PLEASE NOTE THIS IS A PREPRINT OF THE FINAL VERSION PUBLISH IN JOURNAL OF ECOLOGICAL MODELLING https://www.sciencedirect.com/science/article/pii/S0304380018301297 In particular, the published version contain a more thorough discussion of the SIR method.
1
Characterizing the strength of density dependence in at-risk species through Bayesian model
2
averaging
3
Guillaume Bal1,2, Mark D. Scheuerell3 and Eric J. Ward1
4 5 6
1
7
Service, National Oceanic and Atmospheric Association, Seattle, WA 98112, USA 2
8 9 10
Conservation Biology Division, Northwest Fisheries Science Center, National Marine Fisheries
3
Marine Institute, Rinville, Oranmore, Co. Galway, Ireland
Fish Ecology Division, Northwest Fisheries Science Center, National Marine Fisheries Service, National Oceanic and Atmospheric Association, Seattle, WA 98112, USA
11 12 13
Corresponding author: G. Bal, Marine Institute, Rinville, Oranmore, Co. Galway, Ireland. (E-
14
mail:
[email protected], Phone: +353 83 191 9748, Fax: +353 91 387 201)
15 16
17
1
18
Developing effective conservation plans for at-risk species requires an understanding of the
19
relationship between numbers of breeding adults and their subsequent offspring. In particular,
20
establishing the degree to which density-dependence affects population size can be difficult due
21
to errors in the data themselves, uncertainty in model parameters, and possible misspecification
22
of model structure. Here we use a Bayesian framework to obtain posterior probability
23
distributions for model parameters and each of four simple adult-offspring production models
24
that allow for varying types of density dependence. Specifically, we analyzed 42 at-risk
25
populations of anadromous Chinook salmon (Oncorhynchus tshawytscha) from the northwestern
26
United States and found strong evidence that more than two-thirds of the populations exhibit
27
negative density-dependent effects of adults (i.e., compensatory mortality). This was somewhat
28
remarkable given large reductions in adult numbers relative to historical benchmarks, indicating
29
that carrying capacity of spawning habitat has been reduced considerably. Furthermore, we also
30
found non-zero probabilities of positive density-dependent effects of adults (i.e., Allee effects) in
31
about one-third of the populations, which could suggest that cumulative losses of spawning
32
adults over the past century has led to decreased nutrient and energy subsidies from semelparous
33
carcasses, and diminished bio-physical disturbance from nest-digging activity. Importantly, our
34
analysis highlights the utility of Bayesian model averaging in a conservation context wherein
35
errors in choosing the best model may have more severe consequences than errors in estimating
36
model parameters themselves.
37
Keywords: density dependence, Bayesian model averaging, Allee effect, compensation,
38
depensation, salmon.
39
ABSTRACT
40 41
2
INTRODUCTION One of the cornerstones of conservation biology is establishing the relationship between
42
the numbers of parents and the offspring they produce. In particular, the degree to which
43
organisms are affected by population density has important implications for individual fitness
44
and population growth. Negative density dependence (NDD), or compensatory mortality, occurs
45
when density is relatively high and any further increases in density lead to increased transmission
46
of diseases or competition for resources (e.g., food, breeding locations), ultimately causing
47
reductions in per capita growth rates (Hixon et al. 2002, Brook and Bradshaw 2006). Conversely,
48
positive density dependence (PDD), or depensatory mortality, arises when density is relatively
49
low and the addition of more individuals causes increased per capita growth rates as a result of
50
cooperative foraging or defensive behaviors, increased probability of finding a mate, or
51
combinations of these factors (i.e. Allee effects, Courchamp et al. 1999, Berec et al. 2007,
52
Kramer et al. 2009). Understanding whether these phenomena occur and if so, to what extent, is
53
necessary for determining the best options for conserving at-risk species. For example, the
54
presence of NDD could indicate limited habitat availability (i.e., insufficient total area) whereas
55
the existence of PDD might suggest a high degree of habitat fragmentation; rectifying those two
56
types of habitat deficiencies could require rather different approaches.
