SCIENTIA

MANU E T MENTE

ANALYSIS OF RECOVERY-RECAPTURE DATA FOR LITTLE

A thesis submitted for the degree of Doctor of Philosophy

By Leesa A. Sidhu

School of Physical, Environmental and Mathematical Sciences, The University of New South Wales, Australian Defence Force Academy.

November 2007 ORIGINALITY STATEMENT

'I hereby declare that this submission is my own work and to the best of my knowledge it contains no materials previously published or written by another person, or substantial proportions of material which have been accepted for the award of any other degree or diploma at UNSW or any other educational institution, except where due acknowledge- ment is made in the thesis. Any contribution made to the research by others, with whom I have worked at UNSW or elsewhere, is explicitly acknowledged in the thesis. I also declare that the intellectual content of this thesis is the product of my own work, except to the extent that assistancefrom others in the project's designand conceptionor in style, presentation and linguistic expression is acknowledged.'

Signed... x.'as,a

Date r /t /: e?7 COPYRIGHT STATEMENT

'I hereby grant the UNSW or its agents the right to archive and to make available my thesis or dissertation in whole or part in the University libraries in all forms of media, now or here after known, subject to the provisions of the Copyright Act 1968. I retain all proprietary rights, such as patent rights. I also retain the right to use in future works (such as articles or books) all or part of this thesis or dissertation. I also authorise University Microfilms to use the three hundred and fifby word abstract of my thesis in Dissertation Abstract International (this is applicable to doctoral thesesonly). I have either used no substantial portions of copyright material in my thesis or I have obtained permission to use copyright material; where permission has not been granted I have applied/will apply for a partial restriction of the digital copy of my thesis or dissertation'' signed. ../ A.-f(&

Date ?/ u /aa7 z

ll AUTHENTICITY STATEMENT

'I certify that the Library deposit digital copy is a direct equivalent of the final officially approvedversion of my thesis. No emendationof contenthas occurred and if there are any minor variationsin formatting, they are the result of the conversionto digital format.' signed...../.Atk

Date 7fur/aa.a.7.

ltl Abstract

This thesis analyses yearly mark-recapture-recovery information collected over a 36- year period, from 1968 to 2003, for 23686 flipper-banded Little Penguins Eudyptula minor of Phillip , in south-eastern . Such a long-term data set is extremely rare for any species. Few studies of any have been able to model age dependence for the sur- vival, recapture and recovery probabilities simultaneously. I successfully apply such a modelling scheme and obtain biologically realistic age structures for the parame- ters. I also provide illustrations of erroneous results that may arise when analyses fail to consider simultaneous age dependence, or fail to detect annual variations that may mask age dependence. I obtain a low survival estimate of 17% in the first year of life, increasing to 71% in the second year, and around 80% thereafter, and declining gradually after age nine years. First-year survival increases with number of chicks fledged per pair, increases with annual average fledging weight and decreases with mean laying date. An in- crease in first-year survival is associated with warmer sea surface temperatures in the summer and autumn of the previous year, which agrees with biological consid- erations. Irrespective of this inter-year variation, born early in the breeding season, relative to the rest of their cohort, have greatly enhanced first-year survival, when compared to birds born late in that season. Fledglings survive better in years in which the mean fledgling weight is higher, and fledglings of above average weight have a better chance of survival than their underweight counterparts. I next analyse seven years of recapture data from a separate experiment studying the effect of banding on adult Little Penguins. In the year following marking, the

i survival probability of banded birds is 6% lower than that of unbanded birds, while in subsequent years the survival is 4% lower for banded birds. Band loss is negligible. Finally, I compare the survival estimates for with those obtained for a six-year study in . While first-year survival is significantly higher for New Zealand, there is a marked decline over time, coinciding with an increase in population size.

ii Acknowledgements

I wish to acknowledge my supervisors: Ted Catchpole from UNSW at ADFA, and Peter Dann from the Research Group at the Phillip Island Nature Park. I owe an enormous debt to my long-suffering supervisor, Ted Catchpole, who has continued to provide “full-time” supervision since his retirement in September 2004. I have greatly appreciated the way in which Ted has always been willing to offer his expert assistance at any time and on any topic — from statistical queries to LaTeX, to programming in Matlab or R, and so on. I have learned an enormous amount from Ted, and he has been a wonderful role model as a researcher. His kindness and his incredible enthusiasm for all of my achievements have been much appreciated. My PhD project has been greatly enhanced by the input of my co-supervisor, Peter Dann. I began this project with no biological background, and I have been very fortunate to be able to access Peter’s expert biological advice. On numerous occasions, I have sent Peter emails of questions that sometimes require him to retrieve the hand-written data from the early 1970s! I am grateful to Peter for patiently answering my endless questions about Little Penguins and the data, and for his good-humoured advice and continual encouragement throughout this project. I wish to thank Mark Collins for taking on the role of joint supervisor from September 2004 until he left ADFA in July 2006. Mark’s enthusiasm for my work and his continual encouragement and support have been much appreciated. I thank my husband, Harvinder, for always believing in me, and for his encour- agement and advice in helping to get me back on track on the odd occasion when I became slightly de-railed! I greatly appreciate the years he has spent juggling his own work and our children, Kurran and Dhiaan. His efforts have enabled me to complete this thesis.

iii I am grateful to my parents, Jim and Judy McEwan, for the unconditional love and support that they have always given me, and for their numerous car trips from Ipswich to Canberra to help look after the children, so that I could work on my thesis. Those four week trips have been invaluable. I also wish to thank my parents-in-law, Karam Singh and Sukhdev Kaur, for travelling from Singapore to help out with child care and home duties, and giving me time to work on my thesis. I am indebted to staff of the Phillip Island Nature Park, particularly Ros Jessop, Marg Healy and Leanne Renwick, and to members of the Study Group and the Australian and Bat Banding Scheme for providing the data upon which this study is based. I also wish to acknowledge Lynda Chambers (from the Australian Bureau of Meteorology) for providing sea surface temperature data and Figure 1.1, and Dave Houston who provided the data from Oamaru, New Zealand. I am grateful to Edda Johannesen for suggesting that we compare the survival of Little Penguins in Australia and New Zealand, and for providing valuable input into Chapter 8. I am grateful to my examiners for their careful appraisal of this thesis. Their insightful comments and suggestions were very valuable. This work was conducted under the support of an Australian Postgraduate Award, and using the resources of the School of Mathematics and Statistics and later the School of Physical, Environmental and Mathematical Sciences (PEMS) at UNSW @ ADFA. I wish to thank my colleagues in the School of PEMS for their friendship and encouragement throughout my thesis. I am also grateful to Anthony Tate and other members of the computing staff for sorting out those inevitable computer problems, to Annabelle Boag and David Paull for their assistance with some of the Figures, and to the mathematics and statistics staff for allowing me to do some intense “number crunching” on their machines after hours. I wish to acknowledge Professor J.M. (Mike) Cullen who worked with the pen- guin research group on Phillip Island for 25 years, up until his death in March 2001. My PhD project came into being when Mike sent the penguin data to Ted, slightly before I came to ADFA as an Associate Lecturer in 1997. Mike’s keen mind, incred- ible insight, endless hypothesizing and boundless enthusiasm for any result I sent him, no matter how seemingly insignificant, were inspirational. Years later, I have

iv obtained results of biological interest and commented to Ted, “That’s just the way Mike thought it would be!” Maximum likelihood fitting was performed via the MATLAB package, under the GNU-Linux operating system. It is not commonly known that the Linux mascot, “Tux”, is a , chosen by Linus Torvalds (the author of Linux) during a visit to the National Zoo and Aquarium in Canberra.

v Contents

Abstract i

Acknowledgements iii

Chapter 1 Introduction 1 1.1 BackgroundofStudy ...... 1 1.1.1 TheStudySites...... 2 1.1.2 TheDataCollection ...... 4 1.2 Rationaleforstudy ...... 5 1.3 Thesisoverview ...... 6 1.4 Publications arising from the thesis and in preparation ...... 9 1.5 Ananecdote...... 9

Chapter 2 Literature Review 11 2.1 Relevantecologicalresearch ...... 11 2.1.1 Physical characteristics and life cycle of the Little Penguin . 12 2.1.2 Small- and large-scale movements of the Little Penguin... 13 2.1.3 Factors affecting survival and breeding success ...... 15 2.1.4 Survival estimates for penguins ...... 21 2.2 Recaptureandrecoverymethods ...... 24 2.2.1 Somedefinitions...... 24 2.2.2 Anhistoricalperspective ...... 24 2.2.3 Issues in recapture and recovery studies ...... 32 2.2.4 Modelselection ...... 39 2.2.5 Computer software used in recapture and recovery analyses . 45

vi Chapter 3 The Data 47 3.1 TheRawData ...... 47 3.2 Disqualification of birds from raw data set ...... 49 3.2.1 Birdsbandedasadults ...... 49 3.2.2 Birds banded in locations other than Phillip Island . . ... 51 3.2.3 Experimentalbirds ...... 51 3.2.4 Birdsthatfailedtofledge ...... 52 3.2.5 Birds from the final cohort and dead recoveries in last six monthsofstudy...... 52 3.3 Summarizingthedata ...... 53 3.3.1 Rationale for using yearly summarized data ...... 53 3.3.2 Methodofsummarizingthedata ...... 53 3.4 An alternative method of summarizing the data ...... 57 3.5 Breedingdata...... 58 3.6 Somesimplestatistics ...... 59

Chapter 4 Modelling Age Dependence 63 4.1 Introduction...... 63 4.2 Methods...... 65 4.2.1 Mark-recapture-recoverydata ...... 65 4.2.2 Assumptions...... 66 4.2.3 Parameters ...... 67 4.2.4 Sufficientstatistics ...... 68 4.2.5 Thelikelihood...... 69 4.2.6 Fittingthemodels ...... 69 4.2.7 Modelselection ...... 70 4.3 Results...... 71 4.3.1 Leakagebetweenparameters ...... 71 4.3.2 Agestructuresformodelparameters ...... 73 4.3.3 Confoundingageandtime ...... 75 4.3.4 Modelaveraging ...... 76 4.3.5 Parameterestimates ...... 77

vii 4.3.6 Correlationsbetweenparameters ...... 78 4.3.7 Populationmodelling ...... 79 4.4 Overdispersion ...... 80 4.5 Discussion...... 82 4.5.1 Agestructures...... 82 4.5.2 Apparent versus true survival ...... 83 4.5.3 Comparison with other survival estimates ...... 84 4.6 Analyzing subsets of the complete data set ...... 85 4.6.1 Including birds that failed to fledge ...... 85 4.6.2 Recapture-only, recovery-only and recapture-recovery analy- ses using data for birds banded as chicks, those banded as adults,andallbirds ...... 86 4.7 “Strangers”analysis ...... 89 4.8 Estimates using old data summary and current method ...... 92 4.9 Conclusion...... 93

Chapter 5 Modelling Time and Covariate Dependence 95 5.1 Timedependence ...... 97 5.2 Biologicalcovariates ...... 103 5.3 Climaticconditions ...... 110 5.3.1 Localseasurfacetemperatures ...... 110 5.3.2 SouthernOscillationIndex ...... 118 5.4 Correlationsbetweencovariates ...... 121 5.5 Conclusion...... 123

Chapter 6 Individual Covariates 125 6.1 Introduction...... 126 6.1.1 Dateofbanding...... 126 6.1.2 Weightatbanding ...... 126 6.1.3 Sexandbilldepth ...... 127 6.1.4 Currentweight ...... 127 6.2 Calculatingthelikelihood ...... 128 6.2.1 Standard likelihood incorporating individual covariates . . . 128

viii 6.2.2 Missingcovariatevalues ...... 129 6.2.3 Thethree–statelikelihood ...... 131 6.3 Results...... 133 6.3.1 Dateofbanding...... 133 6.3.2 Weightatbanding ...... 140 6.3.3 Sexandbilldepth ...... 145 6.3.4 Currentweight ...... 147 6.3.5 Correlations between covariates ...... 151 6.4 Conclusion...... 153

Chapter 7 Banding Effects 155 7.1 Introduction...... 155 7.2 TheData ...... 156 7.2.1 Collectionofdata...... 156 7.2.2 Sampleofdata ...... 157 7.3 Analysis ...... 161 7.3.1 Notation...... 161 7.3.2 Thelikelihood...... 162 7.4 Results...... 165 7.4.1 Recaptureprobability ...... 166 7.4.2 Survivalprobability...... 167 7.4.3 Bandandtransponderloss...... 177 7.5 Conclusion...... 178

Chapter 8 Comparison with a New Zealand colony 180 8.1 Introduction...... 180 8.1.1 Oamarustudyareas ...... 180 8.1.2 MonitoringinOamaru ...... 181 8.1.3 Dataanalysis ...... 181 8.2 Results...... 182 8.2.1 Age structures and estimates for parameters ...... 182 8.2.2 Timevariationinsurvival ...... 183 8.2.3 Dependence of first-year survival on biological covariates . . 184

ix 8.2.4 Densitydependence...... 186 8.2.5 Correlations between biological parameters ...... 188 8.3 Conclusion...... 189

Chapter 9 Conclusion 191 9.1 Summaryofresults ...... 191 9.2 Futurework...... 195

Appendix A Deriving and programming the likelihood 197 A.1 Derivingthelikelihood ...... 197 A.2 Calculating the likelihood ...... 204 A.2.1 Modelnotation ...... 204 A.2.2 Fittingthemodels ...... 205 A.3 Sufficientstatistics ...... 211

Appendix B Codes for Raw Data 220 B.1 Methods of Encounter and Status Codes ...... 220 B.1.1 Methodsofencounter...... 220 B.1.2 Statuscodes...... 223 B.2 ExperimentalCodes ...... 226 B.3 LocationCodes ...... 228 B.3.1 Phillip Island Location Codes ...... 228 B.3.2 GeneralLocationCodes ...... 229

References 235

x Chapter 1

Introduction

1.1 Background of Study

Phillip Island (38◦30′S, 145◦10′E) is located in , approximately 140 km south-east of Melbourne, (Figure 1.1). The Little Penguins (Eudyptula minor) at Phillip Island come ashore at dusk in groups, cross the beach or rocks, and make their way to their burrows in the sand dunes. This behaviour, together with their small stature and appealing looks, make the Little Penguins a popular wildlife tourist attraction. Visitors were ferried across to Phillip Island to watch the penguins crossing the beach as early as 1928 (Newman, 1992). Indeed, with 500 000 visitors each year, the Phillip Island Penguin Parade is Australia’s most visited wildlife attraction (Penguin Reserve Committee of Management, 1992). In 1966, the Manager of the Penguin Reserve at Phillip Island, Bert West, asked the Australian Bird-Banding Scheme (ABBS) from the Division of Wildlife Research of the Commonwealth Scientific and Industrial Research Organisation (CSIRO) to provide flipper-bands to enable him to tag the Little Penguins (Reilly, 1983). Pauline Reilly was asked by ABBS to go to Phillip Island to explain to West that they could not grant his request since no suitable bands existed. Reilly discovered that although West had been studying the penguins for forty years and a few short-term banding studies had been reported in ornithological literature, no one was able to determine answers to such basic questions as how long they lived, and whether they mated for life. It was found that an ornithologist from New Zealand, Fred Kinsky, had been studying Little Penguins for 15 months, and had managed to purchase reasonably priced bands from a Swedish manufacturer. From

1 Figure 1.1: The location of Phillip Island.

October 1966, Reilly began to organise a research program to be undertaken on Phillip Island, and flipper-banding commenced in 1968.

1.1.1 The Study Sites

The distribution of breeding colonies on Phillip Island has diminished several-fold over the past 100 years, largely due to introduced predators (Dann, 1992). The remaining breeding colony is restricted to the 340 hectare Penguin Reserve on the Summerland Peninsula at the south-western end of Phillip Island (Figure 1.2), which is characterised by woodlands and Poa grasslands across the centre, and suc- culent herblands containing the breeding areas around the perimeter. The current Little Penguin breeding population on Phillip Island has been estimated at 26 000 birds (P. Dann, unpubl. data). The recapture data considered in this thesis were collected in a variety of study sites in the breeding areas on the Summerland Peninsula. The number and location of study sites visited, and the frequency of visits to the burrows in these study sites have varied over the years. Visits to burrows in the initial study area (the Penguin Study Group site, Site 1 in Figure 1.2) varied from weekly (1968–1970) to monthly (1971–1984) to two-weekly during the breeding season and monthly outside the breeding season (1985–2004), and then back to monthly all year round. Additional study sites were added in 1981 (Site 2) and 1984 (Sites 3–8) and visited monthly

2 3 5 4

1 8

6 7

2

Figure 1.2: The distribution of breeding areas (shaded) of Little Penguins on the Summerland Peninsula, Phillip Island in December 2002. The numbers denote study sites, while the hatched region represents the housing estate. until 1999 and then at two-weekly intervals during breeding and infrequently at other times. Sites 7 and 8 are visited fortnightly all year round. These regular visits to Sites 1, 7 and 8 throughout the year enable the researchers to keep track of the timing of the Little Penguins’ breeding activities. At the start of each breeding season, the study site boundaries are “defined”, and perhaps adjusted slightly from previous years to keep the number of burrows approximately constant. The study sites also vary in the proportions of natural/artificial burrows. Several sites have only natural burrows, others (Sites 7 and 8) have only artificial wooden nest boxes, while the remaining sites have a mixture of natural burrows and nest boxes. Sites 7 and 8 (see Figure 1.2) are located in a housing estate. Penguins were living in the area when the estate was opened. However, due to the severe impact of the housing estate on the penguins, the Victorian Government initiated a “buy- back” scheme in 1985, as part of their “Penguin Protection Plan” (Department of the Environment and Heritage, 1993). All 180 buildings and 780 allotments of the housing estate will eventually be bought back and revert to public ownership. By 1999 around 600 allotments had been bought back by the Victorian Government (M. Cullen, pers. comm.). The penguin population in these sites has increased since the implementation of this buy-back scheme.

3 1.1.2 The Data Collection

The data used in this study come from the Phillip Island Nature Park Research Group, headed by Peter Dann, and the Penguin Study Group (PSG), a branch of the Victorian Ornithological Research Group1. The main study includes fledged chicks banded from the breeding season in 1967/1968 until the 2002/2003 breeding season: 23 686 penguins were banded as chicks, with a further 13 075 banded as adults of unknown age. However only known aged birds (that is, birds banded as chicks) are included in the present study. A detailed explanation of the raw data can be found in Chapter 3.

Live recapture data The live recaptures occur during visits to the study sites on Phillip Island. In each study site, the burrows are marked by numbered pegs and their locations are shown on a map of the site. The data collection process is particularly challenging for natural burrows, whereby researchers lie flat on the ground, insert an arm into the burrow and remove its contents (that is, adults, chicks and eggs), while making sure that they do not remove any snakes in the process! Birds are banded if necessary, various information is recorded and they are placed back into their burrow. Any unbanded adults encountered in the study sites up until 2002 were banded with flipper bands. (Since 2002, injected transponders have been used to mark individuals.) A chick is banded prior to fledging at around six weeks of age, once its flippers are sufficiently developed to hold a band and its weight is at least 500 grams. Since the down on the fledgling’s flippers has been replaced by steel-blue feathers by this stage, the band cannot damage the erupting feathers. Due to the nature of the experimental methods, a bird breeding in a study site in a given year will almost always be recaptured in that year. However, if a bird does not breed in a season, or if it breeds outside of the study site, it may not be encountered.

Dead recovery data The recoveries of bands from dead penguins are made by the Phillip Island Nature Park research staff, or by members of the public. Bands are returned to the Aus-

1http://home.vicnet.net.au/∼vorg/lecvorg.htm

4 tralian Bird and Bat Banding Scheme (ABBBS), which is currently administered by the Department of the Environment and Heritage2. Dead recoveries come from the breeding areas on Phillip Island and along the coastline of southern Australia, as far west as Kangaroo Island in South Australia (see Figure 4.1 on p.65).

1.2 Rationale for study

Research into the ecology of the Little Penguins at Phillip Island has been un- derway since 1968 (Reilly and Cullen, 1979, 1981, 1982, 1983; Dann and Cullen, 1990; Dann et al., 1995; Dann and Norman, 2006). Although a great deal has been learned about the Little Penguins and their breeding biology, breeding success, diet, and movements and patterns of mortality at sea, less is known about their annual survival and how this important parameter varies with age and over time. Sev- eral empirical estimates of mortality rates have been made using life-table analyses (Reilly and Cullen, 1979; Dann and Cullen, 1990). However, this current project is the first to carry out a detailed mark-recapture-recovery analysis using the live recapture and dead recovery information, in order to study the age-specific survival of Little Penguins. From a biological viewpoint, the survival probability is the parameter of greatest interest. However, several previous studies of Little Penguins (Reilly and Cullen, 1981, 1982; Dann and Cullen, 1990; Dann et al., 1992) have highlighted the differing lifestyles of these birds as they age, suggesting that their survival, recapture and recovery probabilities are all likely to vary with age. There may also be a senescence effect, whereby the survival probability reduces as adults age. Hence models for these parameters should allow them to vary with age simultaneously. Since birds banded as adults are of unknown age, I consider only birds banded as chicks in this analysis, and attempt to develop age structures for the model parameters that fit the data and make biological sense. I expect the survival probability during the first and subsequent years of life to vary with time, due to conditions (such as food availability) varying from year to year (M. Cullen, pers. comm.). This temporal variation can potentially be described

2http://www.deh.gov.au/biodiversity/science/abbbs/

5 using group covariates such as mean laying date and mean annual fledgling weight. Individual traits of birds (such as individual “birth” date, bill depth and current weight) are also likely to affect their survival.

1.3 Thesis overview

This thesis analyses yearly mark-recapture-recovery information collected over a 36-year period for the Little Penguins Eudyptula minor of Phillip Island, in south- eastern Australia. Such a long-term data set is extremely rare for any species. Chapter 2 of the thesis reviews the relevant ecological literature and provides an overview of the statistical theory that has been developed to address the estimation of annual survival, with particular focus on mark-recapture-recovery methods. The data require very careful consideration, because of their complexity. Chap- ter 3 includes an explanation of the raw data, and a rationale for my methods of summarizing the data. Previous estimates of annual survival for Little Penguins (particularly for pre- breeders) have been tenuous due to the methods used (life tables). Other studies of wild populations have been based on small sample sizes or short-term data sets, so that annual variation in conditions may have masked the age dependence in the sur- vival estimates. Few studies of any animal have been able to model age dependence for the survival, recapture and recovery probabilities simultaneously. However, in Chapter 4, I successfully apply such a modelling scheme and obtain biologically realistic age structures for the model parameters. I also provide illustrations of potentially erroneous results that may arise when analyses fail to consider such si- multaneous age dependence, or fail to detect annual variations that may mask age dependence. From 1968 to 2004, 23 686 penguin chicks were flipper-banded before fledging on Phillip Island, with 2979 birds encountered (dead or alive) after fledging, and 1347 ultimately recovered dead. I obtain a low survival estimate of 17% in the first year of life, increasing to 71% in the second year of life, 78% in the third year, and 83% thereafter, and declining gradually after age nine years. A population model allowing for immigration of birds from areas surrounding the study sites fits the observed stable population in the study sites.

6 In Chapter 5, I use the age structures for survival, recapture and recovery as determined in Chapter 4, and identify the age components of the model parameters that exhibit temporal variation. I next examine several time-varying group (extrin- sic) covariates, in order to explain some of the observed temporal variation in the survival probability. First-year survival increases with number of chicks fledged per pair, increases with annual average fledging weight and decreases with mean laying date (that is, late breeding is associated with low first-year survival). I also use seasonal sea surface temperature (SST) data for various areas of , and Southern Oscillation Index (SOI) data as covariates for first year and adult survival. An increase in first-year survival probability is associated with warmer sea surface temperatures in the summer and autumn of the previous year and in the autumn af- ter fledging. These results agree with biological considerations (Chambers, 2004a), since obtaining a good food supply prior to and during moult puts the parents in a better position to raise their chicks successfully, and the autumn following fledging is the period of greatest mortality of birds in their first year. While conditions that apply to all may certainly affect their probabilities of surviving, or being recaptured or recovered, individual traits are also likely to affect the model parameters. In Chapter 6, I consider intrinsic traits, such as weight at banding, date of banding, and current weight, which vary from one individual to another, as covariates for survival. While good seasons obviously produce chicks that are more likely to survive their first year of life, birds born early in the breeding season, relative to the rest of their cohort, have a greatly enhanced probability of surviving their first year of life, when compared to birds born late in the same season. Fledglings survive better in years in which the mean fledgling weight is higher, and fledglings of above average weight, relative to others in their cohort, have a better chance of survival than their underweight counterparts. However, there appears to be an optimal weight of 1300 grams, beyond which fledglings have a reduced probability of surviving their first year of life. It is possible that the survival probabilities for banded Little Penguins (esti- mated in Chapter 4) are significantly lower than the survival probabilities of un- banded birds, since bands may hamper birds energetically, make them more obvious to predators, or cause injuries (for example, see Sladen et al., 1968; Culik et al.,

7 1993; Jackson and Wilson, 2002). In 1995 the Research Group from the Phillip Island Nature Park began a separate study which sought to determine whether banding affects the mortality of Little Penguins. In Chapter 7, I analyse seven years of recapture data for three groups of birds; one flipper-banded group, one unbanded group that had been implanted with passive-induction transponders and one group with both tags. This enables me to estimate the rate of band loss for adults, so that (if necessary) this can be allowed for in the main study. It also allows me to estimate any reduction in survival due to banding — which is relevant to the future use of banding as a method for studying Little Penguin biology. The annual probability of losing a band is around 0.4%, while the probability of losing a transponder is 5% in the first year of marking, and around 1% in subsequent years. Although the recapture probability varies considerably from year to year, this parameter does not depend on the type of mark used. My results show a po- tential “marking effect”, whereby the survival of penguins in their first year after marking is reduced, regardless of the type of mark used. In the 12 months following marking, the average survival probability of banded birds is 6% lower than that of unbanded birds, while in subsequent years the average annual survival probability of banded birds is around 4% lower than their unbanded counterparts. While the effect of banding is virtually the same for both males and females in the year follow- ing marking, I found that males survive better than females. In fact, banded males survive almost as well as unbanded females. This enhanced survival probability for males could be due to a size effect, as penguins with larger bill depths appear to have higher probabilities of survival. The results of this work will be used by the Department of the Environment and Heritage to determine whether banding of Little Penguins will be allowed to continue in Australia. In Chapter 8, I analyse mark-recapture-recovery data from two locations: Phillip Island, Australia and Oamaru, New Zealand. While the Phillip Island study has been underway for almost 40 years, the New Zealand study consists of six years of data. Here I estimate and compare the survival of penguins in these two locations, and determine to what extent the conclusions on covariate dependence of survival can be extended from Phillip Island to another penguin colony. The first-year survival of the Oamaru birds is significantly higher than that of Phillip Island.

8 However, in Oamaru there is a marked decline in first-year survival over time, coinciding with a rapid increase in population size. Finally, I provide a summary of the study and some suggestions for future work in Chapter 9.

1.4 Publications arising from the thesis and in preparation

1. Sidhu, L.A., Catchpole, E.A., and Dann, P. (2007). Mark-recapture-recovery modelling and age-related survival in Little Penguins Eudyptula minor. The Auk, 124, 815–827. (Based on Chapter 4 of thesis.) 2. Sidhu, L.A., Catchpole, E.A., and Dann, P. Modelling time dependence in the survival, recapture and recovery probabilities for Little Penguins Eudyptula minor. In preparation. (Based on Chapter 5 of thesis.) 3. Sidhu, L.A., Catchpole, E.A., and Dann, P. Individual covariates for the survival probability for Little Penguins Eudyptula minor. In preparation. (Based on Chapter 6 of thesis.) 4. Dann, P., Sidhu, L.A., Catchpole, E.A., Jessop, R., Cullen, M., Renwick, L., Healy, M., Collins, P., and Baker, B. The effects of flipper bands on the survival of Little Penguins Eudyptula minor. In preparation. (Based on Chapter 7 of thesis.) 5. Sidhu, L.A., Johannesen, E., Catchpole, E.A., Dann, P., and Houston, D. First-year survival of Little Penguins Eudyptula minor: A comparison be- tween Phillip Island, Australia and Oamaru, New Zealand. In preparation. (Based on Chapter 8 of thesis.)

1.5 An anecdote

I conclude this chapter with an anecdote illustrating the incredible amount of data collected by the Phillip Island researchers and members of the Penguin Study Group over the past 40 years. Upon examining some of the monthly records, I noted that some birds (such as penguins 17065 and 78098) were banded as chicks and recovered dead within the same month. I concluded that the birds could probably be deemed to have fledged in that short period between being banded and recovered

9 dead, since they were found dead floating in the water or beachwashed. When I asked Mike Cullen for his opinion, his emailed reply was as follows:

Just a point about the birds 17065 and 78098 which, even on the monthly data, only have entries of “2”. The first was banded as a chick on 8 March and found beachwashed on 27 March at Apollo Bay 100 km or so to the west. The general strategy is to band chicks as close as possible to fledging, so it is not surprising that nearly three weeks later it should have left its burrow. The other bird was an experimental one, who had been transported when about to fledge to a foster home in a burrow in part of the colony where we want to build up the breeding numbers. It was installed in its foster home on 1 January 1994, given a meal of pilchards (by the research people) on that day and the next, and had headed off to sea by 3 January. It then turned up next beachwashed on 13 January at Fairhaven, also quite close to Apollo Bay. I suppose these birds reinforce the point that a month is not an instantaneous time window for a penguin.

Ted Catchpole’s reaction to this was that he couldn’t remember what he had for breakfast that morning, but Mike was able to tell me what this penguin had for dinner one evening in 1994!

10 Chapter 2

Literature Review

In this chapter I review research relating to penguins and to the statistical theory on recapture and recovery methods, since both of these fields of research are relevant to the work undertaken in this study. Section 2.1 includes an examination of the pertinent ecological re- search concerning Little Penguins and other seabirds. In particular, I focus on the physical characteristics and life cycles of Little Penguins, as well as issues such as small- and large-scale movements of these birds, philopatry, senescence, and the effects of banding and climate variation. I then examine previous survival estimates for Little Penguins, and for other penguin species. Section 2.2 focuses on the statistical theory that has been developed to address the estimation of annual survival in animal ecology. I begin with a brief historical overview of this field of research, highlighting the major developments in capture-recapture, mark-recovery and joint mark-recapture-recovery analyses. I next discuss issues in recapture and recovery studies, such as age dependence in the model parameters, tag loss, covariate dependence (including the problem of missing values for time-varying individual covariates), parameter redundancy and model selection.

2.1 Relevant ecological research

Research into the ecology of the Little Penguins at Phillip Island has been underway since the commencement of banding in 1968. Since that time, an enormous amount

11 has been learned about the Little Penguins and their breeding biology, breeding suc- cess, diet, and movements and mortality at sea. Some of the more significant studies conducted on Phillip Island include: a comprehensive series of papers by Pauline Reilly and Mike Cullen examining the mortality of adults (Reilly and Cullen, 1979), aspects of breeding such as breeding success, pair bonds and failure to breed (Reilly and Cullen, 1981), dispersal of chicks from their natal colony and survival after banding (Reilly and Cullen, 1982), and the location, timing and duration of the moult (Reilly and Cullen, 1983); as well as important research conducted by Peter Dann and associates which examined the survival, patterns of reproduction and reproductive output of these birds (Dann and Cullen, 1990), population trends and factors influencing the population size (Dann, 1992), reproductive cost (Dann et al., 1995), and the role of intraspecific competition in population regulation (Dann and Norman, 2006).

2.1.1 Physical characteristics and life cycle of the Little Penguin

Little Penguins are found along the coast of New Zealand and southern Australia (Williams, 1995). In Australia they breed on or in inaccessible parts of the coast in southern Australia, from northern New South Wales, around southern Australia and as far west as Perth in Western Australia (Reilly, 1983). The Little Penguin is the smallest penguin species, with a standing height of approximately 33 cm (Reilly, 1983). Although the mean mass of an adult penguin is around 1100 grams for a male and 1000 grams for a female (Dann et al., 1995), penguins have been known to double in weight prior to moult. Little Penguins moult annually after their first year of life. The moult takes place between February and April, and lasts for 15–20 days (Reilly and Cullen, 1983). The average life expectancy of the Little Penguin is approximately 6.5 years (Reilly and Cullen, 1979; Dann and Cullen, 1990; Dann et al., 1995), although the oldest known bird was recaptured at 25 years of age (Dann et al., 2005). As is fairly common with birds, it is difficult for a field worker to determine the sex of a Little Penguin directly. However, the sex of an adult Little Penguin can be determined by measuring the depth of its bill, since males have stouter bills than females. Indeed, birds of

12 breeding age can be sexed to an accuracy of 91% using their bill depth measurements (the bill depth of a male is greater than 13.3 mm, Arnould et al., 2004). The breeding season for the Little Penguin is quite extensive and quite variable from year to year, extending from July or August through to April at Phillip Island (Reilly and Cullen, 1981). While egg-laying usually occurs between August and December, it has been known to commence as early as May in some breeding seasons, or finish as late as January in other seasons (Reilly, 1983). Each clutch contains two eggs, with a second clutch occasionally following a successful first (Gales, 1985). The parents take turns to incubate the eggs over a 35-day period. Once the chicks hatch, one parent stays at the burrow with the chicks while the other goes out during the day to find food. By the time the growing chicks are around two weeks of age, both parents go out daily to find food (Reilly, 1983). The parents come ashore at dusk in groups, waddle across the beach and make their way back to their chicks and the relative safety of their burrows (Klomp and Wooller, 1991). Once young birds “fledge” at around 8 weeks of age, they go to sea and spend most of their time away from the colony (Reilly and Cullen, 1981) until their first or second moult. They begin to breed at around two or three years of age (50% of two-year-olds breed, Dann and Cullen, 1990). Pre-breeders are usually referred to as “juveniles”. However, they are not distinguishable by sight from breeding adults, since their plumage is identical (Reilly and Cullen, 1982).

2.1.2 Small- and large-scale movements of the Little Penguin

The large-scale spread of newly fledged birds away from their colony of birth is called “natal dispersal” (Reilly and Cullen, 1982). Young Little Penguins spend their first winter away from their natal site, possibly returning for their first moult at one or two years of age. All young and many adult Little Penguins disperse to wintering areas between breeding seasons. However, while the numbers of Phillip Island birds in the breeding areas are greatly reduced between breeding seasons, some birds are present throughout the year (M. Cullen, pers. comm.). Dann et al. (1992) shows the different patterns of movement for adults and juveniles in their first year of life.

13 “Philopatry” (the tendency to return to breed in the natal area) is very strong for Little Penguins. Young penguins generally return to their natal site (or somewhere close by) to breed for the first time. Of the birds that were banded as chicks and recaptured later as adults, 93% were recaptured in the area of their natal site (Dann, 1992). However, birds banded as chicks on Phillip Island are occasionally recorded breeding elsewhere. For example, two birds that were banded as chicks on Phillip Island were recaptured alive at the St Kilda colony in the same year, and later became regular breeders at St Kilda (M. Cullen, pers. comm.). Once a bird breeds at a site, it will usually return to the same area to breed for the rest of its life. Although the main study sites on Phillip Island are separated by at most a few hundred metres, birds breeding in a study site almost always return to that site in subsequent years, if they are recorded breeding at all (M. Cullen, pers. comm.). Such an observation is confirmed by Johannesen et al. (2002a), in their study of the effect of breeding success on nest and colony fidelity for Little Penguins in two study sites in Otago, New Zealand. These authors reported that there were no movements of breeding birds from one study site to the other, but that successful breeders had a higher probability of returning to the same site to breed in the following year than unsuccessful breeders. Furthermore, penguins generally breed in the same burrow from one year to the next (Reilly and Cullen, 1981). Johannesen et al. (2002a) reported that burrow fidelity (which was measured at 72%) was higher for more successful breeders, and for birds breeding in wooden nest boxes (as opposed to natural burrows). A bird may move into a nearby burrow if its previous burrow collapsed or was taken over by another breeding pair (M. Cullen, pers. comm.). Birds with burrows near the boundary of a study site may move a few metres outside the site during some breeding seasons, and may consequently fail to be recaptured for one or more years. Around 15% of breeding birds were involved in such small-scale movements out of a study site in one season and returning to the site subsequently (Dann and Cullen, 1990). Larger scale movements between breeding areas on Phillip Island, or emigration to other colonies in Bass Strait are quite rare (Dann, 1992).

14 2.1.3 Factors affecting survival and breeding success

Age and experience For most species of long-lived animals, juvenile survival is lower than that of adults (see for example Stearns, 1992; Saether and Bakke, 2000). The survival probability for Little Penguins appears to exhibit such a juvenile–adult age effect (Dann and Cullen, 1990). This is not surprising since a particularly critical time in the life of a Little Penguin is when the newly-fledged bird goes to sea and starts foraging, without any experience or any parental guidance. Indeed, dead recovery records have indicated that the main period of mortality for juveniles occurs during March or April, three or four months after fledging (Dann et al., 1992). Age and experience also affect the breeding success of Little Penguins. Chiaradia and Kerry (1999), in an investigation of the nest attendance and breeding perfor- mance of Phillip Island Little Penguins, found that birds that bred earlier in the season were more successful than later breeders. Furthermore, they observed a pos- sible age-effect on breeding success, whereby successful breeders were an average of one year older than unsuccessful breeders. An earlier study by Dann and Cullen (1990) also reported that breeding success increased with age, for the first five years of breeding. In their study of the breeding performance and survival of Little Penguins in New Zealand, Johannesen et al. (2003) reported that the incidence of double clutching increased with age. They also found that birds laying two clutches in a season had an increased annual survival probability after the breeding season, suggesting that the tendency to lay double clutches was due to quality (or fitness) difference amongst the breeding birds. There was no significant difference between the post- fledgling survival probabilities for birds whose parents laid single or double clutches. Whereas the survival of fledglings whose parents laid single clutches declined with laying date, this was not the case for young whose parents laid double clutches.

Senescence Another potential cause of age dependence in the survival probability is senescence, whereby the survival probability reduces as adults age. Senescent effects have been recorded in many seabirds. For example, Harris et al. (1994) reported that the

15 survival probability for the European Shag Phalacrocorax aristotelis declined from the age of 14 years, while Bradley et al. (1989) found a reduction in the survival of the Short-tailed Shearwater Puffinus tenuirostris from 13 years, and Aebischer and Coulson (1990) reported a similar decline in survival from 12 years for the Kittiwake Rissa tridactyla. Nichols et al. (1997) pointed out that the small numbers of observations for “older” adults in many studies of wild populations make it difficult for researchers to form valid inferences about age-specific survival. Indeed, these authors rec- ommended that such studies should use large samples of animals, collected over many years, in order to study age dependence. Furthermore, Nichols et al. (1997) claimed that much of the earlier research that showed declines in survival with age was based on estimation methodology that required very restrictive assumptions (e.g., life-table analyses). However, more recent research modelled a linear decline in survival, on a logistic scale, within a capture-recapture framework (Pugesek et al., 1995; Nichols et al., 1997). Fletcher and Efford (2007) studied the effect of senescence on the adult survival estimate when the age was unknown. They remarked that if a species experienced senescence, adult survival would be underestimated, since mark-recapture studies often include individuals that are older than the average of the general population. Nisbet (2001) discussed methodological issues arising from the detection and measurement of senescence in wild birds, with particular focus on the long-lived seabird, the common tern Sterna hirundo. He warned that a spurious apparent de- cline in survival probability with age can arise when the quality of the environment deteriorates during a study, since this causes a decline in the survival probabilities for each age group over time. However, since birds often enter a study as fledglings, the average age of the birds in the study is increasing over time, and so the com- bined effect of these two trends makes it appear that older birds have lower survival rates.

Environmental conditions The lack of experience of newly-fledged birds in finding and catching food makes them more vulnerable to low food availability, so that their survival varies consid- erably according to good and bad years, leading to a higher temporal variability in

16 juvenile survival when compared to adult birds (Saether and Bakke, 2000). Good and bad years for seabirds are mainly determined by the availability of marine prey (see for example Chastel et al., 1993). Marine prey availability is hard to measure directly. Indeed, foraging performance, breeding success and breeding numbers of seabirds have been used as indicators of changes in marine ecosystems (see for ex- ample Bost and Le Maho, 1993; Le Maho et al., 1993; Cherel and Weimerskirch, 1995; Caldow and Furness, 2000). Several studies of Little Penguins on Phillip Island have considered the associa- tions between various measures of breeding success, and the relationships between these measures and post-fledgling survival. Significant correlations have been found between most of the breeding success variables (Mickelson et al., 1992; Chambers, 2004b). In particular, when penguins commence egg-laying earlier, the weight of chicks and the number of chicks fledged per pair were higher. Birds banded in the first half of the season were more likely to survive their first year of life than birds banded later in the season (Reilly and Cullen, 1982), and birds with higher weight at the time of banding were more likely to survive their first year than their underweight counterparts (Reilly and Cullen, 1982; Dann, 1988). Lack of prey availability can also affect the survival of adult Little Penguins. Dann et al. (1992) found that adult Little Penguins are most vulnerable after moult (ie. March or April) and then again in early spring (September) when they some- times succumb to starvation. Similarly, Johannesen et al. (2002b) studied the sea- sonal variation in survival of Little Penguins in Otago, New Zealand, and found that the survival probability for adults was lowest in the months following the moult, but that there were no adult mortalities recorded from egg laying to the time of fledging. Dann et al. (2000) state that a widespread pilchard mortality from March to May 1995 led to an increase in penguin mortality, later egg-laying in the following breeding season, and lower than average chick fledging rate (0.3 chicks fledged per pair compared to a mean of 1.0). There was also a significant reduction in birds coming ashore at Phillip Island, and an increase in the numbers of birds recovered dead.

17 Climate Sandvik et al. (2005), in their study of the effect of climate on North Atlantic seabirds, claimed that few studies have examined the effect of “non-catastrophic climate variation” on the adult survival probability of long-lived seabirds, even though survival is an important parameter in the study of any long-lived species. Although our understanding of the effect of climate on the breeding success and survival of seabirds in Australia is limited, the species of fish upon which Little Penguins feed are acknowledged to be very sensitive to variations in environmental conditions (Stenseth et al., 2002; Chiaradia et al., 2003). Furthermore, since Little Penguins are shallow-diving seabirds, sea surface temperatures (SSTs) are likely to affect their foraging patterns and prey abundance (Jarvis, 1993; Montevecchi, 1993). Fortescue (1998), in a study of Little Penguins from Bowen Island, reported an increase in the abundance of various fish species during periods of warmer water temperatures in Jervis Bay, New South Wales. Bunce (2000) stated that pilchards (a prey species of Little Penguins) moved into Bay during warm water periods. Several studies have shown that local climatic conditions (such as the SSTs in regions near Phillip Island, Mickelson et al., 1992; Chambers, 2004a,b) and global conditions (Chambers, 2004b) have marked effects on various Little Penguin breed- ing parameters. Chambers (2004b) found that local SSTs had a greater influence on breeding parameters than global conditions. She showed that local SSTs affected the breeding success, the mean laying date, the number of chicks fledged per breed- ing pair and the weight of the chicks, whereas global conditions influenced hatching success. Chambers (2004a) reported that warm SSTs in summer and autumn are associated with earlier breeding in the following breeding season. She provided the following biological explanation for this result: in summer, once breeding adults finish breeding and caring for chicks, they feed intensively to build themselves up in preparation for the moult, which lasts for two or three weeks and occurs between February and April (Reilly and Cullen, 1981). At the completion of moult, in au- tumn, penguins again feed intensively. Therefore, a good food supply in summer and autumn has the greatest effect on the birds breeding success in the following breeding season.

18 Mickelson et al. (1992) found that breeding success varied considerably from year to year. They reported that a decreased east–west SST gradient across Bass Strait is associated with earlier breeding, heavier chicks and more chicks per pair. They also found that a decreased SST gradient is associated with an increase in weights of adult penguins four months later. These authors claimed that this effect was due to cooler waters having slightly higher concentrations of nutrients and, as a result, an increase in abundance of penguin prey species. Perriman et al. (2000) studied the effects of climate fluctuations on the breeding of Little Penguins in New Zealand. They found that the presence of La Ni˜na conditions (with higher than average SSTs in New Zealand) was associated with later than usual breeding and a lower incidence of double breeding in the following breeding season. The findings of Mickelson et al. (1992) and Perriman et al. (2000) are in contrast with Chambers (2004a), which highlights the complexity of the relationship between climate variation and breeding success of penguins. Several studies in South America and the Equatorial Pacific have shown that cool water conditions have had a positive effect on penguins and other sea birds. Gal´apagos penguins Spheniscus mendiculus increased in weight during cold surface water conditions (Boersma, 1998), and decreased in weight and failed to breed during warm water periods (Boersma, 1998; Valle and Coulter, 1987). Chambers (2004b) highlights the reasons for the potential differences in the effect of warmer SSTs on seabirds in equatorial regions and regions of higher latitude such as southern Australia. In equatorial regions, higher SSTs are associated with lower seabird productivity and lower survival, since surface water temperatures are too high for fish, making it difficult for surface-foraging seabirds to find food (Montevec- chi, 1993). However, in high latitude regions, colder than average waters leads to longer hatching times for plankton and slower growing demersal fish (Montevecchi, 1993). Therefore warmer SSTs are associated with increased seabird productivity in regions with higher latitudes.

Banding effect For more than 50 years, researchers have been marking penguins with flipper bands (Sladen, 1952). Leg bands cannot be safely used on penguins, as their legs are too

19 short (Department of the Environment and Heritage, 2005). While early research into banding focused on extending the life of bands (e.g., Sladen and LeResche, 1970), more recent research has studied the effect of banding on the mortality and energetics of several penguin species. Excellent reviews of these banding effects studies appear in Stonehouse (1999), Jackson and Wilson (2002) and Petersen et al. (2006). A small number of penguins have been injured or killed as a direct result of their flipper bands: sometimes when a band that had not been properly sealed had partially opened and cut into the birds flesh, when the band became entangled in fishing gear (P. Dann, pers. comm.), through injuries caused by the restrictive band to the swollen flipper during moult (Sladen et al., 1968), or possibly due to the shiny silver band attracting predators (Jackson and Wilson, 2002). Some of these negative effects have been reduced through the design of better bands (Sladen et al., 1968). Indeed, plastic bands are currently being tested by Underhill and associates on African Penguins Spheniscus demersus in Robben Island, South Africa (see Jackson and Wilson, 2002). However, bands can potentially have a much more insidious effect on their wearers. Banded Ad´elie Penguins adeliae have been found to have lower sur- vival rates than unbanded penguins until their first moult after banding (Ainley et al., 1983), and to use more energy than unbanded penguins when swimming (Culik et al., 1993). More recently, Dugger et al. (2006) used capture-recapture methods to study the effect of banding on the foraging behaviour and survival of these penguins. These authors allowed the banding effect to vary between the sexes (due to the differences in body sizes for males and females), and found that male penguins had higher survival than females, in both banded and unbanded groups. They found that the apparent survival of banded male and female birds was 11–13% lower than their unbanded counterparts in 1996–2003. Studies of King Penguins patagonicus have also indicated that banding has a negative effect. Le Maho et al. (1993) found that banded birds arrived at their breeding sites later than unbanded birds, while Froget et al. (1998), in their study of single- and double-banded (on both flippers) King Penguins, found that double-banded birds commenced breeding later than single-banded birds and

20 had a lower recapture rate. Gauthier-Clerc et al. (2004) showed that survival rates of banded chicks were significantly lower than unbanded, transpon- dered chicks. Therefore, it is possible that banding significantly reduces the survival of Little Penguins, and that the survival rates of the general penguin population will be underestimated when data for banded penguins are analysed.

Population size Numerous studies have examined the effect of population pressure on seabirds. Ash- mole’s ‘halo effect’ (Ashmole, 1963) proposed that intense foraging by seabirds in areas surrounding a colony depleted the food supply, affected the breeding suc- cess, and hence regulated the size of the population. Other models for population regulation were: the ‘hungry horde’ model (Furness and Birkhead, 1984) which claimed that the size of the colony was inversely proportional to the numbers of birds from other colonies that shared the same foraging areas, and the ‘hinterland’ model (Cairns, 1989) which assumed that foraging areas of neighbouring colonies did not overlap, and sought to model the relationship between the foraging area (known as the hinterland) and the size of the colony. In their study of population regulation in Little Penguins in south-eastern Aus- tralia, Dann and Norman (2006) found some evidence in support of Ashmole’s ‘halo effect’. However, they found that the ‘hungry horde’ and ‘hinterland’ models did not apply. Dann and Norman (2006) reported a positive association between pop- ulation size and nesting area for colonies on small islands, so that the population size is limited by the available nesting area. However, these authors found that for larger islands such as Phillip Island, the population size of the colony is affected by food availability rather than available nesting area, since competition for nesting sites appears low or non-existent due to an abundance of unused sites.

2.1.4 Survival estimates for penguins

Several empirical estimates of survival rates for the Little Penguins of Phillip Island have been made using life-table analyses (Reilly and Cullen, 1979; Chambers, 1989; Dann and Cullen, 1990). Reilly and Cullen (1979) analysed ten years of data for Little Penguins banded at breeding age, and obtained an overall estimate of annual

21 adult survival of 85%. In their analyses of 20 years of penguin data, Dann and Cullen (1990) obtained a survival estimate of 31% from fledging to breeding age, and an average adult survival of 75%, with survival decreasing with age for penguins four years and older. Chambers (1989) obtained similar results to those of Dann and Cullen (1990), namely, a survival of 33% from fledging to age three years. However, Dann and Cullen (1990) acknowledged that their life-table analysis would “serve as a first approximation prior to a more complex analysis of the kind recommended by Anderson et al. (1985)”. A detailed mark-recapture-recovery analysis using the live recapture and dead recovery information, and producing age-specific survival estimates, has not previously been undertaken for Little Penguins. The survival estimates for other penguin species appear in Table 2.1. Several penguin species (such as the Pygoscelis papua and the ) have adult survival estimates that are comparable to that of the Little Penguins (Croxall and Rothery, 1995; Williams, 1995; Underhill et al., 1999). The Aptenodytes forsteri has the lowest first-year survival estimate of any other species (19%). However, in terms of population stability, this is com- pensated by a much higher annual adult survival (95%) (Mougin and van Beveren, 1979). However, it should be noted that many of these studies were based on ques- tionable estimation methodology (see explanation of some of the drawbacks of life tables in Section 2.2.2 on p.25).

22 Table 2.1: Estimates of annual adult and first-year survival of various penguin species.

Annual First year adult Species survival (%) Source survival (%) Emperor Penguin 95 19 Mougin and van Beveren (1979) Aptenodytes forsteri King Penguin 91-95 40-50 Weimerskirch et al. (1992) Aptenodytes patagonicus Ad´elie Penguin 89 51 Ainley (2002) Pygoscelis adeliae 86 n.a. Carrick (1972) Eudyptes schlegeli Yellow-eyed Penguin 86a-87b 48 Richdale (1957) antipodes Gentoo Penguin 75-85 59 Croxall and Rothery (1995) Pygoscelis papua 89 27-38 Williams (1995) African Penguin 85 n.a. Underhill et al. (1999) Spheniscus demersus

a refers to females and b to males

23 2.2 Recapture and recovery methods

2.2.1 Some definitions

The terms “capture”, “recapture” and “recovery” have very specific meanings in this thesis. Here capture refers to the initial live encounter with an animal, when it is marked with a tag that allows it to be identified later, while recapture denotes a subsequent live encounter with a previously marked animal, and recovery refers to finding a tag from a dead animal. Note that my distinction between recapture and recovery, although very useful, is not universal. In capture-recapture experiments, groups of animals are marked on a number of occasions, released and then recap- tured alive on later occasions. Mark-recovery studies involve the marking of groups of animals and the subsequent recovery of dead marked animals. (If an animal is found dead in the jth year of the study, it is assumed to have survived for j − 1 years and then died in the following year.) The mark-recapture-recovery scenario includes both live recapture and dead recovery information. A “closed population” is a population that is not changing through births, deaths, immigration or emigration for the period of the study, whereas in an “open population” the population size is allowed to change in the study period (see Seber, 1982, p.4).

2.2.2 An historical perspective

The estimation of demographic parameters has been of significant interest to popu- lation ecologists for a long time (see for example Lincoln, 1930; Bailey, 1951), with an individual’s survival and fecundity being fundamental components of studies of population demography (Lack, 1954; Eberhardt, 1985). Indeed, one of the earliest recorded capture-recapture studies appears in Laplace (1786), which estimated the human population of France (see Seber, 1982; Manly et al., 2005a). While fecundity may be measured readily from observations, the measurement of survival presents a significant difficulty for animal ecologists, since wild populations are rarely able to be “followed exhaustively” through time (Lebreton, 2001). In the early 20th cen- tury, bird banding was used to study the migration patterns of birds. By the late 1920’s, ornithologists such as Magee (1928) realized that recoveries of dead banded birds provided information on survival processes. Lack (1943) was the first to use

24 information on the recoveries of dead animals to estimate age-dependent survival probabilities. Most early recapture studies (for example, Schnabel, 1938) assumed that the populations were closed. Furthermore, these studies assumed that all animals had the same probability of being captured on each capture/recapture occasion. Since the closure assumption is really only appropriate for short-term studies, closed population models have very limited applications. The frequent failure of cap- ture/recapture experiments to meet these assumptions led to the development of more complicated models. Life-table analyses, which were at one time widely used in the analysis of re- covery data (see Seber, 1971; Burnham and Anderson, 1979), assume that survival depends only on age, that the recovery rates are constant, and that the fates of the individuals in the population are known. Since these assumptions are almost always invalid, life tables are well known to produce biased estimates of survival (Seber, 1986). Indeed, Burnham et al. (1985), in their paper reviewing the major problems in the use of the life-table analyses, refer to this method as an “invalid procedure” that “should not be used”. Lakhani and Newton (1983) and Anderson et al. (1985) also outlined some of the problems with life-table analyses. While early capture-recapture and mark-recovery analyses used ad-hoc methods, by the 1960s maximum likelihood estimation (MLE) was used to estimate the model parameters (e.g., Seber, 1962). Other methods include Bayesian techniques, which also use the likelihood func- tion. Early work in this field was conducted by Freeman (1990) who demonstrated the usefulness of Bayesian methods, but also highlighted the computation difficul- ties inherent in such an approach. These problems were addressed by using Markov chain Monte Carlo (MCMC) simulation methods (see Vounatsou and Smith, 1995). Bayesian techniques are particularly suitable for random effects and model averag- ing (see Brooks et al., 2000; Barry et al., 2003). Brooks et al. (2000) also provide a useful overview of Bayesian methods. Yet another method of analysis for recapture and recovery experiments is gen- eralised estimating equations, GEE, which extend the generalised linear models framework (Liang and Zeger, 1986), and the related martingale method which was

25 first used in capture-recapture studies by Becker (1984) and Yip (1989). Extensive work in this area has been completed by Huggins, Yip and Chao, primarily in the estimation of population size rather than survival (see for e.g., Yip, 1991; Huggins and Yip, 1999; Chao et al., 2001; Huggins, 2006). Useful reviews of these methods appear in Lloyd and Yip (1991), Schwarz and Seber (1999) and Chao (2001). How- ever, since I do not use Bayesian methods or estimating equations in my thesis, I will not discuss these approaches further. The original aim of capture-recapture studies was to estimate the size of the population being studied, with the survival estimates being regarded as “nuisance” parameters. However, by the 1980’s the emphasis of many studies had changed to the estimation of the survival probability. In fact, the estimation of the survival probability was widely regarded as being “critical to understanding animal popula- tion dynamics” (Lebreton et al., 1992). Since birds are generally able to be banded soon after birth, bird banding studies have the potential to provide information on age-dependent survival probabilities, in particular estimates of juvenile survival. Indeed, banding “remains a key source of information for the study of survival in bird populations” (Lebreton, 2001). In the past 25 years, there have been numerous excellent reviews of the state of research in this field, for example Seber (1982), North and Morgan (1985), Seber (1986), Pollock (1991), Anderson et al. (1993), Lebreton (1995), Schwarz and Seber (1999), and Lebreton (2001). In the following sections I will highlight the major achievements and the important issues in the analysis of recapture and recovery experiments.

