A Novel Approach For Estimating Growth and Mortality of Fish Larvae

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

ICES Journal of Marine Science (2021), 78(8), 2684–2699. https://doi.org/10.

1093/icesjms/fsab161

Original Article
A novel approach for estimating growth and mortality of fish

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


larvae
,*  , 
Charles Hinchliffe , Pierre Pepin , Iain M. Suthers , and Daniel S. Falster

1
Evolution and Ecology Research Centre and School of Biological, Earth, and Environmental Sciences, University of New South Wales, Sydney, NSW,
2052, Australia
2
Northwest Atlantic Fisheries Centre, Fisheries and Oceans Canada, 80 East White Hills, P.O. Box 5667, St. John’s, NL A1C 5X1, Canada
3
Sydney Institute of Marine Science, Chowder Bay Road, Mosman, NSW 2088, Australia

Corresponding author: e-mail: c.hinchliffe@unsw.edu.au
Hinchliffe, C., Pepin, P., Suthers, I. M., and Falster, D. S. A novel approach for estimating growth and mortality of fish larvae. – ICES Journal of
Marine Science, : –.
Received  March ; revised  August ; accepted  August ; advance access publication  September .

Accurate estimates of growth and mortality are needed to understand drivers of production and cohort success. Existing methods for estimating
mortality rates, such as catch-curves, require large sample sizes, as they work by grouping individuals into age-bins to determine a frequency
distribution. Yet, sampling enough larvae is often not possible at fine scales within the constraints of research projects, due to low density of
larvae in pelagic environments. Here, we develop a novel method to simultaneously estimate growth and mortality rates of fish larvae as a
continuous function of size using theory of size-structured populations, eliminating the need to group data into age-bins. We compare the
effectiveness of our model to existing methods by generating data from a known distribution. This comparison demonstrates that while all
models recover correct parameter values under ideal circumstances, our new method performs better than existing methods when sample sizes
are low. Additionally, our method can accommodate non-linear growth and mortality functions, while also allowing growth and mortality to vary
as functions of environmental co-variates. This increased accuracy and flexibility of our method should improve our ability to relate variability
in larval production to environmental fluctuations at finer spatial scales.
Keywords: catch-curve, larval fish, population dynamics, recruitment potential, size-structure, survival.

1984). Yet, despite advances in our understanding of key processes,


Introduction the low availability of information on population metrics such as
Forecasting productivity of fish stocks has been a primary goal of spawner biomass and high variability in survivorship limits relia-
fisheries science for over 100 years (Dannevig 1907; Hjort 1914); bility of model fits and prevents meaningful predictions of recruit-
however, our ability to make accurate forecasts of abundance from ment (Pepin 2015a).
adult stocks is often hindered by the high rates of demographic To improve our understanding of environmental effects on the
variability in the early life stages (Pepin 2015a). Various hypothe- larval period and cohort success, we need to better resolve popula-
ses have been developed to explain what determines the success of tion dynamics, which is difficult in marine ecosystems. The relative
fish from egg to juvenile stages (Hare 2014), based on the variabil- importance of growth and mortality in determining survivorship is
ity of migratory pathways, adult abundances, parental investment, known to be potentially equal (Houde 1987; Beyer 1989), but es-
prey abundance and distribution, dispersal (Hjort 1914), predation timating growth is both more accurate and easier than mortality,
(Bailey and Houde 1989), physiological responses to the environ- which has led to an unequal focus on growth in empirical stud-
ment (Fry 1971; Neill et al., 1994), growth and mortality trade- ies (Takasuka et al., 2017). This discrepancy is likely caused by the
offs (Anderson 1988), fishing mortality (Beverton and Holt 1957), nature of oceanographic studies, where temporal resampling of a
and adult trophic interactions (Odum and Odum 1955; Polovina cohort is impractical or in many cases, impossible for pelagic larvae.

C The Author(s) 2021. Published by Oxford University Press on behalf of International Council for the Exploration of the Sea. All rights
reserved. For permissions, please e-mail: journals.permissions@oup.com
A novel approach for estimating growth and mortality of fish larvae 

Moreover, current “snapshot” techniques require large sample sizes, as such, that area may provide a disproportionate number of sur-
which may be difficult to collect as a result of financial or time con- vivors to the juvenile stages. This could mean that interpreting the
straints on research projects. Estimating mortality rates at different mortality to growth ratio estimate for the entire seascape alone pro-
relative scales in contrast to growth rates (e.g. population versus vides a more dire assessment of potential success than exists.
individual) compounds the problem (Govoni 2005). This imbal- The minimum spatiotemporal scale at which variance in mor-
ance of focus is made more problematic as high rates of predation- tality and growth could be estimated will vary between species
driven mortality have been shown to positively bias sampling of and seascapes. To sample enough larvae to fit catch-curve models
growth rates (Takasuka et al., 2003; Tian et al., 2007). We therefore appropriately (Ricker 1975), the volume of a tow or the number of
need to improve our ability to estimate mortality and increase ef- collated tows should reflect the in situ abundance of the species be-
forts to incorporate both these vital rates into empirical research ing investigated, which will influence the spatial resolution of out-
if we hope to enhance our ability to make predictions of cohort puts. The stability of the seascape may also limit spatial resolution,
success. factoring in bias on mortality estimates caused by aberrant drift
Well-resourced longitudinal studies can follow cohorts through (Houde 2002), as scale dependency may largely be restricted owing

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


space and time to estimate changing rates of growth and mortality to losses via advection (Helbig and Pepin 1998). A significant issue
(Rutherford et al., 1997; Vanalderweireldt et al., 2019), though these which needs considering is variation in the hatch-rate of individ-
methods are heavily dependent on spawning seascape and often not uals, which could violate a key assumption of catch-curve models.
possible outside of estuaries or semi-enclosed bays. Alternatively, Variability in the flux of individuals into the system can propagate
many currently used methods estimate growth and mortality based error in estimates of mortality, which if fluctuating around a mean
on a standardised snapshot of a population (i.e. a survey or program daily value can be largely ignored (Ricker 1975); however, if the
level estimation that does not incorporate time), using subsequent variance in hatch-rate is non-random, for example, shifting because
otolith microstructure analysis where age is estimated from daily of a spawning season ending, estimates of mean mortality may be
growth rings (Stevenson and Campana 1992). Growth can be mea- skewed. For this reason, a good understanding of a species spawn-
sured at the individual level, while mortality is estimated at the sam- ing ecology and habitat is needed. Attempts to resolve mortality and
ple or population level. Sampling mortality requires a large number growth at finer scales should probably be limited to species who ex-
of observations to achieve adequate resolution. Typically, growth is hibit protracted spawning times, such as Pacific sardine (Sardinops
determined from length-at-age relationships using linear or non- sagax) in coastal waters of south-eastern Australia (Sexton et al.,
linear models, such as a Gompertz (Gompertz 1825) or von Berta- 2019) or Atlantic cod (Gadus morhua) in inshore bay areas of
lanffy growth equations (von Bertalanffy 1938). Weight-specific in- southern Newfoundland (Rose and Rowe 2020), and applied to
stantaneous growth may also be estimated using exponential mod- samples from areas that spatially overlap spawning locations.
els (Sassa et al., 2014). Mortality is most commonly estimated from There are, however, other shortcomings in the way mortality
an age-frequency distribution, where the relationship between the rates are estimated using current methods, which prohibit detailed
abundance of individuals in adjacent age-bins is used to determine inferences. First, data are typically divided into relatively large age
an instantaneous mortality coefficient (Pepin 2015b). Early research or length bins, making the methods sensitive to the choice of bin
commonly used length data (rather than ages) directly to estimate widths and sample size (Figure 1). This is true for a variety of com-
mortality, where length was considered to be representative of age mon methods such as catch-curves (Ricker 1975; Maceina 1997)
based on a linear age–length relationship (Pepin 1991; Houde 1997; and the Chapman–Robson method (Chapman and Robson 1960),
Pepin 2015b). Remarkably, these approaches enable researchers to described further in the "Estimators" section. Second, to obtain
measure demographic rates, based only on a one-off snapshot of the larger sample sizes, researchers often combine multiple samples
population from a single survey. across large temporal or spatial distributions to determine a mean
The scale of such snapshots commonly varies between studies, mortality rate (Watanabe et al., 1997; Sassa et al., 2014; Sassa and
from estimates derived by combining multiple tows for a single sur- Takahashi 2018), which hinders our ability to resolve environmen-
vey (e.g. Sassa and Takahashi 2018) to estimates determined from tal effects on cohort success at fine scales relevant to the drivers of
a database comprised of several surveys across multiple years (e.g. instantaneous vital rates (i.e. at the tow/sample level). Third, sepa-
Bailey et al., 1996; Rutherford et al., 1997). In some cases, such as rate models are used to estimate growth and mortality rates which is
for plaice (Pleuronectes platessa) in the North Sea (Bannister et al., problematic in two ways. First, data from the same specimens must
1974), survey-level mortality rates of larvae have accurately pre- be handled in multiple steps using multiple models. Second, un-
dicted subsequent year-class strength when hindcast, although suc- certainty in one model may not be incorporated appropriately into
cessful forward predictions of recruitment remain a rarity. To max- the other. As length-at-age is typically only known for a subset of a
imise the utility of sampling data, we need to estimate rates of mor- sample population, which are then used to determine the ages of all
tality and growth across all possible spatial scales. For example, one individuals in a sample (i.e. using a growth rate) (Pepin 2015b), un-
can imagine a situation where age-size data from an extensive sam- certainty in the length-at-age relationship is not conserved in the es-
pling survey comprised of many tows are used to estimate a sin- timation of mortality. Combined, these shortcomings suggest new
gle rate of mortality and growth for a given seascape. Based on the techniques that simultaneously estimate both vital rates from an ob-
outcome of this work, the ratio of mortality to growth is high (or served size distribution are needed.
growth to mortality is low), thus one concludes that the recruitment Size distributions capture the demographic consequences of
potential for the species is low for that survey period (e.g. Sassa growth and mortality rates (Muller-Landau et al., 2006) (Figure 2).
et al., 2014) and may predict the number of individuals likely to A formal link between growth and mortality rates with size distri-
reach the juvenile stage is also low (Secor and Houde 1995; Houde butions is described mathematically by McKendrick–von Foerster
1997). If variability in the ratio of mortality to growth across the equation (6) (Von Foerster 1959; Sinko and Streifer 1967), where
seascape is more easily observed, it is possible that larvae of some the number of individuals at a given size is related to the flux of
areas are experiencing a much lower mortality to growth ratio, and individuals into a population (i.e. hatch-rate or recruitment rate)
 C. Hinchliffe et al.