57
Density-dependence has been studied extensively in fish populations because of its
58
importance to both the management of healthy stocks and conservation of imperiled populations
59
(Liermann and Hilborn 1997, Barrowman and Myers 2000, Barrowman et al. 2003). In classical
60
fisheries management, the presence of compensatory mortality within a stock implies parental
61
biomass should be harvested to the point where the surplus production of new recruits to the
62
fishery is maximized relative to replacement (Hilborn and Walters 1992). Conversely, the degree
63
of depensatory mortality will determine the rate at which overfished stocks will recover when
64
harvest is reduced. Most conservation practitioners concentrate on the possible existence of
65
positive density dependence. However, negative density dependence at relatively low densities
66
can exist, implying diminished carrying capacity from factors like habitat loss/modification or
67
the presence of non-native species (Achord et al. 2003), although this is often ignored in
68
conservation contexts.
69
Population dynamics models offer a formal means for estimating both positive and
70
negative density dependence (Boyce 1992). For example, Beverton (1957) and Ricker (1954)
71
models have been used to estimate the relationships between parents and offspring for decades.
72
More recently, Barrowman (2000) introduced a form of piecewise regression model (known as
73
the “hockey stick” model) similar to the Ricker and Beverton-Holt curves. The hockey stick
74
model offers potential advantages over other models in a conservation context because it
75
provides more conservative estimates of the maximum density-independent survival (i.e., slope
76
at the origin) and carrying capacity (Barrowman and Myers 2000). In addition, the breakpoints in
77
the hockey stick segments may provide natural reference points for management decisions.
78
However, the hockey stick model does not allow for PDD (depensation).
79
Regardless of model choice, three main types of uncertainties can hinder our ability to
80
infer the underlying relationship between parents and their offspring. First, observation errors
81
arise in the form of sampling and measurement errors. Second, model parameters are rarely
82
known without error and instead must be estimated from the data. Third, uncertainty about the
83
structure of the model itself affects inference about the form of the parent-offspring relationship.
84
The first two concerns are often addressed through appropriate sampling designs and explicit
85
consideration of both process and observation/sampling errors. However, possible
86
misspecification of a particular model is typically ignored and instead the “best” model is chosen
87
based on some model selection measure such as Akaike’s Information Criterion (Burnham and
88
Anderson 2002). In such cases, two models with nearly identical support from the data could
89
produce widely divergent predictions, especially when confronted with new data (Pascual et al.
90
1997). As a guard against this very real possibility, model averaging (MA) offers a formal means
91
for explicitly addressing model-selection uncertainty in problems of inference and prediction
92
(Burnham and Anderson 2002, Wintle et al. 2003).
93
Pacific salmon (Oncorhynchus spp.) are important to human economies and the ecology
94
of coastal ecosystems across the northern Pacific Rim (Schindler et al. 2003). Fisheries for
95
salmon are worth hundreds of millions of US dollars annually (NMFS 2013). Salmon also act as
96
ecosystem engineers by modifying benthic habitats through nest-digging activity (Moore 2006)
97
and serve as keystone species in food webs as food for fish, mammals, birds, and insects
98
(Helfield and Naiman 2006). In many coastal watersheds across western Canada and the US,
99
however, Pacific salmon populations have been reduced to mere fractions of their historical
100
abundances due to changes in habitat, hydropower development, overharvest, and climate
101
(Ruckelshaus et al. 2006). Thus, obtaining a proper understanding of whether NDD, PDD, or
102
both occur among at-risk salmon populations is necessary for determining the best recovery
103
options.
104
Here we examined the strength of NDD and PDD among 42 populations of Chinook
105
salmon (O. tshawytscha) from the northwestern USA that are currently listed as “threatened”
106
under the US Endangered Species Act. To do so, we used four different piecewise-linear models
107
to characterize various forms of density dependence (Fig. 1): 1) unrestricted (density-
108
independent), 2) Allee effect (PDD) only, 3) carrying capacity (NDD) only, and 4) both PDD
109
and NDD. Specifically, we added an additional breakpoint to Barrowman and Myers’ (2000)
110
hockey stock model, which can be very useful in conservation contexts because 1) it tends to
111
provide better estimates of per capita productivity at low population sizes; and 2) it offers more
112
conservative estimates of mortality rates that lead to extinction. In addition, we addressed both
113
parameter and model uncertainty within a unified framework through Bayesian model averaging
114
(BMA), which allowed us to easily combine predictions and uncertainties across all four models.