Capture-recapture analyses Cormack (1964), Jolly (1965) and Seber (1965) provided ground-breaking work on the estimation of demographic traits, particularly survival rates, in animal popula- tions. These authors independently developed a theoretical framework for the anal- ysis of capture-recapture studies for open populations. The Cormack–Jolly–Seber (CJS) model used only recaptures of marked animals, and produced estimates of the survival and recapture probabilities. The recapture probability accounted for the fact that an animal that was alive at the capture/recapture occasion could be missed. However, the Jolly–Seber (JS) model used the ratios of marked to

26 unmarked animals and estimated the population size as well as the survival and recapture probabilities. The difference in the assumptions for the JS and CJS mod- els is as follows: the JS model assumes all animals are a random sample from the population, so that the probability of capturing a marked animal is the same as the probability of capturing an unmarked animal, whereas the CJS model is based only on recapture histories of marked animals (Manly et al., 2005b). While the work of Cormack, Jolly and Seber is still widely cited more than 40 years later, many researchers have extended the JS and CJS models, to make them more realistic and to reduce the restrictions placed on the parameters. Seber (1971) further developed his earlier work by examining age-dependent survival, but with constant recovery probability. Buckland (1980) extended the JS model to include information on dead recoveries. Another extension of the CJS model was the robust design proposed by Pollock (1982), and discussed in detail by Kendall et al. (1995, 1997). Pollock’s robust de- sign includes repeated sampling in several short “windows” of time (so that closure applies), with birth, death, immigration and emigration able to occur during the longer time intervals between one sampling window (capture/recapture occasion) and the next. The survival probability was estimated over the longer time interval. While the CJS model allows only time-dependent estimates for the survival and recapture probabilities, authors such as Pollock (1981), Clobert et al. (1987) and Pollock et al. (1990) expanded the suite of capture-recapture models, to incorporate age and time dependence in the recapture and survival rates. Burnham et al. (1987) provided a comprehensive explanation of the design and analysis of capture- recapture experiments. Lebreton et al. (1992) provided a “common framework” for capture-recapture studies, and further extended the existing models by examining variation in the recapture and survival probabilities with time, age, time since banding, time-specific covariates and categorical variables (such as sex). The widespread availability of more powerful computers and the advancements made in computer software used to estimate population parameters (see Section 2.2.5 on p.45), has led to the development of much more sophisticated models than those of the original JS and CJS models (Manly et al., 2005b). Various restrictions can now be placed on the parameters, and the parameters can be made to depend

27 on covariates, either individual covariates (that is, those relating to each individual sampled) or time-varying group covariates (relating to the population as a whole at the time of sampling). It is now possible to consider a large number of models for a particular data set (see section on AIC weights and model averaging on p.44).

Mark-recovery analyses In many capture-recapture studies of birds banded as chicks (particularly seabirds), survival estimates of young birds cannot be made, since they spend time away from the breeding areas and are rarely recaptured. Indeed, “live recaptures rarely — if ever — give access to survival in the early period of life, when birds are very mobile and rarely amenable to recapture in a restricted area” (Lebreton, 2001). Hence, the analysis of mark-recovery data for birds banded as chicks often provides the only means of estimating juvenile survival (Clobert and Lebreton, 1991). Furthermore, while recaptures provide estimates of “local survival which can be biased by per- manent emigration (confounded with mortality) or affected by local environmental effects”, recoveries provide information regarding “large scale survival” (Lebreton, 2001). While Seber (1962, 1970) began to develop mark-recovery models, Brownie and Robson (1976) and Brownie et al. (1978) pioneered the use of mark-recovery, by fitting an extensive sequence of models. However, these authors did not include age and time dependence in the survival probabilities. Brownie et al. (1985) extended the theory of mark-recovery by including age- and time-dependent survival and re- covery probabilities. They considered a simple age structure of young, sub-adult and adult. The inclusion of time dependence in the recovery and survival probabilities provided a much more realistic modelling framework, and enhanced the reliabil- ity of the parameter estimates (Francis and Saurola, 2002). The mark-recovery analysis was further developed by Freeman and Morgan (1992) who proposed a strategic approach for analysing recovery data for birds banded as young. These authors began with a global model which included time dependence in the recovery and first-year survival probabilities, and age dependence in the survival probabil- ity for older birds, before fitting simpler submodels. Another innovation was that they conducted a simulation to estimate the potential bias in the estimates of the survival rates caused by age dependence in the recovery probability.

28 Catchpole et al. (1995) developed models involving age and time dependence in their analysis of mark-recovery data for birds banded as young. In particular, they were concerned with age dependence in the recovery probability. Catchpole and Morgan (1996) outlined a model selection approach to the anal- ysis of mark-recovery data, by using score tests. They also proposed a succinct notation that could be used to denote a wide range of models. While age and time variation in the survival and recovery probabilities are often required in mark-recovery analyses, such modelling can result in parameter redun- dancy problems (see Section 2.2.3 on p.32).

Combining information from different sources: mark-recapture-recovery methods A well-known problem with the Cormack–Seber mark-recovery model for birds banded as young, is that an additional constraint must be applied in order for the maximum likelihood estimates to be unique (Cormack, 1970; Seber, 1971; Free- man et al., 1992). Since the use of such a constraint may produce misleading parameter estimates (Lakhani and Newton, 1983), many authors have sought to overcome this problem by supplementing mark-recovery data with information from other sources. For example, in order to avoid the use of artificial parameter con- straints, Freeman et al. (1992) and Lakhani (1987) combined the information from mark-recovery studies with additional field information, namely data from separate radio-tracking studies of bird marked as young, and Catchpole et al. (1993) com- bined mark-recovery data for birds banded as young with information from a single recapture occasion at the end of the study. In practice, mark-recovery studies often include some live recapture data, since samples of animals to be marked include some previously marked animals. Although a small number of live recaptures can perhaps be ignored in the mark-recovery analysis, if there are a large number of recaptures, these should be incorporated into the analysis. Prior to the development of a framework for mark-recapture-recovery analyses, when the same experiment yielded both live recapture and dead recovery infor- mation, capture-recapture and mark-recovery analyses were sometimes conducted separately, and then the results were compared (for example, see Anderson and Sterling, 1974).

29 Early research into the mark-recapture-recovery approach appears in Anderson and Sterling (1974), Buckland (1980), Mardekian and McDonald (1981) and Szym- czak and Rexstad (1991). Brownie et al. (1985) used live recapture information, in addition to mark-recovery data, to refine their survival estimates. Mardekian and McDonald (1981) supplemented the dead recovery data, by using live recapture information for the last recapture occasion. They then conducted a simultaneous analysis of the recapture and recovery data, by forcing the recapture data into the framework for mark-recovery studies. Lebreton et al. (1995) analysed independent capture-recapture and mark-recovery data sets on the same species (but not the same birds), using common parameters. These authors showed that the addition of independent live recapture data to a mark-recovery study improved the analysis, by providing information on adult sur- vival. Rather than combining independent samples of recoveries and recaptures, a more efficient approach is to collect dead recovery and live recapture information for the same individuals (Lebreton, 2001). By combining recapture and recovery information, researchers can increase the precision of their estimates, improve the robustness to assumption deviation, and avoid the restrictions placed on model choice due to problems with parameter redundancy (Catchpole and Morgan, 1997; Catchpole et al., 1998a; Freeman et al., 1992; Lebreton, 2001). The standard scenario for mark-recapture-recovery data is a short capture- recapture occasion with no deaths, births or migrations during this time, and with deaths and recoveries occurring between these occasions (Lebreton et al., 1992). A joint mark-recapture-recovery analysis, which analysed mark-recovery and capture-recapture data on the same birds simultaneously, was first considered by Burnham (1993). This author synthesised the two distinct areas of study: capture- recapture and mark-recovery, by generalising the CJS model for capture-recapture and Brownie’s model for mark-recovery. His joint analysis which focused on the pa- rameter in common, the survival probability, was based on the multinomial distri- bution. Burnham (1993) modelled time-dependent parameters, a rather restrictive modelling scenario that was unrealistic for many wild-life studies. Indeed, he ac- knowledged that a theoretical framework was needed to extend his analysis of joint

30 recapture-recovery data to age-dependent models. Such a framework was supplied by Barker (1997) and Catchpole et al. (1998a), enabling the construction of realistic age-dependent models for the survival, recapture and recovery probabilities. Barker (1997) extended the CJS model to incorporate data arising from three sources: capture-recapture, mark-recovery and mark-resighting (which can occur continually in between recapture occasions). He noted that the advantage of joint analyses is that “all sources of information can be used”, resulting in “more precise estimates” using smaller sample sizes, and that such a joint analysis is more efficient than analysing the data separately. He also allowed for random emigration or permanent emigration of animals from the study site. However age dependence in the model parameters was not considered here, since models involved only a single age class. Catchpole et al. (1998a) claimed that integrating capture-recapture and mark- recovery information for the same individuals resulted in “more realistic estimates of the survival probabilities of wild animals” than when separate recapture or re- covery studies were undertaken. These authors provided a general framework for a joint mark-recapture-recovery analysis, which allowed fully age- and time-dependent model parameters. They also provided a means of determining the relative contribu- tions to the analysis of the live recapture and dead recovery information. Another advantage of their approach was the significant improvement in the efficiency of their calculation of the likelihood. Rather than taking the product of the individual likelihood contributions for each animal, this method takes the product of the likeli- hood contributions for each cohort of animals over each capture/recapture occasion. Since data are collected from independent groups (ie. cohorts), the likelihood func- tion is the product of multinomial distributions. The likelihood is calculated using sufficient statistics which form a useful summary of the data. Further details of the methods and notation of Catchpole et al. (1998a) appear in Chapter 4, and an informal derivation of their likelihood appears in Appendix A.1. Of course, this approach precludes the use of individual covariates. Catchpole et al. (2004) allowed for this possibility, at the expense of computer efficiency, by essentially designating each animal as a separate cohort.

31 Catchpole et al. (1998a) and Francis and Saurola (2002) highlighted the benefits of using combined recapture and recovery data, by examining survival estimates for various subsets of data, that is, using live recapture data, dead recovery data and combined recapture and recovery data for birds banded as chicks, for those banded as adults, or for all birds. Francis and Saurola (2002) found that there were potential biases in the parameter estimates when they used either recapture or recovery data alone.

2.2.3 Issues in recapture and recovery studies

Parameter redundancy While it is often desirable to include time and age dependence in the survival and recovery probabilities in mark-recovery studies, such modelling is prone to problems of parameter redundancy (see Catchpole et al., 1996; Catchpole and Morgan, 1997; Catchpole et al., 1998b). A model is “parameter redundant” if it can be expressed in terms of a vector of smaller dimension than the parameter vector (Catchpole et al., 1996). If a model is parameter redundant, not all of the parameters can be estimated separately using the data, and the likelihood surface is flat. Morgan and Freeman (1989) showed that adding further parameters to the model (when the biology warrants it) could alleviate this problem by allowing them to “move away from the flat part of the surface”. Catchpole and Morgan (1991, 1994) studied the shape of the likelihood function in age-dependent mark-recovery models, while Catchpole et al. (1996) showed how to determine by inspection whether mark-recovery models were parameter redun- dant. Catchpole and Morgan (1997) extended this work by providing a means of determining which parameters, in parameter redundant models, can be estimated using the maximum likelihood method.

Covariates Once time dependence of the model parameters was considered, efforts were made to try to explain the temporal variation in terms of varying environmental conditions. Pollock (2002) provided the following rationale for modelling covariate dependence in capture-recapture studies:

32 There are two very important reasons to model covariates. First it en- ables more parsimonious parameterisations and thus precision of all pa- rameter estimates is increased. Second there may be inherent ecological importance in understanding the nature of the relationships between the parameters and specific environmental and individual animal variables.

In the late 1970’s, environmental variables were used as covariates for the first time. North and Morgan (1979), in a landmark paper, used weather data to model the survival of herons in a mark-recovery study. These authors were the first to incorporate environment variables as covariates directly into the likelihood via lo- gistic regression, paving the way for many other studies involving covariates, such as Clobert and Lebreton (1985), Clobert et al. (1987), Morgan and Freeman (1989), Lebreton et al. (1992), and Catchpole et al. (1999). North and Morgan (1979) hypothesized that first year survival would be likely to depend on the yearly vari- ation in the severity of the winter weather. Hence they could incorporate time variation, without the inclusion of a large number of parameters. Such a model is biologically-realistic, as animals in their first year of life are known to be particu- larly vulnerable to climatic variations. While North and Morgan (1979) assumed a separate constant second-year survival probability, a constant annual survival for older age groups and a constant recovery probability, Morgan and Freeman (1989) allowed the recovery probability to vary with time. Besbeas et al. (2002) combined mark-recovery and census data to estimate ani- mal abundance and productivity. These authors found that their combined analysis increased the significance of the dependence of survival on measures of winter sever- ity. Clobert and Lebreton (1985) applied a similar approach to North and Morgan (1979), but for capture-recapture studies (see also Clobert et al., 1987; Kanyamibwa et al., 1990). Pollock et al. (1984), Lebreton et al. (1992) and Pollock (2002) further developed the theory for the inclusion of covariates in capture-recapture studies.

Individual covariates While some intrinsic covariates are time-invariant, or are recorded only once for each individual, others are time-varying and vary from one mark/recapture occasion to

33 the next. Seber and Schwarz (2001) and Pollock (2002) highlighted the potential difficulties when modelling the dependence of the model parameters on individual covariates, particularly when the individual covariates are time varying. These authors also acknowledged the lack of publications in this area. Multi–state models, such as the Arnason–Schwarz model (Arnason, 1973; Schwarz et al., 1993), can deal with certain types of individual covariates, by allowing the model parameters to vary between animals that are in one of several states. The state could relate to a geographical location (such as a study site) or it could repre- sent the levels of a discrete covariate. Although multi–state models can incorporate time-varying covariates and individual covariates, this method cannot deal with continuous covariates (Bonner and Schwarz, 2006). However, Nichols et al. (1992) used a continuous time-varying individual covariate by approximating it a “discrete weight states”.

Missing covariate values For time-invariant individual covariates, that is covariates that are recorded once for each animal, animals that do not have valid covariates values are often disqualified. This is known as “complete case” (c.c.) analysis — see Little and Rubin (1987, Section 3.2) and Molenberghs and Verbeke (2005, Section 27.1). However, the problem of missing covariate values is of particular concern when using time-varying individual covariates, such as current body weight. In cases such as this, a c.c. analysis can result in an enormous amount of wasted data. Indeed, as Catchpole et al. (2008) point out, such a method of analysis can lead to “imprecise and/or strongly biased estimators”. An alternative approach would be to fill in missing values with the last previously known measurement, known as “last observation carried forward” (LOCF) — see Molenberghs and Verbeke (2005, Section 27.1). However, such a method may result in bias if the probability of being missed depends on the value of the covariate. For example, animals that are in poor condition in a particular season may not breed and hence may not be recaptured and weighed in that season. The computer package MARK (Cooch and White, 2006, Chap. 12, p.22) replaces missing covariate values from a particular recapture occasion with the mean covari- ate value for those measured on that occasion. This is known as “unconditional

34 mean imputation” (UMI) — see Little and Rubin (1987, Section 3.4.2) and Molen- berghs and Verbeke (2005, Section 27.3). However, Catchpole et al. (2008) pointed out that such an approach “shrinks individual differences, which is undesirable”. Both LOCF and UMI are imputation methods, which involve imputing the un- known values of the covariates. More sophisticated imputation models are possible. For example, Bonner and Schwarz (2004) dealt with missing values of time-varying individual covariates by modelling the distribution of the missing values conditional on the observed values, while Bonner and Schwarz (2006) developed a diffusion model. As an alternative to imputation, Catchpole et al. (2008) developed a three– state likelihood, which is equivalent to the standard likelihood when there are no missing covariate values. While a c.c. analysis uses no information from animals that have missing covariate values, the three–state method uses all of the available information for each animal, since the unknown conditional transition probabilities are simply deleted from the likelihood, while the known ones are retained (see further explanation in Section 6.2.2 on p.129).

Age dependence in model parameters Age dependence in survival rates in mark-recovery studies has been modelled for more than 35 years (Fordham and Cormack, 1970). While Brownie et al. (1985) claimed that “there is no valid way to estimate age-specific survival rates from only the banding of young” in mark-recovery studies, this point was disputed by Morgan and Freeman (1989), who successfully modelled age dependence in survival (see examples in Freeman and Morgan, 1990). Necessary and sufficient conditions for avoiding parameter redundancy when the survival and the recovery probabilities are both age-dependent in mark-recovery studies are given by Catchpole et al. (1996). Many studies have reported age-specific survival probabilities in the species stud- ied, and some have detected senescence (see p.15 in this chapter). Francis (1990) in a study of Lesser Snow Geese Anser caerulescens, found that first year birds and young adults had lower survival rates than experienced breeders, but found no evidence of senescence. Harris et al. (1994) found that for banded adult Eu- ropean shags, survival in the year after marking was significantly lower than the mean annual survival in later years. They also reported that the survival of shags

35 declined significantly with age, from about 14 years of age. Francis (1995) showed that the first-year and adult survival probabilities for Lesser Snow Geese were quite different, and that they also had differing patterns of time variation. In a 20-year study of Puffins Fratercula arctica, Harris et al. (1997) reported an age-specific sur- vival probability, with survival being lower for older birds. They pointed out that the time-dependence in survival observed in some studies, could in fact be due to “long-term shifts in colony age structure”. As mentioned previously, many earlier studies that reported senescence were based on invalid estimation methodology. Furthermore, even when valid estima- tion procedures were used, it is likely that some of the apparent age-related effects observed were spurious, since careful analysis is needed to detect confounding be- tween age- and time-related changes in the model parameters (as illustrated in Section 4.3.3 on p.75). The assumption that all animals in a population share common parameters, regardless of their age, is widely acknowledged to be extremely restrictive and un- realistic (e.g. see Pr´evot-Julliard et al., 1998; Barker, 1999; Francis and Saurola, 2002). However, while the importance of modelling age dependence in the survival probability is well known, researchers often assume that the recapture and recovery rates are constant or purely time-varying, rather than depending on age. If the lifestyles of young birds are different from their older counterparts, these assump- tions are likely to be invalid. Therefore, when biological considerations suggest that all of the model parameters should vary with age, simultaneous age dependence in the parameters should be considered. Indeed, failing to model age dependence in the parameters simultaneously is a serious flaw in many earlier studies in this field. Such a modelling scheme avoids the possibility of “leakage” between the model parameters, whereby an unmodelled age variation in one parameter can lead to erroneous estimates for the others (Catchpole et al., 1998a, 2004). In Section 4.3.1 (on p.71), I provide an illustration of potentially erroneous results that may arise when researchers fail to consider simultaneous age dependence. Fitting age-dependent models for the survival, recapture and recovery probabili- ties simultaneously in a mark-recapture-recovery study can result in difficulties. For example, Catchpole et al. (2004) were unable to fit simultaneous age dependence

36 for the survival, dispersal and recovery probabilities of red deer, finding that simul- taneous modelling resulted in extremely unstable parameter estimates. However, in Chapter 4, I have successfully applied such a modelling scheme and obtained biologically realistic age structures for the model parameters.

Tag loss While many studies assume that tags are permanently retained by the animals (see Seber, 1982, p.196), others warn that tag loss can negatively bias the survival estimates, unless tag retention probabilities are allowed for (Xiao, 1996; Bradshaw et al., 2000; Conn et al., 2004). Experiments involving double-tagged animals, which were pioneered by Beverton and Holt (1957), allow the estimation of tag loss (for example, Arnason and Mills, 1981; Wetherall, 1982; Johnson et al., 1995; Stobo and Horne, 1994). Such studies also enable researchers to determine the effect of tag loss on the survival estimates (Treble et al., 1993), and the effect of the tags on survival (see Froget et al., 1998; Petersen et al., 2006, and Section 2.1.3 on p.19 of this chapter). Arnason and Mills (1981) showed that under the assumption of homogeneous tag loss, whereby tag loss does not depend on age or size of the animal or the time of tag retention, the estimator of the population size under the JS model is unbiased. Pollock et al. (1990) claimed that tag loss or tag-induced mortality can negatively bias the survival estimates, but that the population size estimator would be unbiased. However, under more realistic tag-loss models, McDonald et al. (2003) showed that the JS estimator of population size was biased due to tag loss or tag-induced mortality. Cowen and Schwarz (2006) extended the JS model by incorporating tag retention probabilities that were assumed to depend on the release group and the time elapsed since marking. Xiao (1996) developed a model to estimate tag loss rates in double-tag experiments, by deriving the probabilities of tag retention under general conditions. Kremers (1988) incorporated tag loss in a joint analysis of mark-recovery and recapture/resighting data. One tag was assumed to be permanent, while the reten- tion rate of the second tag depended on time elapsed since marking. Nichols and Hines (1993) studied capture-recapture and resighting data, and modelled the tag loss rate as varying with time elapsed since marking. These authors pointed out

37 that the easily-observable tags used for resighting often have lower retention rates, so that recapture data were crucial to allow researchers to obtain valid estimates of tag loss. Bradshaw et al. (2000) studied the assumption of homogeneity of tag loss for New Zealand fur seal pups, and found that some individuals tended to lose their tags more readily than others, due to differences in their behaviours. These authors also found that the tag retention probabilities varied for different tag types. Catchpole (unpubl. data) also incorporated tag loss in the likelihood, in his analysis of tag loss in abalone for multiple tag locations. Details of this method appear in Section 7.3 on pp.161–165 of this thesis. While experiments studying tag loss were traditionally based on double-banded animals, Barrowman and Myers (1996) extended the analysis of double-tagged an- imals to include multiple tag types. Conn et al. (2004) and Alisauskas and Lindberg (2002) developed multi-state models for tag loss, with the presence or absence of a tag considered as a state variable, and the transition probabilities modelling tag loss. In order to reduce the problem of tag loss and the effect of tags on survival, McDonald et al. (2003) urged capture-recapture practitioners to handle animals carefully during captures, to use tags with high retention probabilities, to use double tags, and to ensure that the recapture probabilities are as high as possible.

Goodness-of-fit testing Pollock et al. (1985) and Burnham et al. (1987) pioneered goodness-of-fit testing for capture-recapture studies, while Brownie et al. (1985) described such tests for mark- recovery analyses. Freeman and North (1990) and Freeman and Morgan (1990) tested model fit in mark-recovery studies by examining patterns of residuals for individual cells of the recovery matrix. If the residuals have a pattern, then the model does not properly fit the data. If the global model is rejected, the patterns of residuals also help to determine whether the poor fit is due to the model structure or to assumption failure. Many goodness-of-fit tests involve the model deviance which is defined as

D = −2[log(L) − log(Lmax)],

38 where Lmax is the likelihood under the maximal model (McCullagh and Nelder, 1983, p.17). Although goodness of fit can be assessed “by comparing the deviance of the selected model with its asymptotic distribution”, such an approach is invalid when a study involves sparse data matrices (Catchpole et al., 1999). Catchpole et al. (1999), in their study of mark-recovery data for British Lap- wings, used a Monte Carlo test to assess goodness of fit. They generated 99 sim- ulated data sets from the selected model, using the parameter estimates obtained under the assumption that this model was correct. They then fit the model to each simulated data set, and calculated the deviance each time. This provided an “approximation to the distribution of the deviance if the selected model is correct”. Finally, they compared the deviance corresponding to the observed data, with the distribution of deviances from the simulated data. Such an approach can be used to produce an estimate of the variance inflation factor (see Section 4.4 on p.80 of this thesis).

Overdispersion If there is more variation in the data than predicted by the probability model, the underlying assumption of independence may not be valid, resulting in over-dispersed data (Burnham et al., 1987; Lebreton et al., 1992; Burnham and Anderson, 2002, p.67). One biological reason for overdispersion is that, for some species, breeding pairs or young from the same nest behave “almost as an individual”, and so exhibit a lack of independence in their responses, resulting in underestimated sampling variances (Eberhardt, 1978; Burnham and Anderson, 2002). More research is needed into goodness-of-fit testing and the assessment of overdispersion, particularly for mark-recapture-recovery analyses and for models incorporating covariates (Pollock, 2002). For further discussion of overdispersion and its consequences for model selection, see p.42 of this chapter.

2.2.4 Model selection

General principles Model selection from a set of candidate models has always been important. How- ever, the availability of powerful computers has dramatically increased the num-

39 ber of models that can be examined, making the quest for the most appropriate, biologically-sensible model more crucial than ever before (Manly et al., 2005b). Anderson and Burnham (1999) promote the following general strategy for model building and selection in mark-recapture-recovery analyses: A philosophy of thoughtful, science-based, a priori modelling is advo- cated. Biological science plays a lead role in this a priori model build- ing and careful consideration of the problem. The modelling and careful thinking about the problem are critical elements that have often received relatively little attention in ornithology. Instead, there has often been a rush to ‘get to the data analysis’ and begin to rummage through the ringing data and compute various quantities of interest. We advocate a deliberate, focused effort on a priori model building, as this tends to avoid ‘data dredging’ which leads to over-fitted models and finding spurious effects. Lebreton et al. (1992) claim that biological considerations should determine the set of candidate models on which the analysis is based. These authors state: The approach advocated here includes biologically driven model build- ing, model testing and selection, evaluation of model fit, and parameter estimation based on a “best” model. In deciding which models should be considered, the principle of parsimony should be adopted (Box and Jenkins, 1970). That is, the selected model should have the fewest number of parameters, but it should adequately explain the data, by accounting for the “major components of variation in the data” (Lebreton et al., 1992). Seber and Schwarz (2001) warn researchers against the “trend in increasing complexity” in models (which reduces the efficiency of the estimation) and suggest that models with fewer parameters should be considered. Burnham and Anderson (2002, p.18) point out that fitting a model that is too general may give “spurious results”, while an overly simple model may fail “to identify interesting real effects”. However, when a population has “complex dynamics” and there is a large quantity of data, the models selected tend to require a higher number of parameters (Lebreton et al., 1992).

40 Burnham and Anderson (2002) highlight some of the major issues to be con- sidered in model selection in their race car analogy (p.72). They liken identifying the fastest car/driver combination to finding the best model. The “data” are repre- sented by the results of a race. The biological justification that goes into choosing potential candidate models to be examined, is analogous to pre-race trials which determine the car/driver combinations that qualify for the race. While a particu- larly fast car may fail to qualify for the race, so too a “good” model may not appear in the list of candidate models. Hence, it is imperative that biological information on the species be used to limit the number of models to be considered.

Likelihood Ratio Test Historically, the Likelihood Ratio Test (LRT) was used to choose between several “biologically plausible” models (Seber and Schwarz, 2001). Other methods used to decide between candidate models were the Score test which begins with a basic model and indicates the direction in which to adjust this model, by studying the shape of the likelihood (see Catchpole and Morgan, 1996), and the Wald test which tests the statistical significance of an estimate, by comparing the ratio of the es- timate and its standard error with its asymptotic normal distribution (see Wald, 1943). The LRT can be used to compare “nested” models, models 1 and 2 say, where model 2 is a more general form of model 1. Let L1 and L2 be the maximum values of the likelihood functions under models 1 and 2 respectively, and assume that model 2 has v more parameters than model 1. To determine whether model 2 is an improvement on model 1, one can calculate the difference in the deviances, ∆D, where

∆D = 2[ln(L1) − ln(L2)].

If the additional parameters are in fact zero, then ∆D will have (asymptotically) a chi-squared distribution with v degrees of freedom. Therefore, if ∆D is sufficiently large, one can conclude that model 2 describes the data better than model 1.

Akaike’s Information Criterion Since the LRT only applies to groups of nested models, other criteria must be used to compare candidate models that are not nested. More recently, information-

41 theoretic approaches such as Akaike’s Information Criterion, commonly referred to as AIC (Akaike, 1973), have been widely used to select “the best approximating model” from a set of candidate models (Burnham and Anderson, 2002, p.62). Akaike believed that the most significant contribution of his information-theoretic approach was “the clarification of the importance of modelling and the need for substantial, prior information on the system being studied” (Burnham and Anderson, 2002, p.100). The AIC, which provides an estimate of the relative expected Kullback-Liebler distance between the “fitted model and the unknown true mechanism . . . that ac- tually generated the observed data”, is defined as follows:

AIC = −2 ln(L)+2K, where K is the number of estimable parameters in the model. The model with the lowest AIC value can then be considered as being the best, in terms of a compromise between lack of bias (many parameters) and lack of precision (few parameters) (Burnham and Anderson, 1992). Burnham and Anderson (2002, p.20) reject the notion of a “true model”, but instead endorse finding a “best approximating model” from a set of pre-determined candidate models that arise from the sound application of biological knowledge of the population being studied. This best approximating model and other models with similarly low AICs can then be examined further. Biological considerations should be taken into account, and the chosen model should have (close to) the lowest AIC.

Alternatives to AIC Under certain circumstances, the use of alternatives to the AIC, such as a Second- Order AIC (AICc) and the Quasi-AIC (QAIC) are recommended (Burnham and Anderson, 2002, pp.66–69). When there are insufficient observations n relative to the number of parame- ters in the model K (for example, when there are less than 40 observations per

42 parameter, ie. n/K < 40), the AICc should be used. This is defined as

2K(K + 1) AICc = AIC + . n − K − 1

The second-order bias correction term, 2K(K + 1)/(n − K − 1), is negligible when n is large relative to K. If there is more variation in the data than predicted by the probability model, the underlying assumption of independence may not be valid, resulting in over- dispersed data (see also p.39 of this chapter). Such overdispersion could artificially inflate the likelihoods, and exaggerate the differences in the AIC values. That is, when data are overdispersed, the AIC tends to select overfitted models (Anderson et al., 1994). Overdispersion can be allowed for by using QAIC, provided that there is a full model that fits the data well (Burnham and Anderson, 2002, p.70). QAIC is defined as −2 ln(L) QAIC = +2K, cˆ wherec ˆ is an estimate of the variance inflation factor, c, that is, the observed vari- ance over the variance under the assumed model. When there is no overdispersion, c = 1 and QAIC is equivalent to AIC. If an overdispersion factor is added, the number of degrees of freedom must be increased by one (Burnham and Anderson, 2002). The estimation of c is very much a work-in-progress (Cooch and White, 2006, Chap. 5, p.36). In fact, there are “no general, robust procedures” for estimating c (White and Burnham, 2002). One approach to calculatingc ˆ is to divide the deviance of the global model (assuming that it has a chi-square distribution) by the number of degrees of freedom in the model. However, if the deviance does not have a chi-squared distribution, Cooch and White (2006, Chap. 5, pp.21–24) suggest using a goodness-of-fit test (from program RELEASE, see Section 2.2.5 on p.45) or a Monte Carlo approach, based on simulation. I use their latter approach to calculatec ˆ for the penguin data in Chapter 4 (see p.80 of this thesis).

QAIC may also be modified to QAICc for small sample sizes. QAICc is defined as 2K(K + 1) QAIC = QAIC + . c n − K − 1

43 AIC Weights and Model Averaging Despite careful use of model selection criteria, there remains the possibility of bias in the model parameters, and of unrealistically low standard errors, by not allow- ing for the uncertainty inherent in the model selection process. Model averaging over several candidate models can give an indication of the extent of this problem (Buckland et al., 1997; Burnham and Anderson, 2002, p.150). However, care must be taken when the candidate models produce quite different estimates (Seber and Schwarz, 2001).

The Akaike weight, wi, of model i represents the “weight of evidence” in favour of model i being the best model out of a set of M candidate models (Buckland et al., 1997; Burnham and Anderson, 2002, p.75). Weight wi is given by

exp(−∆i/2) wi = M , r=1 exp(−∆r/2) P where ∆i = AICi − min(AIC), the difference between the AIC for model i and the minimum AIC of the candidate models. The Akaike weights can then be used to determine a model averaged parameter estimate based on a weighted average of the estimates from the candidate models

(Burnham and Anderson, 2002, p.75). The model averaged parameter estimate, θˆa, is M

θˆa = wiθˆi, i=1 X with the standard error given by

M 2 ˆ ˆ 2 σˆa = wi σˆi +(θi − θa) , i=1 X q where θˆi andσ ˆi are the parameter estimate and standard error corresponding to model i. The complexity of many long-term mark-recapture-recovery data sets means that many thousands of plausible models could be fitted. Regardless of the criterion used for model selection, one must be guided by biological as well as statistical reasoning. Manly et al. (2005a) recommend that the criteria should be used as

44 “useful guides” for model selection, and that “. . . it is not safe to assume that the mathematics alone will always guide you to the correct choice of models”. They wisely suggest that researchers consider the extent to which estimates obtained agree with known information and biological considerations, including the “biologist’s intuition”.

2.2.5 Computer software used in recapture and recovery analyses

I conclude this chapter with a brief overview of the available computer software for the analysis of recapture and recovery data. This overview is not intended to be comprehensive, since I have chosen to program the models considered in this thesis in Matlab, rather than being constrained by the limitations of any software. Useful reviews of computer software appear in Lebreton et al. (1992) and Schwarz and Seber (1999).

• ESTIMATE and BROWNIE (Brownie et al., 1985) both deal with mark- recovery models. While ESTIMATE allows only time dependence in the pa- rameters, BROWNIE also includes age-specific estimates. • RELEASE (Burnham et al., 1987) and SURGE (Pradel and Lebreton, 1991) allow the analysis of both recapture and recovery data. While RELEASE enables users to program simple models involving time-dependent parameters, SURGE incorporates more general models including age dependence. • JOLLY and JOLLYAGE (see Pollock et al., 1990) were developed to analyse capture-recapture models. JOLLYAGE extends JOLLY by incorporating age dependence in the parameters. • CAPTURE (Rexstad and Burnham, 1992) enables users to fit time-dependent capture-recapture models. • EAGLE (Catchpole, 1995) is a Matlab-based program for the analysis of mark- recovery data which, unusually, is based on score tests. • POPAN1 (Arnason and Schwarz, 1999) allows the analysis of mark-recapture data from open populations. While POPAN has many of the modelling fea- tures of MARK for the estimation of survival and recapture rates, in addition it permits the estimation of abundance and recruitment rates.

1http://www.cs.umanitoba.ca/∼popan

45 • MARK2 (White and Burnham, 1999) is a powerful, interactive computer pack- age that has the capabilities of SURGE, but in addition it can incorporate recovery data using the joint analysis of Barker (1997). A comprehensive user’s guide (Cooch and White, 2006) provides detailed instructions on the use of MARK, as well as some background material on the underlying statis- tical theory. • M–SURGE3 (Choquet, 2004) specifically deals with multistate capture-recapture models and handles a broader range of these models than other more general packages such as MARK. The companion package U–CARE checks the ap- propriateness of the data.

2http://www.warnercnr.colostate.edu/∼gwhite/mark/mark.htm 3http://ftp.cefe.cnrs.fr/biom/Soft-CR

46 Chapter 3

The Data

This chapter highlights the complexity of the long-term Little Pen- guin data set that has been analysed in this study. Since the penguin data form an invaluable record of these birds over almost 40 years, care- ful study is needed to ensure that I fully appreciate and understand the nuances of the data, and so that the discrepancies in the data can be identified and addressed. I begin by explaining the raw data that have been collected, as well as my reasons for disqualifying various groups of birds from the analysis. This is followed by an outline of the way in which the raw data are summarized into annual life histories for each individual bird, as well as an explanation of an alternative, albeit na¨ıve, approach to summarizing the data. Finally, as a brief overview of the data set, I calculate some simple statistics, such as the numbers of penguins recovered dead by age.

3.1 The Raw Data

The raw data for the Little Penguins in south-eastern Australia consist of 196 856 records, with each record corresponding to a live or dead encounter with a bird. Penguins that are banded and never again encountered (dead or alive), have a single record in the database, while other birds have multiple records, sometimes spanning many years. There is a total of 61 984 birds banded as chicks or adults in various locations, with live capture/recapture and dead recovery information extending from February 1968 to June 2004. The fields in the main raw database (which is in dBase format) appear in Table 3.1.

47 Table 3.1: Fields from main raw database. There is one row for each bird on each occasion it is encountered. Fields marked with a † are applicable only for live encounters. If there are any unbanded chicks in the burrow, measurements are taken for fields marked with a *.

DA Date(19941203means 3December 1994) DN Day/Night (“1” for day- and “2” for night-time encounter, “0” unknown) LOC Location code for live or dead encounters (refer to Appendix B.3.2). SITE† Site number (see Figure 1.2) BURR† Burrow number CONT† Burrow contents (eg. an entry of “102” means that the burrow contains one adult, no eggs and two chicks) BAND Band number TRANSP Transponder number for a bird fitted with a transponder (see Chapter 7). Blank entry for birds without transponders. N New=“1” for the original banding record of an individual, and “0” otherwise. BD† Bill depth WT† Weight HL† Head length SX Sex of bird (“3” for chick, “0” for unknown, “1” for male and “2” for female) MO† Stage of moult 1–5 METH Australian Bird and Bat Banding Scheme (ABBBS) method of encounter and status codes (refer to Appendix B.1). CHW1†∗ Chick weight CBL1†∗ Chick bill length CHL1†∗ Chick head length CHS1†∗ Chick stage 1–5 (whether ready to fledge) COM Comment such as experimental code (refer to Appendix B.2). EA Location code for oiled birds (refer to Appendix B.3.2).

As explained in Chapter 1, the records for the live captures and recaptures are entered after the Phillip Island Nature Park research staff or the Penguin Study Group visit the burrows, whereas the dead recovery records result from band re- coveries made by research staff, as well as by members of the general public who return bands to the ABBBS. If a bird has its band replaced, the bird gets a new band number, and all the records under the original band number are shifted to the new one. Birds that are ultimately recovered dead have a final record with a non-zero entry corresponding to a death code in the METH column. The ABBBS method of encounter and status codes are shown in Appendix B.1. For example, a METH code of “9805” means that the bird was “found dead in or near nest, band removed”. However, not all non-zero METH codes correspond to dead recoveries.

48 For example, a bird with a METH code of “813” was “trapped by hand or with a handheld net, released alive with band”, while “2916” means “burnt or scorched by fire, rehabilitated and released alive with band”. Therefore, I must carefully determine whether a record corresponds to a live recapture or a dead recovery, or whether the bird should be disqualified from the study. A sample of records for three birds, with band numbers 368, 691 and 6160, appears in Table 3.2. The penguin with band number 368 was banded as a chick (SX=3) at the Penguin Parade on Phillip Island (LOC=1) on 26 December 1969. It was subsequently recaptured alive once in the 1971/1972 breeding season and on three occasions in the 1974/1975 season, before being recovered dead after being taken by a wild mammal (METH=8405) on 21 May 1977. On 20 December 1969, bird 691 was banded as a chick at the Penguin Parade. It was recaptured on ten occasions, with the final live recapture occurring on 19 September 1987, and then it was never encountered dead or alive again. Bird 6160 was banded as a chick at the Cliffs to the West of Phillip Island (LOC=3) on 28 January 1975. It was found dead floating in the sea or beachwashed (METH=5405) less than one month later at Torquay Beach in Victoria (LOC=18).

3.2 Disqualification of birds from raw data set

Given the enormous amount of effort expended by the Nature Park staff and vol- unteers from the Penguin Study Group in their collection of the penguin data, my decision to disqualify birds from the study was not taken lightly. However, it was necessary to disqualify the following groups of birds, for reasons as outlined below:

3.2.1 Birds banded as adults

In the raw data set, there are 134 615 records, corresponding to 29 161 birds banded as adults. However, since it is not known how to determine the age of an adult Little Penguin (P. Dann, pers. comm.), and I am interested in variation in the model parameters as birds age, I only include birds banded as chicks in the main analysis. (However, I include the data for the birds banded as adults in my estimation of the adult age components of the model parameters in Section 4.6.) I determine the band numbers of birds banded as chicks by choosing records with N=1 and SX=3, before selecting all of the records corresponding to these band numbers. There are 62283

49 Table 3.2: Sample of records from main raw database

DA DN LOC SITE BURR CONT BANDTRANSPN BD WT HL SX MO METH CHW1CBL1CHL1CHS1COM EA 26/12/19690 1 30 0 1 368 1 0 0 030 0 00000NA 02/10/1971 1 1 30 0 100 368 0 0 0 0 1 0 0 0 0 0 0 0 NA 19/10/1974 1 1 30 0 100 368 0 0 0 0 1 0 0 0 0 0 0 0 NA 25/11/1974 1 1 30 0 100 368 0 151 1250 0 1 0 0 0 0 0 0 0 NA 14/12/1974 2 1 30 0 100 368 0 0 0 0 1 0 0 0 0 0 0 0 NA 21/05/1977 1 1 30 0 0 368 0 0 0 0 1 0 8405 0 0 0 0 0 NA 20/12/19690 1 30 0 1 691 1 0 0 030 0 00000NA

50 19/02/1972 1 1 30 0 100 691 0 0 0 0 2 2 0 0 0 0 0 0 NA 16/08/1976 1 1 30 0 200 691 0 0 0 0 2 0 0 0 0 0 0 0 NA 30/10/1976 1 1 30 0 100 691 0 0 0 0 2 0 0 0 0 0 0 0 NA 23/07/1977 2 1 30 0 200 691 0 0 0 0 2 0 0 0 0 0 0 0 NA 03/03/1979 1 1 30 0 100 691 0 0 0 0 2 1 0 0 0 0 0 0 NA 07/02/1980 2 1 30 999 100 691 0 0 0 0 2 0 0 0 0 0 0 0 NA 16/07/1983 1 1 32 0 100 691 0 0 9 0 2 0 0 0 0 0 0 0 NA 10/09/1983 1 1 32 0 100 691 0 0 1 0 2 0 0 0 0 0 0 0 NA 14/04/1984 1 1 32 0 100 691 0 0 9 0 2 0 0 0 0 0 0 0 NA 19/09/1987 1 1 32 0 100 691 0 0 900 0 2 0 0 0 0 0 0 0 NA 28/01/19750 3 26 0 1 6160 1 0 0 030 0 00000NA 23/02/1975 1 18 0 0 0 6160 0 0 0 0 0 0 5405 0 0 0 0 0 NA records corresponding to 32 804 birds banded as chicks from 1968 to 2004. The data selection was performed using the statistical software package R (R Development Core Team, 2006). The R code is available from [email protected].

3.2.2 Birds banded in locations other than Phillip Island

The captures/recaptures of live penguins have occurred reasonably consistently on Phillip Island from 1968 to the present day. In other locations, penguins have been banded from time to time and sometimes recaptured. However, in order for the recapture probability to be meaningful, I have disqualified the 5 239 birds that were banded as chicks in locations other than Phillip Island. The LOC (lo- cation) codes for Phillip Island are shown in Appendix B.3.1. Although the live captures/recaptures analysed in this study have been restricted to Phillip Island, the dead recoveries occur over a wide area — Appendix B.3.2 shows more than 300 location codes.

3.2.3 Experimental birds

To avoid biasing the natural mortality, a total of 3 319 penguins whose survival probabilities were thought to be affected significantly by human intervention were removed from the study. These birds were identified using their METH (method of encounter and status) or COM (comment) entries in the database. (Refer to Appendices B.1 and B.2 for lists of METH and COM codes.) The following groups of birds were disqualified from the study: • birds that were rehabilitated and released alive (Status code of 16), • birds alive in captivity in locations such as Melbourne Zoo (Status code of 19), • birds that were taken for scientific study (Method of encounter code of 63 or COM code of 111 or 122), • birds banded after death as part of a body dumping experiment (Method of encounter code of 53 or COM code of 110 or 210), • birds that were translocated, ie. moved to a different location to build up the colony in that area (COM code of 104 or 204), • birds that were given extra food (COM code of 008, 108, 208 or 308), and • those drenched to reduce parasitic infection (COM code of 09 or 209).

51 An alternative approach would have been to incorporate the capture/reapture data for some of these experimental birds up until the occasion on which they were treated, and then remove them from the study.

3.2.4 Birds that failed to fledge

Although chicks are banded as close to fledging as possible (mostly two weeks or less before fledging), many chicks die between banding and fledging. Since fledging is such a significant event in the life of a Little Penguin, the bird’s probability of surviving from fledging to around 12 months of age is much more meaningful biologically than its survival from banding to 12 months. Hence I disqualify birds that have failed to fledge. It is safe to assume that almost all of the birds that failed to fledge would be recovered dead somewhere near the nests (P. Dann, pers. comm.), and that almost all of the birds that were captured and never seen again did, in fact, fledge and go to sea. Therefore I identify a chick as having fledged or not based on its dead recovery record. Since burrow visits occur every two weeks on average during the breeding season, most birds would be recovered dead within a couple of weeks of their death, usually within about one month of banding. Once a bird has fledged it tends to go out to sea to feed for several months, possibly returning to its natal colony to moult at around 12 months of age. Hence a bird is said to have “failed to fledge” if it is recovered dead in the same location as it was banded, or “near the nest” (Method of encounter code 98), well before the commencement of the next breeding season. 603 birds were disqualified because they failed to fledge.

3.2.5 Birds from the final cohort and dead recoveries in last six months of study

The mark-recapture-recovery information extends from 8 February 1968 until 30 June 2004. However the cohort of 264 birds that were banded in the final year of the study are disqualified, since there are no recapture or recovery data for these birds in subsequent years. While all of the live recapture data are retained, the dead recovery information for the final six months of the study (1 January 2004 until 30 June 2004) are discarded, for reasons that are discussed in Section 3.3. Thus the present study comprises fledged chicks banded from the breeding season of 1967/1968 until the end of the 2002/2003 breeding season (36 cohorts, n=23 686),

52 live recapture data from 1 January 1968 to 30 June 2004, and dead recovery data from 1 January 1968 to 31 December 2003.

3.3 Summarizing the data

3.3.1 Rationale for using yearly summarized data

The present analysis is based on the yearly mark-recapture-recovery data which are summarized from the full raw data set using R. I analyse the yearly summarized data rather than conducting a continuous-time analysis, as I am interested in the annual variation in the model parameters, in particular the survival probability, rather than the variation in the parameters throughout the year. Very little information will be lost, because of the annual nature of the penguin’s life cycle. Young birds mainly die in the month or two following fledging, and adults die in autumn (following the moult) and in early spring (Dann et al., 1992). Therefore the survival probability would most likely depend on the time of the year, rather than purely on a bird’s varying age throughout the year. Indeed, even birds from the same cohort are not exactly the same age at a particular time of the year. Another advantage of using summarized data is that the raw monthly/fortnightly data would result in too many parameters. This is backed up by the results obtained for two colonies in New Zealand. Using a relatively narrow “window” of mark–recapture data for the month of January produced very similar results to those obtained when using summarized data as in this study (E. Johannesen, unpubl. data). I prefer to retain the full (summarized) data rather than use a narrow mark–recapture window and discard the remaining data. A similar approach was used by Catchpole et al. (2004).

3.3.2 Method of summarizing the data

Penguins initially captured as chicks are banded at around six to eight weeks of age (that is, slightly before fledging, which occurs at eight to ten weeks of age). Figure 3.1(a) indicates that the vast majority of chicks (82%) were banded in the summer months of December, January or February, with the initial capture occurring in late December on average, and fledging occurring soon after. The initial live captures for unbanded adults (of unknown age) occur at roughly the same time in the breeding season as the first recapture of banded adults (P. Dann, pers. comm.). The number

53 8000 2000

7000 1800

6000 1600

1400 5000 1200

4000 1000

3000 800

600 Number of birds banded as adults Number of birds banded as chicks 2000

400

1000 200

0 0 Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Month Month (a) Birds banded as chicks (b) Birds banded as adults of unknown age

Figure 3.1: Number of birds banded as chicks or adults by month. of adults banded in each month are displayed in Figure 3.1(b). Although the timing of adult banding is quite variable, the majority of adults (c. 71%) are captured initially from September to January, with the initial capture occurring in mid- November on average. Therefore, I define 1 January to be the nominal census date and birth date. When summarizing the raw monthly/fortnightly data into annual summary data, it is crucial that the parameters of interest retain their meaning. In the present analysis I am working with individual life histories for each penguin, sum- marized over each “penguin year” (penguin year tj is measured from 1 July in calendar year tj to 30 June in calendar year tj+1). Penguin years are used for the captures/recaptures since I wish to group together birds hatched or encoun- tered alive in the same breeding season, centred around 1 January. However, when summarizing the recovery information, I wish to consider the annual recovery prob- abilities in a bird’s first year of life, second year of life, and so on. If a bird is born on 1 January, then any recovery made in January to December of that year occurs in its first year of life. Hence I group recoveries by calendar year. Although this justification is rather simplistic, it provides a basic understanding of my method of summarizing the data. A more thorough explanation follows. Consider the monthly history entries for several birds in Table 3.3(a), where an entry of “0” (denoted by a dot in the table) means that the bird was not encoun- tered during that month, “1” refers to a live capture or recapture, and “2” means

54 Table 3.3: (a) Sample monthly history entries for birds #1 to #7; and (b) the corresponding annual histories. For clarity, a zero entry in (a) is denoted by a dot. (a) 1969 Penguin Year 1970 Penguin Year 1971 Penguin Year 1970 Calendar Year 1971 Calendar Year 1972 Calendar Year JASONDJFMAMJJASONDJFMAMJJASONDJFMAMJJASOND #1 ····· 1 1 · · ······ 1 1 ························ #2 ···· · 1 ············ 1 · 1 ····················· #3 ······ 1 · · 2 ································ #4 ···· 1 1 ········ 2 ··························· #5 ···· · 1 ············· 2 ······················

55 #6 ······ 1 ·················· 2 ················ #7 ···· 1 2 ····································

Penguin Year (b) 1969 1970 1971 1972 #1 1 1 0 0 #2 1 1 0 0 #3 1 2 0 0 #4 1 2 0 0 #5 1 0 2 0 #6 1 0 2 0 #7 1 2 0 0 that the bird was recovered dead. Since all seven birds in Table 3.3(a) were banded in the 1969 breeding season (with nominal census date 1 January 1970), each has a summarized yearly entry of “1” in Table 3.3(b), corresponding to penguin year 1969. Birds #1 and #2 survived their first year of life and were recaptured at ap- proximately one year of age in the 1970 penguin year. These birds have summarized entries of “1 1” corresponding to penguin years 1969 and 1970 (see Table 3.3(b)). If calendar year had been used for recaptures, the entry for bird #1 would have been “1 0”, which would have been misleading, since this bird was encountered alive in the breeding season following banding and was known to have survived its first year of life. Although bird #3 was recovered dead in the 1969 penguin year, and bird #4 was found dead in the 1970 penguin year, neither survived its first year of life. Therefore the monthly history entries for birds #3 and #4 are summarized in the yearly data as “1 2”. Birds #5 and #6 were recovered dead in penguin years 1970 and 1971 respectively, so it might be thought that the summarized entries should be “1 2” for #5 and“1 0 2” for #6. But both have summarized entries of “102”, since they survived their first year of life but failed to survive their second year.

Hence, the annual summarized entry for year tj for any bird is:

1, if the bird is initially banded or recaptured alive during penguin

year tj, and not recovered dead in calendar year tj

2, if it was recovered dead in calendar year tj 0, otherwise

The only exception occurs for birds such as #7, which was banded and recovered dead immediately in the first half of penguin year 1969. Since its initial capture occurred in penguin year 1969 and its dead recovery was in the 1969 calendar year, the above summary method would produce history entries of “. . . 020 . . .” for bird #7. However this bird did not survive its first year of life, so its summarized yearly entry must be of the form “1 2”. Therefore the dead recovery from calendar year 1969 must be assigned to penguin year 1970 in this case. Due to the manner in which the live recaptures and dead recoveries are summarized, I must discard the dead recovery data from 1 January 2004 to 30 June 2004, since I do not have a full 12 months of recovery information for the 2004 calendar year.

56 The implications of this method of coding are discussed further in Section 4.2.2, while Section 3.4 explains the shortfalls of a seemingly “obvious” alternative ap- proach to summarizing the data, that is, summarizing captures, recaptures and recoveries all by penguin year.

3.4 An alternative method of summarizing the data

As noted in Section 3.3.2, careful consideration needed to be given to summarizing the data from the raw monthly/ fortnightly records and to creating the yearly life histories for each bird. This resulted in the capture/recapture records being summarized by penguin year, but the dead recoveries being summarized by calendar year. A na¨ıve approach (and my original method) is to instead summarize both the recapture and recovery information by penguin year. In this section, I explain the consequences of such an approach on the estimation and the interpretation of the model parameters (especially the survival probability), and highlight the importance of thinking carefully about issues such as these before rushing into the analysis of the data (a point emphasised by Burnham and Anderson, 1998, p.2). Consider the monthly history entries for bird #5 in Table 3.3 (see p.55). Using the data summary methods from this thesis, bird #5 has annual history entries of “1 0 2 0”, since it was captured in the 1969 penguin year and recovered dead in the 1971 calendar year. These history entries correctly suggest that the bird survived its first year of life, but failed to survive its second year of life. However, under the na¨ıve approach, the annual history entries for bird #5 would be “1200” (since it was recovered dead in the 1970 penguin year), which incorrectly suggests that the bird did not survive its first year of life. Indeed, a bird banded on 1 January 1970 and recovered dead anytime up until 30 June 1971 would produce history entries of “1 2 0 0” under the na¨ıve approach. Therefore, using this na¨ıve approach, the first-year survival probability actually represents a bird’s survival up to the age of 18 months. In contrast, under my current method of summarizing the data,

φ1 correctly represents the survival probability in the first 12 months of life (see explanation in Section 3.3.2). Bird #1 in Table 3.3 has history entries of “1 1 0 0” under both methods of data summary. Consider now any bird with such a history. If the bird was banded on

57 1 January 1970, it survived at least 6 months (or possibly as long as 18 months), but it was most likely recaptured at around 12 months of age, and so the survival probability for the first year of life is needed in order to calculate the likelihood contribution for such birds. Therefore, the na¨ıve method of data summary outlined in this section is con- ceptually flawed and is inferior to my current approach (outlined in Section 3.3), due to the inconsistencies in the definitions of the “first year” survival probability. Section 4.8 contains a comparison of the parameter estimates obtained using these two methods of summarizing the data.

3.5 Breeding data

In addition to the life history records, breeding data have been collected for each study site in each breeding season since the commencement of the study. The breeding data for Study Site 1 in the 1978 penguin year appears in Table 3.4. Each record, which corresponds to a clutch of eggs laid in the site in that breeding season, includes the following information: a burrow number(Burr), a grid number corresponding to the location of the burrow on the map of the site (Grid No), the band numbers of the mother (Female) and father (Male) (if known), the number of eggs in the clutch (Eggs), the month in which the clutch was laid (Date), the number of chicks hatched (H), the number of chicks fledged (F), the weights of the chicks (Wt1 and Wt2), the band numbers of the chicks (Band1 and Band2), and the Site and Location numbers (Site and LOC). For example, birds 7088 and 7060 laid a clutch of two eggs in burrow 9026 in October 1978. Although both chicks hatched, only one survived to fledging. This chick had a weight of 900 grams and was banded with band number 8668. The 19 clutches laid in the main study site in the 1978 penguin year were laid by 15 breeding pairs, with 4 pairs laying a second clutch in the season. The parents of the clutch laid in burrow 9076 were unknown. These breeding data are used in the “strangers” analysis in Section 4.7. Note also that bird 4144 bred with bird 7094 in burrow 9061 in the early part of the breeding season, before changing partners later in the season. Bird 7094 was encountered alive for the last time in October 1978. Although the fate of this bird is unknown, it is possible that it died, and that 4144 found a new partner by November 1978.

