Fisheries Research 269 (2024) 106860

Contents lists available at ScienceDirect

Fisheries Research
journal homepage: www.elsevier.com/locate/fishres

Catch per unit effort modelling for stock assessment: A summary of

good practices
Simon D. Hoyle a, b, c, *, Robert A. Campbell d, Nicholas D. Ducharme-Barth e, Arnaud Grüss f,
Bradley R. Moore a, James T. Thorson g, Laura Tremblay-Boyer h, Henning Winker i, Shijie Zhou j,
Mark N. Maunder k, l
a
National Institute of Water and Atmospheric Research Ltd (NIWA), 217 Akersten St, Port Nelson, Nelson 7010, New Zealand
b
Indian Ocean Tuna Commission, PO Box 1011, Victoria, Seychelles
c
Department of Statistics, University of Auckland, Private Bag 92019, Auckland, New Zealand
d
CSIRO Environment, Private bag No. 1, Aspendale, Victoria 3195, Australia
e
NOAA National Marine Fisheries Service, Pacific Islands Fisheries Science Center, 1845 Wasp Boulevard, Building 176, Honolulu 96818, HI, USA
f
National Institute of Water and Atmospheric Research Ltd (NIWA), 301 Evans Bay Parade, Greta Point, Wellington 6021, New Zealand
g
Habitat and Ecological Processes Research Program, Alaska Fisheries Science Center, National Marine Fisheries Service, NOAA, 7600 Sand Point Way N.E., Seattle, WA
98115, USA
h
CSIRO Environment, GPO Box 1538, Hobart, TAS 7001, Australia
i
Department of Aquatic Resources, Institute of Marine Research, Swedish University of Agricultural Sciences, 45330, Sweden
j
CSIRO Environment, 306 Carmody Road, St Lucia, Brisbane, QLD 4067, Australia
k
Inter-American Tropical Tuna Commission, 8901 La Jolla Shores Drive, La Jolla, CA 92037-1508, USA
l
Center for the Advancement of Population Assessment Methodology, Scripps Institution of, Oceanography, La Jolla, CA, USA

A R T I C L E I N F O A B S T R A C T

Handled by A.E. Punt Indices of abundance based on fishery catch-per-unit-effort (CPUE) are important components of many stock
assessments, particularly when fishery-independent surveys are unavailable. Standardizing CPUE to develop
Keywords: indices that better reflect the relative abundance requires the analyst to make numerous decisions, which are
Catch-per-unit-effort (CPUE) standardization influenced by factors that include the biology of the study species, the structure of the fishery of interest, the
Indices of relative abundance
nature of the available data, and the objectives of the analysis such as how standardized data will be used in a
Good practices
subsequent assessment model. Alternative choices can substantially change the index, and hence stock assess­
Data preparation
Modelling methods ment outcomes and management decisions. To guide decisions, we provide advice on good practices in 16 areas,
Stock assessments focusing on decision points: fishery definitions, exploring and preparing data, misreporting, data aggregation,
density and catchability covariates, environmental variables, combining CPUE and survey data, analysis tools,
spatial considerations, setting up and predicting from the model, uncertainty estimation, error distributions,
model diagnostics, model selection, multispecies targeting, and using CPUE in stock assessments. Often the most
influential outcome of exploring and analysing catch and effort data is that analysts better understand the
population and the fishery, thereby improving the stock assessment.

1. Introduction costly or impractical (Bishop, 2006). In such situations, the main

approach is to develop models that standardize fishery-dependent catch
Indices of abundance based on fishery catch-per-unit-effort (CPUE) and effort data to produce an index of abundance (often simply referred
are important components of many stock assessments for fish and other to as the ‘index’) that better reflect the relative abundance by accounting
marine species, particularly when fishery-independent sources of in­ for the other factors that influence catch rates (Maunder and Punt, 2004;
formation about population trends, such as research surveys, are un­ Ye and Dennis, 2009). This practice dates at least to Beverton and Holt,
available (Maunder and Punt, 2004). This is the case, for example, in Section 12) (1957).
many fisheries for pelagic or lower value species where surveys are too CPUE standardization requires the analyst to make numerous

* Corresponding author at: National Institute of Water and Atmospheric Research Ltd (NIWA), 217 Akersten St, Port Nelson, Nelson 7010, New Zealand.
E-mail address: [email protected] (S.D. Hoyle).

https://doi.org/10.1016/j.fishres.2023.106860
Received 30 April 2023; Received in revised form 15 September 2023; Accepted 16 September 2023
Available online 30 September 2023
0165-7836/© 2023 Elsevier B.V. All rights reserved.
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

decisions. These decisions are influenced by factors that include the

C qN
biology of the study species, the structure of the fishery of interest, the CPUE = = qD =
E A
nature of the available data, and the objectives of the analysis (including
how standardized data will be used in a subsequent assessment model). where N is the number of fish on the fishing grounds (a three-
Alternative choices can change the indices and, consequently, the stock dimensional volume, usually defined by the two-dimensional surface)
assessment outcomes and resulting advice to fishery managers. To guide and A is the spatial area of the fishing grounds. It follows that changes in
decisions and choices, in this paper we provide advice on good practices CPUE are due either to changes in the stock density or to changes in the
in 16 areas for analysts who develop indices of abundance for use in catchability coefficient. If the changes in q can be accounted for (‘filtered
stock assessments (Table 1, Fig. 1). For each of the 16 areas, we focus on out’), then the remaining changes in CPUE can be related to those in
decision points – occasions where analysts need to select an approach – stock density. This is the basic idea underlying what is known as the
particularly on instances where those choices can affect the resulting standardization of catch rates.
index. Analysts are encouraged to consider the proposed good practices The concept of abundance needs some elaboration, particularly in
and apply them where practical. It is important for the analyst to un­ relation to the concept of availability. Following the definitions pro­
derstand the implications of applying or not applying these practices and posed by Marr (1951), true abundance is the absolute number of in­
to be able to justify their decisions. dividuals in a population, availability is the degree (a percentage) to
We begin by considering the fishery definition in the stock assess­ which a population is accessible to the efforts of a fishery, and apparent
ment – i.e., for what stock component are we estimating CPUE? Next, we abundance is the abundance as affected by availability, or the absolute
consider the available data, including issues associated with data number of fish accessible to the fishery. The concept of availability can
preparation, quality, misreporting, resolution, and aggregation. We then be further decomposed into two components: one that is directly under
discuss processes that can affect catch rates, such as variables associated the influence of the fishing gear (i.e., the selectivity of the gear defined
with catchability versus density. We then move to aspects of the analysis as the probability of capture of any fish dependent on individual traits
itself, including the tools used, methods to identify fishing fleet targeting such as size) and one which is influenced by factors other than the
strategies, issues associated with space, setting up and predicting from fishing gear (see below).
the model, error distributions and uncertainty estimation, model di­ From these definitions, if M represents the true abundance, N mea­
agnostics, and model selection. Finally, we consider approaches for sures the apparent abundance, s measures the selectivity of the fishing
employing CPUE in stock assessments, including the use of index fish­ gear, and a represents the other component of availability, then N = saM
eries, splitting the time series, selectivity changes, and catchability and substituting into the above equation gives:
change.
The advice provided in this paper can be implemented using various saqM
CPUE =
software packages. However, given their free availability and wide­ A
spread use, we primarily focus on examples using the R software envi­ From this equation, it is seen that the relationship between CPUE and
ronment and its associated libraries (R Core Team, 2022). the true abundance of fish within a given spatial region is influenced by
both the selectivity (s) and other components of availability (a) of the
2. Basic equations and definitions fish to the fishing gear, and the efficiency of the fishing gear (q); thus, to
adequately standardize CPUE, the analyst needs to understand what
The relation between catch rates (CPUE) and stock abundance is influences each of these factors. Note, in most of the fishery literature,
based on the catch equation which, as a first order approximation, re­ the selectivity parameter combines the two parameters selectivity and
lates the number of fish in the catch, C, fishing effort, E, and the average availability as defined here (i.e., the distinction between the s and a
fish population density, D, on the fishing grounds: terms is rarely made). Selectivity and availability often differ by factors
such as age, size, sex, and stage and can be informed by the composition
C = qED
data based on these characteristics. To simplify the following de­
where q is a fixed constant of proportionality known as the catchability scriptions and discussions, we simply refer to selectivity, availability,
coefficient and is related to the efficiency of the fishing gear (i.e., the and composition data in the following without specifically referring to
proportion of the stock removed by one unit of effort). From this these characteristics.
equation: Availability will be influenced by the environmental conditions
prevailing at the time of the fishing operation and/or behavioural at­
tributes of the species being targeted. For example, oceanographic
conditions may influence the vertical distribution of the habitats
Table 1 preferred by both pelagic and semi-demersal target species and, conse­
The 16 areas of good practices in catch-per-unit-effort (CPUE) standardization quently, the overlap of these habitats with the fishing gear (Hinton and
discussed in this paper. Nakano, 1996; Maunder et al., 2006; Monnahan et al., 2021). Behav­
Section/subsection Area of good practices ioural characteristics of a species, such as diurnal vertical migrations,
Section 3 Fishery definitions
may also influence the vertical availability of fish to the gear (Abascal
Subsection 4.1 Exploring and preparing data et al., 2010; He et al., 1997). Covariates describing environmental
Subsection 4.2 Misreporting and biases conditions that likely impact local availability to fishing gear are one
Subsection 4.3 Data aggregation type of ‘catchability covariate’. Alternatively, oceanography can also
Subsection 4.4 Density and catchability covariates
affect the local density, i.e., by causing fish to move outside of the fished
Subsection 4.5 Environmental variables
Subsection 4.6 Combining fishery and survey data area and these covariates are often referred to as habitat or ‘density’
Subsection 5.1 Model fitting methods covariates (O’Leary et al., 2020). The distinction between ‘catchability’
Subsection 5.2 Spatial considerations and ‘density’ covariates is discussed more thoroughly in Sections 4.4 and
Subsection 5.3 Multispecies targeting
5.8.
Subsection 5.4 Error distributions
Subsection 5.5 Uncertainty estimation
In addition to localized environmental conditions, catchability will
Subsection 5.6 Model diagnostics be influenced by the types of fishing gear used and how they are
Subsection 5.7 Model selection deployed. The effectiveness of a given fishing gear (i.e., its ability to
Subsection 5.8 Assembling an index from a fitted model catch the available fish being targeted) will also depend upon a range of
Section 6 Using CPUE indices in stock assessments
factors that are under the influence of the fisher. For example, the catch

2
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

Fig. 1. Flow chart summarizing the stages of an analysis of catch per unit effort data.

of broadbill swordfish (Xiphias gladius) increases with the use of light- selectivity, and many assessments assume that selectivity is constant
sticks and squid bait on shallow sets deployed in the afternoon, while through time, even though in practice it tends to vary over time
albacore tuna (Thunnus alalunga) prefer pilchard baits on deep sets (Sampson and Scott, 2012). Wrongly assuming constant selectivity can
deployed in the morning (Campbell, 2019; Campbell et al., 2017). In­ bias the assessment results (Hilborn and Walters, 1992; Martell and
dividual fishers will also strive to improve the effectiveness of the fishing Stewart, 2014), although so can time-varying selectivity in the presence
operations to increase the catch rates, and the combined effects of of unbalanced composition sampling, substantial data gaps, and/or
learning and improving technology over time lead to a phenomenon model misspecification. These biases are driven in three main ways: (1)
commonly known as ‘effort creep.’ Towards this end, fishers will often by affecting the proportion of the population abundance vulnerable to
‘experiment’ with the way the fishing gear is deployed to maximize its the fishery, which will affect model results via the fit to the CPUE; (2) by
effectiveness. To understand this variability and allow analysts to affecting removals from the population and, therefore, the population
standardize the resulting CPUE for these differences, it is important that structure; and (3) by affecting the composition data, resulting in a
the details of how the fishing gear is deployed are fully recorded in the mismatch between the observed and expected size/age/sex composition
vessel logbook. which can substantially affect results (particularly population scaling)
via the composition likelihood. Joint modelling can be used to ensure
3. Fishery definitions that CPUE and composition data are set up consistently in the model (see
below).
A CPUE index in any stock assessment is associated with the part of a To reduce these biases, fisheries should be defined in ways that
stock sampled by a specific fishery (termed a ‘fleet’ or sometimes minimize selectivity change through time. An important source of
‘métier,’ defined here as a specific fleet deploying a specific fishing selectivity variation is changing spatial and seasonal distribution of
method). The assessment model defines the relationship between the fishing effort through time and its interaction with spatiotemporal
index and the overall stock via the catch equation (catchability coeffi­ variation in population structure and availability.
cient and the selectivity function in age/length/sex/stage structured The analyst should explore spatial and seasonal patterns in avail­
assessment models), and for multi-region models, the spatial structure. ability by size, age, and/or maturity to identify fishery definitions that
Defining the fishery (and its representation as a fleet in the assess­ will be robust to changing effort distribution. Tree-based analytical
ment model) is, therefore, a fundamental aspect of CPUE standardiza­ methods can be used to identify optimal spatial arrangements for fish­
tion, requiring a detour into the discussion of selectivity for structured eries by exploring spatial variation in composition and CPUE trends
assessment models. Most structured stock assessments make the sepa­ (Lennert-Cody et al., 2013). Similarly, generalized additive models
rability assumption, i.e., that fishing mortality is the product of catch­ (GAMs) can be employed to identify spatial, seasonal, and environ­
ability, annual fishing effort, and selectivity (Quinn and Deriso, 1999). mental patterns in length, maturity, and sex ratio which are then used to
As shown above, selectivity is a combination of availability and contact define fisheries for CPUE standardization (Devine et al., 2022; Hoyle