115
3
116
3.1
117
MATERIAL AND METHODS Allee Hockey Stick model Barrowman and Myers’ (2000) hockey stick (HS) model assumes the number of
118
offspring increases linearly up to an asymptote, beyond which it is independent of parental
119
abundance. For a given population, the model is described by two linear functions,
𝑂𝑡 = {
𝛽𝑃𝑡 𝑖𝑓 𝑃𝑡 < 𝑃𝑐 } 𝛽𝑃𝑐 𝑖𝑓 𝑃𝑡 > 𝑃𝑐
(1)
120
where is the density-independent per-capita reproductive rate, 𝑃𝑡 is the number of parents
121
breeding in year t, and 𝑃𝑐 is the threshold parental abundance above which compensatory
122
density-dependent effects dominate and restrain the number of offspring. As 𝑃𝑐 approaches
123
infinity, the degree of density-dependence goes to zero and there is no limit to population size.
124 125
To allow for Allee effects at relatively low adult abundance, we modified the original HS model (Barrowman and Myers 2000) to allow for a third linear segment. This new “Allee
126
Hockey Stick” (AHS) model allows growth rates to become depressed when adult abundance is
127
less than some threshold, 𝑃𝑑 𝛽1 𝑃𝑡 𝑂𝑡 = {𝛽1 𝑃𝑑 +𝛽2 (𝑃𝑡 − 𝑃𝑑 ) 𝛽1 𝑃𝑑 +𝛽2 (𝑃𝑐 − 𝑃𝑑 )
𝑖𝑓 𝑃𝑡 < 𝑃𝑑 𝑖𝑓 𝑃𝑑 < 𝑃𝑡 < 𝑃𝑐 𝑖𝑓 𝑃𝑡 ≥ 𝑃𝑐
(2)
128
where 0 < 𝛽1 < 1 < 𝛽2, and 𝑃𝑑 and 𝑃𝑐 are the breakpoint thresholds respectively. By fixing
129
parameters at zero or infinity, the AHS model is capable of producing a range of dynamics,
130
including three less complex models nested within Equation 2 (Fig. 1): unrestricted growth (i.e.,
131
𝛽1 = 0, 𝑃𝑑 and 𝑃𝑐 = ∞), original HS with carrying capacity only (i.e., 𝛽1 = 0, 𝑃𝑑 = 0), and
132
Allee effects only (i.e., 𝑃𝑐 = ∞).
133
3.2
134
Parameter estimation and model weighting Though Bayesian estimation could be done in a 2 step procedure, we illustrate here how
135
to conduct both estimation and model selection in a single framework, using the Sampling –
136
Importance – Resampling (SIR) algorithm (Rubin 1988). This approach aims to approximate the
137
posterior distribution by using an importance function to resample parameter draws from a
138
proposal distribution; we adopted the same version of the algorithm as described and
139
implemented by Moore (2008). We assigned equal probabilities a priori to each of the 4 models,
140
and calculated the posterior marginal likelihood as the mean likelihood across all parameter
141
vectors from the SIR algorithm (Carlin and Louis 2000, Gelman et al. 2004). The SIR algorithm
142
can be run independently for each of the proposed models, and the marginal likelihoods can then
143
be normalized so that ∑𝑖 𝑃(𝑚𝑖 |𝑥) = 1. These normalized marginal likelihoods are interpreted as
144
conditional probabilities of models on observed data. Compared to other Bayesian estimation
145
approaches, such as MCMC, the SIR algorithm offers several advantages. The main is that
146
estimating the mean marginal likelihood mean is more straightforward than other MCMC
147
approaches, such as WinBUGS (Lunn et al. 2000) or JAGS software (Plummer 2003). The SIR
148
algorithm also does not require a burn-in period or assessment of MCMC diagnostics for
149
convergence. All priors used in this study were weakly informative uniform priors and shared
150
among models whenever possible (Table I). Lower and upper boundaries were based on
151
Barrowman’s (2000) work on the congener O. kisutch.