58 Table 3.4: Sample of records from breeding data: Study Site 1 1978/1979

Burr Grid No Female Male Eggs Date H F Wt1 Wt2 Band1 Band2 B Site LOC 9009 B1TR 3996 7083 2 Nov-78 21 011 9019 C3CR 7066 7065 2 Aug-78 2 2 875 975 9565 9566 2 1 1 9026 7088 7060 2 Oct-78 2 1 900 8668 1 1 1 9033 E4BC 11532 3942 2 Sep-78 2 2 9570 1 1 1 9052 A1TR 2990 7027 2 Sep-78 22 011 9061 7094 41442May-7820 011 9061 7094 41441Sep-7800 011 9061 9564 4144 2 Nov-78 2 2 9662 9663 2 1 1 9070 A7TR 11531 12550 2 Nov-78 2 0 0 1 1 9076H10 2Sep-7800 011 9096 C2 7028 115332Aug-7820 011 9096 C2 7028 115332Nov-7820 011 9099 D1TC 7082 7037 2 Jun-78 20 011 9099 D1TC 7082 7037 2 Sep-78 2 2 950 9568 9569 2 1 1 9105 E1 6055 7087 2Sep-7800 011 9110 F7CL 2488 7023 2 Sep-78 20 011 9111 F11BR 951 6830 2 Nov-78 2 2 800 850 9665 9666 2 1 1 9119 A4TC 2992 4132 2Aug-7800 011 9119 A4TC 2992 4132 2 Nov-78 1 1 1200 9664 1 1 1

3.6 Some simple statistics

The numbers of birds banded as chicks in each breeding season varied considerably throughout the years of the study, from 17 chicks in the 1967 penguin year to 2375 birds in 1993 (see Figure 3.2). Of the 23686 penguins banded as chicks, 87% (20 707) were never encountered (dead or alive) after banding. The remaining 2 979 birds were encountered in subsequent years, with 1 788 birds being recaptured at least once (total of 4 368 live recaptures), and 1 347 being ultimately recovered dead. Hence there were only 156 birds that were recaptured and subsequently recovered dead. The numbers of birds recovered dead in each age group appear in Figure 3.3, while Figure 3.4 displays the numbers of birds recaptured alive at each age level. Many penguins in this study have one or more yearly history entries of zero in between their first and last encounters. The zeros that appear between the first and last encounters have a different interpretation than those occurring after the final encounter. In the former case, I know that the bird is still alive, whereas in the latter case, the bird is of unknown fate. There are many possible reasons for a live

59 2200

2000

1800

1600

1400

1200

1000 Size of Cohort

800

600

400

200

1970 1975 1980 1985 1990 1995 2000 Penguin Year

Figure 3.2: Number of penguins banded as chicks in each year of the study.

800

700

600

500

400

Dead recoveries 300

200

100

0 2 4 6 8 10 12 14 16 18 20 Year of Life

Figure 3.3: Number of penguins banded as chicks and recovered dead by age.

60 600

500

400

300 Live recaptures

200

100

0 2 4 6 8 10 12 14 16 18 20 22 Age

Figure 3.4: Number of live recaptures by age for penguins banded as chicks. bird not being encountered: newly fledged birds go out to sea to feed and may not return to the colony until they commence breeding at the age of two or three years; breeders may fail to breed in a season if they are in poor condition; birds may move outside of the study sites for one or more years before returning to the site in a later season; or birds may be breeding in the site but not be encountered during the day-time burrow visits. Of the 2 979 penguins that were encountered again after banding, 1 960 had at least one “0” entry in their history records between their first and final encounters. That is, 1 960 banded penguins were alive but not recaptured on at least one penguin year. The numbers of zero entries between the first and last encounters for each bird range from zero to 19 (Figure 3.5). While the vast majority of birds in the study (20 707) are banded and never seen again, a small number have been encountered for 20 years or more. Figure 3.6 displays the number of years elapsed from banding to the final dead or live encounter (inclusive) for each bird included in the study.

61 500

450

400

350

300

250

Number of birds 200

150

100

50

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Number of zeros in history entries

Figure 3.5: Number of birds with 1, 2,..., 19 zero entries between their first and last encounter.

1000

900

800

700

600

500

400 Number of birds

300

200

100

0 2 4 6 8 10 12 14 16 18 20 22 Number of years of history entries

Figure 3.6: Number of birds with life histories spanning 2, 3,..., 23 years.

62 Chapter 4

Modelling Age Dependence

This chapter models age dependence simultaneously in the survival, recapture and recovery probabilities for a population of Little Penguins, using mark-recapture-recovery data collected over a 36-year period. The successful modelling of simultaneous age dependence is very rare (see p.36). Here I present a clear methodology for developing a demographic age structure for any animal using mark-recapture-recovery data. I also provide illustrations of potentially erroneous results that may arise when researchers fail to consider simultaneous age dependence, or fail to de- tect confounding between age- and time-related changes in the model parameters. The results from this chapter form a manuscript entitled “Mark- recapture-recovery modelling and age-related survival in Little Penguins Eudyptula minor” (Sidhu et al., 2007), which has been accepted for publication in The Auk and will appear in the October 2007 issue.

4.1 Introduction

The primary aim of this chapter is to estimate the annual survival of Little Penguins in relation to age. However, there is every reason to expect age variation in the recapture and recovery probabilities also. Previous studies of Little Penguins at Phillip Island (Reilly and Cullen, 1981, 1982; Dann and Cullen, 1990; Dann et al., 1992) have highlighted the differing lifestyles of these birds as they age. Once young birds fledge, at around eight weeks of age, they go to sea and spend most of their time away from the colony (Reilly and Cullen, 1981), until they begin to breed at around two or three years of age (50% of two-year-olds breed, Dann and

63 Cullen, 1990). Consequently I expect different demographic parameter values for juveniles in their first year of life, for pre-breeders and for breeders. I have chosen to group birds by age rather than breeding status because breeding status was not always known. Once penguins commence breeding they return to the same colony to breed most years (Reilly and Cullen, 1981), and so their recapture probability should be higher than that of the juveniles. Since established breeders produce more chicks than younger breeders (Dann and Cullen, 1990), and consequently would be expected to visit the colony more frequently to feed their chicks, recapture rates of adult birds may increase with age. Another potential cause of age dependence in the parameters is senescence, whereby the survival probability reduces as adults age (see Nisbet, 2001, for senescence in long-lived seabirds). The recapture probability might also decrease with very advanced age, since senescent birds would perhaps be more likely to skip breeding for a year. Radio-tracking studies of adult penguins at Phillip Island (Weavers, 1992; Collins et al., 1999) have shown that the recoveries of dead birds reflect the movements and distribution of the live birds (Dann et al., 1992). The distribution of recoveries for various regions of the Victorian and South Australian coastline, as shown in Figure 4.1, are quite different for juveniles in their first year of life, and for penguins of breeding age. (See also Figure 1.1 on p.2 for the location of Phillip Island). Dann et al. (1992) suggest that the differences in the recovery locations for the various age groups are due to pre-breeders using more distant feeding grounds (e.g. upwellings off the coast of western Victoria or South Australia) as they do not need to return to the colony (except to moult). In contrast, breeding adults use less profitable feeding grounds that are closer to the colony, since they need to make frequent trips back to the colony for the renovation of burrows, pair maintenance, and the incubation and feeding of chicks (Collins et al., 1999). This description is supported by Figure 4.1: while the recovery locations of juveniles are widely dispersed, those of the older birds are concentrated on Phillip Island and in Port Phillip Bay, a highly populated region where members of the public are more likely to find dead penguins washed ashore. As mentioned in Chapter 3, I initially analyse the data for birds banded as chicks only, since I wish to investigate age dependence in the model parameters and birds

64 100 0 to 1 years

50

0

100 3 years and over

50

0 Ka 1 2 3 4 5 6 7 8 9 10 11PPBPh Fr 12 13 14 15 Ki Tas

Figure 4.1: The numbers of recoveries of juveniles (aged 0–1 years) and adults (aged three years and over) for penguins banded as chicks on Phillip Island. Each region represents a 100 km section of the Victorian or the South Australian coastline. The recoveries from Port Phillip Bay (PPB), Phillip Island (Ph), (Fr), (Ki), Kangaroo Island (Ka) and mainland (Tas) are considered separately. banded as adults are of unknown age. Parameter estimates arising from analyses of other subsets of the complete data set appear in Section 4.6.

4.2 Methods

4.2.1 Mark-recapture-recovery data

As mentioned earlier, “mark” refers to the initial banding of a bird, “recapture” denotes a subsequent live encounter, and “recovery” refers to the recovery of a band from a dead bird. On each burrow visit, any unbanded adults, or chicks that were likely to fledge in the next two weeks, were marked with individually numbered flipper bands; also the band number of any previously marked bird was recorded. Data were collected at a number of breeding sites on the Summerland Peninsula on

65 Phillip Island. The number and location of study sites visited, and the frequency of visits, varied over the years, resulting in a variable search effort across time. Visits to burrows in the initial study area varied from weekly (1968–1970) to monthly (1971– 1984) to two-weekly during the breeding season and monthly outside the breeding season (1985–2004). Additional study sites were added in 1981 and 1984, and were visited monthly until 1999, and then at two-weekly intervals during breeding periods and infrequently at other times. Recoveries came from the breeding areas on Phillip Island and along the coastline of southern Australia, as far west as Kangaroo Island in South Australia (see Figure 4.1).

4.2.2 Assumptions

The standard scenario for mark-recapture-recovery data is a short capture-recapture occasion with no deaths, births or migrations during this time, and with deaths and recoveries occurring between these occasions (Lebreton et al., 1992). Although the captures and recaptures of Little Penguins on Phillip Island extend over the whole breeding season, the above scenario is appropriate here since: (i) birds are mostly seen during the peak breeding season, (ii) most of the deaths occur once penguins have gone to sea for extended periods (a few months after fledging for the chicks and in late winter/early spring for the adults, Dann et al., 1992), and (iii) migration between breeding colonies during or between the breeding seasons is rare (Dann, 1992). Standard analyses of yearly recapture/recovery data assume that if a bird dies sometime between the mark-recapture occasions in penguin year j and penguin year j + 1, it will be recovered dead in that period or not be recovered at all. The ABBBS Methods of Encounter and Status codes (listed in Appendix B.1) suggest that this assumption may not always be valid, for example when the band only was found (code 42**), when a partially decomposed body was found, when a skeleton or dried out corpse was found with a band (codes **30, **32), or when a leg (or flipper) and band only was found (code 58**). Whereas dead bodies washed up on the beach usually decay quite rapidly, those in the breeding sites may last for up to six months (M. Cullen, pers. comm.). Since there are only around 60 bodies found as skeletons or dried out corpses (out of a total of around 2300 recoveries for

66 all birds in the data base), my assumption of recovery in the same year as death seems a reasonable approximation of the true situation. The alternative would be to delete these birds, and not to assign a date of death. It should be noted that failing to encounter a penguin in a particular year could merely mean that the bird has moved a few metres outside the study site. However, researchers attempt to address this problem by redefining the study sites at the beginning of each breeding season. That is, if they observe a burrow just outside the study site boundaries, they adjust the boundaries slightly as necessary to keep the number of burrows approximately constant. Despite this, small-scale emigration remains a cause for concern. A more detailed discussion of this issue and its potential effect on the survival estimates appears in Section 4.5 “Apparent survival”.

4.2.3 Parameters

The parameters of interest in this study are the survival, recapture and recovery probabilities. The following notation and definitions were used by Catchpole et al. (1998a).

Let t1,...,tk denote the k mark-recapture occasions, with cohort c (c =1,...,C) consisting of all birds banded at tc.

Further, let φc,j = P(a bird from cohort c, alive at tj, survives until tj+1), and

λc,j = P(a bird from cohort c, which dies in (tj, tj+1), has its death reported), for c =1,...,C and j = c,...,k − 1, while

pc,j = P(a bird from cohort c, alive at tj+1, is recaptured at tj+1) for j = c +1,...,k.

The definition of pc,j is slightly different from that of Catchpole et al. (1998a), in order to remain consistent throughout the thesis. In this study, the recapture probability for a breeder in a given year essentially represents the probability that the bird is breeding in a study site in that year, since a breeding bird will almost always be encountered at some stage in the season (P. Dann, pers. comm.).

67 4.2.4 Sufficient statistics

One of the advantages of the approach developed by Catchpole et al. (1998a) is the significant improvement in the efficiency of their calculation of the likelihood. Rather than taking the product of the individual likelihood contributions for each animal, this method takes the product of the likelihood contributions for each cohort of animals over each mark-recapture occasion. The likelihood is calculated using sufficient statistics, D, V, W, Z, which form a useful summary of the data. These upper triangular matrices are defined as follows:

dc,j = number of animals from cohort c recovered dead in the interval (tj, tj+1), 1 ≤ c ≤ C, c ≤ j ≤ k − 1.

vc,j = number of animals from cohort c marked or recaptured at tj and not seen again during the course of the study, 1 ≤ c ≤ C, c ≤ j ≤ k (with j = c corresponding to the initial banding). (Assumptions: There are no deaths on recapture and recaptured animals are returned immediately to the population.)

wc,j = number of animals from cohort c recaptured at tj+1, 1 ≤ c ≤ C, c ≤ j ≤ k − 1.

zc,j = number of animals from cohort c not recaptured at tj+1 but encountered later, either dead or alive, 1 ≤ c ≤ C, c ≤ j ≤ k − 1. The sufficient statistics corresponding to the Little Penguin data appear in Ap- pendix A.3. Note that the sufficient statistics were named in the following way: V for ‘vanished’, Z for ‘zeros’, W for ‘ones’, and D for ‘dead’ (E. Catchpole, pers. comm.).

Catchpole et al. (1998a) next define αc,j, the probability of an animal from cohort c surviving until tj, as

1, j = c, αc,j =  j−1 φ , c +1 ≤ j ≤ k,  s=c c,s  Q and χc,j as

χc,j = P(a bird from cohort c, alive at tj, is not seen dead or alive after tj).

68 The recursion for χc,j is derived as follows. Note that the following probabilities are conditional on the animal being from cohort c and being alive at tj.

χc,j = P(not seen dead or alive after tj)

= P(dies in (tj, tj+1) and not reported dead)

+ P(survives till tj+1, not recaptured at tj+1, not seen dead or alive after tj+1)

= (1 − φc,j)(1 − λc,j)+ φc,j(1 − pc,j)χc,j+1, for 1 ≤ c ≤ C, c ≤ j ≤ k − 1. All values of χ can be calculated recursively since

χc,k = P(alive in final year of study, not seen dead or alive thereafter) = 1 for 1 ≤ c ≤ C.

4.2.5 The likelihood

I use the maximum likelihood method to estimate the model parameters, and form the likelihood by using the techniques developed by Burnham (1993) and Catch- pole et al. (1998a). Maximum likelihood fitting was performed via the MATLAB package. Code is available from [email protected]. I use the following form of the likelihood equation, as developed by Catchpole et al. (2000):

C k−1 k dc,j wc,j +zc,j wc,j zc,j vc,j L = const × {(1 − φc,j)λc,j} φc,j pc,j (1 − pc,j) χc,j . c=1 "j=c j=c # Y Y Y (4.1) An informal derivation of the likelihood equation and a detailed explanation of the MATLAB programs used for this analysis appear in Appendices A.1 and A.2.

4.2.6 Fitting the models

I am primarily concerned with the annual survival probability and how this impor- tant parameter varies with age. However, I must also construct appropriate models for the recapture and recovery probabilities, since an inappropriate choice for one parameter can lead to erroneous estimates for the others (see Section 4.3.1 and Catchpole et al., 1998a).

69 Fitting age-dependent models for all parameters simultaneously in a mark- recapture-recovery study can result in difficulties. For example, Catchpole et al. (2004) found that simultaneous modelling resulted in extremely unstable estimates (see p.36 for further details). For this reason, I begin by modelling age dependence in one parameter at a time. Next I fit a model with full age dependence (as far as this is reasonable) in all parameters. I then fit an extensive sequence of models, with the age variation in each parameter being gradually simplified, whilst seeking to eliminate the possibility of confounding between age and time for the survival and the recapture probabilities. Since Section 4.3 focuses on age dependence of the parameters, I now use specific notation for age dependence rather than the general notation introduced above.

th The notation φj denotes a penguin’s survival probability in its j year of life. The recovery probability λj is defined similarly. The recapture probability pj represents the probability of recapturing the bird when it is j years of age. Notation such as

φj:l denotes the constraint, φj = φj+1 = . . . = φl, while φj+ denotes φj = φj+1 =

. . . = φ36, since the maximum possible age in this study (in theory) is 36 years.

4.2.7 Model selection

Models are selected by comparing Akaike Information Criterion (AIC) values (Akaike, 1973). The AIC represents a compromise between lack of bias (many parameters) and lack of precision (few parameters) (Burnham and Anderson, 1992). The list of models examined, together with their numbers of estimable parameters and AIC values, appear in Table 4.1. Under certain circumstances, Burnham and Anderson (2002, pp.66–69) recom- mend using alternatives to the AIC, such as the Second-Order AIC, denoted by AICc, or the Quasi-AIC, QAIC. AICc should be used when there are insufficient observations relative to the number of parameters in the model, which is not the case in this study. However, since these data are based, at least in part, on clus- tered individuals (e.g. fledglings coming from the same nest), the assumption of independence may not be valid, resulting in over-dispersed data. Such overdisper- sion could artificially inflate the likelihoods, and exaggerate the differences in the AIC values. Overdispersion can be allowed for by using QAIC, provided that there

70 is a full model that fits the data well. For further details see Section 4.4 on p.80 of this chapter.

4.3 Results

4.3.1 Leakage between parameters

The model with full age dependence for survival (that is, separate survival proba- bilities for each age) and constant recapture and recovery probabilities (model 4.3 in Table 4.1) has extremely wide confidence intervals for the survival estimates of the birds older than 12 years, suggesting that I have insufficient data to support the fully age-dependent structure. I therefore constrain the annual survival proba- bilities to be equal for the 13th and subsequent years of life. For each parameter in turn, I fit this age dependence whilst keeping the remaining two parameters con- stant (models 4.4–4.6 in Table 4.1, and Figure 4.2). I next consider a model with simultaneous age dependence in all parameters (model 4.7 in Table 4.1, and Fig- ure 4.3). A simultaneous model minimises the possibility of “leakage”, whereby an unmodelled age variation in one parameter may cause an apparent age dependence in another (Catchpole et al., 1998a, 2004). The considerable differences between Figures 4.2 and 4.3 provide an illustration of such leakage. For example, in Figure 4.2, models 4.5 and 4.6, which assume that survival is constant across age groups, estimate the survival probabilities as 61% and 65% respectively. These are gross overestimates of first-year survival, and lead to an unrealistically low prediction for the number of mortalities. Therefore model 4.6 greatly overestimates the probabil- ity of a dead first year bird being recovered (λ1). Similarly there is an unrealistically high predicted number of birds available for recapture at age one, leading to un- derestimation in the recapture probability p1 in model 4.5. Substantial leakage also occurs at older ages (see Figures 4.2 and 4.3), although to a lesser extent than in the first age group. Some of the parameters for model 4.7 are highly correlated. In particular there is a strong negative correlation between first and second year survival (r = −0.82). Strong positive correlations also exist between the survival and recovery rates for birds in their second year of life (r = 0.92), third year (r = 0.78), and so on.

71 Table 4.1: AIC values and numbers of parameters K for models examining age structures in φ, p and λ. AICs are shown relative to the minimum AIC of the purely age-dependent models.

Model K ∆AIC 4.1 φ; p; λ ...... 3 7,061.2 4.2 φ1, φ2+; p; λ ...... 4 1,744.1 4.3 φ1, φ2, . . ., φ36; p; λ ...... 38 1,740.7 4.4 φ1, φ2, . . ., φ12, φ13+; p; λ ...... 15 1,706.2 4.5 φ; p1, p2, . . ., p12, p13+; λ ...... 15 2,602.5 4.6 φ; p; λ1, λ2, . . ., λ12, λ13+ ...... 15 6,606.3 4.7 φ1, . . ., φ12, φ13+; p1, . . ., p12, p13+; λ1, . . ., λ12, λ13+ ...... 39 32.4 The following models have survival and recovery structures of φ1, φ2, . . ., φ12, φ13+; λ1, λ2, . . ., λ12, λ13+ 4.8 p1, p2, p3+ ...... 29 102.9 4.9 p1, p2, p3:4, p5+ ...... 30 33.3 −1 4.10 p1, p2, p3:4, p5+(age ) ...... 31 20.5 The following models have survival and recapture structures of −1 φ1, φ2, . . ., φ12, φ13+; p1, p2, p3:4, p5+(age ) 4.11 λ1, λ2, λ3, λ4, λ5, λ6+ ...... 24 9.3 4.12 λ1, λ2:3, λ4+ ...... 21 7.5 The following models have recapture and recovery structures of −1 p1, p2, p3:4, p5+(age ); λ1, λ2:3, λ4+ 4.13 φ1, φ2, φ3, φ4:8, φ9+ ...... 13 15 4.14 φ1, φ2, φ3, φ4:8, φ9+(age) ...... 14 0 The following models have recapture and recovery structures of p1, p2, p3:4, p5+; λ1, λ2:3, λ4+ 4.15 φ age structure from model 4.13 but with φ(age+time) ...... 65 −310.7 4.16 φ age structure from model 4.14 but with φ(age+time) ...... 66 −323.4 The following models have survival and recovery structures of φ1, φ2, φ3, φ4:8, φ9+; λ1, λ2:3, λ4+ 4.17 p age structure from model 4.9 but with p(age+time) ...... 51 −298.6 4.18 p age structure from model 4.10 but with p(age+time) ...... 52 −296.8 The following models have recapture and recovery structures of p1,2,3:4,5+(age+time); λ1, λ2:3, λ4+ 4.19 φ age structure from model 4.13 but with φ(age+time) ...... 105 −402.7 4.20 φ age structure from model 4.14 but with φ(age+time) ...... 106 −414.5 4.21 φ1, φ2, φ3, φ4:8, φ9+(age); p1, p2, p3:4, p5+; λ1, λ2:3, λ4+ ...... 13 13.5

72 1

0.9 (a)

0.8

0.7 Survival probability Recapture probability Recovery probability 0.6

0.5 Probability 0.4

0.3

0.2 (b)

0.1 (c)

0 0 2 4 6 8 10 12 14 Age

Figure 4.2: Age dependence for (a) the survival probability under model 4.4 {φ1, φ2, . . ., φ12, φ13+; p; λ}, (b) the recapture probability under model 4.5 {φ; p1, p2, . . ., p12, p13+; λ}, and (c) the recovery probability under model 4.6 {φ; p; λ1, λ2, . . ., λ12, λ13+}. Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale). Survival for the 13+ age group is shown at age 13.

Therefore the estimates shown in Figure 4.3 are merely a guide to possible age structures to be investigated.

4.3.2 Age structures for model parameters

After simplifying the age structures for each of the model parameters (models 4.8– 4.14 in Table 4.1), the resulting best purely age-dependent model is model 4.14:

−1 φ1,φ2,φ3,φ4:8,φ9+(age); p1,p2,p3:4,p5+(age ); λ1,λ2:3,λ4+.

Indeed, even a simple age structure for survival such as {φ1, φ2+} (model 4.2) is an enormous improvement on a constant survival model (model 4.1). No simplified submodel of model 4.14 has a lower AIC. The notation {φ4:8, φ9+(age)} refers to a “hockey-stick” regression, with φ4 = . . . = φ8 and a linear logistic regression on −1 age, φ9+ = {1 + exp[−(α + βage)]} , with a continuity constraint to ensure that

73 1

0.9

0.8 (a)

0.7 Survival probability 0.6 Recapture probability Recovery probability

0.5 Probability 0.4

0.3 (b)

0.2 (c)

0.1

0 0 2 4 6 8 10 12 14 Age

Figure 4.3: Age dependence for (a) the survival, (b) the recapture and (c) the recovery probabilities under model 4.7 {φ1, . . ., φ12, φ13+; p1, . . ., p12, p13+; λ1, . . ., λ12, λ13+}. Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale). the age dependence is continuous at the “change point” of nine years (see Figure 4.4). The choice of change point is somewhat arbitrary — any choice from eight to 11 years is consistent with the data. (I have counted the change-point location as an extra parameter in Table 4.1.) The senescence implicit in model 4.14 is highly significant statistically (compared with model 4.13, p < 0.0002, likelihood ratio test). However the decline in survival after age nine is only slight, with the survival probability reducing from 83% at age nine years to 76% at age 13. The fully age dependent estimates for survival under the model {φ1, . . ., φ16, φ17+; p1, p2, p3:4, −1 p5+(age ); λ1, λ2:3, λ4+} lie close to the fitted values for model 4.14 (Figure 4.4). I choose age−1 rather than age as the regressor for the recapture probability, as it provides a better fit, and has the intuitively more appealing property of not approaching the limit p = 1 as age increases (see Figure 4.3). Further analysis (see Section 4.3.3) shows that this apparent increase in p with age may be illusory.

74 1

0.9

0.8

0.7

0.6

0.5

0.4 Survival Probability

0.3

0.2

0.1

0 0 2 4 6 8 10 12 14 16 18 Age

Figure 4.4: Age dependence for survival probability for the model {φ1, . . ., φ16, −1 φ17+; p1, p2, p3:4, p5+(age ); λ1, λ2:3, λ4+}. Vertical bars represent 1 standard error each side of the estimate (back-transformed from the logistic scale). Dashed line shows fitted values for model 4.14, in which survival from age 9 years decreases linearly with age (on a logistic scale).

4.3.3 Confounding age and time

Nisbet (2001) notes that a spurious apparent decline in survival probability with age can arise when the quality of the environment deteriorates during a study, since this causes a decline in the survival probabilities for each age group over time; but since birds often enter a study as fledglings, the average age of the birds in the study is increasing over time (that is, age is partially confounded with time). The combined effect of these two trends makes it appear that older birds have lower survival rates. This chapter is concerned with the age dependence of the model parameters. Time dependence of the parameters, and possible covariates that explain such tem- poral variation, will be examined in Chapter 5. However, to account for the pos- sibility of confounding between age and time, I need to single out the effect of age over and above time. I fit two models (models 4.15 and 4.16), both of the form φ(age+time). An age+time model for survival has adult survival varying with time

75 and each of the other age components of survival varying in parallel (on a logistic scale). While models 4.15 and 4.16 are φ(age+time) models, they differ slightly in age dependence: model 4.15 has φ9+ constant, while model 4.16 has φ9+(age). The markedly lower AIC for model 4.16 confirms that, after allowing for possible effects by any time-varying covariates, survival does decline past age nine. In a similar way, I examine the recapture probabilities p, to determine whether these increase after the age of five years. Again I compare similar models for p, both incorporating age and time (models 4.17 and 4.18). A comparison of model

−1 4.17, which has constant p5+, with model 4.18, which includes p5+(age ), shows no improvement in AIC. Therefore the apparent increase in recapture probability with age past five years could be equally well explained as an increase in p with time. Indeed recapture effort did increase substantially during the course of the study, as explained on p.65. The apparent steady increase in p beyond the age of five shown in Figure 4.3 is thus probably illusory. Hence in choosing a final purely age-

−1 dependent model (model 4.21), I keep p5+ constant rather than regress on age . However the recapture age structure cannot be simplified further, as the increase in p up to the age of five years remains. I again confirm that the decrease in survival after age nine remains significant when both φ and p are allowed to vary with time (compare models 4.19 and 4.20). Further consideration of age+time models appears in Chapter 5.

4.3.4 Model averaging

With data such as these, many thousands of plausible models could be fitted. In such circumstances one must be guided by biological as well as statistical reason- ing. Nevertheless, there remains the possibility of bias in the parameters, and of unrealistically low standard errors, by not allowing for the uncertainty inherent in the model selection process. Model averaging (Burnham and Anderson, 1998) over several candidate models can give an indication of the extent of this problem.

For example, the model-averaged estimate of φ1 is 0.1640, with a standard error of 0.0075; where averaging has been taken over models 4.10–4.14 and 4.21. These are close to the values of 0.1647 ± 0.0074 from my chosen model 4.21, and indicate that model selection bias is not a problem.

76 Table 4.2: Estimates (with standard errors in parentheses) for model parameters under model {φ1, φ2, φ3, φ4:8, φ9+(age); p1, p2, p3:4, p5+; λ1, λ2:3, λ4+} (model 4.21).

Year of life Survival Recapture Recovery 1 0.165(0.007) 0.078(0.005) 0.041(0.002) 2 0.711(0.025) 0.216(0.009) 0.121 (0.015) 3 0.780(0.021) 0.294 (0.009) 4 o 5:8 0.829 (0.006)†o 0.194 (0.010) 0.387 (0.008) 9+ ) ) o † For age ≥ 9, this corresponds to the intercept α in the equation φ = {1 + exp[−(α + β(age − 9))]}−1; the slope is β = −0.092 (0.026).

4.3.5 Parameter estimates

The parameter estimates for the chosen model 4.21 (see Table 4.2) indicate a low probability of survival of 0.17 in the first year of life, with 95% confidence interval (CI) from 0.15 to 0.18. For the second and third years of life the values are 0.71 (95% CI from 0.66 to 0.76) and 0.78 (95% CI from 0.74 to 0.82) respectively. Penguins aged three to nine years had a constant annual survival probability of 0.83 (95% CI from 0.82 to 0.84), whereas the survival probability of older adults decreased slightly as the birds aged, from 0.83 at age nine to 0.76 at age 13.

77 4.3.6 Correlations between parameters

For ease of interpretation, I omit the age dependence in φ9+ when calculating the correlations. The matrix of correlation coefficients between logit(φi), logit(pi) and logit(λi) under model {φ1, φ2, φ3, φ4:8, φ9+; p1, p2, p3:4, p5+; λ1, λ2:3, λ4+} appears below. This matrix is calculated using Matlab as diag(se)\inv(H)/diag(se), where se=sqrt(diag(inv(H))) and H is the hessian matrix, evaluated at the MLE using numerical second differences.

φ2 φ3 φ4:8 φ9+ p1 p2 p3:4 p5+ λ1 λ2:3 λ4+

φ1 −0.78 −0.56 0.07 −0.00 −0.58 −0.31 0.06 0.01 0.22 −0.78 0.05

φ2 0.28 −0.14 0.01 0.50 0.07 −0.13 −0.02 −0.19 0.71 −0.10    φ3 −0.25 0.02 0.36 0.36 −0.23 −0.04 −0.14 0.55 −0.17      φ4:8 −0.11 −0.05 0.04 0.22 −0.18 0.02 −0.15 0.21      φ9+ 0.00 −0.00 −0.01 −0.03 −0.00 0.01 0.02      p1  0.20 −0.04 −0.01 −0.14 0.50 −0.03      p2  0.04 0.01 −0.08 0.24 0.03      p3:4 0.06 0.02 −0.14 0.16      p5+ 0.00 −0.02 −0.01      λ1 −0.19 0.01      λ2:3 −0.10     

The highest correlations are between (the logits of) φ1 and φ2, and φ1 and λ2:3 (t = −3.897,p = 0.003). Since these correlations are negative, the data can be explained almost as well by increasing the survival probability for first year birds and decreasing the survival for second year birds (or by increasing φ1 and decreasing

λ2:3). That is, an increase in φ1 would suggest that there are more birds ‘at risk’ at the end of their first year, and hence a lower proportion of at risk birds would need to survive their second year. There are also marginally significant negative correlations between φ1 and φ3 (t = −2.134,p =0.059), and φ1 and p1 (t = −2.241,p =0.049).

If φ1 was increased, then there would be more birds “at risk” at the end of their

first year, so p1 (the probability of recapturing birds at age one year given that they

78 were alive at that stage) and λ2:3 (the annual recovery probability for birds aged 1–3 years) would need to be reduced.

4.3.7 Population modelling

I tested the survival estimates by applying them to a Leslie matrix model (see Caswell, 2001) and checking whether this model matched the observed stability of the penguin population in the study sites. Since this model implicitly includes emigration, via the apparent survival (see Section 4.5.2), it should also include immigration. From 1970 to 2000, 11.6% of the breeders seen in the main study site were “strangers” to the site, that is, birds that hadn’t previously been banded, or birds that had been banded elsewhere (see “Strangers” Analysis in Section 4.7). So, if η denotes the proportion of the adult population of age 2 years and over that are immigrants from the surrounding areas, I estimate η as 0.116. Since most of these birds were of unknown age, I cannot estimate immigration by age group. Here I use a simple model for survival, with a constant annual survival proba- bility for birds aged over three years {φ1,φ2,φ3,φ4+} (model 4.22), since it is not possible to form a Leslie matrix using the survival model {φ1,φ2,φ3,φ4:8,φ9+}, be- cause of the 4–8 age group. I use the recapture and recovery age structures of model

4.21 (that is, {p1,p2,p3:4,p5+; λ1,λ2:3,λ4+}). The annual survival probabilities are estimated as φ1 =0.1638, φ2 =0.7174, φ3 =0.7896, φ4+ =0.8180. Other inputs to the model are as follows: the mean annual production of chicks between 1968 and 2004 is 0.94 fledged per breeding pair or F = 0.47 chicks per individual (P. Dann, unpubl. data); and 50% of 2-year-olds are breeders (b = 0.5) (Dann and Cullen, 1990). I assume that all birds are breeding when three years of age and over, although some birds of breeding age may not breed in particularly bad years (Dann and Cullen, 1990).

Let ni(t) (i = 0, 1, 2) be the number of birds of age i years and let n3(t) be the number of birds of age three years and over immediately after the breeding season in year t. A deterministic model for the number of birds in each age group immediately after the breeding season in year t + 1 is n(t +1) = An(t), where

T n(t)=(n0(t), n1(t), n2(t), n3(t)) , and A is the Leslie matrix

79 0 φ2bF (φ3 + η)F (φ4+ + η)F   φ1 0 0 0 A = .    0 φ2 0 0       0 0 φ3 + η φ4+ + η      Using the estimates above, this matrix has principal eigenvalue λmax = 0.987, and so predicts an almost steady population size, in agreement with observations in the study sites. This result provides some confidence in the population model and, in particular, the age-specific survival estimates reported here.

4.4 Overdispersion

Burnham and Anderson (2002, p.68) recommend checking for overdispersion by estimating c, the variance inflation factor. When there is no overdispersion, c = 1 and QAIC is equivalent to AIC. If there is a statistic that measures the goodness of fit of the data, and which has, when the data are not overdispersed, a chi-square distribution with df degrees of freedom, then the variance inflation factor can be estimated as cˆ = χ2/df

(Cooch and White, 2006, Chapter 5, p.2). Goodness-of-fit testing is problematic, particularly for mark-recapture-recovery analyses and for models incorporating co- variates (Pollock, 2002), and so there are “no general, robust procedures” for esti- mating c (White and Burnham, 2002). However, Cooch and White (2006, Chapter 5, pp.21–24) propose a Monte Carlo approach to estimating c, whereby a large num- ber of data sets are simulated by using the parameter estimates obtained under the assumption that a selected model is correct. The model is fitted to each simulated data set, and the deviance is calculated each time (as in Catchpole et al., 1999). Finally,c ˆ is calculated as

deviance for observed data cˆ = . mean of deviances for simulated data

I now apply this Monte Carlo approach to estimate the variance inflation factor for the penguin data. I begin by choosing a model that fits the data well. Reliable

80 parameter estimates are needed since they will be used to generate the simulated life history data. Modelling time dependence in each age component for the “best” model (model 4.21), that is {φ1,2,3,4:8,9+(age∗time),p1,2,3:4,5+(age∗time),λ1,2:3,4+(age∗time)}, would potentially yield parameter estimates that are not robust, due to the lack of data for some age classes in some years. Although a possible approach is to use random effects for time (see Barry et al., 2003), here I consider a simple model,

{φ1,2,3+(age∗time); p1,2+(age∗time); λ1,2+(age∗time)} (model 4.23), which retains the important age structures in φ, p and λ, and yields more robust parameter estimates than the best model. Furthermore, this simple model includes 247 pa- rameters and takes 27 minutes to run, as compared to 407 parameters and 107 minutes run-time for the “best” model. In order to determine the deviance of this model, I calculate the likelihood of the maximal model, Lmax, using a cohort-based approach, since this approach was used to calculate the likelihood in the main analysis. I calculate Lmax by recognising that the history entries for the birds in each cohort form a multinomial distribution. Lmax is then formed by taking the product over each cohort as follows:

C ui di,j 1 di,j Lmax = Ri! , di,j! Ri i=1 " j=1 # Y Y   where C is the number of cohorts, Ri is the number of birds in cohort i, ui is the th number of unique histories in cohort i, and di,j is the number of times the j unique history occurs in cohort i (for i = 1,...,C and j = 1,...,ui). (Note that the maximal model is not required anywhere else in this thesis.) The constant term in the likelihood for the maximal model is equivalent to the constant term in the cohort-based likelihood used in this analysis. Hence the constants are ignored when calculating the deviances. I now generate 99 simulated data sets using the parameter estimates obtained under the assumption that model 4.23 is correct. I then fit the model to each simulated data set, and calculate the deviance each time. Finally, I estimate the variance inflation factor asc ˆ ≈ 1.002. If the model fits poorly, thenc ˆ may be underestimated. However, since model 4.23 is reasonably close to the best model,

81 andc ˆ is very close to one, I conclude that the data do not appear to be overdispersed. Hence, my use of the AIC for model selection is appropriate.

4.5 Discussion

4.5.1 Age structures

The survival, recapture, and recovery structures of model 4.21 are consistent with the differing lifestyles of birds of various ages (see Section 4.1). While penguins in their first year of life survive poorly, probably due to their lack of food-finding skills, more experienced birds in their second and third years have much higher survival probabilities. Once breeding is established, the survival probability stabilises, until it begins to decline slightly, with the onset of senescence occurring at around nine years of age. One-year-old birds are rarely seen, resulting in a relatively low probability of recapture, while two-year-olds are more likely to be encountered, since they may return to the colony to moult and breed. Breeders have a higher recapture proba- bility, as they spend more time in the colony than juveniles. However, I find strong statistical evidence for considering the recapture probability of three- and four-year- olds separately from older adults (models 4.8 and 4.9, Table 4.1 on p.72). This is perhaps due to experienced birds breeding more successfully (and thus spending more time in the colony raising chicks) than their younger counterparts — Dann and Cullen (1990) reported that breeding success increased with age, and Chiara- dia and Kerry (1999) found that successful breeders were on average one year older than unsuccessful breeders. Since recoveries are often made by members of the public, recovery rates vary by geographic location, and the spatial difference in the recovery distributions of dead birds (Figure 4.1, p.65) provides a plausible explanation for the observed juvenile/adult age variation in the recovery rate that is suggested by the chosen model 4.21. The apparent increase in recapture probability with age beyond five years shown in Figure 4.2, p.73 is most likely an artefact, caused by an increase in search effort during the study. This provides a classic illustration of the confounding between age and time in a study of animals marked at birth, since as time progresses the

82 average age of the animals increases (Nisbet, 2001). This is an important issue to bear in mind when modelling age dependence.

4.5.2 Apparent versus true survival

Mark-recapture-recovery studies often include annual history entries of ‘0’ in be- tween banding and the final (live or dead) encounter. In fact, of the birds that are banded as chicks in this study and encountered dead or alive subsequently, 66% are not seen during at least one year between banding and the final encounter. While some of these zeros occur because the bird was present at the study site but simply not encountered during the burrow visits, others occur because the bird was absent from the site. For example, a relatively small proportion of surviving one-year olds are recaptured (p1 = 8% compared to p5+ = 39%, model 4.21), since they spend their time away from the colony until they return to moult or breed at age two or three years (Reilly and Cullen, 1981; Dann and Cullen, 1990). Other possible reasons for zero entries are: birds may occasionally take a breeding sabbatical, or they may temporarily move out of the study site. Philopatry amongst Little Penguins is very strong, with large-scale movements between Phillip Island and other colonies in Bass Strait appearing to be quite rare (Dann, 1992). Indeed penguins generally breed in the same burrow (or in one close by) from one year to the next (Reilly and Cullen, 1981). However small-scale emigration can occur when birds whose burrows are located on the outer limit of a study site move to new burrows a few metres outside the study site; such birds may consequently fail to be recaptured for one or more years (Dann and Cullen, 1990).

Hence a topic for future research would be modelling pt in terms of pt−1 and pt−2 — is there any pattern in their decision to take a sabbatical? Since there are no existing data that allow the estimation of small- or large- scale emigration out of the study sites, emigration has not been included in my mark-recapture-recovery analysis. Hence φ represents the survival without taking site fidelity (of lack thereof) into account. That is, if e1 is the probability that a bird emigrates (on a large or small scale) from its natal site in its first year of life, then φ1 = s1(1 − e1), where s1 is the true survival probability. Although I expect the true and apparent survival estimates to be closer for breeders than for juveniles

83 (since emigration would most likely occur after penguins fledge and leave their natal colony), local emigration from the study sites may contribute to an underestimation of the true survival. Immigration must be included in the population model: if immigration is omit- ted, the model predicts a population decline of 12.3% per annum: that is, the maximum eigenvalue of the Leslie matrix is λmax = 0.877. Such a small λmax is not caused by sampling error in the parameter estimates — increasing each of the parameter estimates in model 4.22, in turn, by two standard errors increases

λmax to a value ranging from 0.880 for φ3 to 0.886 for φ4+. Thus movement must play a significant role in the population dynamics, and so my survival estimates

(particularly φ1) which implicitly include emigration, must be underestimates of the true survival. However, since the relative magnitude of small-scale emigration for the various age groups is unknown, it is not possible to use the measurement of influx of breeders to the study sites to determine the true survival.

4.5.3 Comparison with other survival estimates

Reilly and Cullen (1979), in their study of ten years of data for Little Penguins banded at breeding age, obtained an overall estimate of annual adult survival of 85%, while Dann and Cullen (1990) in their analysis of 20 years of Phillip Island Little Penguin data, concluded that survival decreased with age from four years of age, with an average adult survival of 75%, and a survival estimate of 31% from fledging to three years of age. However, such life table analyses are well known to produce biased estimates of survival (Seber, 1986). Although the estimate of annual adult survival for Little Penguins obtained in this study (83%) falls within the range of published estimates for other penguin species, the first-year survival probability (17%) is substantially lower than most estimates for these other species (see Table 2.1 on p.23). The Emperor Penguin is the only species with a reported first-year survival (19%) similar to the estimate for Little Penguins. However, the Emperor Penguin’s low first-year survival is com- pensated (in terms of population stability) by a much higher annual adult survival (95%). There are several factors that may contribute to an underestimation of survival:

84 a. Local emigration — As discussed above, the apparent survival estimates for juveniles may be significantly lower than the true survival, due to local emi- gration. b. Band loss — The rate of band loss in juveniles is unknown, but in a sample of adult birds banded and marked with injected transponders, the annual probability of losing a band was found to be less than 0.4% (Chapter 7). This rate of band loss would have very little influence on the survival estimates. c. Banding effect — Recent studies have indicated that survival rates of banded King Penguin chicks are significantly lower than unbanded, transpondered chicks (Gauthier-Clerc et al., 2004), and that banded Ad´elie penguins use more energy than unbanded penguins when swimming (Culik et al., 1993), and may have lower survival rates than unbanded penguins until their first moult after banding (Ainley et al., 1983). I show in Chapter 7 that banded adult Little Penguins have an annual survival probability 6% lower than their unbanded counterparts. Although no such data are available for birds banded as chicks, it is possible that banding significantly reduces the survival of ju- venile Little Penguins, particularly in their first year of life, and that I have consequently underestimated the first-year survival of unbanded penguins. A study to evaluate this possibility is in progress on Phillip Island.

4.6 Analyzing subsets of the complete data set

In this section I begin by using the best purely age-dependent model as determined in Section 4.3, that is, {φ1,φ2,φ3,φ4:8,φ9+(age); p1,p2,p3:4,p5+; λ1,λ2:3,λ4+} (model 4.21), and compare the estimates obtained for the birds banded as chicks on Phillip Island (with the disqualifications as listed in Chapter 3), with those for various subsets of the complete raw data set.

4.6.1 Including birds that failed to fledge

As explained in Chapter 3, birds that failed to fledge have been disqualified from my main analysis, so that the first-year survival estimate actually represents the survival from fledging to one year of age. I now compare the estimates of φ1 and

λ1 obtained in Section 4.3, with those arising when I include birds that failed to fledge in the analysis (see Table 4.3). As expected, the inclusion of data for birds

85 Table 4.3: Estimates of φ1 and λ1 when birds that failed to fledge are included in, or disqualified from, the analysis.

φ1 λ1 Birds that failed to fledge included 0.159 (0.008) 0.067 (0.003) Birds that failed to fledge disqualified 0.165 (0.007) 0.041 (0.002) that failed to fledge results in a lower estimate of the first-year survival probability and a higher probability of recovery of birds in their first year.

4.6.2 Recapture-only, recovery-only and recapture-recovery analyses using data for birds banded as chicks, those banded as adults, and all birds

The main analysis focuses on the mark-recapture-recovery data for birds banded as chicks. This section analyses various subsets of the full data set, as in Catchpole et al. (1998a). For each of the subsets of the data, that is for birds banded as chicks (bachs), for those banded as adults (baads), or for all birds (both bachs and baads), I carry out a combined mark-recapture-recovery analysis, an analysis using only the dead recovery data, and an analysis using the information on the live recaptures only. Thus I analyse nine subsets of data. Appendix A.2 includes an explanation of the way in which the MATLAB programs calculate the sufficient statistics and the likelihood corresponding to the subset of data being considered. Since the birds banded as adults are of unknown age, I cannot impose an age structure on the model parameters for these birds. Therefore, the information cor- responding to these birds is only used when calculating the oldest age components of the survival, recapture and recovery probabilities. For this reason I have simpli-

fied the age structure for survival in model 4.21 from {φ1, φ2, φ3, φ4:8, φ9+(age)} to

{φ1,φ2,φ3,φ4+} (model 4.22). The parameter estimates corresponding to the nine subsets of the data appear in Table 4.4.

Compare the estimate for the adult survival probability (φ4+) for birds banded as chicks under a combined recovery-recapture analysis (0.818 ± 0.005), with that obtained under a recapture-only analysis (0.817 ± 0.005), and with recovery only

(0.842 ± 0.012). Since the estimates of φ4+ obtained under recovery-recapture and recapture-only analyses are almost identical and there is no improvement in the standard error when the recovery information is added, the information appears

86 Table 4.4: Parameter estimates for model {φ1,φ2,φ3,φ4+; p1,p2,p3:4,p5+;λ1,λ2:3,λ4+} (model 4.22) using various subsets of the data. Standard errors are given in parentheses.

Type of analysis φ1 φ2 φ3 φ4+ p1 p2 p3:4 p5+ λ1 λ2:3 λ4+ recov-recap: bachs 0.164 0.717 0.790 0.818 0.078 0.215 0.291 0.389 0.041 0.125 0.191 (0.007) (0.025) (0.021) (0.005) (0.005) (0.009) (0.008) (0.008) (0.001) (0.015) (0.010) recov-recap: baads - - - 0.770 - - - 0.522 - - 0.118 - - - (0.002) - - - (0.003) - - (0.003) recov-recap: all 0.161 0.741 0.821 0.777 0.080 0.213 0.282 0.501 0.041 0.141 0.127 (0.007) (0.025) (0.020) (0.002) (0.006) (0.009) (0.008) (0.003) (0.001) (0.018) (0.003) recap only: bachs 0.151 0.701 0.762 0.817 0.085 0.240 0.336 0.454 - - -

87 (0.009) (0.048) (0.029) (0.005) (0.007) (0.012) (0.010) (0.009) - - - recap only: baads - - - 0.753 - - - 0.568 ------(0.002) - - - (0.004) - - - recap only: all 0.151 0.700 0.811 0.762 0.085 0.240 0.324 0.551 - - - (0.009) (0.048) (0.030) (0.002) (0.007) (0.012) (0.009) (0.003) - - - recov only: bachs 0.606 0.612 0.679 0.842 - - - - 0.087 0.025 0.074 (0.500) (0.000) (0.000) (0.012) - - - - (0.500) (0.000) (0.000) recov only: baads - - - 0.846 ------0.127 - - - (0.007) ------(0.004) recov only: all 0.343 0.622 0.692 0.845 - - - - 0.052 0.045 0.127 (0.000) (0.000) (0.000) (0.006) - - - - (0.000) (0.000) (0.004) to come almost entirely from the live recaptures. This is confirmed by comparing the survival estimates (φ4+) and standard errors for birds banded as adults under a recapture-recovery analysis (0.770±0.002), a recapture-only analysis (0.753±0.002) and a recovery-only analysis (0.846 ± 0.007), and noting the reduction in standard error when recapture information is incorporated. Recall that the data for birds banded as adults contribute only towards the parameter estimates for the oldest age group. The recapture probability estimate for adults aged five years and over (p5+) is much higher when the data for birds banded as adults are included (e.g. using recovery-recapture data, p5+ = 52% for baads and 39% for bachs). This is to be expected since most birds banded as adults are likely to be breeders (P. Dann, pers. comm.), and once breeding commences, birds tend to breed in the same area and so they are more likely to be recaptured in subsequent breeding seasons. Since the estimates of p5+ are in the ratio of 3:4 1 for bachs and baads, this may suggest that around 4 of birds emigrate during their first four years of life.

The estimate and standard error for φ4+ under a recapture-recovery analysis are much higher for bachs (0.818 ± 0.005) than for baads (0.770 ± 0.002). This is perhaps due to locals (bachs) knowing the area better and hence surviving better than the new recruits (baads). I would expect the recovery-only analysis to yield higher survival estimates than those arising from the recapture information, since the survival estimates based on the recaptures could be biased by permanent emigration (see p.28 of this thesis, and Lebreton, 2001). The first-year survival estimate should be most affected, as permanent emigration would most likely occur after penguins fledge and leave their natal colony. However, since it is very important to model the survival and recovery probabilities in the first year of life separately from the older age classes, resulting in the parameter redundancy problem discussed below, this issue cannot be adequately assessed.

Parameter redundancy The parameter estimates in Table 4.4 corresponding to the recovery-only analyses are unreliable due to a parameter redundancy problem. This problem arises since the age structures for φ and λ “step up” at the same age (Catchpole et al., 1996),

88 Table 4.5: Estimates of φ and λ for birds banded as chicks, birds banded as adults and all birds, using only recovery information under model {φ1,φ2,φ3,φ4+; λ}.

φ1 φ2 φ3 φ4+ λ bachs 0.445 (0.014) 0.786 (0.017) 0.862 (0.016) 0.842 (0.012) 0.062 (0.002) baads - - - 0.846(0.007) 0.127(0.004) all 0.457 (0.014) 0.795 (0.016) 0.868 (0.015) 0.829 (0.005) 0.080 (0.002)

that is in model {φ1,φ2,φ3,φ4+; λ1,λ2:3,λ4+}, both φ and λ “step up” at ages one and three years. Parameter redundancy can be avoided by ensuring that the pa- rameters step up at different ages, by making first-year survival vary with time, or by making the recovery probability constant (Catchpole et al., 1996). The re- sults under model ({φ1,φ2,φ3,φ4+; λ}) appear in Table 4.5. However, it should be noted that this model is not ideal since the recovery probability was found to vary considerably with age (see Section 4.3). Note that the first-year survival estimates are much higher than those obtained for the mark-recapture-recovery analysis. This is possibly due to leakage between the model parameters (see Section 4.3).

4.7 “Strangers” analysis

There are no existing data which enable me to estimate small-scale immigration into the study sites. However, in this section I obtain a measure of the influx of birds into a study site from the surrounding areas, by studying the recapture/recovery data and the breeding data for birds from the main study site (Site 1), which has been studied intensively since the Phillip Island study commenced. This is done in a fairly na¨ıve way. However, it is possible that the methods of Pradel (1996), which analysed time-reversed data so that “death” became “recruitment”, could be extended to estimate immigration (B. Morgan, pers. comm.). I wish to determine the proportion of breeders that are “strangers” to Site 1 in year j, that is, birds that bred in Site 1 in that year but were: (i) previ- ously unbanded, (since birds previously breeding in Site 1 would almost certainly have been recaptured during that season, see p.67) or (ii) banded elsewhere (as chicks or adults) in a previous year. In this strangers analysis, I have ignored cap- tures/recaptures of strangers that were not breeding in Site 1 in that season, since

89 these include birds that were merely passing through the site on their way out to sea or to their own burrows outside of the site. Indeed, an average of 30% of adults captured/recaptured alive in Site 1 in a given year had never been encountered in that site before, but this figure is a gross overestimate of the true small-scale immigration, due to the presence of live encounters with transient birds. Hence, to obtain a more realistic estimate of the influx of birds from surrounding areas, it is necessary to supplement the raw mark-recapture-recovery data with the breeding data (see Section 3.5 on p.58). Again I use the R package to extract the required information from the main raw data set, using the code strangers.R1 I begin by restricting the data set to include only records from Site 1, and then disqualify birds as outlined in Section 3.2. For each bird, I form a live encounter history for Site 1 over the years of the study. I also determine the year of banding for a bird banded as an adult in Site 1, and the year of first encounter in Site 1 if it was banded elsewhere in a previous year. From this information, I generate the band numbers of birds that were strangers to the site in each year. Using the breeding data sheets, I obtain lists of the band numbers of all breeders in Site 1 in each year. I then compare these to the information generated using the mark-recapture information for Site 1, to determine how many of the breeders were in fact strangers to the site. Table 4.6 shows the total number of breeders in Site 1 and the number of breeders that were strangers to the site from 1970 to 2000. Over this time period, 11.6% of the breeders seen in Site 1 were strangers to the site. Therefore, an estimate of the influx of breeders into a study site from the surrounding areas is 11.6% per annum. This value is used as an estimate of immigration into the study sites in the population modelling in Section 4.3.7. Occasionally the researchers were unable to determine the parents of the eggs or chicks that they found in a burrow. These unidentified parents were referred to as “unknown breeders” (eg. the parents in burrow 9076, Table 3.4 on p.59). Unknown breeders were ignored in calculating the proportion of breeders that were strangers. If all of these birds were also strangers to the site, then an average of 13.9% (127/915) of the breeders would have been strangers.

[email protected]

90 Table 4.6: Total number of breeders and numbers of strangers and unknown breed- ers breeding in Site 1 from 1970 to 2000.

Number of breeders Year Total Strangers Unknown 1970 48 4 0 1971 56 2 0 1972 44 4 3 1973 34 2 0 1974 41 2 1 1975 43 3 0 1976 48 5 0 1977 41 2 0 1978 27 2 2 1979 36 4 0 1980 28 3 0 1981 22 2 0 1982 18 2 0 1983 14 2 0 1984 16 4 2 1985 13 3 1 1986 19 2 1 1987 10 4 0 1988 18 9 0 1989 16 4 0 1990 18 2 0 1991 20 2 0 1992 24 2 2 1993 32 4 0 1994 35 4 0 1995 35 6 1 1996 33 8 1 1997 19 3 4 1998 36 5 1 1999 27 0 3 2000 20 2 2 Totals 891 103 24

91 Table 4.7: Estimates (with standard errors in parentheses) for model parameters under model {φ1, φ2, φ3, φ4:8, φ9+; p1, p2, p3:4, p5+; λ1, λ2:3, λ4+} when captures, recaptures and recoveries are summarized by penguin year (na¨ıve method).

Year of life Survival Recapture Recovery 1 0.1503 (0.0064) 0.0951 (0.0063) 0.0373 (0.0014) 2 0.7871 (0.0297) 0.2542 (0.0103) 0.0512 (0.0081) 3 0.7652 (0.0217) 0.3397 (0.0093) 4 o 0.8446 (0.0069) 0.1431 (0.0103) 5:8 o 0.4179 (0.0090) 9+ o0.7965 (0.0120) o o 4.8 Estimates using old data summary and current method

Section 3.3 outlines the method used in this thesis to summarize the raw fort- nightly/monthly mark-recapture-recovery information into yearly data, that is, the captures/recaptures are summarized by penguin year and the dead recoveries by calendar year. In Section 3.4, I explain an alternative, na¨ıve method of summa- rizing the data, whereby the recaptures and recoveries are both summarized by penguin year. Here I compare the parameter estimates obtained using these two methods of data summary under a simple age-dependent model

{φ1,φ2,φ3,φ4:8,φ9+; p1,p2,p3:4,p5+; λ1,λ2:3,λ4+} (see Tables 4.7 and 4.8). As explained in Section 3.4, when the captures, recaptures and recoveries are summarized by penguin year (na¨ıve method), a bird banded on 1 January 1970 and recovered dead by 30 June 1971 would have history entries of “1 2 0 0”. Although the bird could have survived its first 18 months of life, the history entries suggest that it did not survive its first year of life. If the recoveries are summarized by calendar year (as in my current method), the dead recoveries from 1 January 1971 to 30 June 1971 would be assigned to the 1971 penguin year and these birds would have history entries of “1 0 2 0”. Therefore, under the na¨ıve method of data summary,

φ1 is artificially depressed (since some birds that die in their second year of life are recorded as dying in their first year of life). Similarly the na¨ıve method gives a φ2 which is too high (since there are fewer birds recorded as dying in their second year of life) and a λ2:3 that is too low (due to the apparent decrease in recoveries of dead birds aged 1–2 years).