3
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

et al., 2017). Care needs to be taken when the spatial definition does not stock. Such choices are a topic for ongoing research.
include the whole population (see below).
Such exploratory analyses are also an effective way to improve un­ 4. Data issues
derstanding of the biological structure of the stock and the fisheries. This
understanding is often more important for assessment outcomes than 4.1. Exploring and preparing data
changes in trends of the indices themselves.
Gear selectivity can also change through time, the effect of which can Before standardizing CPUE it is essential to thoroughly explore the
be minimized by separating gear types with different selectivity into catch and effort dataset, to develop understanding of the data, generate
different fisheries. For example, a change from J-hooks to circle-hooks in hypotheses, and inform the standardization strategies to consider. Un­
the Australian longline fishery for broadbill swordfish resulted in fewer derstanding can be developed by graphically summarising the data
larger fish being caught (Campbell et al., 2019; Pilling and Brouwer, (both the response variable and covariates), exploratory modelling (e.g.,
2017). Managing selectivity variation with fishery definitions is a useful using random forest (RF) models), and by consulting with key stake­
way to account for the major sources of selectivity variation and to holders. There should be special focus on identifying changes in the
predict catch and CPUE with plausible selectivity for strata with gaps in distributions of covariates through space or time, as covariates that
the composition data time series. However, the number of fishery defi­ change are the most likely to affect the index (Bentley et al., 2012). This
nitions can be limited by the difficulty of managing them with the stock preliminary exploration will also allow the analyst to anticipate issues
assessment package being used, and by the information available to with model variables such as data entry errors, missing data, and
distinguish them and to estimate selectivity. outliers.
An alternative or complementary approach to account for selectivity Exploratory analyses should consider the sources of the data and the
variation is to jointly model the CPUE and composition data (Maunder constraints imposed by data collection and storage methods through
et al., 2020) using a spatiotemporal model (STM) to create a joint index time. Logbooks and observer forms often change through time in ways
of relative abundance and stock composition. The approach augments that affect data quality, such as in the recording of spatial and temporal
the standardization with the observed composition data to generate two resolution, which covariates are reported, the precision and detail re­
sets of predicted composition data: one to estimate the index fleet ported, species resolution, whether catches are recorded in number or
selectivity and the other to estimate the extraction fleet selectivity. The weight, whether a species is recorded, and reporting of sex and size. Data
index composition data are spatially weighted by the predicted stan­ storage methods may also be influential.
dardized abundance (the index), while the extraction fleet composition Graphical summaries can be undertaken in consecutive steps from
data are spatially weighted by the catch. The index fleet can be designed low-to-high resolution. Aggregated summaries (e.g., histograms or
to have stable selectivity through time by accounting for spatial and density plots) of each variable across the full dataset provide an overall
seasonal covariate effects on availability, as well as catchability varia­ description of the response variable and potential covariates. These
tion associated with vessels, equipment, and fishing technique. In explorations should consider the proportions of zeroes (where relevant)
contrast, extraction fleet selectivity will vary through time. Some con­ and the range of the data prior to cleaning. Response variable summaries
straints with the joint modelling approach are that STMs can be data will inform the set of error distributions to consider (Campbell, 2015;
hungry and computationally slow and that their implementation re­ Hoyle et al., 2014b); see also Section 5.4, including delta (hurdle) and
quires more expertise than traditional approaches to set up and run. mixture models (Langley, 2019). Aggregated summaries also help
These factors have tended to limit their application, but simplifications identify outliers and data entry errors (see below) and the need for
are available to reduce computational demands (e.g., don’t model the possible data transformation. Examining individual values taken by
composition data but use the raw compositions). In addition, complex­ continuous variables is often necessary to identify and remove outliers.
ities such as spatial variation in covariate effects may be difficult to Spatial and temporal trends in the catch data and candidate cova­
parameterize. This approach will be further discussed in Sections 5 and riates should be described through, for instance, time-series of boxplots
6. and maps of summary metrics (mean, median, upper and lower quan­
It is important to consider the spatial domain of the stock assessment: tiles). Where possible, maps of the variable of interest should be dis­
how this domain relates to the population domain of the species, how aggregated by temporal strata to help detect any spatiotemporal trends.
this domain can be subdivided into a hypothetical ‘stock-wide sampling Spatial GAMs can be useful for exploring patterns and identifying sta­
frame’ composed of spatial sampling units, and what portion of the tistically meaningful trends (see Section 5.2).
sampling frame is informed by CPUE data for a given fleet. If the fleet To help identify factors that may influence catch rates, relationships
samples only a small proportion of this stock-wide sampling frame, then between nominal CPUE and candidate covariates should be explored.
there is likely to be substantial non-random variation through time in These data analyses will inform the type of modelling framework to use
the proportion of the assessed stock available to that fleet. This variation for continuous covariates, as not all model types allow non-linear re­
in availability through time is represented as variation in catchability for lationships (see Section 5.1). Exploratory modelling using a flexible
a standardized index from the fleet if it is used to represent changes in modelling framework such as RFs or boosted regression trees (BRTs)
stock-wide abundance or size/age/sex composition (Wilberg et al., could also be undertaken to identify covariates explaining the highest
2009). proportion of the variability in the nominal CPUE and the potential
The best approach to define fisheries and their associated CPUE will shape of the relationship (e.g., linear or quadratic).
depend on the specifics of each stock assessment, the available data, and Analysts should maximize the number of records available for the
the analyst’s skills and time. Indices that represent a higher proportion standardization to improve model performance and estimates of un­
of the stock are more likely to be representative (i.e., fewer processes certainty. As such, the treatment of missing data in covariates needs
could drive changes in availability that are confounded with density). careful consideration as most standardization methods can only include
However, there may be constraints that restrict the components that can records that have values for all the model covariates (Forrestal et al.,
be included in a representative index such as spatiotemporal patterns or 2019). Analysts should tabulate missing values for each covariate and
covariate effects that are too complex to include in a single STM. It may try to identify causes and correlates of missingness. Options for records
be useful to combine the use of fishery definitions to account for large with missing values include dropping the full record, dropping the co­
and stable effects on selectivity with STMs to account for remaining variate, inferring the value from other information (e.g., a missing
variation within each fishery. On the other hand, it may be better to fishing gear value may be inferred from the gear typically used by the
combine multiple fisheries by calibrating them (with catchability and fisher), or imputation (e.g., median values per vessel). Approaches for
selectivity estimated) in an STM to try and span the spatial range of the imputing values range from simple rule-based procedures (Walters,

4
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

2003) to more sophisticated geostatistical (Munoz et al., 2010; Thorson fisheries, Langley, 2019).
et al., 2015) and Bayesian algorithms (Shemla and McAllister, 2006). Bycatch species are more likely to be under-reported when they are
When including a covariate with a high proportion of missing entries of little interest to the crew (due to their low commercial value), and
(>5–10%) that changes the standardized index, the analyst should rates may vary between vessels and change through time. For example,
confirm that the change is due to the covariate itself (e.g., by using in­ rates of shark bycatch reporting in Japanese southern bluefin tuna
fluence plots) and not due to the removal of records where the covariate (Thunnus maccoyii) longline fisheries varied substantially between ves­
was missing. sels and increased substantially in 2008 for reasons that remain unclear
The identification of outliers and their treatment is a key step in (Hoyle et al., 2017). When rates of misreporting vary between vessels,
preparing data for CPUE standardization. Model predictive performance analysts can focus on vessels that report more reliably (Grüss et al.,
is likely to be poorer in the lower and upper tails of covariates given 2022). It should be noted that not all data may be needed to estimate a
insufficient data, and outlying values are more likely to be reporting precise index. Limiting the analysis to reliable data that has less vari­
errors. Common approaches include omitting values outside ranges that ability in catchability and selectivity will be adequate if they have suf­
are either fixed or pre-determined based on quantiles. The analyst ficient spatial, temporal, and covariate coverage.
should also check whether outliers are true records of infrequent fishing Under-reporting can also occur due to discarding or high-grading,
behaviour (for instance) or data entry mistakes. which can be based on size or relative commercial value (e.g., tuna
Lastly, stakeholder engagement is an important, yet often over­ discards in the Indian Ocean, Huang and Liu, 2010), or on prescribed
looked, component of developing a CPUE standardization. It supports all conservation measures. The stage in the fishing process when discarding
the steps outlined above. Fishers, fishery observers, and even fish sellers occurs can affect whether it is recorded as part of the catch. Analyses
can provide valuable insights to clarify potential data entry issues, may need to be adjusted to account for changes in discard rates (e.g.,
confirm or invalidate data trends identified in exploratory analyses, Hoyle et al., 2019b). Changes in reporting procedures, such as the
advise on key covariates to include and their bounding values (e.g., net introduction of electronic monitoring, may also change the rates of
length, number of hooks per line), and explain motivations for changes reporting of discard species (Emery et al., 2019). This is important if
in fishing behaviour (Tesfamichael et al., 2014). When detailed logbook discards are included in the total catch used in CPUE analyses. Depre­
data are available, fishers can also highlight the targeting and gear dation of fishing sets by large marine predators (e.g., Peterson et al.,
setting practices they consider will influence the catch rates of the 2014; Roche et al., 2007) can also result in under-reporting or
species being targeted. Importantly, engagement can improve trust in zero-inflation of catch and may require additional modelling or adjust­
the analysis as well as facilitate constructive dialogue when the final ment of the data.
indices are considered for management. Misreporting is difficult to address but there are options in some
circumstances. If available, analysts can compare catch rates between
4.2. Misreporting and biases data types to identify groups of unreliable records, such as comparisons
between observer data and logbook data (Hoyle et al., 2017), or between
Misreporting catch of the species of interest is a consistent and logbooks before and after the introduction of electronic monitoring
sometimes major problem in many fisheries (Pitcher et al., 2002; Rudd (Emery et al., 2019). Since catch rates vary with covariates and target­
and Branch, 2017), which can affect the data used in CPUE standardi­ ing, predicting expected logbook catch rates based on models fitted to
zation. Indices are more prone to bias when rates of misreporting are observer data is more reliable and powerful than simply comparing raw
higher or change through time. To understand the potential for bias, the catch rates between datasets (Hoyle et al., 2017; Kai, 2019).
analyst needs to become familiar with fishing industry dynamics and the
data collection processes and to explore the data. It is useful to construct 4.3. Data aggregation
a timeline of changes (e.g., regulations, gear specifications, logbook
forms, observer training, market conditions etc.) that may affect Analysis of aggregated CPUE data may be required or convenient
reporting for the species and fishery being evaluated. This can help to given the availability and quality of data. Due to privacy concerns,
generate hypotheses to consider while modelling, particularly when a public domain fisheries data are often available only at aggregated
sharp change is observed between time periods. spatial and temporal scales (Hinz et al., 2013). Aggregation may begin
One cause of misreporting is errors in species identification, which is with the fishing logbook, e.g., with effort reported at the day or even the
particularly important for bycatch species. Such errors can also affect trip level. Data may also be aggregated across species due to identifi­
target species in both commercial (e.g., Beerkircher et al., 2009; Peat­ cation problems or for convenience. Although analyses of aggregated
man et al., 2019; Webber and Starr, 2022) and recreational (Jones, data may be the only option when finer-scale data (e.g., set-by-set or
2004) fisheries. The ways species are reported can also vary. For species level data) are unavailable, their results should be treated with
example, sharks can be recorded either at the species level or at a more caution and critically evaluated given the issues raised in the following
generic grouping level. CPUE indices may need to be restricted to pe­ paragraphs.
riods with reliable species identification (e.g., Noriega et al., 2011). Indices derived from data aggregated spatially, temporally, or across
Sampling bias can affect catch estimates based on estimates of spe­ species may not vary in proportion to the target stock (or complex).
cies composition made by both fishers and observers. Examples include Covariates that affect catch rates at the set or vessel level (e.g., gear
biases associated with grab sampling (Peatman et al., 2019), inconsis­ settings) or summarize the effects on catchability of multiple factors (e.
tent reporting and possible layering in the catch (Webber and Starr, g., vessel IDs) are often unavailable when CPUE data are aggregated to
2022), and the stratification used for statistical analysis (Duparc et al., coarser spatial or temporal strata. This loss of information may limit the
2020). ability of CPUE standardization to correct for changes in catchability
Changes in misreporting can also be linked to administrative factors over time (e.g., effort creep: Kleiven et al., 2022; Palomares and Pauly,
such as changes in logbooks, regulations, the introduction of e-moni­ 2019). Aggregating data temporally and using a coarser temporal defi­
toring, or observer training. Both under-reporting and over-reporting nition of effort (e.g., day, trip, quarter) than the higher-resolution
can result from attempts to avoid quota limits, such as when the har­ effective unit of effort (e.g., fishing sets, pot throws, or hours
vest location is misreported (e.g., Hoyle et al., 2015b). Deliberate searched) can change the relationship between CPUE and abundance (i.
over-reporting can be linked to attempts to establish a catch history in e., the covariate changes over a smaller temporal or spatial scale, often
anticipation of future quota allocation (Buck, 1995). Under-reporting intentionally to improve catch rates, but these changes are not apparent
can be a feature of a management system, such as when only the top in the aggregated data). For example, defining CPUE as catch per day
N captured species are reported (e.g, New Zealand inshore trawl could mask a population decline if search time within each day or the