152
3.3
153
Model Application For management purposes, Pacific salmon species are grouped into evolutionarily
154
significant units (ESUs), defined as a group of salmon that (1) is reproductively isolated from
155
other conspecific populations, and (2) represents an important component in the evolutionary
156
legacy of the species (Waples 1991). We estimated the strength of density dependence among 42
157
Chinook salmon populations within 3 distinct ESUs from Washington, Idaho, and Oregon in the
158
northwestern USA (Fig. 2). Chinook salmon within the Puget Sound ESU are “ocean-type”,
159
which means juveniles spend less than 1 year in fresh water before migrating to the ocean, but
160
salmon from the Snake River and Upper Columbia ESUs are “stream-type” and spend 1 full year
161
in fresh water before migrating to sea (Taylor 1990). This offers us an opportunity to compare
162
whether the type or strength of density dependence varies between the contrasting life histories.
163
We used the total number of spawning adults in a given year as our estimate of parent
164
abundance. Because salmon from the same cohort mature at different ages, our estimates of
165
offspring are then the sum of subsequent adults that were born in a given year, but that vary in
166
age and return over sequential years (e.g., the offspring of Snake River stream-type adults that
167
spawned in 2000 return as 3-6 years later in 2003-2006). All data used here were compiled by
168
the National Marine Fisheries Service and are available upon request. To ensure time series are
169
long enough to potentially detect positive density dependence (Brook and Bradshaw 2006,
170
Gregory et al. 2010), we restricted our analysis to only include populations with at least 15 years
171
of data. This cutoff yielded 42 time series (Table 2).
172
4
173
RESULTS For both life history types, we found the most support from the data for a model with
174
only NDD (i.e., the median posterior model probabilities were 0.72 and 0.56 for ocean- and
175
stream-type Chinook salmon, respectively; Fig. 3). We found very little evidence for models
176
containing Allee effects (PDD) only, but models that included both PDD and NDD had median
177
probabilities of being the best across ESUs of 0.10 and 0.075 for ocean- and stream-type salmon,
178
respectively.
179
Among populations within a specific life history type, however, the relative support for
180
the different models varied considerably (Fig. 4). Although there were a few instances where
181
model probabilities were evenly split among the four model forms (e.g., 4, 15), most populations
182
showed strong support in favor of one particular model. For ocean-type Chinook salmon, most
183
populations showed strong indications of NDD, but a few populations had very little evidence
184
for either form of density dependence (e.g., 7, 18, 20); we found similar results for stream-type
185
Chinook. Notably, the evidence for models with both PDD and NDD was reasonably strong in a
186
few populations (e.g., 25, 26).
187
For each population, the model averaged relationship differed slightly from the best
188
model. For example, the best models selected for the Duwamish, Skykomish, and Imnaha rivers
189
(i.e., populations 4, 15, 25 in Table 2) were the AHS model, linear model, and HS model,
190
respectively. Applying BMA to estimate the shape of this curve may yield a result similar to the
191
single “best” model, but can also result in more conservative results (Fig. 5). In the case of the
192
Duwamish River, a simple model selection would select a linear relationship whereas the shape
193
coming from BMA exhibits Allee effects when the population size falls under 865 fish. In the
194
Imnaha River, the estimated carrying capacity from the best model is about 50% higher than the
195
model-averaged estimate. For the Skokomish River, however, there is very little difference
196
between the best and MA models.
197
5
198
DISCUSSION Most management decisions for harvested, threatened and/or endangered populations are
199
based on estimates from a single population model, but the estimated relationship between
200
population size and per capita growth is driven by the choice of models used in estimation
201
(Wintle et al. 2003). Although prediction intervals can be constructed based on parameter
202
uncertainty, ignoring model uncertainty creates overconfidence in the model prediction and may
203
obfuscate decision-making (Hoeting et al. 1999, Wintle et al. 2003, Ellison 2004). In examining
204
support for processes such as positive or negative density dependence, most previous approaches
205
have simply identified whether or not data provide evidence for density dependence being
206
present. Model averaging improves on this by allowing a range of forms of density dependence
207
to be evaluated simultaneously.
208
Our results reveal two important messages regarding the role of density-dependence in
209
the population dynamics of Chinook salmon in the Pacific Northwest. First, although we found
210
that linear relationships between offspring and parents with no density dependence were
211
probable, especially within the Snake River ESU (median model posterior probability = 0.27),
212
we found much stronger support for models with NDD (median model posterior probability =
213
0.56) even though parental abundances across all populations have been greatly reduced from
214
historical numbers and hence we expected less evidence for NDD. This suggests that future
215
restoration efforts ought to focus on increasing overall carrying capacity. Unfortunately we lack
216
much of the age-specific data for many of these populations to assess where in the life-cycle
217
capacity is most limiting.