92 Table 4.8: Estimates (with standard errors in parentheses) for model parameters under model {φ1, φ2, φ3, φ4:8, φ9+; p1, p2, p3:4, p5+; λ1, λ2:3, λ4+} when cap- tures/recaptures are summarized by penguin year and recoveries by calendar year (current method).

Year of life Survival Recapture Recovery 1 0.1648 (0.0074) 0.0777 (0.0054) 0.0409 (0.0015) 2 0.7116 (0.0249) 0.2158 (0.0092) 0.1214 (0.0146) 3 0.7809 (0.0213) 0.2935 (0.0085) 4 o 0.8288 (0.0063) 0.1942 (0.0103) 5:8 o 0.3876 (0.0076) 9+ o0.7970 (0.0096) o o 4.9 Conclusion

In this chapter, I successfully model age dependence for the survival, recapture and recovery probabilities simultaneously. The age structures for the parameters in model {φ1,φ2,φ3,φ4:8,φ9+(age); p1,p2,p3:4,p5+; λ1,λ2:3,λ4+} (model 4.21) make biological sense and are consistent with the differing lifestyles of birds of various ages (see Section 4.1). While penguins in their first year of life survive poorly (with a survival probabil- ity of 17%), probably due to their lack of food-finding skills, more experienced birds in their second and third years have much higher survival probabilities (71% and 78% respectively). Once breeding is established, the survival probability stabilises (at 83%), until it begins to decline slightly, possibly due to senescence, at around nine years of age. One-year-old birds are rarely seen, resulting in a relatively low probability of recapture (8%), while two-year-olds are more likely to be recaptured (22%), since they may return to the colony to moult and breed. Breeders have a higher recapture probability, as they spend more time in the colony than juveniles. However, new breeders must be considered separately from older adults, perhaps because experi- enced breeders are more successful and spend more time in the colony than their younger counterparts (p3:4 = 29%, p5+ = 39%). The apparent increase in recapture probability with age beyond five years is most likely an artefact, caused by an in- crease in search effort during the study. This illustrates the confounding between age and time that can occur in a study of animals marked at birth.

93 Recovery rates vary by geographic location, and the spatial difference in the recovery distribution of dead birds provide a plausible explanation for the observed juvenile/adult age variation in the recovery rate that is suggested by model 4.21.

The recovery locations of first year birds are widely dispersed (λ1 = 4%), while those of young adults are less dispersed (λ2:3 = 12%), and the recoveries of established breeders are concentrated on Phillip Island and in Port Phillip Bay (λ4+ = 19%). Since there are no existing data that allow the estimation of small- or large-scale emigration out of the study sites, emigration has not been included in the analysis. Hence φ represents the net survival after site fidelity has been taken into account. Although I expect the true and apparent survival estimates to be closer for breeders than for juveniles (since emigration would most likely occur after penguins fledge and leave their natal colony), local emigration from the study sites may contribute to an underestimation of the true survival. However, since the relative magnitude of small-scale emigration for the various age groups is unknown, it is not possible to determine the true survival. Using the breeding data for the main study site, I found that an average of 11.6% of the breeders encountered in this site in any given year were strangers to the site, that is birds that were previously unbanded, or banded elsewhere (as chicks or adults) in a previous year. A population model allowing for immigration of birds from areas surrounding the study sites fits the observed stable population in the study sites.

94 Chapter 5

Modelling Time and Covariate Dependence

From a biological perspective, I am primarily concerned with the survival probability — how this parameter varies with time and age, and possible causes of this variation. However, it is imperative that ap- propriate age structures and dependences for the recapture and recovery probabilities are applied, since an inappropriate choice for one parame- ter can lead to unrealistic estimates for the others, as I have shown in Chapter 4 (see also Catchpole et al., 1998a). This chapter uses the age structures for the survival, recapture and recovery probabilities determined in Chapter 4. The age components of the model parameters that exhibit temporal variation are identified, by allowing each component to vary in turn with time. In an attempt to explain the observed temporal variation in the survival probability, I consider several time-varying group covariates, that is, extrinsic co- variates that apply to all animals and that vary from year to year. The following biological data are used as covariates for survival: the mean an- nual banding/fledging weight, the mean annual number of chicks fledged per pair, and the mean annual laying date for first clutches. Other group covariates discussed here are the mean annual number of beach crossings of adult penguins at the Penguin Parade (a measure of population pres- sure), as well as seasonal sea surface temperature data, and Southern Oscillation Index data. The covariate values appear in Figure 5.1 and the correlations between them can be found in Section 5.4 on p.121.

95 Chick weight at banding (g) Chicks fledged per pair 1100 2

1000 1.5

900 1

800 0.5

700 0 1970 1980 1990 2000 1970 1980 1990 2000 Mean laying date (as a decimal) Mean annual parade counts 12 800

11 600 10 400 9

8 200 1970 1980 1990 2000 1970 1980 1990 2000 Mean annual SST in chick region Mean annual SOI 6.5 10 6 0 5.5 −10 5

4.5 −20 1970 1980 1990 2000 1970 1980 1990 2000 Year Year

Figure 5.1: The variation over time of the covariates for survival: current weight at banding, chicks fledged per pair, mean laying date, mean annual parade counts, mean annual sea surface temperature (SST), and mean annual southern oscillation index (SOI). Only one of the SST graphs is shown.

96 5.1 Time dependence

By modelling age dependence simultaneously in the survival, recapture and recovery probabilities, Chapter 4 confirmed that all three model parameters exhibit very rich age structures. The best model (model 4.21, see Table 4.1 on p.72) had different annual probabilities for each of the age components of the model parameters; for survival, for birds aged 0–1 years, 1–2 years, 2–3 years, and 3–9 years, and with survival decreasing linearly with age (on a logistic scale) for birds aged nine years and older; for recapture for birds of age one year, two years, three and four years, and those aged five years and over; and, for recovery, for birds aged 0–1 years, 1–3 years and those four years and over. This model is denoted by {φ1, φ2, φ3, φ4:8,

φ9+(age); p1, p2, p3:4, p5+; λ1, λ2:3, λ4+}. In this section, I consider time dependence for the various age components of the model parameters. For consistency, the survival probability for adults aged nine years and over is kept constant rather than decreasing linearly with age, since it becomes unnecessarily complicated to include time dependence in φ9+ while still maintaining continuity in survival at age nine years. Given the very small gradient, any bias resulting from this will be small. Hence the age structures is fixed as

{φ1,φ2,φ3,φ4:8,φ9+; p1,p2,p3:4,p5+; λ1,λ2:3,λ4+} (model 5.1).

The first cohort of birds (consisting of 17 birds banded in penguin year 1967) has been omitted, since banding began part-way through the breeding season, and the time-varying covariates (such as mean laying date, mean weight at banding) are not meaningful for this year due to the small size of the cohort. Tables 5.1–5.4 show the models considered, together with their AIC values. I begin by introducing temporal variation for each of the age components of the survival probability in turn (models 5.2–5.6 in Table 5.1), before considering time dependence for survival in all age groups independently (model 5.7 in Table 5.1). The strong temporal variation for first-year survival is expected (compare models 5.1 and 5.2), since important factors such as food availability, which varies from year to year, can have a major effect on the survival of newly–fledged birds, either directly, or indirectly via the effect of these factors on their parents. Time dependence is

97 Table 5.1: AIC values (after subtracting 44 000) and numbers of parameters K for models involving time dependence in the survival probability. All models have recapture and recovery structures of {p1, p2, p3:4, p5+, λ1, λ2:3, λ4+}.

Model K AIC 5.1 φ1, φ2, φ3, φ4:8, φ9+ ...... 12 1034 5.2 φ1(time), φ2, φ3, φ4:8, φ9+ ...... 46 672 5.3 φ1, φ2(time), φ3, φ4:8, φ9+ ...... 45 837 5.4 φ1, φ2, φ3(time), φ4:8, φ9+ ...... 44 995 5.5 φ1, φ2, φ3, φ4:8(time), φ9+ ...... 43 1023 5.6 φ1, φ2, φ3, φ4:8, φ9+(t) ...... 38 999 5.7 φ1,2,3,4:8,9+(age∗time) ...... 168 616 5.8 φ1,2,3,4:8,9+(age+time) ...... 64 687 stronger for the first-year survival than for the survival of the older groups, in terms of model fit as measured by the AIC (compare model 5.2 with models 5.3–5.6). This result corresponds to predictions made from life-history theory (for example Saether and Bakke, 2000, and references therein). Although temporal variation is still present for φ2 and φ3, this becomes less important for the older age groups. However, it should be noted that the AIC depends on the amount of available information on the survival of the various age groups, as well as the variation of these probabilities over time. Figure 5.2 shows the yearly variation in survival under an age∗time modelling scheme (denoted by {φ1,2,3,4:8,9+(age∗time)}, model 5.7), which models time-dependent survival rates for all age-classes independently. The estimates for first-year survival vary considerably over time, with 1983 being a particularly good year, and 1986 being particularly bad. It is interesting to observe the reduced survival probabil- ities for all age groups, following the widespread mortality of their primary food source, pilchards, in 1995 (Dann et al., 2000). The variation in standard errors for

φ1 in Figure 5.2 can be explained by the yearly differences in the number of birds in each cohort. For example, there were 2375 birds banded as chicks in the 1993/1994 breeding season and only 100 in the 1997/1998 season. Hence the estimate for φ1 in the 1998 calendar year has a large standard error. The very wide confidence intervals for some of the yearly survival estimates of birds in the older age groups indicate that there is insufficient information to model yearly variation in survival for these age classes. Although the adult survival also exhibits temporal variation,

98 Survival Probabilities 0.5

Age 0−1 0 1970 1975 1980 1985 1990 1995 2000 1

0.5

Age 1−2 0 1970 1975 1980 1985 1990 1995 2000 1

0.5

Age 2−3 0 1970 1975 1980 1985 1990 1995 2000 1

0.5

Age 3−8 0 1970 1975 1980 1985 1990 1995 2000 1

0.5 Age 8+ 0 1970 1975 1980 1985 1990 1995 2000 Calendar Year

Figure 5.2: Survival probabilities versus time for model φ1,2,3,4:8,9+(age∗time), p1, p2, p3:4, p5+, λ1, λ2:3, λ4+ (model 5.7). Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale). the survival estimates are much higher, and vary less, than those of the first year birds. The inclusion of time dependences for all age groups greatly increases the number of parameters, and thus the computer run-time. For example, changing the survival model from {φ1(time), φ2,φ3,φ4:8,φ9+} (model 5.2 in Table 5.1) to an age∗time model for survival (model 5.7) leads to a considerable improvement in model fit (∆AIC=56). However this substantially increases the complexity of the model, adding an extra 122 parameters. This problem could be overcome by using random effects for time (Barry et al., 2003). As a compromise, model 5.8 (in Table 5.1) has an age+time structure for sur- vival, denoted by φ1,2,3,4:8,9+(age+time). This implies that time-varying factors af- fecting survival (for example climate) affect all age groups similarly. That is, there is the same proportional change in survival odds, so that the survival varies in parallel (on a logistic scale) for each age component. (Appendix A.2.2 on p.205 shows how to program an age+time model.) An age+time model for survival (model 5.8) leads

99 Table 5.2: Numbers of parameters under models φ1,2,3,4:8,9+(age∗time) (model 5.7) and φ1,2,3,4:8,9+(age+time) (model 5.8). The recapture and recovery structure is {p1, p2, p3:4, p5+, λ1, λ2:3, λ4+}.

Model for survival No. of parameters φ p λ Total age∗time 161 4 3 168 age+time 5743 64

Survival Probabilities 0.5

Age 0−1 0 1970 1975 1980 1985 1990 1995 2000 1

0.5 Age 1−2 0 1970 1975 1980 1985 1990 1995 2000 1

0.5 Age 2−3 0 1970 1975 1980 1985 1990 1995 2000 1

0.5

Age 3−8 0 1970 1975 1980 1985 1990 1995 2000 1

0.5 Age 8+ 0 1970 1975 1980 1985 1990 1995 2000 Calendar Year

Figure 5.3: Survival probabilities versus time for model {φ1,2,3,4:8,9+(age+time); p1, p2, p3:4, p5+; λ1, λ2:3, λ4+} (model 5.8). Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale). to a considerable improvement in model fit (∆AIC=347) when compared to the constant model (model 5.1), however the φ(age∗time) model (model 5.7) performs considerably better than φ(age+time) (compare models 5.7 and 5.8, ∆AIC=71). Nevertheless, an advantage of the age+time model is that it allows each of the age components of the parameter to vary temporally whilst avoiding the inclusion of a large number of parameters (see Figure 5.3). Whereas the age∗time model has a total of 168 parameters, the age+time model has only 64 parameters (see Table 5.2). Another useful feature of the age+time model is the decreased tendency for boundary estimates, as illustrated in Figures 5.2 and 5.3.

100 Table 5.3: AIC values (after subtracting 44 000) and numbers of parameters K for models involving time dependence in the recapture probability. All models have survival and recovery structures of φ1, φ2, φ3, φ4:8, φ9+, λ1, λ2:3, λ4+.

Model K AIC 5.1 p1, p2, p3:4, p5+ ...... 12 1034 5.9 p1(time), p2, p3:4, p5+ ...... 46 1039 5.10 p1, p2(time), p3:4, p5+ ...... 45 980 5.11 p1, p2, p3:4(time), p5+ ...... 44 931 5.12 p1, p2, p3:4, p5+(time) ...... 42 848 5.13 p1,2,3:4,5+(age∗time) ...... 141 652 5.14 p1,2,3:4,5+(age+time) ...... 51 687

I next consider time dependence for each component of the recapture (models 5.9–5.12 in Table 5.3) and the recovery probabilities (models 5.15–5.17 in Table 5.4) in turn, while keeping each age component of φ constant. Using constant rather than time-dependent age components of φ will cause bias in the estimates of p or

λ, since φ1 will most likely be low in a bad year, and so it will be overestimated by a constant φ1 model. (An overestimated φ1 would result in an unrealistically low estimate for p1, since it assumes that more first year birds survive and are available for recapture.) However, this leakage is not a cause for concern, since the results in Tables 5.3 and 5.4 are used only as an indication of potential temporal variation in the age components of p or λ.

Temporal variation is not as important for p as it was for juvenile survival φ1 (compare model 5.2 with models 5.9–5.12 in Table 5.3). Indeed, time dependence for p1 is no improvement on the constant model (compare models 5.1 and 5.9). This is not surprising since newly-fledged birds spend most of their time away from the colony until their first moult (Reilly and Cullen, 1981) (resulting in large standard errors for p1 in Figure 5.4), and the varying conditions from year to year would be unlikely to affect the probability of recapturing birds at one year of age.

The recapture probabilities for older birds, particularly p5+, show more time dependence than that of the younger birds (compare models 5.9–5.12). Established breeders spend more time in the colony than younger birds. Since varying conditions would affect their breeding success (or determine whether or not they breed at all), their recapture probability would be more likely to vary from year to year.

101 Table 5.4: AIC values (after subtracting 44 000) and numbers of parameters K for models involving time dependence in the recovery probability. All models have survival and recapture structures of φ1, φ2, φ3, φ4:8, φ9+, p1, p2, p3:4, p5+.

Model K AIC

5.1 λ1, λ2:3, λ4+ ...... 12 1034 5.15 λ1(time), λ2:3, λ4+ ...... 46 940 5.16 λ1, λ2:3(time), λ4+ ...... 45 1018 5.17 λ1, λ2:3, λ4+(time) ...... 43 974 5.18 λ1,2:3,4+(age∗time) ...... 110 863 5.19 λ1,2:3,4+(age+time) ...... 48 937

The model p(age∗time) (model 5.13), allows all age components of the recapture probability to vary with time. Varying all of the recapture rates independently with time significantly increases model fit (compare AICs for models 5.1 and 5.13). The recapture probability for penguins aged five years and over (p5+) reduced in the early 1980’s when the number of study sites was increased significantly (see Figure

5.4). Since then, p5+ has increased steadily as the search effort has intensified. An age∗time model for the recapture probability is an improvement on p(age+time) (compare models 5.14 and 5.13, ∆AIC=35). However, the gains are not as signif- icant for the recapture probability as for the survival probability (models 5.8 and 5.7, ∆AIC=71). Some of the temporal variation in the recapture probability would be due to changes in experimental methods such as the frequency of burrow vis- its. Changing experimental methods would probably affect all age groups similarly, making an age+time model for p a reasonably good approximation of the true sit- uation. However, when conditions are particularly poor, birds of breeding age may not breed or may have little success. In this case, the effect of the conditions on the recapture probability could be different for breeders and pre-breeders. This con- trasts with survival, where changing environmental conditions are likely to affect the young and vulnerable birds more than experienced birds. The introduction of time dependence for the recovery probabilities of the various age groups leads to improved model fit for birds in their first year of life (model 5.15 in Table 5.4, ∆AIC=94), and to a lesser extent the adults of age 4 years and over (model 5.17, ∆AIC=60), when compared to the model with all recovery probabilities held constant (model 5.1). First year birds are not often recaptured,

102 Recapture Probabilities

0.2

0.1 Age 1

0 1970 1975 1980 1985 1990 1995 2000 1

0.5 Age 2

0 1970 1975 1980 1985 1990 1995 2000 1

0.5 Age 3:4 0 1970 1975 1980 1985 1990 1995 2000 1

0.5 Age 5+ 0 1970 1975 1980 1985 1990 1995 2000 Calendar Year

Figure 5.4: Recapture probabilities versus time for model φ1, φ2, φ3, φ4:8, φ9+, p1,2,3:4,5+(age∗time), λ1, λ2:3, λ4+ (model 5.13). Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale). but they are recovered dead on the western Victorian or South Australian coastlines (see Figure 4.1 on p.65). Although there are gains to be made by allowing all age components of the recovery rate to vary with time simultaneously (compare models 5.1 and 5.18, with ∆AIC=171, and see Figure 5.5), the recovery probability does not exhibit strong time dependence. An age+time model for recovery does not perform better than a model with time-dependent λ1 only (models 5.15 and 5.19 in Table 5.4, ∆AIC=3). Perhaps varying environmental conditions affect the recovery of first year birds more than the older birds that are concentrated on Port Phillip Bay and Phillip Island.

5.2 Biological covariates

In order to explain the observed temporal variation in the survival probability, I consider the following biological parameters: mean annual weight at banding, mean annual number of chicks fledged per pair and mean annual laying date for first clutches. These biological covariates provide a measure of the yearly variation in

103 Recovery Probabilities 0.2

0.15

0.1

Age 0−1 0.05

0 1970 1975 1980 1985 1990 1995 2000

1

0.5 Age 1−3

0 1970 1975 1980 1985 1990 1995 2000

1

0.5 Age 3+

0 1970 1975 1980 1985 1990 1995 2000 Calendar Year

Figure 5.5: Recovery probabilities versus time for model φ1, φ2, φ3, φ4:8, φ9+, p1, p2, p3:4, p5+, λ1,2:3,4+(age∗time) (model 5.18). Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale). the quality of environmental conditions (such as food availability, weather, etc.) for the parents. In this chapter, the biological parameters have been used as time- varying group covariates, whereas Chapter 6 considers individual covariates for the model parameters. When allowing the first-year survival to depend on the biological covariates, age+time models are imposed on the recapture and recovery probabilities, and the remaining components of the survival probability (in line with the results of Tables

5.1, 5.3 and 5.4). That is, I consider models of the form {φ1(V), φ2,3,4:8,9+(age+time), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)}. In this way, φ, p and λ are allowed to vary with time (so that bias in the survival is minimised), but such a model has fewer parameters than an age∗time structure. The Wald test is used to check the significance of the regression coefficients (ie. whether the regression coefficients are significantly different from zero) and the Likelihood Ratio Test (LRT) compares the covariate-dependent model with the constant case.

104 Making φ1 depend on each of the biological covariates in turn leads to some improvement in model fit when compared to the model with constant first-year survival (see AICs for models 5.22–5.28 in Table 5.5).

The mean annual laying date (model 5.24) is the best predictor of φ1, followed by the mean annual chick weight at banding (model 5.23) and mean number of chicks fledged per pair (model 5.22). The regression coefficient corresponding to mean laying date is negative, indicating that φ1 decreases with mean laying date. That is, early breeding is associated with high first-year survival (see Figure 5.6). This result is in accordance with Reilly and Cullen (1982) who found that birds banded in the first half of the season were more likely to survive their first year of life than birds banded later in the season. First-year survival increases with mean annual weight at banding (as found by Reilly and Cullen, 1982; Dann, 1988) and with number of chicks produced per pair (see Figures 5.7 and 5.8). This is not surprising, since “good years” (that is, years with favourable conditions) should produce heavier chicks, more chicks that survive to fledging, and fledglings that are more likely to survive their first year of life. The pair of covariates that best fits the data (in terms of AIC) are weight at banding and laying date (model 5.27). However, the time variation in first-year survival is not adequately explained by the variation in these biological covariates (compare models 5.22–5.28 with model 5.21). Hence there must be other unmodelled factors which contribute to the tem- poral variation in survival. In fact, number of chicks per pair, mean banding weight and mean laying date together account for only 28% of the time variation (compare ∆AICs between models 5.20 and 5.21, and 5.20 and 5.28). It may be worthwhile considering quadratic models here also. I next consider covariate dependence for the adult survival probability. In the interests of parsimony (see Tables 5.1, 5.3 and 5.4), I again choose a model with age+time structures for p and λ and for the survival probability of birds up to the age of three years. There are several options for modelling φ4:8 and φ9+. The model φ4+(V) would be unsuitable if there was an interaction between the age groups and the covariates. Since the effect of a biological covariate is likely to be the same for φ4:8 and φ9+, a model such as φ4+(age+V) should be sufficient, rather

105 Table 5.5: AIC values (after subtracting 44 000) and regression coefficients (with standard errors in parentheses) for models involving biological covariates for φ1. The model is {φ1(V), φ2,3,4:8,9+(age+time), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)}, where V=chicks per pair (chpp), weight at banding (bw), mean laying date (mld), and so on. The p-values for the regression coefficients (from Wald tests) are stated if < 0.05.

Covariate for φ1(V) Regr. coeff. (s.e.) p-value AIC 5.20 constant — 436 5.21 time — 135 5.22 chpp 0.2905 (0.0993) p=0.0034 430 5.23 bw 3.3189 (0.4829) p< 10−11 393 5.24 mld −0.4812 (0.0585) p< 10−15 373 5.25 chpp+bw chpp: −0.0079 (0.1072) 395 bw: 3.3382 (0.5489) p< 10−8 5.26 chpp+mld chpp: −0.1208 (0.1072) 374 mld: −0.5174 (0.0668) p< 10−14 5.27 bw+mld bw: 2.2293 (0.5099) p< 10−4 356 mld: −0.3894 (0.0618) p< 10−9 5.28 chpp+bw+mld chpp: −0.3167 (0.1114) p=0.0045 350 bw: 2.7970 (0.5470) p< 10−6 mld: −0.4618 (0.0665) p< 10−11

0.5

0.4

0.3

0.2 First year survival probability 0.1

0 1 Sep 1 Oct 1 Nov 1 Dec Mean laying date

Figure 5.6: First-year survival probability versus mean laying date under model 5.24. Yearly estimates of survival under a φ1(time) model (model 5.2) are also shown. Vertical bars represent 1 standard error each side of the estimate (back- transformed from the logistic scale).

106 0.5

0.4

0.3

0.2 First year survival probability

0.1

0 750 800 850 900 950 1000 1050 1100 Mean annual weight at banding (grams)

Figure 5.7: First-year survival probability versus mean weight at banding under model 5.23. Yearly estimates of survival under a φ1(time) model (model 5.2) are also shown. Vertical bars represent 1 standard error each side of the estimate (back- transformed from the logistic scale).

0.5

0.4

0.3

0.2 First year survival probability

0.1

0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Mean number of chicks fledged per pair

Figure 5.8: First-year survival probability versus mean number of chicks fledged per pair under model 5.22. Yearly estimates of survival under a φ1(time) model (model 5.2) are also shown. Vertical bars represent 1 standard error each side of the estimate (back-transformed from the logistic scale).

107 than φ4+(age∗V). Therefore the basic model considered here is {φ1,2,3(age+time),

φ4:8,9+(age+V), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)}. The mean weight at banding, the mean annual number of chicks fledged per pair and the mean laying date in the previous breeding season are indicators of the conditions for penguins in that year. Using number of chicks fledged per pair and mean laying date as covariates for adult survival are no improvements on the constant model (compare model 5.29 with models 5.31 and 5.33 in Table 5.6). It is possible that mean banding weight might influence annual adult survival (LRT, p- value=0.014). However, the regression coefficient corresponding to banding weight is negative, indicating that low adult survival is associated with higher mean annual banding weight. This does not make biological sense, since high banding weights are indicative of good conditions for penguins, which should result in enhanced survival. Therefore, the relatively low statistical significance of this result, together with the counter-intuitive negative regression coefficient suggests that it may not be biologically significant. As expected, the dependence on the biological covariates is much stronger for first-year survival than for adult survival: inexperienced juveniles are more affected by varying environmental conditions than adults. In their study of population regulation in Little Penguins, Dann and Norman (2006) found that for larger breeding sites such as Phillip Island, the population size of the colony is affected by food availability rather than available nesting area. In order to determine whether the variation in the size of the penguin population on Phillip Island had any effect on the survival of the birds, I consider the mean annual number of penguins crossing the beach at the Penguin Parade as a covariate for first year and adult survival. These data were based on counts taken in the first 50 minutes of the nightly Penguin Parade from the 1977 penguin year onwards. The same base model is used as in the biological covariate analysis. There is no evidence of an effect of density dependence on first-year survival (see Table 5.7). However, this measure of population density may not accurately reflect the yearly variation in the population size of the Little Penguin colony on Phillip Island. Furthermore, it might be worth considering the mean number of beach crossings in the breeding season before the chicks are born as a covariate for first-year survival, since high numbers of adults in the previous season may exhaust

108 Table 5.6: AIC values (after subtracting 44 000) and regression coefficients (with standard errors in parentheses) for models involving biological covariates for adult survival. The model is {φ1,2,3(age+time), φ4:8,9+(age+V), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)}. That is φ4:8 and φ9+ both depend on the covariate but they vary in parallel (on a logistic scale). The p-values for the regression coefficients (from Wald tests) are stated if < 0.05.

Covariate for φ4:8,9+(age+V) Regr. coeff. (s.e.) p-value AIC 5.29 constant — 216 5.30 time — 157 5.31 chpp −0.1693 (0.1300) 217 5.32 bw −1.8243 (0.7422) p=0.0140 212 5.33 mld −0.0914 (0.0779) 217 5.34 chpp+bw chpp: 0.0455 (0.1658) 214 bw: −1.9787 (0.9287) 5.35 chpp+mld chpp: −0.3151 (0.1467) 214 mld: −0.1802 (0.0858) 5.36 bw+mld bw: −2.6705 (0.8071) p=0.0009 208 mld: −0.2147 (0.0863) 5.37 chpp+bw+mld chpp: −0.0747 (0.1734) 210 bw: −2.4511 (0.9572) mld: −0.2236 (0.0881)

Table 5.7: AIC values (after subtracting 38 000) and regression coefficients (with standard errors in parentheses) for models with mean annual number of beach crossings in the first 50 minutes of the Penguin Parade as a covariate for φ1. The model is {(φ1(V), φ2,3,4:8,9+(age+time), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)}. The following models are based on data from 1977 onwards.

Covariate for φ1(V) Regr. coeff. (s.e.) AIC 5.38 constant — 306 5.39 beach crossings 0.0900 (0.3256) 308 the food supply in the vicinity of Phillip Island, making it difficult for parents to find food for their chicks in the next season. I next consider the mean number of beach crossings as a covariate for adult survival (see Table 5.8). The results are somewhat contradictory: the negative regression coefficient is highly significant (p< 10−10, Wald test), but model 5.41 does not fit the data significantly better than model 5.40 (p=0.083, LRT). Hence there is some evidence to suggest that higher mean numbers of adults in the Penguin Parade

109 Table 5.8: AIC values (after subtracting 38 000) and regression coef- ficients (with standard errors in parentheses) for models with mean annual number of beach crossings in the first 50 minutes of the Penguin Parade as a covariate for adult survival. The model is {φ1,2,3(age+time), φ4:8,9+(age+V), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)}. The following models are based on data from 1977 onwards.

Covariate for φ4:8,9+(age+V) Regr. coeff. (s.e.) AIC 5.40 constant — 166 5.41 beach crossings −2.1615 (0.3261) 165 area are associated with low adult survival. Therefore, the effect of population pressure on penguin survival remains a topic for future study.

5.3 Climatic conditions

Studies such as Mickelson et al. (1992) and Chambers (2004a,b) have shown that climatic conditions (such as the sea surface temperatures in regions near Phillip Island) have marked effects on various Little Penguin breeding parameters. In this section I will examine the effect of climatic conditions on the juvenile and adult survival probabilities of these birds. As in Chambers (2004b), I consider the local sea surface temperatures (SSTs) in the Bass Strait region, as well as a global-scale measurement, the Southern Oscillation Index (SOI). In order to minimise the bias in the survival probability and keep the numbers of parameters manageable, I consider models of the form {φ1(V), φ2,3,4:8,9+(age+time), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)} when modelling covariate dependence for first-year survival (see explanation p.104), and

{φ1,2,3(age+time), φ4:8,9+(age+V), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)} when modelling covariate dependence for adult survival (see p.108).

5.3.1 Local sea surface temperatures

Local SST values may have a direct effect on the survival of young birds, or affect them indirectly, via an increase or decrease in the abundance of their prey. In addition, SSTs may influence the availability of food for the parents, thus affecting a young bird’s condition at fledging. The highly complex nature of the potential effect of sea temperatures and other environmental factors on Little Penguins, and on the abundance of their prey species (such as pilchards and anchovies), make

110 it virtually impossible to predict a relationship between survival and SST with any certainty. Therefore, my intention in this section is to investigate the possible dependence of the annual survival probabilities of first year birds and adults on the SSTs in various regions of Bass Strait and with various time lags (see p.112), with the aim of generating rather than validating hypotheses. The SST data used in this study are based on the U.S. National Centers for Environmental Prediction (NCEP) data (up to 2001) and corrected satellite data (Reynolds and Smith, 1994) in recent years, and have a 1 degree x 1 degree resolution (for further explanation of the SST data, see Chambers, 2004b). The grid of SST values runs from 138◦E to 152◦E and from 35◦S to 45◦S. For each month from January 1949 to May 2004, there are 10 rows and 14 columns of data, corresponding to the monthly mean SSTs at the midpoints of each grid square. Cells where the centres of the grid boxes are on land are ignored. I begin by determining the appropriate regions of SST to be used as covariates for first year and adult survival. Figure 4.1 on p.65 shows that pre-breeders were recovered dead over a wide region of the coastline of southern Australia, as far west as Kangaroo Island in South Australia, whereas breeding adults were found dead mostly in Port Phillip Bay and Phillip Island (see Figure 1.1 on p.2). As discussed in Chapter 4, radio-tracking studies of Phillip Island penguins (Weavers, 1992; Collins et al., 1999) have shown that the recoveries of dead birds reflect the movements and distribution of the live birds. Therefore, the locations of dead recoveries for first year and adult birds will be used to determine the regions of SST to be considered as covariates for first year and adult survival. For first-year survival, I use the SST values from the grid boxes adjacent to the coastline from Kangaroo Island to Wilson’s Promontory (see the shaded region in Figure 5.9), whereas for adult survival I use a smaller region centred around Port Phillip Bay. Since the area of ocean adjacent to 142 degrees of longitude is very productive, and currents move this water across Bass Strait (P. Dann, pers. comm.), SSTs in these grid boxes have been included in my analysis for adult survival. Hence for adult survival I consider the SSTs adjacent to the coastline from Warrnambool to Wilson’s Promontory (see the shaded region in Figure 5.10). I also consider a second, smaller region, consisting of the grid boxes closest to Port Phillip Bay and Phillip Island (the shaded region

111 Figure 5.9: SST region to be used as a covariate for first-year survival, φ1. in Figure 5.11), as covariates for adult survival. I have chosen to average the SST values over the appropriate region rather than using a statistical data reduction technique, such as principal component analysis or cluster analysis (L. Chambers, pers. comm.), since the SST values over the entire region are meaningful. Since the SSTs in the regions described above are highly correlated (see Section 5.4), I could use any of the regions as covariates for φ1 and φ4+. However, I have chosen to use the regions appropriate to first year and adult birds since these make biological sense. The next issue to be addressed is which months (or groups of months) of SSTs should be used as covariates for annual survival. It is possible that the penguins’ food source may move to places with warmer or cooler water (Bunce, 2000), or that there may be a “lag” of several months in the effect of sea surface temperature on survival, since the water temperature may affect the production of fish food such as plankton, and hence the abundance of fish. Mickelson et al. (1992) found that increased weights of adults were associated with decreased sea temperature gradients four months earlier. Therefore, to explain the temporal variation in φ1

112 Figure 5.10: SST region to be used as a covariate for adult survival, φ4+.

Figure 5.11: Smaller SST region, close to Phillip Island and Port Phillip Bay, to be used as a covariate for adult survival, φ4+.

113 Table 5.9: AIC values (after subtracting 44 000) and regression coefficients (with standard errors in parentheses) for models involving seasonal sea surface temper- atures in the region specified in Figure 5.9 as covariates for φ1. The model is {φ1(V), φ2,3,4:8,9+(age+time), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)}.

Covariate for φ1(V) Regr. coeff. (s.e.) AIC 5.20 constant — 436 5.21 time — 135 5.44 SST summer previous year 0.2782 (0.0591) p< 10−5 416 p< 10−5 5.45 SST autumn previous year 0.2495 (0.0625) p< 10−4 422 p< 10−4 5.46 SST winter previous year −0.2604 (0.1212) 433 5.47 SST spring previous year −0.0468 (0.1000) 437 5.48 SST summer current year −0.0495 (0.0652) 437 5.49 SST autumn current year 0.2121 (0.0650) p=0.0011 427 p=0.0013 5.50 SST winter current year −0.0187 (0.1205) 438 5.51 SST spring current year 0.0124 (0.1053) 438 in calendar year 1983, say, I could consider the SST for the appropriate region in January 1982, then February 1982, . . ., December 1983. However, to lessen the amount of data dredging and the possibility of obtaining spurious results, I will consider the effect of the mean seasonal SSTs for spring ( September–November), summer (December–February), autumn (March–May) and winter (June–August), rather than the monthly mean, on the survival probability. That is, I will consider, as covariates for φ1 in calendar year 1983, the mean SST for each season from summer 1981/1982 to spring 1983. I refer to these as “SST in the summer of the previous year” through to “SST in the spring of the current year”. Indeed, a seasonal SST analysis may be more appropriate than a monthly analysis, since mortality events tend to span months, but are less likely to span seasons. For example, Dann et al. (1992) found that adult Little Penguins are most vulnerable to death after moult (ie. March or April) and then again in early spring (September) when they sometimes succumb to starvation, while the peak mortality period for first year birds is in March or April, three or four months after fledging.

Dependence of first-year survival on SST Table 5.9 displays the AICs, and the regression coefficients (and standard errors) when the mean seasonal SST values are used as covariates for first-year survival, φ1.

The strongest relationship is between φ1 and the mean SST in the summer of the

114 0.6

0.5

0.4

0.3

0.2 First year survival probability

0.1

0 6 6.5 7 7.5 8 Mean SST for summer in previous year

Figure 5.12: The relationship between time dependent first-year survival under model {φ1(time), φ2,3,4:8,9+(age+time), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)} (model 5.21) and the sea surface temperature for the summer in the previous year. The line shows predicted values for first-year survival when the SST in the summer of the previous year is used as a covariate (model 5.44).

−5 previous year (p < 10 ) (see Figure 5.12). φ1 is also associated with the autumn SSTs in the previous year and in the current year (p< 10−4 and p=0.0011 respec- tively). The significant regression coefficients are positive in each case, suggesting that increased survival of first year birds is associated with warmer sea surface temperatures in those seasons. Perhaps warm SSTs in the summer prior to penguin breeding aides the larval development of the fish, resulting in a good stock of juvenile fish for young penguins to feed on after fledging (P. Whittington, pers. comm.). Several Australian studies have shown increases in the abundance and movement of prey species during warm water periods. Fortescue (1998), in a study of Little Penguins from Bowen Island, reported an increase in the abundance of various fish species during periods of warmer water temperatures in Jervis Bay, New South Wales. Bunce (2000) stated that pilchards (a prey species of Little Penguins) move into Port Phillip Bay during warm water periods. Chambers (2004b) highlights the reasons for the potential differences in the effect of warmer SSTs on seabirds in

115 Equatorial regions and regions of higher latitude such as southern Australia (see explanation in Chapter 2 on p.19). Chambers (2004a) who studied the effect of climate on the timing of Little Pen- guin breeding, found negative correlations between the mean laying date for Little Penguins and the SSTs in the summer and autumn prior to breeding, suggesting that “penguins tend to breed earlier when sea temperatures are warm during these periods”. Furthermore, Reilly and Cullen (1981) and Chambers (2004b) found that an early start to the breeding season is associated with a more successful breeding season (that is, higher banding weights and more chicks fledged per pair), and Sec- tion 5.2 (p.103) showed that first year birds survive better in years in which breeding occurs earlier. Chambers (2004a) provides the following explanation regarding the significance of the SSTs in the previous summer and autumn: “obtaining a good supply of food during the end of the previous breeding season (summer) and prior to and during moult (autumn) places the penguins in a good position for the sub- sequent breeding season”. Hence, if the parents are in better condition prior to the breeding season, they would be more likely to produce healthier chicks that have a higher probability of surviving their first year of life. Therefore, the positive associa- tions between φ1 and the SSTs in the summer and autumn of the previous year that have been obtained here, correspond closely to the results of the aforementioned Australian studies.

The significant relationship between φ1 and the autumn SST in the current year also makes biological sense. Most chicks are banded slightly before fledging in December–February (see p.54). By the beginning of autumn, almost all of the newly-fledged birds from the previous breeding season will have gone to sea. There- fore it is not surprising that autumn is the peak mortality period for these young birds that are inexperienced in finding food (Dann et al., 1992), and that the au- tumn sea surface temperatures would affect the survival of first year birds. However, the association between warmer water temperatures and enhanced survival (this study), or more successful breeding (Chambers, 2004a), is in contrast to the results of some other studies undertaken in Australia and New Zealand. Mickelson et al. (1992) reported that a decreased east–west sea temperature gradient across Bass Strait is associated with earlier breeding, heavier chicks and more chicks

116 Table 5.10: AIC values (after subtracting 44 000) and regres- sion coefficients (with standard errors in parentheses) for mod- els involving seasonal sea surface temperatures in the region speci- fied in Figure 5.10 as covariates for adult survival. The model is {φ1,2,3(age+time), φ4:8,9+(age+V), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)}.

Covariate for φ4:8,9+(age+V) Regr. coeff. (s.e.) AIC 5.29 constant — 216 5.30 time — 157 5.54 SST summer previous year 0.1327 (0.0813) 216 5.55 SST autumn previous year 0.0506 (0.0731) 218 5.56 SST winter previous year −0.1374 (0.1169) 217 5.57 SST spring previous year 0.1511 (0.1005) 216 5.58 SST summer current year −0.1704 (0.0842) p=0.0431 214 p=0.0455 5.59 SST autumn current year 0.1670 (0.0746) p=0.0252 213 p=0.0253 5.60 SST winter current year 0.3036 (0.1154) p=0.0085 211 p=0.0082 5.61 SST spring current year 0.0967 (0.1018) 217 per pair. They claimed that this effect was due to cooler waters having slightly higher concentrations of nutrients and, as a result, an increase in abundance of penguin prey species. In their study of the effect of climate fluctuations on Little Penguins in New Zealand, Perriman et al. (2000) found that La Ni˜na conditions (with higher than average sea temperatures in New Zealand) were associated with later breeding and a reduction in double breeding. So far I have considered the SST in the chick region (see Figure 5.9) as a covariate for φ1. However, since the strongest relationship between φ1 and SST is for the summer in the previous year, that is, around 12 months before the birds have fledged, the SST must affect the parents rather than affecting the chicks directly. Therefore, particularly in the months prior to fledging, I would expect the adult

SST region (Figure 5.10) to be a better covariate for φ1 than the chick SST region. Indeed this is the case — the SSTs for the summer and autumn in the year prior to fledging are highly significant (with p< 10−9 and p< 10−5 respectively), more significant than for the chick region. However, since the SSTs in the adult and chick regions are highly correlated (see Section 5.4), this is not worth pursuing further.

117 Dependence of adult survival on SST Table 5.10 displays the AICs, and the regression coefficients (and standard errors) when the mean seasonal SST (in the region shaded in Figure 5.10) is used as a covariate for adult survival. The strongest relationships are for the current autumn and winter (p=0.025 and p=0.009 respectively, Wald tests), and to a lesser extent, the current summer (p=0.043). That is, the adult survival probability in the 1972 calendar year is best explained by the mean SSTs in the autumn and winter of 1972, and so on. The reasons for expecting these relationships follow. During the moult in February or March, adults fast for 15–20 days, and then go back to sea to feed (Reilly and Cullen, 1983). Since penguins are quite vulnerable after moulting, variations in SST in autumn may affect the well-being of these birds. Dann et al. (1992) report that adult mortality peaks in autumn after moult, and again in early spring. Therefore it is likely that the winter SST also influences penguin survival. However, although the relationships are significant, the relatively low significance and the degree of data dredging that has gone on would lead to doubts about the reality of this association. Furthermore, the regression coefficient for summer is negative, suggesting that cooler SSTs in summer are associated with increased adult survival probability, which is the reverse of the relationship between SST and first-year survival. In fact, if we use instead the region just outside Phillip Island and Port Phillip Bay, the p-value for winter increases from 0.009 to 0.012 — further reason to doubt the reality of this association. It is interesting to note that the larger SST region for adults (Figure 5.10) explains slightly more of the temporal variation than the smaller region (Figure 5.11). Hence, the surface temperature of the highly productive region of ocean near 142 degrees of longitude may be a determinant of adult survival (as hypothesized by P. Dann, pers. comm.).

5.3.2 Southern Oscillation Index

The Southern Oscillation Index (SOI) measures the fluctuation in the air pressure difference between Tahiti and Darwin (Australian Bureau of Meteorology). The SOI values lie between −35 and 35. El Ni˜no episodes are characterised by sus- tained negative SOI values, and are associated with weaker Pacific Trade winds,

118 Table 5.11: AIC values (after subtracting 44 000) and regression coeffi- cients (with standard errors in parentheses) for models involving the sea- sonal Southern Oscillation Indices (SOIs) as covariates for φ1. The model is {φ1(V), φ2,3,4:8,9+(age+time), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)}.

Covariate for φ1(V) Regr. coeff. (s.e.) AIC 5.20 constant — 436 5.62 SOI summer previous year −0.0065 (0.0036) 434 5.63 SOI autumn previous year 0.0031 (0.0033) 437 5.64 SOI winter previous year 0.0049 (0.0039) 436 5.65 SOI spring previous year −0.0099 (0.0034) p=0.0034 429 p=0.0038 5.66 SOI summer current year −0.0184 (0.0035) p< 10−6 410 p< 10−6 5.67 SOI autumn current year −0.0090 (0.0033) p=0.0062 430 p=0.0065 5.68 SOI winter current year −0.0022 (0.0040) 437 5.69 SOI spring current year −0.0127 (0.0037) p< 10−3 426 p< 10−3 reduced rainfall in eastern and northern Australia, and a cooler SST in Australian waters. Sustained highly positive SOIs are indicative of a La Ni˜na episode, which is associated with stronger Pacific Trade winds, an increase in rainfall in eastern and northern Australia, and a warmer SST. In this analysis, I average the monthly SOI data from the Australian Bureau of Meteorology1, to obtain seasonal SOIs. The seasonal SOIs are then used as covariates for φ1 and φ4+. Tables 5.11 and 5.12 display the AICs, and the regression coefficients (and standard errors) corresponding to the SOI, when the mean seasonal SOIs are used as covariates for first year and adult survival respectively. The strongest relationship is between first-year survival and the SOI for the

−6 current summer (p< 10 , Table 5.11, Figure 5.13). That is, φ1 in the 1969 calendar year is associated with the SOI in the summer of 1968/1969. It is possible that there is a greater lag in the effect of global-scale climatic conditions, such as the SOI, than in the local SST, since φ1 in the 1969 calendar year was strongly associated with the SST in autumn 1969. In fact, there is a significant, positive correlation between the SOIs in summer and the SSTs in autumn (correlation coefficient=0.374, p=0.025). The significant regression coefficients are negative, which means that a higher first-year survival probability is associated with lower SOIs. Highly negative SOIs indicate the presence of El Ni˜no conditions, which are generally associated

1http://www.bom.gov.au/climate/glossary/soi.shtml

119 0.7

0.6

0.5

0.4

0.3

0.2 First year survival probability

0.1

0 −30 −25 −20 −15 −10 −5 0 5 10 15 20 Mean SOI for summer in current calendar year

Figure 5.13: The relationship between time dependent first-year survival under model {φ1(time), φ2,3,4:8,9+(age+time), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)} (model 5.21) and the southern oscillation index for the summer in the current year. The line shows predicted values for first-year survival when the SOI in the current summer is used as a covariate (model 5.66).

Table 5.12: AIC values (after subtracting 44 000) and regression coefficients (with standard errors in parentheses) for models involving the seasonal South- ern Oscillation Indices (SOIs) as covariates for adult survival. The model is {φ1,2,3(age+time), φ4:8,9+(age+V), p1,2,3:4,5+(age+time), λ1,2:3,4+(age+time)}.

Covariate for φ4:8,9+(age+V) Regr. coeff. (s.e.) AIC 5.29 constant — 216 5.70 SOI summer previous year −0.0024 (0.0051) 218 5.71 SOI autumn previous year 0.0081 (0.0045) 215 5.72 SOI winter previous year 0.0143 (0.0055) p= 0.0088 211 p=0.0091 5.73 SOI spring previous year 0.0121 (0.0049) p=0.0135 212 p=0.0128 5.74 SOI summer current year 0.0081 (0.0052) 216 5.75 SOI autumn current year 0.0048 (0.0046) 217 5.76 SOI winter current year 0.0010 (0.0056) 218 5.77 SOI spring current year 0.0043 (0.0048) 217 with lower SSTs in the Australian region. This result appears to contradict those obtained for local-scale sea surface temperatures, again highlighting the complexity of the effect of climatic conditions on the survival of Little Penguins. When the SOIs are considered as covariates for adult survival (Table 5.12), the only significant results are for winter and spring in the previous year. That is,

120 φ4+ in the 1972 calendar year is associated with the SOIs in winter and spring in 1971. The significant regression coefficients are positive, whereby a higher adult survival probability is associated with higher SOIs. Highly positive SOIs indicate the presence of La Ni˜na conditions, which are generally associated with warmer SSTs in the Australian region. This result corresponds to that of Fortescue (1998) who found that positive phases of the SOI were associated with increased adult survival from one season to the next. It is interesting that the regression coefficients corresponding to SOI for φ1 and φ4+ again reverse sign, as they did for the SSTs. Future work will consider maximum, minimum or threshold values for SSTs and SOIs, rather than the raw values, since extremely high or low SOIs (La Ni˜na or El Ni˜no conditions) may affect the survival, whereas moderate SOIs may have no effect.

5.4 Correlations between covariates

The correlation coefficients for the covariates considered in this chapter appear in Table 5.13. As expected, the SSTs in the chick and adult regions are highly positively correlated. The SSTs in successive seasons (and the SOIs in successive seasons) are also positively correlated. There is a positive correlation between mean numbers of chicks fledged per pair and weight at fledging, and a negative correlation between number of chicks per pair and mean laying date, that is more chicks per pair is associated with earlier laying. The autumn SST in the previous year is correlated positively with the number of chicks per pair and negatively with the mean laying date. That is, warm autumn sea surface temperatures in the previous breeding season are associated with more chicks per pair and earlier laying (as found in Chambers, 2004a). Finally, autumn SSTs are positively correlated with autumn SOIs.

121 Table 5.13: Correlation coefficients between covariates. soi sst ch sst ad bw mld sum aut win spr sum aut win spr sum aut win spr chpp 0.568 −0.469 0.017 0.182 0.186 0.058 0.204 0.437 0.250 0.042 0.222 0.451 0.268 0.006 bw −0.354 −0.098 −0.022 −0.035 −0.172 0.196 0.325 0.136 −0.090 0.162 0.341 0.101 −0.158 mld −0.196 −0.240 −0.102 −0.044 −0.265 −0.582 −0.140 0.050 −0.320 −0.558 −0.134 0.081 sum 0.600 −0.034 −0.009 0.316 0.455 0.334 0.133 0.275 0.374 0.240 0.106 soi aut 0.574 0.555 0.247 0.497 0.307 0.157 0.276 0.466 0.320 0.239 122 win 0.811 0.099 0.084 0.059 0.135 0.106 0.029 −0.003 0.115 spr 0.097 0.062 0.224 0.301 0.132 0.025 0.169 0.248 sum 0.543 0.310 0.265 0.965 0.492 0.206 0.185 sst ch aut 0.488 0.122 0.576 0.966 0.494 0.106 win 0.544 0.283 0.458 0.853 0.401 spr 0.250 0.153 0.447 0.883 sum 0.561 0.272 0.263 sst ad aut 0.547 0.191 win 0.502 5.5 Conclusion

This chapter shows that several of the age components of the model parameters exhibit considerable temporal variation. As expected, time dependence is stronger for first-year survival than for the survival of the older groups, since changing envi- ronmental conditions (such as food availability) are likely to affect the young and vulnerable birds more than experienced birds. Temporal variation is not as important for the recapture probability as it was for juvenile survival. Some of the time variation in the recapture probability would be due to changing experimental methods (such as the frequency of burrow visits), which would most likely affect all age groups similarly. However, the older birds show more time dependence in recapture probability than do the first year birds. This is not surprising, since newly–fledged birds spend most of their time away from the colony, and so varying conditions would be unlikely to affect the probability of recapturing birds at one year of age. In contrast, varying conditions would affect the breeding success of older birds (or determine whether or not they breed at all), and so their recapture probability would be more likely to vary from year to year. The recovery probability does not exhibit strong time dependence. First-year survival increases with number of chicks fledged per pair, increases with annual average fledging weight and decreases with mean laying date (that is, late breeding leads to low first-year survival). As expected, the dependence on the biological covariates is much stronger for first-year survival than for adult survival, that is, inexperienced juveniles are more affected by varying environmental conditions than adults. There is no evidence of an effect of density dependence on first-year survival. However, there is some evidence to suggest that higher numbers of adults in the Penguin Parade area are associated with low adult survival. An increased first-year survival probability was associated with warmer sea sur- face temperatures in the summer and autumn of the previous year and in the autumn after fledging. These results are in accordance with Chambers (2004a), who provides a strong biological justification for the observed relationships. Using seasonal SST as a covariate for adult survival does not produce any convincing results.

123 First-year survival is associated with the SOI in the previous summer. However, the regression coefficient is negative, meaning that a higher first-year survival is associated with lower SOIs, which generally produce lower SSTs in the Australian region. This result appears to contradict those obtained for local-scale SSTs, high- lighting the complexity of the effect of climatic conditions on the survival of Little Penguins.

124 Chapter 6

Individual Covariates

In Chapter 5 I found that several of the age components of the model parameters (particularly first-year survival) exhibited considerable tem- poral variation, and I sought to explain this variation by using extrinsic covariates that apply to all animals (eg. annual mean laying date, sea surface temperature). While conditions that apply to all animals may certainly affect the probability of a bird surviving, or being recaptured or recovered, it is likely that individual traits or characteristics would also affect these parameters. This chapter focuses on intrinsic covariates, that is, covariates that vary from one individual to another, and their effect on the survival probability. The following intrinsic covariates are considered: weight at banding, date of banding, weight at previous recapture occasion, bill depth, and sex. While some intrinsic covariates are time-invariant (e.g. weight at banding, banding date), or are recorded only once for each individual (e.g. bill depth or sex), others are time-varying and vary from one mark/recapture occasion to the next (eg. current body weight). Time-varying covariates typically contain missing values, a problem which is also addressed in this chapter.

125 6.1 Introduction

The following subsections introduce the individual covariates for survival that will be examined in this chapter.

6.1.1 Date of banding

Each penguin included in this study was banded as a chick between six and eight weeks of age, slightly before fledging, which occurs at eight to ten weeks of age (see Chapter 3). Although the dates of hatching and fledging are generally unavailable, banding dates are recorded for all birds. The banding date, measured as the number of days elapsed from 1 July to the day on which a bird was banded, provides an indication of its date of hatching. It should be noted that the banding date is a reflection of the bird rather than the behaviour of the bander (see Chapter 1). Although most birds were banded as chicks between November and February (see Figure 3.1 on p.54), the date of commencement of breeding, which depends on environmental conditions that affect the penguins, varies from year to year (Reilly and Cullen, 1981). A bird’s banding date anomaly which is defined as its banding date minus the mean banding date for all birds in that breeding season, provides a measure of how early or late a bird was banded (and hatched) relative to other chicks from that cohort. Also, since some breeding seasons are of longer duration than others, it may be useful to consider the standardized banding date anomaly, which is obtained by dividing the banding date anomaly by a measure of spread. Banding of the first cohort commenced in February 1968. Since banding began part-way through the 1967 breeding season, the mean banding date for this cohort is artificially high. Hence the 1967 cohort, which comprises only 17 chicks, is not included in the analyses in this chapter.

6.1.2 Weight at banding

A bird’s weight at banding provides an indication of its condition prior to fledging. Therefore it is plausible that the individual’s banding weight may affect its survival in its first year of life. However, since banding occurs at six to eight weeks of age, the chick’s weight may also depend on its exact (unknown) age at the time of banding.

126 A bird’s survival may also be affected by its banding weight anomaly (the banding weight minus the mean banding weight of chicks in that season), which indicates whether it was lighter or heavier than average for that season. Banding weights were recorded reasonably consistently from 1982 onwards. Of the 23 686 birds banded as chicks that are included in this analysis, 18 406 were weighed at the time of banding. The remaining 5280 birds are disqualified from this analysis.

6.1.3 Sex and bill depth

Unlike weight, bill depth does not depend on a bird’s condition in a particular season. Hence bill depth is likely to be a better index of the “size” of a bird than its weight. Adult Little Penguins can be sexed with an accuracy of 91% by using their bill depth measurements: a bird with a bill depth greater than 13.3mm is classified as a male, while a bird with a smaller bill depth is said to be a female (Arnould et al., 2004). Once sex is accounted for, the bill depth measurement provides an indication of the relative size of the bird, with bigger bill depth measurements corresponding to larger birds (P. Dann, pers. comm.). Bill depth measurements were also recorded for a small number of chicks at the time of banding. However, the bill depths cannot be used to determine the sex of immature birds, since their bills continue to grow until they reach breeding age (Stahel and Gales, 1987). Hence sex and bill depth will only be considered as individual covariates for adult survival. Only 336 birds have bill depth measurements recorded. However, 1686 birds have a sex entry recorded in the database, since many birds were sexed “by sight”, whereby the researchers determined the sex of each bird based on a visual appraisal of the size of the bill.

6.1.4 Current weight

An adult bird’s weight at the previous recapture occasion, henceforth referred to as its current weight, provides some indication of its condition and hence may affect its probability of survival. The usual assumption is that heavier birds have a higher chance of survival. However, there may be an upper limit (M. Cullen, pers. comm.), whereby large birds find it more difficult to catch their food and

127 maintain their bodies, especially under poor conditions. A bird’s survival may also be affected by its weight anomaly, which indicates whether it was lighter or heavier than average compared to other adults encountered during that season. Care must be taken when interpreting the results, since better conditions (eg. more fish) lead to heavier birds and improved chances of survival, but higher weight may not result in enhanced survival. Until the mid-1980s banded birds were not routinely weighed upon recapture as adults. Furthermore, since many birds have one or more entries of “0” in their life histories, there are a large number of missing weight values (see Section 6.2.2).

