative importance of each of the three steps may vary depend-
ing upon the modeling objectives. Step 1 will receive primary attention
for fundamental ecological studies. If the management purpose is of pri-
mary interest, more modeling effort will be devoted to Step 3.
Throughout this book, we will exemplify how the Bayesian setting
enables one to combine these three steps (model design, model fitting
and predictive simulations) within a single unified and rigorous statistical
framework. A deeper discussion on the Bayesian setting is postponed to
Section 1.3.2.

1.1.2 Models are intrinsically complex, highly uncertain

and partially observed
The growing interest in solving ecological problems has provided a
new impetus for the development of complex models, which raised statis-
tical challenges for quantitative modeling analyses. Complexity in ecolog-
ical models arises from two main reasons: the motivation to get enough
realism and the need to account for various sources of uncertainty.

1.1.2.1 Increasing model realism and dimension

Improving the realism of ecological models inherently results in an in-
crease of model dimension and complexity. Various components are now
increasingly used in recent ecological models, such as dynamic or spatial
components (e.g., spatial heterogeneity, migration patterns), and multi-
ple interactions forms (e.g., nonlinear intra-specific density dependence,
inter-specific interactions) often combined with the effects of covariates
to trigger the main influences of the environment (e.g., changing forcing
conditions as a climate under global warming).
8 Introduction to Hierarchical Bayesian Modeling for Ecological Data

1.1.2.2 Accounting for different sources of variability and un-

certainty
The field of environmental sciences is becoming increasingly aware
of the importance of accurately accounting for the multiple sources of
variability and uncertainty when modeling environmental processes and
making forecasts. This development is motivated in part by the desire
to provide an accurate picture of the state of knowledge of ecosystems,
and to be able to better assess the quality of predictions and decisions
resulting from the models. Models are functional caricatures of (com-
plex) natural systems which are not easily observed. Uncertainties in
ecological models are inherent to our incomplete knowledge, due to both
the difficulty to gather exhaustive observation data on natural systems
and to our relatively poor ability to propose mechanistic models with
good predictive power. Uncertainty is hence a rather difficult notion be-
cause it encompasses the elements that fluctuate due to unpredictable
or unaccessible varying factors (uncertainty by essence) and those that
are just partially known (uncertainty by ignorance).
In ecology, uncertainties stem from three main sources (see [63], [133],
and [173], but also [252] for a general discussion about uncertainty): i)
model errors; ii) process stochasticity; iii) observation errors.

1. Model errors
As any model is a functional simplification of a real process, it thus
provides an imperfect and a potentially misleading representation
of how the real system works. Model errors (also referred to as
structural errors) reflect our ignorance about the very nature of
the system. Model error may of course yield major consequences if
the model is used for forecasting (in particular when extrapolating
beyond the range of the data that have been collected). The role
of hypotheses is to restrict the number of plausible models. By
convention, a zero probability is put on the models that do not
belong to the set delimited by the hypotheses. The model error
term is to be quantified only for the members of this remaining
set.
2. Process stochasticity
Models with deterministic process are useful to learn from the in-
teraction of complex processes. However, when confronting models
with data, models must account for unpredictable variability or
stochasticity that cannot be explained by deterministic processes.
Process stochasticity can take different forms of which the most
common are demographic stochasticity (due to random differences
between individuals or statistical groups of individuals with the
 Bayesian hierarchical models in statistical ecology 9

same characteristics) or environmental stochasticity (due to un-

predictable random variations of parameters) ([100]).

3. Observation or measurement errors

Observation or measurement errors stem inherently from the im-
possibility to accurately and exhaustively observe Nature. They
are the consequence of imperfection in the data records. Field ob-
servations are rarely issued from an optimized sample scheme, and
the expense of data collection restrict the collection of as much
data as might be desirable. Data are often incomplete, and the
amount of sampling and measurement errors resulting from either
sampling, measurement or estimation mistakes is often large and
unknown. In the words of J. Schnute ([272]), “counting fish is like
counting trees except they are invisible and they move.”Hence, the
Ecological Detective in Figure 1.1 has to cope with indirect, noisy
and incomplete observations which have to be put in coherence
with some hypotheses about the hidden process that underlines
the observed phenomenon.

In statistical ecology, uncertainty due to model errors generally rep-

resents the major source of uncertainty ([252]; [277]). Model weighting
or model averaging offers some tools to handle model errors ([139]; [182];
[287]). In statistical models, uncertainty due to process stochasticity is
generally irreducible, in the sense that it does not decline when the
sample size increases. The uncertainty due to estimation errors embed
uncertainty stemming from measurement errors, and results from impre-
cision (quantified by statistical inference) when estimating the unknown
of the systems from the available data. In contrast with the uncertainty
due to process stochasticity, estimation errors can be reduced by improv-
ing the data collection procedure. In theory, they decline asymptotically
when more data become available (only under the stringent hypotheses
of a perfect model and a stationary phenomenon).

1.1.3 Embedding ecological models within a statistical

approach
As these different sources of errors and variability are conceptually
and quantitatively different, modern statistical ecology seeks to propose
quantitative tools to separate out the uncertainty due to the process
stochasticity to the one due to the observation sampling process. This
requires flexible approaches to fusing models with data, approaches that
can accommodate uncertainties in the way ecological processes operate
and the way we observe them.
10 Introduction to Hierarchical Bayesian Modeling for Ecological Data

1.1.4 Motivating examples

The Biomass production model
Biomass production models ([137]; [244]) have a long history in quan-
titative fisheries sciences and continue to prove useful in stock assessment
under a data-poor environment. Although they are based on a crude sim-
plification of the fish population dynamics, they remain a tool of choice
for providing fisheries management advice in situations in which:

• There are insufficient age- and size-composition data to permit the

use of assessment methods based on age- or size-structured models;

• One is mostly interested in estimating the maximum sustainable

yield (CM SY ), and its associated biomass (BM SY ) (mathematical
details about M SY calculations are postponed to Chapter 7, see
Fig. 7.5);
• It is worthwhile predicting the biomass levels and surplus produc-
tion for the future and sometimes in the past, to impute missing
data records. The latter is particularly important to test the perfor-
mance of alternative management scenarios when efforts are being
made to control the level of catches and promote a sustainable
harvest policy.

FIGURE 1.2: Biomass dynamics of an exploited fish stock.

The motivating example of the Biomass production model illustrates

the three complementary steps (Fig. 1.1).
 Bayesian hierarchical models in statistical ecology 11

1. Step 1. Propose a tentative model for the process under concern.

The dynamic biomass production model is a crude but useful sim-
plification of the dynamics of a harvested fish population. The key
variable of interest is the total biomass of the population at each
time. The dynamics is modeled in discrete time, most often on a
year-to-year basis. The biomass at the beginning of time step t + 1,
denoted Bt+1 , is obtained from Bt through a rather simple balance
equation:
Bt+1 = Bt + g(Bt ) − ct (1.1)
where ct is the observed harvest between t and t+1 and g(Bt ) is the
production function. It quantifies the balance between recruitment
(arrival of new individuals in the stock biomass), growth (weight),
natural mortality, and possibly emigration-immigration. The most
classic choice for the production function is the logistic one with
two parameters, the population growth rate r and the carrying
capacity K, which are generally hypothesized constant over time
(see also Chapter 11 for more details):

Bt
g(Bt ) = r · Bt · (1 − ) (1.2)
K
A Log-Normal random noise term is generally added to capture the
biological variability due to (unpredictable) environmental varia-
tions:
Bt+1 = (Bt + g(Bt ) − ct ) · et+1
(

iid
(1.3)
εt+1 ∼ N ormal(0, σ 2 )
with the ∼ sign meaning that t+1 , the logarithm of the perturba-
tion, is a normally distributed N (0, σ 2 ) random term standing for
the environmental noise. Mathematics for the Normal Distribution
will be developed in Chapters 3 and 6.
2. Step 2. Learning from observations.
Let t = 1, ..., n denote the epochs for which observations are
available. Available data in that case typically consist of a series
of observed catches c1 , ..., cn and of abundance indices i1 , ..., in ,
the latter are often assumed proportional to the current biomass
it = q · Bt , ∀t ∈ {1, ..., n}.
Abundance indices can be of different nature. They can be derived
from commercial data such as commercial catches per unit of fish-
ing effort, or from a scientific survey such as scientific trawling or
acoustic surveys. The abundance indices it=1:n are often assumed
proportional to the current biomass it = q · Bt , ∀t ∈ {1, ..., n} with
12 Introduction to Hierarchical Bayesian Modeling for Ecological Data

a catchability parameter q, hypothesized constant over time. A

common, although simplifying assumption is that observed catch
and abundance index of each year t are respectively related to the
unobserved true catch and unobserved biomass through stochastic
observation models:
(
ct = Ct · eω1,t
(1.4)
it = q · Bt · eω2,t

with ω1,t and ω2,t normally distributed N (0, τ1,o 2 ) and N (0, τ2,o 2 )
random terms describing the uncertainty in the observed catches
and abundance indices due to measurement and sampling error
(observation error).
Fisheries scientists would typically rely on the available knowledge
(expertise and past data) and observation (data at hand) to pro-
vide answers to the following questions of interest: i ) Is the logistic
growth function appropriate or does another form of production
function fit the observed data better? ii ) What are the credible
values for the parameters (r, K) and the associated uncertainty?
Can the growth rate r and the carrying capacity K be elicited from
some probabilistic prior judgmental expertise? iii ) What are the
credible values for the historical trajectory of the biomass level,
say B1 , ..., Bn and what is the level of the Biomass depletion over
the time series B B1 ?
n