218
Second, the existence of PDD among these Chinook salmon populations is a possibility,
219
which agrees with previous studies of other salmon species (Liermann and Hilborn 1997,
220
Barrowman et al. 2003). Although we cannot identify the exact cause of PDD, some reasonable
221
hypotheses exist. First, salmon gain about 90% of their adult biomass in the ocean, and therefore
222
act as net importers of marine-derived nutrients and energy to freshwater ecosystems (Schindler
223
et al. 2003). Low adult abundance could therefore reduce survival during earlier juvenile life
224
stages. Second, nest-digging activity by female salmon mobilizes fine sediment (Moore et al.
225
2004) and decreases the probability of stream-bed scour and excavation of buried salmon eggs or
226
embryos (Montgomery et al. 1996). Thus, losses of adult salmon may reach a point whereby the
227
ecosystem cannot support the juvenile production it once did (Achord et al. 2003, Scheuerell et
228
al. 2005).
229
For salmon in particular, absence of observed Allee effects could be linked to high
230
straying rates and colonization capacities (Makhrov et al. 2005, Correa and Gross 2008, Launey
231
et al. 2010, Westley et al. 2013). Alternatively, it could indicate that availability of juvenile
232
habitat is more important than direct mortality from predators (Mogensen and Post 2012). For
233
Puget Sound populations, in particular, much of the lowland areas have been converted from
234
forest to agriculture, which has greatly decreased the carrying capacity for juvenile Chinook
235
salmon (Scheuerell et al. 2006). Furthermore, increasing evidence points to NDD among adult
236
life stages during their ocean residency owing to competition with hatchery fish, (Connors et al.
237
2012, Ruggerone et al. 2012), which shifts the focus away from habitat actions and more toward
238
hatchery reforms (Buhle et al. 2009).
239
The varying degrees of support for different forms of density dependence among the
240
many populations provide some important insights for conservation. Decisions to list Pacific
241
salmon as threatened or endangered under the Endangered Species Act are made collectively at
242
the relatively large ESU level, but recovery plans are developed and implemented at much
243
smaller scales relating to specific populations. Here we have used posterior model probabilities
244
to assess the degree to which PDD, NDD, or a combination thereof exists for each population
245
within an ESU. Managers could potentially use this information to further direct location-
246
specific actions that might decrease the intensity of PDD or NDD.
247
Our analysis highlights the utility of model averaging for estimating relationships among
248
parents and their offspring. Model averaged results were more conservative than the best model
249
alone, and therefore applying model averaging may be particularly important for populations of
250
conservation concern. The use of the HS and AHS models was largely chosen based on the taxa
251
in our analysis; other model sets may be developed for other species (and unlike our analysis,
252
there is no requirement that all candidate models be nested). While there are a number of ways to
253
implement Bayesian model averaging, the SIR approach is best for relatively simple models like
254
those we used here. Increased model complexity will decrease the speed of the SIR approach
255
exponentially, however. Using other estimation routines such as Markov Chain Monte Carlo
256
(MCMC), the model-averaged procedure could also be constructed hierarchically across
257
populations, to estimate common strengths of density dependence across populations.
258
6
259
This research was performed while Dr. Guillaume Bal was a post-doctoral research associate
260
sponsored by the National Research Council. We thank Brian Burke and Kevin See for helpful
261
comments on an earlier draft of the manuscript.
262
ACKNOWLEDGMENTS
263
7
TABLES
264
Table I. Uniform distributions used as Bayesian priors for the parameters in the four competing
265
models described by equation 2 in the methods. Pmax and Omax are the maximum numbers of
266
observed parents and offspring, respectively, for a given population.
Parameter
Linear
β1 β2
Allee Unif(0, 1)
Unif(1, 25)
Pd
267
Allee Hockey Stick Unif(0, 1)
Unif(1, 25)
Unif(1, 100)
Unif(1, 100)
Unif(0, Pmax / 3)
Unif(0, Pmax / 3)
Unif(0, Pmax / 3)
Pc CV
Hockey Stick
Unif(Omax / 5, Omax) Unif(Omax / 5, Omax) Unif(0, 3)
Unif(0, 3)
Unif(0, 3)
Unif(0, 3)
Table 2. Summary of parent and offspring data for Chinook salmon used in the study. PS, S, and UC refer, respectively, to Puget Sound, Snake, and Upper Columbia Evolutionary Significant Units (ESU).