6.2 Calculating the likelihood

6.2.1 Standard likelihood incorporating individual covariates

While the analyses in Chapters 4 and 5 (and later in Chapter 8) calculate the likelihood by taking the product of the contributions for each yearly cohort of birds and each mark/recapture occasion (Catchpole et al., 2000), this chapter uses a conditional probability approach, looking separately at each individual and each occasion (Catchpole et al., 2004). Consider a mark-recapture-recovery experiment on n birds over k occasions, at times t1, t2,...,tk. If bird i is banded at ci, then the model parameters can be defined as follows:

φi,j = P(alive at tj+1 | alive at tj)

pi,j = P(recaptured at tj+1 | alive at tj+1)

λi,j = P(found dead during (tj, tj+1) | died during (tj, tj+1)), where j = ci,...,k − 1.

Sufficient statistics, D, V, W and Z, are defined as on p.68, but with row i of each matrix corresponding to bird i rather than the ith cohort of birds. Hence the likelihood used here is identical to that developed by Catchpole et al. (2000) (and derived in Appendix A.1), except subscript i now refers to bird i. The likelihood is

128 given by

n k−1 k di,j wi,j +zi,j wi,j zi,j vi,j L = const× {(1 − φi,j)λi,j} φi,j pi,j (1 − pi,j) χi,j , (6.1) i=1 "j=c j=c # Y Yi Yi

Here χi,j is defined as

χi,j = P(bird i not seen after tj | alive at tj), (6.2)

= (1 − φi,j)(1 − λi,j)+(1 − pi,j)φi,jχi,j+1, for 1 ≤ i ≤ n and 1 ≤ j ≤ k − 1. χi,j can be calculated recursively since χi,k = 1, as tk is the final mark/recapture occasion.

Furthermore, since χi,k =1 for 1 ≤ i ≤ n, equation 6.1 can be re-written as n k−1 di,j wi,j +zi,j wi,j zi,j vi,j L = const × {(1 − φi,j)λi,j} φi,j pi,j (1 − pi,j) χi,j . (6.3) i=1 "j=c # Y Yi 6.2.2 Missing covariate values

Time-invariant covariates When covariates are recorded once for each individual (for example, date of band- ing, weight at banding, bill depth or sex), birds with missing covariate values are disqualified from the analysis. While all birds in the study had banding dates recorded, 5280 birds did not have weights recorded at their times of banding, and were consequently disqualified. Since bill depth and sex are not measured until birds are recaptured as adults, these quantities are only considered as covariates for adult survival. Of the 23 686 Little Penguins banded as chicks, 1788 birds were subsequently recaptured alive as adults, and bill depth measurements were recorded for only 336 birds. One might think that the 20 707 birds that were banded and never encountered again (dead or alive) can be included in the analysis, since first-year survival is not taken to depend on sex or bill depth. However if a bird is banded in year tj and never seen again, its likelihood contribution would include the term χi,j. Since χi,j is defined recursively (refer to definition above), it will implicitly include the survival estimates for all of the age components. So if adult survival is allowed to vary with bill depth or sex, these covariate values are required even for birds that are never again encountered

129 after banding (providing they would have reached “adult” age by the end of the study). Therefore birds without valid covariate values are disqualified from the analysis.

Time-varying covariates The problem of missing covariate values is of particular concern when using co- variates such as current weight, which vary with time for each individual. Out of 23 686 birds banded as chicks in this study, and recaptured from penguin years 1968 to 2003, only 3748 weight measurements were recorded. Furthermore, 87% of birds banded as chicks were never encountered again after the initial capture, and 66% of those encountered dead or alive in subsequent years had at least one “0” (not seen) entry in their yearly history records between their initial and final encounters. Therefore, the method of disqualifying birds with any missing covari- ate values would result in the disqualification of almost all of the birds, leading to biased estimates (Catchpole et al., 2008). In fact, the only birds to be included in the analysis would be those with “complete weight records”. This notion of com- pleteness depends on the age structure and the model for the survival probability.

If the model for survival is {φ1, φ2, φ3, φ4:8, φ9+(weight)}, birds recaptured and weighed each breeding season from the initial capture until occasion k − 1, or until they are recovered dead, would have complete weight records. Due to the recursive nature of χ, if a bird vanishes prior to tk−1, χi,j will include the probability φ9+ which depends on the bird’s unknown weight on occasion j. The only exception will be if birds vanish at an earlier age towards the end of the study, so that χ only involves “younger” survival probabilities, which do not depend on weight. In cases such as this, standard approaches involving disqualifying animals with missing covariate values can result in an enormous amount of wasted data. Indeed, as Catchpole et al. (2008) point out, such a method of analysis can lead to “imprecise and/or strongly biased estimators”. An alternative approach would be to fill in missing values with the last previously known measurement (as in Chapter 7 and Catchpole et al., 2004). However, such a method may result in bias if the probability of being missed depends on the value of the covariate. For example, birds that are in poor condition (eg. sick or

130 underweight) in a particular season may not breed and hence may not be recaptured and weighed in that season. Another method is to use the three–state likelihood developed by Catchpole et al. (2008) and outlined in the next section, whereby, for each bird, only those tran- sition probabilities in the likelihood that depend on missing covariates are deleted. This approach is used in the analysis in Section 6.3.4.

6.2.3 The three–state likelihood

Again consider a mark-recapture-recovery experiment on n animals over k occasions, at times t1, t2,...,tk. The model parameters, φ, p and λ, are defined as before. Catchpole et al. (2008) consider the mark-recapture-recovery experiment as a three–state process, with the three states corresponding to the possible history entries

0, if animal i was not seen at tr,  hi,r = 1, if animal i was seen alive at tr,   2, if animal i was found dead in (tr−1, tr).   They extend the definition of χi,r (from Equation 6.2 on p.129) to

χi,r,s = P(not found, alive or dead, from tr+1 to ts inclusive | alive at tr), for s = r +1,...,k, with χi,r,r = 1, so that χi,r = χi,r,k.

The recurrence relation for χi,r,s is now

χi,r,s = (1 − φi,r)(1 − λi,r)+ φi,r(1 − pi,r)χi,r+1,s, (6.4) where ci ≤ r

Catchpole et al. (2008) then define the transition probability from state a to state b as follows:

πi,r(a, b)=P(hi,r+1 = b | hi,r = a). (6.5)

In the transitions from state 0, l is taken to be the last occasion, prior to occasion r, on which the bird was seen (assuming that it was then alive). The transition probabilities are given by:

131 1, if hi,l =2 πi,r(0, 0) = (6.6a)  χ /χ , if h =1  i,l,r+1 i,l,r i,l r−1  πi,r(0, 1) = φi,s(1 − pi,s) × φi,rpi,r/χi,l,r, (6.6b) s=l Yr−1

πi,r(0, 2) = φi,s(1 − pi,s) × (1 − φi,r)λi,r/χi,l,r, πi,r(1, 0) = χi,r,r+1, (6.6c) s l Y= πi,r(1, 1) = φi,rpi,r, (6.6d)

πi,r(1, 2) = (1 − φi,r)λi,r, (6.6e)

πi,r(2, 0)=1. (6.6f)

Note that πi,r(0, 0) = 1 once a bird has been recovered dead.

Brief derivations of Equations 6.6(a) and 6.6(b) follow:

Let A = {not seen at tr+1}

B = {not seen at tl+1,...,tr}

C = {seen alive at tl}

D = {seen alive at tr+1}

πi,r(0, 0) = P(not seen at tr+1 | (not seen at tl+1, ..., tr ∩ seen alive at tl))

= P(A | (B ∩ C)) P((A ∩ B)|C) = P(B|C) χ = i,l,r+1 , χi,l,r

132 πi,r(0, 1) = P(seen alive at tr+1 | (not seen at tl+1, ..., tr ∩ seen alive at tl)) = P(D | (B ∩ C)) P((D ∩ B)|C) = P(B|C) r− 1 φ p = φ (1 − p ) × i,r i,r . i,s i,s χ s l i,l,r Y=

The derivation of πi,r(0, 2) is very similar to that of πi,r(0, 1), while the remaining transition probabilities are trivial. Catchpole et al. (2008) then use the life-history data for each bird to calculate the likelihood as follows:

n k−1 2 2 xi,r(a,b) L = πi,r(a, b) , (6.7) i=1 r=c a=0 b Y Yi Y Y=0

1, if hi,r = a and hi,r+1 = b, where xi,r(a, b)=  0, otherwise.  The three–state likelihood is equivalent to the standard likelihood when there are no missing covariate values. Whereas the standard likelihood cannot be calculated for animals that have missing covariate values, the three–state method uses all of the available information for each animal, since the unknown conditional transition probabilities can simply be deleted from the likelihood, while the known segments are retained (Catchpole et al., 2008).

6.3 Results

In this section I consider the best purely age-dependent model as determined in

Chapter 4, that is, {φ1,φ2,φ3,φ4:8,φ9+; p1,p2,p3:4,p5+; λ1,λ2:3,λ4+} (model 5.1). The survival estimates (with standard errors in brackets) for this model appear in Table 6.1. I now fix this age structure and consider various individual covariates for the age components of the survival probabilities.

6.3.1 Date of banding

The list of models considered, together with their AIC values and the regression coefficients (and standard errors), appear in Table 6.2. I first consider a model

133 Table 6.1: Estimates (with standard errors in parentheses) for model parameters under model {φ1, φ2, φ3, φ4:8, φ9+, p1, p2, p3:4, p5+, λ1, λ2:3, λ4+} (model 5.1).

Year of life Survival Recapture Recovery 1 0.165 (0.007) 0.078 (0.005) 0.041 (0.002) 2 0.712 (0.025) 0.216 (0.009) 0.121 (0.015) 3 0.781 (0.021) 0.294 ( 0.009) 4 o 0.829 (0.006) 5:8 o 0.194 (0.010) 0.388 (0.008) 9+ o 0.797 (0.010) ) o

0.7

0.6

0.5

0.4

0.3

First year survival probability 0.2

0.1

0 50 100 150 200 250 300 Banding date

Figure 6.1: First-year survival probability versus number of days from the beginning of the penguin year to each individual’s date of banding under model 6.1. The range of banding dates in the graph corresponds to the observed values.

with φ1 depending on the individual’s banding date (denoted by bd), that is the number of days elapsed from the beginning of the penguin year to the bird’s date of banding (model 6.1). While the average survival probability for birds in their first year of life is 16.5% (model 5.1, Table 6.1), Figure 6.1 indicates that an individual’s banding date has a very strong effect on its survival probability in the first year of life. For birds banded very early in the breeding season, the predicted probability of survival is greatly enhanced (around 60%) when compared to birds banded later in the breeding season (φ1 as low as 5%).

134 Table 6.2: AIC values (after subtracting 44 000) and regression coefficients (with standard errors in parentheses) for models involving banding date as a covariate for φ1. The model is {φ1(V), φ2, φ3, φ4:8, φ9+, p1, p2, p3:4, p5+, λ1, λ2:3, λ4+}, where V=band date (bd), band date anomaly (bda), mean band date (mbd), band date anomaly standardized (bdas), and so on.

Covariate for φ1(V) Regr. coeff. (s.e.) AIC 6.1 bd bd: −1.53 (0.08) 622 6.2 bda bda: −1.53 (0.09) 722 6.3 mbd mbd: −1.60 (0.15) 933 6.4 bda+bda2 bda: −1.49 (0.10) 684 bda2: 1.29 (0.23) 6.5 mbd+bda mbd: −1.60 (0.16) 624 bda: −1.51 (0.09) 6.6 bdas bdas: −0.47 (0.03) 706 6.7 mbd+bdas mbd: −1.64 (0.16) 603 bdas: −0.47 (0.03)

There are two components of banding date: (i) variation between seasons and (ii) variation within a season. Further investigation is required to determine whether the relationship between φ1 and the banding date is due to (i), (ii) or both. The variation between seasons, which can be measured by the mean banding date (mbd) for all birds banded in each breeding season, is most likely due to changing conditions. For example, breeding is known to commence early in breeding seasons in which conditions are favourable, and late when there are poor conditions (a late onset of breeding was observed following the pilchard crash of 1995, Dann et al., 2000). In Chapter 5 I found that φ1 was higher in years with earlier mean laying date, perhaps because young birds survive better and breeding commences earlier in years in which there are good conditions. Variation within a season is measured by an individual’s band date anomaly (bda), which tells whether a bird was banded early or late in the season relative to other birds in its cohort. These statistics are defined as follows:

(banding dates for cohort j) mean band date (j) = , number of birds in cohort j P band date anomaly (i, j) = banddate(i, j) − mean band date(j).

135 0.5

0.45

0.4

0.35

0.3

0.25

0.2

First year survival probability 0.15

0.1

0.05

0 150 160 170 180 190 200 210 220 230 Mean banding date

Figure 6.2: First-year survival probability versus mean banding date under model 6.3. Note that the mean banding dates have a much smaller range than the indi- vidual banding dates. Yearly estimates of survival under a φ1(time) model (model 5.2) are also shown. Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale).

Figures 6.2 and 6.3 display graphs of φ1 versus mean band date (model 6.3, Table 6.2) and band date anomaly (model 6.2) respectively. While the regression coefficients corresponding to mean band date and band date anomaly are both highly significant (Wald test, p< 10−25 and p< 10−64 respectively) and of similar magnitude, the individual band date anomaly has a much greater effect on φ1 than the mean band date. That is, φ1 decreases from 25% to 8% as the mean band date increases, whereas φ1 declines from 67% to 3% for increasing band date anomaly (refer to Figures 6.2 and 6.3), over the observed data range. Figure 6.2 also displays yearly estimates of first-year survival under a φ1(time) model (model 5.2). (Note that this is not possible for the individual covariate graphs.)

I next consider a model with φ1 depending on both mean band date and band date anomaly (model 6.5). The regression coefficients (and standard errors) cor- responding to mean band date and band date anomaly under this model are still highly significant and almost identical to those obtained when φ1 was allowed to

136 0.7

0.6

0.5

0.4

0.3

First year survival probability 0.2

0.1

0 −200 −150 −100 −50 0 50 100 150 Banding date anomaly

Figure 6.3: First-year survival probability versus banding date anomaly (bda) under model 6.2. The curve stretches over the range of observed values of bda. vary separately with mean band date and band date anomaly (compare model 6.5 with models 6.2 and 6.3, Table 6.2). The effects of these two covariates on φ1 are the same when considered separately or together, which is not surprising, as I would expect the two components of banding date to be uncorrelated. Therefore, in addition to good seasons producing chicks that are more likely to survive their first year of life, penguins hatched earlier than average in any breeding season have a much better chance of surviving than those hatched relatively late in the season. There are two possible biological factors which might account for this. Firstly, a bird’s parents are likely to breed earlier than average in a season if they are in good condition, in which case they are more likely to produce healthy offspring, and to be in a better position to care for their chicks. Secondly, while the adults are caring for chicks, they must find food near Phillip Island, so that they can return to the breeding areas at night (Dann et al., 1992). This might cause the food supply close to Phillip Island to diminish by late in the breeding season (Dann and Norman, 2006), so that birds that fledge then might have difficulty finding food. Indeed, Chiaradia and Kerry (1999) found that chicks hatched late in the breeding season failed to fledge due to starvation.

137 1

0.9

0.8

0.7

0.6

0.5

0.4 First year survival probability

0.3

0.2

0.1 −200 −150 −100 −50 0 50 100 150 Banding date anomaly

Figure 6.4: First-year survival probability versus banding date anomaly under model 6.4 (quadratic on a logistic scale).

I next impose a quadratic relationship (on a logistic scale) between φ1 and indi- vidual band data anomaly (model 6.4, Figure 6.4). A comparison of the likelihood values for model 6.2 (which imposed a linear relationship) and model 6.4 indicates that the quadratic function fits the data better than the linear function. However, the apparent increase in φ1 for band date anomalies larger than about 50 days (see Figure 6.4) does not make biological sense and is probably artificial. It might be worthwhile investigating other, possibly nonlinear or threshold, models. However, one must beware of overfitting the data. Since the durations of some breeding seasons are longer than others, I next calculate the standardized band date anomaly for a bird in cohort j by dividing its band date anomaly by the sample standard deviation for the band dates in breeding season j, that is

band date anom (i, j) standardized band date anomaly (i, j)= . standard deviation for cohort j

Figure 6.5 displays the relationship between φ1 and the standardized band date anomaly under model 6.6. So once the variation in the widths of the breeding season is accounted for, the survival declines from 53% to 2% for birds banded increasingly

138 0.7

0.6

0.5

0.4

0.3

First year survival probability 0.2

0.1

0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 Standardized banding date anomaly

Figure 6.5: First-year survival probability versus standardized banding date anomaly under model 6.6.

Table 6.3: Regression coefficients (with standard errors in parenthe- ses) and p-values (if p<0.05) for each age component of φ for model {φ1,2,3,4:8,9+(age∗V), p1, p2, p3:4, p5+, λ1, λ2:3, λ4+}, where V= band date anomaly.

Age component Regr. coeff. (s.e.) p-value −41 φ1 −1.5065 (0.1110) p< 10 φ2 −0.0062 (0.2284) φ3 −0.1962 (0.2723) φ4:8 −0.0037 (0.1268) φ9+ −0.0684 (0.1808)

later relative to other birds in that breeding season. Note that a φ1(mbd+bdas) model (model 6.7) fits the data better than φ1(mbd+bda) (model 6.5), suggesting that the duration of the breeding season should be accounted for when calculating the band date anomaly. When each age component of the survival probability is modelled as a function of banding date anomaly (model 6.8), only the first year regression coefficient is significant (see Table 6.3). Hence banding date has a significant effect on survival only for first year birds.

139 Table 6.4: AIC values (after subtracting 33 000) and regression coefficients (with standard errors in parentheses) for models involving weight at banding as a covariate for φ1. The model is {φ1(V), φ2, φ3, φ4:8, φ9+, p1, p2, p3:4, p5+, λ1, λ2:3, λ4+}, where V=weight at banding (bw), band weight anomaly (bwa), mean band weight (mbw), and so on.

Covariate for φ1(V) Regr. coeff. (s.e.) AIC 6.9 bw bw: 2.43(0.16) 690 6.10 mbw mbw: 3.31 (0.46) 905 6.11 bwa bwa: 0.23(0.02) 744 6.12 mbw+bwa mbw: 3.46 (0.46) 686 bwa: 0.23(0.02) 6.13 bw+bw2 bw: 10.73(1.22) 644 bw2: −4.03 (0.58) 6.14 mbw+mbw2 mbw: −13.61 (3.59) 896 mbw2: 9.02(1.96) 6.15 bwa+bwa2 bwa: 0.26(0.02) 724 bwa2: −0.03 (0.01) 6.16 mbw+bwa+bwa2 mbw: 3.52 (0.46) 663 bwa: 0.27(0.02) bwa2: −0.03 (0.01)

6.3.2 Weight at banding

I next consider the weight of each penguin at the time of banding (denoted by bw) as a covariate for its first-year survival probability, φ1 (model 6.9 in Table 6.4). As can be seen in Figure 6.6, φ1 increases dramatically (from 4% to 68%) for increasing weight at banding. This result is in accordance with Reilly and Cullen (1982) and Dann (1988). Note that the AIC values listed in Tables 6.2 and 6.4 cannot be compared, as they are not based on the same data sets. As mentioned in Section 6.1.2, while all 23 686 birds have dates of banding recorded, only 18 406 have weights recorded at the time of banding. As with banding date, this is not necessarily an individual effect. Better con- ditions result in heavier chicks and increased survival, but individual birds may not survive better because they are heavier. To determine whether the apparent relationship between φ1 and weight at banding is due to conditions in the season or to the individual weight at banding in relation to the rest of the cohort, I consider mean weight at banding (mbw), which varies with the conditions, and individual

140 0.7

0.6

0.5

0.4

0.3

First year survival probability 0.2

0.1

0 200 400 600 800 1000 1200 1400 1600 1800 2000 Banding weight (g)

Figure 6.6: First-year survival probability versus individual weight at banding under model 6.9. banding weight anomaly (bwa), the weight of an individual relative to other birds in its cohort. These statistics are defined in the same way as for banding date.

The graphs of φ1 versus mean banding weight (model 6.10) and banding weight anomaly (model 6.11) appear in Figures 6.7 and 6.8 respectively. While the regres- sion coefficients for both covariates are highly significant (Wald test, p< 10−12 and p< 10−29 respectively), the individual banding weight anomaly has a much greater effect on φ1 than the mean banding weight, due to the greater range for the band- ing weight anomalies. Figure 6.7 also displays yearly estimates of first-year survival under a φ1(time) model (model 5.2).

I next allow φ1 to vary with mean banding weight and banding weight anomaly (model 6.12). The regression coefficients corresponding to the two covariates are still highly significant. Therefore it appears that fledglings survive better in years when the chicks are heavier, and that penguins of above average weight have a better chance of survival than their underweight counterparts.

In models 6.13, 6.14 and 6.15, I again allow φ1 to vary with weight at banding, mean weight at banding and banding weight anomaly respectively. However, this time I impose a quadratic relationship (on a logistic scale) between φ1 and each of the covariates (see Figures 6.9, 6.10 and 6.11). Note that the three quadratic

141 0.5

0.4

0.3

0.2 First year survival probability

0.1

0 700 750 800 850 900 950 1000 1050 1100 1150 Mean banding weight (g)

Figure 6.7: First-year survival probability versus mean weight at banding under model 6.10. Yearly estimates of survival under a φ1(time) model (model 5.2) are also shown. Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale).

0.7

0.6

0.5

0.4

0.3

First year survival probability 0.2

0.1

0 −800 −600 −400 −200 0 200 400 600 800 1000 Banding weight anomaly (g)

Figure 6.8: First-year survival probability versus individual banding weight anomaly under model 6.11.

142 0.35

0.3

0.25

0.2

0.15

First year survival probability 0.1

0.05

0 200 400 600 800 1000 1200 1400 1600 1800 2000 Banding weight (g)

Figure 6.9: First-year survival probability versus individual weight at banding under model 6.13. models fit the data better than the corresponding linear models, models 6.9, 6.10 and 6.11 (Wald tests for the quadratic terms, p< 10−11, p =0.0047 and p < 10−5 respectively). In Figures 6.9 and 6.11, individuals have little chance of surviving their first year of life if their weight is low at the time of banding. Their probability of survival increases considerably if their weight is in the “middle range”. Fledglings with very high weight have a reduced probability of surviving their first year (as hypothesized by M. Cullen, pers. comm.). It is interesting to note that the φ1 versus mean weight at banding curve (Figure 6.10) is quite different from Figures 6.9 and 6.11. Whereas individuals seem to survive better if they are “medium” sized, Figure 6.10 shows that survival increases with mean weight at banding, that is, the higher the average chick weight in a particular season (the better the conditions), the better they survive. However, the monotonic increase in φ1 with mean weight at banding could be due to the fact that the mean weight at banding never reaches the “optimal” value of around 1300 grams.

2 I next consider model φ1(mbw+bwa+bwa ) (model 6.16), which fits the data 2 much better than model φ1(bwa+bwa ) (model 6.15). When each age component of the survival probability is modelled as a function of banding weight ({φ1,2,3,4:8,9+(age∗banding weight)}, model 6.17), only the first

143 0.35

0.3

0.25

0.2 First year survival probability

0.15

0.1 700 750 800 850 900 950 1000 1050 1100 1150 Mean banding weight (g)

Figure 6.10: First-year survival probability versus mean weight at banding (quadratic on a logistic scale) under model 6.14.

0.35

0.3

0.25

0.2

0.15

First year survival probability 0.1

0.05

0 −800 −600 −400 −200 0 200 400 600 800 1000 Banding weight anomaly (g)

Figure 6.11: First-year survival probability versus individual banding weight anomaly under model 6.15.

144 Table 6.5: AIC values (after subtracting 33 000) and regression coefficients (with standard errors in parentheses) for models involving banding date and weight at banding as covariates for φ1. The model is {φ1(V), φ2, φ3, φ4:8, φ9+, p1, p2, p3:4, p5+, λ1, λ2:3, λ4+}, where V=bw+bda, and so on.

Covariate for φ1(V) Regr. coeff. (s.e.) AIC 6.18 bw+bdas bw: 1.98(0.16) 512 bdas: −0.41 (0.03) 6.19 bw∗bdas bw: 1.96(0.16) 506 bdas: −0.23 (0.07) int: -0.69(0.25) 6.20 mbd+bdas+bw+bw2 mbd: −1.97 (0.20) 384 bdas: −0.41 (0.03) bw: 8.56(1.27) bw2: −3.21 (0.61)

−16 year regression coefficient is significant (Wald test, p< 10 for φ1, while p> 0.110 for the older age components of survival). That is, as with banding date, banding weight has a significant effect on survival only for first year birds. I next consider weight at banding and the standardized banding date anomaly together (with and without an interaction term) as covariates for first-year survival (models 6.18 and 6.19, in Table 6.5). The model without an interaction term is de- noted by “bwa+bdas”, while “bwa∗bdas” includes the main effects and interaction (Wilkinson and Rogers, 1973). Considering these two covariates together is an im- provement (in terms of AIC) on modelling them separately, and the interaction term is significant (Wald test, p =0.006). I next concatenate two of the best models for

2 banding date and weight at banding, and consider model φ1(mbd+bdas+bw+bw ) (model 6.20). This improves the model fit when compared to the model with a quadratic relationship between φ1 and banding weight (model 6.14).

6.3.3 Sex and bill depth

Birds without valid sex (sx) and bill depth (BD) covariate values are disqualified from the analysis in this section. Since this results in the disqualification of all birds that were not recaptured at age one year or older (refer to Section 6.2.2), I impose the constraints φ1 = 1 (as all of the birds in the restricted data set must have survived their first year of life) and λ1 = 0 (as none of the birds recaptured as adults could have been recovered dead in their first year of life). In this section I

145 0.89

0.88

0.87

0.86

0.85

0.84 Adult survival probability

0.83

0.82

0.81 10 11 12 13 14 15 16 17 Bill depth (mm)

Figure 6.12: Adult survival probability (for penguins aged three years and older) versus bill depth under model 6.21. consider the same age structure as in model 5.1, except that I form an adult survival for penguins aged three years and older, that is φ4+.

When φ4+ depends on bill depth (model 6.21 in Table 6.6), the predicted adult survival probability increases with bill depth from approximately 81% to 89% over the range of observed bill depths (see Figure 6.12). The regression coefficient cor- responding to bill depth is not significant (Wald test, p=0.08). As well as the modelled dependence on bill depth being quite low, the small number of birds in this analysis (n = 336) contributes to this lack of significance. As explained on p.127, the sex of a bird is determined as follows: 1, bill depth≤ 13.3 mm (female), sex =  0, bill depth> 13.3 mm (male). 

I now consider a model with φ4+ depending on sex (model 6.22). Adults aged three years and over had estimated survival probabilities of 86% for males and 84% for females. However, the regression coefficient corresponding to sex was not significant (p=0.14).

Model 6.24 again has φ4+ depending on sex, but this time the sex is determined “by sight” (denoted by sxs). This model produces adult survival estimates of 84% for males and 82% for females. This analysis is based on many more birds (n =

146 Table 6.6: AIC values and regression coefficients (with standard errors in paren- theses) for models with adult survival (φ4+) depending on bill depth and sex. The model is {φ1, φ2, φ3, φ4+(V), p1, p2, p3:4, p5+, λ1, λ2:3, λ4+}, with φ1=1 and λ1=0. Here V=sex(sx), bill depth(BD), sex by sight (sxs) and so on. Model 6.24 is based on a different set of data from models 6.21–6.23.

Covariate for φ4+(V) Regr. coeff. (s.e.) AIC 6.21 BD BD: 1.02(0.59) 4711 6.22 sx sx: −0.20 (0.13) 4714 6.23 sx∗BD sex: −0.20 (0.13) 4715 BDf: 0.32(1.48) BDm: 1.57(1.52) 6.24 sxs sxs: −0.10 (0.07) 19605

1686), but once again the regression coefficient corresponding to sex is not significant (p=0.13). For the bill depth to be a meaningful measure of the size of the bird, males and females must be considered separately. I next consider a model with φ4+ depending on {sx+BDf+BDm} (model 6.23), where bill depth, if bird is female, BDf =  0, otherwise,  and BDm is defined similarly. This model is denoted by φ4+(sx∗BD) (Wilkinson and Rogers, 1973). However, the regression coefficients corresponding to sex, BDf and BDm are not significant (Wald tests, p=0.14, 0.83 and 0.30 respectively).

6.3.4 Current weight

As explained in Section 6.2.2, there are very few complete weight records in the Little Penguin data. Therefore, standard approaches for dealing with missing co- variate values will almost certainly fail: disqualifying birds with any missing values will result in the disqualification of almost all of the birds in the study, while using previous recorded weights will result in bias if the probability of being missed in a season depends on the weight value in that season. To overcome these potential problems, the analysis in this section uses the three–state method of Catchpole et al. (2008), as outlined in Section 6.2.3.

147 The weight of bird i in year j is denoted by w(i, j), while the mean weight (mw) in year j is the mean weight of all birds whose weights are recorded at tj. The weight anomaly (wa) for bird i in year j is then defined as:

wa(i, j) = w(i, j) − mw(j).

I now consider various models using the weights recorded during the breeding season (with census date 1 January) as covariates for adult survival, φ4+, in the following calendar year (see Table 6.7). The regression coefficient corresponding to weight is significant for model φ4+(w) (model 6.25, Wald test, p=0.0002), with the annual adult survival probability increasing from 55% to 100% over the range of the observed weights (see Figure 6.13). Model φ4+(wa) (6.26) also produces a significant result (Wald test, p=0.0009), with φ4+ increasing from 14% for underweight birds to 100% for birds that are much heavier than average in that season (see Figure 6.14). Considering mean weight (ie. variation in adult weights between seasons) as a covariate for adult survival, does not produce a significant result (model 6.27). This result is unexpected — I would expect higher mean weights to occur in “good” sea- sons, and that the favourable conditions in that season would result in a higher annual adult survival probability. It is interesting to note that the regression co- efficient estimates are similar for models 6.25 and 6.27, but the standard error for model 6.27 is much higher.

Since φ4+ approaches one for models 6.25 and 6.26, I next consider quadratic relationships first between adult survival and weight (model 6.28) and then between

φ4+ and weight anomaly (model 6.29). Neither of these models improves the model fit. Figure 6.15 suggests that the adult survival probability approaches one for birds weighing 1400 grams and over. There is little improvement in model fit by

2 considering models φ4+(mw+wa) and φ4+(mw+wa+wa ) (models 6.30 and 6.31). It would be interesting to consider weight as a covariate for the recapture prob- ability, since heavier birds are likely to be breeders and are therefore more likely to be recaptured.

148 1

0.95

0.9

0.85

0.8

0.75

0.7 Adult survival probability

0.65

0.6

0.55 600 800 1000 1200 1400 1600 1800 2000 2200 Weight (g)

Figure 6.13: Adult survival probability (for penguins aged three years and older) versus weight under three–state model 6.25.

1

0.9

0.8

0.7

0.6

0.5

0.4 Adult survival probability

0.3

0.2

0.1 −1000 −800 −600 −400 −200 0 200 400 600 800 1000 Weight anomaly (g)

Figure 6.14: Adult survival probability (for penguins aged three years and older) versus weight anomaly under three–state model 6.26.

149 1

0.9

0.8

0.7

0.6

0.5

0.4 Adult survival probability

0.3

0.2

0.1 600 800 1000 1200 1400 1600 1800 2000 2200 Weight (g)

Figure 6.15: Adult survival probability (for penguins aged three years and older) as a quadratic function of weight (on a logistic scale) under three–state model 6.28.

Table 6.7: AIC values and regression coefficients (with standard errors in parenthe- ses) for models with adult survival (φ4+) depending on current weight. The model is {φ1, φ2, φ3, φ4+(V), p1, p2, p3:4, p5+, λ1, λ2:3, λ4+}, where V=weight (w), weight anomaly (wa), mean weight (mw), and so on.

Covariate for φ4+(V) Regr. coeff. (s.e.) AIC 6.25 w w: 4.39(1.18) 5312 6.26 wa wa: 0.40(0.12) 5315 6.27 mw mw: 5.50(4.77) 5324 6.28 w+w2 w: −5.06 (5.74) 5311 w2: 5.08(2.98) 6.29 wa+wa2 wa: 0.60(0.17) 5314 wa2: 0.04(0.03) 6.30 mw+wa mw: 7.66(3.64) 5313 wa: 0.41(0.12) 6.31 mw+wa+wa2 mw: 8.06(3.73) 5314 wa: 0.61(0.16) 5313 wa2: 0.04(0.03)

150 800

600

400

200

0

Banding weight anomaly (g) −200

−400

−600

−150 −100 −50 0 50 100 Banding date anomaly (days)

Figure 6.16: Banding weight anomaly versus banding date anomaly.

6.3.5 Correlations between covariates

There is a highly significant negative correlation between banding weight anomaly and banding date anomaly (r = −0.2450, p< 10−10). Biologically, this makes sense. A bird that fledges earlier than average has a higher weight at banding relative to other birds in its cohort, since the food supply is presumably more plentiful for parents raising chicks earlier in the season, regardless of the conditions in that year. Later in the season, when the food stocks around Phillip Island have diminished (Dann and Norman, 2006), the parents find it more difficult to find food for their chicks, resulting in lighter weight chicks. Figure 6.16 displays a plot of banding date anomaly versus banding weight anomaly. There is a negative correlation between the mean banding date and the mean weight at banding (r = −0.2238), since birds breed earlier and chicks are heavier in good seasons. However there is insufficient evidence (p = 0.1961) to reject the null hypothesis of zero correlation between these statistics (see Figure 6.17). This p-value is highly conservative since it ignores the fact that these are sample means.

151 1150

1100

1050

1000

950

900

850 Mean banding weight (g)

800

750

700 150 160 170 180 190 200 210 220 230 Mean banding date (days)

Figure 6.17: Mean banding weight versus mean banding date.

152 6.4 Conclusion

In this chapter I have found that birds banded very early in the breeding season have a greatly enhanced probability of surviving their first year of life (around

60%) when compared to birds banded later in the breeding season (φ1 as low as 5%). Considering the two orthogonal components of variation in the banding date (mean banding date and individual banding date anomaly) as covariates for first- year survival yields the following results: (i) seasons with earlier mean banding dates produce chicks that are more likely to survive their first year of life, because these tend to be “good” seasons, and (ii) “early” chicks (that is, penguins hatched earlier than average in a breeding season) have a much better chance of surviving their first year (φ1=67%) than “late” chicks (φ1=3%). A chick’s weight at banding also has a strong effect on its survival in its first year of life. φ1 increases dramatically (from 4% to 68%) for increasing weight at banding. Again considering the variation in banding weight between seasons (mean banding weight) and within the season (banding weight anomaly) as covariates for first-year survival, shows that (i) newly-fledged birds survive better in years when the chicks are heavier (again this is a reflection of “good” seasons), and (ii) birds of above average banding weight have a better chance of survival than their underweight counterparts. However, while first-year survival increases with weight at banding, the remaining age components of survival do not depend on banding weight. Some interesting results were observed when I considered quadratic relation- ships between φ1 and banding weight, banding weight anomaly and mean banding weight. I found that individuals have little chance of surviving their first year of life if their weight is low at the time of banding, but that their survival probability increases considerably if their weight is in the “middle range”. Fledglings with very high banding weights (or with banding weights much higher than average for that season) have a reduced probability of surviving their first year. There are two pos- sible explanations for this result: (i) there may be an optimum weight for chicks, and (ii) the effect of banding could be greater for large birds (see p.173 in Chapter 7). However, this result does not hold for mean weight at banding. Survival in- creases with mean banding weight, that is, the higher the average chick weight in a particular season, the better they survive.

153 Considering sex and bill depth as covariates for the annual adult survival prob- ability did not produce any significant results. However, current weight and weight anomaly were found to have strong effects on the survival probability for adults. Adult survival increased from 55% to 100% over the range of observed weights, and from 14% for underweight birds to 100% for those that were heavier than average.

154 Chapter 7

Banding Effects

While Chapters 4–6 of this thesis have analysed 36 years of mark- recapture-recovery data for 23 686 Little Penguins banded as chicks on Phillip Island, this chapter analyses band and transponder data from a separate, smaller study of 2483 birds marked as adults, to determine the rates of tag loss and the effect of banding on the survival of adult Little Penguins.

7.1 Introduction

For more than 50 years, researchers have been marking penguins with flipper bands (Sladen, 1952). While early research into banding focused on extending the life of bands (eg., Sladen and LeResche, 1970), more recent research has studied the effect of banding on the mortality and energetics of penguins (Ainley et al., 1983; Culik et al., 1993; Gauthier-Clerc et al., 2004; Dugger et al., 2006). A small number of penguins are known to have been injured or killed as a direct result of their flipper bands: when a band that was not properly sealed partially opened and cut into the birds flesh, when the band became entangled in fishing gear, or through injuries caused by the restrictive band to the swollen flipper during moult (P. Dann, pers. comm.). Some of these negative effects have probably been reduced through the design of better bands (Sladen et al., 1968). However, bands can have a more insidious effect on their wearers. Ainley et al. (1983) found that banded Ad´elie Penguins had lower survival rates than unbanded penguins until their first moult after banding. A study by Culik et al. (1993) showed that banded Ad´elie penguins used more energy than unbanded penguins when swimming. More

155 recently, Gauthier-Clerc et al. (2004) indicated that survival rates of banded King Penguin chicks were significantly lower than unbanded, transpondered chicks. This chapter analyses seven years of mark-recapture data for three groups of birds; one flipper-banded group, one unbanded group that had been injected with passive-induction transponders, and one group that had been marked with both devices. Since only adults (of unknown age) were marked in this transponder study, the effect of banding on young birds remains unknown.

7.2 The Data

7.2.1 Collection of data

On 86 occasions during the 1994/95 to the 2000/01 breeding seasons, Little Pen- guins were captured at dusk as they came ashore on Cowrie Beach at Phillip Island, and crossed into the breeding areas along four main tracks on the beach. The birds were caught using walk-in traps made from a shadecloth fence approximately 15m in length. At the time of first capture all birds had their bill depths measured, and were weighed with a spring balance. Birds were also weighed at each recap- ture. On each capture occasion, unmarked penguins were assigned to the various groups as follows: the first and fourth unmarked birds captured were assigned to the band only (B) group, the second and fifth unmarked birds to the transpon- der only (T) group, the third and sixth unmarked birds to the group with both band and transponder (BT), and so on, to allow an equal and unbiased selection of birds in each group. Flipper bands were supplied by the ABBBS, and Trovan transponders (2mm × 11mm) and hand-held readers were used. A total of 2483 birds were marked in three groups (nT = 829, nB = 826, and nBT = 828) over four years (1994–1997). The group sizes were slightly uneven through errors made during the initial marking. The study was continued for an additional three years after marking had stopped in 1998, to collect recapture data. The transponders, which were individually sealed in sterile needles, were in- jected under the skin between the scapulae. There were no apparent injuries to the penguins caused by the injection of the transponders. Although the initial inten- tion was to seal the injection site with wound glue, the wounds were so small, and the feather density so great, that the researchers were unable to locate the wound

156 on the bird’s skin. During the study several transpondered penguins were killed by feral foxes Vulpes vulpes and were examined for transponder side effects. No internal damage was detected and the transponders had remained at the injection site. The mark-recapture data for birds marked with both a band and a transponder (ie. those in the BT group) provide useful information on band and transponder loss rates for birds marked as adults. A transponder can be “lost” if it comes out of the wound at the injection site, if it malfunctions, or if it migrates from the injection site in such a way that it cannot be detected by the scanner (although this is quite unlikely). Band loss would most likely be caused by the incorrect application of a band. Although bands are sometimes lost through band wear, this is unlikely in a short-term study such as this. In the first year of the study, researchers marked almost every bird that was caught. Very few of these birds had been marked previously, due to the separation between Cowrie Beach and the six study sites from the main study. Hence there was perhaps a fairly random sample from the age-distribution of adults in the first cohort. In subsequent years, many birds encountered during the beach visits were already marked, and the unmarked birds that were captured would have been mostly new birds returning to the colony to moult or breed for the first time, although some would have been established breeders that were missed or had taken a sabbatical in earlier years of the study. Therefore the cohorts of birds marked in later years of the study were likely to be younger than those marked in the first year (P. Dann, pers. comm.).

7.2.2 Sample of data

The data consist of 7077 records, with each record corresponding to a live encounter with a bird. Recoveries of dead birds have not been included in the analysis as they were few in number, and there were no dead recovery data available for the transponder-only birds. The live captures were carried out from 4 January 1995 to 22 January 1998, while the recaptures extend from 16 January 1995 to 16 January 2001. The fields in the raw database appear in Table 7.1, while Table 7.2 lists the

157 Table 7.1: Fields from band and transponder database

DATE Date (19941203 means 3 December 1994) BAND Band number for bird fitted with a band, “0” otherwise TRANSP Transponder number for bird fitted with a transponder, “0” otherwise CORRAL Corral or track number of beach where bird was encountered N New=“1” for initial marking record of an individual, and “0” for subse- quent encounters TBERROR Transponder or band error, eg. TLOST for lost transponder (see Table 7.2) ET Error type: numerical code for TBERROR (see Table 7.2) BD Bill depth WT Weight SX Sex of bird — “0” for unknown, “1” for male and “2” for female (see p.170) METH ABBBS method of encounter and status code (refer to Appendix B.1) COM Comment (“13” for band only (B) group, “113” for transponder only (T), “213” for both (BT))

Table 7.2: Errors in data for banding effects study.

Error Error type Meaning code

ET=1 BMISSED Band not observed when BT bird recaptured, but seen later. ET=2 TMISSED Scanner failed to detect transponder of BT bird, but transponder detected later. ET=3 TLOST Scanner failed to detect transponder of BT bird. ET=4 BMISREAD Band number misreads corrected for BT birds. ET=5 TMISREAD Already corrected as scanning device records numbers of scanned transponders. ET=6 TBWRONG Scribing error ET=8 BLOST Similar to TLOST. types of errors encountered. An explanation of these errors appears on the following pages.

When the scanning device fails to detect the transponder of a BT bird at tj, and the transponder is not detected later, an error of TLOST is recorded at tj. If the missed transponder is detected later, a TMISSED error is recorded at tj. If the scanner misses the transponder of a T bird, this bird would not be recorded as having been recaptured. Indeed, on the odd occasion, a previously transpondered

158 bird was re-tagged as it was mistakenly identified as an unmarked bird. Such birds were removed from the study. BMISREAD errors are corrected for BT birds, since the band numbers can be crosschecked against the transponder numbers. For B birds, researchers correct band numbers where possible, by substituting numbers that sound (or look) similar, ensuring that they obtain a valid band number, and that the sex and corral numbers match previous records for that individual. Although the transponder numbers are automatically stored in the scanning de- vice, the scribes record these numbers in their field notes. The recorded transponder numbers are later checked against the stored numbers and, if necessary, they are marked as TMISREAD errors and corrected. An example of TBWRONG follows: The scribe confused two birds and recorded them as 1. BAND=91573, TRANSP=135C9B1, and 2. BAND=92706. However, these entries should have been 1. BAND=91573, and 2. BAND=92706, TRANSP=135C9B1.

The records for four birds appear in Table 7.3. These birds were encountered in Corral 1 on Cowrie Beach and were marked with a band and transponder on 4 January 1995. Three of the birds were females (birds 1, 3 and 4) and one was a male (bird 2). Determination of sex is explained on p.173. While birds 2 and 4 had their transponders intact on each occasion on which they were encountered, bird 1 lost its transponder between the 1995 and 1996 breeding seasons, and bird 3 lost its transponder between the 1995 and 2000 seasons. Bird 1 vanished with only its band intact after the 1997 season, and bird 2 vanished with both band and transponder intact after the 1996 season. Bird 4 was alive and well, with band and transponder intact, at the end of the study. (See p.163 for the summarised yearly data for these birds.) The numbers of birds marked each year in each group appear in Table 7.4, while the number of recaptures are shown in Table 7.5.

159 Table 7.3: Mark-recapture records for four birds

BIRD DATE BAND TRANSP CORRALN TBERRORET BD WT SX METH COM 19950104 85026 1DAFA4D 1 1 0 0 118 1250 2 0 213 19951206 85026 1DAFA4D 1 0 0 0 0 0 2 0 213 1 19970113 85026 1DAFA4D 1 0 TLOST 3 0 1460 2 0 213 19971110 85026 1DAFA4D 1 0 TLOST 3 0 0 2 0 213 19971206 85026 1DAFA4D 1 0 TLOST 3 0 1260 2 0 213 19950104 85027 12A6CA8 1 1 0 0 152 1170 1 845 213 19950116 85027 12A6CA8 1 0 0 0 0 0 1 0 213 2 19951113 85027 12A6CA8 1 0 0 0 0 1440 1 845 213 19960206 85027 12A6CA8 1 0 0 0 0 0 1 845 213 19961203 85027 12A6CA8 1 0 0 0 0 1230 1 845 213 19970107 85027 12A6CA8 1 0 0 0 0 1110 1 845 213 19950104 85028 1368BAD 1 1 0 0 116 1050 2 0 213 3 19951126 85028 1368BAD 1 0 0 0 0 1120 2 0 213 20010115 85028 1368BAD 1 0 TLOST 3 0 0 2 0 213 19950104 85029 1D3B445 1 1 0 0 116 1030 2 845 213 4 19950116 85029 1D3B445 1 0 0 0 0 950 2 845 213 19960104 85029 1D3B445 1 0 0 0 0 0 2 845 213 20001216 85029 1D3B445 1 0 0 0 0 990 2 845 213

Table 7.4: Number of birds by group and cohort.

Penguin year B T BT Banded Unbanded Totals 1994 245 245 246 491 245 736 1995 149 151 151 300 151 451 1996 247 247 249 496 247 743 1997 185 186 182 367 186 553 Totals 826 829 828 1654 829 2483

Table 7.5: Number of recaptures by group and year.

Penguin year B T BT Banded Unbanded Totals 1995 20 22 25 45 22 67 1996 94 105 105 199 105 304 1997 122 145 142 264 145 409 1998 333 354 332 665 354 1019 1999 279 341 300 579 341 920 2000 95 114 89 184 114 298 Totals 943 1081 993 1936 1081 3017

160 7.3 Analysis

7.3.1 Notation

The data are analysed using a mark-recapture analysis, with models that include year, time elapsed since marking, and covariate dependence in the model param- eters. The data have been summarized by “penguin year” (as defined on p.54 in Chapter 3), so that birds encountered alive in the same breeding season are grouped together. Maximum likelihood methods are used to estimate the model parameters, and the likelihood is formed using the techniques of Burnham (1993), Catchpole et al. (1998a) and Catchpole (unpubl. data).

As before, φi,j denotes the annual survival probability of penguin i from tj to tj+1, and pi,j represents the probability of recapturing penguin i at tj+1. Also, the probabilities of penguin i retaining its band or transponder from tj to tj+1 are denoted by ρB,i,j and ρT,i,j respectively. Let k be the number of capture/recapture occasions (here k = 7) and assume that bird i is marked on occasion ci. The numbers of birds in the three groups are nB, nT and nBT .

The statistics VB, WB, and ZB are defined for bird i in group B (for 1 ≤ i ≤ nB) as follows:

1 if bird i is captured or recaptured at tj and not seen again, VB,i,j =  0 otherwise,   for ci ≤ j ≤ k, and

1 if bird i is recaptured at tj+1, WB,i,j =  0 otherwise,  and 

1 if bird i is not recaptured at tj+1 but encountered later, ZB,i,j =  0 otherwise,   for ci ≤ j ≤ k − 1. In a similar manner, VT , WT and ZT , and WBT and ZBT are defined for birds in groups T and BT respectively.

161 Additional statistics, V10, V01, V11, W1, W2, Z1, Z2, VOB and VOT , are required for double-tagged BT birds.

V10 is defined as:

1 if bird i is recaptured with band only at tj and not seen again, V10,i,j =  0 otherwise.  V and V are defined similarly, with “band only” replaced by “transponder only” 01 11 and “band and transponder” respectively.

While WBT and ZBT use all of the recapture data for each BT bird, W1 and Z1 use only the recapture information when BT birds are recaptured with their band only or with both their band and transponder. That is, encounters with BT birds that have lost their bands are ignored in the calculations of W1 and Z1. Similarly

W2 and Z2 use only the recapture information gained from the transponder.

We define temporary variables vB and oB by:

j if band is seen on bird i for the last time at tj and band is known to be lost, vB,i =  0 if band is not known to be lost,   j if transponder is seen at tj for the first time after bird i lost band, oB,i =  0 if band is not known to be lost,  with vTand oT defined similarly. Note that a band is “known to be lost” if the bird is encountered later with its transponder only. Now VOB is defined as:

1 if vB(i) > 0 and oB(i) > 0 and vB(i) ≤ j ≤ oB(i) − 1, VOB,i,j =  0 otherwise. 

In other words, VOB indicates those occasions (if any) from when the band vanishes up until, but not including, the occasion when the other tag (ie. the transponder) is next seen. VOT is defined similarly.

7.3.2 The likelihood

For birds in the B group, the history matrix is defined as follows:

1 if bird i was captured or recaptured at tj, HB,i,j =  0 otherwise.   162 Table 7.6: Typical tag histories

Bird Penguin Year 1994 1995 1996 1997 1998 1999 2000 1 11 11 10 10 00 00 00 2 11 11 11 00 00 00 00 3 11 11 00 00 00 00 10 4 11 11 00 00 00 00 11

Similarly for HT and HBT . In addition, each BT bird is assigned a 2-digit “tag history” code for each year of the study as follows: the first digit refers to the presence/absence of a band and the second refers to the transponder. So the tag history for bird i at tj is given by

11 if bird i was encountered with band and transponder at tj,  10 if bird i was encountered with band only at t ,  j Ti,j =   01 if bird i was encountered with transponder only at t ,  j 00 otherwise.    Here “encountered” means captured or recaptured. The tag histories for the four birds from Table 7.3 appear in Table 7.6, and the likelihood contributions are as follows. For clarity, the subscripts referring to the bird numbers are omitted.

Bird1: L= ρB,1ρB,2ρB,3ρT,1(1 − ρT,2)φ1φ2φ3p1p2p3χB,4,

2: L= ρB,1ρB,2ρT,1ρT,2φ1φ2p1p2χBT,3,

3: L= ρB,1 ...ρB,6ρT,1(1 − ρT,2 ...ρT,6)φ1 ...φ6p1(1 − p2)(1 − p3)(1 − p4)(1 − p5)p6,

4: L= ρB,1ρB,2 ...ρB,6ρT,1ρT,2 ...ρT,6φ1φ2 ...φ6p1(1 − p2)(1 − p3)(1 − p4)(1 − p5)p6,

where χB,i,j, χT,i,j and χBT,i,j are defined as

χB,i,j = P(penguin i not seen after tj | seen alive with band only at tj),

χT,i,j = P(penguin i not seen after tj | seen alive with transponder only at tj),

χBT,i,j = P(penguin i not seen after tj | seen alive with both band and transponder at tj).

163 The recursion for χB,i,j is derived below. Note that the probabilities in the derivation are conditional on bird i being seen alive with its band only at tj.

χB,i,j = P(lost band in (tj, tj+1)) + P(retained band in (tj, tj+1), died in (tj, tj+1))

+ P(retained band, surv (tj, tj+1), not recap at tj+1, not seen after tj+1)

= (1 − ρB,i,j)+ ρB,i,j(1 − φi,j)+ ρB,i,jφi,j(1 − pi,j)χB,i,j+1.

The recursion for χT,i,j is similar, while the recursion for χBT,i,j is:

χBT,i,j = P(lost band and transponder in (tj, tj+1))

+ P(didn’t lose both band and transponder in (tj, tj+1), died in (tj, tj+1))

+ P(retained band, lost transp., surv (tj, tj+1), not recap tj+1, not seen after tj+1)

+ P(lost band, retained transp., surv (tj, tj+1), not recap tj+1, not seen after tj+1)

+ P(retained band & transp., surv (tj, tj+1), not recap tj+1, not seen after tj+1)

= (1 − ρB,i,j)(1 − ρT,i,j)+(ρB,i,j + ρT,i,j − ρB,i,jρT,i,j)(1 − φi,j)

+ρB,i,j(1 − ρT,i,j)φi,j(1 − pi,j)χB,i,j+1 + (1 − ρB,i,j)ρT,i,jφi,j(1 − pi,j)χT,i,j+1

+ρB,i,jρT,i,jφi,j(1 − pi,j)χBT,i,j+1.

Note that χB,i,k = χT,i,k = χBT,i,k = 1. All recursions are valid for ci ≤ j ≤ k − 1, where ci is the initial marking occasion for bird i and tk is the final recapture occasion. Over the course of the study, 43 birds from the BT group lost their transponders, while five lost their bands. Of the 43 birds that lost their transponders, six were recaptured without their transponders within a month or two of marking. Five of these were never encountered again, while the sixth bird was encountered again two years later. Using the summarised tag history data, these five birds would have likelihood contributions of χBT,1, which assumes that they disappeared with their bands and transponders intact. However, this is not using the full information, since I know that they did not retain their transponders in the first year after marking and that they disappeared with only their bands intact. To make use of the full information, the likelihood contributions are adjusted to (1 − ρT,1)χB,1. Similarly,

164 the bird that was recaptured immediately without its transponder and was not seen again for two years has the tag retention part of its likelihood contribution adjusted from ρB,1ρB,2(1 − ρT,1ρT,2) to ρB,1ρB,2(1 − ρT,1). The likelihood for all of the birds in the study is formed by taking the product of the likelihoods for the three groups of birds, that is L = LBLT LBT . The derivations of LB and LT are the same as that of the likelihood from Chapter 4 (see Appendix A.1), except that subscript i refers to bird i rather than cohort i. The likelihood for the double-tagged birds, LBT , which is given in Catchpole (unpubl. data), can be derived by considering the likelihood contributions for birds 1–4 given above.

Neglecting constants, LB, LT and LBT are given by:

nB k−1 WB,i,j +ZB,i,j WB,i,j ZB,i,j VB,i,j LB = {ρB,i,jφi,j} pi,j (1 − pi,j) χB,i,j , i=1 "j=c # Y Yi

nT k−1 WT,i,j +ZT,i,j WT,i,j ZT,i,j VT,i,j LT = {ρT,i,jφi,j} pi,j (1 − pi,j) χT,i,j , i=1 "j=c # Y Yi

nBT k−1 W1,i,j +Z1,i,j W2,i,j +Z2,i,j WBT,i,j +ZBT,i,j WBT,i,j ZBT,i,j LBT = ρB,i,j ρT,i,j φi,j pi,j (1 − pi,j) i=1 "j=c Y Yi k−1 k−1 V10,i,j V01,i,j V11,i,j VOB,i,j VOT,i,j × χB,i,j χT,i,j χBT,i,j (1 − ρB,i,j )(1 − ρT,i,j ) . # j=c j=c ! Yi Yi

For ease of programming, the lower limit j = ci can be replaced by j = 1 in each expression, since each of the exponents is zero when j

7.4 Results

As all of the birds in this study were marked as adults of unknown age, age depen- dence in the parameters cannot be included. However, since there is frequently a tagging effect on bird survival (Lebreton et al., 1992), and the loss of a transponder would most likely occur immediately after marking (P. Dann, pers. comm.), the survival probability and the probabilities of retaining the tags are likely to depend

165 on time elapsed since marking. Therefore, in this chapter the term “age” in the model notation denotes “time elapsed since marking” rather than chronological age. I consider separate annual survival probabilities and probabilities of retaining the tags in the first year after marking and in all subsequent years, since there is no biological justification for considering more refined age structures for these parameters. The general notation of Section 7.3.1 is adapted here to suit the models fitted.

