Size- and stage-dependence in cause-specific mortality of migratory brown trout

1. Evidence-based management of natural populations under strong human influence frequently requires not only estimates of survival but also knowledge about how much mortality is due to anthropogenic vs. natural causes. This is the case particularly when individuals vary in their vulnerability to different causes of mortality due to traits, life history stages, or locations. 2. Here, we estimated harvest and background (other cause) mortality of landlocked migratory salmonids


| INTRODUC TI ON
Population dynamics-particularly of long-lived species-are often highly sensitive to changes in mortality (Saether & Bakke, 2000).
Mortality can have a wide variety of causes (e.g. starvation, predation, disease, harvest), and vulnerability to cause-specific mortality may depend on individual factors such as age or life stage (Ronget, Garratt, Lemaître, & Gaillard, 2017). As a consequence, population-level responses to changes in mortality may vary greatly depending on the underlying cause, and disentangling different causes of mortality may provide insights crucial for population management and conservation (Williams, Nichols, & Conroy, 2002). This is particularly important in populations where a significant portion of mortality is linked to human activity; in such cases, knowledge about the relative impact of human-induced mortality and its effects on other mortality sources is crucial for developing sustainable and successful management strategies (Hilborn & Walters, 2013;Koons, Rockwell, & Aubry, 2014).
Studies of marked individuals constitute a highly valuable source of demographic data for wild animal populations and are essential for estimating survival, as well as cause-specific mortality. The recovery of a dead marked animal often provides information on the cause of death. However, unless animals are marked with radio-or satellite transmitters, most dead individuals will not be found, and this imperfect detection needs to be accounted for when estimating mortality parameters. Moreover, when considering multiple mortality causes, detection probability frequently depends on the cause of mortality, and some causes of mortality may not be observable at all. This is usually the case for natural mortality when dead recoveries are exclusively based on the reports of hunters or fishers (e.g. Koons et al., 2014;Servanty et al., 2010). Schaub and Pradel (2004) developed a multistate mark-recapturerecovery framework that allows separately estimating mortality from different causes while accounting for cause-dependent detection probabilities. Specifically, cause-specific mortalities are estimated as transitions from an 'alive' state to several 'dead from cause of interest' states. When this framework is extended to also include multiple 'alive' states, it becomes possible to estimate differences in vulnerability to cause-specific mortality depending on, for example, an individual's life stage (e.g. juveniles vs. adults, Schaub & Pradel, 2004) or location (Fernández-Chacón, Moland, Espeland, & Olsen, 2015). Such group-level differences in mortality can be tremendous and accounting for them is crucial for modelling population dynamics (Ronget et al., 2017). However, in addition to that, vital rates and population dynamics are often strongly affected by individual differences in continuous, dynamic traits such as body size (De Roos, Persson, & McCauley, 2003;Vindenes & Langangen, 2015). Particularly in species that are harvested and/or have indeterminate growth (e.g. fish species), cause-specific mortality is expected to depend strongly on body size. Fernández-Chacón, Moland, Espeland, Kleiven, and Olsen (2017) demonstrated this by estimating cause-specific mortalities for different sizes of Atlantic cod (Gadus morhua). However, they did so by lumping individuals into either of two size classes ('small' or 'large'), thus foregoing the possibility of investigating the continuous relationship between body size and mortality from different causes. Knowledge about the relationships between continuous traits and vital rates is, however, invaluable for studying population-level trait dynamics (e.g. using integral projection models; Ellner & Rees, 2006).
Migratory salmonid fishes are extensively studied due to their ecological, cultural and economical value (Drenner et al., 2012).
Nonetheless, studies at the population level are frequently hindered by a lack of knowledge about the mortality of adults residing in the sea or large lakes (Piccolo, Norrgård, Greenberg, Schmitz, & Bergman, 2012).
Many salmonid populations are heavily impacted by human activity, not only in the form of harvesting, but also through pollution, fish farming, habitat fragmentation and hydro-electrical power production (dams) in rivers (Aas, Klemetsen, Einum, & Skurdal, 2010), making the study of population-level consequences of such impacts a priority.
Here we study a population of migratory brown trout (Salmo trutta, hereafter 'trout') which inhabits a river-lake system in Eastern Norway and has been a popular target for fishing for decades due to its large body size. The spawning river is dammed, and trout migrating to spawning grounds above the dam face additional risks linked to dam passage on their up-and down-river migrations. Trout spawning below the dam, on the other hand, completely avoid these risks but may, in turn, incur costs related to poor river condition and crowding on the spawning grounds below the dam. Mortality risks are thus likely associated with spawning location in addition to individual body size and environmental conditions. To account for this heterogeneity, we re-parameterized mark-recapture models for cause-specific mortality in terms of mortality hazard rates (Cox, 1972;Ergon, Borgan, Nater, & Vindenes, 2018;Quinn, 2003) and extended the framework to include a continuous individual-and time-varying trait (body size) as a predictor of vulnerability within groups of individuals with different migration patterns. Fitting the resulting model to a unique 50-year time series of recaptures and recoveries of marked trout enabled us to investigate the effects of individual (size, spawning location, origin) and environmental (river discharge) factors on, and temporal variation in, several key vital rates: the vulnerability of adult trout to mortality due to harvest, natural causes and dam passage, as well as the probability of using a fish ladder within the dam to access upriver spawning areas. intuitive way to estimate temporal patterns within and correlation among different mortality sources.