Index
ESU
River name
Time period
Parent range
Offspring range
1
PS
Cascade R.
1981-2005
83-625
92-2458
2
PS
Cedar Creek
1965-2005
126-1896
8-9099
3
PS
Dungeness R.
1986-2005
43-955
47-1486
4
PS
Duwamish R.
1968-2005
2014-11551
657-52907
5
PS
Elwha R.
1986-2005
164-5309
418-5485
6
PS
Lower Sauk R.
1952-2005
112-3896
321-22021
7
PS
Lower Skagit R.
1952-2005
409-9263
577-62260
8
PS
North Fork Nooksack R.
1984-2005
10-7473
13-5337
9
PS
North Fork Stillaguamish R.
1974-2005
330-1849
353-89489
10
PS
Puyallup R.
1968-2005
527-5387
74-23354
11
PS
Sammamish R.
1983-2005
34-550
20-1708
12
PS
South Fork Nooksack R.
1984-2005
103-625
100-1364
13
PS
South Fork Stillaguamish R.
1974-2005
73-391
15-14265
14
PS
Skokomish R.
1968-2005
189-3184
39-16734
15
PS
Skykomish R.
1965-2005
1681-7703
110-110621
16
PS
Snoqualmie R.
1965-2005
324-3603
149-24714
17
PS
Suiattle R.
1952-2005
167-1804
87-10407
18
PS
Upper Sauk R.
1952-2005
108-3345
57-27343
19
PS
Upper Skagit R.
1952-2005
3586-20040
6146-114870
20
PS
White R.
1965-2005
7-2131
9-7779
21
S
Big Creek
1957-2002
5-1858
6-3349
22
S
Camas Creek
1964-1998
3-554
1-1611
23
S
Catherine Creek
1955-2002
28-3161
4-6515
24
S
Grand Ronde R. (Upper Mainstream)
1956-2002
3-1028
2-1991
25
S
Imnaha R.
1954-2002
258-6267
197-5361
26
S
Lemhi River
1957-2002
10-3357
13-7633
27
S
Lostine River
1959-2002
37-1585
35-3372
28
S
Marsh Creek
1957-2002
13-1845
4-4461
29
S
Minam River
1954-2002
54-4104
26-4068
30
S
Pahsimeroi River
1989-2002
27-306
64-1187
31
S
Secesh River
1957-2002
48-1395
21-1581
32
S
Snake River (Lower Mainstem)
1957-2002
11-4888
24-2444
33
S
Snake River (Upper Mainstem)
1962-2002
18-3554
41-5267
34
S
South Fork Salmon River (East Fork)
1958-2002
23-1257
17-1444
35
S
South Fork Salmon River (Mainstem)
1958-1999
224-5029
138-6763
36
S
Tucannon River
1979-2002
11-897
11-16050
37
S
Wenaha River
1964-2002
48-2682
46-5838
38
S
Salmon River Lower Mainstem
1963-2000
11-432
52-636
39
S
Salmon River Upper Mainstem
1964-2000
18-2047
137-2545
40
UC
Entiat River
1962-1998
18-714
35-1111
41
UC
Methow River
1962-1998
43-2813
141-3894
42
UC
Wenatchee River
1960-2000
58-3523
166-6126
8
FIGURES
Figure 1: Shape of the four alternative population production models described in the Methods. Thin grey line is the 1:1 replacement line.
Figure 2. Map of the study area. Grey shading indicates ocean-type populations within the Puget Sound ESU; black shading represents stream-type populations from the Upper Columbia and Snake River ESUs.
Figure 3. Boxplots of models probabilities for the two different life-history types of Chinook salmon. Whiskers and hinges indicate the 95% and 50% credible intervals, respectively, around the median.
Figure 4 Stacked barplots of the posterior probability for each of the four competing models: linear (L), linear with Allee effect (A), hockey stick showing compensation (HS), hockey stick with depensation plus Allee Effect (AHS)
Figure 5. Plot of the stock recruitment data (in thousands) together with the posterior median of predicted values coming from model the Bayesian model model averaing and the most credible model only.