That is, φ1 denotes the survival probability in the first year after marking, and φ2+ is the annual survival probability in subsequent years. The annual probability of retaining the band is ρB,1 in the first year after marking and ρB,2+ in subsequent th years, with ρT,1 and ρT,2+ defined similarly. The recapture probability in the (j+1) breeding season is pj. I now define factors “tag”, with three levels corresponding to the three types of tags used, that is, band only (B), transponder only (T) and both devices (BT), and “band”, which has two levels, namely banded (ie. birds from B and BT groups) and unbanded (birds from T group). The AIC values and the total number of estimable parameters for the various models appear in several tables in this Section. Note that when calculating the number of estimable parameters in each model, I must allow for the fact that when φ and p are time-dependent, these parameters cannot be estimated separately in the final year of the study, due to a well-known model identifiability problem (Lebreton et al., 1992). Furthermore, φ1 cannot be estimated in the final two calendar years of the study since birds were not marked in these years, and φ2+ cannot be estimated in the first year of the study, as there were no birds in this category then.

7.4.1 Recapture probability

Transponders may fail to be detected when a bird is recaptured and scanned, whereas bands are quite obvious and would (almost always) be noticed by the researchers. Thus the recapture probability p might depend on “band”, with p for B and BT birds being higher than that of T birds. However there is no reason to expect this effect to vary from one year to the next, because the same scanners were used throughout the study and the scanning was done in the same way. Hence a

166 Table 7.7: AIC values (after subtracting 11 500) and total number of es- timable parameters, K, for various models for the recapture probability. In each case the model for the probabilities of survival and of retaining the tags is {φ1, φ2+; ρB,1, ρB,2+, ρT,1, ρT,2+}.

Model for p K AIC 7.1 p 7 1114 7.2 p(time) 12 101 7.3 p(band+time) 14 100 p(band+time) model, whereby recapture varies by year but the effect of group is the same every year, would be more appropriate than a p(band∗time) model. While the recapture probability is strongly time-dependent (compare models 7.1 and 7.2 in Table 7.7), it does not depend on band: a p(band+time) model does not fit the data significantly better than a purely time-dependent recapture model (compare models 7.2 and 7.3, p =0.068, LRT). Therefore, I will use p(time) while constructing an appropriate model for φ, before checking the model for p once again.

7.4.2 Survival probability

In this section, I consider various models for the survival probability φ. Here the recapture probability and the probabilities of retaining the tags are modelled as

{p(time); ρB,1, ρB,2+, ρT,1, ρT,2+}.

Modelling survival separately for φ1, the survival probability in the first year after marking, and φ2+, the annual survival in subsequent years, is a substantial improvement on a model with constant survival φ (compare models 7.4 and 7.5 in Table 7.8). Furthermore, allowing the components of the survival probability to vary with “tag” or “band” (see models 7.6–7.9, Table 7.8) is a significant improvement on the {φ1, φ2+} model (compare models 7.5 and 7.9, p=0.0001, LRT). However, there is no significant improvement in model fit by considering the B and BT groups separately (compare models 7.8 and 7.9, p=0.099, LRT). This implies that banding affects the survival probability, but there is no evidence that the transponder has any additional effect on the survival of banded birds. Therefore, henceforth I consider only the effect of the factor “band” on the survival probability. The estimates (and standard errors) for the annual survival probabilities for banded and unbanded birds in their first year after marking and in subsequent

167 Table 7.8: AIC values (after subtracting 11 500) and total number of estimable parameters, K, for various models for survival involving the two-level factor “band” and the three-level factor “tag”. In each case the model for the probabilities of recapture and of retaining the tags is {p(time); ρB,1, ρB,2+, ρT,1, ρT,2+}.

Model for φ K AIC 7.4 φ 11 123 7.5 φ1, φ2+ 12 101 7.6 φ1(tag), φ2+(tag) 16 88 7.7 φ1(band), φ2+(tag) 15 89 7.8 φ1(tag), φ2+(band) 15 86 7.9 φ1(band), φ2+(band) 14 87

Table 7.9: Survival estimates (and standard errors) under model 7.9: {φ(age∗band); p(time); ρB,1, ρB,2+, ρT,1, ρT,2+}.

Annual survival probability Unbanded Banded first year after marking 0.8042 (0.0260) 0.7401 (0.0178) subsequent years 0.8983 (0.0158) 0.8577 (0.0127) years (model 7.9) appear in Table 7.9. In the 12 months following marking, the av- erage survival probability of banded birds is 6% lower than that of unbanded birds, while in subsequent years the average annual survival probability of banded birds is around 4% lower than their unbanded counterparts. These results indicate a po- tential marking effect, suggesting that marking does cause trauma to the penguins, since the survival of these birds in their first year after marking is reduced, regard- less of the type of mark used. Since the survival may vary from year to year, the standard errors in Table 7.9 may exaggerate the precision of the survival estimates.

Time variation in survival

I now consider a model with both φ1 and φ2+ varying separately with time (model 7.10 in Table 7.10). The survival estimates for this model vary considerably from year to year (see Figure 7.1), perhaps due to poor conditions in the early part of the study. Indeed, the conditions for penguins in southern Australia were particularly poor in the early years of the study, following the widespread mortality of their major food source, pilchards, in 1995 (Dann et al., 2000). Again, the survival in the first year after marking is consistently lower than the survival estimates in subsequent years.

168 Table 7.10: AIC values (after subtracting 11 500) and total number of estimable parameters, K, for various models for survival involving time. In each case the “age” structure for survival is {φ1,φ2+} and the model for the probabilities of recapture and of retaining the tags is {p(time); ρB,1, ρB,2+, ρT,1, ρT,2+}.

Model for φ K AIC 7.10 φ(age∗time) 18 63 7.11 φ(age∗{band+time}) 20 47 7.12 φ(age∗band∗time) 27 41

To determine whether the apparent time dependence in survival is actually caused by the changing chronological ages of the cohorts over the years of the study, I consider the results of model 7.10, together with the age-dependent sur- vival estimates from my analysis of the main data set. Since the first cohort of birds marked in the 1995 penguin year consists of a reasonably random sample from the age-distribution of adult birds, the φ1 estimate for 1995 should correspond to the annual survival for birds aged approximately three years and over. In contrast, the later cohorts mostly comprise new breeders (as mentioned above), so the φ1 esti- mates for later years should be based on the data for younger birds. But since the survival estimates for established breeders are higher than those of the new breeders

(φ2−3 =0.78 and φ3−8 =0.83, see p.77), the age distribution in 1995 should produce a slightly higher φ1 estimate than for the later cohorts of new breeders. Therefore, the lower φ1 estimate in 1995 is likely to be due to conditions in that year (such as the pilchard crash), rather than any genuine age effect. I next consider band+time models for the survival probability in the first year after marking and in subsequent years (model 7.11 in Table 7.10). That is, φ1 depends on whether or not the bird was banded, and varies from year to year, but the effect of banding is the same every year, and similarly for φ2+. However, this band+time model does not fit the data as well, according to the AIC, as a band∗time modelling scheme (model 7.12), which allows the effect of banding on survival to vary over the years (LRT, p=0.005). A possible reason for this is that the effect of banding on survival is greater in a “bad” year than in a “good” year. However, this is not borne out by Figure 7.2 which shows the yearly variation in the survival estimates for banded and unbanded birds under model 7.12. While

169 1 Subsequent years

0.9

0.8

0.7 Survival probability

0.6 First year after marking

0.5 1995 1996 1997 1998 1999 Year

Figure 7.1: Survival probability versus time under model {φ1(time), φ2+(time), p(time), ρB,1, ρB,2+, ρT,1, ρT,2+} (model 7.10). Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale). the general pattern is “φ for banded birds is less than φ for unbanded birds”, the reverse is true for φ1 in 1996. The reason for this anomalous result is unknown.

Sex effect Recall that Little Penguins can be sexed with an accuracy of 91% by using their bill depth measurements (see p.127 in Chapter 6). I now allow the annual survival probabilities in the first year after marking and in subsequent years to vary with sex (models 7.13 and 7.14 in Table 7.11). The regression coefficient corresponding to sex is significant for φ1 (model 7.13, Wald test, p-value=0.007), but not for φ2+ (model 7.14, p-value= 0.656). Therefore, the annual survival probability in the year following marking depends on sex (φ1=79% for males and 73% for females), but the sex effect does not appear in subsequent years (φ2+=87% for all birds). It is not clear at this stage whether this effect is a genuine sex difference or if it is actually a size effect. I will investigate this later in this chapter. I now determine whether the effect of sex varies with time, via a sex∗time model for the survival probabilities in the first year after marking and in subsequent years

170 1 1 Unbanded 0.95 0.95

0.9 Unbanded 0.9

0.85 0.85

0.8 0.8

0.75 0.75 Banded

0.7 Banded 0.7 Survival probability Survival probability 0.65 0.65

0.6 0.6

0.55 0.55 1995 1996 1997 1998 1996 1997 1998 1999 Year Year (a) (b)

Figure 7.2: Time-varying annual survival probability for (a) the first year after marking and (b) more than one year after marking under model {φ(age∗band∗time); p(time); ρB,1, ρB,2+, ρT,1, ρT,2+} (model 7.12). Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale).

Table 7.11: AIC values (after subtracting 11 500) and total number of estimable parameters, K, for various models for survival involving sex. In each case the “age” structure for survival is {φ1,φ2+} and the model for the probabilities of recapture and of retaining the tags is {p(time); ρB,1, ρB,2+, ρT,1, ρT,2+}.

Model for φ K AIC 7.13 φ1(sex), φ2+ 13 95 7.14 φ1, φ2+(sex) 13 103 7.15 φ(age∗sex∗time) 27 70 7.16 φ(age∗{sex+time}) 20 60 7.17 φ(age∗{sex+band}) 16 83 7.18 φ1(sex∗band), φ2+(band) 16 84 7.19 φ1(sex+band), φ2+(band) 15 82 7.20 φ1({sex+band}∗time), φ2+(band∗time) 31 42 7.21 φ1(sex+band+time), φ2+(band+time) 21 44 7.22 φ1(sex∗band+time), φ2+(band+time) 22 46 7.23 φ1(sex∗band∗time), φ2+(band∗time) 35 48

171 1

0.95

0.9 Males 0.85

0.8

0.75

Survival probability 0.7

0.65

Females 0.6

0.55 1995 1996 1997 1998 Year

Figure 7.3: Time-varying survival probability for the first year after marking un- der model {φ(age∗sex∗time), p(time), ρB,1, ρB,2+, ρT,1, ρT,2+} (model 7.15). Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale).

(model 7.15), whereby the effect of sex on φ1 is allowed to vary from year to year, and similarly for φ2+. Results from this model show that males consistently survive better than females in the year following marking (refer to Figure 7.3). As expected, there is no significant difference in the survival probabilities of males and females in subsequent years. Since the effect of sex appears not to vary much from year to year, I next consider model 7.16: {φ1(sex+time), φ2+(sex+time)}, which indeed fits the data better than the sex∗time model.

The results for model 7.18: {φ1(sex∗band), φ2+(band)}, appear in Table 7.12. It is interesting to note that banded males survive the twelve months following marking almost as well as the unbanded females. This is consistent with results obtained by Dugger et al. (2006) who found that male Ad´elie Penguins had higher survival than females, in both banded and unbanded groups. Table 7.12 also indicates that the effect of banding is virtually the same for both males and females in the year following marking. Therefore I next consider a φ1(sex+band) model (model 7.19) which, as expected, fits the data better than model 7.18.

172 Table 7.12: Survival in the year following marking varies by sex and band, while the annual survival in subsequent years varies by band only (φ1(sex∗band), φ2+(band), model 7.18).

Annual survival probability First year after marking Subsequent years Male Female Banded 0.7751 (0.0234) 0.7079 (0.0228) 0.8571 (0.0127) Unbanded 0.8322 (0.0329) 0.7757 (0.0335) 0.8976 (0.0158)

As was the case with the band models in Table 7.10, the addition of time as a factor in {sex, band} models substantially improves the fit (see models 7.20–7.23 in Table 7.11), without shedding any more light on the underlying biology.

Bill depth Bill depth (BD) is confounded with sex, in the sense that a bird with a bill depth greater than 13.3mm is classified as a male, while a bird with a smaller BD is said to be a female (see p.127). Therefore, I consider the effect of BD, while allowing for sex. Once sex is accounted for, the bill depth measurement provides an indication of the size of the bird within each sex.

I first consider a model with both φ1 and φ2+ depending on band+sex∗BD (model 7.24 in Table 7.13). Here the effect of banding is taken to be the same for males and females, which is consistent with my previous results from Tables 7.11 and 7.12, and the slope is assumed to be the same for unbanded and banded females, and similarly for unbanded and banded males. This model does not perform better than model 7.17 which had φ1 and φ2+ depending on sex+band. However, since model 7.24 is biologically interesting (see discussion below), I have included Figure 7.4, which shows the annual survival estimates in the first year of marking and in subsequent years against bill depth for banded and unbanded males and females under this model. The regression coefficient corresponding to bill depth for females is marginally significant for φ1 (Wald test, p=0.0663), while the remaining regression coefficients for bill depth are not significant. Figure 7.4 shows that, in the first year after marking, unbanded birds survive better than their banded counterparts, and males have higher survival probabilities than females, within the banded/unbanded groups. While Figure 7.4(b) suggests an optimum size for penguins (as hypothesized by M. Cullen, pers. comm.), these results are not statistically significant.

173 Table 7.13: AIC values (after subtracting 11 500) and total number of estimable parameters, K, for various models for survival involving bill depth. In each case the “age” structure for survival is {φ1,φ2+} and the model for the probabilities of recapture and of retaining the tags is {p(time); ρB,1, ρB,2+, ρT,1, ρT,2+}.

Model for φ K AIC 7.24 φ(age∗{band+sex∗BD}) 20 83 7.25 φ(age∗{band+sex∗BD}∗time) 54 61 7.26 φ(age∗band∗sex∗BD) 26 92 7.27 φ(age∗band∗sex∗BD∗time) 81 97

0.9 0.92 Unbanded 0.85 males 0.9 Unbanded Unbanded females males 0.8 Unbanded 0.88 0.75 females Banded 0.7 males 0.86

0.65 0.84

0.6 Survival probability Banded Survival probability 0.82 females 0.55 Banded females Banded 0.8 males 0.5

0.45 0.78 80 100 120 140 160 180 80 100 120 140 160 180 Bill depth Bill depth (a) (b)

Figure 7.4: Annual survival probability versus bill depth for (a) the first year after marking and (b) more than one year after marking (φ(age∗{band+sex∗BD}), model 7.24 in Table 7.13).

Model 7.24 fits the data considerably better than a model with φ1 and φ2+ depending on band∗sex∗BD (model 7.26), where the effect of banding is different for males and females, and there are distinct slopes corresponding to bill depth for

φ1 and φ2+ for each of the four groups: banded females, unbanded females, banded males, and unbanded males. While adding time variation to model 7.26 does not improve the fit (compare models 7.26 and 7.27), a {band+sex∗BD}∗time model for

φ1 and φ2+ (model 7.25) is an improvement on model 7.24. However, model 7.25 does not perform as well as model 7.20.

Weight and weight anomaly Weight (wt) measurements are recorded when the penguins are captured or recap- tured as they cross Cowrie Beach. A bird’s weight can vary significantly throughout

174 Table 7.14: AIC values (after subtracting 11 500) and total number of estimable parameters, K, for various models for survival involving weight. In each case the “age” structure for survival is {φ1,φ2+} and the model for the probabilities of recapture and of retaining the tags is {p(time); ρB,1, ρB,2+, ρT,1, ρT,2+}.

Model for φ K AIC 7.28 φ1(sex+band+wt), φ2+(band+wt) 17 80 7.29 φ1(sex+band+wt+time), φ2+(band+wt+time) 23 46 7.30 φ1({sex+band+wt}∗time), φ2+({band+wt}∗time) 40 50 7.31 φ1(sex+band+wtanom), φ2+(band+wtanom) 17 82 the breeding season. In fact, it can double in weight prior to the moult in Febru- ary or March (P. Dann, pers. comm.). If a bird is not recaptured in a particular year, its previous recorded weight is used (see discussion below). I also consider the weight anomalies (wtanom) as covariates for survival. Since missing weight values are replaced by the last recorded values, the weight anomaly for year j is calculated by subtracting the mean weight in the year that it was recorded, rather than the mean for year j. That is, the weight anomaly for individual i in year j is defined as: wtanom(i, j) = weight of bird i when last recorded

− mean weight of recaptured same sex birds in year bird i last recorded.

A bird’s weight anomaly provides a measure of its condition relative to other birds of the same sex in that season. Since weight anomaly allows for the difference in weight between the sexes, there is no need to consider the males and females separately.

Model 7.28 (in Table 7.14) has a survival model of {φ1(sex+band+wt), φ2+(band+wt)}. This model fits the data slightly better than the corresponding model without weight (model 7.19), although the regression coefficients for weight are not sig- nificant (p=0.148 and p=0.246 for φ1 and φ2+ respectively, Wald tests), suggesting that weight may be only of marginal use as a covariate for survival. A model al- lowing for variation in the effect of weight with time, model 7.30, does not fit the data as well as the corresponding model without weight (model 7.20).

175 Table 7.15: AIC values (after subtracting 11 500) and total number of estimable parameters, K, for various models for recapture. In each case the model for the probabilities of survival and of retaining the tags is {φ1({sex+band}∗time), φ2+(band∗time); ρB,1, ρB,2+, ρT,1, ρT,2+}.

Model for p K AIC 7.32 p 27 556 7.20 p(time) 31 42 7.33 p(band+time) 32 44

I next consider model 7.31: φ1(sex+band+wtanom), φ2+(band+wtanom). Since this model is inferior to model 7.28 (in terms of AIC), I do not consider weight anomaly further. Rather than replacing missing weight values with previously recorded values, I could have used the three-state method of Catchpole et al. (2008) (as outlined on p.131 in Chapter 6). However, since the three-state programs would require considerable modifications to enable them to be used for B, T and BT birds, and the results obtained using weight as a covariate are not significant, such an approach is probably not worth pursuing here.

Revisiting the recapture probability I now check whether the p(time) model is still appropriate once survival has been modelled adequately. The results from Table 7.15 confirm that the recapture prob- ability varies strongly with time (compare models 7.32 and 7.20). Once again, the recapture probability does not depend on band, since a p(band+time) model does not fit the data as well as the p(time) model (compare models 7.20 and 7.33). Figure 7.5 displays the recapture estimates under model 7.20. The recapture probability was very low (14%) in the 1995/1996 breeding season, it was still rela- tively low (around 35%) in 1996 and 1997, but then picked up substantially in 1998 and 1999 (70%). Note that for models with time dependence in the recapture and survival probabilities, φ and p are unable to be estimated separately in the final year of the study, due to a well-known parameter redundancy problem (Catchpole et al., 1996; Lebreton et al., 1992). Although time dependence in the recapture probability is often due to a varying recapture effort over time, this is not the case in the current study. A more likely

176 1

0.9

0.8

0.7

0.6

0.5

0.4

Recapture probability 0.3

0.2

0.1

0 1995 1996 1997 1998 1999 Penguin year

Figure 7.5: Recapture probability versus time under model 7.20: {φ1({sex+band}∗time), φ2+(band∗time); p(time); ρB,1, ρB,2+, ρT,1, ρT,2+}. Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale). explanation for the observed temporal variation in the recapture probability would be a change in the numbers of birds breeding. Researchers are likely to encounter breeding penguins at some stage during the breeding season, since these birds go out to sea to feed during the day and return to the burrows to feed their hungry chicks at dusk. However, birds that do not breed in a particular season are less likely to be recaptured by the researchers. As mentioned earlier, the conditions for penguins were particularly poor in the early years of the study, following the widespread pilchard mortality in 1995 (Dann et al., 2000). According to life history strategy, when conditions are poor, many animals try to increase their chance of survival by not breeding in that season (Saether and Bakke, 2000; Tavecchia et al., 2005).

7.4.3 Band and transponder loss

Table 7.16 displays the estimates of the probability of retaining a mark (model 7.20 in Table 7.11). The annual probability of losing a band is around 0.4%, while the

177 Table 7.16: Estimates for the annual probability of retaining the mark under model {φ({sex+band}∗time), φ2+(band∗time); p(time); ρB,1, ρB,2+, ρT,1, ρT,2+} (model 7.20).

Annual probability of retaining mark Type of mark First year after marking Subsequent years Band 0.9953 (0.0049) 0.9969 (0.0024) Transponder 0.9496 (0.0104) 0.9884 (0.0039) probability of losing a transponder is 5% within the first year of marking and an annual probability of around 1% in subsequent years.

7.5 Conclusion

This analysis indicates that banding does have a negative effect on the survival of adult Little Penguins, with banded birds having an annual survival probability 6% lower than their unbanded counterparts. Indeed, the survival probability in the first year after marking is considerably lower than in subsequent years for both banded and unbanded birds, suggesting that marking results in trauma for the bird, regardless of the type of mark used. However, there is no significant difference in the survival probabilities for B and BT birds, indicating that applying a transponder to banded birds does not result in any additional negative effect. Male penguins (as determined by bill depth) survive better than females in the year following marking (79% for males and 73% for females). In fact, banded males survive the twelve months following marking as well as unbanded females. However, this sex effect was not present in subsequent years. There is some evidence that females with larger bill depths survive better than smaller birds. However, the corresponding results for males are equivocal. There is a possibility of an optimum size for penguins, but this hypothesis is not backed up by statistical evidence. A bird’s weight (or weight anomaly) does not provide any further explanation of the variation in its survival probability. Little Penguins are more likely to lose a transponder than a band, and they tend to lose it in the first year after marking rather than in subsequent years. The annual probability of losing a band is around 0.4%, while the probability of losing a transponder is 5% in the first year after marking, and around 1% in subsequent

178 years. The very small rate of band loss (for adults at least) suggests that no correction for this is required in the major study. The annual recapture probability does not depend on the type of tag, but varies considerably over the years of the study, from 14% in penguin year 1995, to around 35% in the next two years, then to 70% in the later years of the study. Since the recapture effort is constant over the years of the study, the temporal variation in the recapture probability is most likely due to changes in the numbers of birds breeding, as a result of varying conditions for penguins. Conditions were particularly poor in the early years of the study following the 1995 pilchard “crash” (Dann et al., 2000). Since banding has been shown to have a detrimental effect on the survival of adult penguins, the adult survival estimates obtained in the main study (see Table 4.2 on p.77) are likely to underestimate the true survival of the general penguin population. Although no such data are available for birds banded as chicks, it is possible that banding also significantly reduces the survival of juvenile Little Penguins, particularly in their first year of life, and that the first-year survival estimates stated in Chapter 4 are underestimates of the true survival probability for unbanded birds. A study examining the effect of banding on chicks is currently underway on Phillip Island (P. Dann, pers. comm.).

179 Chapter 8

Comparison with a New Zealand colony

In this chapter I use mark-recapture-recovery data taken from two populations of Little Penguins, in Phillip Island, Australia and Oamaru, New Zealand, to compare the survival of juvenile Little Penguins from the two locations, and to determine whether the conclusions on covariate dependence for the survival of Phillip Island birds can be extended to other locations. While the Phillip Island study comprises 35 annual cohorts of birds banded as chicks, the New Zealand study consists of six such cohorts. For both colonies, the first-year survival probability varies consider- ably from year to year. While the survival estimate for first year birds in Oamaru is significantly higher than that of Phillip Island, there is a marked decline in first-year survival over time in Oamaru, coinciding with a rapid increase in population size. In order to explain the ob- served temporal variation in first-year survival, I consider demographic parameters, such as mean annual weight at banding, mean annual num- bers of fledglings produced per pair and mean laying date, as indicators of good and bad years. For both Oamaru and Phillip Island, first-year survival increases with mean banding weight, increases with numbers of fledglings per pair, and decreases with mean laying date.

8.1 Introduction

8.1.1 Oamaru study areas

While tourists have visited Phillip Island to watch the Little Penguins crossing the beach since 1928 (Newman, 1992), a smaller scale penguin eco-tourism operation

180 commenced in the town of Oamaru (in south-eastern New Zealand) in 1993 (Per- riman et al., 2000; Johannesen et al., 2003). In Oamaru, there are approximately 1000 breeding Little Penguins (D. Houston, unpubl. data), found along a 3 km section of coast alongside the town and harbour. Little Penguins have been studied intensively at two sites located 1 km apart at each end of the Oamaru harbour. One of the sites, Oamaru Quarry, consists of approximately 100 pairs, while the other, Oamaru Creek, has more than 100 pairs (the exact number is unknown). At the Quarry site, the number of breeding pairs has increased from 33 in 1993 to 98 in 1999, concurring with an increase in the number of available nest boxes (from 87 to 193). There are a negligible number of natural nests in this site. The Creek site had 79 wooden nest boxes during this study, and an unknown, but large number of natural nests.

8.1.2 Monitoring in Oamaru

In this chapter, I analyse mark-recapture-recovery data for 1523 birds banded as chicks in Oamaru from the 1993 to the 1998 breeding seasons. Live recapture data to the end of the 1999 penguin year and dead recoveries up to the end of the 1999 calendar year are included. (The updated Oamaru data will be analysed as soon as they are available.) The live recaptures occur during the weekly visits to the burrows in the study sites. Only nest boxes were monitored as the natural nests were deep and inaccessible. The dead recoveries occur along the Otago coastline.

8.1.3 Data analysis

The Oamaru data are summarised into yearly life history data (as in Chapter 3) and analysed using the methods of Chapter 4. Mark-recapture-recovery models including age- and time-variation are used to model the survival, recapture and recovery probabilities. I next consider three biological variables that provide an indication of the quality of the year: mean laying date of the first clutches, mean number of fledglings produced per pair, and mean banding weight. All models are compared using the AIC values, as in earlier chapters.

181 Table 8.1: Recapture and recovery probability estimates (with standard errors in brackets) for Phillip Island and Oamaru, under model {φ1, φ2+; p; λ}.

Phillip Island Oamaru Recapture 0.260 (0.005) 0.513 (0.018) Recovery 0.058 (0.002) 0.049 (0.007)

8.2 Results

8.2.1 Age structures and estimates for parameters

The overall recapture and recovery probability estimates for Phillip Island and Oamaru appear in Table 8.1. The recapture estimate is much higher in Oamaru (around 51%) than on Phillip Island (26%). This is not surprising since the Phillip Island sites consist of a mixture of natural and box burrows (see p.3), whereas only box burrows are included in the Oamaru study. Birds present in the box burrows will be recaptured during the burrow visits, but those present in natural burrows may occasionally be missed. Furthermore, the Phillip Island sites are visited less often than the Oamaru sites. Unlike the recapture probabilities, the overall recovery estimates are similar for both sites (around 6% for Phillip Island and 5% for Oamaru).

The best age structure for the recapture and recovery probabilities is {p1, p2, p3+, λ1, λ2+} for the Oamaru penguins, and {p1, p2, p3:4, p5+, λ1, λ2:3, λ4+} for Phillip Island. The recapture and recovery probabilities increase with age for both colonies. The age structures selected for the Phillip Island recapture and recovery probabilities are more complicated than for Oamaru. This could be an artefact of the much shorter data set for Oamaru, or it may be due to genuine biological differences in the birds from the two colonies. While one-year-old birds are rarely seen in either colony (hence the need for a separate p1), two-year-olds (with recapture probability p2) may return to the colony to moult, or to breed. The majority of Oamaru penguins commence breeding at the age of two years (or occasionally at one year of age, Johannesen et al., 2003).

Hence, the recapture probability for established Oamaru breeders, p3+, should not vary much with age. In contrast, 50% of the Phillip Island penguins commence breeding at age two, with the remainder usually entering the breeding population

182 at three years of age (Dann and Cullen, 1990). Therefore, for Phillip Island, p3:4 is the probability of recapturing new breeders and p5+ is the recapture probability for established breeders. Established breeders breed more successfully than their younger counterparts (Dann and Cullen, 1990; Chiaradia and Kerry, 1999), and so are more likely to be recaptured.

The recovery age structure for Phillip Island {λ1, λ2:3, λ4+} is indicative of the gradual change in the amount of time the birds spend in various locations as they age (see p.64). Since older birds tend to spend more time near the highly populated Port Phillip Bay, they have a higher recovery probability than younger birds. Less is known about the foraging locations of the penguins from Oamaru. However, the data suggest that the dead recovery probability structure for Oamaru has two age classes {λ1, λ2+}, with a higher recovery probability for the older age class. Since the Little Penguins of Oamaru start spending time ashore from one year of age, they are thus more likely to be found dead than first year birds. Most of the dead recoveries of these older birds occur within Oamaru itself. I next fix the recapture and recovery age structures as explained above and model the survival probabilities for the colonies. For the Phillip Island penguins, the most important aspect of age-related survival is the distinction between the first-year survival and that of older birds (see Table 4.1 on p.72). The best age structure for survival for the Oamaru penguins is {φ1, φ2+}. Therefore, in order to compare the survival estimates across the two colonies, I use a common age structure of {φ1, φ2+} for survival for both Phillip Island and Oamaru, and run the same set of survival models.

8.2.2 Time variation in survival

Several models (models 8.1–8.5 in Table 8.2) are considered in order to assess the relative importance of the temporal variation in the first-year survival and the survival of older birds. The best model for Phillip Island is φ(age∗time) (model 8.1), whereas for Oamaru a model with first-year survival varying with time (model 8.3) fits the data slightly better than the φ(age∗time) model. Figure 8.1 shows the variation in the first-year survival estimates over time for Phillip Island and Oamaru. The first-year survival estimates are quite different for the two colonies,

183 Table 8.2: AIC values (after subtracting 44 000 for Phillip Island and 3000 for Oa- maru) for models with two age-classes for survival, that is {φ1, φ2+}, and with various structures for time-variation in survival. For Phillip Island the age struc- tures for the recapture and recovery probabilities are {p1, p2, p3:4, p5+, λ1, λ2:3, λ4+}, whereas the corresponding structures for Oamaru are {p1, p2, p3+, λ1, λ2+}. All age components of the recapture and recovery probabilities are set to be constant. K is the number of parameters for the survival component of each model.

Model for survival Phillip Island Oamaru K AIC K AIC 8.1 φ1,2+(age∗time) 69 600 11 956 8.2 φ1,2+(age+time) 36 724 7 965 8.3 φ1(time), φ2+ 36 694 7 955 8.4 φ1, φ2+(time) 35 960 6 1014 8.5 φ1, φ2+ 2 1053 2 1013 with a lower average survival in Phillip Island than in Oamaru. Fledglings are thought to survive better in Oamaru due to favourable conditions for penguins in this location (see Section 8.2.4). Varying conditions are likely to be more important for the Phillip Island penguins than for the Oamaru birds (Johannesen et al., 2003), as resources near Phillip Island are depleted towards the end of the breeding season

(Dann and Norman, 2006). However, while there is no apparent trend in φ1 for

Phillip Island, there is a marked decline in φ1 for Oamaru birds over the course of the study, to the level of the Phillip Island estimates. For Phillip Island, temporal variation is more important for first-year survival (compare models 8.3 and 8.5, ∆AIC=359) than for the older birds (compare models 8.4 and 8.5, ∆AIC=93). This is consistent with the results from Oamaru, and corresponds to predictions made from life-history theory (for example Saether and Bakke, 2000, and references therein). It should be noted that the AIC depends on the amount of available information on first year and adult survival, as well as the temporal variability of these two probabilities.

8.2.3 Dependence of first-year survival on biological covariates

Models for φ1(V) are shown in Table 8.3. First-year survival increases with annual mean banding weight (measured in kg) in both Phillip Island (regression coefficient: 3.23 (0.38), p< 10−16, Wald test) and Oamaru (regression coefficient: 10.61 (4.86), p=0.03) (see Figure 8.2). The mean weights in Oamaru are higher than those of

184 Phillip Island Oamaru

0.5

0.4

0.3

0.2 First year survival probability

0.1

0 1970 1975 1980 1985 1990 1995 2000 Year

Figure 8.1: First-year survival of Little Penguins plotted against year for Phillip Island (denoted by ◦, model 8.1) and Oamaru (denoted by ∗, model 8.3). Vertical bars represent 1 standard error on each side of the estimate (back-transformed from the logistic scale).

Phillip Island, perhaps due to better conditions for the penguins in Oamaru (see Section 8.2.4). However, the very small range in weights for Oamaru makes it difficult to identify any regression. Indeed, since the p-value for banding weight in

Oamaru is only marginally significant, the apparent trend in φ1 for this colony is not particularly convincing. The effect of banding weight on first-year survival may be direct, that is when the mean banding weight is high, individual fledglings have a higher weight on average and are in better condition, resulting in a higher chance of survival post- fledging (for example, Magrath, 1991; Olsson, 1997; Ringsby et al., 1998; Sagar and Horning Jr, 1998; Van der Jeugd and Larsson, 1998). Alternately, mean banding weight might affect first-year survival indirectly, that is a high mean banding weight would indicate good food availability leading up to fledging, with the abundance continuing past fledging, making it easier for newly fledged chicks to find food and thus to survive their first year. In Chapter 6, I found that the variation in banding weight between seasons and within seasons both affected a bird’s survival in its

185 first year of life. Future work will examine the effect of individual covariates such as date of banding and weight at banding for Oamaru penguins (E. Johannesen and D. Houston, unpubl. data). An increase in the number of chicks produced per pair is associated with an increase in first-year survival in Oamaru (regression coefficient:0.67 (0.48), not sig- nificant), and Phillip Island (regression coefficient:0.33 (0.07), p< 10−5). Further- more, first-year survival decreases with mean laying date for both colonies, that is late breeding is associated with low first-year survival in both Oamaru (regression coefficient:−0.011 (0.006), p=0.07) and Phillip Island (regression coefficient:−0.33 (0.04), p< 10−15). Thus the relationships between the first-year survival and the demographic vari- ables are consistent between the two colonies. For both colonies, years with low first-year survival had low mean annual banding weights, few young fledged per pair and a late onset of breeding (although these results were not all significant for Oamaru). Models 8.9–8.12 in Table 8.3 include more than one of the variables as covariates. For both Phillip Island and Oamaru, the annual mean laying date and the annual mean banding weight are the two variables that together best explain the yearly variation in first-year survival (model 8.10 in Table 8.3). The inclusion of all three variables improves the model fit for both locations. In fact, for Oamaru, this final model fits the data as well as the model with full time variation in first-year survival (model 8.3). However, since the temporal variation in the first-year survival at Phillip Island has not been adequately explained by the demographic covariates used in this study, the search for potential covariates continues.

8.2.4 Density dependence

The higher first-year survival and the higher incidence of double clutching at Oa- maru (Perriman et al., 2000) when compared to Phillip Island suggest more favourable conditions for Little Penguins at Oamaru. This is perhaps due to a plentiful food supply in the New Zealand colony during the breeding season and the provision of nest boxes in the breeding sites (Johannesen et al., 2003). However, Figure 8.1 indicates that there has been a decline in juvenile survival towards the level

186

Phillip Island Oamaru 0.5

0.4

0.3

0.2 First year survival probability

0.1

0 750 800 850 900 950 1000 1050 1100 1150 Annual average fledgling weight (grams)

Figure 8.2: The relationship between mean annual banding weight and first-year survival probability. The relationship predicted by the Phillip Island data is shown as a solid line (model 8.8). The estimates for models with yearly variation in first- year survival are denoted by ◦ for Phillip Island (model 8.1) and ∗ for Oamaru (model 8.3).

Table 8.3: AIC values (after subtracting 44 000 for Phillip Island and 3000 for Oamaru) for first-year survival probability modelled as dependent on biological parameters under survival model {φ1(V), φ2+}. Age structures for the recapture and recovery probabilities are as in Table 8.2.

Model for φ1 AIC Phillip Oamaru Island 8.3 time 694 955 8.5 constant 1053 1013 8.6 laying date 988 1012 8.7 chicks per pair 1033 1013 8.8 banding weight 981 1010 8.9 chicks per pair, banding weight 983 991 8.10 banding weight, laying date 950 969 8.11 chicks per pair, laying date 990 1014 8.12 chicks per pair, laying date, banding weight 945 955

187 of Phillip Island. This decline is correlated with an increase in the number of breeding penguins at Oamaru. In fact, for the Quarry site data only, the number of breeding pairs explained the variation in first-year survival better than any of the other demographic variables examined, and was the most parsimonious of all models examined. At present it cannot be determined whether this relationship is causal (whereby density affects juvenile survival negatively, by increasing pressure on the food stock, or leads to increased emigration of juveniles, etc.), or whether it is merely due to a density-independent worsening of the conditions in the later years. However, it is worth noting that the probability of double clutching in Oa- maru has also decreased from 30–50% in 1993–1997 to 2–4% in 1998–2000. The observed decrease in apparent survival might also be due to increased emigration as population density increases (B. Morgan, pers. comm.). There was little evidence of an effect of population pressure on first-year survival for Phillip Island when mean annual number of beach crossings at the Penguin Parade was used as a covariate for survival (see p.109 in Chapter 5, and Dann and Norman, 2006). However, this measure of population density may not accurately reflect the yearly variation in the population size of the Little Penguin colony on Phillip Island.

8.2.5 Correlations between biological parameters

The correlation structure amongst the biological covariates is shown in Table 8.4. For both colonies, the annual mean laying date of first clutches was negatively correlated with the number of fledglings per pair, that is a late onset of breeding resulted in the production of fewer chicks (see also Chambers, 2004b). In Oamaru, where double brooding is common (for example, Perriman et al., 2000), a late onset of breeding results in a lower proportion of pairs laying two clutches, and hence a reduction in the number of chicks produced per pair (Johannesen et al., 2003). In Phillip Island, where double clutching is less common (Reilly and Cullen, 1981), the correlation between late onset of breeding and chicks produced per pair is primarily due to a lower survival rate from birth to fledging for chicks produced later in the breeding season (Chiaradia and Kerry, 1999).

188 Table 8.4: Correlation coefficients between the biological parameters. The p-values are the probabilities of getting correlations as large as those observed by chance.

Biological parameters Phillip Island Oamaru laying date, chicks per pair −0.51, p=0.002 −0.78, p=0.07 banding weight, chicks per pair 0.57, p=0.0004 −0.85, p=0.03 banding weight, laying date −0.40, p=0.017 0.87, p=0.02

Banding weights were much higher, on average, in Oamaru than on Phillip Island (see Figure 8.2). However, in Oamaru, years with high banding weight had a late onset of breeding and fewer young produced per pair (although these results were only marginally significant, see Table 8.4). There is some evidence to suggest that, at Oamaru, chicks from single clutches have higher banding weights than those from double clutches, although this relationship is weak (E. Johannesen, unpubl. data). This may partly explain why in years with late onset of breeding, there are fewer double breeders (and consequently fewer young produced per pair), but the average weight of the chicks is higher. Yearly variation in food availability seems to contribute to the variation in several demographic parameters at Phillip Island. For instance, the pilchard stock crash at Phillip Island in the 1995 season resulted in a late onset of breeding, the production of few chicks and low banding weights (Dann et al., 2000).

8.3 Conclusion

The main aim of this chapter was to compare the survival of juvenile Little Penguins in Phillip Island with that of another colony in Oamaru, New Zealand, and to determine to what extent conclusions on covariate dependence of survival extended from Phillip Island to another location. While the Phillip Island data comprised 35 annual cohorts of birds banded as chicks, the Oamaru data consisted of six such cohorts (with data extending until the end of 1999). Future work will analyse the updated Oamaru data when they are available. It will then be possible to fit more realistic models. For both locations, I found a low probability of annual survival in the first year of life, and a high probability thereafter. While the first-year survival probability was generally higher for Oamaru than for Phillip Island, this parameter varied

189 considerably over time for both colonies. In Oamaru there was a marked decline in first-year survival over the years of the study, coinciding with a rapid increase in population size. The results for covariate dependence for Oamaru penguins were consistent with those for Phillip Island (see Chapter 5). That is, first-year survival increased with mean banding weight, increased with number of fledglings per pair and decreased with mean laying date. While the differences in the results for the two colonies (such as the differing recapture and recovery age structures) could be an artefact of the shorter data set for Oamaru, these differences are consistent with the results of existing research into the biological aspects of these populations.

190 Chapter 9

Conclusion

9.1 Summary of results

This thesis analysed a long-term mark-recapture-recovery data set for 23686 Little Penguins that were banded as chicks on Phillip Island from 1968 to 2004. Although various aspects of the study (such as number of study sites and frequency of visits) have varied over the years, the basic experimental methods have remained the same for almost 40 years, making this an extraordinarily rare and valuable data set. Numerous studies have been conducted on Phillip Island and much has been learned about the Little Penguins and their breeding biology, breeding success, diet, and movements and patterns of mortality at sea (Reilly and Cullen, 1979, 1981, 1982, 1983; Dann and Cullen, 1990; Dann et al., 1995; Dann and Norman, 2006). Indeed, several empirical estimates of mortality rates have been made using life-table analyses (Dann and Cullen, 1990; Reilly and Cullen, 1979). However, this current project is the first to carry out a detailed mark-recapture-recovery analysis, using the live recapture and dead recovery information, in order to study the age- specific survival of Little Penguins. While few studies of any animal have been able to model age dependence for the survival, recapture and recovery probabilities simultaneously, I have success- fully applied such a modelling scheme (Chapter 4). I also provided illustrations of potentially erroneous results that may arise when researchers fail to consider such simultaneous age dependence, or fail to detect annual variations that may mask age dependence. The resulting age structures for the model parameters made biological sense and were consistent with the differing lifestyles of birds of various ages. I found that

191 penguins in their first year of life survived poorly, with a survival probability of 17%, perhaps due to their lack of food-finding skills. However, first-year survival could be underestimated due to emigration (see below). More experienced birds in their second and third years had much higher survival probabilities (71% and 78% respectively). Once breeding was established, the survival probability stabilized (at 83%), until it began to decline slightly, possibly due to senescence, at around nine years of age. Since there were no existing data that allowed the estimation of small- or large- scale emigration out of the study sites, emigration was not included in the analysis, and so φ represented the survival without taking site fidelity into account. There- fore, local emigration from the study sites may have contributed to an underesti- mation of the true survival. This is expected to have the greatest effect on the first year survival estimates, since emigration would most likely occur after penguins fledge and leave their natal colony. A population model allowing for immigration of birds from areas surrounding the study sites fit the observed stable population in the study sites. However, since the relative magnitude of small-scale emigration for the various age groups is unknown, it is not possible to determine the true survival. In Chapter 5 I found that several of the age components of the model parameters exhibited considerable temporal variation. The survival in the first year of life varied with time more than the other age components of survival, perhaps because changing environmental conditions, such as food availability, affected the young and vulnerable birds more than the experienced birds. In order to explain the observed temporal variation in the survival probability, I considered several time-varying group covariates for survival. First-year survival increased with number of chicks fledged per pair, increased with annual average fledging weight and decreased with mean laying date. I also considered seasonal sea surface temperature (SST) data for various areas of Bass Strait, and Southern Oscil- lation Index (SOI) data as covariates for first year and adult survival. An increased first-year survival probability was associated with warmer sea surface temperatures in the summer and autumn of the previous year and in the autumn after fledg- ing. These results are in accordance with Chambers (2004a), who provided strong biological justifications for the observed relationships.

192 Chapter 6 considered traits of individual birds as covariates for survival. I found that birds banded very early in the breeding season had a greatly enhanced probability of surviving their first year of life (around 60%) when compared to birds banded later in the breeding season (φ1=5%). By considering the two orthogonal components of variation in the banding date (mean banding date and individual banding date anomaly) as covariates for first-year survival, I found that seasons with earlier mean banding dates (“good” seasons) produced chicks that were more likely to survive their first year of life, and birds born earlier in a breeding season than the rest of their cohort had a much better chance of surviving their first year

(φ1=67%) than “late” chicks (φ1=3%). A chick’s weight at banding also had a strong effect on its survival in its first year of life. First-year survival increased from 4% to 68% as weight at banding increased from 370 to 1950 grams. Considering the variation in banding weight between seasons (mean banding weight) and within the season (banding weight anomaly) as covariates for first-year survival, showed that newly-fledged birds survived better in years in which the mean fledgling weight was heavier, and birds of above average banding weight had a better chance of survival than their underweight counterparts. Interesting results were observed when I considered a quadratic relationship between

φ1 and the banding weight. Individuals had little chance of surviving their first year of life if their weight was low at the time of banding, but their survival probability increased considerably if their weight was in the “middle range”. Fledglings with very high banding weights had a reduced probability of surviving their first year. There are two possible explanations for this result: (i) there may be an optimum weight for chicks, and (ii) the effect of banding could be greater for large birds. Chapter 7 suggested that the latter explanation might be true. However, this result was not statistically significant. Current weight and weight anomaly were found to have strong effects on the survival probability for adults, as well as for chicks. Adult survival increased from 55% to 100% over the range of observed weights, and from 14% for underweight birds to 100% for those that were heavier than average. Considering sex and bill depth as covariates for the annual adult survival probability did not produce any significant results.

193 In Chapter 7 I considered a separate data set from an experiment studying the effect of banding on the survival of adult Little Penguins. There were seven years of recapture data for three groups of birds: one flipper-banded group, one unbanded group that had been implanted with transponders, and one group that had both devices. The annual probability of losing a band was around 0.4%, while the probability of losing a transponder was 5% in the first year of marking, and around 1% in subsequent years. The recapture probability varied considerably from year to year (from 14% to around 70%), most likely due to changes in the numbers of birds breeding, as a result of varying conditions for penguins. However, the recapture probability did not depend on the type of mark used. A “marking effect” was evident, whereby the survival of penguins in their first year after marking was considerably lower than in subsequent years for both banded and unbanded birds, suggesting that marking caused the birds trauma, regardless of the type of mark used. Banding did appear to have a negative effect on the survival of adult Little Penguins. In the 12 months following marking, the average survival probability of banded birds was 6% lower than that of unbanded birds, while in subsequent years the average annual survival probability of banded birds was around 4% lower than their unbanded counterparts. Although the effect of banding in the year following marking was virtually the same for males and females, males survived better than females. In fact, banded males survived the twelve months following marking as well as unbanded females. This could be due to a size effect, as females with larger bill depths survived better than smaller birds. In Chapter 8 I compared the survival estimates for the Phillip Island penguins with those obtained for another Little Penguin colony in Oamaru, New Zealand, by analysing six years of mark-recapture-recovery data for the birds from Oamaru. For both locations, I found a low probability of annual survival in the first year of life, and a high probability thereafter. While the first-year survival probability was generally higher for Oamaru than for Phillip Island, this parameter varied considerably over time for both colonies. In Oamaru there was a marked decline in first-year survival over the years of the study, coinciding with a rapid increase in population size.

194 The results for covariate dependence for Oamaru penguins were consistent with those for Phillip Island, that is, first-year survival increased with mean banding weight, increased with number of fledglings per pair and decreased with mean laying date. While the differences in the results for the two colonies (such as the differing recapture and recovery age structures) could have been an artefact of the shorter data set for Oamaru, these differences were consistent with the results of existing research into the biological aspects of these populations.

9.2 Future work

This thesis provided many insights into interesting aspects of the lives of Little Penguins. However, due to the vast amount of Little Penguin data available, there are many other issues that can potentially be addressed. The recapture probability for a breeder in a given year essentially represents the probability that the bird is breeding in a study site in that year, since a breeding bird will almost always be encountered at some stage in the season. Thus far, I have assumed that the annual survival probability does not depend on whether or not birds were breeding in the previous breeding season. Breeding success information, which is available for only some of the banded Little Penguins on Phillip Island, is harder to access than the life history information, since there are separate breeding sheets for each study site in each season. However, future work will compile breeding histories for birds recorded as breeding in the study sites, and will use these data, together with the mark-recapture-recovery data for these birds, to obtain estimates of survival for birds that bred in the previous season, and for those that did not. Furthermore, I will determine whether previous breeding success (both recent and long-term, as in Moyes et al., 2006) affects the annual survival probability. I will also determine whether the effect of breeding on survival is greater during periods of “severe environmental conditions”, as proposed by Tavecchia et al. (2005). While Little Penguins have been known to take breeding sabbaticals in years in which conditions are poor, little is known about possible patterns in their decision to take a sabbatical, and the effect of senescence on this. Future work will model the recapture probability on occasion j (a proxy for the probability of breeding in that year) in terms of the recapture probabilities on occasions j − 1 and j − 2.

195 Since penguins go to sea after fledging, and rarely return to the colony until they moult or begin to breed, it may be useful to consider a multi-state model of recruitment to the breeding population, as developed by Crespin et al. (2006). These authors developed capture-recapture methods which modelled recruitment as a process where birds “move from an unobservable state at sea, through a non- breeding state present in the colony, to the breeding state”. Chapters 4–6 analysed 36 years of mark-recapture-recovery data for Little Pen- guins banded as chicks, whereas Chapter 7 analysed band and transponder data from a separate, smaller study of birds marked as adults. Future work will anal- yse the combined data for birds banded as chicks and for birds banded as adults, including those data arising from the banding effects study. When estimating annual survival rates from live re-sightings of African Penguins, P. Whittington (pers. comm.) found that the survival estimate was lower in the year following banding compared to subsequent years. To further investigate this, he replaced the true banding date with that of the first live re-sighting, provided that it was at least one year after the true banding date. The next re-sighting, if there was one, became the first re-sighting after the new artificial banding date. Whittington still recorded lower survival probabilities in the year following the artificial banding date when compared to subsequent years. Future work will determine whether this phenomenon occurs with the Little Penguin dataset. Chapter 7 showed that the survival probability of banded adult Little Penguins was considerably lower than that of unbanded adults. It is likely that banding also has a significant negative effect on the survival of juvenile Little Penguins, particu- larly in their first year of life, and that the first-year survival estimates obtained in this thesis are underestimates of the true survival probability for unbanded birds. Although no such data are currently available for birds banded as chicks, a study examining the effect of banding on chicks is underway on Phillip Island (P. Dann, pers. comm.). The data arising from this work will allow us to estimate the rate of band loss for chicks, and to determine the effect of banding on juvenile penguins.

196 Appendix A

Deriving and programming the likelihood

A.1 Deriving the likelihood

The likelihood function is formed by taking the product of the probabilities corre- sponding to the life histories of each of the individuals. I now provide an informal justification of the likelihood developed by Catchpole et al. (1998a), by grouping together birds whose individual histories have a particular form, working out how many birds fit into this category and so on. Here I use the notations and definitions of Catchpole et al. (1998a), as outlined on p.67. Each bird has a final non-zero history entry of either “2” (if it was recovered dead) or “1” (if it was last sighted alive). I begin by considering birds whose final non-zero entry was “2”.

Consider birds from cohort c = 2 (that is birds initially captured at time t2).

Possible histories for birds recovered dead in the interval (t4, t5) are as follows:

t1 t2 t3 t4 t5 t6 ... tk 010020 . . . 0 011020 . . . 0 010120 . . . 0 011120 . . . 0

Each of these birds has a likelihood contribution of φ2,2φ2,3(1 − φ2,4)λ2,4, or

α2,4(1 − φ2,4)λ2,4, corresponding to the survival and recovery parameters, regardless of their individual recapture histories. Since there are d2,4 birds from the second cohort which were recovered dead in (t4, t5), this expression must be raised to the power of d2,4 in the likelihood equation. Note that I have not yet included the recapture probabilities.

197 dc,j Hence an expression of the form {αc,j(1− φc,j)λc,j} would arise for birds from cohort c (1 ≤ c ≤ C) recovered in the interval (tj, tj+1), for c ≤ j ≤ k − 1, and so, all of the birds from cohort c recovered dead anytime during the study would produce a contribution of

k−1 dc,j {αc,j(1 − φc,j)λc,j} , for 1 ≤ c ≤ C (A.1) j=c Y corresponding to the survival and recovery components of the likelihood. I now consider birds whose final non-zero entry was “1”. Let c = 2 and let their

final year of recapture be t5. The possible recapture histories for this group of birds are:

t1 t2 t3 t4 t5 t6 ... tk 010010 . . . 0 011010 . . . 0 010110 . . . 0 011110 . . . 0 Ignoring the recapture probabilities at this stage, each of these birds has a likelihood contribution of the form α2,5χ2,5. Since there are v2,5 birds in this group,

v2,5 they contribute {α2,5χ2,5} to the likelihood. Therefore the appropriate term for the birds from cohort c seen alive for the last time on recapture occasion tj

vc,j is {αc,jχc,j} (1 ≤ c ≤ C, c ≤ j ≤ k). Hence the section of the likelihood corresponding to the birds of cohort c that were last encountered alive is

k vc,j {αc,jχc,j} , for 1 ≤ c ≤ C. (A.2) j=c Y I next work out the recapture terms for all of the birds in cohort c. Each bird has a pc,j or (1−pc,j) entry corresponding to every recapture occasion from tc+1 until the occasion before they are recovered dead or the occasion they are recaptured for the last time. That is, I only need a pc,j or (1−pc,j) corresponding to the mark-recapture occasion if the bird is known to be alive at that point. There are wc,j birds from cohort c recaptured at tj+1, and zc,j birds from this cohort not recaptured at this time but encountered later. Each bird from cohort c alive at tj+1 has a probability

198 of being recaptured of pc,j and a probability of not being recaptured of (1 − pc,j).

Therefore pc,j appears wc,j times and (1−pc,j) appears zc,j times. Hence the section of the likelihood corresponding to the recapture probabilities over all occasions for birds in cohort c is k−1 wc,j zc,j pc,j (1 − pc,j) . (A.3) j=c Y Neglecting constant multiples, the likelihood equation corresponding to all the birds of cohort c (1 ≤ c ≤ C) is simply the product of expressions A.1–A.3. That is, k−1 k k−1 dc,j vc,j wc,j zc,j {αc,j(1 − φc,j)λc,j} {αc,jχc,j} pc,j (1 − pc,j) . j=c j=c j=c Y Y Y Now taking the product over all of the cohorts, I obtain the following expression as given by Catchpole et al. (1998a):

C k−1 k k−1 dc,j vc,j wc,j zc,j L = const × {αc,j(1 − φc,j)λc,j} {αc,jχc,j} pc,j (1 − pc,j) . c=1 "j=c j=c j=c # Y Y Y Y (A.4) Using the likelihood in Equation A.4, I next derive an equivalent form as stated in Catchpole et al. (2000). Equation A.4 can be rewritten as

C k−1 k−1 k k k−1 dc,j dc,j vc,j vc,j wc,j zc,j L = const× {(1 − φc,j)λc,j} αc,j αc,j χc,j pc,j (1 − pc,j) . c=1 "j=c j=c j=c j=c j=c # Y Y Y Y Y Y (A.5)

k−1 dc,j k vc,j Now let Ac = j=c αc,j j=c αc,j . Therefore Equation A.5 becomes Q Q C k−1 k−1 k dc,j wc,j zc,j vc,j L = const × {(1 − φc,j)λc,j} (Ac) pc,j (1 − pc,j) {χc,j} . c=1 "j=c j=c j=c # Y Y Y Y (A.6)

199 If I append a column of zeros to matrix D so that D is now a C × k matrix with dc,k =0 for 1 ≤ c ≤ C, then I can write

k k−1 dc,j dc,k dc,j αc,j = αc,k × αc,j j=c j=c Y Y k−1 dc,j = αc,j . j=c Y

Therefore Ac can now be written as

k dc,j +vc,j Ac = αc,j j=c Y k j−1 dc,j +vc,j = φc,s by the definition of αc,j. j=c s=c Y Y

Now by taking the ln of |Ac| and then interchanging the order of summation I obtain

k j−1

ln |Ac| = (dc,j + vc,j) ln φc,s j=c s=c X X k−1 k

= (dc,j + vc,j) ln φc,s. s=c (j=s+1 ) X X

k I now verify that j=s+1(dc,j + vc,j)= wc,s + zc,s (for 1 ≤ c ≤ C) by considering the definitions of matricesP D, V, W and Z. In the following argument I consider only birds from cohort c (that is those banded initially in tc) for fixed c (1 ≤ c ≤ C).

wc,s represents the number of birds recaptured at ts+1, that is the set of indi- viduals with entries of “1” in the history column corresponding to ts+1, but not including the original banding occasion. It should be noted that these birds may or may not have been encountered in later years.