K E Y W O R D S
Bayesian statistics, dam, harvesting, hazard rate, mark-recapture, mortality, NIMBLE, trout 2 | MATERIAL S AND ME THODS

| Study system and data
The studied population of landlocked migratory (potamodromous) brown trout inhabits the lake Mjøsa and its main inlet river, Gudbrandsdalslågen, in Eastern Norway. Eggs are deposited in the river in fall and develop over winter. After hatching in spring, juvenile trout remain in the river for an average of 4 years before smolting and migrating to the lake. They typically mature after 2-3 years of piscivorous diet and fast growth in the lake, and from that point on migrate up the river to spawn every other year (usually in August/ September, Figure 1). See Aass, Nielsen, and Brabrand (1989) for a more detailed description of the life history. The population consists of wild-hatched trout and stocked (first-generation hatchery-reared) trout, which are recognizable by their clipped adipose fin. Stocked trout are released into the river and lake as smolts but then follow the same general life history as wild-hatched individuals (Aass, 1993).
Shortly after the river was dammed in the 1960s, a fish ladder was installed to restore connectivity to the spawning grounds above the dam. Depending on body size and hydrological conditions, trout may either pass the dam by using the fish ladder on their upriver spawning migration, or reproduce below the dam (Aass et al., 1989;Haugen, Aass, Stenseth, & Vøllestad, 2008). Trout spawning above the dam have to pass the dam again on their return migration to the lake (in October/ November or in the following spring). Since the fish ladder cannot be used for moving downriver, these trout must pass either through the floodgates or the turbine shaft. Whether or not an individual uses the fish ladder thus determines not only its spawning location, but also the potential risks it encounters during the return to the lake.
From 1966 to 2016 a trap was operated within the fish ladder, allowing for all trout passing the ladder to be captured, measured, and individually marked. Thus, all adult trout were marked with Carlin tags (Carlin, 1955) when they used the fish ladder on an upriver spawning migration for the first time, and could be recaptured on subsequent spawning migrations given that they passed the ladder again. Subsequent spawning runs occur 2 years later for the majority of fish (98.5%), which adhered to a strictly biennial spawning cycle (Figure 1).
Over the 50-year time period, 13,975 adult trout were marked and 2,106 of these were recaptured in the ladder later. Since the population has been exposed to fishing over the entire time period, an additional 2,322 marked trout were reported dead by fishers. For more details on the marking scheme, sampling protocol and resulting data from the mark-recapture-recovery study, see Moe et al. (2020).
In this study we performed mark-recapture analyses over intervals of 2 years, as estimating parameters for spawning and non-spawning years separately was not possible (due to trout being unobservable in non-spawning years, Figure 1). We thus summarized the data into individual capture histories y i,t , in which each time index t corresponds to a 2-year time step (interval from current spawning year to next spawning year). For each time step, we coded three types of observations: 1 = alive and captured in the ladder, 2 = dead from harvest and reported and 3 = not observed. We set y i,t = 1 when an individual was captured in the fish ladder in any month during time interval t. Harvest of trout happens year-round ( Figure S1.1) and if an individual was harvested and reported at any point during interval t we set y i,t = 2, unless (a) the individual had also been caught in the fish ladder during interval t or (b) the harvest happened after August in the second year within the interval t. If either (a) or (b) was the case, we moved the harvest observation to the next interval such that y i,t+1 = 2. Furthermore, we excluded all individuals that did not follow a strictly biennial spawning cycle (1.5% of all individuals), did not have a single size measurement taken (<1%) or were of unknown origin (wild vs. stocked, <1%). The analyses presented here are based on the remaining 13,003 capture histories containing 1,498 trap recaptures and 2,252 harvest recoveries from both wild-hatched and stocked (hatchery-reared) trout.