Figure 1.3 relies on a graphical representation based on different

layers to illustrate the conceptual difference between the stochas-
tic process for the biomass dynamics and the sampling model for
the observations. Statistical inference aims at using the informa-
tion contained in the data (it , ct )t=1:n to learn about the unknown
biomass dynamics.
3. Step 3. Use the model in a deductive way as a decision tool.
Quantities that interest the natural resources manager, such as the
management reference points related to long-term equilibrium, can
be directly derived from the parameters (r, K) of Eq. (1.2):

r·K

 CM SY =

4 (1.5)

B K
M SY =
2
What are credible values for the management reference points
CM SY and BM SY and their associated uncertainty? Do they match
 Bayesian hierarchical models in statistical ecology 13

FIGURE 1.3: Biomass production model. Hierarchical representation

to explicitly disentangle the stochastic model for the hidden dynamics
from the sampling model for the observations. The arrow from the hidden
dynamic to the observations represents the sampling process giving rise
to observations (it , ct )t=1:n .

the common sense of fishery scientists? Fisheries scientists may

also be interested in deriving predictions of future trajectories
of the biomass and catches over time t = n + 1, ..., n + k un-
der alternative management scenarios. Simulations typically aim
at comparing different harvest control rules, e.g., different fishing
mortalities Fn+1 , ..., Fn+k leading to a series of catches Cn+1 =
Fn+1 Bn+1 , ..., Cn+k = Fn+k Bn+k . Implementation uncertainties
can also be modeled at this stage to evaluate the consequences
of the unavoidable discrepancies between a recommended manage-
ment policy and its actual implementation in operational practice.
Predictive simulations are also of interest to help implement ex-
perimental designs. Suppose for instance that, due to budget limi-
tations, only half of the scientific campaigns previously planned in
the next ten years will be undertaken. When and how should the
measurements be taken to make the best of the future information?
14 Introduction to Hierarchical Bayesian Modeling for Ecological Data

Atlantic salmon stage-structured population dynamic model

As a second example, we develop a more elaborate model, based on
age-structured salmon data. Stage-structured population dynamics mod-
els have become standard tools for analyzing population dynamics ([50];
[51]; [300]). They are grounded on biological fundamentals and provide
some flexibility to mimic complex dynamic behavior. Such models can
also be employed to evaluate the performance of a wide range of manage-
ment options. Rivot et al. ([259]) consider a model for Atlantic salmon
populations dynamics in the northwest of France which focuses on two
of the fundamental events in the life history of A. salmon: Smoltification
and spawning migration (Fig. 1.4).

FIGURE 1.4: Simplified Atlantic salmon life cycle model with only one
sea-age class. Spawners Sp1 return and spawn one year before spawners
Sp2 of the same cohort.

Spawning occurs in the river in late fall and early winter and most
spawners die shortly after their first reproduction. Eggs hatch in the
gravel and alevins emerge during the next spring. After emergence,
young-of-the-year salmon (denoted 0+ in Fig. 1.4) spend one or sev-
eral years in the river before undergoing smoltification in the spring. At
this time, they migrate downstream to the sea as smolt. Adults return to
their home river for spawning after one or several years at sea. In France,
a major portion of the juveniles become smolts at one year of river age
(1+ Smolt, denoted Sm1 in Fig. 1.4), in the spring following their emer-
gence as young-of-the-year salmon, leaving behind the smaller juveniles
(P arr1) to spend an additional year in the river before seaward migra-
tion (2+ Smolt, denoted Sm2 in Fig. 1.4). Fish from the two smolt age
classes return as spawners after one or two years at sea. Other possible
life histories (early maturation in freshwater before seaward migration,
 Bayesian hierarchical models in statistical ecology 15

adults spending more than 2 winters at sea, spawners surviving after

spawning) are rare and only have a negligible influence on population
dynamics.
In this motivating example, we will consider a simplified life cycle
model by forgetting the variability of sea age classes and consider that
fish from the two smolt age classes return as spawners after only one year
at sea as shown in Fig. 1.4 (Sp1 and Sp2, respectively). Two additional
simplifying hypotheses are made: i ) We suppose that the life history is
not an inheritable characteristics; ii ) We use a common life cycle for
male and female fish.
Below we illustrate what could be the three complementary steps of
the quantitative modeling approach (Fig. 1.1) applied to this A. salmon
population dynamics model.

1. Step 1. Propose a tentative model for the process under consider-

ation.
Hereafter a mathematical representation of the life cycle model is
proposed in Fig. 1.4. The model is built on a discrete, yearly basis
time step. The number of spawners, eggs, young-of-the-year, pre-
smolts, 1+ smolts and Parr 1 and 2+ smolts at each time step t
are denoted Spt (Spt = Sp1t + Sp2t ), Wt , 0+t , P Smt , Sm1t , P 1t
and Sm2t , respectively.

(a) Spawners → Eggs. The number of eggs spawned by the adults

returning in year t, Wt , is modelled as a deterministic func-
tion of the number of spawners, the proportion of females pf
and of the mean fecundity of these females denoted f ec, both
considered as known and constant over time:

Wt = Spt · pf · f ec (1.6)

(b) Eggs → 0+ juveniles. A density-dependent process is used to

model the freshwater production of juveniles resulting from
the reproduction of the spawners returning in year t. Den-
sity dependence is modelled by the widely used dome-shaped
Ricker curve with unknown parameters (α, β). Environmen-
tal variability renders the stock-recruitment process stochas-
tic. Again, this is classically introduced via independent and
identically distributed LogNormal errors:

0+t+1 = α · Wt · e−β·Wt · et

(

iid
(1.7)
εt ∼ N (0, σ 2 )

As in Eq. (1.3), t is a normally distributed N (0, σ 2 ) random

16 Introduction to Hierarchical Bayesian Modeling for Ecological Data

term standing for the environmental noise (see also Chapter

7).
(c) 0+ juveniles → Smolts. The young-of-the-year 0+t+1 will
then survive to the next spring of year t + 2 as pre-smolts
P Smt+2 , with probability γ0+ (considered as invariant over
time). Survival can be modeled as a Binomial process that
captures the demographic stochasticity:

P Smt+2 ∼ Binomial(0+t+1 , γ0+ ) (1.8)

In each cohort, a proportion θSm1 (also assumed time invari-

ant) of the pre-smolts will migrate as 1+ Smolts, the remain-
ing part will stay one additional year as 1+ Parrs. Life history
choice for smoltification is also commonly modeled using a Bi-
nomial distribution:

Sm1t+2 ∼ Binomial(P Smt+2 , θSm1 ) (1.9)

1+ Parrs will survive (with a survival rate γP arr1 ) and will mi-
grate as 2+ Smolts. Demographic stochasticity in the survival
of resident parrs is also modelled using a Binomial distribu-
tion:

Sm2t+3 ∼ Binomial(P arr1t+2 , γP arr1 ) (1.10)

(d) Smolts → Returning spawners. Hypothesizing a strict homing

of adults to their native stream, the smolt-to-spawner transi-
tion can be modeled as the result of Binomial processes with
γSm the survival probability at sea, considered as invariant
over time and between the two smolt age classes:
(
Sp1t+3 ∼ Binomial(Sm1t+2 , γSm )
(1.11)
Sp2t+4 ∼ Binomial(Sm2t+3 , γSm )

The spawners returning in year t + 3 and t + 4 will contribute

to the reproduction of year t + 3 and t + 4, respectively.

2. Step 2. Learning from observations.

All parameters (α, β, σ, γ0+ , θSm1 , γP arr1 , γSm ) in the A. Salmon
stage-structured population dynamic model are usually unknown.
They could ideally be estimated from time series of observations
of the number of fish in each of the development stages. However,
these numbers are generally not directly observable. For instance,
the true number of 1+ Smolt migrating downstream each year t
 Bayesian hierarchical models in statistical ecology 17

is unknown. Typically, only partial knowledge is available through

a trapping experiment. The available observations will then con-
sist of a number of smolts caught in a downstream trap, denoted
CSm1,t . Assuming a standard Binomial trapping experiment with
trap efficiency πSm (considered constant), such an observation pro-
cess follows a Binomial distribution:

CSm1,t ∼ Binomial(Sm1t , πSm ) (1.12)

Similar Binomial counting processes with eventually different cap-

ture efficiencies can be connected at stages Sm2, Spawners Sp1
and Sp and 0+ juveniles, thus providing information on the dy-
namics of all development stages but accounting for the sampling
uncertainty. Statistical inferences will consist of using this informa-
tion to learn about the A. salmon population dynamics. In other
words, we want a probabilistic judgment concerning the various
stages and parameters given the data. Chapter 11 is devoted to
this analysis.
3. Step 3. Use the model in a deductive way as a decision tool.
Although the recreational fishery (angling) is often the main source
of salmon exploitation, associated harvest rates could be rather
high (near 50%) and have different harvest control rules (limited
fishing periods, quotas, etc.) which can be used to regulate ex-
ploitation. Such a population dynamic model can be a useful tool
to assess the performance of different management strategies. This
will be developed extensively in Chapter 12.