ticularly for the early life stages, given we can quantify the size, age,
and density of larvae based on the estimated volume of a plankton
tow. This approach could help to improve upon shortfalls of com-
mon mortality estimation; first by modelling growth and mortal-
ity within a single framework thereby conserving uncertainty from
the length- or weight-at-age relationship in mortality estimates, and
second by removing the need to arbitrarily set age-bin sizes, thereby
making the approach less sensitive to sample size (Muller-Landau
et al., 2006).
Here, we develop variants of a size distribution model to deter-
mine rates of growth and mortality for three data handling scenar-
ios common in fisheries literature: (i) the age of every fish in a sam-
ple is known; (ii) the size of every fish in a sample is known and

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


a literature-derived growth rate is applied, and (iii) the size of ev-
ery fish in a sample is known and the age of fish is estimated for
a subset of the sample. We then test the efficacy of applying our
size distribution model to empirically derive rates of growth and
mortality by generating populations with known values of larval
Figure 1. Schematic demonstrating the influence of bin size on the hatch rate, growth, mortality, and size-at-hatch. We compare the
slope of classic catch-curve. Data were generated with sample size skill of our model in estimating these values accurately, compared
varying from  to , and a classic catch-curve model was run to three other commonly used methods of mortality estimation
using the same data with -day and -day bin width sets. z̄ set at . (Nelson 2019a): the classic catch-curve (Ricker 1975); a weighted
for all model runs, represented by dotted black lines. Solid black line catch-curve (Maceina 1997; Maceina and Bettoli 1998), and the
represents a slope of  (i.e. -day model = -day model). Chapman–Robson method (Chapman and Robson 1960). We sys-
tematically test the bias of each of these models by adding variability
and the integral of mortality over growth. This equation has been to individual rates around mean values used in the data generating
widely used in fish and plankton research to numerically describe population model.
trophic and disturbance dynamics (Ebert et al., 1993; Smith et al.,
1998; Law et al., 2009; Datta et al., 2011; Heneghan et al., 2016; Blan-
chard et al., 2017; Law and Plank 2018), and to compare/estimate Methodology
rates of recruitment, mortality, and growth from changes in plant Our goal is to estimate the mean mortality (z̄) experienced by fish
density with size within forest plots (Coomes et al., 2003; Kohyama caught in a one-off sample (i.e. not tracked over time). All the meth-
et al., 2003; Muller-Landau et al., 2006). As the theory is not taxon ods applied here achieve this by comparing the abundances of fish
specific, similar methods can be applied to empirical fisheries, par- at different size and/or ages, assuming that the temporal pattern in

Figure 2. Schematic demonstrating size-structured population dynamics metrics of a larval fish cohort. The x-axis shows the size s of individual
fish, which can be described in either terms of length (mm) or weight (mg). Growth ḡ drives individuals across the x-axis from left to right, from
their size-at-hatch s towards their size-at-recruitment into the fishery. The y-axis describes n, the abundance of individuals per unit size per
unit volume (e.g., abundance mm- m ). Mortality z̄ determines the decrease on the y-axis from no (the n of individuals at s ), which is itself
determined by the larval hatch rate R.
A novel approach for estimating growth and mortality of fish larvae 

Table 1. Notation used for models, data generation, and output from analyses of the catch-curve, weighted catch-curve, Chapman-Robson
method, and all three variants of the size distribution models.

Parameter Description Units

n Abundance as a continuous measure from a Individuals per unit size or age per volume
density function, with respect to size or age
N Abundance with respect to a size- or age-class. Individuals per unit size or age per volume
a The age of each fish  day
s Size of an individual Standard length (mm)
s0 Size-at-hatch for an individual. Standard length (mm)
s0 Sample or population mean size-at-hatch. Standard length (mm)
g Instantaneous growth rate. mm day−
ḡ Sample or population mean growth rate. mm day−
day−

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


z Instantaneous mortality rate.
z̄ Sample or population mean mortality rate. day−
p The probability an individual will survive to the –
specified day. Used in generating sample
populations.
q The age of each fully recruited age-bin (i.e. fully days
vulnerable to sampling gear), defined as each
age-bin in a sample that is equal in age or older
than the most abundant age-bin (i.e. the modal
age). Used in the Chapman-Robson estimator.
u The oldest age-bin in a sample, used in the
Chapman-Robson estimator.
R The rate at which new individuals arrive in a Individuals day− m−
sample/population (i.e. the larval hatch-rate or
recruitment rate).

spawning is constant over the time interval from the smallest to in the “Estimators” section. The first scenario is entirely age-based,
largest sizes. Note that while length is the most common metric of while scenarios 2 and 3 use a size-based approach in combination
size used in larval fish studies, we use the more generic notation with a growth rate estimated from the literature or a subset of age-
of “size” because weight measurements can also be applied using at-size data.
the same models. Weight is often used when assessing the “recruit-
ment potential” of a cohort (Houde 1997; Sassa et al., 2014; Sassa
and Takahashi 2018). Scenario 1—age is known
The following notation is used (Table 1). For an individual, we The first scenario is where the age of every individual larval fish
let a be age, s be size, s0 be size-at-hatch, g be instantaneous growth is known. Most commonly, ages are obtained through otolith mi-
rate, and z be instantaneous mortality rate. Within a population, let crostructure analysis of daily growth rings (Stevenson and Cam-
R be the larval hatch rate, z̄ be the mean mortality rate, ḡ be the pana 1992; Quiñonez-Velázquez et al., 2000).
mean growth rate, and n(a) be the abundance of fish with age a.
Note n(a) is a continuous measure with respect to a, and so gives
the number of fish per unit age per unit volume (or area) of water. Scenario 2—age is estimated from a published relationship
Several of the methods require estimates of total abundance within In scenario 2, ages are estimated as a function of size using pub-
an age class, which we indicate by Na . lished or estimated growth equations from the literature (Houde
Assuming larval hatch rates are relatively constant with time, fish 2002; Pepin 2015b). In this case, the size of every individual in a
abundances naturally decline with time, according to the classic sur- sample is measured, and a species-specific growth rate ḡ and size-
vival equation at-hatch s0 (estimated from the intercept of a published growth rate
model) are used to calculate ages. For the case of linear growth rates,
n(a) = n0 exp (− z a) . (1) age is estimated as
Equation (1) works when abundance is measured as a continu- s − s0
a = . (2)
ous density function (i.e. n(a)), or as counts in age classes (Na ) and ḡ
underpins all of the methods applied.

Scenario 3—age is known for a subset of fish


Data handling scenarios Scenario 3 is perhaps the most common when trying to estimate
In fisheries research, there are a variety of data handling procedures growth and mortality. When obtaining age information for an entire
that may be used when trying to determine stock or population dy- sample is too costly, the size of every individual in a sample is mea-
namic rates. Here, we consider three data handling scenarios that sured and a subset of the sample is processed to obtain ages through
are used to estimate mean rates of instantaneous mortality from a otolith microstructure analysis (Coggins Jr et al., 2013). The rela-
sample of either adult or larval fish, using the methods described tionship between age and size from the subset of the sample is then
 C. Hinchliffe et al.

used to determine a mean growth rate, and this mean growth rate Size distribution models
is then used to estimate ages of the remaining sample via (2) (Sassa Our model involves two innovations. First, we use the theory of age-
et al., 2014; Sassa and Takahashi 2018). and size-structured populations to derive exact predictions of n, for
a given combination of growth and mortality functions. In this anal-
ysis, we restrict ourselves to the simplest possible functions: linear
Estimators growth (2) and constant mean mortality. This is so that our results
To assess the efficacy of our size distribution models, we compare are directly comparable with the existing methods outlined above,
them to three commonly used methods for estimating z̄. These in- where these assumptions also apply. In other words, our focus is on
clude the classic catch-curve based on a simple linear regression improving the statistical fit of the model and not on changing as-
(Ricker 1975), a weighted catch-curve based on a weighted linear sumptions about the dynamics of the system. We therefore make
regression (Maceina and Bettoli 1998), and the Chapman–Robson the same assumptions as the catch-curve with the aim of demon-
method (Chapman and Robson 1960). strating that our size distribution model provides more accurate and
precise statistical estimates of the underlying parameters. However,