| General model structure
We analysed the trout mark-recapture-recovery data in a multistate mark-recapture framework (Lebreton, Almeras, & Pradel, 1999) with both 'alive' and cause-specific 'newly dead' states ( Figure 2). Since trout are marked in the fish ladder while passing the dam on an upriver spawning migration, all individuals are in state 1, 'spawning upriver', at the start of their first 2-year time interval. State 1 individuals i may F I G U R E 1 Illustration of the biennial spawning cycle and mark-recapture scheme of the studied trout population. All individuals are marked in the fish ladder while passing the dam on an upriver spawning migration. Two years later they may be recaptured on the next spawning migration, but only if they pass the fish ladder to spawn above the dam (if they spawn below the dam, they are unobservable). Trout remain in the lake and are unobservable during nonspawning years survive from the current (t) to the next (t + 1) spawning migration with probability S 1,i,t and will then either use the fish ladder (probability p i,t+1 ) to spawn above the dam again, or remain below the dam for spawning (probability 1 − p i,t+1 ). Individuals using the ladder and thus remaining in state 1 are guaranteed to be observed, while individuals not using the ladder transition to state 2, 'spawning downriver', and are unobservable. Since spawning location may have a considerable effect on mortality, state 2 individuals have their own survival probability S 2,i,t , but we assume that their probability of using the fish ladder during the next spawning run (p i,t+1 ) does not differ from that of state 1 individuals.
When deaths of marked individuals can be observed and attributed to a cause, multistate mark-recapture models can be used to estimate the probability of dying from cause X as the transition from an 'alive' state to 'newly dead from cause X' state (Schaub & Pradel, 2004;Servanty et al., 2010). For the studied trout population, deaths due to harvest are clearly distinguishable from deaths due to other causes since fishers may report catching marked trout. Extending the model with the state 'newly dead from harvest' (state 3) thus allows us to include the probability of an individual i in state n (n = 1 for above-dam spawners, n = 2 for below-dam spawners) dying due to harvest, Ψ H n,i,t , and dying due to other causes, Ψ O n,i,t , over the time interval t to t + 1. Individuals that have recently died due to harvest (state 3) may be reported by fishers with reporting rate r t . Individuals that die due to other causes are not observable and therefore transition directly to the 'dead' state (state 4; see Figure 2). The resulting multistate model for the trout mark-recapturerecovery data can be expressed with the following state transition matrix and associated observation probabilities:

| Parameterization by mortality hazard rates
Different cause-specific mortality probabilities (Ψ) are not independent of one another; if a certain cause of mortality becomes more prevalent (e.g. due to some event or change in the environment), not only will the probability of dying from that cause increase, but the probability of dying from any other cause will decrease at the same time. This confounding complicates inference (e.g. Cooch, Guillemain, Boomer, Lebreton, & Nichols, 2014), but Ergon et al. (2018) recently re-emphasized that this can be avoidedalso in the context of discrete-time mark-recapture analyses-by parameterizing with mortality hazard rates instead of probabilities (Cox, 1972;Quinn, 2003). Assuming that the intensities of mortality from different causes remain proportional within time intervals, we can define the survival and mortality probabilities in the trout model using harvest (m H ) and other-cause (hereafter 'background') mortality hazard rates (m O ): (see derivation in Ergon et al., 2018).
In the present implementation, we further constrained harvest mortality to be the same for trout spawning above and below This constraint was necessary to obtain an identifiable model, but also biologically reasonable since spawning location is expected to have only minor (if any) effects on 2-year harvest mortality for two reasons: (a) the time trout spend in the river, i.e. the spawning run, constitutes a relatively small proportion of the 2-year analysis interval relative to the time all trout spend together in the lake, presumably experiencing the same harvest pressure, and (b) during the spawning run itself, differences in harvest pressure of above-and below-dam spawners are likely small due to general fishing restrictions in the river in fall.