200 Therefore wc,s includes birds that were

• seen alive for the final time at ts+1,

• seen alive at ts+1, and seen alive for the final time at ts+2, or ts+3, ··· , or tk

(note that the birds may or may not have been encountered between ts+1 and their final encounter),

• seen alive at ts+1, and then found dead in the interval (ts+1, ts+2) (that is an

entry of “2” corresponding to the ts+2 column), or found dead in (ts+2, ts+3),

··· , or found dead in (tk−1, tk) (again birds may or may not have been en-

countered between ts+1 and the time they are found dead).

zc,s represents the number of birds not recaptured at ts+1 but seen again dead or alive later, that is birds with “0” entries in the ts+1 column but “1” or “2” in later years. Therefore, zc,s includes birds that were

• not seen in ts+1, but seen alive for the final time at ts+2, or ts+3, ··· , or tk

(note that the birds may or may not have been encountered between ts+1 and their final encounter),

• not seen in ts+1, but found dead in the interval (ts+1, ts+2) (with a “2” in the

ts+2 column), or dead in (ts+2, ts+3), ··· , or dead in (tk−1, tk) (again birds may

or may not have been encountered between ts+1 and the time they are found dead).

201 Hence wc,s +zc,s represents the number of birds recaptured at ts+1, or not recaptured at ts+1 but seen dead or alive later. More specifically I can think of wc,s + zc,s as

wc,s + zc,s = number seen alive for final time at ts+1

+number seen alive at ts+1, and seen alive for final time at ts+2

+number not seen in ts+1, but seen alive for final time at ts+2

+number seen alive at ts+1, and found dead in the interval (ts+1, ts+2)

+number not seen in ts+1, but found dead in the interval (ts+1, ts+2)+ ···

= number seen alive for final time at ts+1

+(number seen alive at ts+1, encountered dead/alive for final time at ts+2

+number not seen at ts+1, encountered dead/alive for final time at ts+2)

+ ···

= number seen alive for final time at ts+1

+number encountered dead or alive for final time at ts+2

+ ···

k I can rewrite j=s+1(dc,j + vc,j) as follows P k k

(dc,j + vc,j) = (vc,j + dc,j) j=s+1 j=s+1 X X = vc,s+1 +(dc,s+1 + vc,s+2)+(dc,s+2 + vc,s+3)+ ··· + dc,k.

Now consider the meanings of the terms in the above equation. Recall that dc,k is defined as 0, and that vc,s+1 represents the number of birds seen alive for the final time at ts+1. I also note that dc,s+1 is the number of birds recovered dead in the interval (ts+1, ts+2) (that is the number of birds with a “2” in the history column corresponding to ts+2), whereas vc,s+2 is the number of birds recaptured for the final time at ts+2. Therefore dc,s+1 + vc,s+2 represents the number of birds encountered

202 dead or alive for the final time at ts+2. I can interpret dc,s+2 + vc,s+3 and so on in a similar manner. Therefore k

(dc,j + vc,j)= wc,s + zc,s, j=s+1 X and so

k−1

ln |Ac| = (wc,s + zc,s) ln φc,s s=c Xk−1 wc,s+zc,s = ln φc,s . s=c X

Hence

k−1 wc,s+zc,s Ac = φc,s . s=c Y

Now substituting this expression for Ac back into the likelihood expression in Equation A.6 gives

C k−1 k dc,j wc,j +zc,j wc,j zc,j vc,j L = const × {(1 − φc,j)λc,j} φc,j pc,j (1 − pc,j) χc,j , c=1 "j=c j=c # Y Y Y (A.7) the likelihood equation derived by Catchpole et al. (2000) and appearing as Equa- tion 4.1 on p.69. This is the form of the likelihood that I have used in my current work. Ignoring constants, the log(likelihood) is now given by

C k−1 C k−1

ln(L) = dc,j{ln(1 − φc,j) + ln(λc,j)} + (wc,j + zc,j) ln φc,j c=1 j=c c=1 j=c X X X X C k−1 C k

+ (wc,j ln pc,j + zc,j ln(1 − pc,j)) + vc,j ln (χc,j) . c=1 j=c c=1 j=c X X X X

203 The log(likelihood) can be easily programmed in MATLAB (refer to program lik.m1 as

sum(sum(D. ∗ (log(1 − Phi) + log(Lam)))) + sum(sum((W + Z). ∗ log(Phi)))

+sum(sum(W. ∗ log(P )+ Z. ∗ log(1 − P ))) + sum(sum(V. ∗ log(χ))), (A.8) where 1 is a matrix of ones.

A.2 Calculating the likelihood

The above likelihood equation is written in terms of the sufficient statistics D, V, W, Z. The form of the likelihood allows the easy analysis of various subsets of the data, namely: for the birds banded as chicks, or those banded as adults, or for both chicks and adults, and also using dead recovery information only, or live recaptures, or combined recapture–recovery data. There are thus nine possible subsets of data that can be analysed. For example, I will often want to consider the combined recapture–recovery data for birds banded as chicks. I must therefore set up sufficient statistics for birds banded as chicks (with a subscript of “ch”), and those banded as adults (with subscript “ad”), as well as statistics when only dead recoveries are used (subscript “dead”) and when only live recaptures are considered (“live”).

A.2.1 Model notation

In the MATLAB programs, models are specified in the order Survival/Recapture/ Recovery, or φ/p/λ. For example, Age/C/T means that the survival probability φ is age dependent, the recapture probability p is constant, and the recovery probability λ is time dependent.

φage contains the age structure for the survival probability, whereas φvar refers to the dependence of the survival probabilities for each of the “age groups” defined in φage. (I define page, λage, pvar and λvar in a similar manner.)

For example, φage=[1;2;33] means that I define separate survival probabilities for penguins from three age groups: those in their first year of life, those in their next

[email protected]

204 two (or second and third) years of life, and then those in their 4th and subsequent years of life. Furthermore, if φvar=[T;C;C], then the survival probability for birds in their first year is time dependent, whereas it is constant for birds in their second and third years, and for those in their later years.

Furthermore, if φage=[1;1;4;30] then the survival probability is divided into com- ponents corresponding to the first year of life, the next (or second) year of life, the next four (or 3rd–6th) years of life, and the next 30 (or 7th–36th) years of life.

Therefore, if φvar=[T;C;C;T], first year survival is time dependent, second year sur- vival is constant, while for those in their 3rd–6th years of life the annual survival probability is the same for each age level and is constant over time, and the annual survival probability is the same for each age level but varies over time for those in their 7th–36th.

A.2.2 Fitting the models

The MATLAB programs used for the analysis in Chapters 4, 5 and 8 are based on code written by E. Catchpole to analyse sheep data in Catchpole et al. (2000). The main difference between the sheep analysis and my current work is that some of the sheep are banded as adults of known age, whereas the penguins that were banded as adults were of unknown age. Therefore the programs were carefully adjusted in order to allow the incorporation of adults of unknown age into the penguin analysis.

Specifying the model information My results program results.m begins by specifying the type of analysis I want to carry out, for example using combined recovery–recapture data for birds banded as chicks. I define the global string variable “type” which can take on the values ‘recov-recap’ (when I want to include all information on live recaptures and dead recoveries), ‘recovery’ (for dead recoveries only) or ‘recapture’ (for live recaptures only), and “data” which holds values of ‘all’ (if I wish to include all banded birds), ‘ch’ (for birds banded as chicks only) or ‘ad’ (for birds banded as adults only).

If φ is to be age dependent I next enter appropriate values for the φage and φvar vectors which specify the age structure for φ and the various dependences for each of the age components. (Similarly for p and λ.)

205 If there is any covariate dependence (either for a parameter, or for an age com- ponent of a parameter), I enter the covariate matrices as necessary. Here phiCOV refers to the covariate matrix for φ, pCOV1 refers to the covariate matrix for the first age component of p and so on. For example, phiCOV1 may be a matrix with three columns — column 1: annual mean fledgling weight, column 2: annual mean laying date, column 3: mean number of chicks fledged per pair.

Calculating the total number of parameters I next specify the model to be fitted using the notation outlined in the previous sec- tion, and invoke the model function model.m to calculate the number of parameters needed. For example I may specify model(’Age/Age/C’). Function model.m reads in the model name, Age/Age/C, and uses string commands to pick out the sections of the model name corresponding to φ, p and λ. It then cycles through φ, p and λ one by one and works out the number of parameters corresponding to each. Note that the maximum possible age of a bird banded as a chick in the first year of the study, t1, and still alive in the final year of the study, tk, is k − 1. Now consider the φ loop, since the same principles hold for p and λ. If the model name for φ is ‘C’ (that is survival is taken to be constant), the number of parameters for φ is simply ‘1’, whereas φ obviously has k − 1 parameters if it is taken to be time-dependent. If φ is covariate dependent with covariate matrix phiCOV, then the number of parameters required for φ is the number of columns of phiCOV plus one (that is, one parameter corresponding to the constant term and one for each of the regression coefficients corresponding to each column of the covariate matrix). The most complicated part of model.m deals with the case when φ is age depen- dent. In this case, the program works out the number of parameters corresponding to each group of age cohorts of φ by checking the appropriate entry of φvar to de- termine the type of dependence for that age component. It then increments the number of parameters for φ accordingly. If the type of dependence for a group of cohort is ‘C’ it increments the number of parameters for φ by one, whereas if the dependence is ‘V’, it adds (1+number of columns of the covariate matrix) param- eters. If the dependence is ‘T’ it adds the following number of parameters: (k − 1 − the number of cohorts already accounted for in the previous groups of cohorts).

For example, if φage=[1;2;33] and φvar=[C;C;T], then there would be one parameter

206 each for the groups of cohorts 0–1 and 1–3 and then 35 − 3 = 32 parameters for the 3–36 age group. φ estimates for 1967, 1968 or 1969 are not required for the birds in the 3–36 group, since the birds banded as chicks in the first year of the study do not reach this age group until 1970. Note that if I include the data for birds banded as adults only, I consider only the number of parameters corresponding to the final entries of φage and φvar, that is the age component for the oldest group of birds. If I wish to include birds banded as chicks together with those banded as adults, the total number of parameters for φ will be as explained above. The only exception is when the dependence for the adult group is ‘T’. Again consider φage=[1;2;33] and φvar=[C;C;T]. If I want to include birds banded as adults, I need an additional set of parameters for φ in 1968, 1969 and 1970. Once the same procedure has been repeated to calculate the number of param- eters for p and λ, model.m determines the total number of parameters required for any model specified in the results program. If I am using the data for dead recoveries only, the number of parameters for p is set to zero, whereas the number of parameters for λ is set to zero for a live recaptures only analysis. Note that these programs use the logit scale to force the actual parameter values (which correspond to probabilities) to be between 0 and 1.

The likelihood program In program lik.m I calculate the negative log-likelihood corresponding to the model that is being fitted. I begin by setting the p matrices for birds banded as chicks and adults to zero if I am using recovery only data. The corresponding λ matrices would be set to zero for an analysis of live recaptures only. Again I cycle through the φ, p and λ parameters, determine the form of the parameter matrices corresponding to the model for each and finally calculate the negative log-likelihood by substituting the appropriate parameter matrices and the sufficient statistic matrices (calculated in setup3.m) in the code for the log-likelihood specified in Equation A.8. (See explanation below for a discussion of how this is done for the various subsets of data.)

207 The lik.m function calls up the appropriate Pmat function to determine the form of the parameter matrix corresponding to the model name for that parameter. For example PmatC.m is called up when the parameter is constant, PmatT.m when it is time dependent, PmatV.m for covariate dependence and PmatAge.m for age depen- dence. Note that these matrices are comparable to the parameter index matrices in SURGE or Program MARK. Consider possible forms of the upper-triangular φ parameter matrix. Recall that I am actually working with model parameters on a logit scale, so that θ = logit(φ), say, where the dimension of vector θ corresponds to the number of parameters for φ as determined by program model.m. I abbreviate ilogit(θ) as ilt(θ). If survival is taken to be constant over time, then the parameter matrix for th survival, φ, has an (i, j) entry of φi,j = ilogit(φ1). For time-dependent survival, the corresponding parameter matrix for φ has φi,j = ilogit(φj), where ilogit(θ1) corresponds to the survival from the 1968 calendar year, and so on. Next let survival depend on one or more covariates. For example, assume that the covariate matrix, phiCOV, is of the form

w1 l1 . . phiCOV =  . .  ,    w36 l36      where the columns represent the annual mean chick weight and the annual mean laying date, respectively. The parameter matrix for φ has φi,j = ilogit(θ1 + wjθ2 + ljθ3), where θ1, θ2 and θ3 are the regression coefficients corresponding to the constant term, chick weight and laying date respectively. The covariate values must be suitably scaled to avoid a potential problem with the magnitude of the covariates. If the magnitudes are too large, MATLAB will issue the error message — Warning: log of zero. This is because a large negative value for (θ1 + wjθ2 + ljθ3) will result in the ilogit of this expression going to zero, which in turn causes difficulty in calculating the log of φ. When survival is age dependent, the matrix obviously becomes more compli- cated. Let us consider the following age structure and dependence for survival:

208 φage=[1;2;33] and φvar=[V;T;C], with phiCOV=[w1; ··· ; w36]. The parameter ma- trix for φ is of the form:

ilt(θ1 + w1θ2) ilt(θ3) ilt(θ4) ilt(θ38) ··· ilt(θ38)  .. .  ilt(θ1 + w2θ2) ilt(θ4) ilt(θ5) . .  ......   . . . ilt(θ38)    ,  .. ..   . . ilt(θ37)     ..   . ilt(θ37)       ilt(θ1 + w36θ2)    with the diagonals of the parameter matrix corresponding to the various age groups of birds. Here θ1 and θ2 are the regression coefficients corresponding the survival for the 0–1 age group, ilt(θ3), ··· , ilt(θ37) correspond to the survivals from 1969–2003 for the 1–3 age group, and ilt(θ38) is the constant survival for the 3–36 age group. To work out the φ parameter matrix corresponding to birds banded as chicks, I simply use the methods outlined above. However, if survival is taken to be age dependent and I am including birds banded as adults of unknown age, I must consider the oldest age group of cohorts, as specified by the final elements of φage and φvar. If the dependence for this “adult” group is ‘C’, I use function PmatC, and so on. Care was needed to ensure that the correct components of the parameter estimate vector were passed into the Pmat functions. For further details refer to the code for lik.m. In forming the likelihood for the dead recovery only data, I delete the terms involving the recapture (p) parameter matrix (sum(sum(W. ∗ log(P )+ Z. ∗ log(1 − P )))) from the likelihood (A.8), whereas for live recaptures only, I delete the term sum(sum(D. ∗ (log(1 − Phi) + log(Lam)))).

Estimation of parameters The model parameters are estimated using the maximum likelihood method. This is obviously equivalent to maximising the log-likelihood, which in turn corresponds to minimising the negative log-likelihood. In this analysis I use the constr (or fmincon) function in MATLAB’s Optimisation Toolbox which finds the constrained minimum of a multivariable function. It should be noted that although constr is capable of minimising a function subject to certain specified equality constraints, I have chosen to include the constraints more directly by programming design matrices

209 when required. However, I use lower and upper bounds (denoted by lb and ub) on the parameter estimates, to stop them from going to −∞ or +∞ (lb=0 and ub=1 on a probability scale).

Programming a linear decrease in survival with age

I next consider a model Age/C/C with φage=[1;1;34] and φvar=[C;C;V(age)]. The matrix of survival probabilities, with elements φc,j defined as above, is set up as follows: Each element of the main diagonal of the φ matrix (corresponding to the survival for birds of age 0–1) is of the form ilogit(θ1), while the terms in the sub- diagonal above the main diagonal, which correspond to the 1–2 age group, are of the form ilogit(θ2). The remaining terms in the φ matrix, which contain the adult survival probability, are of the form φa = ilogit(θ3 + θ4age). Note that I impose a continuity constraint, whereby the first survival estimate in the linear decrease is equal to the previous constant estimate. That is, in this example, the survival estimate for birds aged 2–3 is set to be equal to the survival estimate for age 1–2.

Programming an Age+time model

Consider a model with the following age structure, {φ1, φ2, φ3, φ4:8, φ9+; p1, p2, p3:4, p5+; λ1, λ2:3, λ4+}, but with an Age+time dependence for p (model 5.14).

The adult recapture probability, p5+, is allowed to be fully time-dependent and I fix the other components of the recapture probability to be equal to a constant plus the adult probability in that year (on a logit scale). An Age*time model for p (model 5.13) would result in a total of 141 parameters (five for φ, 133 for p and three for λ). When p is Age+time dependent, there are 51 parameters (with 43 parameters corresponding to p). To program this model, I set up a design matrix which transforms the reduced set of 51 parameters into the 141 parameters of the standard Age*time model. In this way, the logit of the first year recapture probability in a particular year is made to equal a constant plus the logit of the adult recapture rate in that year. The notation p1,j refers to the recapture probability for th the one-year-old birds in the j breeding season, p5+,j denotes the adult recapture probability (for birds aged five years and older) in the jth season, and so on. Hence the equations corresponding to the recapture probabilities in the design matrix are as follows:

210 logit(p1,j) = θj + k1,

logit(p2,j) = θj + k2,

logit(p3:4,j) = θj + k3,

logit(p5+,j) = θj, for j =5,..., 35.

It should be noted that the recapture rate in the first breeding season is only th defined for the first year birds, and p5+ is only defined from the 5 season onwards. Therefore, I need to define additional recapture probabilities for the other age groups in the earlier years (ie. p1,1,...,p1,4,p2,2,...,p2,4,p3:4,4).

A.3 Sufficient statistics

The following pages contain the values of the sufficient statistics, Dch, Vch, Wch, Zch,Zch live,Zch dead, Vch live, Vch dead, that have been used in the analyses in Chapters 4, 5 and 8 for the Phillip Island penguins.

211 Table A.1: Matrix Dch 10000000000000000000000000 0000000000 2000000000000000000000000 0000000000 2432003032230100100000000 0000000000 26201140230202000000000 0000000000 11510512250000000000000 0000000000 1513220010200000000000 0000000000 21110000100000000000 0000000000 33231172201000001000 0000000000 1411110112000010000 0000000000 12000100000000000 0000000000 17521201220010010 0000100000 9110201410100120 0000000000 181412022000001 0000000000 15914010000110 0210000000 1701000120000 0200000000 93210310111 0220000000

212 6000100101 0010000000 1431000010 1010000000 551221210 0000000100 29531231 0100001001 2321220 0000101000 362611 2120011003 2410 4 1 3 3 3 11 1 1 2 01 45 610 310 4 63 1 5 3 13 10 5 313 2 37 1 5 0 02 50 13 8 3 04 3 3 1 02 15512 3 13 1 3 2 13 44 6 02 4 2 1 14 40 10 2 3 2 00 175 1 2 1 33 3 1 0 0 00 14 9 2 3 1 16 1 0 2 14 4 1 18 1 7 Table A.2: Matrix Vch 160000000000 0000000 00000 00 0 0 0000000000 65024000000 3110000 00000 00 0 0 0000000000 41414623217 1001100 01100 01 0 0 0000000000 3703314255 3021001 20000 00 0 0 0000000000 602531484 3262021 00000 00 0 0 0000000000 34710214 0110100 00000 00 0 0 0000000000 701002 0120100 01000 00 0 0 0000000000 821026 1171423 01000 00 0 0 0000000000 38803 1220111 00000 00 1 1 0000000000 791 0100000 01000 00 0 0 0000000000 4196115212621 42000 11 0 0 030000000 246341420 21000 22 0 0 0000000000 295331053 31012 41 0 0 4010000000 3713775 37103 55 1 1 1201000000 453146 24433 03 0 5 3000000000 20828 75134 23 5 3 3000110000

213 1591 01120 21 0 1 1000100020 247 03301 10 0 1 1101201000 11532215 64 1 4 1121011010 703374 76 5 1 4402101110 77024 55 4 6 7422101202 971 3 21 14 9 14 15 4 7 10 1 1 2 2 3 3 993 13 12 9 11 14 4 2 3 6 3 4 7 2 13 1341 9 21 21 28 12 17 11 8 5 11 9 10 18 687 5 16 22 9 9 8 5 5 9 2 4 11 1451 13 32 6 11 7 9 12 8 6 8 25 2072 13 13 12 7 14 12 11 10 10 17 1582 10 8 12 8 6 5 4 1 16 336 43 3 3 1 1 07 623 5 5 7 9 3 8 18 91 1 2 1 0 01 591 8 5 9 8 20 478 2 6 10 12 508 4 7 11 414 4 19 373 12 Table A.3: Matrix Wch 000000000000000000000000000000000000 13610012141100000000000000000000000 212848891244231101110111000000000000 575101381094211012000000000000000000 10610121615 8 5 7 2 0 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2363522101000000000000000000000 200411202011100000000000000000 381387926240100000000000000000 1533214221112211210000000000 101000000100000000000000000 122412 81910 5 3 7 5 3 4 3 4 3 2 2 2 3 0 0 0 0 0 0 0 6538576312032000000000000 4122113 810 9 6 3 6 6 3 2 2 5 1 1 0 0 0 0 0 0 0 72417151016 9 91312 9 4 4 2 2 0 1 0 0 0 0 0 0 4 812 51511 77 17 4 63 00 0 00 00 0 0 314131712 813 5 8 9 5 4 1 0 0 1 1 0 0 0 0

214 22685363123211211120 1530343433323201000 5551097464222331010 51814201712 9 8 6 5 5 3 3 2 2 2 0 6131916151516 9 8 6 6 4 2 4 1 2 15 54 50 35 39 35 22 19 18 9 11 6 9 5 3 27 41 36 33 37 22 20 22 22 18 16 16 10 13 39 84 78 85 59 59 45 41 43 36 30 23 18 13 39 52 29 34 26 18 17 16 9 9 11 30 58 37 41 45 39 37 30 27 26 25 20 37 47 35 47 37 31 29 23 17 18 23 30 30 22 16 12 10 16 4 610 8 5 7 6 7 9 16 25 24 20 21 18 222111 13 12 20 20 20 3 12 17 12 6 11 11 10 19 12 Table A.4: Matrix Zch 000000000000000000000 0 0 0 0 0 0000000000 108345543411000000000 0 0 0 0 0 0000000000 5036343226201710 9 6 5 4 4 3 4 2 2 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 00 50393630222413 9 9 8 7 4 5 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00 5246383522141513 4 3 3 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00 20171110544330000000 0 0 0 0 0 0000000000 10109455330100000 0 0 0 0 0 0000000000 45382826251513 6 6 1 2 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 00 211614121095522221 0 1 1 0 0 0000000000 532221111000 0 0 0 0 0 0000000000 6239383621181614 7 5 5 3 4 3 2 2 2 2 1 1 0 0 0 0 00 382523151496675 7 4 2 0 0 0000000000 503822191712 8 811 7 5 4 4 3 0 0 0 0 0 0 0 0 00 6536352626161615 11 9 6 5 4 5 2 1 0 0 0 0 0 00 40352627151514 9 12 6 6 4 2 0 0 0 0 0 0 0 00 62 46 37 25 25 25 16 20 14 9 7 5 3 2 2 1 0 0 0 0 0

215 15141078 8 5 5 6 4 2222111100 21141212 9 7 7 5 6 43 2 11 10 0 00 383531 23 18 12 10 7 5 6 5 3 1 1 2 1 00 6039 33 22 16 12 9 9 6 3 3 3 2 2 1 00 50 39 28 24 18 14 7 7 4 4 1 2 3 0 10 114 70 47 47 33 21 18 15 9 8 4 7 2 4 0 107 70 59 52 34 32 27 22 18 15 13 7 6 0 196 136 111 80 68 52 43 33 22 19 11 8 0 133 97 65 53 37 33 26 21 12 10 8 0 144 90 71 58 43 38 28 20 14 9 0 128 86 60 59 37 32 23 12 7 0 72 51 36 22 18 16 14 11 0 26 19 12 9 6 1 1 0 61 44 29 21 15 8 0 4 2 0 0 00 52 36 21 9 0 30 18 7 0 21 8 0 14 0 0 Table A.5: Matrix Zch live 000000000000000000000 00 0 0 0 0000000000 108345543411000000000 00 0 0 0 0000000000 4029262418139432323232211 00 0 0 0 0000000000 36313024171710 6 5 4 3 2 3 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00 41403330191111 9 4 3 3 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00 151276322110000000 00 0 0 0 0000000000 777344220100000 00 0 0 0 0000000000 3226181615994401000 00 0 0 0 0000000000 14987652322221 01 1 0 0 0000000000 221111111000 00 0 0 0 0000000000 492828271310 9 9 4 2 2 1 2 1 1 1 1 1 0 0 0 0 0 0 00 2117159732232 41 0 0 0 0000000000 4029161313 8 5 710 6 4 3 3 3 0 0 0 0 0 0 0 0 00 5028272222131312 8 5 3 3 2 2 1 1 0 0 0 0 0 00 35302223111010 710 4 4 2 0 0 0 0 0 0 0 0 00 534032212019 1114 9 5 4 2 1 2 2 1 0 0 0 00

216 1413967 74 5 5 4 2222111100 151099 64 4 3 4 3221110000 302826 2016 12 10 7 5 6 5 3 1 1 1 1 00 4730 2717 12 10 8 8 5 2 2 2 1 2 1 00 45 3524 22 17 13 6 6 3 2 0 1 3 0 10 104 61 41 41 28 18 15 14 8 6 3 6 1 2 0 82 54 45 39 24 25 21 16 13 11 10 6 5 0 163 106 88 59 54 40 35 27 17 14 9 6 0 104 73 41 37 23 22 21 16 11 9 6 0 112 70 59 48 32 31 23 18 13 8 0 102 71 47 47 28 24 18 9 5 0 57 40 25 12 10 10 9 7 0 1812 5 4 4 1 10 48 36 21 15 9 5 0 3 2 0 0 00 38 30 17 8 0 29 18 6 0 16 7 0 13 0 0 Table A.6: Matrix Zch dead 000000000000000000000000000000000000 00000000000000000000000000000000000 20171515151212 9 7 5 2 2 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00000 231715151413 9 9 7 4 4 2 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00000 2116151510 9 7 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00000 111075333220000000000000000000000 432111110000000000000000000000 2018151413 6 4 2 2 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 00000 9876554311111000000000000000 311110000000000000000000000 18131110 8 8 7 5 3 3 3 2 2 2 1 1 1 1 1 1 0 00000 2312121010 9 5 4 4 3 3 3 2 0 0 0 0 0 0 0 00000 13128755311111100000000000 201110 6 6 5 5 5 5 5 4 3 3 3 1 0 0 0 00000 6655554222222000000000 1714121111 8 7 7 6 5 4 4 2 0 0 0 00000

217 44443332211100000000 7433333221100000000 1097542111111111000 181310 9 7 4 3 3 2 2 2 2 21110 9764222222110000 2018121110 8 7 5 5 5 43330 31 21 17 16 13 10 7 6 5 4 3 1 1 0 55 49 39 36 26 22 16 13 12 7 4 3 0 41 36 33 20 18 15 8 7 2 2 2 0 37 24 16 13 13 9 6 3 2 2 0 29 17 14 13 10 9 6 4 3 0 20 14 14 12 8 6 5 4 0 8 7 7 52000 15 10 9 7 6 3 0 1 00000 15 6 4 1 0 3 2 2 0 5 1 0 1 0 0 Table A.7: Matrix Vch live 1700000000000000000 000 0 0 00 0 0 0000000000 670240000003110000 000 0 0 00 0 0 0000000000 448176253291101100 011 0 0 01 0 0 0000000000 38633145556321001 200 0 0 00 0 0 0000000000 6245317975362021 000 0 0 00 0 0 0000000000 367224140210100 000 0 0 00 0 0 0000000000 7520020120100 010 0 0 00 0 0 0000000000 8671372471423 010 0 0 00 0 0 0000000000 409131220111 000 1 0 00 1 1 0000000000 8310100000 010 0 0 00 0 0 0000000000 449912531262 142 0 0 01 1 0 0030000000 272541522 310 0 0 22 0 0 0000000000 323341153 410 1 2 41 0 0 4010000000 4015875 371 0 4 55 1 2 1201000000 475146 254 3 3 03 0 5 3000000000 22629 765 3 5 23 5 3 4000110000

218 1661 012 2 0 22 0 1 1100100020 267 133 0 1 10 0 1 1101201000 121622 1 5 65 1 4 1121021010 7454 7 4 87 5 1 5402101120 798 3 5 56 4 6 7432101202 1017 4 24 15 9 14 16 4 7 11 1 2 2 3 4 3 1042 14 14 9 11 14 6 3 3 6 3 4 7 2 13 1419 12 24 22 31 14 19 12 8 10 11 10 11 18 726 5 19 27 9 9 9 6 6 9 3 4 11 1533 14 32 7 12 7 10 12 8 6 9 25 2253 14 14 12 7 14 12 11 10 11 17 1641 12 8 13 10 6 5 4 1 16 384 43 3 3 1 1 07 653 5 6 7 10 3 8 18 95 1 2 1 0 01 619 9 5 9 8 20 495 2 7 11 12 527 4 7 11 433 4 19 380 12 Table A.8: Matrix Vch dead 1600000000000000000 000 0 0 00 0 0 0000000000 760000000000000000 000 0 0 00 0 0 0000000000 446000000000000000 000 0 0 00 0 0 0000000000 40200000000000000 000 0 0 00 0 0 0000000000 6430000000000000 000 0 0 00 0 0 0000000000 358000000000000 000 0 0 00 0 0 0000000000 7800000000000 000 0 0 00 0 0 0000000000 8490000000000 000 0 0 00 0 0 0000000000 401000000000 000 0 0 00 0 0 0000000000 8200000000 000 0 0 00 0 0 0000000000 4750000000 000 0 0 00 0 0 0000000000 267000000 000 0 0 00 0 0 0000000000 33600000 000 0 0 00 0 0 0000000000 4230000 000 0 0 00 0 0 0000000000 491000 000 0 0 00 0 0 0000000000 25600 000 0 0 00 0 0 0000000000

219 1720 000 0 0 00 0 0 0000000000 262 000 0 0 00 0 0 0000000000 118600 0 0 00 0 0 0000000000 7500 0 0 00 0 0 0000000000 817 0 0 00 0 0 0000000000 1080 0 00 0 0 0000000000 1096 0 0 0 0 0 0 00 0 0 0 0 00 1521 0 0 0 0 0 00 0 0 0 0 00 792 0 0 0 0 00 0 0 0 0 00 1588 0 0 0 00 0 0 0 0 00 2191 0 0 00 0 0 0 0 00 1652 0 00 0 0 0 0 00 358 00 0 0 0 0 00 6780 0 0 0 0 00 96 0 0 0 0 00 641 0 0 0 0 0 508 0 0 0 0 530 0 0 0 437 0 0 385 0 Appendix B

Codes for Raw Data

B.1 Methods of Encounter and Status Codes

B.1.1 Methods of encounter

00 METHOD NOT COVERED BY CODE

01 PROBABLY TRAPPED DEVICE UNKNOWN

02 TRAPPED DEVICE UNKNOWN

03 TRAPPED IN MIST NET

04 TRAPPED IN CAGE TRAP

05 TRAPPED WITH CANNON NET

06 TRAPPED IN CLAP TRAP, SPRUNG TRAP, ETC

07 TRAPPED WITH BAL-CHATRI/NOOSE CARPET

08 TRAPPED BY HAND OR WITH HANDHELD NET

09 TRAPPED USING LIGHT DEVICE

0A TRAPPED WITH A DHO-GHAZA

0R LOCATED USING RADIO TELEMETRY

10 TRAPPED IN HARP TRAP

11 TRAPPED IN MONOFILAMENT MIST NETS

12 TRAPPED WITH TRIP WIRE OVER WATER

1A HAND CAUGHT AT NEST

13 HAND CAUGHT AT ROOST OR NEST

14 TRAPPED AS ATTRACTED TO DOMESTIC BIRDS

15 DELIBERATELY TRAPPED FOR THE AVIARY

16 TRAPPED BECAUSE BAND TANGLED IN NATURAL OBJECT

17 TRAPPED BECAUSE BAND TANGLED IN HUMAN OBJECT

220 18 TRAPPED BECAUSE BAND TANGLED IN FISHING GEAR

19 TRAPPED BECAUSE BIRD TANGLED IN NATURAL OBJECT

20 TRAPPED BECAUSE BIRD TANGLED IN HUMAN OBJECT

21 TRAPPED BECAUSE BIRD TANGLED IN FISHING GEAR

2A CAUGHT ON LONGLINE

22 TRAPPED ACCIDENTALLY IN TRAP FOR TERRESTRIAL ANIMALS

23 TRAPPED ACCIDENTALLY IN MARINE/AQUATIC ANIMAL TRAP

24 TRAPPED USING NARCOTIC DRUGS

25 FOUND SICK OR INJURED

26 EXHAUSTED

27 INJURED BY BAND

28 OILED

29 BURNT OR SCORCHED BY FIRE

30 FOUND NEAR ELECTRICITY WIRES

31 COLLIDED WITH A MOVING ROAD VEHICLE

32 COLLIDED WITH A MOVING TRAIN

33 COLLIDED WITH A MOVING AIRCRAFT

34 COLLIDED WITH A MOVING SHIP

35 COLLIDED WITH A LIGHTHOUSE OR STATIONARY NIGHT LIGHT

36 COLLIDED WITH A WINDOW OR OTHER TRANSPARENT MATERIAL

37 COLLIDED WITH A BUILDING,NON-WIRE FENCE,IMMOBILE VEHICLE

38 COLLIDED WITH A MAST,TOWER,POLE,WIRE FENCE,AERIAL,SPRINKLER

39 FOUND ON HIGHWAY/ROAD; BUT NOT CERTAINLY HIT BY CAR

3A COLLIDED WITH A NATURAL OBJECT EG TREE, CLIFF

40 BAND FOUND ON BIRD, NO FURTHER DATA ON METHOD OF ENCOUNTER

41 BAND RETURNED, NOT REPORTED IF BAND ON BIRD

42 BAND ONLY FOUND (USE STATUS=02)

43 BAND NUMBER ONLY REPORTED

44 BAND LOST

45 BAND DESTROYED OR DAMAGED

46 COLOUR MARKING SIGHTED IN FIELD (COHORT ONLY) (USE STATUS=26)

47 BAND NUMBER READ IN FIELD (BIRD NOT TRAPPED) (USE STATUS=26)

221 48 COLOUR MARKING SIGHTED IN FIELD (BAND NO. INFERRED) (USE STATUS=26)

49 BAND NUMBER/COLOUR MARKING SIGHTED ON BIRD ON NEST (USE STATUS=26)

50 CAPTIVE BRED BIRD/BAT

51 SUCKLING YOUNG HAND RAISED (BATS ONLY)

52 NESTLING HAND RAISED (ABANDONED, ORPHANED OR NEST DESTROYED)

53 BANDED AFTER DEATH FOR EXPERIMENT

54 FOUND FLOATING IN SEA OR FRESHWATER OR BEACHWASHED

55 FOUND IN/ON CAR,SHIP,ETC PROBABLY ENCOUNTERED ELSEWHERE

56 TRAPPED/KILLED BECAUSE IT WAS BANDED

57 BAND FOUND ON SPECIES DIFFERENT TO THAT BANDED (SEE FILE)

58 LEG (OR WING) AND BAND ONLY FOUND (USE STATUS=02)

59 COLOUR MARKER FOUND NOT ON BIRD (BAND NO. INFERRED)

5A COLOUR MARKER FOUND ON WING ONLY (BAND NO. INFERRED) (USE STA-

TUS=05)

60 READABLE BAND SIGHTED, NO. ON STANDARD BAND INFERRED (USE STATUS=26)

61 SHOT (REASON UNKNOWN)

62 SEIZED FOR LAW ENFORCEMENT REASONS

63 TAKEN FOR SCIENTIFIC STUDY (NOT BANDING)

64 TAKEN TO PROTECT CROPS

65 TAKEN TO PROTECT DOMESTIC ANIMALS

66 TAKEN FOR AIRCRAFT STRIKE PREVENTION PROGRAM

67 TAKEN FOR FOOD, FEATHERS, CEREMONIAL REASONS

68 SHOT FOR SPORT/FOOD

69 TAKEN FOR HUMAN HEALTH REASONS

6A TAKEN FOR NATURE CONSERVATION REASONS

70 SHOT WITH ARROW OR SPEARED

71 ACCIDENTLY INJURED/KILLED IN EXPLOSION

72 POISONED - UNKNOWN IF INTENTIONAL

73 POISONED - NATURAL SOURCE

74 UNINTENTIONALLY POISONED BY BAIT FOR OTHER ANIMALS

75 UNINTENTIONALLY POISONED BY AERIAL SPRAYING OF CROPS

76 UNINTENTIONALLY POISONED BY INDUSTRIAL WASTES

222 77 INTENTIONALLY POISONED WITH BAITS

78 INTENTIONALLY POISONED BY AERIAL SPRAYING FOR BIRDS

79 LEAD POISONED (LEAD SHOT)

80 TAKEN BY UNKNOWN ANIMAL

81 TAKEN BY DOMESTIC OR WILD CAT

82 TAKEN BY DOMESTIC OR WILD DOG

83 TAKEN BY DOMESTIC ANIMAL (SPECIES?)

84 TAKEN BY A WILD MAMMAL (SPECIES?)

85 TAKEN BY A WILD BIRD (SPECIES?)

86 TAKEN BY A WILD FISH (SPECIES?)

87 TAKEN BY A WILD REPTILE (SPECIES?)

88 CARCASS BEING EATEN BY SCAVENGING BIRDS

89 FOUND DEAD OR INJURED AFTER A STORM

90 OBSOLETE CODE

91 OBSOLETE CODE

92 INJURED OR KILLED BY HUMAN (NOT FOR FOOD)

93 INJURED/DIED DURING EXPERIMENTAL ACTIVITIES

94 ELECTROCUTED

95 FOUND IN STILL WATER

96 CAPTIVE BIRD/BAT (WAS FROM WILD) (USE STATUS=13)

97 FOUND INSIDE A MAN MADE STRUCTURE

98 FOUND DEAD IN/NEAR A NEST (PULLI AND ADULTS)

99 FOUND DEAD, CAUSE UNKNOWN

9A BANDING DATA UNKNOWN (SEE ’NO SCHEDULES’ FILE)

B.1.2 Status codes

00 STATUS OF BIRD/BAT AND BAND IS UNKNOWN

01 STATUS OF BIRD/BAT IS UNKNOWN AND THE BAND WAS LEFT ON

02 STATUS OF BIRD/BAT IS UNKNOWN AND THE BAND WAS REMOVED

03 WAS DEAD AND THE STATUS OF THE BAND IS UNKNOWN

04 WAS DEAD AND THE BAND WAS LEFT ON

223 05 WAS DEAD AND THE BAND WAS REMOVED

06 WAS MERCY KILLED AND THE STATUS OF THE BAND IS UNKNOWN

07 WAS MERCY KILLED AND THE BAND WAS LEFT ON

08 WAS MERCY KILLED AND THE BAND WAS REMOVED

09 REHABILITATION ATTEMPTED BUT BIRD/BAT DIED, STATUS OF BAND UNKNOWN.

10 REHABILITATION ATTEMPTED BUT BIRD/BAT DIED, BAND LEFT ON

11 REHABILITATION WAS ATTEMPTED BUT BIRD/BAT DIED, BAND WAS REMOVED

12 WAS RELEASED ALIVE, STATUS OF BAND IS UNKNOWN

13 WAS RELEASED ALIVE WITH THE BAND

14 WAS RELEASED ALIVE AND THE BAND WAS REPLACED DUE TO WEAR, ETC

15 WAS REHABILITATED & RELEASED ALIVE, BAND STATUS IS UNKNOWN

16 WAS REHABILITATED & RELEASED ALIVE WITH THE BAND

17 WAS REHABILITATED & RELEASED ALIVE, BAND WAS REMOVED

18 IS ALIVE IN CAPTIVITY AND STATUS OF BAND IS UNKNOWN

19 IS ALIVE IN CAPTIVITY WITH BAND

20 IS ALIVE IN CAPTIVITY AND BAND WAS REMOVED

21 ALIVE: UNKNOWN IF RELEASED OR CAPTIVE, BAND STATUS UNKNOWN

22 ALIVE BUT UNKNOWN IF RELEASED OR CAPTIVE, BAND WITH BIRD/BAT

23 ALIVE: UNKNOWN IF RELEASED OR CAPTIVE, BAND REMOVED

24 TRANSPORTED TO NEW SITE AND RELEASED WITH BAND

25 TRANSPORTED TO NEW SITE AND BAND REMOVED

26 WAS ALIVE IN THE WILD WITH THE BAND

27 PARTIALLY DECOMPOSED AND BAND STATUS UNKNOWN

28 PARTIALLY DECOMPOSED AND BAND LEFT ON

29 PARTIALLY DECOMPOSED AND BAND REMOVED

30 WAS SKELETON/DRIED OUT CORPSE, BAND STATUS UNKNOWN

31 WAS SKELETON/DRIED OUT CORPSE, BAND LEFT ON

32 WAS SKELETON/DRIED OUT CORPSE, BAND REMOVED

33 FLEW AWAY WITHOUT THE BAND

34 RELEASED ALIVE & THE BAND WAS REPLACED DUE TO INJURY

35 STATUS OF BIRD/BAT UNKNOWN

36 RELEASED ALIVE, BAND REMOVED AND NOT REPLACED DUE TO INJURY

224 37 RELEASED ALIVE CARRYING 2 OR MORE BANDS

38 DIED DURING CAPTURE, BAND STATUS IS UNKNOWN

39 DIED DURING CAPTURE, BAND LEFT ON

40 DIED DURING CAPTURE, BAND REMOVED

41 KILLED IN NET BY PREDATOR, BAND STATUS IS UNKNOWN

42 KILLED IN NET BY PREDATOR, BAND LEFT ON

43 KILLED IN NET BY PREDATOR, BAND REMOVED

44 RELEASED ALIVE AND THE BAND WAS REMOVED

45 RELEASED ALIVE WITH BAND AND ELECTRONIC TAG

99 DIED BEFORE BANDING

225 B.2 Experimental Codes

101 Comment in field notes 102 ExA Stomach flush 103 ExR Radio-tracking 104 Translocation 204 Translocation Recovery 105 ExB CuSulphate Ex 106 ExW Injected and bled (metabolism) 107 ExA + ExR 008 Extra Food 108 Extra food (given) 208 Extra food Recovery 308 Extra food Recovery 109 Drenched 209 Drenched Recovery 110 Body Dump 210 Dump Recovery 111 Experiment other institution 112 Reband inc. transponder 013 Band only in transp. Exp. 113 Tag only in transp. Exp. 213 Tag and band in transp. Exp. 114 Fix Band 115 Band Wear 116 Bled for Genetic Study 117 30 Burrows for disturbance Exp. 118 Additional 20 Burrows for disturbance Exp. 119 Depth Gauge Exp. 120 Nobbies Development Relocation 121 Alison Kemp chicks (honours) 122 Returned from other institution 123 Darkened Bands

226 124 Stomach flushed + Blood Taken 125 Carpark Relocation chicks 126 Band Recovery - no information known

227 B.3 Location Codes

B.3.1 Phillip Island Location Codes 1 Penguin Parade 3 Cliffs to West 4 Northern Shore 25 Cowes 28 Sunderland Estate, North of Leonards Rd 29 Sunderland Estate, South of Leonards Rd 65 Sunderland Bay/Surf Beach 114 Smith Beach 214 Red Rocks Beach 222 229 Newhaven Beach 235 Kitty Miller Bay 237 Forrest Caves 265 Berry Beach 271 Cape Woolamai 273 Ventnor Beach 288 Thorny Beach 298 Cleeland Bight 307 Storm Bay 322 Woolshed Bight 332 Rhyll 362 Flynns Reef Penguin Reserve 363 Pyramid Rock 372 Swan Lake 383 Silverleaves

228 B.3.2 General Location Codes

Latitude Longitude LOC NAME DESCRIPTION DEGS MINS DEGE MINE 1 PENGUINPD PENGUINPARADEPHILLIPISLANDVIC 38 31 145 8 2 CHALKYIS CHALKYISLANDTAS 40 5 147 53 3 CLIFFS WST PHILLIP ISLAND CLIFFS TO WEST VIC 38 31 145 8 4 NTHSHORE NORTHERNSHOREPHILLIPISLANDVIC 38 31 145 7 5 PT CAMPBEL PORT CAMPBELL (LONDON BRIDGE) VIC 38 37 142 55 6 PTCAMPBEL PORT CAMPBELL (12 APOSTLES) LOCH ARD 38 40 143 6 VIC 7 CATARAQUI CATARAQUI POINT KING ISLAND TAS 40 3 143 51 8 GABOIS GABOISLANDVIC 37 33 149 54 9 GLENNIEIS GREATER GLENNIE ISLAND WILSONS PROM 38 5 146 13 VIC 10 RABBIT IS RABBIT ISLAND VIC 38 54 146 30 11 RABBITROCK RABBIT ROCK WILSONS PROMONTORY VIC 38 54 146 29 12 WLDNESSCK KINGISLANDWILDERNESSCREEKTAS 40 4 143 52 13 GRASSYIS GRASSYISLANDKINGISLANDTAS 40 3 144 4 14 EASTCOVE EASTCOVEDEALISLANDKENTGROUPTAS 39 28 147 18 15 LTLESQUAL LITTLESQUALLYCOVEDEALISKENTGROUP 39 29 147 20 TAS 16 GNDCOVE GARDEN COVE 39 27 147 19 TAS 17 BOOBIALLA BOOBIALLACOVEDEALISLANDKENTGROUP 39 28 147 19 TAS 18 TORQUAY TORQUAY BEACH TORQUAY VIC 38 20 144 19 19 MILE IS MILE ISLAND TAS 40 7 147 54 20 MELBZOO MELBOURNEZOOLOGICALGARDENSVIC 37 47 144 57 21 BREAKWATER ST KILDA MARINA BREAKWATER VIC 37 53 144 58 22 HAMPTON HAMPTON PORT PHILLIP BAY VIC 37 56 144 59 23 LONSDALE PTLONSDALEPORTPHILLIPBAYVIC 38 18 144 38 24 MUD IS PORT PHILLIP BAY VIC 38 17 144 36 25 COWES COWES PHILLIP ISLAND VIC 38 27 145 14 26 LAWRE ROCK LAWRENCE ROCK NEAR PORTLAND VIC 38 24 141 40 27 WATTLE IS WATTLE ISLAND WILSONS PROMONTORY VIC 39 8 146 21 28 SLDESTN SLDESTATENORTH OFSTLEONARDSROAD 38 31 145 8 VIC 29 SLDESTS SLDESTATE SOUTHOFSTLEONARDSROAD 38 31 145 8 VIC 30 PORTLAND PORTLAND HARBOUR VIC 38 21 141 38 31 MIDDLE IS MIDDLE ISLAND SOUTH WARRNAMBOOL VIC 38 24 142 28 32 MERRI IS MERRI ISLAND SOUTH WARRNAMBOOL VIC 38 24 142 28 33 DINGHYCOVE DINGHY COVE VIC 38 24 142 0 34 SEAL BAY SEAL BAY LADY JULIA PERCY ISLAND VIC 38 25 142 2 35 PENGUINIS PENGUINISLANDBEACHPORTSA 37 29 140 0 36 GRANITE IS GRANITE ISLAND VICTOR HARBOUR SA 35 33 138 37 37 ALBATROSS ALBATROSS ISLAND TAS 40 22 144 39 38 SPENCERF SPENCES REEFS LADY BARRON FLINDERS IS 40 13 148 14 TAS 39 TOOLINA CV TOOLINA COVE GREAT AUSTRALIAN BIGHT 32 44 125 1 WA 40 CITADELIS CITADEL IS GLENNIE GROUP WILSONS PROM 39 6 146 14 VIC 41 MCHUGHIS MCHUGH IS GLENNIEGROUP WILSONSPROM 39 7 146 14 VIC 42 AIREYSINLT AIREYS INLET VIC 38 28 144 6 43 DOLEYKI DOLEYKANGAROOISLANDSA 35 47 137 51 44 PTMARSDEN POINTMARSDENKANGAROOISLANDSA 35 33 137 37 45 PENGUIN PENGUINTASMANIATAS 41 7 146 4 46 THREE HUMM TAS 40 26 144 54 47 GREEN IS BIG GREEN ISALND TAS 40 11 147 58 48 PASSAGE IS PASSAGE ISLAND FURNEAUX GROUP TAS 40 30 148 20 49 RODONDOIS RODONDOISLANDNRWILSONSPROMTAS 39 13 146 23 50 ERITH IS DEAL GROUP TAS 39 26 147 17 51 CPGANTHEA CAPEGANTHEAUMEKANGAROOISLANDSA 36 4 137 27 52 VIVONNE BH VIVONNE BEACH KANGAROO ISLAND SA 35 59 137 12 53 PENNESHAW CHRISTMAS COVE PENNESHAW KANGAROO 35 42 137 56 ISLAND SA 54 BROWNS BCH BROWNS BEACH KANGAROO ISLAND SA 35 47 137 50 55 NORMAN IS WILSONS PROMONTORY VIC 39 1 146 14 56 CP CASSINI CAPE CASSINI KANGAROO ISLAND SA 35 34 137 19 57 FORT IS SOUTH CHANNEL ISLAND (THE FORT) VIC 38 18 144 48 58 SEALISLAN SEAL GROUP WILSONS PROM 38 55 146 39 VIC 59 NOTCHIS NOTCHISLANDSEALGROUPWILSONSPROM 38 56 146 40 VIC

229 60 HANSONBAY HANSONBAYKANGAROOISLANDSA 36 1 137 51 61 BARRALLIER BARRALLIERISLANDWESTERNPORTVIC 38 17 145 19 62 AMERICAN R AMERICAN RIVER KANGAROO ISLAND SA 35 47 137 45 63 RAGISLAND RAGISLANDSEALGROUPWILSONSPROMVIC 38 57 146 40 64 KANGAROOIS KANGAROOISLAND(GENERAL)SA 35 43 137 55 65 SUNDERLAND SUNDERLANDBAY/SURFBCHPHILLIPISVIC 38 30 145 16 66 MAIDSTONE MAIDSTONE VIC 37 47 144 52 67 KANOWNA KANOWNAISLANDWILSONSPROMVIC 39 9 146 18 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 XE PORT PHILLIP DUMP EXP VIC 38 10 144 45 96 XD WP DUMP EXPERIMENT NOBBIES-FLINDERS 38 31 145 4 VIC 97 XA WP DUMP EXPERIMENT COWES-TORTOISE 38 27 145 15 HEAD VIC 98 XB WP DUMP EXPERIMENT TANKERTON- 38 22 145 14 FAIRHAVEN VIC 99 XC WPDUMPEXPERIMENTLONGIS-QUAILIS 38 16 145 17 VIC 100 FAIRHAVEN FAIRHAVEN FRENCH IS WESTERN PORT VIC 38 20 145 17 101 LAKES ENTR LAKES ENTRANCE VIC 37 52 147 59 102 KILCUNDA KILCUNDA VIC 38 33 145 28 103 FISHERMANS FISHERMANS BEACH MORNINGTON PENNIN- 38 13 145 1 SULA VIC 104 POWLETT POWLETT RIVER NEAR DAYLESTON VIC 38 35 145 36 105 SEAFORD BH SEAFORD BEACH PORT PHILLIP BAY VIC 38 7 145 8 106 GLENELG MOUTH OF GLENELG RIVER OCEAN GROVE 38 3 141 0 VIC 106 NELSON NELSONMOUTHOFGLENELGRIVERVIC 38 3 141 1 107 SHOREHAM SHOREHAM BEACH PORT PHILLIP BAY VIC 38 26 145 3 108 WYE RIVER WYE RIVER GREAT OCEAN ROAD VIC 38 37 143 53 109 CPE JERVIS CAPE JERVIS SA 35 36 138 5 110 OCEANGRAN OCEANGRANGEGIPPSLANDLAKESVIC 37 57 147 45 111 PTMCDONNEL PORT MACDONNELL SA 38 3 140 41 112 ROBE ROBE SA 37 9 139 45 113 CPBRGWATER CAPE BRIDGEWATER VIC 38 23 141 24 114 SMITH BEAC SMITH BEACH PHILLIP ISLAND VIC 38 31 145 15 115 EASTVIEW EASTERNVIEWNEARAIREYSINLETVIC 38 28 144 3 116 NIRRANDA NIRRANDA VIC 38 30 142 46 117 SKENES CRK SKENES CREEK GREAT OCEAN ROAD VIC 38 43 143 43 118 KERRY GREE BEACH KERRY GREEN CAPE SCHANK VIC 38 30 144 53 118 CAPE SCHAN CAPE SCHANCK VIC 38 30 144 53 119 NEWSURF NEW SURFERS BEACH SA 35 31 138 45 120 DOUGLASPT DOUGLASPOINTNENEVALLEYSA 38 1 140 34 121 SOMERS SOMERS WESTERN PORT VIC 38 23 145 9 122 SWAN LAKE SWAN LAKE DISCOVERY BAY VIC 38 12 141 18 123 CARPENTRKS CAPRENTER ROCKS NEAR MOUNT GAMBIER 37 54 140 24 SA 124 PAMBULA BH PAMBULA BEACH NSW 36 55 149 54 125 WILLIAMSON WILLIAMSONS BEACH WONTHAGGI VIC 38 36 145 35 126 BLACK R BLACK RIVER SOUTHERN TASMANIA TAS 40 50 145 17 127 SOMERSET SOMERSETTAS 41 2 145 49 128 WHISKY BH WHISKY BAY WILSONS PROMONTORY VIC 39 1 146 17 129 GOOLWA GOOLWA SA 35 30 138 46 130 BREAMLEA BREAMLEA VIC 38 17 144 23