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


it is worth noting that more complex functions can be used with
Simple catch-curve the new approach, which we address in the "Discussion" section.
The basic catch-curve is based on a loglinear relationship between Second, we use an improved statistical technique to estimate the
the number of fish caught in age classes (Na ) and their age. Log parameters of the models (including z̄) without having to group in-
transformation of (1) gives dividuals into larger bins.
There are three variants of our model, one for each of the dif-
log (Na ) = log (N0 ) − z̄ a. (3) ferent data handling scenarios. For scenario 1, the model is simply
that shown in (1). The other two models are derived from the gen-
eral McKendrick-von Foerster partial differential equation (PDE)
Weighted catch-curves for size structured population dynamics (Von Foerster 1959; Sinko
The weighted catch-curve uses a two-stage process, where a simple and Streifer 1967). Further, n is written as a function of size s, not
catch-curve (3) is first used to estimate log abundance-at-age. Us- age (i.e. we consider how the abundance of larvae changes with size
ing the same data, a weighted linear regression is then fit, using the rather than age). Assuming constant larval hatch rate, the PDE can
fitted values of log abundance-at-age from the simple catch-curve be solved to give an equation for the density individuals of size s as
as weights. Mortality is estimated from the slope coefficient of this a(s)
δg (s (a ) , E)


second weighted regression (Nelson 2019a). This technique has the n(s) = n(s0 ) exp − ∫ + z s a , E da ,
effect of assigning smaller weights to older, rarer fish reducing their 0 δs
contribution to the regression sum of squares (and has the oppo- (6)
site effect for younger, more abundant fish) (Maceina 1997; Maceina
and Bettoli 1998) such that where g(s, E) and z(s, E) are the growth and mortality rates of in-
dividuals of size s in environment E, n(s0 ) is the initial density of
log (Na ) = log (N0 ) − z̄ · b1 (a) , (4) recruits, s(a ) is the size of individuals age a , and a(s) is the age of
individuals size s. By linking the flux of individuals across the lower
where b1 is the partial regression coefficient estimated from the pre- bound, s0 of the size distribution to the rate R at which new individ-
dicted natural log of abundance-at-age using (3). uals arrive (i.e. larval hatch or recruitment rate), we can substitute
R
n(s0 ) = . (7)
Chapman–Robson method g(s0 )
The Chapman–Robson method is commonly applied to adult Note that the age distribution in (1) can be derived as a special
stocks to estimate annual mortality rates, and has recently been case of the more general distribution in (6), where size is measured
shown to be robust to sampling biases driven by gear-type (Nelson as age a (i.e. s = a), such that g = 1 and δg δs
= 0. This serves as
2019a). This method models age as a geometric distribution, giving the basis for our size distribution model used for scenario 1.
(Chapman and Robson 1960; Nelson 2019a) For scenarios 2 and 3, we can reorganize (6) so that the integral
⎛ u ⎞ is expressed in terms of size, not age. Following equation (22) of de
q=0 qNq
Roos (1997), we get
z̄ = log ⎝   ⎠
u s z (s , E)
N+ q=0 qNq − 1
R 
n(s) = exp − ∫ ds . (8)
s0 g (s , E)
g (s, E) 
⎛ ⎞
(N − 1) (N − 2)
−⎝       ⎠, Substituting in functions for z and g then leads to exact equa-
N N+ u
q = 0 qNq − 1
u
q = 0 qNq +1 tions for n(s). Assuming z(s, E) = z̄ and g (s, E) = ḡ leads to a
predicted size distribution (Muller-Landau et al., 2006)
(5)
R (t ) z̄
n(s) = exp − (s − s0 ) . (9)
where q is the age of each fully recruited age-bin (i.e. fish that are ḡ ḡ
considered to be fully vulnerable to sampling gear), which is defined In scenario 2, ḡ is specified based on available literature values.
as each age-bin in a sample that is equal in age or older than the For scenario 3, ḡ is estimated alongside z̄, by fitting the growth
most abundant age-bin (i.e. the modal age), and u is the oldest age- model (2)
bin. The ages of the fully recruited fish are standardized to begin at
q = 0. s = s0 + ḡa (10)
A novel approach for estimating growth and mortality of fish larvae 

using a subset of the sample with available age a and size s. This oc- ures as worked examples are available at https://github.com/charlie
curs simultaneously alongside fitting of the size distribution model hinchliffe/models_for_larval_fish_growth_and_mortality.
(using the size s of all individuals in the population), which esti- The catch-curve, weighted catch-curve, and Chapman–Robson
mates ratios of Rḡ and z̄ḡ . With ḡ constrained by the growth model estimates of z̄ were obtained using the "agesurv" function from
this leads to an estimate of R and z̄. The advantage of fitting both the "fishmethods" package (Nelson 2019b), applying the "lr", "wlr",
models simultaneously is that uncertainty in the growth rate is and "crcb" methods, respectively. Age-bins used in these models
preserved. were set at 1-day widths. Size distribution models were built in the
To fit (1) and (9) to data, we use the following likelihood model. Bayesian inference language Stan and implemented in R using the
First, we count the number of individuals recorded at each size package "rstan" (Stan Development Team 2019). Model param-
along the measurement scale, binning only at the scale of measure- eters were estimated using Markov chain Monte Carlo (MCMC)
ment resolution, i.e. Nsi , where si for i = 1, . . . , n are the size in- sampling, with three chains and 999 iterations. For each variant
tervals. We limited the precision of size measurements to 0.1 mm, of the data generation parameters described, models were run 500
so individuals are counted in 0.1 mm bins. Note that while we have times. For all models, age resolution was set at 1-day integers, and

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


selected 0.1 mm as our measurement resolution, the model imple- size resolution was set at 0.1 mm, to reflect appropriate finest scale
mentation is identical for data which may have a coarser resolution measurements made on fish larvae under laboratory conditions
(e.g. 0.5–1 mm). With small sample sizes, most values of Nsi are (Mullaney and Suthers 2013). Prior values for z̄, ḡ , and R were
zero. The expected number of individuals in each interval is ob- drawn from a Cauchy distribution with mean of 10 and wide
tained by integrating the density function from si−1 to si . For sce- variance, truncated to positive values.
narios 2 and 3, integrating (9) gives the number of individuals in To observe the influence of variability in each parameter on
each size class Nsi as model efficacy, we simulated populations such that ideal conditions
were evident in all parameters except for the parameter being inves-
R z̄ z̄
Nsi = − exp − (si − s0 ) − exp − (si−1 − s0 ) . tigated (i.e. large sample size and no variability in parameters). For
z̄ ḡ ḡ each simulation, R was set at 500 (when not being investigated), s0 at
(11) 3 mm, ḡ at 0.21, and z̄ at 0.25. Models that returned estimates closer
to z̄ = 0.25 and ḡ = 0.21, with less variability in estimates, were
These estimates are then compared to the observed counts using a considered more skilful than those that gave either biased or more
Poisson probability distribution variable estimates. For ease of interpretation, we present relative z̄

values here (z̄r ), where z̄r is the percentage difference between the
Nsi ,obs ∼ Poisson Nsi .
estimated z̄ and the true z̄ used to simulate the population (i.e. an
In scenario 3, ḡ is simultaneously constrained by comparing the estimated z̄ of 0.25 is equivalent to a z̄r of 0%, an estimated z̄ of 0.2
predicted sizes (from 10) to observed sizes, using a Gaussian prob- is equivalent to a z̄r of –20%, and an estimated z̄ of 0.3 is equivalent
ability distribution. to a z̄r of 20%). The same applies to growth estimates. Performance
A similar approach is used to fit data to scenario 1, but using age- is greatest when mean z̄r and ḡr estimates are closer to 0% and the
bins integrating (1) gives standard deviation in z̄r and ḡr is smaller.
R
To investigate the influence of R, we simulated populations with
Nai = − exp (−z̄ ai ) − exp (−z̄ ai−1 ) . (12) R equal to 5, 10, 20, 50, 100, and 500. To investigate the influence

of variability around log10(s0 ), log10(ḡ), and log10(z̄), the standard
deviation (σs0 , σg , and σz , respectively) was set at 0, 0.005, 0.01, 0.02,
Data generation 0.03, 0.05, and 0.1. These standard deviations were set across this
To test the efficiency of the different models, we generated data range such that the level of variance that bias was introduced into
with known values for z̄, ḡ , and R, and assessed how well the four model estimates was able to be observed for each parameter.
models could recover the known values. We used a basic popula- To observe the effect of applying a theoretical species-specific
tion model that allows for daily larval hatch (or recruitment), in- growth rate from published literature that differs from the “real”
dividual growth, and survival to vary around specified mean val- growth rate of a sampled population in scenario 2, we increased the
ues. The daily number of recruits was randomly generated from a bias between the theoretical literature g in the models and ḡ used to
Poisson model with rate R. These individuals were then grown to simulate data by 0% (i.e. 0.21), 1, 3, 5, 10, 25, and 50% (i.e. 0.315).
time of sampling according to (10). At this point, a random num- To understand how the number of individuals that are selected for
ber of individuals were allowed to survive, with probabilities p = aging in scenario 3 affects estimates of ḡ and z̄, we set coarse size
exp(−z̄ a). Each individual was initiated with values for z, g, and bins of 1.5 mm and then randomly selected 1, 3, 5, 10, 20, 40, and
s0 drawn from normal or log-normal distributions with a speci- 80 individuals (Ngrowth ) from each size bin to be used in the growth
fied mean and standard deviation. This data generation method estimation models. We consistently applied variability of σ = 0.3
assumes that individual-level mortality rates and individual-level to individual ages, and σ = 1 for individual sizes within growth
growth rates are serially correlated through time—assumptions models, while maintaining ideal conditions for each generated
that are also implicit in the common catch-curve methods investi- population.
gated here (Chapman and Robson 1960; Ricker 1975; Maceina and While running the analyses presented here, we observed that
Bettoli 1998). generating populations with even low levels of variance in size-at-
hatch (i.e. σs0 > 0) caused model estimates of z̄ to quickly diverge
from the true value of z̄ used to generate populations (see the "Sce-
Simulations nario 2" section under "Results"). We therefore added a function
All statistical analyses were performed using R version 3.6.1 (R Core into our data generation methods to subset data by removing all
Team 2019). Source code for reproducing all the analyses and fig- individuals smaller than the modal size class when specified. The
 C. Hinchliffe et al.