| Individual and temporal variation in vital rate parameters
Body size and hydrological conditions are often key determinants of vital rate variation in freshwater fish, including our study population (e.g. Carlson, Olsen, & Vøllestad, 2008;Haugen et al., 2008;Letcher et al., 2015). We thus used individual body size (length; mm) at the beginning of the time interval and average river discharge during the relevant season as covariates for mortality and ladder usage in our model. We further accounted for potential effects of hatchery origin and additional among-year variation in all parameters x using intercept offsets for stocked individuals ( x stock ) and temporal random effects ( x t ) respectively. Random effects on all parameters were assumed to be independently normally distributed on the link scale (but see Supporting Information S6 for a model extension with correlated random effects).
Harvest in our study system has been done mostly using fishing rods and gillnets; the selectivity of the former is often positively correlated with body size (Lewin, Arlinghaus, & Mehner, 2006) while the latter typically have bell-shaped selectivity curves (Hamley, 1975). Since we here pooled harvest by all gear types, we modelled linear and quadratic effects of size on harvest hazard rate on the log-scale: where µ H is the median harvest hazard rate, H S and H SS are slope parameters for linear and quadratic size effects respectively. size i,t is the individual length at spawning and origin i is a binary variable taking values of 1 for stocked fish and 0 for wild-hatched fish.
Background mortality is expected to depend not just on body size, but also on spawning location and river discharge, as above-and below-dam spawners encounter different hydrological conditions during/after spawning and only the former need to pass the dam on their downriver migration. Mortality associated with the spawning migration in general, and passing of the dam in particular, may also depend on body size. We thus modelled background mortality hazard rate as: Here O n is the median background mortality hazard rate of state n, discF t is the average discharge during the fall when many postspawned trout are expected to migrate downriver (October to November), and O D,n and O S,n are slope parameters for discharge and size effects respectively. Stocking effects ( O stock ) and temporal random effects ( O t ) for background mortality were assumed to be shared across states n.
The probability of using the fish ladder and thus spawning above the dam was previously found to depend on a complex interplay of individual body size and river discharge (Haugen et al., 2008). We adopted the basic model structure from this earlier analysis and extended it by allowing for stocking effects and random among-year variation such that: The discharge covariate used here, discS t , differs from the one used above and represents the average discharge over the summer season when trout undertake their upriver spawning migration (July to October).

| Size imputation under imperfect detection
Using continuous, time-varying individual traits such as body size as covariates in mark-recapture models is problematic due to imperfect detection: information on body size will be missing for sampling occasions when an individual is not captured (Pollock, 2002).
There are several ways to approach this problem, including integrated growth models (e.g. Bonner, Morgan, & King, 2010) and inter-/extrapolation using other available data and/or separate models. Due to the prohibitively large computational demands of an integrated analysis, we here adopted the latter approach and used a detailed growth model previously developed for the study population of brown trout  to impute missing values in the individual size covariate. Specifically, we re-fitted the growth model of Nater et al. (2018) to an extended set of growth data from 6,843 individuals spanning the years 1952-2003 and used the resulting parameter estimates, plus a correction factor, to calculate all missing entries in the body size covariate.
The imputation procedure, as well as implementation and results of the growth analysis, are described in detail in Supporting Information S5.

| Autoregressive reporting rate model
Time-dependent reporting rate r t can be expected to vary considerably over a period of 50 years. To accommodate this, we followed the example of Zhao, Boomer, and Kendall (2018) and used a flexible, autoregressive model for time-dependent reporting rates: where r t are independently normally distributed random effects. For details on the implementation of the autoregressive model in logit r t = logit r t−1 + r t , the context of the overlapping 2-year time intervals in our model, we refer readers to the model code (Supporting Information file nimbleDHMM.R).

| Implementation with NIMBLE
We implemented the model in a Bayesian framework in NIMBLE (de Valpine et al., 2017). Building on the work of Turek, de Valpine, and Paciorek (2016)

| Model identifiability and fit
With increasing model complexity, and particularly when unobserved states are included, it is not obvious whether all parameters within a multistate mark-recapture model can be estimated (Gimenez, Choquet, & Lebreton, 2003;Lebreton & Pradel, 2002).
Using an extended (hybrid) symbolic method (Choquet & Cole, 2012;Cole, 2012 From each simulated dataset, we then extracted several test statistics representing numbers and size distributions of recaptured/ harvested trout and compared them to the same quantities obtained from the real data using visual tools and Bayesian p-values.
Methodology and results of the PPCs are described in detail in Supporting Information S4.