1.2 Conditional reasoning, graphs and hierarchical

models
1.2.1 Use conditional probability distributions to model
probabilistic transitions
Conditional models are networks of components (variables, also
called nodes in graph theory jargon), but many of the nodes are un-
known. Those variables are linked by deterministic or probabilistic con-
ditional dependencies ([152]). In this book, conditional probability distri-
butions are used to mimic uncertainty and stochastic influences. We will
use the Gelfand’s ([115]) bracket notation for probability distributions.
18 Introduction to Hierarchical Bayesian Modeling for Ecological Data

Let
[V1 | V2 ]
denotes the probability of event V1 given the event V2 has occurred.
It states that, given V1 , the issue V2 is uncertain and described by
probabilistic bets. The bracket notation will be used indifferently for
probability distributions of discrete variables and for probability density
functions (pdf) of continuous variables.

Example (continued): A. salmon population dynamic model

The A. salmon stage-structured model can be viewed as a conditional
model. Following our bracket notation above, the Eggs → 0+ juveniles
transition (Eq. (1.7)) can be written as a conditional probability dis-
tribution under the form [0 +t+1 |Wt , α, β, σ]. Here, the conditional pdf
represents Log-Normal environmental variability.
All mechanisms of the 0+ juveniles → Smolts transitions were mod-
eled as Binomial distributions that can be written as conditional prob-
ability distributions [P Smt+2 |0+t+1 , γ0+ ], [Sm1t+2 |P Smt+2 , θSm1 ] and
[Sm2t+3 |P arr1t+2 , γP arr1 ]. Here, the conditional pdf represents demo-
graphic stochasticity.
The Binomial smolt-to-spawner survival transitions can also be writ-
ten as conditional probability distributions [Sp1t+3 |Sm1t+2 , γSm ] and
[Sp2t+4 |Sm2t+3 , γSm ].

1.2.2 Graphical models

Graphical models, which are often called Bayesian networks or Bayes
nets, are a useful metaphor for such conditional models ([73]). The use
of graphical models significantly improves the modeling process in two
ways: i) They are useful in the two tasks of statistical modeling ([285]),
deductive way of thinking when designing the model structure (cause-to-
effect) and inference when seeking to learn about the unknowns (effect-
to-cause); ii) They help visualize the factorization of complexity; subse-
quently, they proved a valuable tool for catalyzing interactions between
experts during the process of model building.
Figure 1.5 sketches the possible basic conditional relationships be-
tween three variables V1 , V2 , V3 (to get all configurations, consider all
permutations since the ordering 1, 2, 3 does not matter).
In graphical models, conditional reasoning is efficiently summarized
and pictured by means of ellipses, boxes and arrows. Graphical conven-
tions help to explain the statuses of the variables under the inferential
context, and the conditional dependencies and independencies between
the variables: i) Ellipses are reserved for variables that are defined by a
 Bayesian hierarchical models in statistical ecology 19

FIGURE 1.5: Some acyclic conditional links between three variables.

probability distribution. Nonobserved variables are represented as white

ellipses, and observed variables (data) as shaded ellipses. An ellipse with
no parents is often called a parameter ; ii) Shaded rectangular boxes can
be used to represent fixed constant or covariates which may physically
help to understand the structure of the model; iii) Arrows are used to
model dependencies between the variables. Solid line arrows represent
stochastic transitions between variables, and dotted line arrows repre-
sent logical transition (deterministic). The direction of the arrows follows
the line of reasoning the modeling phase ([215]).
In these graphs, no cycle is allowed, so that nodes are unambigu-
ously defined conditionally to their parents, following the arrows. The
very stringent conditions for the existence of a joint distribution [V1 , V2 ],
when the two conditionals [V1 |V2 ] and [V2 |V1 ] are given (i.e., there ex-
ists a cycle in the previous graph), are detailed in [8]. Graphical models
conveniently designed without cycle via successive conditioning links are
called Directed Acyclic Graphs (DAG). They have proved to be valuable
tools for many ecological studies ([34]; [127]; [167]; [171]; [188]; [203];
[239]; [255]; [257]; [259]; [303]; [304]).
20 Introduction to Hierarchical Bayesian Modeling for Ecological Data

1.2.3 Conditional reasoning and factorization of the

complexity
Many components may interact in a model and global complexity
emerges from these local interactions. In a stochastic world, the com-
plete interactions are described by the full joint distribution of all the
components. If the model is formed from n components or random vari-
ables V1 , ..., Vn , the full joint distribution of V1 , ..., Vn is written as:

[V1 , ..., Vn ] (1.13)

However, the global complexity is made from a network of much sim-

pler local interactions. Each variable Vi has been introduced in turn by
considering its dependence directly upon only a reduced number of vari-
ables (the direct parents of Vi ). The global complexity, materialized by
the full joint probability distribution of all the variables, becomes easily
tractable when seen as the combination of these small local interactions.
Following the arrows from these parameter nodes down to the other
nodes of the DAG, the joint distribution of Eq. (1.13) can be factorized
into more simple parent-child interactions:
n
Y
[V1 , ..., Vn ] = [Vi |pa(Vi )] (1.14)
i=1

The parameters have been previously defined as variables without

parents; hence, they cannot be defined as conditional distributions in
Eq. (1.14) and we must make the convention that pa(Vi ) is the null set
for the i’s corresponding to parameters. Distributions for these entry
nodes of the graph are typically called a priori distributions or priors.
It is worth noting that the interpretation of probabilistic conditional
models combines naturally with the Bayesian paradigm that quantifies
uncertainty by means of probabilistic judgments ([26]; [82]; [153]; [268];
[299]). As random variables generated by priors, parameters naturally
appear as white ellipses in the DAG and the corresponding terms in
Eq. (1.14) collapse to [V ∗ |pa(V ∗ )] = [V ∗ ], and Eq. (1.14) writes:
Y
[V1 , ..., Vn ] = [V ∗ ] × [Vi |pa(Vi )] (1.15)
Vi 6=V ∗

Finally, one will merely find the various elementary conditional con-
figurations illustrated in Fig. 1.5 as the small pieces of interaction issued
from such decomposition (Eq. (1.15)) of the joint distribution.
Constants and forcing conditions introduced into the analysis as de-
terministic covariates have no probabilistic status (conversely to the ran-
dom variables V1 , ..., Vn ) and should not in principle be considered in
 Bayesian hierarchical models in statistical ecology 21

V1 V2

V3

V4

FIGURE 1.6: A simple example of DAG with four variables and one
covariate. V1 and V2 have no parents and are considered as param-
eters with independent unconditional distributions [V1 ] and [V2 ]. The
full joint distribution has been designed according to the factorization
[V1 , V2 , V3 , V4 |X] = [V1 ] × [V2 ] × [V3 |V1 , V2 ; X] × [V4 |V3 ].

probabilistic computations. However, it is often useful to make them ap-

pear in a DAG in order to precisely detail the model structure. Yet, note
that we used shaded rectangular boxes instead of ellipses as graphical
conventions to emphasize their particular status. Once they appear in
the DAG, the full joint distribution of the model is implicitly conditioned
by the constants and covariates which are part of the deterministic model
structure. As far as the bracket notation is concerned, they will be subse-
quently mentioned in the conditioning side of a probability distribution
when necessary (but one may also wish to separate them by a semicolon
from the random variables). Figure 1.6 provides a very simple example
of conditional model with four variables (V1 , V2 , V3 , V4 ) (in which V1 and
V2 are parameters) and one covariate X.