(a) R
5 10 20 50 100 200 500
200

100
zr
0

−100

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


−200

(b) σz
0 0.005 0.01 0.02 0.03 0.05 0.1
20
zr

−20

Figure 3. Distribution of z̄r estimates from scenario 1. Horizontal red lines at z̄r = % represent the “true” . z̄ used to generate populations. In
each cell from left to right, violins represent z̄r estimates from the catch-curve, weighted catch-curve, Chapman–Robson, and size distribution
models, respectively. Horizontal lines within violins represent , , and % quantiles. (a) Larval hatch rate R per day. (b) Values show σ of
individual z around log(z̄).

modal size class is defined as the most abundant 0.1 mm size bin Scenario 
in the population (i.e. where N is at its maximum across the dis- As expected, the general trend across all four models demonstrated
tribution of sizes). This function was used where variability in s0 more precise and accurate estimates of z̄ as R increases under data
created size structured populations in which the modal size class handling scenario 1 (Figure 3a). The size distribution model re-
was not the smallest size class in the population (i.e. 3–3.1 mm), turns better estimates of z̄ at low levels of R (and sample size),
with the aim of returning more accurate z̄ estimates (i.e. closer to with z̄r = –2.37% (±32.52% standard deviation, presented as brack-
z̄ = 0.25). eted ±values hereon) when R = 5 (sample size ≈ 25), a pattern con-
sistent across all levels of R tested here (Figure 3). The weighted
catch-curve performed the next best with z̄r = –7.58% (±66.69%),
followed by the classic catch-curve and Chapman–Robson mod-
Results els [z̄r = –51.48% (±31.72%) and 55.29% (±64.09%), respectively].
Under ideal conditions, all models across the three data han- The classic catch-curve method underestimated z̄ at all levels of R
dling scenarios perform well, returning accurate and precise tested here except the highest (R = 500), with model z̄r estimates
estimates of ḡ and z̄. The bias and standard deviation of all the being more than one standard deviation away from z̄r = 0%.
size distribution model z̄r estimates ranged from –71.57 to 20.59% All four models were robust to variability in individual z
and ±0.01 to ±32.52%, respectively. The Chapman–Robson (Figure 3b). All models returned z̄r estimates within one standard
method had a range of z̄r estimates from –65.06 to 56.45% and a deviation from 0%, up until σz was set at 0.05 around log10(z̄)
range in standard deviations from ±1.40 to ±64.32%. The catch- (equivalent to σ of ±12% around z̄). Above σz set at 0.05, bias in
curve had a range of z̄r estimates from –58.57 to 10.46% with a range estimates of z̄ started to increase.
in standard deviations from ±0.01 to ±36.65%, while the weighted
catch-curve had a range of z̄r estimates from –73.42 to 19.96% and a
range in standard deviations from ±2.14 to ±71.43%. The estimates Scenario 
and standard deviation from every configuration of the models run Results from data handling scenario 2 were similar to scenario 1,
here can be found in Supplementary Table 1. with all four models giving more precise and accurate estimates of
A novel approach for estimating growth and mortality of fish larvae 

z̄ as R increases, and a high degree of robustness to variability in in- to the level of σs0 = 0.02 (as opposed to 0.005 in scenario 2). The
dividual z (Figure 4a and b). The size distribution model returned catch-curve and weighted catch-curve returned mean z̄r estimates
better estimates of z̄ at low levels of R (and sample size), with z̄r = – at the highest level of σs0 = 0.1 within one standard deviation of 0%
0.04% (±30.91%) and –2.42% (±18.71%) when R = 5 and 10 (sam- (z̄r = 3.31% (±8.29%) and –5.31% (±7.43%), respectively), while
ple size ≈ 25 and 50), respectively. The weighted catch-curve per- the Chapman–Robson and size distribution model estimates were
formed the next best with z̄r = –7.63% (±65.63%) at R = 5 and z̄r more than one standard deviation away from 0% when σs0 = 0.03
= –11.52% (±26.96%) at R = 10. The classic catch-curve method, (z̄r = –7.58% (±3.1%) and –9.2% (±5.73%), respectively). The use
again, underestimated z̄ at all but the highest levels of R tested, with of a modal subsetting function (see the "Methods" section) im-
z̄r estimates being more than one standard deviation away from proved z̄ estimates across all models and all levels of σs0 (Figure 5d),
0%. All models returned z̄r estimates with low variability that were with all models at all levels of σs0 returning mean z̄r estimates within
within one standard deviation from 0% when variability in individ- one standard deviation from 0% except the catch-curve where z̄r =
ual z was low, up to σz set at 0.05 (equivalent to σ of ±12% of z̄). 10.46% (±10.17%) when σs0 = 0.1.
Variability in individual s0 was problematic for all models, even All models were relatively robust to individual variability in g,

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


at low levels of σs0 (Figure 4c). With σs0 around log10(s0 ) set at with all mean z̄r estimates within one standard deviation from 0%
0.005 (equivalent to σ of just ±1% of s0 ), all models underestimated (Figure 5e) and all ḡr estimates within one standard deviation of 0%
z̄r by between 2.3 [size distribution model -11.49% (±4.93%)] (Figure 6a) up to level σg = 0.03. As σg around log10(ḡ) increases
and 13.6 [Chapman–Robson –19.1% (±1.4%)] standard deviations above 0.03, mean z̄r estimates begin to deviate above 0% increas-
away from 0%. The bias from z̄r = 0% increased further for all ing to 5.59% (±8.62%), 19.96% (±9.98%), 22.3% (±9.72%), and
models as σs0 increased. The use of modal subsetting function im- 20.59% (±7.61%) at σg = 0.1 for the catch-curve, weighted catch-
proved z̄r across all models and all levels of σs0 (Figure 4d), with curve, Chapman–Robson, and size distribution models, respec-
every model returning a z̄r within one standard deviation from 0% tively. Model ḡr estimates across the same σg range demonstrated
up to σs0 = 0.03, equivalent to σ of ±7% of s0 [weighted catch- a similar pattern, increasing above 0% when σg is greater than 0.03,
curve z̄r = 0.04% (±3.37%), size distribution z̄r = 0.33% (±5.93%), with variability in ḡr estimates from the linear growth model used
catch-curve z̄r = 0.7% (±4.56%), Chapman–Robson z̄r = 1.22% in the catch-curve, weighted catch-curve, and Chapman–Robson
(±2.77%)]. At σs0 = 0.05 (equivalent to σ of ±12% of s0 ), the classic methods (i.e. length-at-age regression) being greater than the size
catch-curve, weighted catch-curve, and the size distribution mod- distribution model ḡr estimates. The ḡr estimate from the linear
els had mean z̄r within one standard deviation from 0% [z̄r = – growth model was equal to 7.37% (±5.9%) at σ = 0.05 and 25.11%
2.62% (±4.3%), –2.48% (±4.04%) and –4.83% (±7.01%), respec- (±9.72%) at σ = 0.1, while size distribution model ḡr = 6.18%
tively) while the Chapman–Robson methods did not (z̄r = –5.75% (±3.47%) and 22.62% (±6.91%) at the same levels of σ .
(±3.9%)]. As the number of individuals used in the estimation ḡ increased,
When the difference between the literature g rate applied in sce- the variability in z̄ estimates from all models decreased as expected
nario 2 and ḡ used to generate the population increased, a fast de- (Figure 5f). The size distribution model had the least variable z̄r
cline in the ability of all models to return accurate estimates of z̄ estimate at lower levels of Ngrowth selected from each rough size bin,
was observed (Figure 4e). At literature g set at 5% above ḡ (i.e. lit- with z̄r = 4.26% (±18.3%), 2.02% (±11.51%), 2.05% (±9.58%), and
erature g = 0.2205), all z̄r estimates were more than one standard 2.24% (±7.13%) at levels Ngrowth = 1, 3, 5 and 10, respectively. The
deviation from 0% [catch-curve z̄r = –5.07% (±3.91%), weighted Chapman–Robson model had the next least variable estimates of z̄r
catch-curve z̄r = –3.64% (±3.2%), Chapman–Robson z̄r = –3.56% = 0.62% (±18.3%), –0.19% (±18.74%), –0.19% (±14.67%), and –
(±2.09%), size distribution z̄r = –5.43% (±4.32%)]. When litera- 1.29% (±10.16%) across the same levels. Estimates of ḡ from the
ture g is set 50% above ḡ, z̄r decreased by more than 30%. classic linear growth model and size distribution model had the
same trend of decreasing variability with increasing Ngrowth . The size
distribution model returned less variable estimates of ḡ than classic
Scenario  growth models, with size distribution model ḡr = 5.12% (±17.78%),
Data handling scenario 3 results also showed all four models giv- 2.26% (±10.66%), 2.13% (±8.4%), and 2.44 (±5.58%) at levels
ing more precise and accurate estimates of z̄ as R increases, and a Ngrowth = 1, 3, 5, and 10, respectively. The linear growth models re-
high degree of robustness to variability in individual z (Figure 5a turned ḡr = –1.2 (±34.28%), –1.7% (±18.98%), –1.55 (±14.74%),
and b). The size distribution model returned better estimates of z̄ and –2.67% (±10.18%) across the same levels of Ngrowth .
at low levels of R (and sample size) than other models, with z̄r =
1.08% (±29.04%) and –3.32% (±19.45%) when R = 5 and 10 (sam-
ple size ≈ 25 and 50), respectively. The weighted catch-curve per-
formed the next best with z̄r = –5.38% (±71.43%) at R = 5 and Discussion
z̄r = 14.42% (±26.87%) at R = 10. The classic catch-curve method Our results highlight the value of the novel statistical techniques
again under predicted z̄ at all levels of R tested here except the high- and model structure we present for estimating rates of growth
est (R = 500), with the mean z̄r estimates being more than one stan- and mortality in snapshot samples of fish larvae. In all three data
dard deviation away from 0%. All models returned z̄r estimates with handling scenarios, the size-based modelling approaches we assess
low variability that were within one standard deviation from 0% here were more accurate at low sample size than the three com-
when variability in individual z was low, up to and above σz around mon methods we compared them with. This highlights the ad-
log10(z̄) set at 0.03 (equivalent to σ of ±7% of z̄). vantage of removing the need to bin data and modelling abun-
Individual variability in s0 caused z̄r estimates to deviate from 0% dance as continuous with respect to age or size. As species-specific
for all models as σs0 increased, though scenario 3 methods appear larval abundances often will not exceed more than a few fish
to be more robust than scenario 2 methods (Figure 5c). All mean per 100 m3 or double digits for a single tow (Hinchliffe et al.,
model z̄r estimates were within one standard deviation from 0% up 2021), any improvements made in our ability to estimate growth
 C. Hinchliffe et al.