| Model identifiability and fit
We found that in the absence of random effects, the only model structures that were intrinsically identifiable were those in which harvest mortality depended on an individual time-varying covariate (e.g. body size) and background mortality was either constant or dependent on an environmental covariate (Table S3.1). However, all models (irrespective of covariate structure) became identifiable when random year effects were included on at least harvest hazard or reporting rates (Table S3.1). Prior-posterior overlaps were below 35% for all parameters except r 1 , indicating no major problems with near-identifiability (Supporting Information S3.3).
Posterior predictive checks indicated that overall, the model pro- For detailed PPC results, refer to Supporting Information S4.3.

| Size-dependent fish ladder usage
Posterior distributions for all estimated parameters are plotted in Figures S1.2-S1.10. Numerical results in the following text are displayed as median [95% credibility interval].

F I G U R E 3
Predictions of the effects of body size on ladder usage probability at different levels of river discharge. Grey, dashed = low discharge (mean − SD). Grey-blue, solid = average discharge (mean). Blue, longdashed = high discharge (mean + SD). Lines represent median prediction, ribbons indicate 95% credibility intervals. The blue density kernel above the plot visualizes the size distribution of trout caught in the ladder (data) The probability of using the fish ladder-and thus spawning above the dam-depended strongly on individual size and, to a lesser degree, on river discharge (Figure 3). Intermediate-sized trout (600-700 mm) were most likely to pass the dam under any discharge conditions. Small to intermediate-sized trout had a higher probability of using the ladder when river discharge was high, whereas the probability decreased markedly with size for larger trout irrespective of hydrological conditions. Ladder usage probability fluctuated considerably over time ( Figure 4c) and was predicted to be lower for stocked  Figure S1.11).

| Cause-and size-dependent mortality
Median mortality hazard rates were estimated at 1.  (Figure 5a). Background mortality hazard rate, while mostly lower than harvest hazard rate, decreased with size for above-dam spawners and increased with size for belowdam spawners (Figure 5a). Consequently, total survival probability increased with size for all trout up to 870 mm, but flattened out for larger below-dam spawners (Figure 5b). River discharge was predicted to increase background mortality of above-dam spawners only ( Figure S1.2). Residual among-year random variation was substantial in harvest and especially background mortality, with hazard rates at the 97.5 percentile being 1.28-and 69.67-fold higher than at the 2.5 percentile respectively, but no temporal trends were evident in either mortality cause (Figure 4a,b).
Model results did not support differences in harvest-or background mortality due to trout origin: hazard ratios of stocked and F I G U R E 4 Estimates for time-dependent (a) log harvest hazard rate, (b) log background mortality hazard rate (above-dam spawners), (c) ladder usage probability, and (d) reporting rate (calculated using random variation and discharge effects). Lines represent median predictions, ribbons indicate 95% credibility intervals Reporting rate (2 years) (d)

F I G U R E 5
Predictions of the effects of body size on (a) harvest and background mortality hazard rates and (b) survival probabilities (under consideration of both mortality sources). Red and blue curves apply to individuals that have last spawned above and below the dam respectively. The black curve (harvest) applies to all individuals irrespective of their last spawning location. Lines represent median predictions, ribbons indicate 95% credibility intervals. Density kernels above the panels illustrate the informative data range: red = size distribution of individuals captured in the fish ladder (above-dam spawners, raw data), blue = simulated size distribution of unobservable below-dam spawners

| Temporal patterns in reporting rate
A clear decrease in estimates of reporting rate over the 50-year time period was evident (Figure 4d), with values exceeding 50% in early years but dropping below 10% towards the end of the time series.

| D ISCUSS I ON
Individuals can differ greatly in their vulnerability to mortality from different causes depending on traits like body size and variation in exposure to mortality risk (e.g. as a consequence of reproductive state or location). Particularly when some mortality causes are directly linked to human activity, understanding and accounting for such individual differences in vulnerability can be crucial for management and conservation. In this study, we combined recent advances in mark-recapture methodology and Bayesian modelling to investigate factors determining vulnerability of large migratory brown trout to harvest and background mortality in a system heavily impacted by fishing and hydropower production.