5
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

number of schools fished changes to maintain stability in catch magni­ 4.4. Density and catchability covariates
tude (Ducharme-Barth et al., 2022; Hsu et al., 2022). Spatial and tem­
poral aggregation of CPUE data may create a mismatch between the Since Beverton and Holt, Section 12) (1957), stock assessment sci­
actual oceanographic conditions and the oceanographic covariate entists have standardized fishery CPUE to remove the confounding ef­
associated with the aggregated CPUE data. If oceanographic covariates fects of vessel size and power. For example, Maunder and Punt (2004)
are modelled as density covariates (see Section 4.4), this mis-match state “Explanatory variables should, however, be considered in an analysis
could lead to spurious estimated relationships between predicted den­ only if there is an a priori reason that they may influence catchability.”
sity and the covariate and, if using STMs, could result in inappropriate However, spatial ecologists do the exact opposite by including cova­
spatial imputations of density (Ducharme-Barth et al., 2022). For riates and then conditioning upon their effect when predicting densities
example, fishers may move to find a specific habitat type within a larger across space. To alleviate this confusion, Thorson (2019a) distinguishes
spatial stratum to increase catch rates of a certain species. If fishers between ‘catchability’ and ‘density’ covariates: both are included in a
target fish aggregations or areas of preferred fish habitat such as eddies linear predictor to explain catch-and-effort data, but only density (and
or seamounts at a finer resolution than the spatial scale of the aggregated not catchability) covariates are conditioned upon when predicting
CPUE data, then localized, sequential depletion of aggregations could densities across space. The index is constructed by aggregating pre­
introduce hyperstability into the index (Cardinale et al., 2011; Sadovy dictions across a specified spatial and temporal domain such that the
and Domeier, 2005). Lastly, aggregating CPUE data across species is effect of catchability is removed from the index.
likely to exaggerate fishery-induced changes to the combined abun­ Distinguishing between catchability and density covariates leads to
dance, given that catchability varies across species (Kleiber and the question: how does the analyst know whether a covariate affects
Maunder, 2008). Similarly, if productivity varies across the aggregated density, catchability, or both? For example, species might migrate to
species, fishery-induced decline in abundance of the less productive maintain a desired daytime foraging temperature (Lehodey et al., 1997)
species might be masked by the abundance trend of the more productive such that local temperature predicts population densities. Simulta­
species (Dulvy et al., 2000). neously, temperature might affect digestion and metabolic rates, leading
In addition to the risk of bias, analysing aggregated data poses sta­ to a greater attraction towards a baited fishing hook (and hence a higher
tistical modelling challenges, such as applying an appropriate variance CPUE for a given density and soak duration). In this case, an argument
structure. Commonly used error structures (e.g., Lognormal and Pois­ could be made that temperature affects both catchability and density. To
son) for CPUE standardization models are generally robust to hetero­ resolve these hypothesized mechanisms, an analyst could combine
scedastic data (e.g., non-constant variance assumption where, typically, fishery CPUE with auxiliary information, e.g., by measuring behaviour
the variance increases in proportion to the mean). However, this mean- directly using satellite and/or conventional tags. Similarly, analysts can
variance relationship is inverted in aggregated data if areas of high catch make inferences about density, catchability, and availability via:
rates also attract higher concentrations of effort, thus reducing the
variance and violating distributional assumptions (e.g., Hoyle, 2021). 1. Local depletion: In freshwater sampling, analysts can sample and
Furthermore, spatially aggregating data makes it difficult for models to remove individuals in multiple survey passes, using the decline and
accurately represent the spatiotemporal correlation structure. For quantity of removals to identify catchability. Similarly, in sessile
instance, in GAMs, where the spatial correlation is modelled implicitly marine organisms, a decline in fishery CPUE over a short season can
by varying the degree of the spatial smoothing applied, aggregating data be compared with preceding fishery removals to estimate
spreads the spatial and temporal extent of abundance hotspots, and catchability.
generates less flexible spatial smooths (e.g., fewer degrees of freedom) 2. Paired sampling: Similarly, bottom trawl data can be paired with
which may further inappropriately ‘smear’ the hotspots when making vertically disaggregated acoustical backscatter to identify vertical
spatial predictions. distribution and infer availability to gear operating at a given vertical
Perhaps most importantly, aggregating data can remove fine-scale layer (Monnahan et al., 2021).
information that allows analysts to understand fishing behaviour. The 3. Process studies: Finally, an analyst might use a small-scale experiment
key process of data exploration is limited by loss of information about, (e.g., temperature-dependent feeding experiments) to identify how
for example, individual vessel behaviour, fine-scale fish and vessel dis­ temperature might affect bait attraction and then assume this rela­
tributions, and movement patterns. Analysis methods to identify tar­ tionship to ‘subtract out’ this effect.
geting strategy based on species composition (see Section 5.3) may
remain possible (e.g., Fu et al., 2016), but with less resolution and Regardless of the process, we recommend that analysts explicitly
sensitivity than with operational data. outline their rationale when specifying that a covariate affects density or
Although they do not balance the concerns above, there can also be catchability. In cases where the correct process is difficult to determine,
advantages associated with aggregating data before standardization, analysts should explore the implications of excluding the covariate from
particularly for fisheries where pseudo-replication and serial autocor­ the index and including it as either density or catchability. These
relation are issues. For example, a vessel that obtains high CPUE along alternative hypotheses can form the basis of alternative CPUE index
an oceanic front may continue to fish the front for as long as high catches scenarios considered in the stock assessment.
persist, so that sets cannot be considered independent. Similar effects are
known from sequential trawl hauls. It is often computationally chal­ 4.5. Environmental variables
lenging to apply the appropriate autocorrelation or random effect
structure to address the consequent violation of independence, espe­ As noted in Section 4.4, aspects of the aquatic environment, such as
cially for large datasets. Given the risks of bias due to aggregation on the water temperature, surface chlorophyll-a concentration, oxygenation,
one hand, and pseudo-replication on the other, aggregation by vessel, and light at depth, affect both fish density and catchability. The task of
month, statistical cell and/or other factors may in some cases be the CPUE standardization is to eliminate the effect of catchability variables
lesser of two evils. A further option is to subsample the data in a random so the analyst can employ CPUE data to infer differences in fish density
or structured way to reduce pseudo-replication. Trade-offs can be across time and space. Given this estimate of density across time and
explored through simulation, or by working with more computationally space, the analyst can then sum density in each area to generate an
tractable subsets of the operational data. abundance index for use in stock assessment, or average over years to
get an estimate of habitat utilization for use in spatial management.
Regardless of whether analysts specify a covariate as affecting either
density or catchability, all covariates are used to inform model fit to

6
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

observations. Treatment between the two types of covariates differs at taken by each analyst, particularly for variable categorization and model
the prediction stage. Catchability covariates are fixed at a base value selection (Forrestal et al., 2017; Forrestal et al., 2019).
when predicting density across space and time, thereby ‘filtering out’ Two research areas closely related to CPUE standardization are
variation associated with different values in the fitted data. By contrast, species distribution modelling (SDM) and habitat suitability modelling
a density covariate has an assigned value at every location across a (HSM). Both SDM and HSM heavily use environmental variables to
modelled spatial and temporal domain, and the analyst conditions upon model fish distributions and their habitat preferences (Bosch et al.,
these values when predicting densities. It should be noted that catch­ 2018; Lee and Terrell, 1988; Maunder et al., 2006; Pickens et al., 2021;
ability covariates are only needed for catch events, while density Rowden et al., 2017; Zhang and Li, 2017). These studies involve similar
covariates are needed for all temporal and spatial strata included in the data and modelling techniques to CPUE analyses. They also often
index irrespective of whether there is catch. Including a variable that consider a third spatial dimension, depth, which can be particularly
affects density as a catchability covariate can result in biased inference. important for some gears such as pelagic longline or for bottom-fish
For example, if recruitment decreases as the average temperature in­ species. (see also Hinton, 1996); Hinton and Nakano (1996) developed
creases (e.g., due to climate change), including annual water tempera­ a method to calculate indices of abundance for pelagic longline that
ture as a catchability covariate in an index-standardization model and match the spatiotemporal-vertical habitat with the longline gear and
subsequently using a constant value in the prediction will mask the true species habitat preference. This method was extended into a statistical
abundance decline over time. However, including average temperature framework by Maunder et al. (2006).
as a density covariate (and using its value when predicting densities) can
result in improved estimates and forecasts (e.g., O’Leary et al., 2020). 4.6. Combining fishery and survey data
The converse is also true — wrongly attributing to density an effect that
is linked to catchability will introduce bias. In some jurisdictions, analysts produce a separate CPUE index for
Several environmental variables are commonly used in CPUE each fishery fleet or métier. This then results in a multitude of abun­
models. Variables such as sea surface temperature, sea surface salinity, dance indices, and equally weighting these indices implicitly results in
sea level height, chlorophyll-a concentration, and dissolved oxygen the assessment model averaging them. To resolve and simplify this sit­
concentration have been found to be significant predictors of spatial uation, Conn (2010) used a state-space model to combine multiple
distribution for some species (Campbell, 2015; Han et al., 2022; Liu indices into a single ‘consensus’ index, an approach that has been refined
et al., 2022; Tian et al., 2009), particularly for pelagics. Some environ­ using Dynamic Factor Analysis (DFA) (Peterson et al., 2017; Peterson
mental variables, such as moon phase, eddies, water turbidity, and cloud et al., 2021). However, DFA has several drawbacks, including that: (1) it
cover, may affect the availability and/or catchability of certain species must assign some implicit weight to each constituent index when
and gear types. combining them, and these constituent indices representing small or
Environmental variables are often not recorded by fishers at the time large areas are often given equal weight; and (2) residual variation in
of fishing, but rather obtained from other sources such as meteorological constituent indices is ignored, and the analyst must make some decision
organisations. This is particularly true for density covariates, which about which DFA index represents changes in the stock or is attributed to
must be available for all locations and time periods. Variables with a correlated variation in catchability.
long time-lag (e.g., temperature in the months preceding the fishing Given the difficulty of reconciling differences in multiple indices
event) should be used with caution in CPUE models. They may affect within an assessment, or pre-processing using DFA, analysts increasingly
productivity during the period leading to fishing and therefore local seek to include data from multiple fisheries and/or surveys during index
density (a density covariate), but are often autocorrelated, where cur­ standardization. Joint analysis expands data coverage spatially and
rent values may affect catchability (a catchability covariate). Similarly, temporally. It also ensures consistency of analytical methods, thereby
analysts sometimes use covariates that are derived from a large spatial removing a key source of differences between indices. This occurs, e.g.,
area or obtained from satellite or regional ocean modelling systems (e.g., in combining nearshore and offshore survey data (Perretti and Thorson,
Phillips et al., 2014). In all cases, it is important to provide some 2019), combining multiple surveys to achieve a basin-scale index
ecological justification for (1) whether those variables are affecting (Maureaud et al., 2021; Ono et al., 2018), or fleets from multiple nations
density or catchability and (2) why those are better than using to increase the spatial and temporal coverage of an index (Ducharme-­
vessel-based measurements of environmental conditions, if available. Barth et al., 2020; Hoyle et al., 2019a; Hoyle et al., 2015b). Prior to joint
The error in measuring the covariates may also need to be considered in analyses of data from multiple fleets, individual national datasets should
the model, particularly if the error varies among spatiotemporal strata. be thoroughly explored (e.g., Hoyle et al., 2015a; Hoyle and Okamoto,
Despite being important factors affecting fish distribution, many 2015; Hoyle et al., 2015c) to identify and eliminate sources of data
environmental variables are confounded with spatial and seasonal co­ conflict. Joint indices for Atlantic tropical tunas (Hoyle et al., 2019a;
ordinates that are often employed as predictors in CPUE models. Hence, Hoyle et al., 2019c) were judged to have improved the resulting stock
spatial-temporal effects can act as proxies for comprehensive dynamic assessments (Anonymous, 2019; Walter et al., 2020) by reducing data
environmental variables, especially if interactions between these effects conflicts, improving model diagnostics, and ensuring broad and
are included to help account for temporal changes in environmental consistent spatial and temporal coverage.
conditions at any spatial location. If the environmental conditions However, there is less research regarding how to combine data from
generally exhibit consistent spatial and temporal patterns, which is often multiple fisheries and/or surveys. As one exception, Grüss and Thorson
the case, explicitly including these environmental variables may explain (2019) combined data from different surveys to generate an abundance
little additional variation. In the cases where using environmental var­ index for Gulf of Mexico red snapper (Lutjanus campechanus). Notably,
iables does increase accuracy, their inclusion may also increase the the authors first analysed each data source individually to confirm that
annual coefficients of variation (CVs) compared to the models without any apparent conflict in the indices could be explained by differences in
the environmental variables, likely due to the added requirement of the spatial extent of each fleet. Rufener et al. (2021) combined fishery
estimating a relatively imprecise relationship between catch rates and and survey data, after confirming minor apparent data conflict.
environmental variables (Forrestal et al., 2017; Forrestal et al., 2019). Given these successes, we suspect that there will be ongoing efforts to
Goodyear (2016) explored relationships between environmental combine CPUE from multiple fisheries and/or surveys. We recommend
factors, three-dimensional variation in habitat, and longline CPUE in a the following good practices:
case study for blue marlin (Makaira nigricans). Results of analysing data
generated from this model generally favoured the inclusion of envi­ 1. The analyst should standardize each dataset individually using
ronmental and habitat variables but were affected by the approaches broadly consistent methods and compare resulting density maps and