(a)

(b)

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


(c)

(d)

(e)

Figure 4. Distribution of z̄r estimates from scenario 2. Horizontal red lines at z̄r = % represent the “true” . z̄ used to generate populations. In
each cell from left to right, violins represent z̄r estimates from the catch-curve, weighted catch-curve, Chapman–Robson and size distribution
models, respectively. Horizontal lines within violins represent , , and % quantiles. (a) Larval hatch rate R per day. (b) Values show σ of
individual z around log(z̄). (c) Values show σ of individual s around log(s ). (d) Values show σ of individual s around log(s ) of
generate populations which have been subsetted to remove individuals smaller than modal size. (e) Values show percentage increase between
the literature applied g and ḡ used to generate populations.

and mortality rates at low sample size will be beneficial. This is With this in mind, these results suggest that use of the traditional
particularly critical if our goal is to resolve the effects of envi- catch-curve should likely be avoided in favour of another estima-
ronmental drivers of growth and mortality at meaningfully fine tor when sample sizes are limited, as it underestimated z̄ across all
scales (Pepin 2004), such that we could predict cohort success for three data handling scenarios in populations with less than ∼400
oceanographic scales that characterize a range of spawning events. individuals.
A novel approach for estimating growth and mortality of fish larvae 

(a)

(b)

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


(c)

(d)

(e)

(f)

Figure 5. Distribution of z̄r estimates from scenario 3. Horizontal red lines at z̄r = % represent the “true” . z used to generate populations. In
each cell from left to right, violins represent z̄r estimates from the catch-curve, weighted catch-curve, Chapman–Robson and size distribution
models, respectively. Horizontal lines within violins represent , , and % quantiles. (a) Larval hatch rate R per day. (b) Values show σ of
individual z around log(z̄). (c) Values show σ of individual s around log(s̄0 ). (d) Values show σ of individual s around log(s̄0 ) of
generating populations that have been subset to remove individuals smaller than model size. (e) Values show σ of individual g around log(ḡ).
(f) Values show number of randomly selected individuals per . mm bin-width (Ngrowth ) used in the ḡ model.
 C. Hinchliffe et al.

(a)

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


(b)

Figure 6. Distribution of ḡr estimate from scenario 3. Horizontal red lines at ḡr = % represent the “true” . ḡ used to generate populations. In
each cell, the violins on the left represent the ḡr estimates from the length-at-age regression used in the catch-curve, weighted catch-curve, and
Chapman–Robson methods. The violins on the right-hand side of each cell represent the size distribution model ḡr estimates. Horizontal lines
within violins represent , , and % quantiles. (a) Values show σ of individual g around log(ḡ). (b) Values show number of randomly
selected individuals per . mm bin-width (Ngrowth ) used in the ḡ model.

Scenario  While increasing variability in individual probability of mortal-


In a scenario where the ages of every fish in a sample have been de- ity reduced the accuracy of mortality estimates, as a result of bias
termined, modelling abundance as a continuous density function caused by the persistence of an increased number of individuals
(as per (1) or (8) using the statistical techniques we have presented with very low z, the level of bias appears to be equal across the four
here, should provide the most reliable estimate of mean cohort mor- models tested here. Although previous studies have utilised size-
tality. As all models perform adequately at large samples sizes (e.g. or growth-selective losses to provide a measure of relative individ-
∼2000 fish), but at lower sample sizes, the continuous model per- ual mortality based on variability in growth (Takasuka et al., 2004,
forms markedly better (up to ∼200 fish), it is appropriate to rec- 2017; Fontes et al., 2011), it is difficult, or impossible, to transform
ommend its use over the other models we tested, provided the as- those inferences into true estimates of individual mortality risk in
sumptions regarding the underlying biology used to generate popu- a natural system. This implies that while we can infer rough mea-
lations in this study reflect the system being analysed. For instance, sures of differential survival from growth rates (i.e. fast growers
the model we suggest had improved mean relative mortality esti- have higher survival), this does not precisely inform us about the
mates around 5, 9, and 2% compared to the next best model (and variability of the individual probabilities of survival within groups
52, 41, and 30% compared to the worst performing model) when of larvae experiencing the same environmental conditions. Fur-
samples sizes were relatively small (i.e. 20, 40, and 80, respectively). thermore, this information cannot be discerned from variability in
As these sample sizes are within the typical range of a single plank- mean mortality rates among cohorts. If σ in individual growth rates
ton tow, these improvements in accuracy would likely make a sig- of a recently spawned cohort accurately reflect the variability in in-
nificant impact on our effort to link environmental drivers of mor- dividual mortality risk (Takasuka et al., 2004), it may be unlikely
tality at fine scales by helping to reduce noise in a highly stochastic that the σ in individual mortality risk would exceed the >10% of the
system. cohort mean z̄ which is necessary to make the model estimates un-
A novel approach for estimating growth and mortality of fish larvae 

reliable (Figure 3b). Researchers using mortality estimators should to vary between larval cohorts, both temporally and spatially, by
be aware of the substantial effect that high variance in individual an order of magnitude bigger than the 5% difference (Houde 1997)
mortality risk could have on reliability of their estimates, while we here needed to return z̄ estimates from all four models that were
remain limited in our ability to estimate this parameter in natural more than a standard deviation away from the true z̄. For exam-
populations. ple, Sassa et al. (2014) show that the cohort mean growth rate (mm
day–1 ) of larval jack mackerel (Trachurus japonicus) can vary over
60% between years as a result of environmental fluctuations in the
Scenario  southern East China Sea. Given this, the method in scenario 2 is
Observing the influence of variable R (and sample size) and individ- likely inappropriate for resolving mortality rates at fine scales and
ual z on modelling z̄ from a sample where the size of every fish is should not be used unless estimating a very crude mortality rate
known and a literature growth rate is applied demonstrated compa- based on a growth rate that is determined by averaging results from
rable advantages of our modelling approach to scenario 1, highlight- multiple studies that covers the environmental range of the species
ing the robustness of the size distribution method to sample size. At or population in focus.