| Size-dependence of cause-specific mortality
Size-dependent survival is well documented for salmonid fishes like brown trout, but direction and strength of size effects vary widely across habitats, populations, years and life history stages (Carlson et al., 2008;Drenner et al., 2012). Here, we were able to not only reproduce previous findings of positively size-dependent survival for the studied trout population (Figure 5b, Haugen et al., 2008), but also able to disentangle the underlying contributions from mortality due to harvest and other causes.
Model results supported our initial expectation of nonlinear dependence of vulnerability to harvest and body size: harvest mortality was highest for trout with sizes of around 550 mm and decreased for both smaller and larger individuals (Figure 5a). Bell-shaped selectivity curves such as this are typical for gillnets (Hamley, 1975), which have been commonly used in our study area. The low harvest mortality of large trout, however, may seem surprising given that 44% of the reported harvests were due to angling, which often targets larger fish (Lennox et al., 2017). This may indicate that large trout escape harvest either through their individual behaviour (e.g. different foraging habitats and prey preferences, or learning, Arlinghaus, Klefoth, Kobler, & Cooke, 2008;Lewin et al., 2006) or because cohort selection favours more cautious fish, allowing them to survive and grow to large sizes (Lennox et al., 2017).
Effects of body size on background mortality were predicted to be markedly different for trout spawning above and below the dam, in particular for larger trout (Figure 5a). Trout spawning above the dam generally had low background mortality, possibly indicating limited mortality risk associated with dam passage for adult fish. Nonetheless, smaller individuals were slightly more vulnerable to dying from non-harvest causes than larger ones (92% of posterior samples indicated a negative effect of size on background mortality, Figure S1.2).
Two mechanisms that may be (partially) responsible for this are turbine mortality and energetic costs of dam passage. During downriver migration after spawning above the dam, trout have to pass through the floodgates or the turbine shaft to return to the lake. As on many hydroelectric dams, racks are installed in front of the Hunderfossen power plant's turbine intake to prevent fish from entering, but small fish may slip through the grid and subsequently suffer severe injury and die passing the turbine (e.g. Fjeldstad, Pulg, & Forseth, 2018).
Alternatively, smaller fish may have reduced survival following dam passage due to large energy expenditures resulting from dam passage (on up-and/or down-river migration) itself (e.g. Roscoe, Hinch, Cooke, & Patterson, 2011) or as a consequence of migration delays, particularly if these force individuals to overwinter in the river (Fjeldstad et al., 2018). Both of these mechanisms are plausible here when also considering that background mortality of above-dam spawners was predicted to increase at higher levels of river discharge ( Figure S1.2): stronger water flow could increase both the risk of being swept into the turbine shaft and the energetic costs of passage.
Unlike trout spawning above the dam, trout spawning below the dam were predicted to be more vulnerable to background mortality at larger sizes (Figure 5a). Many mechanisms may be responsible for this; one possibility is related to trout density downriver of the dam, which can be very high during the spawning season (Kraabøl, 2006) and likely results in elevated levels of stress, aggressive interaction and disease transmission. Mortality below the dam could increase with body size if larger individuals (due to their size, age or other traits correlated with large body size) are less able to cope with these challenges and/or increased their investment into reproduction at the cost of survival under adverse conditions. At the same time as having higher background mortality below the dam, large trout were also much more likely to spawn below the dam in the first place ( Figure 3), and thus incur the resulting higher mortality. The hydropower dam therefore has the potential to function as an ecological trap (Schlaepfer, Runge, & Sherman, 2002) via its size-selective fish ladder and adverse conditions on downriver spawning grounds, particularly when considering that the reproductive output of large fish is often central to the viability of salmonid populations (Jonsson & Jonsson, 2011).
A second, more practical consequence of the selectivity of the fish ladder is that it substantially limited comparisons of background mortality of above-and below-dam spawning trout of the same size in this study. With small and large trout predominantly spawning above and below the dam respectively, direct comparisons are only informative for a relatively narrow size range (~700-850 mm).
Within this range, predictions for above-and below-dam spawners mostly overlap, with the exception of the largest sizes ( Figure 5).
Additional data-particularly on the fates of individuals spawning below the dam-would be necessary for a more detailed assessment of the interactive effects of hydropower production and spawning location on mortality and for investigating potential mechanisms explaining higher mortality among large fish below relative to above the dam. What our approach did allow, however, was an unbiased quantitative comparison of size-dependent harvest and background mortality: the risk of dying due to fishing was higher than the risk of dying due to any other cause for almost the entire size range, suggesting fishing as the main source of adult mortality in this population (see Kleiven et al., 2016, for a similar result on Atlantic cod).