7
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

abundance indices both visually and using goodness-of-fit criteria (e. Random effects have limited application for categorical covariates
g., Alglave et al., 2022; Rufener et al., 2021); with too few levels to estimate the shape of the distribution, e.g., bait
2. Differences between dataset should be explained mechanistically type. However, they can be useful for covariates with multiple (> 12)
where possible, such as by including covariates that can explain the levels that may be expected to belong to a common, normal distribution,
differences in catchability. such as a vessel effect. In these instances, some levels might have too few
records to reliably estimate their model coefficients as fixed effects.
These recommendations arise from the expectation that data con­ Poorly sampled levels are not an issue in a mixed-effect framework as
flicts imply model misspecification (the “Law of Conflicting Data”, they are shrunk towards the coefficients of better-informed levels.
Maunder and Piner, 2017). The analyst should also consider that Nevertheless, fixed effect levels with very low record numbers often
selectivity differences between surveys and fisheries may require have minimal influence on the indices if retained and may be omitted if
simultaneously modelling the catch rate and length composition data to they cause model convergence problems. Note also that vessel effects
estimate differences in selectivity, in which case calibrating catchability will not be normally distributed if, for example, there are groups of
and selectivity may require spatial overlap. vessels with similar catchability or if there are temporal trends in the
catchability of vessels joining and leaving the fishery. In such cases, the
5. Analysis analyst may add structure to the random effects model or may prefer to
use fixed effects. Simulation testing is useful for selecting a strategy that
5.1. Modeling framework provides a reliable index.
A common strategy for vessels is to define a ‘core fleet’ comprised of
As outlined in previous sections, the objective of CPUE standardi­ vessels meeting an arbitrary set of activity thresholds (e.g., as a function
zation is to remove the effects of changes in fishing practice through of catch and/or time present in the fleet; (Kendrick and Bentley, 2011;
time (i.e., changes to catchability) from observed catch rates such that McKenzie and Parsons, 2012). The core vessels are more likely to have
the trend in standardized CPUE reflects the true relative abundance of characteristics, such as consistent targeting strategies and reporting
the stock component selected by the fishery. In practice, this is accom­ behaviour, that indicate stable catchability. Treating vessel ID as a
plished by building a statistical model that predicts CPUE as a function random effect can be an alternative to defining a core fleet since it avoids
of a set of variables thought to impact local abundance (density) or making arbitrary decisions about threshold rules for the core fleet and
catchability. Recent standardizations mostly use generalized linear expands the number of records available to the analysis (Grüss et al.,
models (GLMs) or an expansion thereof, implemented in the R Statistical 2023a). Care should be taken when defining a ‘core fleet’ as more
Computing system (R Core Team, 2022); this section focuses on tools experienced operators may be able to maintain high catches even if
available within this framework. abundance declines.
There are five key factors to consider when selecting a modelling The relationship between CPUE and continuous covariates can take a
framework which will limit the set of approaches available to the ana­ variety of shapes, including linear, saturating, and dome shaped. In most
lysts: (1) the form of the relationship between CPUE and candidate instances, the relationship can be assumed to be non-linear as a starting
explanatory variables; (2) the probability (error) distribution of the point, as linearity is a strong constraint. Non-linear relationships are best
CPUE; (3) the type of correlation that might be present within the CPUE implemented via splines, which are a useful improvement over poly­
dataset; (4) whether CPUE will be jointly standardized with composition nomials as they are not constrained in the shapes they can take. Poly­
data; and (5) how results are processed to yield an overall index of nomials have other undesirable properties such as the fit in one data
abundance. Often, the analyst will have to prioritize one of these five range being affected by data in other parts of the range (Harrell, 2001;
factors, as different software packages do not always allow the imple­ Magee, 1998). The splines library (e.g., via the ns function) allows GLMs
mentation of all modelling structures jointly. There are also logistical to include splines with a user-specified number of knots, including in
considerations (e.g., computational power) for large datasets that may base-R packages such as stats::glm.
prevent the implementation of some features. A preferred alternative for fitting non-linear relationships is the R
A CPUE standardization will likely consider both categorical and package mgcv which implements GAMs using penalized splines (Wood,
continuous covariates. Vessel identity, bait type, and area are examples 2017). With penalized splines, the analyst specifies a maximum of knots,
of categorical covariates; hooks-between-floats, net length, and sea and the algorithm attempts to maximize fit to the data while minimising
surface temperature are examples of continuous covariates. Continuous spline ‘wiggliness.’ The inclusion of splines has improved the perfor­
covariates can sometimes be implemented as categorical, depending on mance of CPUE standardization models by predicting more realistic
the desired statistical treatment (e.g., Grüss et al., 2019). In the GLM relationships between CPUE and continuous covariates. However, care
framework and its relatives, categorical covariates are typically must still be taken to prevent overfitting or extrapolating far beyond the
modelled as deviates from a ‘baseline’ intercept level. In a conventional range of available data. If the splines are allowed to be too flexible, the
GLM, categorical covariates are estimated as fixed effects, that is, there CPUE standardization model could be capturing noise. In mgcv, the
are no constraints on the value of the coefficient assigned to each level. default optimization method GCV tends to overfit, and REML is rec­
This is the default treatment for a categorical covariate and will be ommended instead (Wood, 2017). The relative penalty on ‘wiggliness’
available in most modelling frameworks. can be controlled via the gamma parameter which scales the effective
An alternative is to fit categorical covariates as random effects. In a sample size. Even with data that are independent and identically
mixed-effect (or hierarchical) framework (e.g., generalized linear mixed distributed (i.i.d.), it is recommended to increase gamma to 1.4 from the
model (GLMM); generalized additive mixed model (GAMM)), categori­ default mgcv value of 1 when using method=GCV (Wood, 2017). The
cal covariates can be fitted as fixed or random effects. With random analyst can also elect to allow for fewer knots, e.g., for one-dimensional
effects, level coefficients are constrained to belong to a distribution splines; three to four knots will be enough to realistically represent most
(usually Gaussian) of the expected values of the coefficients (Thorson relationships between CPUE and a continuous covariate (Grüss et al.,
and Minto, 2015). This results in the ‘shrinkage’ of covariate levels to­ 2019; Roberts et al., 2016). However, restricting the number of knots
wards an overall mean, with more shrinkage for values less informed by will worsen the fit to the data and may increase the number of covariates
data, which are otherwise more likely to be extreme. Mixed effects are retained in the model. The mgcv package implements many other useful
now available in an increasing number of R packages, including nlme, features, such as cyclic (a.k.a. periodic) regression splines (see `?mgcv::
lme4, glmer, glmmTMB, and mgcv (see also Bolker, 2022), and are smooth.terms’) which constrain the effects fitted to the minimum and
straightforward to implement in bespoke Template Model Builder maximum values of a continuous covariate to be equal. This is useful
(TMB) code (see below). when (for example) time of day or month of year are included as