Example (continued): A. salmon population dynamic model

Figure 1.7 gives simple examples of DAG for the different demo-
graphic transitions that characterize the A. salmon life cycle, and pro-
vides the DAG for the full life cycle model running for one cohort orig-
inated from eggs spawned at year t. The DAG is built by connecting
22 Introduction to Hierarchical Bayesian Modeling for Ecological Data

a) b) c)

α β σ γ0+ θSm1

Wt 0+t+1 0+t+1 PSmt+2 PSmt+2 Sm1t+2

Parr1t+2
d) e)
γSm
γParr1

Sm1t+2 Sp1t+3

Parr1t+2 Sm2t+3 Sm2t+3 Sp2t+4

α β σ γ0+ θSm1 γParr1 γSm

Wt 0+t+1 PSmt+2 Sm1t+2 Sp1t+3

Parr1t+2 Sm2t+3 Sp2t+4

FIGURE 1.7: Step-by-step construction of a complex model by articu-

lating blocks to describe the full A. salmon life cycle model. Upper panel:
(a) Ricker stock-recruitment relationship with environmental noise; (b)
Survival from 0+ juveniles to pre-smolts; (c,d) Smoltification transitions
with demographic stochasticity; (e) Post-smolts marine survival rates.
Bottom panel: DAG for the entire life cycle describing the dynamic of
one cohort originated as eggs spawned at year t.
 Bayesian hierarchical models in statistical ecology 23

the sub-modules that mimic the density-dependent eggs → 0+ juveniles

survival (Ricker relationship), the 0+ juveniles → pre-Smolts survival,
the smoltification phase and the marine survival of post-smolts. This
illustrates how models can be built step-by-step, by first building some
bricks independently that will be connected together in a second step.
The basic properties of DAGs simply derive from standard proba-
bility theory. Consider for instance the set of all variables in Fig. 1.7.
(α, β, σ, Wt , γ0+ , θSm1 ) have no parents; the parents of 0+t+1 are Wt
and (α, β, σ); the parents of Sm1t+2 are P Smt+2 and θSm1 and so on
for the ongoing of each cohort. The full joint distribution of all unknown
variables for one cohort originated from eggs spawned at year t, denoted
Jt , can be factorized relying on probability theory. Following the local
dependencies in the DAG (Fig. 1.7), and assuming that all parameters
are independent, this factorization of the joint probability distribution
reads:

Jt =[α] × [β] × [σ] × [Wt ] × [γ0+ ] × [θSm1 ] × [γP arr1 ] × [γSm ]

× [0 +t+1 |Wt , α, β, σ] × [P Smt+2 |γ0+ , 0+t+1 ]
× [Sm1t+2 , P arr1t+2 |P Smt+2 , θSm1 ]
× [Sm2t+3 |P arr1t+2 , γP arr1 ]
× [Sp1t+3 |Sm1t+2 , γSm ] × [Sp2t+4 |Sm2t+3 , γSm ] (1.16)

1.2.4 Embedding observation processes into a proba-

bilistic conditional model
Following the above section, the DAG is a graphical representation
of probabilistic cause-to-effect relationships. Looking at the graph by
following the direction of the arrows emphasizes the child-parent depen-
dence (and eventually conditional independence) of the variables. This
corresponds to the modeling way of thinking. The DAG depicts the ten-
tative structure that has been proposed for the model. The mathematical
(probabilistic) result is the factorization of the full joint distribution as
in Eq. (1.16).
Such a probabilistic conditional model can be further developed
by including probabilistic observation process. Indeed, observation pro-
cesses are rarely perfect, especially in ecological sciences. Sampling and
measurement errors may occur in many experimental protocols, and are
also modeled via conditional probability distributions. For instance, a
variable Vi in a model can be an observable, and the conditional dis-
tribution [Vi |pa(Vi )] stands for the observation process or a sampling
process, that is the stochastic mechanism that leads to an observation.
The DAG depicts the structure that is proposed for both the process and
24 Introduction to Hierarchical Bayesian Modeling for Ecological Data

the observation model. As far as probability is concerned, the observa-

tion process is simply additional (observed) variables fully integrated
into the conditional probability model.

1.2.5 Hierarchical modeling: Separating out process

stochasticity from observation errors
Variables in a complex graphical model that include both a process
for state variables and an observation process, can advantageously be
classified into three categories to emphasize the so-called Hierarchical
Modeling structure.

a) Simple model b) Hierarchical Model

θ θ Parameters

Latent
Z variables

Y Y Observables

FIGURE 1.8: Directed acyclic graph (DAG) of a prototype Hierarchi-

cal Model.

Nodes without any parent, nonobservable nodes with parents, and

observed nodes are denoted θ, Z, and Y, respectively. The nonobservable
nodes with parents, denoted Z, are also called latent variables.
The conventional notation (θ, Z, Y ) splits the DAG variables into
three layers as depicted in Fig. 1.8 where, conversely to DAG, we con-
sider each layer as a block. Using the generic set of variables from the
previous section V1 , ..., Vn , let us for convenience sake suppose that the
p first nodes V1 , ..., Vp are not observables, and that the last n − p nodes
Vp+1 , ..., Vn are observables. We hence have distinguished Y , a block
notation for all the observables in the graph (those previously termed
 Bayesian hierarchical models in statistical ecology 25

Vp+1 , ..., Vn ) and the other nonobserved quantities V1 , ..., Vp =(θ, Z) that
are unknown. Among these unknown quantities, we further make the dis-
tinction between the parameters (suppose for convenient notation that
the k first variables V1 , ..., Vk are nodes without parent) that are entry
points to the graph and are denoted θ and the p − k inner ones (the re-
maining Vk+1 , ..., Vp ) that are denoted Z. Parameters θ are often called
state of nature while latent variables Z are commonly named state of
the system. Such a three-layers structure as depicted by Fig. 1.8 is the
core of Bayesian Hierarchical Modeling (HBM).
Keeping in mind the groupings θ = (V1 , ..., Vk ), Z = (Vk+1 , ..., Vp ),
Y = (Vp+1 , ..., Vn ), and assuming that parameters have an unconditional
prior probability distribution [θ], the full joint distribution of the HBM
may be factorized as in Eq. (1.17):

[θ, Z, Y ] = [θ] × [Z|θ] times[Y |θ, Z] (1.17)

Equation 1.18 offers a more explicit interpretation of the factorization

of a hierarchical model, which is illustrated in Figs. 1.8 and 1.12:

[P arameters, P rocess, Observables] =

[P arameters]
× [P rocess|P arameters]
× [Observables|P rocess, P arameters] (1.18)

It is worth noting that [Z|θ] and [Y |Z, θ] are compact representations of

eventually highly complex probabilistic state and observation processes.
These process and observation equations might be constructed from the
combination of many more simple local interactions, which are them-
selves reorganized through parent-child conditional dependencies using
Eq. (1.14).

Example (continued): A. salmon population dynamic model

In the A. Salmon stage-structured population dynamic model, the
number of fish in each of the development stage are unknown but are
observed indirectly through capture experiments during migration in
downstream (for smolts) and upstream (for adults) traps and through
electrofishing experiments for freshwater resident 0+ juveniles. As de-
scribed previously, those capture experiments can be modeled by Bi-
nomial sampling probability distributions, with probabilities of capture
26 Introduction to Hierarchical Bayesian Modeling for Ecological Data

denoted πSm , πSp , and π0+ , and considered constant for all years t:


 C0+,t+1 ∼ Binomial(0+t+1 , π0+ )

 CSm1,t+2 ∼ Binomial(Sm1t+2 , πSm )



CSm2,t+3 ∼ Binomial(Sm1t+3 , πSm ) (1.19)

CSp1,t+3 ∼ Binomial(Sp1t+3 , πSp )





CSp2,t+4 ∼ Binomial(Sp2t+4 , πSp )


Equation (1.19) brings an additional block to be linked to the pre-

vious DAG. Figure 1.9 extends Fig. 1.7 where Binomial counting pro-
cesses have been connected at stages 0+ juveniles, smolts and spawners.
Following our graphical conventions, the corresponding ellipses for the
Binomial catches in the graph are shaded if the catches are observed,
and the catchabilities, considered known, appear in shaded boxes.

FIGURE 1.9: DAG describing the A. salmon life cycle model with
Binomial observation process. Parameters (θ following our notations)
appear in the upper layer. The middle layer describes the dynamics of
hidden states ([Z|θ]). The observation model ([Y |Z, θ]) is presented in
the bottom layer.

Keeping in mind our block notation (θ, Z, Y ) for a generic hierarchical

model, the vector of all the parameters θ would correspond to the vector
of transition probabilities between each stage (survival and mortality
 Bayesian hierarchical models in statistical ecology 27

ratios, Ricker parameters, etc.). The latent states Z (the other nodes in
between parameters and observables) would be mainly the number of in-
dividuals in the possible stages for all time indices, and the observations
y will be the catches:

• Nodes without parents θ = (α, β, σ, Wt , γO+ , γP arr1 , θSm1 , γSm )

• Nonshaded nodes with parents Z = (O+t+1 , P Smt+2 , Sm1t+2 ,

P arr1t+2 , Sm2t+3 , Sp1t+3 , Sp2t+4 )
• Shaded observed nodes Y = (cO+t+1 , cSm1t+2 , cSm2t+3 , cSp1t+3 ,
cSp2t+4 ).

Note: In this example, all the catchabilities π are considered known.

But the π’s could also be estimated. This is easily performed by ex-
panding the observation process to include capture-mark-recapture ex-
periments designed to provide information on both capture efficiencies
π and total numbers (for details, see [255], [257] and [259]).