| Temporal variation over 50 years
The present analysis extended over half a century, in which the riverlake system experienced variation in abiotic and biotic factors due to river regulation, lake restoration and changes in climate and human activities (Hobaek et al., 2012). It is therefore unsurprising that we found high among-year variation in cause-specific mortality and fish ladder usage over the course of the 50-year study period (Figure 4a-c).
Background mortality in particular was subject to large fluctuations and displayed a marked increase during the period 1997-2001 (also visible in overall mortality and survival, Figure S1.12). This coincides with a documented outbreak of a fungal disease in the study population (Saprolegnia spp. infections, possibly in combination with ulcerative dermal necrosis, Johnsen & Ugedal, 2001). This suggests that disease may be a key driver of changes in adult trout mortality and has the potential to substantially affect population viability (Hudson, Rizzoli, Grenfell, Heesterbeek, & Dobson, 2002). Since freshwater ecosystems are particularly vulnerable to infectious diseases (Okamura & Feist, 2011), studying fungal disease dynamics and how they are affected by harvest, river regulation and other environmental factors (e.g. temperature, Letcher et al., 2015) represents an important venue for future research.
Unlike cause-specific mortality and ladder usage, which displayed strong fluctuations but no obvious trends, fisher's reporting rate decreased clearly and rapidly over time: from over 50% of catches being reported in the beginning of the study period to less than 10% in the last two decades (Figure 4d). Declining fisher engagement over time is a known problem in tagging studies without reward tags (Piccolo et al., 2012), and highlights the importance of maintaining volunteer participation in long-term studies by providing appropriate feedback and keeping up with technological development of tools and platforms for reporting (Dickinson et al., 2012).

| Model limitations
When analysing long-term ecological data even complex hierarchical models, like the ones used here, can fail to sufficiently capture heterogeneity (overdispersion) in the data, resulting in lack of model fit (Richards, 2008). PPCs (Conn et al., 2018) showed that overall our final model fit the data reasonably well, but also revealed that goodness-of-fit varied substantially across the study period. Particularly the early years in the data, which correspond to the first two decades following dam construction, were characterized by relatively poorer model fit ( Figures S4.3 and S4.7).
Many individuals present during this period were born before the damming of the river and the implementation of the stocking programme. They may have experienced environmental conditions vastly different from individuals later in the time series, possibly resulting in long-lasting cohort effects not uncommon for salmonid fishes (e.g. Vincenzi, Mangel, Jesensek, Garza, & Crivelli, 2016).
Furthermore, given the profound changes in harvest practices (gradual shift from gillnet to rod fishing, Aass & Kraabøl, 1999), river regulation (flow regimes, turbine intake grid sizes, etc.) and disease prevalence during the 50-year study period, it is also not unlikely that size-dependence of mortality and migratory behaviour itself has changed over time. Overdispersion in our data could thus be related to changes in selection pressures, something that may warrant attention in future studies.
Both parameter estimates and resulting model fit were sensitive to the way we imputed body size, illustrating that covariate imputation remains the main challenge of mark-recapture models with continuous individual time-varying covariates like body size (Bonner et al., 2010;Pollock, 2002). Imputing body size using mean estimates from an externally run growth model, as we have done here, comes with several limitations. First, data used to estimate growth may not be representative of the individuals contained in the mark-recapture data. In our case, most data on growth in the lake pertain to the subadult life stage (prior to maturation) and resulting growth estimates may thus be less well suited for the mature, spawning trout that make up the mark-recapture data. Second, growth data are only available for 53% of individuals and 74% of years (only up to 2003) contained in the mark-recapture data. Size imputation for a nonrandom sample of individuals was thus lacking estimates of year and individual random effects. Finally, and perhaps most importantly, by directly imputing size using mean estimates of growth model parameters, we omitted all uncertainty in size estimates arising from residual variation in growth (stochasticity) and parameter uncertainty.
Since the reduced growth model we used matched well with observations ( Figure S5.1) and fit of the mark-recapture-recovery model was overall decent, it is unlikely that the results we present here are biased to a degree as to invalidate any of the main conclusions.
However, as a result of direct size imputation and likely related lack of model fit, some of the patterns and effects may be estimated with inflated precision and this has to be considered when interpreting the presented relationships.