8
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

continuous covariates. CPUE standardization raises several issues. First comes the question of
It is up to the analyst to select a probability distribution for the how exactly space should be considered in a model and whether a
response variable (see Section 5.4). The choice of distributions depends complex model (e.g., an STM) will necessarily perform better than a
firstly on whether the response variable is discrete (e.g., catch occur­ simpler model (e.g., a GAM). Second comes the question of how to factor
rence, catch in number of individuals) or continuous (e.g., catch in in the existence of sub-stocks (i.e., stock components that are distin­
weight, the ratio of catch to a measure of effort). Common modelling guished for management purposes and whose productivities may differ).
frameworks (GLMs, GAMs) can now handle most standard error distri­ Classically, CPUE standardization is performed using a GLM, which
butions, including the normal, log-normal, Gamma, binomial, Poisson, is very often a two-step, delta GLM to account for the presence of many
and negative binomial distribution models. For the Tweedie distribution zeros in the data (Lo et al., 1992; Stefansson, 1996). The simplest way to
(see Section 5.4), an additional R package (Tweedie: Dunn, 2017) may consider space in a GLM consists of dividing the study region into area
be needed, or it can be approximated using a Poisson-linked delta model strata and including the fixed effect of area stratum – a categorical co­
(Thorson, 2018). The Weibull distribution, a positive continuous dis­ variate – in the GLM (e.g., Forrestal et al., 2017). The inclusion of area
tribution that can handle overdispersion in the data, can be imple­ strata in a model aims to account for spatial heterogeneity in stock
mented in the R package survival (Therneau et al., 2022). The R package density, and stock density is assumed to be homogeneous within each
gamlss allows the analyst to fit a diverse set of error distributions area stratum (Bishop, 2006); areas should be sufficiently small and
(including zero-inflated applications of more common distributions and numerous to accommodate strong spatial patterns in CPUE. A time
some that are less commonly available such as the beta distribution) and + area model assumes that the temporal variability is the same in each
specify model structure explicitly for the different distribution param­ area and only the means differ. We refer to this approach as the ‘GLM’
eters (Rigby and Stasinopoulos, 2005). Alternatively, the R package brms approach (Table 2). Often, the GLM employed for CPUE standardization
(see below) allows the analyst to specify custom distributions (e.g., includes a time-area interaction to account for potentially different
Tremblay-Boyer and Neubauer, 2019). The choice of error distribution temporal trends among area strata (e.g., Campbell, 2004; Carruthers
should be validated by inspecting residual diagnostics (See Section 5.6). et al., 2011; Nakano, 1998). The resulting index should be calculated by
Most statistical frameworks assume by default that residuals are i.i.d. weighting the effect for each area by the size of that area. However,
However, correlation between records in fisheries datasets arises from different trends among areas (i.e., the year-area interaction is signifi­
multiple sources. Spatial correlation can be handled implicitly by cant) may imply that a spatial stock assessment model is required.
specifying a two-dimensional spline for longitude and latitude in a GAM, Alternatively, the time-area interaction can be integrated into a GLMM
or explicitly in a geostatistical or spatiotemporal modelling framework as a random effect (Chang, 2003; Maunder and Punt, 2004; Miyabe and
by estimating the Matérn covariance function, for example, using the R Takeuchi, 2003), which treats it as a nuisance parameter; we refer to this
packages VAST (Thorson, 2019a). The R-INLA package (Lindgren and approach as the ‘GLMMint’ approach.
Rue, 2015) can also be used for other geostatistical or spatiotemporal Even though the ‘GLM’ and ’GLMMint’ approaches represent (sim­
modelling applications (e.g., Cosandey-Godin et al., 2015; Pinto et al., ple) ways to consider space in CPUE standardization, they come with the
2019; Zhou et al., 2019) and provides additional options to account for issue of defining the area strata, which can substantially affect model
assumptions about spatial and/or temporal correlation. Correlation performance. Ideally, data are rich enough to define fine-scale spatial
structure between records belonging to the same strata (e.g., fishing trip,
observer) can also be specified with generalized estimating equations (e.
g., Coelho et al., 2020; Peatman and Nicol, 2021) using the R package Table 2
geepack (Højsgaard et al., 2006). Overview of the catch-per-unit-effort (CPUE) standardization approaches with
spatial considerations, along with a selection of R packages mentioned in the
Model fitting can be implemented in a frequentist or Bayesian sta­
text.
tistical framework using maximum likelihood estimation (MLE) or
Markov Chain Monte Carlo (MCMC), respectively. Most GLM/GAM tools Approach Overview R packages
are implemented via MLE or some derivation thereof. TMB (library GLM Method using generalized linear stats::glm, mgcv, brms,
TMB: Kristensen et al., 2016) can be employed for large datasets as models (GLMs) that integrate fixed TMB
year and area effects.
model fitting will be considerably faster due to automatic differentiation
GLMMint Method using generalized linear TMB, brms, nlme,
and Laplace approximation; but the analyst will have to specify the mixed models (GLMMs) that integrate lme4, glmer,
CPUE model manually (e.g., model matrix, likelihood function) in the fixed year and area effects and a glmmTMB, mgcv
C+ + code embedded in TMB files, unless appropriate code is already random year-area interaction term.
available (e.g., VAST; Thorson, 2019a). An alternative is to fit CPUE in a Spatial GAM Method using generalized additive mgcv, brms, TMB
models (GAMs) that integrate an
Bayesian framework using MCMC with the R package brms (Bürkner,
interaction term between longitude
2017). Key advantages of this approach include (1) the option to specify and latitude representing spatial
priors to inform or constrain effects, and (2) a more intuitive and better variation (long-term latent variation)
integrated estimate of index uncertainty. One downside is that model at a broad spatial scale.
fitting will be slower, especially for large datasets with complicated Spatiotemporal Method using GAMs that integrate an mgcv, brms, TMB, R-
GAM interaction term between longitude, INLA
covariate structures and may be less stable if some relationships are latitude and year representing
poorly informed by the available data. spatiotemporal variation (latent
Machine learning methods such as artificial neural networks variation that changes over time) at a
(Maunder and Hinton, 2006), support vector machines (Li et al., 2015), broad spatial scale, in addition or in
lieu of an interaction term between
regression trees (Watters and Deriso, 2000), and RFs (Chambers and
longitude and latitude.
Hoyle, 2015; Li et al., 2015) have been used to model CPUE but rela­ STM Method using models that explicitly TMB, R-INLA, VAST,
tively infrequently to date. They can achieve high predictive perfor­ specify a correlation or semi-variance sdmTMB
mance due to their flexibility, but results may be difficult to interpret, function for latent variables
and they are prone to overfitting. representing unmeasured processes
that vary over space and time. These
can be specified with similar structure
5.2. Spatial considerations to a spatio-temporal GAM but also
allow more flexibility for explicit
All fish populations and fisheries exhibit spatial structure to some movement, size/age structure, or
other ecological mechanisms.
extent. Nevertheless, the introduction of spatial considerations into

9
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

strata that do not constrain estimates. For example, tuna longline fishery to apply for exploration and may have the flexibility to fit model
CPUE analyses usually define spatial strata as 5◦ cells (Anon, 2013). structures that are unavailable in the VAST framework. Thus, we suggest
However, some degree of aggregation is often required and there are that the analyst should not assume a priori that STMs will outperform
several approaches that can be used. In the ‘ad hoc’ approach, area strata the spatial or the spatiotemporal GAM approach, or even the GLMMint
are the management areas employed in the study region or are based on approach, in their case study. Instead, the analyst should ideally carry
environmental layers such as bottom depth contours (Forrestal et al., out a simulation experiment to better understand the capabilities of
2019; Huang et al., 2020; Huang et al., 2007). In the ‘binary recursive STMs in their case study under different scenarios. Such simulation ex­
partitioning approach,’ an algorithm is used to divide the study region periments permit evaluation of the accuracy, error, and confidence in­
into several area strata in a sequential and recursive manner (Ichino­ terval coverage of STMs compared to simpler approaches. When
kawa and Brodziak, 2010). Finally, in the ‘spatial clustering’ approach, a operating models are not available to generate simulated data, the fish
k-medoids algorithm is employed to partition a spatial grid covering the abundances needed for the simulation experiment can be obtained by
study region into several area strata based on the proximity and mean fitting an STM without any vessel effect or other catchability effects, as
value of CPUE in each spatial grid cell (Ono et al., 2015). Using a in Thorson et al. (2015) and Hsu et al. (2022).
simulation experiment based on Pacific saury (Cololabis saira) data, Hsu When the stock of interest is made of several sub-stocks, analysts may
et al. (2022) found that the GLMM relying on the ad hoc approach had develop a separate CPUE standardization model for each one (McKenzie
the poorest performance and that the GLMMs relying on the spatial and Parsons, 2012). This approach is not necessarily warranted with
clustering approach had the best performance. Moreover, regarding the STMs. Specifically, the fine-scale spatial areas occupied by each
spatial clustering approach, Hsu et al. (2022) found that, under prefer­ sub-stock can be identified in the STM input, and the STM can then
ential sampling (a classical situation with CPUE data), assigning equal compute an abundance index for each sub-stock (Grüss et al., 2023a;
weights to spatial proximity and mean CPUE values in the k-medoids Thorson, 2022). Even in the presence of stock-structure, a single STM
analysis was preferable to giving more weight to mean CPUE values. that appropriately weights composition data spatially by the estimated
A more flexible approach than the ‘GLM’ and ’GLMMint’ approaches CPUE may be able to implicitly adjust for any spatial structure in the
to considering space in CPUE standardization is the ‘spatial GAM’ stock and the fishery (Maunder et al., 2020). However, further research
approach, which consists of fitting a GAM including an interaction term is needed to explore this issue.
between longitude and latitude (Braccini et al., 2021; Grüss et al., 2019; If covariate effects or data availability vary spatially or through the
McKechnie et al., 2013). The spatial interaction term (e.g., a tensor time series, it may be difficult to develop a single STM that includes all
product smooth) depicts spatial variation (long-term latent variation) at influential covariate effects. It is important to include relevant covariate
a broad scale (Denis et al., 2002; Grüss et al., 2021). The spatial GAM effects since models that fail to include them will often produce biased
approach can be extended to a ‘spatiotemporal GAM’ approach via the indices of abundance. In general, analysts should explore multiple
inclusion of an interaction term between longitude, latitude, and time, modelling approaches to ensure that the resulting indices are consistent
which represents spatiotemporal variation (latent variation that changes with the best available information about abundance trends.
over time) at a broad scale (Hoyle, 2020; Zhou et al., 2019). By ac­ Interestingly, instead of being employed for standardizing CPUE,
counting for spatial and/or spatiotemporal variation at a broad scale, STMs can instead be used to identify stock structuring (Lindegren et al.,
such GAMs can generate more stable predictions for spatial areas where 2022), investigate existing stock structure hypotheses (Grüss et al.,
data are sparse (Hoyle, 2020; McKechnie et al., 2013). Using a simula­ 2023a), or evaluate the impacts of modifying sub-stock spatial bound­
tion experiment based on blue marlin data from the Atlantic Ocean, aries. They can also be applied to estimate relative population scale
Grüss et al. (2019) found that spatial GAMs tended to outperform across regions for use in multi-region stock assessments (Ducharme-­
simpler CPUE standardization models. Barth et al., 2020; Hoyle and Langley, 2020). Inclusion of density
STMs constitute an even more flexible approach for CPUE stan­ covariates when using an STM to develop relative population scale
dardization by depicting spatial and spatiotemporal variation at a very across regions (i.e., ‘regional scaling’) should be carefully considered
fine scale, thereby yielding very precise estimates (Anderson et al., given that they can influence predicted values in poorly sampled areas of
2022; Shelton et al., 2014; Thorson et al., 2015). Over the recent years, the model domain (Ducharme-Barth et al., 2020). Given the influence of
STMs implemented with R package VAST (Thorson, 2019a) have been ‘regional scaling’ in multi-region, spatially explicit stock assessments,
increasingly employed to standardize CPUE data (e.g., Cao et al., 2017; we recommend developing and exploring alternative CPUE index sce­
Ducharme-Barth et al., 2022; Grüss et al., 2019; Kanamori et al., 2021; narios with different implied ‘regional scaling’.
Xu et al., 2019). Because STMs represent latent spatial and spatiotem­ When choosing the sampling frame for the CPUE analysis, analysts
poral variation at a very fine scale, the inclusion of habitat variables that must consider whether to include regions at the edge of the fishery that
are inherently spatial or spatiotemporal (e.g., sea surface temperature) are fished infrequently. It may be uncertain whether these regions are
in these models generally does not improve their predictive capabilities unfished due to low CPUE or for unrelated reasons (e.g., they are too far
(e.g., Han et al., 2021; Hsu et al., 2022; Thorson, 2015). Using a simu­ from port). They may also be subject to spatial changes in stock distri­
lation experiment based on skipjack tuna (Katsuwonus pelamis) data for bution due to environmental events. Including such regions in the
the western and central Pacific, Ducharme-Barth et al. (2022) found that modelling frame will increase uncertainty, as the abundance in such
including local environmental covariates or regional oceanographic regions will be inferred from limited information. One option is to limit
indices in STMs did not meaningfully improve model performance the spatial domain to a ‘core’ fishery defined as those regions where
beyond what was achieved using spatiotemporal random effects, and most of the catch is taken, or that are fished for a majority of the
even degraded model performance in some cases. However, we note that modelling period (Campbell, 2015; Grüss et al., 2023b; Xu and
other STM case studies have suggested small but important improve­ Lennert-Cody, 2022). For example, Campbell (2015) limited the spatial
ments resulting from including density covariates (Thorson, 2019b). domain to those one-degree squares where the total catch of swordfish
They may have the most influence when predicting into areas with poor was greater than 500 fish over the period modelled, while Xu and
sampling coverage. Lennert-Cody (2022) defined a core fishing ground for skipjack for the
The simulation experiment conducted by Grüss et al. (2019) indi­ floating object and unassociated fisheries in the eastern Pacific Ocean as
cated that, overall, the STM approach implemented using VAST per­ all one-degree squares with at least 11 and 6 years of CPUE data between
formed better than simpler approaches, including the spatial GAM 2000 and 2021, respectively. Sensitivity of the resulting abundance
approach. However, in some instances, the VAST and spatial GAM ap­ index to different spatial domain definitions should be considered, as
proaches performed similarly or the spatial GAM approach performed limiting the analysis only to a core region could mask hyperstability, or
slightly better (Grüss et al., 2019). Spatiotemporal GAMs are often easier changes in the species’ range.