1.3 Bayesian inferences on hierarchical models

1.3.1 The intuition: What if a node is observed?
The main intuition for Bayesian statistical learning (in other words
Bayesian inference) can be advantageously presented via fluxes of in-
formation on a DAG. What happens when an observation becomes ac-
tually available? Let us, for convenience, suppose that the observable
nodes Y (e.g., the n − p last nodes in the graph Vp+1 , ..., Vn , have been
observed and take the value y. Following our graphical conventions, the
corresponding ellipses in the graph then become shaded. Once the phe-
nomenon Y has been observed (note we are using latin capital letters
for the random phenomenon and lowercase ones for data), this certainly
impacts the probability distributions of all the other nodes in the graph.
When looking at the flow of information in the DAG, the information
contained in Y = y will propagate in the reverse direction of the ar-
rows. An effect was observed: How will the distribution of the differ-
ent parents of y be modified? This is the inferential way of thinking.
In other words, and following our block notations (θ, Z, Y ), observable
nodes in the block Y are no longer random but they have been set to
fixed observed values Y = y, and the range of possible values of the other
nodes (θ, Z) = (V1 , ..., Vp ) in the graph will subsequently be restrained
 Another Random Document on
Scribd Without Any Related Topics
semble donc être après la première dynastie des demi-dieux. On a
reconnu dans ces Nekyes ou mânes les Khouou des textes religieux
égyptiens, divinités secondaires qui constituent la troisième ennéade
héliopolitaine, d’abord les quatre génies funéraires, les Enfants
d’Horus, Amset, Hapi, Douamoutef et Kebhsenouf, puis un autre
Horus, Khent-Khiti, et ses quatre fils.
Après les dynasties divines et semi-divines, calquées sur le modèle
des trois cycles de dieux héliopolitains, et qui servent en quelque
sorte de cadre aux souvenirs relatifs à ces époques très anciennes,
Manéthon en énumère trois autres qui sont composées de rois d’une
essence plus rapprochée de la nôtre, et considérés sans doute
comme de simples hommes: d’abord ce sont des rois dont il
n’indique ni l’origine ni le nombre et qui régnèrent en tout 1.817
ans, puis trente rois memphites, pendant 1.790 ans et enfin dix rois
thinites, dont les règnes successifs durèrent 350 ans. Au papyrus de
Turin, la division de cette période était un peu différente, et dans le
fragment qui s’y rapporte, on peut reconnaître qu’il avait parlé de six
dynasties au moins; les noms des rois n’étaient pas donnés, mais
seulement la mention qu’ils s’étaient succédé de père en fils et que
parmi eux se trouvaient sept femmes ayant régné; les chiffres,
donnant la somme des années de chaque dynastie, sont trop mutilés
pour que nous puissions en tenir compte.

C. LA CHRONIQUE LÉGENDAIRE

En résumé, toute cette période fabuleuse se divisait en plusieurs

époques, celle des dieux cosmogoniques et organisateurs de
l’humanité, celle des demi-dieux dont le rôle très effacé a plutôt un
caractère transitoire, et enfin celle des hommes-rois; pour les
Egyptiens eux-mêmes, les souverains à partir de la IIme dynastie,
donc les demi-dieux, les mânes et les hommes formaient un seul
grand groupe, celui des Shesou-Hor, ou suivants d’Horus, auxquels
Manéthon attribue une durée totale de règne de 11.000 ans, tandis
que les dieux eux-mêmes auraient occupé le trône pendant 13.900
ans. Cela donnerait pour tous les rois antérieurs à Ménès une
somme de 24.900 ans, chiffre qui paraissait très exagéré à Eusèbe,
aussi préférait-il adopter l’explication de Panodore, que ces années
n’étaient autres que des années lunaires de 30 jours, des mois, ce
qui réduisait donc la durée des rois mythiques à 2.206 ans. Cette
interprétation fantaisiste est du reste dénuée de tout fondement, et
l’on voit qu’au papyrus de Turin il s’agit bien d’années ordinaires,
d’années solaires; si les chiffres ne sont pas ici exactement les
mêmes que ceux de Manéthon, ils leur correspondent dans les
grandes lignes. La somme totale des règnes est en effet ici de
23.200 ans au lieu de 24.900, et sur des chiffres pareils l’écart n’est
pas très considérable; pour la période des Shesou-Hor, le papyrus
compte 13.420 ans, chiffre équivalant à peu près à celui que donne
Manéthon pour les dieux, et il est possible qu’il y ait eu une
interversion dans un des documents qu’il avait entre les mains. La
question a du reste peu d’importance pour nous, puisqu’il s’agit de
chiffres absolument fantaisistes.
Les Egyptiens avaient donc au sujet de leurs origines une tradition
qui nous paraît simple et pleine de renseignements précis, si nous la
comparons à celles des autres peuples, souvent remplie de détails
charmants et inutiles, de digressions qui nuisent à la clarté de
l’ensemble, et font perdre facilement le fil conducteur. Ici c’est une
légende pour ainsi dire quintessenciée, prenant le monde à ses
débuts, l’humanité à sa création même, la suivant à travers les
grandes commotions géologiques qui bouleversèrent la vallée du Nil
avant le début de l’histoire. Nous pouvons, en coordonnant ces
traditions, suivre les progrès, le travail lent, mais sûr, de la
civilisation que les réactions brutales ne peuvent anéantir. Au
commencement, ce sont les dieux qui dirigent le mouvement
progressif de l’humanité qu’ils ont eux-mêmes mis en branle, puis
peu à peu ils s’effacent, passant la main à des êtres moins sublimes,
moins éloignés par leur nature même de la race qu’ils ont à
gouverner, et enfin à de vrais hommes, arrachés définitivement à la
sauvagerie primitive et capables en une certaine mesure, après des
milliers d’années d’efforts, de s’affranchir de la tutelle directe des
dieux. Ces débuts des hommes furent obscurs et sans doute
difficiles, et il fallut encore de longs siècles avant que l’un d’entre
eux pût saisir d’une main ferme les rênes du pouvoir et donner à
l’Egypte cette puissante organisation qui devait durer plus longtemps
que celle d’aucun autre pays. Les rois locaux antérieurs à Ménès
n’ont pas laissé de traces dans l’histoire, mais il est possible qu’un
certain nombre de leurs noms aient été conservés: en effet, au
premier registre de la pierre de Palerme, on voit représentés toute
une série de personnages portant la couronne rouge, l’insigne des
rois de la Basse Egypte, au-dessus desquels sont gravés quelques
signes qui peuvent fort bien être des noms, mais des noms bizarres
qui ne ressemblent guère aux noms égyptiens ordinaires. Seka,
Khaaou, Taou, Tesh, Neheb, Ouazand, Mekha. Ce serait le seul
document précis relatif à la fin de la période légendaire, à ces rois
memphites dont parle Manéthon. Quant aux rois de la Haute Egypte,
leurs compétiteurs, peut-être devons-nous en reconnaître quelques-
uns parmi les monuments d’Abydos qu’on attribue généralement à la
Ire dynastie: il s’y trouve en effet quelques noms de rois difficiles à
lire et à identifier et qui peuvent appartenir à certains des
prédécesseurs immédiats de Ménès.

Fig. 13. Les enfants d’Horus

(d’après Budge. Pap. of Ani, pl. VIII).
 Fig. 14. Poignard en silex
(d’après J. de Morgan. Recherches sur les origines de l’Egypte, I, fig.
174).
 CHAPITRE III