230 131 MIDPARK BH MIDDLE PARK BEACH PORT PHILLIP BAY VIC 37 51 144 57 132 TAMBOON TAMBOON INLET VIC 37 45 149 7 133 SAFETY BEA SAFETY BEACH MORNINGTON PORT PHILLIP 38 19 144 59 BAY VIC 134 JAN JUC JAN JUC BEACH NEAR TORQUAY VIC 38 20 144 19 135 TOWRADGI TOWRADGI BEACH NSW 34 23 150 54 136 NEWCASTLE MAIN BEACH NEWCASTLE NSW 32 56 151 46 137 REDRIVER REDRIVERNRWINGANINLETVIC 37 44 149 30 138 CPLIPTRAP CAPELIPTRAPNEARWILSONSPROMONTORY 38 54 145 55 VIC 139 CONRAN CAPE CONRAN VIC 37 48 148 43 140 LEWISHAM LEWISHAMTAS 42 49 147 36 141 ROSEBUD ROSEBUD BEACH PORT PHILLIP BAY VIC 38 21 144 54 142 HALLETT HALLETTCOVESA 35 4 138 29 143 PTLONSDAL BACK BEACH PT LONSDALE (OCEAN BEACH) 38 18 144 37 VIC 144 JAMIESON JAMEISON CREEK GREAT OCEAN ROAD VIC 38 35 143 56 145 PORT LATTA PORT LATTA TAS 40 51 145 22 146 DANGERPT DANGERPOINTSA 38 2 140 48 147 CHILDERS CHILDERS COVE VIC 38 29 142 40 148 SEALERS SEALERSCOVEWILSONSPROMONTORYVIC 39 0 146 26 149 EDEN EDENNSW 37 3 149 54 150 90 MILE BH 90 MILE BEACH VIC 38 5 147 35 151 PATTERSON CAPE PATTERSON NEAR WONTHAGGI VIC 38 40 145 36 152 153 SANDRINGHA SANDRINGHAM BEACH PORT PHILLIP BAY VIC 37 56 145 0 154 SHELLEY SHELLEY BEACH CAPE OTWAY (ROTTEN 38 21 141 26 POINT) VIC 155 STHELENS STHELENSPOINTTAS 41 16 148 21 156 OPOSSUM BY OPOSSUM BAY BEACH TAS 42 59 147 24 157 SAN REMO SAN REMO VIC 38 31 145 22 158 CORIO BAY CORIO BAY PORT PHILLIP BAY VIC 38 6 144 24 159 OBERON BAY OBERON BAY WILSONS PROMONTORY VIC 39 4 146 20 160 SURPRISE B SURPRISE BAY KING ISLAND TAS 40 7 143 54 161 DARBY BCH DARBY BEACH WILSONS PROMONTORY VIC 38 58 146 16 162 PORTARLING PORTARLINGTON PORT PHILLIP BAY VIC 38 7 144 39 163 CUMBERLAND CUMBERLAND RIVER GREAT OCEAN ROAD 38 34 143 54 VIC 164 WHITEMARK WHITEMARK TAS 40 7 148 0 165 SANDY PT B SANDY POINT BEACH WARRATAH BAY VIC 38 49 146 7 166 LOWHEAD LOWHEADTAS 41 3 146 47 167 SENTIMENTA SENTIMENTAL POINT FLINDERS ISLAND TAS 40 1 147 51 168 MOTHERS BH MOTHERS BCH MORNINGTON PORT PHILLIP 38 12 145 2 BAY VIC 169 ELWOOD BCH ELWOOD BEACH PORT PHILLIP BAY VIC 37 53 144 59 170 EAST BEACH EAST BEACH TAS 41 3 146 48 171 GOLDEN BCH GOLDEN BEACH 90 MILE BEACH VIC 38 13 147 24 172 TOWER HILL TOWER HILL BAR BEACH WARRNAMBOOL VIC 38 20 142 21 173 GRANTVILLE GRANTVILLE HARBOUR TAS 41 49 145 1 174 BRUSHISLA BRUSHISLANDNSW 35 31 150 25 175 TATLOWS BH TATLOWS BEACH STANLEY TAS 40 46 145 17 176 ADVENT BAY ADVENTURE BAY TAS 43 21 147 20 177 RYE OCEAN RYE OCEAN BEACH VIC 38 25 144 49 178 LAKETYRES LAKETYRESGIPPLANDLAKESVIC 37 50 148 8 179 NORTH PT NORTH POINT FLINDERS ISLAND TAS 39 42 147 56 180 CURRIE CURRIE KING ISLAND TAS 39 55 143 51 181 BADGER BOX BADGER BOX CREEK BEACH KING ISLAND TAS 39 58 143 52 182 MULGRAVE MULGRAVE FREEWAY NEAR HAMPTON PARK 38 2 145 16 VIC 183 HARMERS HA HARMERS HAVEN VIC 38 39 145 34 184 MALLACOOTA MALLACOOTA VIC 37 33 149 45 185 SHIPWRECK SHIPWRECK CREEK VIC 37 37 149 39 186 DALMENY YABRA BEACH DALMENY NSW 36 9 150 7 187 WALLAGOOT WALLAGOOT BEACH NSW 36 47 149 57 188 TATHRA BCH TATHRA BEACH NSW 36 43 149 59 189 WAIRO BCH WAIRO BEACH NSW 35 24 150 26 190 DENHAMS DENHAMS BEACH NEAR BATEMANS BAY NSW 35 44 150 13 191 SYDENHAM SYDENHAMINLETVIC 37 45 148 58 192 PARADISE PARADISE BEACH VIC 38 11 147 25 193 WYNYARD WYNYARD TAS 40 59 145 44 194 MERIMBULA TURA BEACH MERIMBULA NSW 36 49 149 56 195 THE COVE THE COVE NEAR WARRNAMBOOL VIC 38 29 142 40 196 WINGANINL WINGAN INLET CROAJINALONG NATIONAL 37 44 149 30 PARK VIC 197 FORTESCUE FORTESCUE BAY TAS 43 8 147 57 198 CLIFTON CLIFTON SPRINGS PORT PHILLIP BAY VIC 38 9 144 34 199 FITZROY FITZROY RIVER VIC 38 9 141 42 200 BLANKET BY BLANKET BAY VIC 38 49 143 35 201 MT WILLIAM MT WILLIAM NATIONAL PARK TAS 40 52 148 13 202 ALDINGA ALDINGABEACHSA 35 17 138 26

231 203 NULLAWARRE OCEAN BEACH SOUTH OF NULLAWARRE VIC 38 32 142 44 204 MTDEFIANCE MOUNT DEFIANCE VIC 38 36 143 57 205 CANADIAN CANADIAN BAY MT ELIZA PORT PHILLIP BAY 38 10 145 5 VIC 206 NARACOOPA NARACOOPA KING ISLAND TAS 39 55 144 7 207 PTMELBOURN PORT MELBOURNE VIC 37 50 144 56 208 JOHANNA JOHANNA BEACH GREAT OCEAN ROAD VIC 38 46 143 23 209 CHEVIOT CHEVIOT BEACH PORT PHILLIP BAY VIC 38 18 144 40 210 PT COOK PT COOK PIER PORT PHILLIP BAY VIC 37 56 144 48 211 INVERLOCH INVERLOCH VIC 38 37 145 43 212 ANSONSBAY ANSONSBAYTAS 41 2 148 16 213 GLENELGBH GLENELGBEACHADELAIDESA 41 2 148 16 214 RED ROCKS RED ROCKS BEACH PHILLIP ISLAND VIC 38 27 145 12 215 APOLLO BAY APOLLO BAY VIC 38 44 143 40 216 GELLIBRAND GELLIBRAND RIVER GREAT OCEAN ROAD VIC 38 42 143 10 217 PT ELLIOT HORSE-SHOE BAY/KNIGHTS BEACH/BOOMER 35 32 138 41 BCH SA 218 CP MONTESQ CAPE MONTESQUIEU VIC 38 8 141 10 219 YOUNGHUSB SALT CREEK YOUNG HUSBAND PENNINSULA 36 7 139 36 SA 220 GLENAIRBH GLENAIRBEACHSA 38 47 140 26 221 MCINTYRE MCINTYRE BEACH SA 37 35 139 8 222 SEAL ROCKS SEAL ROCKS OFF PHILLIP ISLAND VIC 38 31 145 6 223 BRIGHTON BRIGHTON YACHT CLUB PORT PHILLIP BAY 37 56 144 58 VIC 224 DOUBLE CRK DOUBLE CREEK FLINDERS VIC 38 29 145 2 225 BOOLAGOON BOOLAGOON HEATH CONSERVATION PARK SA 37 58 140 41 226 WHITES BCH WHITES BEACH VIC 38 19 144 21 227 VICTOR HAR SURFERS BEACH SOUTH VICTOR HARBOR SA 35 38 138 29 228 SOUTHMELB SOUTH MELBOURNE BEACH PORT PHILLIP 37 51 144 57 BAY VIC 229 NEWHAVEN NEWHAVEN BEACH PHILLIP ISLAND VIC 38 31 145 21 230 231 INDENTED INDENTED HEAD PORT PHILLIP BAY VIC 38 8 144 43 232 MENTONE MENTONE BEACH PORT PHILLIP BAY VIC 37 59 145 4 233 PTNEPEAN POINTNEPEANVIC 38 18 144 39 234 MT MARTHA MOUNT MARTHA PORT PHILLIP BAY VIC 38 16 145 0 235 KITTY MILL KITTY MILLER BAY PHILLIP ISLAND VIC 38 30 145 10 236 WERRIBEE WERRIBEE SOUTH PORT PHILLIP BAY VIC 37 57 144 43 237 FORREST FORREST CAVES PHILLIP ISLAND VIC 38 31 145 18 238 WILLIAMSTO WILLIAMSTOWN PORT PHILLIP BAY VIC 37 52 144 54 239 BLACK ROCK BLACK ROCK BCH BEAUMARIS PT PHILLIP 37 58 145 1 BAY VIC 240 BARWON HEA BARWON HEADS VIC 38 16 144 29 241 GRANGEBCH GRANGEBEACHSA 34 55 138 29 242 MCCRAE MCCRAE BEACH PORT PHILLIP BAY VIC 38 21 144 55 243 WOODSWELL WOODSWELL(CATTLEISLAND)SA 36 0 139 28 244 KOONYA KOONYA OCEAN BEACH VIC 38 21 144 45 245 YAMBUK YAMBUK NEAR PT FAIRY VIC 38 18 142 3 246 PORT FAIRY PORT FAIRY VIC 38 22 142 15 247 FLIND BK B FLINDERS BACK (OCEAN) BEACH VIC 38 29 145 2 248 BLAIRBOWRI BLAIRGOWRIE VIC 38 21 144 46 249 PORKY BCH PORKY BEACH KING ISLAND TAS 39 50 143 52 250 KILLARNEY KILLARNEY BEACH NEAR WARRNAMBOOL 38 21 142 18 VIC 251 NARRAWONG NARRAWONG VIC 38 15 141 42 252 LORNE LORNE VIC 38 32 143 58 253 WRIGHT BAY WRIGHT BAY SA 37 2 139 44 254 OCEAN GVE OCEAN GROVE VIC 38 15 144 31 255 ALTONA ALTONA BEACH PORT PHILLIP BAY VIC 37 52 144 49 256 KENNETT KENNETT RIVER GREAT OCEAN ROAD VIC 38 40 143 52 257 ANGLESEA ANGLESEAVIC 38 24 144 11 258 MARLO MARLO NEAR BAIRNSDALE VIC 37 47 148 31 259 DISASTER DISASTER BAY BEN BOYD NATIONAL PARK 37 15 149 58 NSW 260 QUEENSCLIF QUEENSCLIFF PORT PHILLIP BAY VIC 38 16 144 39 261 VENUS BAY VENUS BAY VIC 38 39 145 42 262 PORTSEA PORTSEA PORT PHILLIP BAY VIC 38 19 144 42 263 HAYBOROUGH HAYBOROUGHSA 35 32 138 38 264 SQUEAKY BH SQUEAKY BEACH WILSONS PROMONTORY VIC 39 1 146 18 266 BERRY BCH BERRY BEACH PHILLIP ISLAND VIC 38 31 145 12 267 DISCOVERY DISCOVERY BAY VIC 38 18 141 20 268 LAKE BONNY LAKE BONNEY MILLICENT SA 37 43 140 19 269 NORMAN BAY NORMAN BAY WILSONS PROMONTORY VIC 39 2 146 19 270 GARFISH BY GARFISH BAY CAPE OTWAY VIC 38 51 143 31 271 CAPE WOOLA CAPE WOOLAMAI PHILLIP ISLAND VIC 38 33 145 21 272 KANGAROO KANGAROO ISLAND SA 35 48 137 51 273 VENTNOR BH VENTNOR BEACH PHILLIP ISLAND VIC 38 28 145 10 274 RICKETTS RICKETTS POINT PORT PHILLIP BAY VIC 37 59 145 2 275 CARRUM BCH CARRUM BEACH PORT PHILLIP BAY VIC 38 4 145 7

232 276 NINE MILE NINE MILE BEACH KING ISLAND TAS 39 43 144 6 277 DROMANA DROMANA BEACH PORT PHILLIP BAY VIC 38 18 144 59 278 PT TURTON POINT TURTON SA 34 56 137 21 279 SMITHTON SMITHTON TAS 40 55 144 40 280 FRANKSTON FRANKSTON BEACH PORT PHILLIP BAY VIC 38 9 145 8 281 282 EDDYSTONE POINT TAS 40 59 148 21 283 NEWTONS NEWTONSBEACHNADGEENATURERESERVE 37 21 149 57 NSW 284 SEASPRAY SEASPRAY PORT PHILLIP BAY VIC 38 22 147 11 285 PETERBOROU PETERBOROUGH VIC 38 36 142 52 286 287 PT CAMPBEL PT CAMPBELL TOWN BEACH (BAY) VIC 38 37 142 59 288 THORNY BCH THORNY BEACH PHILLIP ISLAND VIC 38 30 145 10 289 STLEONARD STLEONARDSPORTPHILLIPBAYVIC 38 10 144 43 290 PTHENRY PTHENRYNEARGEELONGPORTPHILLIPBAY 38 7 144 25 VIC 291 EASTERN EASTERN BEACH GEELONG PORT PHILLIP 38 8 144 22 BAY VIC 292 HAZARDS HAZARDS BEACH FREYCINET PENNINSULA 42 10 148 17 TAS 293 294 295 MOGGS/FAIR BETWEEN MOGGS CREEK AND FAIRHAVEN 38 30 144 0 VIC 296 PTIMPOSSI POINTIMPOSSIBLENEARBARWONHEADSVIC 38 18 144 22 297 ASPENDALE ASPENDALEPORTPHILLIPBAYVIC 38 1 145 6 298 CLEELAND CLEELANDBIGHTPHILLIPISLANDVIC 38 32 145 20 299 CROPPIES CROPPIES POINT (NORTH EAST TAS) TAS 40 51 147 35 300 RIVERNOOK RIVERNOOKPOINTRONALDVIC 38 43 143 10 301 ANDERSON ANDERSONBAYTAS 40 57 147 25 302 CORNER INL VIC 38 46 146 19 303 TROUBRIDGE TROUBRIDGE SA 35 7 137 46 304 SORRENTO SORRENTO BACK BEACH VIC 38 20 144 44 305 PENNINGTON PENNINGTONBAYKANGAROOISLANDSA 35 51 137 44 306 STONY POIN STONY POINT WESTERN PORT VIC 38 22 145 13 307 STORM BAY STORM BAY PHILLIP ISLAND VIC 38 31 145 12 308 MERRICKS MERRICKS BEACH WESTERN PORT VIC 38 23 145 6 309 LADY BAY LADY BAY BEACH WARRNAMBOOL VIC 38 23 142 29 310 CRAYFISH CRAYFISH BAY CAPE OTWAY VIC 38 51 143 32 311 MOONLIGHT MOONLIGHT BEACH CAPE OTWAY VIC 38 44 143 12 312 TYRONE TYRONE FORESHORE BEACH NEAR RYE VIC 38 22 144 48 313 SISTERS SISTERS BEACH NORTH WEST TASMANIA TAS 40 55 145 33 314 SHIREHALL SHIREHALLBEACHMORNINGTONPTPHILLIP 38 13 145 2 B VIC 315 NORA NORA CRIENA BAY SA 37 19 139 50 316 UMPHERSTON UMPHERSTONEBAYCAPEDOUGLASSA 38 1 140 36 317 CINEMA PT CINEMA POINT VIC 38 29 144 2 318 WHITE BCH WHITE BEACH NUBEENA TAS 43 7 147 44 319 MIDDLETON SURFERS BEACH MIDDLETON SA 35 30 138 43 320 WAITPINGA WAITPINGA BEACH NEWLAND HEAD CONS 35 38 138 29 PARK SA 321 PORTLAND SHELLEY BEACH PORTLAND VIC 38 21 141 26 322 WOOLSHED WOOLSHEDBIGHTPHILLIPISLANDVIC 38 28 145 9 323 CANUNDA CANUNDANATIONALPARK(OILRIGSQUARE) 37 39 140 13 SA 324 POINTLEO POINTLEOVIC 38 25 145 4 325 ALLESTREE ALLESTREE BEACH NEAR PORTLAND VIC 38 17 141 38 326 GREEN PT GREEN POINT BEACH TAS 40 53 144 41 327 KINGSTON KINGSTON SE SA 36 50 139 51 328 TEA TREE TEA TREE CROSSING NEAR SALT CREEK SA 36 11 139 14 329 COTTERS BH COTTERS BEACH WILSONS PROMONTORY VIC 38 56 146 45 330 MORDIALLOC MORDIALLOC BEACH PORT PHILLIP BAY VIC 38 0 145 5 331 FRENCH ISL FRENCH ISLAND (NEAR TANKERTON) VIC 38 21 145 21 332 RHYLL RHYLL PHILLIP ISLAND VIC 38 28 145 18 333 LITTLE DIP LITTLE DIP CONSERVATION PARK SA 37 15 139 49 334 DESCARTES DESCARTES BAY VIC 38 20 141 22 335 GREENHEAD GREENHEADTASMANPENINSULATAS 42 57 147 40 336 GUNNAMATTA GUNNAMATTA SURF BEACH VIC 38 27 144 52 337 MASCOT GRI MASCOT GRID BLOCK BOTANY BAY NSW 33 0 151 11 338 PTADDIS POINTADDISNEARANGLESEAVIC 38 23 144 15 339 SUNDAY ISL CORNER INLET VIC 38 42 146 37 340 WEST IS WEST ISLAND VICTOR HARBOUR SA 35 36 138 35 341 CAPE FAREW CAPE FAREWELL KING ISLAND TAS 39 35 143 55 342 PHOQUES BY PHOQUES BAY KING ISLAND TAS 39 39 143 54 343 CASTLE COV CASTLE COVE NEAR CAPE OTWAY VIC 38 47 143 25 344 PARKER RIV PARKER RIVER NEAR CAPE OTWAY VIC 38 49 143 32 345 SPECTACLE SPECTACLE ISLAND FREDERICK HENDRICK 42 52 147 36 BAY TAS

233 346 TEMMA TEMMA WEST COAST TASMANIA TAS 41 13 144 41 347 PT MOOROWI PORT MOOROWIE YORK PENNINSULA SA 35 6 137 31 348 PELICAN PT PELICAN POINT NEAR MOUNT GAMBIA SA 37 56 140 25 349 CAPE NELSO MURRELB BEACH CAPE NELSON VIC 38 26 141 32 350 DUCK POINT DUCK POINT FRENCH ISLAND VIC 38 48 146 16 351 FRIENDLY FRIENDLY BEACH FREYCINET PENINSULA 42 1 148 16 TAS 352 LOCHARD LOCHARDGORGEPTCAMPBELLNPVIC 38 38 143 4 353 CAPE JAFFA CAPE JAFFA SA 36 57 139 40 354 SOUTHEND SOUTHENDFORESHORESA 37 34 140 8 355 BOGGY CK BOGGY CREEK KING ISLAND TAS 40 2 143 53 356 SOMERTON SOMERTON BEACH SA 35 0 138 31 357 ARDROSSAN ARDROSSANSA 34 28 137 55 358 PTWELSHPO PORTWELSHPOOLVIC 38 42 146 27 359 PORT PHILL PORT PHILLIP BAY VIC 38 6 144 53 360 BLACKFORD BLACKFORD DRAIN SA 36 47 139 51 361 DRUM IS DRUM ISLAND CORNER INLET VIC 38 43 146 39 362 FLYNNSREE FLYNNSREEFPENGUINRESERVEVIC 38 30 145 9 363 PYRAMID RK PYRAMID ROCK PHILLIP ISLAND VIC 38 31 145 13 364 365 KIAMA KIAMA NSW 34 39 150 51 366 MARSEY R MARSEY RIVER DEVONPORT TAS 41 10 140 22 367 DALMENY DALMENY BEACH NSW 36 9 150 7 368 FOOTSCRAY FOOTSCRAY IN MELBOURNE VIC 37 48 144 54 369 DANDENONG DANDENONGINMELBOURNEVIC 37 59 145 12 370 ARTHUR RIV ARTHUR RIVER TAS 41 5 144 40 371 MARTHA LAV MARTHA LAVINIA BEACH KING ISLAND TAS 39 39 144 4 372 SWAN LAKE SWAN LAKE PHILLIP ISLAND VIC 38 30 145 9 373 BAXTERS BAXTERS BEACH WONTHAGGI VIC 38 38 145 32 374 BASS STR BASS STRAIT (REHABILITATION) VIC 38 50 146 42 375 COOBOWIE COOBOWIE SA 35 2 137 43 376 BALES BCH BALES BEACH KANGAROO ISLAND SA 35 59 137 21 377 MCLOUGHLIN MCLOUGHLINS BEACH CORNER INLET VIC 38 36 146 53 378 SWAN IS NEAR QUEENSCLIFF VIC 38 15 144 41 379 RED BANKS RED BANKS KANGAROO ISLAND SA 35 44 137 43 380 PRESTON PRESTON IN MELBOURNE VIC 37 44 145 0 381 BARAGOOT BARABOOT BEACH NSW 36 28 150 4 382 BOAG ROCK BOAG ROCK NEAR GUNAMATTA VIC 38 26 144 50 383 SILVERLEAV SILVERLEAVES NEAR COWES VIC 38 27 145 15 384 PEARLPT PEARLPOINTWESTOFBEMMRIVERVIC 37 47 148 53 385 CHURCHILL NR PHILLIP ISLAND VIC 38 30 145 20 386 CP EVERARD CAPE EVERARD NR PT HICKS VIC 37 48 149 16 387 BROWN BAY BROWN BAY 15KM E PORT MACDONNELL SA 38 3 140 50 388 PICCANINNI PICANINNIE PONDS CREEK OUTLET SA 38 2 140 56

234 References

Aebischer N J and Coulson J C (1990). Survival of the Kittiwake in relation to sex, year, breeding experience and position in the colony, Journal of Animal Ecology, 59, 1063–1071. Ainley D G (2002). The Ad´elie Penguin: Bellwether of Climate Change, Columbia University Press, New York. Ainley D G, LeResche R E and Sladen W J L (1983). Breeding Biology of the Ad´elie Penguin, University of California Press, Berkeley. Akaike H (1973). Information theory and an extension of the maximum likelihood principle, in B N Petrov and F Csaki, eds, Second International Symposium on Information Theory, Academiai Kiado, Budapest, pp. 267–281. Alisauskas R T and Lindberg M S (2002). Effects of neckbands on survival and fidelity of white-fronted and Canada geese captured as non-breeding adults, Jour- nal of Applied Statistics, 29, 521–537. Anderson D R and Burnham K P (1999). General strategies for the analysis of ringing data, Bird Study, 46 suppl., 261–270. Anderson D R, Burnham K P and White G C (1985). Problems in estimating age-specific survival rates from recovery data of birds ringed as young, Journal of Animal Ecology, 54, 89–98. Anderson D R, Burnham K P and White G C (1994). AIC model selection in overdispersed capture-recapture data, Ecology, 75, 1780–93. Anderson D R and Sterling R T (1974). Population dynamics of molting pintail drakes banded in south-central Saskatchewan, Journal of Wildlife Management, 38, 226–274. Anderson D R, Wotawa M A and Rexstad E A (1993). Trends in the analysis of recovery and recapture data, in J D Lebreton and P M North, eds, Marked

235 Individuals in the Study of Bird Population, Birkh¨auser Verlag, Basel, pp. 373– 386. Arnason A N (1973). The estimation of population size, migration rates, and sur- vival in a stratified population, Research in Population Ecology, 15, 1–8. Arnason A N and Mills K H (1981). Bias and loss of precision due to tag loss in Jolly-Seber estimates for mark-recapture experiments, Canadian Journal of Fisheries and Aquatic Sciences, 38, 1077–1095. Arnason A N and Schwarz C J (1999). Using POPAN–5 to analyse banding data, Bird Study, 46 suppl., 157–168. Arnould J P Y, Dann P and Cullen J M (2004). Determining the sex of Little Penguins Eudyptula minor in northern Bass Strait using morphometric measure- ments, Emu, 104, 261–265. Ashmole N P (1963). The regulation of numbers in tropical oceanic birds, Ibis, 103B, 458–473. Bailey N T J (1951). On estimating the size of mobile populations from capture- recapture data, Biometrika, 38, 293–306. Barker R J (1997). Joint modelling of live recapture, tag-resight and tag-recovery data, Biometrics, 53, 666–677. Barker R J (1999). Joint analysis of mark-recapture, resighting and ring-recovery data with age-dependence and marking-effect, Bird Study, 46 suppl., 82–90. Barrowman N J and Myers R A (1996). Estimating tag-shedding rates for experi- ments with multiple-tag types, Biometrics, 52, 1410–1416. Barry S C, Brooks S P, Catchpole E A and Morgan B J T (2003). The analysis of ring-recovery data using random effects, Biometrics, 59, 54–65. Becker N G (1984). Estimating population size from capture-recapture experiments in continuous time, Australian Journal of Statistics, 26, 1–7. Besbeas P, Freeman S N, Morgan B J T and Catchpole E A (2002). Integrat- ing mark-recapture-recovery and census data to estimate animal abundance and demographic parameters., Biometrics, 58, 540–547. Beverton R J H and Holt S J (1957). On the dynamics of exploited fish populations, Fishery Investigations Series II, 19, 1–533.

236 Boersma P D (1998). Population trends of the Gal´apagos penguin: impacts of El Ni˜no and La Ni˜na, Condor, 100, 245–253. Bonner S J and Schwarz C J (2004). Continuous time-dependent individual covari- ates and the Cormack–Jolly–Seber model, Animal Biodiversity and Conservation, 27, 149–155. Bonner S J and Schwarz C J (2006). An extension of the Cormack–Jolly–Seber model for continuous covariates with application to Microtus pennsylvanicus, Bio- metrics, 62, 142–149. Bost C A and Le Maho Y (1993). Seabirds as bio-indicators of changing marine ecosystems: new perspectives, Acta Oecologia, 14, 463–470. Box G E P and Jenkins G M (1970). Time series analysis: forecasting and control, Holden-Day, London. Bradley J S, Wooller R D, Skira I J and Serventy D L (1989). Age-dependent sur- vival of breeding short-tailed shearwaters Puffinus tenuirostris, Journal of Animal Ecology, 58, 175–188. Bradshaw C J A, Barker R J and Davis L S (2000). Modeling tag loss in New Zealand fur seal pups, Journal of Agricultural, Biological and Environmental Statistics, 5, 475–485. Brooks S P, Catchpole E A and Morgan B J T (2000). Bayesian animal survival estimation, Statistical Science, 15, 357–376. Brownie C, Anderson D R, Burnham K P and Robson D S (1978). ‘Statistical inference from band recovery data: a handbook’, US Fish and Wildlife Resource Publication 131. Brownie C, Anderson D R, Burnham K P and Robson D S (1985). ‘Statistical inference from band recovery data: a handbook’, US Fish and Wildlife Resource Publication 156. 2nd edition. Brownie C and Robson D S (1976). Models allowing for age-dependent survival rates for band-return data, Biometrics, 32, 305–323. Buckland S T (1980). A modified analysis of the Jolly-Seber capture-recapture model, Biometrics, 36, 419–435. Buckland S T, Burnham K P and Augustin N H (1997). Model selection: an integral part of inference, Biometrics, 53, 603–618.

237 Bunce A (2000), Population dynamics of Australasian Gannets (Morus serrator) breeding in Port Phillip Bay, Victoria: Competition with fisheries and the po- tential use of seabirds in managing marine resources, PhD thesis, University of Melbourne, Australia. Unpublished. Burnham K P (1993). A theory for combined analysis of ring-recovery and recapture data, in J D Lebreton and P M North, eds, Marked Individuals in the Study of Bird Population, Birkh¨auser Verlag, Basel, pp. 199–213. Burnham K P and Anderson D R (1979). The composite dynamic method as evidence for age-specific waterfowl mortality, Biometrics, 43, 356–366. Burnham K P and Anderson D R (1992). Data-based selection of an appropriate biological model: the key to modern data analysis, in D R McCullough and R H Barret, eds, Wildlife 2001: populations, Elsevier Science Publishers, London, pp. 16–30. Burnham K P and Anderson D R (1998). Model Selection and Inference. A Practical Information-theoretic Approach, Springer, New York. Burnham K P and Anderson D R (2002). Model Selection and Multimodel Inference. A Practical Information-theoretic Approach, 2nd edn, Springer, New York. Burnham K P, Anderson D R and White G C (1985). Problems in estimating age-specific survival rates from recovery data of birds ringed as young, Journal of Animal Ecology, 54, 89–98. Burnham K P, Anderson D R, White G C, Brownie C and Pollock K H (1987). ‘Design and analysis methods for fish survival experiments based on release- recapture’, American Fisheries Society, Monograph 5, Bethesda, Maryland. Cairns D K (1989). The regulation of seabird colony size: a hinterland model, American Naturalist, 134, 141–146. Caldow R W G and Furness R W (2000). The effect of food availability on the foraging behaviour of breeding great skuas Catharacta skua and Arctic skuas Stercorarius parasiticus, Journal of Avian Ecology, 31, 367–375. Carrick R (1972). Population Ecology of the Australian Black-backed Magpie, Royal Penguin, and Silver Gull, in Population Ecology of Migratory Birds: A Symposium, U. S. Department of the Interior Wildlife Research Report 2, p. 278.

238 Caswell H (2001). Matrix Population Models: Construction, Analysis and Interpre- tation, Sinauer Associates, Sunderland, Massachusetts. 2nd edition. Catchpole E A (1995). Matlab: an environment for analyzing ring-recovery and recapture data, Journal of Applied Statistics, 22, 801–816. Catchpole E A, Fan Y, Morgan B J T and Coulson T N (2004). Sexual dimor- phism, survival and dispersal in red deer, Journal of Agricultural, Biological, and Environmental Statistics, 9, 1–26. Catchpole E A, Freeman S N and Morgan B J T (1993). On boundary estimation in ring recovery models and the effect of adding recapture information, in J D Le- breton and P M North, eds, Marked Individuals in the Study of Bird Populations, Birkh¨auser Verlag, Basel, pp. 215–228. Catchpole E A, Freeman S N and Morgan B J T (1995). Modelling age variation in survival and reporting rates for recovery models, Journal Applied Statistics, 22, 597–609. Catchpole E A, Freeman S N and Morgan B J T (1996). Steps to parameter redundancy in age-dependent recovery models, Journal of the Royal Statistical Society B, 58, 763–774. Catchpole E A, Freeman S N, Morgan B J T and Harris M P (1998a). Integrated recovery/recapture data analysis, Biometrics, 54, 33–46. Catchpole E A and Morgan B J T (1991). A note on Seber’s model for ring-recovery data, Biometrika, 78, 917–919. Catchpole E A and Morgan B J T (1994). Boundary estimation in ring-recovery models, Journal of the Royal Statistical Society B, 56(2), 917–919. Catchpole E A and Morgan B J T (1996). Model selection in ring-recovery models using score tests, Biometrics, 52, 664–672. Catchpole E A and Morgan B J T (1997). Detecting parameter redundancy, Biometrika, 84(2), 187–196. Catchpole E A, Morgan B J T, Coulson T N, Freeman S N and Albon S D (2000). Factors influencing Soay sheep survival, Journal of Applied Statistics, 49, 453– 472. Catchpole E A, Morgan B J T and Freeman S N (1998b). Estimation in parameter redundant models, Biometrika, 85, 462–468.

239 Catchpole E A, Morgan B J T, Freeman S N and Peach W J (1999). Modelling the survival of British lapwings Vanellus vanellus using ring-recovery data and weather covariates, Bird Study, 46 suppl., 5–13. Catchpole E A, Morgan B J T and Tavecchia G (2008). A new method for analysing discrete case-history data with missing values, Journal of Royal Statistical Society B, to appear. Chambers L E (1989). Population statistics of penguins, with particular reference to the Little Penguin, Honours thesis, La Trobe University, Australia. Chambers L E (2004a). Delayed breeding in Little Penguins — Evidence of climate change?, Australian Meteorological Magazine, 53, 13–19. Chambers L E (2004b). The impact of climate on Little Penguin breeding success, Technical Report BRMC No. 100, Bureau of Meteorology, Australia. Chao A (2001). An overview of closed capture-recapture models, Journal of Agri- cultural, Biological and Environmental Statistics, 6, 158–175. Chao A, Yip P S F, Lee S M and Chu W (2001). Population size estimation based on estimating functions for closed capture-recapture models, Journal of Statistical Planning and Inference, 92, 213–232. Chastel O, Weimerskirch H and Jouventin P (1993). High annual variability in reproductive success and survival of an Antarctic seabird, the snow petrel Pa- gadroma nivea, Oecologia, 94, 278–285. Cherel Y and Weimerskirch H (1995). Seabirds as indicators of marine resources: black-browed albatrosses feeding on ommastrephid squids in Kerguelen waters, Marine Ecology Progress series, 129, 295–300. Chiaradia A, Costalunga A and Knowles K (2003). The diet of Little Penguins (Eudyptula minor) at Phillip Island, Victoria, in the absence of a major prey — Pilchard (Sardinops sagax), Emu, 103, 43–48. Chiaradia A F and Kerry K R (1999). Daily nest attendance and breeding perfor- mance in the little penguin Eudyptula minor at Phillip Island, Australia, Marine Ornithology, 27, 13–20. Choquet R (2004). M–SURGE: new software specifically designed for multistate capture-recapture models, Animal Biodiversity and Conservation, 27, 207–215.

240 Clobert J and Lebreton J D (1985). D´ependance de facteurs de milieu dans les estimations de taux de survie par capture-recapture, Biometrics, 41, 1031–1037. Clobert J and Lebreton J D (1991). Estimation of demographic parameters in bird populations, in L J D Perrins, C M and G J M Hirons, eds, Bird Population Studies: Relevance to Conservation and Management, Oxford University Press, London, pp. 75–104. Clobert J, Lebreton J D and Allaine D (1987). A general approach to survival rate estimation by recaptures or resightings of marked birds, Ardea, 75, 133–142. Collins M, Cullen J M and Dann P (1999). Seasonal and annual foraging movements of Little Penguins from Phillip Island, Victoria, Wildlife Research, 26, 705–721. Conn P B, Kendall W L and Samuel M D (2004). A general model for the analysis of mark-resight, mark-recapture, and band-recovery data under tag loss, Biometrics, 60, 900–909. Cooch E G and White G W (2006). Program MARK: A gentle introduction. 5th edition. URL: http://www.phidot.org/software/mark/docs/book/. Cormack R M (1964). Estimates of survival from the sighting of marked animals, Biometrika, 51, 429–438. Cormack R M (1970). Statistical appendix to Fordham’s paper, Journal of Animal Ecology, 39, 24–27. Cowen L and Schwarz C J (2006). The Jolly–Seber model with tag loss, Biometrics, 62, 699–705. Crespin L, Harris M P, Lebreton J D, Frederiksen M and Wanless S (2006). Recruit- ment to a seabird population depends on environmental factors and on population size, Journal of Animal Ecology, 75, 228–238. Croxall J P and Rothery P (1995). Population change in Gentoo Penguins Pygoscelis papua at Bird Island, South Georgia: potential roles of adult survival, recruitment and deferred breeding, in P Dann, F I Norman and P N Reilly, eds, The Penguins: Ecology and Management, Surrey Beatty, Sydney, pp. 26–38. Culik B M, Wilson R P and Bannasch R (1993). Flipper-bands on penguins: what is the cost of a life-long commitment?, Marine Ecology Progress Series, 98, 209–214. Dann P (1988). An experimental manipulation of clutch size in the Little Penguin Eudyptula minor, Emu, 88, 101–103.

241 Dann P (1992). Distribution, population trends and factors influencing the popu- lation size of Little Penguins Eudyptula minor on Phillip Island, Victoria, Emu, 91(5), 263–272. Dann P, Carron M, Chambers B, Chambers L, Dornom T, McLaughlin A, Sharp B, Talmage M E, Thoday R and Unthank S (2005). Longevity in Little Penguins Eudyptula minor, Marine ornithology, 33, 71–72. Dann P and Cullen J M (1990). Survival, patterns of reproduction and lifetime reproductive output in the Little Blue Penguins (Eudyptula minor) on Phillip Island, Victoria, Australia, in L S Davis and J T Darby, eds, Penguin Biology, Academic Press, San Diego, pp. 63–84. Dann P, Cullen J M and Jessop R (1995). Cost of reproduction in Little Pen- guins, in P Dann, F I Norman and P N Reilly, eds, The Penguins: Ecology and Management, Surrey Beatty, Sydney, pp. 39–55. Dann P, Cullen J M, Thoday R and Jessop R (1992). Movements and patterns of mortality at sea of Little Penguins Eudyptula minor from Phillip Island, Victoria, Emu, 91(5), 278–286. Dann P and Norman F I (2006). Population regulation in Little Penguins, (Eu- dyptula minor): the role of intraspecific competition for nesting sites and food during breeding, Emu, 106, 289–296. Dann P, Norman F I, Cullen J M, Neira F J and Chiaradia A F (2000). Mortality and breeding failure of Little Penguins, Eudyptula minor, in Victoria, 1995–96, following a widespread mortality of pilchard, Sardinops sagax, Marine and Fresh- water Research, 51(4), 355–362. Department of the Environment and Heritage (1993). Resource Assessment Com- mission Coastal Zone Inquiry, Australian Government. URL: http://www.deh.gov.au/coasts/publications/rac/chap4-regulatory.html Department of the Environment and Heritage (2005). About bird and bat bands: band design, Australian Government. URL: http://www.environment.gov.au/biodiversity/science/abbbs/bands.html Dugger K M, Ballard G, Ainley D G and Barton K J (2006). Effects of flipper bands on foraging behaviour and survival of Adelie penguins (Pygoscelis adeliae), The Auk, 123(3), 858–869.

242 Eberhardt L L (1978). Appraising variability in population studies, Journal of Wildlife Management, 42, 207–238. Eberhardt L L (1985). Assessing the dynamics of wild populations, Journal of Wildlife Management, 49, 997–1012. Fletcher D and Efford M G (2007). Effect of senescence on estimation of survival rate when age is unknown, EURING 2007, in press. Fordham R A and Cormack R M (1970). Mortality and population change of Dominican Gulls in Wellington, New Zealand, Journal of Animal Ecology, 39, 13– 27. Fortescue M (1998), The marine and terrestial ecology of a northern population of the Little Penguin, Eudyptula minor, from Bowen Island, Jervis Bay, PhD thesis, University of Canberra, Australia. Unpublished. Francis C M (1990), Patterns of variation in survival rates of lesser snow geese, PhD thesis, Queen’s University, Canada. Unpublished. Francis C M (1995). Estimating survival rates from recoveries of birds ringed as young, Journal of Applied Statistics, 22, 567–577. Francis C M and Saurola P (2002). Estimating age-specific survival rates of tawny owls – recaptures versus recoveries, Journal of Applied Statistics, 29, 637–647. Freeman S N (1990), Statistical analysis of avian breeding and survival, PhD thesis, University of Kent, England. Unpublished. Freeman S N and Morgan B J T (1990). Studies in the analysis of ring-recovery data, The Ring, 13(1–2), 271–288. Freeman S N and Morgan B J T (1992). A modelling strategy for recovery data from birds ringed as nestlings, Biometrics, 48, 217–236. Freeman S N, Morgan B J T and Catchpole E A (1992). On the augmentation of ring-recovery data with field information, Journal of Animal Ecology, 61, 649– 657. Freeman S N and North P M (1990). Estimation of survival rates of British, Irish and French Grey Herons, The Ring, 13, 139–165. Froget G, Gauthier-Clerc M, Le Maho Y and Handrich Y (1998). Is penguin banding harmless?, Polar Biology, 20, 409–413.

243 Furness R W and Birkhead T R (1984). Seabird colony distributions suggest com- petition for food supplies during the breeding season, Nature, 311, 655–666. Gales R (1985). Breeding seasons and double brooding of the little penguin Eudyp- tula minor in New Zealand, Emu, 85, 127–130. Gauthier-Clerc M, Gendner J P, Ribic C A, Fraser W R, Woehler E J, Descamps S, Gilly C, Le Bohec C and Le Maho Y (2004). Long-term effects of flipper bands on penguins, Proceedings of the Royal Society of London B (Supplement), Biology Letters 271, 423–426. Harris M P, Buckland S T, Russell S M and Wanless S (1994). Year- and age-related variation in the survival of adult European shags over a 24-year period, Condor, 96, 600–605. Harris M P, Freeman S N, Wanless S, Morgan B J T and Wernham C V (1997). Factors influencing the survival of Puffins Fratercula arctica at a North Sea colony over a 20-year period, Journal of Avian Biology, 28, 287–295. Huggins R M (2006). Semiparametric Estimation of Animal Abundance Using Capture-Recapture Data from Open Populations, Biometrics, 62, 684–690. Huggins R M and Yip P S F (1999). Estimation of the Size of an Open Population from Capture-Recapture Data Using Weighted Martingale Methods, Biometrics, 55, 387–395. Jackson S and Wilson R P (2002). The potential costs of flipper-bands to penguins, Functional Ecology, 16, 141–148. Jarvis P J (1993). Environmental changes, in R W Furness and J J D Greenwood, eds, Birds as monitors of environmental change, Chapman and Hall, London, pp. 42–85. Johannesen E, Houston D and Russell J (2003). Increased survival and breeding performance of double breeders in Little Penguins Eudyptula minor, New Zealand: evidence for individual bird quality?, Journal of Avian Biology, 34, 198–210. Johannesen E, Perriman L and Steen H (2002a). The effect of breeding success on nest and colony fidelity in the Little Penguin (Eudyptula minor) in Otago, New Zealand, Emu, 102, 241–247.

244 Johannesen E, Steen H and Perriman L (2002b). Seasonal variation in survival, weights, and population counts of blue penguins (Eudyptula minor) in Otago, New Zealand, New Zealand Journal of Zoology, 29, 213–219. Johnson S R, Schiek J O and Searing G F (1995). Neck band loss rates for lesser snow geese, Journal of Wildlife Management, 59, 747–752. Jolly G M (1965). Explicit estimates from capture-recapture data with both death and immigration-stochastic models, Biometrika, 52, 225–247. Kanyamibwa S, Schierer A, Pradel R and Lebreton J D (1990). Changes in adult an- nual survival rates in a western European population of the White Stork Ciconia ciconia, Ibis, 132, 27–35. Kendall W L, Nichols J D and Hines J E (1997). Estimating temporary emigration using capture-recapture data with Pollock’s robust design, Ecology, 78, 563–578. Kendall W L, Pollock K H and Brownie C (1995). A likelihood approach to capture- recapture estimation of demographic parameters under the robust design, Bio- metrics, 51, 293–308. Klomp N I and Wooller R D (1991). Patterns of arrival and departure by breeding Little Penguins at Penguin Island, Western Australia, Emu, 91, 32–35. Kremers W K (1988). Estimation of survival rates from a mark-recapture study with tag loss, Biometrics, 44, 117–130. Lack D (1943). The age of the blackbird, British Birds, 36, 166–175. Lack D (1954). The Natural Regulation of Animal Numbers, Clarendon Press, Oxford. Lakhani K H (1987). Efficient estimation of age-specific survival rates from ring- recovery data of birds ringed as young, augmented by field information, Journal of Animal Ecology, 56, 969–987. Lakhani K H and Newton I (1983). Estimating age-specific bird surival rates from ring recoveries — can it be done?, Journal of Animal Ecology, 52, 83–91. Laplace P S (1786). Sur les naissances, les mariages et les morts, Histoire de l’Acad´emie Royale des Sciences, Ann´ee 1783, Paris, p.693. Le Maho Y, Gendner J P, Challet E, Bost C A, Gilles J, Verdon C, Plumere C, Robin J P and Handrich Y (1993). Undisturbed breeding penguins as indicators of changes in marine resources, Marine ecology progress series, 95, 1–6.

245 Lebreton J D (1995). The future of population dynamic studies using marked individuals: a statistician’s perspective, Journal of Applied Statistics, 22, 1009– 1030. Lebreton J D (2001). The use of bird rings in the study of survival, Ardea, 89(1), 85– 100. Lebreton J D, Burnham K P, Clobert J and Anderson D R (1992). Modeling survival and testing biological hypotheses using marked animals: a unified approach with case studies, Ecological Monographs, 62, 67–118. Lebreton J D, Morgan B J T, Pradel R and Freeman S N (1995). A simultane- ous survival rate analysis of dead recovery and live recapture data, Biometrics, 51, 1418–1428. Liang K Y and Zeger S L (1986). Longitudinal data analysis using generalized linear models, Biometrika, 73, 13–22. Lincoln F C (1930). Calculating waterfowl abundance on the basis of banding returns, U.S. Department of Agriculture, 118, 1–4. Little R J A and Rubin D B (1987). Statistical analysis with missing data, Wiley, New York. Lloyd C J and Yip P (1991). A unification of inference from capture-recapture stud- ies through martingale estimating equations, in V P Godambe, ed., Estimating Functions, Oxford University Press, Oxford, pp. 65–88. Magee M J (1928). How long do Purple Finches live?, Northeastern Bird-Banding Assoc., 4, 132–136. Magrath R D (1991). Nestling weight and juvenile survival in the blackbird, Turdus merula, Journal of Animal Ecology, 60, 335–351. Manly B F J, Amstrup S C and McDonald T L (2005a). Capture-recapture methods in practice, in S C Amstrup, T L McDonald and B F J Manly, eds, Handbook of capture-recapture analysis, Princeton University Press, Princeton, pp. 266–274. Manly B F J, McDonald T L and Amstrup S C (2005b). Introduction to the Handbook, in S C Amstrup, T L McDonald and B F J Manly, eds, Handbook of capture-recapture analysis, Princeton University Press, Princeton, pp. 1–21. Mardekian S Z and McDonald L (1981). Simultaneous analysis of band-recovery and live-recapture data, Journal of Wildlife Management, 45, 484–488.

246 McCullagh P and Nelder J A (1983). Generalized Linear Models, Chapman and Hall, London. McDonald T L, Amstrup S C and Manly B F J (2003). Tag loss can bias Jolly–Seber capture-recapture estimates, Wildlife Society Bulletin, 31, 814–822. Mickelson M J, Dann P and Cullen J M (1992). Sea temperature in Bass Strait and breeding success of the Little Penguins Eudyptula minor at Phillip Island, south-eastern Australia, Emu, 91(5), 355–368. Molenberghs G and Verbeke G (2005). Models for Discrete Longitudinal Data, Springer, New York. Montevecchi W A (1993). Birds as indicators of change in marine prey stocks, in R W Furness and J J D Greenwood, eds, Birds as monitors of environmental change, Chapman and Hall, London, pp. 217–266. Morgan B J T and Freeman S N (1989). A model with first-year variation for ring-recovery data, Biometrics, 45, 1087–1101. Mougin J L and van Beveren M (1979). Structure et dynamique de la population de Manchots Empereur Aptenodytes forsteri de la colonie de l’archipel de Pointe Geologie, Terre Adelie, Compte Rendus Academie Science de Paris, pp. 157–60. Moyes K, Coulson T, Morgan B J T, Donald A, Morris S J and Clutton-Brock T H (2006). Cumulative reproduction and survival costs in female red deer, OIKOS, 115, 241–252. Newman G (1992). Studies on the Little Penguin Eudyptula minor in Victoria: an introduction, Emu, 91(5), 261–262. Nichols J D and Hines J E (1993). Survival rate estimation in the presence of tag loss using joing analysis of capture-recapture and resighting data, in JD Lebreton and P M North, eds, Marked individuals in the study of bird population, Birkh¨auser Verlag, Basel, pp. 229–243. Nichols J D, Hines J E and Blums P (1997). Tests for senescent decline in annual survival probabilities of common pochards, Aythya ferina, Ecology, 78, 1009– 1018. Nichols J D, Sauer J R, Pollock K H and Hestbeck J B (1992). Estimating transi- tion probabilities for stage based population projection matrices using capture- recapture data, Ecology, 73, 306–312.

247 Nisbet I C T (2001). Detecting and measuring senescence in wild birds: experience with long-lived seabirds, Experimental Gerontology, 36(4–6), 833–843. North P M and Morgan B J T (1979). Modelling heron survival using weather data, Biometrics, 35, 667–681. North P M and Morgan B J T (1985). The general context of statistics in ornithol- ogy, in B J T Morgan and P M North, eds, Statistics in ornithology, Springer- Verlag, New York, pp. 1–9. Olsson O (1997). Effects of food availability on fledgling condition and post-fledgling survival in king penguin chicks, Polar Biology, 18, 161–165. Penguin Reserve Committee of Management (1992). Penguins on Parade, Scancolor Australia, Moorabbin, Victoria. Perriman L, Houston D, Steen H and Johannesen E (2000). Climate fluctuation effects on breeding of blue penguins (Eudyptula minor), New Zealand Journal of Zoology, 27, 261–267. Petersen S L, Branch G M, Ainley D G, Boersma P D, Cooper J and Woehler E J (2006). Is flipper banding of penguins a problem?, Marine Ornithology, 33, 75–79. Pollock K H (1981). Capture-recapture models allowing for age-dependent survival and capture rates, Biometrics, 37, 521–529. Pollock K H (1982). A capture-recapture design robust to unequal probabilities of capture, Journal of Wildlife Management, 46, 757–760. Pollock K H (1991). Modeling capture, recapture, and removal statistics for estima- tion of demographic parameters for fish and wildlife populations: past, present and future, Journal of the American Statistical Association, 86, 225–238. Pollock K H (2002). The use of auxiliary variables in capture-recapture modeling: an overview, Journal of Applied Statistics, 29, 85–102. Pollock K H, Hines J E and Nichols J D (1984). The use of auxiliary variables in capture-recapture and removal experiments, Biometrics, 40, 329–340. Pollock K H, Hines J E and Nichols J D (1985). Goodness-of-fit tests for open capture-recapture models, Biometrics, 41, 399–410. Pollock K H, Nichols J D, Brownie C and Hines J E (1990). Statistical Inference for capture-recapture experiments, Wildlife Monographs, 107, 1–97.

248 Pradel R (1996). Utilization of capture-mark-recapture for the study of recruitment and population growth rate, Biometrics, 52, 703–709. Pradel R and Lebreton J D (1991). User’s manual for program SURGE. Ver- sion 4.1., Centre d’Ecologie Fonctionnelle et Evolutive, CNRS, BP 5051, 34033 Montpellier-Cedex, France. Pr´evot-Julliard A C, Lebreton J D and Pradel R (1998). Re-evaluation of adult survival of black-headed gulls (Larus ridibundus) in the presence of recapture heterogeneity, The Auk, 115, 85–95. Pugesek B H, Nations C, Diem K L and Pradel R (1995). Mark-resighting analysis of a California gull population, Journal of Applied Statistics, 22, 625–639. R Development Core Team (2006). R: A Language and Environment for Statistical Computing, R Foundation for Statistical Computing, Vienna, Austria. URL: http://www.R-project.org. Reilly P N (1983). Fairy Penguins and earthy people, Lothian, Melbourne. Reilly P N and Cullen J M (1979). The Little Penguin Eudyptula minor in Victoria. I: Mortality of adults, Emu, 79, 97–102. Reilly P N and Cullen J M (1981). The Little Penguin Eudyptula minor in Victoria. II: Breeding, Emu, 81, 1–19. Reilly P N and Cullen J M (1982). The Little Penguin Eudyptula minor in Victoria. III: Dispersal of chicks and survival after banding, Emu, 82, 137–142. Reilly P N and Cullen J M (1983). The Little Penguin Eudyptula minor in Victoria. IV: Moult, Emu, 83, 94–98. Rexstad E A and Burnham K (1992). User’s guide for interactive program CAP- TURE, Colorado Cooperative Fish and Wildlife Research Unit, Colorado State University, Fort Collins. Reynolds R W and Smith T M (1994). Improved global sea surface temperature analysis using optimum interpolation, Journal of Climate, 7, 929–948. Richdale L (1957). A Population Study of Penguins, Oxford University Press, Ox- ford. Ringsby T H, Saether B E and Solberg E J (1998). Factors affecting juvenile survival in House sparrow Passer domesticus, Journal of Avian biology, 29, 241–247.

249 Saether B E and Bakke O (2000). Avian life history variation and contribution of demographic traits to the population growth rates, Ecology, 81, 642–653. Sagar P M and Horning Jr D S (1998). Mass-related survival of fledgling Sooty shearwaters Puffinus griseus at the Snares, New Zealand, Ibis, 140, 329–331. Sandvik H, Erikstad K E, Barrett R T and Yoccoz N G (2005). The effect of climate on adult survival in five species of North Atlantic seabirds, Journal of Animal Ecology, 74, 817–831. Schnabel Z E (1938). The estimation of the total fish population of a lake, American Mathematical Monthly, 45, 348–352. Schwarz C J, Schweigert J F and Arnason A N (1993). Estimating migration rates using tag-recovery data, Biometrics, 49, 177–193. Schwarz C J and Seber G A F (1999). Estimating animal abundance: review III, Statistical Science, 14, 427–456. Seber G A F (1962). The multi-sample single recapture census, Biometrika, 49, 339– 350. Seber G A F (1965). A note on the multiple recapture census, Biometrika, 52, 249– 259. Seber G A F (1970). Estimating time-specific survival and reporting rates for adult birds from band returns, Biometrika, 57, 313–318. Seber G A F (1971). Estimating age-specific survival rates from bird-band returns when the reporting rate is constant, Biometrika, 58, 491–497. Seber G A F (1982). The Estimation of Animal Abundance and Related Parameters, 2nd edn, Griffin, London. Seber G A F (1986). A review of estimating animal abundance, Biometrics, 42, 267– 292. Seber G A F and Schwarz C J (2001). Capture-recapture: before and after EURING 2000, Journal of Applied Statistics, 29, 5–18. Sidhu L A, Catchpole E A and Dann P (2007). Mark-recapture-recovery modeling and age-related survival in Little Penguins Eudyptula minor, The Auk, 124, 815– 827. Sladen W J L (1952). Notes on methods of marking penguins, Ibis, 16, 541–543.

250 Sladen W J L and LeResche R E (1970). New and developing techniques in Antarctic ornithology, in M W Holdgate, ed., Antarctic Ecology, Academic Press, London, pp. 585–596. Sladen W J L, Wood R C and Monaghan E P (1968). The USARP bird band- ing program, 1958–1965, in O L Austin, ed., Antarctic bird studies: Antarctic Research Series 12, American Geophysical Union, Washington, pp. 213–262. Stahel C and Gales R (1987). Little Penguin, NSW University Press, Kensington. Stearns S C (1992). The evolution of life histories, Oxford University Press, Oxford. Stenseth N C, Mysterud A, Ottersen G, Hurrell J W, Chan K S and Lima M (2002). Ecological effects of climate fluctuations, Science, 297, 1292–1296. Stobo W T and Horne J K (1994). Tag loss in grey seals (Halichoerus grypus) and potential effects on population estimates, Canadian Journal of Zoology, 72, 555– 561. Stonehouse B (1999). Penguin banding: time for reappraisal?, Marine Ornithology, 27, 115–118. Szymczak M R and Rexstad E A (1991). Harvest distribution and survival of a gadwall population, Journal Wildlife Management, 55, 592–600. Tavecchia G, Coulson T, Morgan B J T, Pemberton J M, Pilkington J C, Gulland F M D and Clutton-Brock T H (2005). Predictors of reproductive cost in female Soay sheep, Journal of Animal Ecology, 74, 201–213. Treble R J, Day R W and Quinn II T J (1993). Detection and effects on mortal- ity estimates of changes in tag loss, Canadian Journal of Fisheries and Aquatic Sciences, 50, 1435–1441. Underhill L G, Bartlett P A, Baumann L, Crawford R J M, Dyer B M, Gildenhuys A, Nel D C, Oatley T B, Thornton M, Upfold L, Williams A J, Whittington P A and Woolfaardt A C (1999). Mortality and survival of African Penguins Spheniscus demersus involved in the Apollo Sea oil spill: an evaluation of rehabilitation efforts, Ibis, 141, 29–37. Valle C A and Coulter M C (1987). Present status of the flightless cormorant, Gal´apagos penguin and greater flamingo populations in the Gal´apagos islands, Ecuador, after the 1982–83 El Ni˜no, The Condor, 89, 276–281.

251 Van der Jeugd H P and Larsson K (1998). Pre-breeding survival of barnacle geese Branta leucopsis in relation to fledgling characteristics, Journal of Animal Ecol- ogy, 67, 953–966. Vounatsou P and Smith A F M (1995). Bayesian analysis of ring-recovery data via Markov chain Monte Carlo simulation, Biometrics, 51, 687–708. Wald A (1943). Tests of statistical hypotheses concerning several parameters when the number of observations is large, Transactions of the American Mathematical Society, 54, 426–482. Weavers B W (1992). Seasonal foraging ranges and travels at sea of Little Penguins Eudyptula minor, determined by radiotracking, Emu, 91(5), 302–317. Weimerskirch H, Stahl J C and Jouventin P (1992). The breeding biology and population dynamics of King Penguins Aptenodytes patagonicus on the Crozet Islands, Ibis, 134, 107–117. Wetherall J A (1982). Analysis of double-tagging experiments, Fishery Bulletin, 80, 687–701. White G C and Burnham K P (1999). Program MARK: survival estimation from populations of marked animals, Bird Study, 46 suppl., 120–39. White G C and Burnham K P (2002). Discussion comments on: the use of aux- iliary variables in capture-recapture modelling: an overview, Journal of Applied Statistics, 29, 103–106. Wilkinson G N and Rogers C E (1973). Symbolic description of factorial models for analysis of variance, Journal of Applied Statistics, 22, 392–399. Williams T D (1995). The Penguins, Oxford University Press, Oxford. Xiao Y (1996). A general model for estimating tag-specific shedding rates and tag interactions from exact or pooled times at liberty for a double tagging experiment, Canadian Journal of Fisheries and Aquatic Sciences, 53, 1852–1861. Yip P (1989). An inference procedure for a capture and recapture experiment with time-dependent capture probabilities, Biometrics, 45, 471–479. Yip P (1991). A martingale estimating equation for a capture-recapture experiment in discrete time, Biometrics, 47, 1081–1088.

252