10
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

Non-stationarity of the spatiotemporal correlation also needs to be contained in the catch and effort data.
considered. Rates of spatiotemporal correlation are unlikely to be con­ Subsetting approaches aim to select records where effort was likely
stant either within habitat types or between them. However, approaches allocated towards the species of interest (Hoyle et al., 2022; Stephens
that allow nonstationary spatiotemporal correlation are rare, and it may and MacCall, 2004). These include using catch proportions by species to
be more practical to do separate analyses for different areas (i.e., habitat determine a threshold for subsetting the data (Biseau, 1998; Helle et al.,
types) and subsequently compare the estimates of correlation structure 2015; Klaer and Smith, 2012) to only include records from a core area or
to determine whether combining them with a stationary correlation specific fishing season (Hoyle et al., 2022). However, such subsetting
structure is appropriate. If not, then the results of the separate analyses can be sensitive to the subjective choice of the threshold and associated
may need to be appropriately combined. with risks introducing hyperstability into the standardized index. For
With fishery-dependent data, there is a growing need to account for example, the effect of subsetting to non-zero catches can be illustrated
preferential sampling (the likely correlation between sampling location via a delta-lognormal model in which the binomial distribution model is
and abundance) which can bias indices that fail to account for it used to estimate the encounter probability, and the lognormal model is
(Ducharme-Barth et al., 2022). Preferential sampling can be accommo­ employed to estimate the scale of positive catch rates. For any situation
dated using joint models for sampling intensity and density (Alglave in which the estimated indices from both model components show a
et al., 2022; Conn et al., 2017; Pennino et al., 2019; Rufener et al., 2021) unidirectional trend, an index based on a subset of positive CPUE alone
which make the strong assumption that abundance should approach would be hyperstable relative to standardized CPUE from the
zero in un-sampled areas. Research is needed to develop approaches for delta-lognormal model. Unsurprisingly, simulation experiments yield
modelling preferential sampling in a CPUE standardization framework biased abundance trends when zeros are excluded, with the exception of
either as an inhomogeneous Poisson process (Diggle et al., 2010) or those species that are ubiquitous and occur in almost all (> 95%) of the
using random-utility models to consider fisher sampling location due to CPUE records (Langley, 2019). This bias is likely increased if thresholds
economic or regulatory factors (Girardin et al., 2017). are set higher (e.g., 30–50% of the catch weight) and most severe for less
Finally, the spatial dynamics of a fishery may change significantly abundant species given that encounter probabilities below 30% become
over time (for example, effort may contract spatially in response to approximately proportional to abundance (Kerwath et al., 2019).
economic or management factors while the stock may spatially contract Similarly, a species with declining abundance may contract towards its
in response to overfishing) and relative abundance indices based on core area where CPUE may then remain hyperstable compared to the
catch and effort data can become biased unless consideration is given to overall decline (Harley et al., 2001; Thorson et al., 2016c). Optimally,
such changes. Campbell (2016) developed a general framework for the target species would be known for a random subset of the data, and
indices of stock abundance which uses alternative hypotheses to this could be used to determine the best model to predict targeting for
consider uncertainty about how to structure the analysis, particularly the trips with unknown target.
concerning the spatiotemporal dynamics of the fishery. Stephens and MacCall (2004) applied a logistic regression of multi­
species presence–absence information to subset trip records to locations
5.3. Multispecies targeting (habitats) where the species under assessment was likely to be present
but may not have been caught (‘true’ zeros) and locations where the
CPUE standardization aims to account for the non-random catch­ species was unlikely to occur (‘false’ zeros). This approach therefore
ability change caused by targeted fishing operations, which typically aims to retain ‘true’ zeros in the abundance trend. By design, this
seek to optimize their profits. This is further complicated in multispecies approach is most suitable for species that co-occur within multispecies
fisheries in which fishers can make discrete operational choices to in­ assemblages, where the mere presence-absence of co-occurring species
crease the catchability of specific species or species complexes. Allo­ can provide a strong predictor of habitat suitability. Another approach is
cating targeted effort to one species over another can be related to the to identify ‘indicative’ vessels based on vessel characteristics and catch
choice of fishing-ground, habitat-type, fishing-technique, or the way the history of the species of interest (Helle et al., 2015; Punt et al., 2000).
gear is deployed (Palmer et al., 2009; Pelletier and Ferraris, 2000; This approach requires a detailed understanding of the fleet and assumes
Winker et al., 2013). In the best case, unaccounted variation in targeting that the indicative vessels employ consistent fishing tactics. The selec­
only inflates the uncertainty in the index of abundance, but large tion criteria for indicative vessels rely on a degree of subjectivity which
short-term shifts or long-term trends in targeting can systematically may affect the CPUE indices (Helle et al., 2015). Using indicative vessels
change catchability and, therefore, severely bias abundance indices. The alone may be insufficient for CPUE standardization if vessels change
effects on CPUE of variability in targeting must be accounted for to targeting over time. For example, skipper behaviour may be more
stabilise catchability and estimate reliable abundance indices for influential than the vessel effect (Palmer et al., 2009), but economic (e.
multispecies fisheries. Targeting may include the choice to fish at the g., fish price, market demands, fuel costs) or regulatory factors (e.g.,
season and locations in which high catches of a sought-after target quota limitations, area closures, or by-catch restrictions) may also lead
species are expected. Such target changes can be effectively accounted to changes in targeting (Abbott et al., 2015).
for with adequate STMs (Thorson et al., 2016b). However, other oper­ As an alternative to subsetting, several approaches have been pro­
ational adjustments are non-spatial (e.g., bait type, fishing depth, gear posed that use covariates to account for variations in fishing tactics. One
deployment) or occur at much finer spatiotemporal levels (e.g., dynamic approach is to use the catch rates of alternative target or bycatch species
temperature gradients, time of day, or habitat features) than are re­ as covariates to correct for the effort directed away from the target
ported. For example, many small-scale and recreational hook and line species or species under consideration (Glazer and Butterworth, 2002;
fisheries report spatial information at the coarse scale of a single loca­ Su et al., 2008). Although the catch rates of alternative species do not
tion, whereas multiple fishing locations and fishing techniques may hold direct information about the catch of the species of interest, the
have been employed during the same trip. Information about targeting information in the predictor variables derived from these covariates is
strategies may be unreported, or reported infrequently and inconsis­ not entirely independent of the response CPUE and may have unpre­
tently among skippers, locations, and time periods. In other words, the dictable impacts on the standardized CPUE trends. Another approach is
factors that might indicate the target/fishing strategy are latent vari­ to derive categorical covariates from ranked catch ratios of two species
ables. Consequently, finer scale variations in targeting are often unob­ as an indicator of targeting strategy. These catch ratios, which often
servable and are hereafter referred to ‘fishing tactics’ (Okamura et al., include the species of interest, are then grouped by percentile frequency
2017; Thorson et al., 2016b; Winker et al., 2013). The principal idea in descending order, either over the entire time series (e.g., Mejuto et al.,
behind the various approaches employed to account for these latent 2009) or by year (e.g., Hiraoka et al., 2012). Each percentile group
changes in fishing tactics is to make use of the multispecies information represents a factor level of the categorical variable. However, if the

11
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

ratios are grouped over the entire time series, then there will inevitably for a deterministic proof-of-concept scenario under somewhat idealized
be confounding effects between changes in abundance and the catch conditions with two species and two fishing tactics for which previous
ratio grouping. If the abundance of the species of interest changes over approaches failed (Okamura et al., 2017). However, there is a lack of
time, the true trend abundance tends to be removed from the CPUE follow-up case studies on the efficiency of the DRM in the presence of
index, but if the other species changes, this can introduce a biased trend zeros and over-dispersion, which are common features of real-world
into the abundance index for the species of interest (Chang et al., 2011). datasets. Moreover, the ‘spatial dynamic factor analysis’ (SDFA)
Further, such bias is likely to be aggravated if individual CPUE records (Thorson et al., 2016b), which is now embedded in the R package VAST,
are aggregated. However, if the ratios are grouped by year, it can be presents a statistically coherent method for simultaneously estimating
demonstrated that the catch ratio covariate has no influence on the spatial-temporal variation, random vessel effect, fishing tactics, and
year-effect of interest and is, thus, completely ineffective for removing relative fish abundance for multispecies species within a single model
targeting-induced variation from the index of abundance (Hoyle et al., where the latent fishing tactic effect is estimated from remaining un­
2014b). explained residual correlations among the multispecies catch rates.
If discrete fishing tactics result in distinctive species composition in Simulations demonstrated that SDFA performs well when targeting is
the catch, one can employ clustering techniques to categorize multi­ linked to the spatial allocation of fishing effort. However, the fishing
species CPUE records into groups with similar catch compositions (He tactics model showed limited ability to correct for targeting effects in
et al., 1997). The identified clusters are assumed to represent fishing cases where spatial reporting was aggregated to a coarser resolution
tactics which may be treated as categorical variables in the standardi­ than the ‘true’ simulated dynamics or if spatial information was absent.
zation model to adjust for differences in catchability associated with Further research is needed to better understand the data requirements
each cluster (He et al., 1997). This two-stage approach has been widely for SDFA and to evaluate the impacts of missing covariates on DRM’s
applied for tuna CPUE standardization of large pelagic longline fisheries ability to account for fishing tactics.
(Carvalho et al., 2010; Hoyle et al., 2022; Hoyle et al., 2015b). A related
two-stage ordination approach is the ‘Direct Principal Component’
(DPC) procedure (Winker et al., 2013; Winker et al., 2014), which uses 5.4. Error distributions
continuous principal component scores (PCS), derived from a Principal
Component Analysis (PCA) of the catch composition data, as either As described in the Maunder and Punt Sections 2.1, 3.1, 3.3 and 4.3)
linear or nonlinear predictor variables to adjust for the effect of varia­ (2004) review, it remains important to select an error distribution and a
tions in fishing tactics. The DPC method relaxes the assumption of modelling approach that match the structure of the data. The analyst
fishing tactics being discrete and simulations indicated that the DPC can should first determine an appropriate class of error distribution for the
also perform adequately if CPUE records originated from mixtures of data. An appropriate error structure is needed to accurately represent
fishing tactics (Winker et al., 2014). Both ordination approaches can variance around the standardized index and to determine the trend
provide valuable insights into the targeting dynamics. For example, (Dick, 2004) and scale (Thorson et al., 2021) of the standardized index.
clusters or PCA species loadings can be visually explored at different For discrete observations (e.g., catches recorded as counts), the
spatiotemporal scales (e.g., seasons, area) or by individual vessels. Poisson distribution may be suitable if the variance is approximately
However, several statistical caveats require careful consideration. If the equal to the mean although, in practice, this is rarely the case. If variance
species of interest represents a dominant component of a fishing tactic, it is larger than the mean, then the negative binomial distribution may be
typically needs be included in the catch composition used to identify appropriate (Walsh and Brodziak, 2015), while the flexible
targeting, and the information contained in the resultant predictor Conway-Maxwell-Poisson distribution can account for either over- or
variables composition is not entirely independent from the response in under-dispersed data (Lynch et al., 2014). Additionally, if data are
the standardization model and is, thus, not strictly orthogonal. There­ collected discretely, then effort can be included as an offset (Maunder
fore, additional transformations (e.g., arcsine-square-root or fourth and Punt, 2004), or the numerical catch can be converted to catch rate
root) of the species composition data have been recommended (Camp­ (e.g., number of fish caught per 1000 hooks), which can be modelled
bell et al., 2017; He et al., 1997; Winker et al., 2014). Simulation testing continuously. If the numbers of individuals caught per observation are
has indicated that this confounding effect was not a major concern for large, then they can be approximated as continuous variables (Maunder
multispecies scenarios of moderate complexity (He et al., 1997; Winker and Punt, 2004).
et al., 2014). A drawback of this non-independence is that model se­ Continuous observations (e.g., catch per unit effort or catch recorded
lection criteria (e.g., AIC, Bayesian Information Criterion (BIC)) are as biomass) are typically right-skewed; an appropriate distribution
unlikely to be appropriate since they select too many non-meaningful choice can be informed using Taylor’s power law (Taylor, 1961),
targeting signatures, leading to confounding and over-precision in the var(Y) = αμpY . If the variance is proportional to the square of the mean
abundance trend (Winker et al., 2014). This makes it generally chal­ (p = 2), then the lognormal or Gamma distribution may be appropriate
lenging to identify meaningful interactions between targeting effects and, if proportional to the cube of the mean (p = 3), the data might be
and other covariates. Considering that several subjective decisions must fitted using the inverse Gaussian distribution model (Dick, 2004).
be made (Campbell et al., 2017), the identified targeting signatures However, a recent study tested the different distribution models in STMs
should be consistent with prior knowledge that distinct targeting prac­ and found that the Gamma and Tweedie distributions provided the best
tices exist. Otherwise, there is a risk of mistaking spatiotemporal performance for estimating index scale, followed by the lognormal, and
abundance patterns for targeting, the inverse Gaussian performed worst (Thorson et al., 2021). Alternative
Recently, two approaches have been proposed to overcome the distributions may be distinguished using residual analysis,
confounding effects of the approaches discussed above (Okamura et al., mean-variance plots, and Akaike’s Information Criterion (AIC) (Dick,
2017; Thorson et al., 2016b). Both infer the latent fishing tactic from the 2004).
residual structure. Okamura et al. (2017) proposed the two-stage When selecting the error distribution, the proportions of zero ob­
‘directed residual mixture’ (DRM) approach which entails first fitting servations in the data should be scrutinized. If there are more zeros than
a standardization model to each species with all covariates other than predicted by the distribution, they may be dealt with using a mixture
fishing tactics, then applying a Gaussian mixture model to estimate modelling approach. A discrete variable with excess zeros is typically
discrete latent factors for fishing tactics from the resulting mixing pro­ fitted with a zero-inflated model which uses a separate likelihood
portions of residuals. In the third stage, the estimated components are component to account for the extra probability of a zero observation
included as additional factorial covariates in a final standardization (Lambert, 1992; Minami et al., 2007; Walsh and Brodziak, 2015; Zuur
mode for the individual species. DRM provided an unbiased estimator et al., 2012). Given a continuous response variable, excess zeros can be