L’ÉGYPTE ARCHAÏQUE

Les grands travaux exécutés dans la vallée du Nil au cours du

siècle dernier avaient amené la découverte d’un tel nombre de
monuments datant des époques historiques, édifices, sculptures,
peintures, objets d’art, inscriptions, instruments de toute sorte, que
l’attention des égyptologues devait nécessairement se concentrer sur
ces restes pharaoniques et ne pas aller chercher plus loin des
documents dont, malgré leur abondance considérable, on
connaissait à peine l’existence et dont surtout on ne pouvait encore
soupçonner la valeur. On se contentait de relever les grands
monuments apparents, temples ou tombeaux, de fouiller des
nécropoles riches et le plus souvent bien connues, on ne se livrait
pas encore à une exploration méthodique du pays et l’on n’accordait
aucune attention à des objets sans grande apparence, les silex
taillés, que dans d’autres contrées on recueille avec tant de soin et
qu’ici on ne se donnait même pas la peine de ramasser. Il est vrai
cependant que des archéologues, comme Arcelin et le Dr Hamy, au
cours d’un voyage dans la vallée du Nil, en avaient réuni un certain
nombre et avaient cru pouvoir parler du préhistorique égyptien et
d’un âge de la pierre, d’après ces documents qui étaient du reste
trop insuffisants pour qu’on pût en tirer des conclusions sérieuses;
les égyptologues n’eurent donc pas de peine à leur prouver de la
façon la plus péremptoire que ces instruments n’avaient rien de
préhistorique: n’avait-on pas, en effet, trouvé des silex taillés dans
des tombes de la XIIme dynastie?
 La question semblait donc jugée et, si invraisemblable que cela
paraisse maintenant, on croyait qu’il n’existait en Egypte aucun
monument, aucun objet datant d’une époque antérieure à celle du
fabuleux Ménès: les deux premières dynasties humaines n’ayant
laissé aucune trace autrement que dans la tradition, à plus forte
raison la période qui les précédait devait-elle rester à jamais
inconnue. On devait cependant admettre que dans un pays où tout
se conserve, comme l’Egypte, il eût été naturel qu’on retrouvât
quelque chose au moins des débuts d’une civilisation aussi originale,
et on en était venu, pour expliquer en une certaine mesure cette
lacune apparente, à émettre l’hypothèse que les ancêtres directs des
Egyptiens avaient pu se développer ailleurs, dans le Bahr-bela-mà,
par exemple, le fleuve sans eau, une vallée du désert libyque, ou
bien dans le pays des Somâlis ou plus loin encore. Par conséquent,
et malgré les affirmations catégoriques des Egyptiens d’époque
historique, la civilisation égyptienne ne pouvait être autochtone: une
lacune insondable devait précéder l’histoire, il ne pouvait être
question de paléolithique ni de néolithique, l’Egypte n’avait jamais
connu l’âge de la pierre, et tout au plus pouvait-on considérer les
premières dynasties comme appartenant à la période du bronze.
On en était là quand, vers 1896, cette théorie simpliste reçut de
plusieurs côtés à la fois un choc qui devait non seulement l’ébranler,
mais l’enterrer à tout jamais. A ce moment, des fouilles entreprises
dans des endroits encore inexplorés vinrent révéler à MM. Petrie et
Amélineau l’existence de civilisations très différentes de celles qu’on
connaissait, tandis que les recherches plus méthodiques de M. de
Morgan l’amenaient à la certitude qu’il s’agissait là d’une révélation
inattendue, celle du préhistorique égyptien auquel personne ne
voulait croire. Du même coup l’on voyait réapparaître les premiers
habitants du pays avec leurs armes de silex, leur céramique très
particulière, leurs tombeaux et même leurs villages, et les rois des
deux dynasties encore inconnues, avec le métal et les premiers
monuments de l’écriture hiéroglyphique. Les preuves étaient si
évidentes qu’en peu de temps tous les égyptologues se rallièrent aux
nouvelles théories établies par M. de Morgan, les confirmèrent et les
complétèrent par d’autres recherches, si bien que maintenant on
peut se rendre compte de façon à peu près certaine de ce qu’étaient
les plus anciens occupants de la vallée du Nil.

L’époque préhistorique ne se présente pas en Egypte, comme

dans nos pays européens, avec des divisions nettement marquées
qui sont caractérisées par les procédés employés dans la fabrication
des armes et des outils et par la forme même de ces derniers. A
peine peut-on faire un groupe distinct pour les instruments les plus
anciens et les plus rudimentaires, qui correspondent à peu près
comme type et comme taille à notre Chelléen, mais à partir de cette
époque très reculée, tous les silex présentent à peu de chose près le
même caractère: si nous les comparons aux silex européens, ils
pourraient se ranger aussi bien dans les séries paléolithiques que
dans le néolithique. Les noms de Moustérien, Solutréen,
Magdalénien, qui s’appliquent chez nous à des périodes bien
définies, très différentes les unes des autres, ne correspondent à
rien en Egypte, et leur emploi n’aurait aucune raison d’être pour tout
ce qui concerne les origines de ce pays.
Si donc nous mettons à part une première période, celle du
paléolithique proprement dit, une civilisation qui a dû être
interrompue brusquement par un cataclysme quelconque, nous
trouvons ensuite des séries de monuments préhistoriques qui,
malgré leur grande variété, présentent une parfaite homogénéité.
Les seules différences que nous pouvons remarquer dans la
fabrication des outils de pierre sont de nature purement locale, ainsi
les silex du Fayoum ne sont pas les mêmes que ceux de Negadah,
pas plus que ceux d’Hélouan ne ressemblent à ceux d’Abydos ou
d’autres endroits, mais il n’y a pas lieu de tirer de ce fait des
conclusions au point de vue chronologique, car rien ne peut faire
croire que les uns soient antérieurs aux autres. Les ateliers
employaient des procédés légèrement différents, et surtout des
modèles qui variaient d’un endroit à l’autre; les uns, dans les lieux
où les habitants se livraient principalement à la chasse ou à la
pêche, faisaient surtout des armes, couteaux, pointes de lances, de
javelots ou de flèches, tandis que les autres, dans les centres
agricoles, fabriquaient plutôt des outils, mais ces différences sont de
nature géographique et non historique, et on ne peut en tenir
compte pour scinder la période quaternaire en un plus ou moins
grand nombre d’époques distinctes.
L’évolution de la céramique, chez les peuples primitifs, suit
toujours une marche parallèle à celle des instruments de pierre, et
l’on peut, par ce moyen, contrôler les conclusions fournies au point
de vue historique par l’étude de la forme et des procédés de
fabrication des silex. Il en est de même en Egypte, c’est-à-dire que
dans le domaine de la céramique archaïque, on remarque bien un
développement, un progrès, mais cette transformation est lente,
graduelle, sans secousses. Les anciens modèles cèdent la place à de
nouveaux, mais pas de façon brusque; ils coexistent pendant
longtemps et se retrouvent les uns à côté des autres dans les
mêmes tombes. On peut arriver à constater que tel type est plus
ancien que tel autre, on ne peut dire qu’il caractérise une époque ou
une phase de la civilisation préhistorique. La céramique égyptienne
est du reste tout à fait spéciale et très différente de toutes celles
qu’on rencontre en Europe aux époques primitives, aussi n’y
retrouve-t-on aucun des caractères spécifiques qui permettent aux
préhistoriens de classer ces dernières: les potiers égyptiens avaient
poussé cet art à un haut degré de perfection dès les plus anciens
temps, et nous leur devons des séries très variées, tant au point de
vue de la technique que de la forme et de la décoration.
La céramique, qui est un des éléments les plus importants pour la
classification des restes préhistoriques, ne donne donc lieu ici à
aucun rapprochement, et nous devons nous en tenir aux données
que nous fournissent les armes et les outils de pierre; or nous avons
vu que tous ces objets sont en pierre taillée et qu’ils se rattachent,
pour les formes comme pour les procédés de taille à nos instruments
paléolithiques et néolithiques en silex, tout spécialement aux types
du Solutréen et du Moustérien. Ce qui caractérise chez nous la
période néolithique, l’âge de la pierre polie, manque absolument en
Egypte: on a récolté dans ce pays, pendant ces dernières années,
des centaines de mille et peut-être des millions de silex, et dans
cette masse énorme on aurait peine à trouver cent haches polies, ou
autres outils pouvant rentrer dans la même catégorie. Nous ne
constatons cependant aucune solution de continuité entre la période
dite préhistorique et celle des débuts de l’histoire, aussi pouvons-
nous dire avec certitude que non seulement il n’y a pas de divisions
spéciales à établir dans l’époque paléolithique, mais qu’il n’y a même
pas lieu de distinguer celle-ci de l’âge néolithique. Si donc nous
devions conserver ces deux noms qui ont une certaine valeur
pratique pour la classification, il faudrait leur donner, pour tout ce
qui concerne l’Egypte, un sens un peu différent de celui qu’ils ont
pour l’Europe, réserver le mot paléolithique aux objets les plus
anciens, à ceux qui pour la forme et la facture se rapprochent du
chelléen, et ranger tout le reste dans l’âge néolithique ou même
plutôt énéolithique qui précède immédiatement l’âge historique.
Dans nos pays septentrionaux, où le développement des peuples
suivit une marche toute différente, on range encore dans le
préhistorique la période des métaux et l’on fait succéder l’âge du
cuivre, l’âge du bronze, puis l’âge du fer, à celui de la pierre. Ici il n’y
a aucune distinction semblable à établir puisque les dynasties
thinites suivent immédiatement l’âge de la pierre, sans aucune
transition apparente: les Egyptiens prédynastiques sont déjà en
possession des métaux, ou tout au moins du cuivre qu’ils emploient
presque sans alliage et qu’ils arrivent peu à peu à travailler avec la
plus grande habileté, en même temps qu’ils poussent l’industrie du
silex à un degré de perfection qui ne fut atteint en aucun endroit du
monde. C’est donc au cours de l’époque précédant immédiatement
l’histoire que les Egyptiens apprirent à connaître le cuivre, dont
l’usage ne remplaça que très lentement celui de la pierre taillée;
c’est aussi tout à fait graduellement que les métallurgistes arrivèrent
à doser les alliages grâce auxquels ils devaient obtenir le bronze,
très supérieur au cuivre pur. Quant au fer, nous n’avons aucun
document qui nous permette de fixer l’époque à laquelle il fut
introduit dans la vallée du Nil. Il n’y a donc en Egypte ni âge du
cuivre, ni âge du bronze, ni âge du fer, à proprement parler: la
première de ces trois divisions se confond avec la période
prédynastique, et les deux autres, qui ne sont pas nettement
caractérisées, appartiennent à l’époque historique.
Ménès, le fondateur de la monarchie pharaonique, symbolise pour
nous le début d’une civilisation nouvelle, l’organisation définitive du
pays, et les premiers documents écrits qui paraissent à ce moment-
là, montrent bien qu’une ère nouvelle commence. La transformation
ne s’opéra cependant pas d’une façon subite dans tous les
domaines, elle se fit graduellement, lentement, comme dans les
périodes précédentes, car l’Egypte a toujours été et sera sans doute
toujours le pays le moins révolutionnaire qu’il y ait au monde. Dans
la vie civile surtout, que nous connaissons fort bien, puisque une
grande quantité d’objets de toute sorte nous sont parvenus, le
progrès est presque insensible, la céramique est à peu près la même
qu’auparavant, à peine un peu détrônée par l’usage toujours plus
répandu des vases de pierre, et l’on devait continuer pendant de
longs siècles encore à fabriquer des armes et des outils en silex, bien
qu’on connût déjà fort bien les instruments de métal, dont la
supériorité était évidente. Enfin, si les rois et les grands personnages
commencent à se faire construire des tombeaux monumentaux et
adoptent des coutumes funéraires plus compliquées, les populations
rurales continuent à creuser à la limite des sables du désert de
petites fosses pour leurs morts, qu’ils ensevelissent accroupis et
couchés sur le côté, ou démembrés complètement, avec le même
mobilier funéraire que par le passé.
J’ai employé jusqu’ici, pour désigner les âges primitifs de l’Egypte,
le mot de préhistorique, mais, en ce qui concerne ce pays, ce mot a
une signification trop précise et indique une scission trop nette avec
le temps où commence l’histoire proprement dite; or, comme nous
l’avons vu, cette scission n’existe pas en Egypte. Le terme d’âge de
pierre ne convient pas non plus, puisque l’emploi des instruments de
silex est encore constant sous les premières dynasties et se perpétue
jusqu’au Moyen Empire. J’adopterai donc dorénavant un terme plus
élastique et dont le sens est néanmoins très clair, celui de période
archaïque, qu’on emploie maintenant de préférence, et je diviserai
cette période en deux groupes comprenant, l’un, les âges les plus
anciens, l’éolithique et le paléolithique, l’autre, l’époque beaucoup
plus connue, précédant immédiatement les dynasties, et qu’on peut
appeler prédynastique.