9
LITERATURE CITED
Achord, S., P. S. Levin, and R. W. Zabel. 2003. Density-dependent mortality in Pacific salmon: the ghost of impacts past? Ecology Letters 6:335-342. Barrowman, N. J., and R. A. Myers. 2000. Still more spawner-recruitment curves: the hockey stick and its generalizations. Canadian Journal of Fisheries and Aquatic Sciences 57:665676. Barrowman, N. J., R. A. Myers, R. Hilborn, D. G. Kehler, and C. A. Field. 2003. The variability among populations of coho salmon in the maximum reproductive rate and depensation. Ecological Applications 13:784-793. Berec, L., E. Angulo, and F. Courchamp. 2007. Multiple Allee effects and population management. Trends in Ecology & Evolution 22:185-191. Beverton, R. J. H., and S. J. Holt. 1957. On the dynamics of exploited fish populations, fishery investigations series II volume XIX, Ministry of Agriculture. Fisheries and Food 22. Boyce, M. S. 1992. Population Viability Analysis. Annual Review of Ecology and Systematics 23:481-506. Brook, B. W., and C. J. A. Bradshaw. 2006. Strength of evidence for density dependence in abundance time series of 1198 species. Ecology 87:1445-1451. Buhle, E. R., K. K. Holsman, M. D. Scheuerell, and A. Albaugh. 2009. Using an unplanned experiment to evaluate the effects of hatcheries and environmental variation on threatened populations of wild salmon. Biological Conservation 142:2449-2455. Burnham, K. P., and D. R. Anderson. 2002. Model Selection and Multi-Model Inference. Springer, New York. Carlin, B. P., and T. A. Louis. 2000. Bayes and Empirical Bayes Methods for Data Analysis. 2 edition. Chapman & Hall, New York. Connors, B. M., D. C. Braun, R. M. Peterman, A. B. Cooper, J. D. Reynolds, L. M. Dill, G. T. Ruggerone, and M. Krkosek. 2012. Migration links ocean-scale competition and local ocean conditions with exposure to farmed salmon to shape wild salmon dynamics. Conservation Letters 5:304-312. Correa, C., and M. R. Gross. 2008. Chinook salmon invade southern South America. Biological Invasions 10:615-639. Courchamp, F., T. Clutton-Brock, and B. Grenfell. 1999. Inverse density dependence and the Allee effect. Trends in Ecology & Evolution 14:405-410. Ellison, A. M. 2004. Bayesian inference in ecology. Ecology Letters 7:509-520 %508 Jun %509 Review %! Bayesian inference in ecology. Gelman, A., J. B. Carlin, H. S. Stern, and D. B. Rubin. 2004. Bayesian Data Analysis. Chapman & Hall, New York. Gregory, S. D., C. J. A. Bradshaw, B. W. Brook, and F. Courchamp. 2010. Limited evidence for the demographic Allee effect from numerous species across taxa. Ecology 91:2151-2161.
Helfield, J. M., and R. J. Naiman. 2006. Keystone interactions: Salmon and bear in riparian forests of alaska. Ecosystems 9:167-180. Hilborn, R., and C. J. Walters. 1992. Quantitative Fisheries Stock Assessment: Choice, Dynamics, and Uncertainty. Kluwer Academic Publishers, London. Hixon, M. A., S. W. Pacala, and S. A. Sandin. 2002. Population regulation: Historical context and contemporary challenges of open vs. closed systems. Ecology 83:1490-1508. Hoeting, J. A., D. Madigan, A. E. Raftery, and C. T. Volinsky. 1999. Bayesian model averaging: A tutorial. Statistical Science 14:382-401. Kramer, A. M., B. Dennis, A. M. Liebhold, and J. M. Drake. 2009. The evidence for Allee effects. Population Ecology 51:341-354. Launey, S., G. Brunet, R. Guyomard, and P. Davaine. 2010. Role of Introduction History and Landscape in the Range Expansion of Brown Trout (Salmo trutta L.) in the Kerguelen Islands. Journal of Heredity 101:270-283. Liermann, M., and R. Hilborn. 1997. Depensation in fish stocks: a hierarchic Bayesian metaanalysis. Canadian Journal of Fisheries and Aquatic Sciences 54:1976-1984. Lunn, D. J., A. Thomas, N. Best, and D. Spiegelhalter. 