12
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

accommodated using a two-stage hurdle or delta modelling approach uncertainty in the abundance index than predicted by the observation
where catch-rate is modelled conditional on a positive encounter (Lo error. This variability cannot be measured directly but can be estimated
et al., 1992). The Tweedie distribution can take the shape of several indirectly with stock assessment models (Francis et al., 2003). This
distributions in the exponential family depending on the value of p. variation is often estimated or assumed to be larger than the observation
When p is in the range (1 < p < 2), then the Tweedie is equivalent to a error in the time effects (Maunder and Punt, 2004), particularly for large
compound Poisson-Gamma distribution which can simultaneously ac­ industrial fisheries with low observation error due to high effort data
count for zero observations and positive catch (Shono, 2008a). Applying sample sizes.
either a Poisson-link delta-gamma or Tweedie model structure may be a
sensible default option (Thorson et al., 2021).
In recent years, research has focused on more appropriately model­ 5.6. Diagnostics
ling and accounting for correlation in the data generation process. These
approaches should be incorporated into routine CPUE standardization Traditional model validation diagnostics such as residual analysis are
analyses where appropriate. ‘Poisson-link delta models’ represent a important for determining if the analyst has specified the model
computationally efficient alternative to the compound Poisson-gamma correctly, treated zero observations appropriately, and selected the
distribution and can explicitly account for potential positive correla­ proper error distribution (i.e., the model assumptions are not violated).
tion between encounter rate and positive catch (Thorson, 2018). However, non-Gaussian error structures, mixture (e.g., delta or zero
Newly developed flexible modelling approaches (e.g., VAST; Thor­ inflated) distributions, and/or mixed-effects frameworks commonly
son 2019) more readily model multivariate data within a CPUE stan­ applied in contemporary CPUE standardization analyses complicate
dardization framework. Multivariate approaches can account for using traditional diagnostic approaches. One solution is to define re­
correlation between modelled categories such as species (Thorson and siduals as probability-integral-transform (PIT) residuals (Warton et al.,
Barnett, 2017; Thorson et al., 2016a); or age/length/sex/stage cate­ 2017). PIT-residuals are generated in a ‘model-free’ bootstrap by
gories (Kai et al., 2017; Maunder et al., 2020). Accounting for correla­ comparing observations to a distribution of predicted values for the
tion between species can improve predictions of spatial density for rare given observation generated from the fitted model. Prior to calculating
species (Thorson and Barnett, 2017), and accounting for correlation the PIT-residual, a simple check of the distribution of the observations
across size classes can allow for size-class specific indices (Kai et al., against the distribution of the bootstrapped/predicted values (e.g.,
2017) or the standardization of composition data (Maunder et al., 2020; Parker et al., 2017) can be used to identify if the distributional as­
Thorson and Haltuch, 2019). sumptions are being met (e.g., are coverage levels similar?). R package
DHARMa (Hartig, 2020) is useful for calculating PIT-residuals and has
5.5. Uncertainty estimation several tests for assessing if observations match the distributional as­
sumptions of the model, outlier detection, identification of over- or
The uncertainty associated with the CPUE indices used in a stock under-dispersion, and zero-inflation. In a mixed effects framework, if
assessment includes observation error associated with the estimated models are solved using the Laplace approximation to integrate across
time effects in the index and process error associated with variation in the random effects (e.g., TMB; Kristensen et al., 2016), the reduced
catchability (the relationship between the time effect and abundance) Laplace approximation for strongly non-linear models may limit the
(Francis et al., 2003). utility of PIT-style residuals (Thygesen et al., 2017). Accuracy of the
Observation error in the temporal effects can be estimated as part of Laplace approximation can be determined by refitting the model to
the model fitting process and constructing the index, as discussed in simulated data (Rufener et al., 2021; Thygesen et al., 2017). Alterna­
Section 5.8. Many GLM analyses estimate categorical time effects rela­ tively, residuals can be calculated using the posterior estimate from
tive to the base time (e.g., the first year in the time series). These can be MCMC samples (Rufener et al., 2021) or by using one-step predictions
recalculated as canonical confidence intervals to associate uncertainty (Thygesen et al., 2017). One-step prediction residuals are a useful tool
with all temporal effects (Francis 1999). Observation error for two-step for validating modelʻs fit to sequential data such as CPUE analyses,
delta or hurdle models can be estimated using parametric (Shono, where an observation Yt is compared to a prediction in time t, Ŷ t|t− 1 ,
2008b) or bootstrap (Ichinokawa and Takeuchi, 2012) methods, and conditioned on data up to t − 1.
subsequently applied within the predict-then-aggregate process. Alter­ In addition to model validation, diagnostics can be used to develop
natively, uncertainty can be characterised by draws from the posterior intuition on how the CPUE standardization model is transforming the
predictive distribution if the model is fitted in a Bayesian framework (e. nominal index. Step plots (sequentially plotting how the index changes
g., using the R package brms; (Bürkner, 2017)). with each additional covariate or model component) show incremental
However, standardization usually underestimates observation error changes to the index (Bentley et al., 2012). However, if there are in­
in the temporal effects because it does not fully account for de­ teractions between covariates, then their impact could be masked within
pendencies in the input data (Hoyle et al., 2014b). Dependencies occur a step plot. Influence and coefficient-distribution-influence (CDI; Bent­
among consecutive sets by a vessel and among sets by vessels from the ley et al., 2012) plots are useful in combination with step plots. CDI plots
same company which may communicate with one another. There is also quantify the relative influence of each covariate on the final fitted model
overdispersion due to both aggregation of fish and the unavailability to by incorporating both the estimated coefficients for individual cova­
the analyst of important factors affecting catch rates, such as local riates and the sampling across covariate levels each year. Influence and
environmental factors that affect fish distribution. Some of this over­ CDI plots can be readily created for fixed effect GLMs using the influ
dispersion can be accounted for by including random effects in the CPUE package (https://github.com/trophia/influ) in R. Hsu et al. (2022)
standardization model (Rufener et al., 2021; Thorson and Minto, 2015; extended the influ package to STMs implemented with the R package
Xu et al., 2019). Many sources of catchability variation are not amenable VAST. Applied in a spatial or spatiotemporal modelling framework, a
to standardization (Wilberg et al., 2009). As a result, uncertainty esti­ large annual influence value for the year-area interaction term could
mates from CPUE standardization models (observation error) are often indicate a misspecification of the spatial structure that failed to account
seen as, at best, useful for suggesting relative uncertainty among year for a large shift in spatial sampling (Hsu et al., 2022). Counterfactual
effects. They can also be useful as an estimate of the minimum uncer­ analyses (Hansell et al., 2022; Pearl, 2009) can also be employed to
tainty, and to determine relative uncertainty between indices. evaluate the impact of key structural assumptions (e.g., fitting a model
Process errors occur when the average catchability across the fleet with spatial random effects held at zero and comparing the resultant
varies between time intervals. Effectively, the relationship changes be­ index with a model freely estimating spatial effects in order to identify
tween the abundance index and the true abundance leading to more their impact on the standardized index).

13
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

Additional diagnostics can be applied to specifically interrogate independence. Operational data usually represent a time series of fishing
STMs. Calculated residuals should be evaluated for temporal, spatial (i. events by the same vessels, and there may be information sharing be­
e., Moran’s I test), and spatiotemporal patterns. In addition to being tween vessels. The ‘intra-vessel’, ‘intra-trip’, or other correlation can be
useful for evaluating model performance, cross-validation can be used to represented by treating that factor (e.g., vessel or trip) as a random ef­
identify spatial outliers (Conn et al., 2018; Marshall and Spiegelhalter, fect. Such a treatment is appropriate to account for the reduction in
2003). Spatial and spatiotemporal random effects are often assumed to effective sample size resulting from simple correlation in the observa­
be normally distributed with mean-zero. The estimated random effects tions. However, random effect distributional assumptions may be inac­
should be checked to ensure that they match the assumed distributions. curate given, for example, trends in fishing power or targeting through
Lastly, STMs can predict abundance (or density) into un-sampled areas. time.
Predictions in un-sampled areas will be influenced by adjacent obser­ Operational datasets can have very high sample sizes, and that
vations given the modelled spatial/spatiotemporal covariance structure, combined with lack of independence can make almost any variable
and by any modelled relationships with density covariates. These pre­ statistically significant based on likelihood, even if it has negligible in­
dictions should be scrutinized for consistency with ‘common sense’ ex­ fluence on the year effect. Since the goal is to obtain an index of abun­
pectations of what abundance should be given understanding of the dance, including a covariate that neither explains much deviance nor
species biology and fishery dynamics. Departures from expectations may influences the index of abundance is not usually a problem. However,
indicate the need to change how the model makes predictions into alternative data selection criteria are often used to help develop models
unfished areas, such as by applying a preferential sampling model or that are simple, manageable, and fast to compute.
changing other aspects of the model structure to account for unfished When the objective is accurate prediction for the year effect, a pri­
areas caused by economic or regulatory factors (e.g., with the use of a mary consideration for variable selection is influence. Influence is
random utility model). affected by a combination of the extent to which the covariate affects the
Hinton and Maunder (2004) propose an ‘omnibus test’ that simply response and the extent to which its value changes through time. For
compares the total likelihood from the stock assessment model for all example, a variable that affects catch rates but is completely balanced
data types combined and uses one index of abundance versus using an (such as moon phase in many cases) may have no influence on the
alternative index of abundance. This test measures which index of resulting abundance index. Such variables can be useful nonetheless if
abundance is most consistent with the stock assessment model and the they improve precision and help estimate other effects. They are also
other data. We recommend further research regarding omnibus-tests useful for understanding the biology of the species and the nature of the
when specifying CPUE standardization models. fishery, an important benefit of modelling CPUE.
One commonly employed criterion is the proportion of deviance
5.7. Model selection explained by a variable (R2). This criterion is useful because it is related
to influence and robust to lack of independence, but the appropriate
Model selection is an important part of the analysis process. Analysts threshold level will be case-dependent. For example, covariates will
often employ automated procedures such as stepwise forward or back­ explain a higher proportion of residual variance when data are more
ward selection algorithms, although these can be problematic (Sribney, aggregated, allowing the threshold to be set at a higher level. However,
1996; Wiegand, 2010). Throwing all covariates into an automated se­ conditional R2 (and consistent AIC) showed inconsistent model selection
lection procedure is unlikely to identify the most appropriate model and in the simulation experiment conducted by Hsu et al. (2022).
neglects an opportunity to generate useful understanding. Henderson Sometimes it is important to retain covariates that do not meet se­
and Velleman (1981) noted that automated techniques often hide lection thresholds because it is clear a priori that they will affect CPUE.
important features of the data from the analyst and provided the axiom For example, catch rates almost always vary consistently between ves­
that “the data analyst knows more than the computer.” Modelling data sels, and the vessels in the fleet usually change through time. Analysts
interactively can change understanding and lead to different inferences. may, therefore, assume that catchability varies among vessels even if it
The analyst should first consider the objective of their CPUE stan­ does not reach the model selection threshold and should confirm that
dardization. When generating an index rather than making inferences vessel ID is not influential before dropping it from a model.
about the covariates themselves, the goal is to correct for covariate in­ Some tools have internal automatic variable selection, such as the
fluence and avoid potential bias in the index (i.e., covariates that explain GAMs implemented with R package mgcv, which can automatically
annual variation in abundance should not be used to explain catch­ choose the degrees of freedom for a spline and can be configured to
ability). The focus on prediction rather than estimation has implications shrink effects to zero simultaneously without the need for a stepwise
for model selection, making AIC an appropriate tool if statistical as­ process (Marra and Wood, 2011). Model selection in R package mgcv is
sumptions are met (Aho et al., 2014; Akaike, 1973). CPUE analyses for based on prediction error criteria or likelihood-based methods, with the
understanding are also important for stock assessment but have different likelihood-based methods less prone to local minima (Wood, 2017).
objectives and may use different analytical approaches.
Model selection and inference based on information criteria are well- 5.8. Assembling an index from a fitted model
established with clear guidelines available (e.g., Burnham and Ander­
son, 2004). These include identifying plausible predictor variables and Previous reviews of CPUE standardization typically assumed that the
combinations thereof, evaluating model performance, and averaging index is constructed by first fitting a statistical model to available data
across plausible models. Evaluating model performance should include (see Section 5.1) and then extracting a coefficient that represents the
considering the goodness of fit of the selected models (Mac Nally et al., partial effect of year to use as the index; we call this the ‘year-effect as
2018). ’index’ method in the following. For example, Maunder and Punt (2004)
The analyst should determine whether a parameter is likely to affect state: “Most methods used to standardize catch and effort data estimate a
CPUE, a priori, and whether it affects density or catchability. Starting year effect on which an index of abundance can be based.” However, the
with a set of plausible models saves analysis time and reduces the risk of ‘year-effect as index’ method has many limitations. Most generally,
overfitting (a.k.a. data dredging or p-hacking). treating the partial effect of year as the index does not include the effects
Predictive cross-validation methods can be useful for selecting be­ of other covariates that vary among years. One interpretation of this
tween models with different structures (e.g., Charsley et al., 2022; practice is that all the other covariates are implicitly treated as ‘catch­
Maunder and Hinton, 2006; Shono, 2008a). They can be ability covariates’, and their effect is filtered out when treating the year
time-consuming but are very flexible and broadly applicable. effect as the abundance index.
Model selection based on likelihood is affected by lack of One special case of concern arises when including a year x area

14
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

interaction as a predictor variable in the CPUE-standardization model. predict-then-aggregate method is required.