I. PALÉOLITHIQUE

Les vestiges des tout premiers habitants de l’Egypte sont rares et

incertains. La tendance actuelle est de rechercher partout la trace de
l’homme tertiaire; à défaut de preuves absolument convaincantes de
son existence, comme le serait la découverte d’un squelette dans
une couche géologique appartenant à cette période, on voudrait
retrouver des indices de son activité sur la terre, aussi a-t-on créé la
classe des éolithes, les instruments de l’homme antérieur à l’âge
paléolithique. Ces éolithes sont de simples galets de silex ou des
éclats accidentels sur lesquels on remarque ou croit remarquer des
traces d’usage, et qui auraient été les premiers instruments de
l’homme alors qu’il ne savait pas encore tailler la pierre et devait se
contenter des éclats naturels, plus ou moins appropriés à ses
besoins, qu’il trouvait sur le sol. Ce n’est pas ici le lieu de discuter
cette théorie toute générale, qui est encore très sujette à
controverse; nous nous bornerons à constater qu’elle a aussi été
appliquée à l’Egypte et qu’on a recueilli dans ce pays un certain
nombre d’échantillons de ces éolithes qui ont évidemment pu être
employés par des hommes encore à l’état de sauvagerie, comme
marteaux, grattoirs ou couteaux, bien que rien ne le prouve de façon
absolue.
 Fig. 15–18. Instruments paléolithiques
(d’après J. de Morgan. Recherches sur les origines de l’Egypte, I, fig.
17, 24, 20, 31).

Les silex taillés du type chelléen se retrouvent non seulement en

Europe, mais un peu partout, en Palestine, aux Indes, chez les
Touaregs; on en rencontre aussi en Egypte, sinon en grande
abondance, du moins assez fréquemment. L’objet le plus
caractéristique de cette époque est, ici comme dans les autres
gisements, le coup-de-poing, un grand galet de silex amygdaloïde,
sur lequel on a enlevé par percussion de gros éclats, de manière
qu’une des extrémités forme une pointe plus ou moins prononcée,
tandis que l’autre reste arrondie et épaisse, et sert de poignée. A
côté de cet instrument qui en même temps est une arme
dangereuse, on trouve encore des outils plus petits, ayant pu servir
de hachettes ou de racloirs; et surtout des pointes ou poinçons,
parfois très aigus, du même travail un peu rudimentaire, sans
retouches fines.
Ces silex se trouvent soit à la surface du sol, sur les plateaux
couronnant les premiers contreforts du désert et au sommet des
petits monticules qui sont situés un peu au-dessous, soit dans les
alluvions entraînées par les pluies jusque dans la vallée, très
rarement dans la zone sablonneuse qui sépare les terres cultivables
de la montagne. On en a découvert depuis les environs de la 1re
cataracte jusque près du Caire, ainsi que sur les routes qui
conduisent à travers le désert vers les oasis, et enfin, ce qui est plus
important au point de vue de la date, dans les alluvions très
anciennes, contemporaines du commencement de l’époque
quaternaire, qui est en effet le moment où l’on place l’âge chelléen.
D’après la position où ont été trouvés ces silex, on pourrait conclure
que les Egyptiens primitifs habitaient de préférence, non pas dans la
vallée même, mais sur les monticules avoisinants et sur la crête des
montagnes peu élevées qui bordent le désert. Nulle part on ne voit
de traces d’habitations construites; ils devaient donc vivre soit en
plein air, soit sous de légers abris en branchages. C’est sur ces
plateaux, où les indigènes trouvaient en abondance les rognons de
silex qui servaient à la fabrication de leurs outils, qu’ils établissaient
leurs ateliers de taille: ainsi le plateau qui sépare la Vallée des Rois
du cirque de Deir-el-Bahari, en face de Louxor, où l’on trouve encore
en quantité des éclats n’ayant sans doute jamais servi et qui doivent
être considérés comme des déchets de fabrication. La réalité est
sans doute un peu différente, et si nous ne sommes pas mieux
renseignés sur cette population primitive, sur son habitat et ses
coutumes funéraires, c’est pour la raison qu’elle est antérieure à un
de ces bouleversements géologiques qui dévastèrent et
dépeuplèrent une partie du monde et qui sont restés célèbres dans
la tradition sous le nom de Déluge. L’Egypte en particulier fut
atteinte, la vallée fut entièrement submergée pendant une période
dont nous ne pouvons évaluer la durée et toute trace d’occupation
humaine fut effacée; les hauts plateaux stériles et le désert
émergeaient encore, mais nous ne savons si quelques restes de la
population purent s’y maintenir pour former le noyau de la race
égyptienne prédynastique, ou si celle-ci vint d’ailleurs quand la
région redevint habitable.

II. PRÉDYNASTIQUE
 A. MONUMENTS

Autant cette première période est encore obscure, autant les

documents abondent pour celle qui la suit, et qui, précédant
immédiatement l’époque historique, est souvent désignée par le nom
de prédynastique. Ces documents peuvent se classer en trois
catégories, dont les données combinées nous fournissent des
renseignements d’ensemble et même de détail sur l’état de la vallée
du Nil avant les Pharaons. Ce sont d’abord les objets épars à la
surface du sol, les silex, puis les vestiges des établissements
humains, monticules de débris où l’on reconnaît la trace des villages
primitifs, et enfin les tombeaux qui nous donnent, en plus des
renseignements anthropologiques, des lots très considérables de
céramique, l’élément le plus important pour la classification
générale. Nous prendrons l’un après l’autre chacun de ces points
avant d’aborder l’ethnographie proprement dite, l’étude de la race
prédynastique et de sa civilisation.

Fig. 19–21. Haches et herminettes en silex

(d’après J. de Morgan. Recherches sur les origines de l’Egypte, I, fig.
91, 60, 73).

Silex
 Les couches sédimentaires qui bordent la vallée du Nil sont
extrêmement riches en rognons de silex, qui atteignent parfois de
très grandes dimensions; sur les plateaux, le sol est couvert de
galets de silex, d’agate et de cornaline. Naturellement la qualité de
la pierre varie suivant les endroits, mais partout elle se prête à la
taille et les premiers habitants du pays avaient sous la main, d’un
bout à l’autre du pays, la matière première de laquelle ils pouvaient
tirer leurs armes et leurs outils. C’est vers le nord de l’Egypte, au
Fayoum en particulier, que le silex est le moins abondant, mais les
cailloux du diluvium peuvent le remplacer, et les indigènes en ont
tiré un très bon parti.

Fig. 22–25. Couteaux et grattoirs en silex

(d’après J. de Morgan. Recherches sur les origines de l’Egypte, I, fig.
106, 123, 98, 153).