| Outlook: Data integration and population perspective
The fundamental issues arising from imputing missing individual covariate values can be addressed through integrated analysis of growth and survival/state transition processes (Bonner et al., 2010;Letcher et al., 2015), which allows imputation of the 'true' latent body size and estimation of its effects on vital rates without bias and under full consideration of uncertainty. In our case, not just one but two distinct data sources provide information on growth: length measurements from trout captured in the fish ladder (markrecapture data) and lengths back-calculated from scale year rings of a subset of marked individuals. This provides a unique opportunity for integrated analysis of multiple data sets which is likely to result in more precise estimates of vital rates, more comprehensive understanding of variation therein and insights into potential discrepancies among different types of data (Plard, Fay, Kéry, Cohas, & Schaub, 2019;Saunders et al., 2019).
The large drawback of Bayesian integrated analysis is its high computational costs, and in the case of the present data and model, computational demands precluded a fully integrated analysis.
However, in Supporting Information S2 we have shown how implementing the mark-recapture-recovery model with a custom distribution in NIMBLE can lead to dramatic increases in computational efficiency (32-times faster MCMC than with standard JAGS). With the continuing development of both computational power and flexible, user-friendly MCMC software, large integrated analyses will likely become more feasible in the future.
More efficient computational solutions are also becoming invaluable when looking beyond single vital rates (growth, survival) and towards more holistic models of population dynamics. Several of the results presented here may have important implications for brown trout management but questions such as whether the high harvest mortality of adult trout has consequences for population viability or whether the dam does indeed function as an ecological trap, can only be addressed by adopting a population perspective.
The framework of integrated population models (Plard, Fay, et al., 2019) in general, and recent extensions for populations structured by continuous traits in particular (Plard, Turek, Grüebler, & Schaub, 2019), lend themselves well to the study of these questions for our system and will follow naturally from the integration of growth and survival estimation. Fully integrated, size-structured population models will further provide new opportunities to study the joint impacts of harvesting, stocking, habitat alteration, climate change and disease dynamics (Plard, Fay, et al., 2019) and are thus highly relevant for future studies aiming to improve understanding and inform management of the trout in lake Mjøsa and of animal population inhabiting ecosystems heavily impacted by human activity in general.

| CON CLUS IONS
Multistate mark-recapture models are powerful tools for estimating and understanding survival in animal populations that experience mortality from both natural and anthropogenic causes (Schaub & Pradel, 2004). We used such a model to disentangle harvest-and background mortality of adult brown trout and showed that (a) harvest generally outweighed all other sources of mortality and (b) that vulnerability to both mortality causes was determined by individual differences in body size and migration pattern (dam passage). The use of a novel hazard rate parameterization  and data from both recaptures and harvest recoveries allowed to estimate size-dependence and among-year variation in cause-specific mortality, state transition probabilities and reporting rate without confounding. This framework, including its computationally efficient implementation, is highly applicable to other studies of cause-specific mortality in populations whose vital rates are strongly affected by continuous traits, and may prove particularly valuable also in the context of estimating correlation among different sources of mortality. Finally, we illustrated that the use of an appropriate year random effects structure can be a prerequisite to establishing identifiability of complex mark-recapture models and is therefore crucial to obtain reliable estimates of vital rate parameters. In practice, such random effects can only be estimated when data are collected over a sufficient number of years, emphasizing the importance of investing in the (continued) collection of individual-based data over long time periods (Clutton-Brock & Sheldon, 2010).

ACK N OWLED G EM ENTS
We thank three anonymous reviewers for constructive feedback and suggestions during the review process. This work was supported by the Research Council of Norway (project SUSTAIN, 244647/E10).
We thank all individuals and institutions that have been involved in collecting, maintaining and organizing the trout data (contributions detailed in Moe et al., 2020), and the Norwegian Water and Energy Directorate (NVE) for providing river discharge data. Model fitting was performed on the Abel Cluster (University of Oslo and UNINETT Sigma2 AS).

AUTH O R S ' CO NTR I B UTI O N S
C.R.N., T.E., Ø.L., Y.V. and L.A.V. conceived the ideas; C.R.N. and T.E. designed the methodology; C.R.N., P.A., S.J.M. and A.R. prepared the data for analysis; C.R.N. analysed the data and led the writing of the manuscript; D.T. developed and tested the custom likelihood and drafted Supporting Information S2. D.C. designed identifiability analyses and drafted Supporting Information S3. All authors contributed critically to the drafts and gave final approval for publication.

DATA AVA I L A B I L I T Y S TAT E M E N T
The complete mark-recapture-recovery and growth data sets are available on the Dryad Digital Repository: https://doi.org/10.5061/ dryad.9cnp5 hqf4 , and documented in the accompanying data paper .