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


samples sizes of around 20, 40, and 80 individuals, the size distri-
bution yielded more accurate mean relative mortality estimates of
around 8, 9, and 3%, respectively, compared to the next best model Scenario 
(and 56, 41, and 30% compared to the worst performing model). Estimating mortality using our size distribution approach when the
Our findings emphasize the importance of understanding two ages of a subset of fish are known, along with the size of every
aspects of a population before relying on model estimates of z̄ if this fish in a sample, demonstrated very similar results to scenario 2.
approach is used. First, variability in s0 can heavily and negatively Our model returned the best estimates of z̄ at lower R and was ro-
bias estimates of z̄ even at levels of σ of just ±1% of s0 (Figure 4c), bust to variation in individual g and z. We found that with sam-
which is well below what might be expected in a natural cohort. ple sizes around 20, 40, and 80 individuals, the improvement in
This is likely because variability in s0 will manifest as bias and un- mean relative mortality estimates compared to the next best model
certainty in both growth (i.e. the intercept of length-at-age regres- is around 4, 11, and 3%, respectively (and 55, 43, and 31% com-
sion) and mortality models by either incorrectly aging larvae (i.e. pared to the worst performing model). Furthermore, biases caused
error in growth model affecting catch-curve, weighted catch-curve, by variability in s0 were of lesser magnitude than in scenario 2 with
or Chapman–Robson estimates) or shifting the distribution of sizes mortality estimates remaining relatively accurate above levels of σ
so that smallest size interval in a sample is no longer the most abun- of ±3% of s0 , in the range of what might be expected of some nat-
dant (i.e. the size distribution model). The bias caused by variability ural populations (Bengtson et al., 1987; Chambers et al., 1988; Jor-
in s0 is largely driven by shifts in the intercept of the growth model daan et al., 2006). Use of subsetting to trim data at the modal size
and the smallest sizes of the size distribution model because small largely corrected remaining bias caused by variability in s0 across all
differences in s0 correspond to greater relative difference in ages for levels of σ .
younger individuals than older individuals. The bias caused by vari- In addition to improving z̄ estimates with fewer fish, we found
ability in s0 highlights a potentially under considered aspect of the that our approach has utility in estimating ḡ more effectively than a
biology of larvae in previous studies attempting to resolve mortality standard linear model. We found that not only was the size distribu-
rates using the common methods available. Larvae hatched under tion model more robust to variability in individual g, the standard
laboratory conditions commonly show levels of σ in individual s0 in deviation of ḡr estimates was between 43 and 48% less for the size
the range of ±3–10% of s0 (Bengtson et al., 1987; Llanos-Rivera and distribution model than the standard linear model used to estimate
Castro 2004; Jordaan et al., 2006; Garrido et al., 2015), which is well ages for the other mortality estimators. This trend occurred across
above the level of variability in s0 needed to bias mortality estimates all levels of N per size class used to select the subset of individuals
in scenario 2. Bias caused by variability in s0 can be largely corrected with age and size data available in the growth models (Figure 6).
by removing all individuals smaller than the modal size class for all This finding highlights the applicability of our approach not just
models (Figure 4d) such that reasonable estimates of z̄ are returned to studies interested in mortality or recruitment indices but also
within a range of σ in s0 that might be expected of a natural pop- to those solely concerned with estimating growth rates. As mod-
ulation (i.e. less than ±10% of s0 ). Subsetting data in this way is els exploring environmental effects on growth will likely have high
already a common approach for dealing with biases in the sampled variability in their residuals as a result of environmental stochas-
size and age distribution of larvae caused by the low catchability of ticity, reducing uncertainty in growth estimates as response vari-
small young larvae (i.e. extrusion) (Morse 1989; Sassa et al., 2014; ables in these models will improve our confidence in their final
Raya et al., 2020), so applying this approach can be justified here estimates.
and may have likely reduced the impact of high σs0 on mortality
estimates in previous studies.
Second, bias in estimates of z̄ caused by even small deviations Limitations
in the literature growth rate from ḡ used to generate data here un- In the present study, we systematically observe the influence of bias
derscore the risk of this data handling approach. This bias occurs caused by variability in parameters of population dynamics and
because the survival of larvae is linked to the mortality rate ex- found that variability in individual traits affect model estimates rel-
perienced over the time it takes to grow through a given length, atively equally across the four models tested here. However, due to
whereby a positively biased literature growth rate results in an esti- the number of possible populations produced from combinations
mated higher proportion of older individuals than actually exist in of parameters with different variability levels (here, we have four
the population and a decreased estimate of mortality. The inverse is parameters and tested seven levels of variability, 74 is 2401 possible
true for a negatively biased growth rate, which would result in a pos- population settings), we have not considered interactions between
itively biased mortality. Intraspecific mean growth rates are known variability in different parameters (i.e. interactions between varying
 C. Hinchliffe et al.

levels of σs0 , σg , and σz ). While this approach is common (Saltelli Future directions
et al., 2002; Smith et al., 2019), it is possible that we have missed A major advantage of the approach we suggest is that it allows for
potential susceptibilities or strengths of the models. For example, models to be easily adjusted to assess non-linear growth and mor-
it may be that as σs0 and σg increase, the biases they both cause on tality scenarios. Muller-Landau et al., (2006) demonstrated how
model estimates cancel each other out (or are additive) and this ef- growth and mortality rates that vary as power functions of size can
fect could be non-linear. be incorporated to predict corresponding size distributions. For ex-
This study also omits the use of age-length keys, a common ample, by deriving size-specific growth from non-linear size-at-age
approach to aging a population from a subsample. This is largely equations such as a Gompertz (Gompertz 1825) or von Bertalanffy
because they require a large, representative, simultaneously sam- growth equation (von Bertalanffy 1938), we can fit models to pop-
pled, aged “parental” group collected to avoid bias (Westrheim and ulations shaped from a wider range of biological conditions. A next
Ricker 1978; Coggins Jr et al., 2013; Ailloud and Hoenig 2019), logical step in this research would therefore be to develop a pack-
which is often unattainable at the tow or sample level, and as such age wherein multiple growth and mortality models could be fit to
has limited use in resolving rates at fine scales of interest. Future data, and compared, with the aim of retaining the best model based

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


developments of the size distributions models could be adapted to on relevant model selection criteria (Ward 2008). Achieving this
use age-length keys but given the results from scenario 1 where all would allow widespread application of the size-based modelling
ages are known, the continuous model we present (1) would again approaches presented here beyond larvae populations into adult
offer the best alternative. stocks, where growth is very unlikely to ever be linear (Mommsen
There are several limitations which we have not considered 2001).
which are related to the underlying biology of the systems being As well as improving estimates of ḡ and z̄, these models may add
modelled, and it should be noted that these are of concern for all to our understanding of the fine scale variability in cohort success
the models being tested here. In need of particular attention are the by relating cohort dynamics at the tow or sample level directly to
common assumptions of constant flux of individuals into the pop- in situ environmental parameters. The structure of (1) and (8) al-
ulation being sampled, and the limited loss of individuals owing lows for environmental effects to directly influence estimates, by
to the advection. As we have alluded to in the introduction, non- including additional submodels in a hierarchical framework (Wikle
random directional shifts in larval supply can propagate as bias in 2003; Gelman et al., 2015) that specify how ḡ and z̄ vary as a func-
z̄ estimates (Ricker 1975; Houde 2002). In attempting to apply the tion of commonly sampled parameters such as temperature, salin-
models presented to resolve larval ḡ and z̄ at finer scales, a good un- ity, chlorophyll a, or zooplankton biomass. This will be a valuable
derstanding of specific–specific spawning biology is required. Lim- next step in maximising the utility of these models.
iting the age and size range of individuals that mean mortality is An important step in verifying the applicability and utility of the
integrated over may alleviate bias introduced by shifts in spawning size distribution models we present here requires their application
(e.g. only include larvae up to 14 days), while understanding bias to field data. In approaching this, future studies should first assess
driven by advection requires a strong estimation of current fields variability in accuracy and precision of z̄ and ḡ estimates taken at
in the sampling area (Helbig and Pepin 1998). We have established the single tow, station (i.e. repeated tows deployed at the same lo-
that the size-based modelling approach is as appropriate as cur- cation) and entire survey level. In doing so, they ought to evaluate
rent methods for estimating mortality in circumstances where other how reasonable it may be to quantify vital rates at the finest scales
common methods are already applied (i.e. for survey/multiple sur- available for a given dataset. Appropriate scales likely vary between
vey level estimates) but further empirical field study is needed to species based on their biology (e.g. short or protracted spawning
establish the feasibility of applying the models to finer-scale data times, abundances) and among seascapes (e.g. levels of retention or
(see the "Future directions" section). aberrant drift, convergence-divergence eddy structure). The neces-
Other biological limitations relate to situations where growth sary precision of these estimates may also vary between projects and
and mortality rates are not constant. For example, mortality rates study goals. Where studies have lots of tows available, they may be
are likely to be non-constant through development resulting from less affected by higher error in estimates. For example, a study may
aspects of the growth–survival paradigm (Takasuka et al., 2017), wish to relate variability in z̄ and ḡ with environmental drivers using
while some species may exhibit non-constant growth through their the hierarchical framework mentioned above. If error in tow level
development as a consequence of changes in feeding ability and estimates (i.e. noise) is not biased in any particular direction, and
success (Pepin et al., 2015; Vasbinder and Ainsworth 2020). Addi- many tow level replicates are available, the bias may be averaged
tionally, problems may arise through the omission of density (or out and the direction and size of effects (i.e. signal) could still be
frequency) dependence in the models, whereby increased popu- estimated. One advantage of a hierarchical size distribution mod-
lation sizes reduce resource availability, thereby increasing intra- elling approach is that it allows for easy indexation of the models
specific competition, and reducing growth and survival of larvae such that a tow, station (an area within a survey where multiple
(Stige et al., 2019). This ultimately leads to non-constant popula- tows are taken) and entire survey level estimate of z̄ and ḡ can be
tion rates through time. Density-dependent effects on the early-life gained from a single modelling exercise. Estimating vital rates fol-
stages of fish have long been thought to impact recruitment (Cush- lowing a hierarchy of scales in such a way to address issues of bias
ing 1974; Shepherd and Cushing 1980; Stige et al., 2019) and failure would greatly benefit from the inclusion of a framework, like that
to account for this may cause bias across all the models presented suggested by Helbig and Pepin (1998), incorporating theoretical ex-
when tested on real populations. Again, these non-constant rates pressions evaluating the quality of spatial sampling based on physi-
will bias estimates from all the models here. However, the easily cal oceanographic processes. Already existing databases containing
modifiable size distribution model that we describe here presents size data comprised of extensive sampling across a seascape (e.g. the
a new opportunity to determine how rates of growth and mortal- California Cooperative Oceanic Fisheries Investigations (CalCOFI)
ity change throughout development and model them accordingly sampling of the California current since 1951; Bograd et al., 2003),
within a single framework. may provide the best avenue for this exploration.
A novel approach for estimating growth and mortality of fish larvae 