Quelle que soit la matière employée, qu’il s’agisse du beau silex

blond translucide d’Abydos, du silex brun de Louxor ou du grossier
galet du Fayoum, le procédé de taille est toujours le même, et ne
diffère pas de celui qui a été en usage dans le monde entier. Le
nucleus, ou noyau préparé pour l’enlèvement des éclats, s’obtenait
d’une façon très simple: on brisait une partie d’un rognon de silex ou
d’un galet, de manière à déterminer une surface unie servant de
plan de frappe, puis on enlevait des éclats normalement à cette
surface, en se servant d’un percuteur, boule de pierre dure employée
comme marteau; les premiers éclats, portant une partie de la
gangue, étaient mis au rebut, et les suivants employés pour divers
usages selon leur forme et leur dimension; ceux qui étaient longs et
minces devenaient des couteaux, ceux qui étaient épais et larges,
des haches ou des herminettes, les petits donnaient des ciseaux, des
poinçons, des pointes de flèches; tous devaient subir de longues et
soigneuses retouches. On travaillait ces éclats soit par percussion,
soit par pression le long des arêtes au moyen d’un autre silex, et les
Egyptiens étaient arrivés très loin dans cet art et modelaient pour
ainsi dire leurs silex au moyen de ces petites retouches, de manière
à leur donner exactement la forme voulue. A côté de ces
instruments, certains éclats, très minces et naturellement
tranchants, pouvaient être utilisés, presque sans retouches, comme
outils, grattoirs ou couteaux.

Fig. 26–29. Pointes de flèches en silex

(d’ap. J. de Morgan. Recherches sur les origines de l’Egypte, I, fig.
176, 190, 181, 185).

On trouve de tout cela dans les gisements de silex égyptiens, sur

la bande sablonneuse qui s’étend d’un bout à l’autre du pays, entre
les terres arrosées et cultivées et les premiers contreforts de la
montagne: d’abord les percuteurs, boules qui ont en général la
grosseur d’une pomme et qui portent des traces très évidentes
d’usage, puis les nuclei à tous les états, depuis celui qui a été mis au
rebut après qu’on en eut détaché quelques éclats seulement, jusqu’à
celui qui, complètement épuisé, n’est plus qu’un petit noyau conique
à facettes; ensuite les éclats eux-mêmes, les uns, informes ou mal
venus, rejetés comme inutilisables, les autres, très tranchants et
sans retouches ou retravaillés seulement à une extrémité; enfin les
outils brisés au cours de la fabrication par suite d’un accident, et
ceux qui portent la trace d’un long emploi ou qui, très usés, ont été
retaillés pour pouvoir être employés de nouveau.
Chaque localité, chaque gisement a pour ainsi dire son propre
type, ou ses types de silex taillés, et l’on ne peut en tirer des
conclusions au point de vue de la classification chronologique; il est
possible, probable même, que dans beaucoup de ces endroits, la
fabrication se soit continuée sans grande modification, pendant des
siècles ou des milliers d’années, comprenant non seulement toute la
période archaïque, mais empiétant aussi sur les époques historiques.
Nous aurons l’occasion de revenir plus loin sur les différents modèles
d’outils et d’armes, sur leurs formes et leur emploi.

Villages

Dans les mêmes régions, en bordure de la vallée, à la lisière du

désert, on remarque en certains endroits de légères surélévations
qui se distinguent à peine du sable environnant par une teinte un
peu plus foncée. Quelques coups de pioche suffisent pour constater
qu’il y a là quelque chose de tout à fait analogue à ce que dans nos
stations préhistoriques européennes, celles du Danemark en
particulier, on appelle des Kjoekkenmoeddings, ou «débris de
cuisine»; ce sont en effet des vestiges d’établissements humains,
datant d’une époque où les populations étaient déjà plus ou moins
sédentaires, mais où elles ne savaient pas encore construire de
vraies maisons: ces restes sont beaucoup trop importants pour être
ceux de simples campements provisoires et passagers, et
contiennent des quantités de détritus qui ont dû mettre fort
longtemps à s’amonceler. D’un autre côté on ne rencontre pas dans
ces monticules de décombres la moindre trace de mur, ni en pierre,
ni en briques crues, ni même en terre pilée: les constructions
devaient donc être très légères, en bois ou même en branchages, de
simples huttes du modèle le plus primitif, suffisantes du reste dans
un climat aussi chaud.
Ces amas de détritus ne renferment guère d’objets en bon état, à
part quelques outils de silex, mais ils nous livrent des
renseignements très importants sur la vie même de ces peuplades
de l’Egypte prédynastique; os d’animaux d’après lesquels on peut,
en partie, reconstituer la faune de l’Egypte à cette époque,
excréments de bestiaux montrant qu’on s’occupait d’élevage, traces
de céréales grâce auxquelles nous apprenons qu’on connaissait déjà
l’agriculture. Ces documents qui ont si peu d’apparence et paraissent
négligeables sont donc extrêmement précieux, puisqu’ils font
connaître les occupations ordinaires, la nourriture, la vie privée des
premiers Egyptiens.

Tombea
ux

Si nous ne connaissons qu’un petit nombre de ces restes de

villages, dont la plupart ont dû entièrement disparaître ou bien sont
trop peu apparents pour qu’on puisse les distinguer, nous avons en
revanche une quantité considérable de sépultures appartenant à la
même époque. Ces tombes ne sont jamais isolées, mais forment des
nécropoles plus ou moins vastes, situées elles aussi au bord du
désert, près des terrains cultivés, donc à proximité immédiate des
habitations des vivants: en effet, chaque fois que nous
reconnaissons l’emplacement d’un kjoekkenmoedding, nous sommes
sûrs de trouver à peu de distance, quelques centaines de mètres à
peine, un cimetière qui est vraisemblablement celui des habitants du
village.
 Fig. 30. Tombeau prédynastique
(d’après Ayrton. El-Mahasna, pl. VI, fig. 26).

Ces nécropoles d’un type tout spécial ont très longtemps passé
inaperçues et elles semblent en effet, au premier abord, fort difficiles
à reconnaître. C’est avec le jour frisant du soir ou du matin qu’on
peut le mieux distinguer ces groupes de dépressions très légères, à
peine perceptibles en plein soleil, qui sont à la surface plus ou moins
inégale du terrain le seul indice extérieur des tombeaux archaïques.
Les sépultures sont de simples fosses creusées dans les bancs de
cailloux roulés qui s’étendent au pied de la montagne et qui forment
un terrain suffisamment consistant pour qu’il ne fût pas nécessaire
de soutenir, au moyen d’un mur ou d’un enduit, les bords de
l’excavation: leur forme générale est irrégulière, à peu près ovale ou
même presque ronde, et leur profondeur d’un mètre à deux au plus,
tandis que l’ouverture dépasse à peine un mètre cinquante dans sa
plus grande dimension. A côté de celles-là il en existait de plus
grandes, à peu près rectangulaires et atteignant jusqu’à quatre
mètres sur deux, sans que la profondeur en soit augmentée. Après
l’ensevelissement, les grandes comme les petites fosses étaient
simplement comblées avec du sable et des galets et se confondaient
avec le terrain environnant; il n’y a jamais la moindre superstructure,
pas même une pierre tombale.

Fig. 31. Tombeau prédynastique

(d’ap. J. de Morgan. Recherches sur les origines de l’Egypte, II, fig.
464).

Les dimensions des petites tombes, qui sont de beaucoup les plus
nombreuses, ne permettaient pas d’y déposer le mort étendu tout de
son long, comme on le fit plus tard pour les momies aux époques
historiques; les coutumes funéraires étaient en effet très différentes
et nous pouvons distinguer deux stages, deux modes
d’ensevelissement qui semblent correspondre à deux périodes. Dans
les plus anciennes sépultures, le mort est couché sur le côté gauche,
dans la position dite embryonnaire ou assise, c’est-à-dire avec les
membres repliés de manière que les mains se trouvent devant la
figure, les genoux à la hauteur de la poitrine et les pieds près du
bassin. Etant donnée l’orientation des tombeaux, qui du reste n’est
pas partout rigoureusement exacte, la tête est généralement au sud
la face tournée vers l’ouest.
 Fig. 32. Tombeau prédynastique
(d’après J. de Morgan. Recherches sur les origines de l’Egypte, II, fig.
468).

Le deuxième mode d’inhumation, qui paraît être un peu plus

récent, quoique appartenant toujours à la période prédynastique, est
beaucoup plus curieux: ici, et la chose a été constatée dans de très
nombreuses tombes, le corps était entièrement démembré avant
d’être déposé dans la fosse; les os ne sont ni cassés ni coupés, mais
ils sont placés pêle-mêle, et souvent il en manque un certain
nombre. Il ne s’agit pas d’un dépècement du mort au moment du
décès, ni de cannibalisme, comme on pourrait le croire, mais d’une
coutume qui se retrouve ailleurs qu’en Egypte, dans tout le bassin
de la Méditerranée, en Crète, dans les îles de l’Archipel, au sud de
l’Italie, celle de l’inhumation secondaire: on enterrait provisoirement
le mort, puis au bout de deux ou trois ans, quand les chairs s’étaient
putréfiées et désagrégées, on l’exhumait et on rassemblait les os
pour les déposer dans le tombeau définitif. La transition entre ces
deux coutumes funéraires, qui paraissent si différentes, est marquée
Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.

More than just a book-buying platform, we strive to be a bridge

connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.

Join us on a journey of knowledge exploration, passion nurturing, and

personal growth every day!

ebookbell.com