2000. WinBUGS - A Bayesian modelling framework: Concepts, structure, and extensibility. Statistics and Computing 10:325-337. Makhrov, A. A., E. Verspoor, V. S. Artamonova, and M. O'Sullivan. 2005. Atlantic salmon colonization of the Russian Arctic coast: pioneers from North America. Journal of Fish Biology 67:68-79. Mogensen, S., and J. R. Post. 2012. Energy allocation strategy modifies growth-survival tradeoffs in juvenile fish across ecological and environmental gradients. Oecologia 168:923933. Montgomery, D. R., J. M. Buffington, N. P. Peterson, D. SchuettHames, and T. P. Quinn. 1996. Stream-bed scour, egg burial depths, and the influence of salmonid spawning on bed surface mobility and embryo survival. Canadian Journal of Fisheries and Aquatic Sciences 53:1061-1070. Moore, J. W. 2006. Animal ecosystem engineers in streams. Bioscience 56:237-246. Moore, J. W., D. E. Schindler, and M. D. Scheuerell. 2004. Disturbance of freshwater habitats by anadromous salmon in Alaska. Oecologia 139:298-308. Moore, J. W., and B. X. Semmens. 2008. Incorporating uncertainty and prior information into stable isotope mixing models. Ecology Letters 11:470-480. NMFS (National Marine Fisheries Service). 2013. Annual Commercial Landing Statistics. Pascual, M. A., P. Kareiva, and R. Hilborn. 1997. The influence of model structure on conclusions about the viability and harvesting of Serengeti wildebeest. Conservation Biology 11:966-976. Plummer, M. 2003. JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling. Pages 20-22.
Ricker, W. E. 1954. Stock and recruitment. Journal of the Fisheries Research Board of Canada 11:559-623. Rubin, D. B. 1988. Using the SIR algorithm to simulate posterior distributions. Bayesian statistics 3:395-402. Ruckelshaus, M. H., K. P. Currens, W. H. Graeber, R. R. Fuerstenberg, K. Rawson, N. J. Sands, and J. B. Scott. 2006. Independent populations of Chinook salmon in Puget Sound. U.S. Dept. Commer., NOAA Tech. Memo. NMFS-NWFSC-78, 125 p. Ruggerone, G. T., B. A. Agler, and J. L. Nielsen. 2012. Evidence for competition at sea between Norton Sound chum salmon and Asian hatchery chum salmon. Environmental Biology of Fishes 94:149-163. Scheuerell, M. D., R. Hilborn, M. H. Ruckelshaus, K. K. Bartz, K. M. Lagueux, A. D. Haas, and K. Rawson. 2006. The Shiraz model: a tool for incorporating anthropogenic effects and fish-habitat relationships in conservation planning. Canadian Journal of Fisheries and Aquatic Sciences 63:1596-1607. Scheuerell, M. D., P. S. Levin, R. W. Zabel, J. G. Williams, and B. L. Sanderson. 2005. A new perspective on the importance of marine-derived nutrients to threatened stocks of Pacific salmon (Oncorhynchus spp.). Canadian Journal of Fisheries and Aquatic Sciences 62:961-964. Schindler, D. E., M. D. Scheuerell, J. W. Moore, S. M. Gende, T. B. Francis, and W. J. Palen. 2003. Pacific salmon and the ecology of coastal ecosystems. Frontiers in Ecology and the Environment 1:31-37. Taylor, E. B. 1990. Phenotypic Correlates of Life-History Variation in Juvenile Chinook Salmon, Oncorhynchus-Tshawytscha. Journal of Animal Ecology 59:455-468. Waples, R. S. 1991. Pacific salmon, Onchorhynchus spp., and the definition of "species" under the Endangered Species Act. Marine Fisheries Review 53:11-22. Westley, P. A. H., T. P. Quinn, and A. H. Dittman. 2013. Rates of straying by hatchery-produced Pacific salmon (Oncorhynchus spp.) and steelhead (Oncorhynchus mykiss) differ among species, life history types, and populations. Canadian Journal of Fisheries and Aquatic Sciences 70:735-746. Wintle, B. A., M. A. McCarthy, C. T. Volinsky, and R. P. Kavanagh. 2003. The use of Bayesian model averaging to better represent uncertainty in ecological models. Conservation Biology 17:1579-1590.