Conclusion Bannister, R. C. A., Harding, D., and Lockwood, S. J. 1974. Larval mor-
This study demonstrates the potential for available techniques that tality and subsequent year-class strength in the plaice (Pleuronectes
platessa L.). In: The Early Life History of Fish, pp. 21–37. Ed. by
remain novel in fisheries science to improve upon our ability to Blaxter, J. H. S. Springer, Berlin, Heidelberg.
empirically determine rates of population dynamics. Achieving ac- Bengtson, D. A., Barkman, R. C., and Berry, W. J. 1987. Relationships
curate measurements of vital rates that are fundamental to under- between maternal size, egg diameter, time of spawning season, tem-
standing population dynamics in highly variable ecosystems is es- perature, and length at hatch of Atlantic silverside, Menidia menidia.
sential if we hope to someday accurately predict success of the lar- Journal of Fish Biology 31: 697–704.
val stages from seascape conditions, and thereby reduce a great in- Beverton, R. J. H., and Holt, S. J. 1957. On the Dynamics of Exploited
Fish Populations. Ministry of Agriculture and Fisheries, London.
hibitor to predicting cohort success in fish. By adopting an approach Beyer, J. E. 1989. Recruitment stability and survival: simple size-specific
that models abundance as a continuous measure with respect to age theory with examples from the early life dynamics of marine fish.
or size, and removes the need to bin data, we have shown how the Dana 7: 45–147.
accuracy and precision of estimates can be enhanced. Furthermore, Blanchard, J. L., Heneghan, R. F., Everett, J. D., Trebilco, R., and Richard-
we have shown that substituting the underlying age distribution of son, A. J. 2017. From bacteria to whales: using functional size spec-

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


catch-curves for a size distribution can result in models that not tra to model marine ecosystems. Trends in Ecology & Evolution 32:
174–186.
only have the advantage of treating growth and mortality within a
Bograd, S. J., Checkley, D. A., and Wooster, W. S. 2003. CalCOFI: a half
single framework, thereby conserving uncertainty between the two century of physical, chemical, and biological research in the Cali-
rates, but also produces more accurate and precise estimates. While fornia Current System. Deep Sea Research Part II: Topical Studies
the necessary future work mentioned above will improve our con- in Oceanography 50: 2349–2353.
fidence in, and applicability of, these models, we encourage fish- Chambers, R. C., Leggett, W. C., and Brown, J. A. 1988. Variation in and
eries scientists to consider these methods as another tool to improve among early life history traits of laboratory-reared winter flounder
upon our collective understanding of larval success. Pseudopleuronectes americanus. Marine Ecology Progress Series 47:
1–15.
Chapman, D. G., and Robson, D. S. 1960. The analysis of a catch curve.
Biometrics 16: 354–368.
Data availability Coggins, L. G. Jr, Gwinn, D. C., and Allen, M. S. 2013. Evaluation of
The data underlying this article will be shared on reasonable request age–length key sample sizes required to estimate fish total mortal-
to the corresponding author. ity and growth. Transactions of the American Fisheries Society 142:
832–840.
Coomes, D. A., Duncan, R. P., Allen, R. B., and Truscott, J. 2003. Dis-
turbances prevent stem size-density distributions in natural forests
Author contributions from following scaling relationships. Ecology Letters 6: 980–989.
CH and DSF designed the analysis, CH performed the analysis and Cushing, D. H. 1974. The possible density-dependence of larval mor-
led the writing of the manuscript. All authors helped conceive ideas, tality and adult mortality in fishes. In: The Early Life History of Fish,
contributed critically to the drafts, and gave final approval for pub- pp. 103–111. Ed. by Blaxter, J. H. S. Springer, Berlin, Heidelberg.
Dannevig, H. C. 1907. On some peculiarities in our coastal winds and
lication.
their influence upon the abundance of fish in inshore waters. Journal
of the Royal Society of NSW 41: 27–15.
Datta, S., Delius, G. W., Law, R., and Plank, M. J. 2011. A stability anal-
Supplementary Data ysis of the power-law steady state of marine size spectra. Journal of
Supplementary material is available at the ICESJMS online version Mathematical Biology 63: 779–799.
of the manuscript. de Roos, A. M. 1997. A gentle introduction to physiologically struc-
tured population models. In: Structured-Population Models in Ma-
rine, Terrestrial, and Freshwater systems. pp. 119–204, Ed. by Tul-
japurkar, S. and Caswell, H. Springer, Boston, MA.
Acknowledgements Ebert, T. A., Schroeter, S. C., and Dixon, J. D. 1993. Inferring demo-
CH is funded by an Australian Government Research Training graphic processes from size-frequency distributions: effect of pulsed
Program Scholarship. DSF is funded by ARC Future fellowship recruitment on simple models. Fishery Bulletin 91: 237–243.
(FT160100113). We thank Jian Yen for useful discussions, and two Fontes, J., Santos, R. S., Afonso, P., and Caselle, J. E. 2011. Larval growth,
size, stage duration and recruitment success of a temperate reef fish.
anonymous reviewers for their helpful comments.
Journal of Sea Research 65: 1–7.
Fry, F. E. J. 1971. The effect of environmental factors on the physiology
of fish. Fish Physiology 1–98.
References Garrido, S., Ben-Hamadou, R., Santos, A.M.P., Ferreira, S., Teodósio,
Ailloud, L. E., and Hoenig, J. M. 2019. A general theory of age-length M.A., Cotano, U., Irigoien, X. et al. 2015. Born small, die young:
keys: combining the forward and inverse keys to estimate age com- intrinsic, size-selective mortality in marine larval fish. Scientific re-
position from incomplete data. ICES Journal of Marine Science 76: ports 5: 1–10.
1515–1523. Gelman, A., Lee, D., and Guo, J. 2015. Stan: a probabilistic program-
Anderson, J. T. 1988. A review of size dependent survival during pre- ming language for Bayesian inference and optimization. Journal of
recruit stages of fishes in relation to recruitment. Journal of North- Educational and Behavioral Statistics 40: 530–543.
west Atlantic Fishery Science 8: 55–66. Gompertz, B. 1825. XXIV. On the nature of the function expressive of
Bailey, K. M., Brodeur, R. D., and Hollowed, A. B. 1996. Cohort sur- the law of human mortality, and on a new mode of determining
vival patterns of walleye pollock, Theragra chalcogramma, in She- the value of life contingencies. In a letter to Francis Baily, Esq. F. R.
likof Strait, Alaska: a critical factor analysis. Fisheries Oceanogra- S. andc. Philosophical Transactions of the Royal Society of London
phy, 5: 179–188. 115: 513–583.
Bailey, K. M., and Houde, E. D. 1989. Predation on eggs and larvae of Govoni, J. J. 2005. Fisheries oceanography and the ecology of early life
marine fishes and the recruitment problem. Advances in marine bi- histories of fishes: a perspective over fifty years. Scientia Marina 69:
ology 25: 1–83. 125–137.
 C. Hinchliffe et al.

Hare, J. A. 2014. The future of fisheries oceanography lies in the pursuit Nelson, G. A. 2019b. fishmethods: Fishery science methods and models.
of multiple hypotheses. ICES Journal of Marine Science 71: 2343– R package version 1.11.
2356. Odum, H. T., and Odum, E. P. 1955. Trophic structure and productivity
Helbig, J. A., and Pepin, P. 1998. Partitioning the influence of physi- of a windward coral reef community on Eniwetok Atoll. Ecological
cal processes on the estimation of ichthyoplankton mortality rates. Monographs 25: 291–320.
I. Theory. Canadian Journal of Fisheries and Aquatic Sciences 55: Pepin, P. 1991. Effect of temperature and size on development, mortal-
2189–2205. ity, and survival rates of the pelagic early life history stages of marine
Heneghan, R. F., Everett, J. D., Blanchard, J. L., and Richardson, A. J. fish. Canadian Journal of Fisheries and Aquatic Sciences 48: 503–
2016. Zooplankton are not fish: improving zooplankton realism in 518.
size-spectrum models mediates energy transfer in food webs. Fron- Pepin, P. 2004. Early life history studies of prey-predator interactions:
tiers in Marine Science 3: 201. quantifying the stochastic individual responses to environmental
Hinchliffe, C., Smith, J. A., Everett, J. D., Falster, D. S., Lara-Lopez, variability. Canadian Journal of Fisheries and Aquatic Sciences 61:
A., Miskiewicz, A. G., Richardson, A. J. et al. 2021. Modelling the 659–671.
distribution of larval fish in a western boundary current using a Pepin, P. 2016. Death from near and far: alternate perspectives on size-
multi-voyage database. Reviews in Fish Biology and Fisheries 31: dependent mortality in larval fish. ICES Journal of Marine Science:

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


399–415. Journal du Conseil 73: 196–203.
Hjort, J. 1914. Fluctuations in the great fisheries of Northern Eu- Pepin, P. 2016. Reconsidering the impossible — linking environmen-
rope viewed in the light of biological research. Rapports et Procès- tal drivers to growth, mortality, and recruitment of fish. Canadian
Verbaux des Réunions, Conseil International pour l’Exploration de Journal of Fisheries and Aquatic Sciences 73: 205–215.
la Mer 20: 1–13. Pepin, P., Robert, D., Bouchard, C., Dower, J.F., Falardeau, M., Fortier,
Houde, E. D. 1997. Patterns and trends in larval-stage growth and mor- L., Jenkins, G.P. et al. 2014. Once upon a larva: revisiting the rela-
tality of teleost fish. Journal of Fish Biology 51: 52–83. tionship between feeding success and growth in fish larvae. ICES
Houde, E. D. 2002. Mortality. In: Fishery Science: The Unique Contri- Journal of Marine Science 72: 359–373.
butions of Early Life Stages, pp. 64–87. Ed. by Fuiman, L. A. and Polovina, J. J. 1984. Model of a coral reef ecosystem. Coral reefs 3: 1–11.
Werner, R. G. John Wiley & Sons, New York City, NY. Quiñonez-Velázquez, C., Nevarez-Martı´nez, M. O., and Gluyas-
Houde, ED. 1987. Fish early life dynamics and recruitment variability. Millán, M. G. 2000. Growth and hatching dates of juvenile pacific
American Fisheries Society Symposium Series 2: 17–29. sardine Sardinops caeruleus in the Gulf of California. Fisheries Re-
Jordaan, A., Hayhurst, S. E., and Kling, L. J. 2006. The influence of tem- search 48: 99–106.
perature on the stage at hatch of laboratory reared Gadus morhua R Core Team 2019. R: A Language and Environment for Statistical
and implications for comparisons of length and morphology. Jour- Computing. R Foundation for Statistical Computing, Vienna.
nal of Fish Biology 68: 7–24. Raya, V., Salat, J., and Sabatés, A. 2020. The box-balance model: a new
Kohyama, T., Suzuki, E., Partomihardjo, T., Yamada, T., and Kubo, T. tool to assess fish larval survival, applied to field data on two small
2003. Tree species differentiation in growth, recruitment and allom- pelagic fish. Marine Ecology Progress Series 650: 289–308.
etry in relation to maximum height in a Bornean mixed dipterocarp Ricker, W. E. 1975. Computation and interpretation of biological statis-
forest. Journal of Ecology 91: 797–806. tics of fish populations. Bulletin of the Fisheries Research Board of
Law, R., and Plank, M. J. 2018. Balanced harvesting could reduce Canada 191: 1–382.
fisheries-induced evolution. Fish and Fisheries 19: 1078–1091. Rose, G. A., and Rowe, S. 2020. Evidence of near year-around spawning
Law, R., Plank, M. J., James, A., and Blanchard, J. L. 2009. Size-spectra in Atlantic cod off southern Newfoundland: implications for man-
dynamics from stochastic predation and growth of individuals. agement. Canadian Journal of Fisheries and Aquatic Sciences 77:
Ecology 90: 802–811. 1969–1976.
Llanos-Rivera, A., and Castro, L. R. 2004. Latitudinal and seasonal egg- Rutherford, E. S., Houde, E. D., and Nyman, R. M. 1997. Relationship
size variation of the anchoveta (Engraulis ringens) off the Chilean of larval-stage growth and mortality to recruitment of striped bass,
coast. Fishery Bulletin 102: 207–212. Morone saxatilis, in Chesapeake Bay. Estuaries 20: 174–198.
Maceina, M. J. 1997. Simple application of using residuals from catch- Saltelli, A., Tarantola, S., Campolongo, F., and Ratto, M. 2002. Sensitivity
curve regressions to assess year-class strength in fish. Fisheries Re- analysis in Practice: A Guide to Assessing Scientific Models. John
search 32: 115–121. Wiley & Sons, New York City, NY.
Maceina, M. J., and Bettoli, P. W. 1998. Variation in largemouth bass Sassa, C., and Takahashi, M. 2018. Comparative larval growth and mor-
recruitment in four mainstream impoundments of the Tennessee tality of mesopelagic fishes and their predatory impact on zooplank-
River. North American Journal of Fisheries Management 18: 998– ton in the Kuroshio region. Deep Sea Research Part I: Oceano-
1003. graphic Research Papers 131: 121–132.
Mommsen, T. P. 2001. Paradigms of growth in fish. Comparative Bio- Sassa, C., Takahashi, M., Nishiuchi, K., and Tsukamoto, Y. 2014. Dis-
chemistry and Physiology Part B: Biochemistry and Molecular Bi- tribution, growth and mortality of larval jack mackerel Trachurus
ology 129: 207–219. japonicus in the southern East China Sea in relation to oceano-
Morse, W. W. 1989. Catchability, growth, and mortality of larval fishes. graphic conditions. Journal of Plankton Research 36: 542–556.
Fishery Bulletin 87: 417–446. Secor, D. H., and Houde, E. D. 1995. Temperature effects on the tim-
Mullaney, T. J., and Suthers, I. M. 2013. Entrainment and reten- ing of striped bass egg production, larval viability, and recruitment
tion of the coastal larval fish assemblage by a short-lived, subme- potential in the Patuxent River (Chesapeake Bay). Estuaries 18: 527–
soscale, frontal eddy of the East Australian Current. Limnology and 544.
Oceanography 58: 1546–1556. Sexton, S. C., Ward, T. M., Stewart, J., Swadling, K. M., and Huveneers,
Muller-Landau, H.C., Condit, R.S., Harms, K.E., Marks, C.O., Thomas, C. 2019. Spawning patterns provide further evidence for multiple
S.C., Bunyavejchewin, S., Chuyong, G. et al. 2006. Comparing trop- stocks of sardine (Sardinops sagax) off eastern Australia. Fisheries
ical forest tree size distributions with the predictions of metabolic Oceanography 28: 18–32.
ecology and equilibrium models. Ecology Letters 9: 589–602. Shepherd, J. G., and Cushing, D. H. 1980. A mechanism for
Neill, W. H., Miller, J. M., Van Der Veer, H. W., and Winemiller, K. density-dependent survival of larval fish as the basis of a stock-
O. 1994. Ecophysiology of marine fish recruitment: a conceptual recruitment relationship. ICES Journal of Marine Science 39:
framework for understanding interannual variability. Netherlands 160–167.
Journal of Sea Research 32: 135–152. Sinko, J. W., and Streifer, W. 1967. A new model for age-size structure
Nelson, G. A. 2019. Bias in common catch-curve methods applied to age of a population. Ecology 48: 910–918.
frequency data from fish surveys. ICES Journal of Marine Science Smith, B. D., Botsford, L. W., and Wing, S. R. 1998. Estimation of growth
76: 2090–2101. and mortality parameters from size frequency distributions lacking
A novel approach for estimating growth and mortality of fish larvae 

age patterns: the red sea urchin (Strongylocentrotus franciscanus) as Tian, T., Fiksen, Ø., and Folkvord, A. 2007. Estimating larval fish growth
an example. Canadian Journal of Fisheries and Aquatic Sciences 55: under size-dependent mortality: a numerical analysis of bias. Cana-
1236–1247. dian Journal of Fisheries and Aquatic Sciences 64: 554–562.
Smith, J. A., Eyre, B. D., Rosentreter, J. A., and Taylor, M. D. Vanalderweireldt, L., Winkler, G., Mingelbier, M., and Sirois, P. 2019.
2019. Modelling estuarine stocking density for crustaceans us- Early growth, mortality, and partial migration of striped bass (Mo-
ing net ecosystem metabolism. Bulletin of Marine Science 95: rone saxatilis) larvae and juveniles in the St. Lawrence estuary,
217–238. Canada. ICES Journal of Marine Science 76: 2235–2246.
Stan Development Team 2019. RStan: the R interface to Stan. R package Vasbinder, K., and Ainsworth, C. 2020. Early life history growth in fish
version 2.19.1. reflects consumption-mortality tradeoffs. Fisheries Research 227:
Stevenson, D. K., and Campana, S. E. 1992. Otolith Microstructure Ex- 105538.
amination and Analysis. Department of Fisheries and Oceans. Ot- von Bertalanffy, L. 1938. A quantitative theory of organic growth (in-
tawa, Canada. quiries on growth laws. II). Human biology 10: 181–213.
Stige, L.C., Rogers, L.A., Neuheime, A.B., Hunsicker, M.E., Yaragina, Von Foerster, H. 1959. Some remarks on changing populations. In: The
N.A., Ottersen, G., Ciannelli, L. et al. 2019. Density- and size- Kinetics of Cellular Proliferation, pp. 382–407. Ed. by Stohlman, F.
dependent mortality in fish early life stages. Fish and Fisheries 20: Grune and Stratton, New York, NY.

Downloaded from https://academic.oup.com/icesjms/article/78/8/2684/6364353 by guest on 03 March 2022


962–976. Ward, E. J. 2008. A review and comparison of four commonly used
Takasuka, A., Aoki, I., and Mitani, I. 2003. Evidence of growth-selective Bayesian and maximum likelihood model selection tools. Ecolog-
predation on larval Japanese anchovy Engraulis japonicus in Sagami ical Modelling 211: 1–10.
Bay. Marine Ecology Progress Series 252: 223–238. Watanabe, Y., Oozeki, Y., and Kitagawa, D. 1997. Larval parameters de-
Takasuka, A., Aoki, I., and Mitani, I. 2004. Three synergistic growth- termining preschooling juvenile production of Pacific saury (Colo-
related mechanisms in the short-term survival of larval Japanese an- labis saira) in the northwestern Pacific. Canadian Journal of Fish-
chovy Engraulis japonicus in Sagami Bay. Marine Ecology Progress eries and Aquatic Sciences 54: 1067–1076,
Series 270: 217–228. Westrheim, S. J., and Ricker, W. E. 1978. Bias in using an age–length
Takasuka, A., Sakai, A., and Aoki, I. 2017. Dynamics of growth-based key to estimate age-frequency distributions. Journal of the Fisheries
survival mechanisms in Japanese anchovy (Engraulis japonicus) lar- Research Board of Canada 35: 184–189.
vae. Canadian Journal of Fisheries and Aquatic Sciences 74: 812– Wikle, C. K. 2003. Hierarchical Bayesian models for predicting the
823. spread of ecological processes. Ecology 84: 1382–1394.

Handling Editor: David Secor

You might also like