Maunder and Punt (2004) dealt with this topic in detail and outlined a For a response variable that was transformed before fitting the
few alternative treatments, including averaging across year x area esti­ model, predictions must be back transformed before aggregating. For
mates or fitting separate models to different areas. However, research two-stage approaches such as hurdle models, sampling unit predictions
since then has converged upon a generic approach which we here call are obtained by combining predictions from both stages before
the ‘predict-then-aggregate’ method for index construction (Walters, aggregating.
2003 provides a good foundation for the method). This approach is For a hurdle model with the binomial component fitted using a logit
based on sampling theory and involves the following steps. transformation, the reference levels of the catchability covariates affect
the predicted annual probabilities of non-zero catch. This in turn affects
1. Explicitly define a sampling frame that ideally corresponds to the the final index because the probability of non-zero catch is constrained
stock being assessed. Sampling units in the frame might be defined by a maximum of 1. When binomial predictions are back-transformed,
spatially (i.e., subdividing the stock range into a set of non- index variability can be reduced depending on the reference level
overlapping strata) or via other partitions (i.e., port). selected for the catchability covariates. To avoid bias in index trends and
2. Predict CPUE that would have occurred for each sampling unit, variability, analysts can adjust the mean of the index either via the
conditioning upon values of density covariates that occur at each choice of reference levels or preferably by adding a constant to the logit-
sampling unit and dropping the partial effect of catchability cova­ scale predictions before back-transforming to the probability scale
riates (e.g., by setting them at a defined reference level). Note that (Hoyle et al., 2022). In most situations, we recommend adjusting the
predictions of CPUE into un-sampled sampling units should be scale so that the annual mean of the binomial component of the index
carefully scrutinized and considered (see McKechnie et al., 2013; equals the mean of the annual proportions of nonzero catch. The pre­
Walters, 2003). For example, predictions may need to use one dicted probabilities of non-zero catch may themselves be used in some
approach for areas that are un-sampled because low catch rates make cases to indicate abundance after transforming (Hoyle et al., 2011).
fishing uneconomic and a different approach for areas where man­
agement measures exclude fishing effort. 6. Using CPUE indices in stock assessments
3. Aggregate across sampling units which can often be done by taking
the area-weighted (i.e., the area of available habitat) sum of CPUE As discussed earlier, each CPUE index in a stock assessment is
across these units. associated with a fleet. Section 3 mentioned an approach described by
4. Calculate the variance of the aggregate, either using the delta Maunder et al. (2020) to jointly standardize CPUE and composition data
method, quantiles from MCMC samples in a Bayesian model, and use the results to define two separate fleets in the model with
nonparametric bootstrapping of the data, or other methods. We different purposes. One fleet is associated with the CPUE index and is
expect that these methods will generally give similar results (Mag­ assigned composition data weighted by the spatial distribution of the
nusson et al., 2013), but emphasize that it is necessary to propagate CPUE (representing abundance) and (usually) time-invariant selectivity
information about covariance among predictions to correctly calcu­ so that changes in these composition data represent changes in the
late the variance of the ‘predict-then-aggregate’ approach. composition of the underlying population. The other fleet (not associ­
ated with an index) is assigned composition data weighted by the
For examples of this approach see Campbell (2004) and Campbell time-varying spatial distribution of the catch so that time-varying
(2015). selectivity can be employed to extract fish of the appropriate size/age
Area weighting should, where possible, use the area of available composition.
habitat in each spatial cell (Maunder and Punt, 2004), assuming that Often two or more indices or other datasets provide information that
density is uniform within each cell. For pelagic fishery CPUE, approxi­ is to some extent in conflict. The ‘law of conflicting data’ (Maunder and
mate ocean areas are often determined from the area of grid cells at a Piner, 2017) states that since data are facts, conflict between datasets
latitude, after subtracting any land area. However, the habitat available implies model misspecification. If information from different sources is
to individual species in the pelagic ocean can vary seasonally, with in conflict, the analyst should try to identify the cause and resolve it. If
environmental variables, and long-term with climate change (Goodyear, resolution is not possible, the analyst should create a set of alternative
2016). Inclusion of oceanographic ‘density’ covariates in the standard­ models that omit conflicting datasets; in each model the remaining
ization model and or applying a post-hoc environmental filter can be datasets are well fitted (Francis, 2011; Schnute and Hilborn, 1993).
used to restrict index calculation to viable cells/spatial areas. Methods When conflicts occur, there may be a need to weight the alternative
for determining these areas are an area of active research. Similarly, hypotheses, which introduces some subjectivity. In general, information
cells that are too large can cause bias when they include areas of both about population trends in CPUE should have priority over information
high and low density, with fishing concentrated in the high-density in the composition data (Francis, 2011) because it is usually more
areas. Uncertainty about habitat areas is not often considered but can reliable.
be included via alternative scenarios or Monte Carlo simulation. It is a relatively common practice in some assessment cultures to split
Habitat areas in benthic and reef fisheries are often difficult to CPUE series by time into separate non-overlapping sections and to
determine, and uncertainty can be large. As a result, strata are some­ include each index in a separate fishery. Splitting can occur in response
times weighted by catch (aggregated across the time series), as a proxy to changes in availability of species-specific catch data (e.g., North Pa­
for relative area (e.g., Ralston, 1999). Catch weighting tends to intro­ cific blue shark, ISC Shark Working Group, 2022), evidence of changing
duce bias because fishing intensity always varies spatially. Higher effort operational patterns that potentially change catchability or selectivity
usually occurring in areas with higher catch rates, but some areas may that can’t be included in the CPUE standardisation, or changes to log­
be closed to fishing and others less fished because they are harder to books such as the addition of new data fields (e.g., North Pacific striped
access. Catch weighting is likely to give more weight to areas that are marlin, ISC, 2019). However, splitting is often counterproductive and
more heavily fished, so will tend to exaggerate depletion. should be avoided where possible unless there is another reliable index
When the index-standardization model does not include any in­ of abundance to cover the gap (e.g., Hoyle et al., 2012; Hoyle, 2011).
teractions of year with other covariates, we expect that the ‘predict- Splitting the time series can waste much of the abundance information
then-aggregate’ method will provide the same index as the ‘year-effect in the CPUE data and often substantially changes model outcomes. At
as index’ method (Campbell, 2015). However, when a density covariate best, it increases uncertainty, but it can introduce considerable bias. This
interacts with the year effect (e.g., year x area), or the value of a density should only be the case if something in the model is misspecified (e.g.,
covariate varies through time (e.g., temperature), the misspecification of growth and the associated fits to the length

15
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

composition data which affect model scaling). However, in practice, research or from fishers’ interrogation of their own stored catch and
assessment models always include some degree of misspecification, such effort data. Catchability change associated with vessel turnover can be
as conflict across the time series between different periods of composi­ accounted for by including the vessel ID in the model (e.g., Hoyle and
tion data. Without the stabilizing constraint through time provided by Okamoto, 2011), but datasets rarely indicate when a particular vessel
the continuous index, the model estimates abundance scale indepen­ installed a particular piece of technology or when the behaviour of a
dently before and after the split by adjusting catchability. Splitting the vessel, fishing company, or fleet adapted to use the available informa­
time series therefore often results in abundance changes that reduce the tion more effectively.
internal data conflict but may be inconsistent with expectations about Where catchability increases are considered likely but estimates are
long-term catchability. At minimum, analysts should consider whether unavailable, ignoring them will positively bias stock status estimates (e.
the implied catchability change at the split is consistent with g., Han et al., 2023; Ye and Dennis, 2009). Wilberg et al. (2009)
expectations. recommend a default assumption that catchability varies over time and
There are usually better alternatives to splitting a CPUE series. If the multiple methods of including time-varying catchability should be
split is due to data availability, analysts should strive to retain the same applied. To allow for uncertainty about fishing power, stock assessments
or similar catchability and selectivity for both indices since there is no (particularly for target species) should consider a range of reasonable
evidence that either has changed. If the split is due to catchability scenarios regarding long-term catchability trends, from low to high but
changes linked to new technology, analysts should consider whether the noting that 0% is rarely plausible.
changes estimated in the assessment are plausible. If a change that likely
increased catchability has the opposite effect in the model, it is usually 7. Concluding remarks
better not to split the series. If selectivity has changed with effort moving
to a spatially or seasonally different part of the stock, a better approach CPUE standardization is an influential component of the stock
is to analyse the dataset as separate spatial or seasonal CPUE series that assessment process, as well as a powerful tool for generating under­
each take a consistent part of the stock so that the change in effort standing of the stock and the fishery. Rather than a statistical exercise, it
distribution does not affect either index. Alternatively, a joint model of should be seen as a core part of the fisheries research program. As such,
composition and CPUE data (Maunder et al., 2020) can be used as the initial analytical steps of exploring and characterising the dataset,
described earlier in this section. including talking to fishers, are key. In parallel, the understanding
Transitions such as these often involve alternative hypotheses about developed during CPUE standardization can often motivate changes in
how the changes have affected catchability. Given this uncertainty, a the assessment model structure itself, a contribution often more influ­
useful approach for the analyst is, rather than splitting the index, to ential for assessment outcomes than small changes in the CPUE indices
identify a set of plausible hypotheses about how catchability has themselves.
changed and construct alternative indices based on these hypotheses An important development in the field of CPUE standardisation since
(Campbell, 2016). This approach has more widespread value because Maunder and Punt (2004) has been the increase of STM applications to
analysts often end up with two or more alternative indices, each of fisheries datasets. However, while STMs are attractive models with
which has support. Given the importance of CPUE indices for deter­ powerful features, they do not constitute the best option in all situations.
mining stock assessment outcomes and the many ways that estimated They are also complicated to implement for new analysts because of the
CPUE indices can diverge from the true abundance trend, it is important different statistical concepts involved. We recommend that simpler ap­
to consider the range of plausible options as alternative stock assessment proaches be tried first to develop understanding before, if warranted,
scenarios, preferably within a model ensemble. transitioning to more sophisticated ones (like STMs), if only to under­
Trends in catchability over time can bias CPUE indices and can be of stand the influence of spatio-temporal structure on one’s
great importance for stock assessments. Increases in catch efficiency, or catch-and-effort dataset. Of note, a course based on the VAST STM
fishing power, have played a critical role in the history of fisheries (Thorson, 2022; Thorson, 2019a) is available (https://github.
(Scherrer and Galbraith, 2020; Squires and Vestergaard, 2013). Wilberg com/James-Thorson/2018_FSH556), as are examples in the VAST
et al. (2009) describe causes of increased catchability such as changes in Wiki (https://github.com/James-Thorson-NOAA/VAST). In addition, as
fishing practices and technology (Garrod, 1964). Technological creep STMs may be sensitive to model settings and choices, we encourage
can mask declines in abundance by increasing catchability (Kleiven analysts to consult the growing literature (e.g., Commander et al., 2022;
et al., 2022); and it is observed in almost all analyses involving time Dambly et al., 2023) describing potential analytic trade-offs. We also
series of fishing effort, particularly if the analyses exceed one decade in emphasize that simulation experiments represent valuable tools for
temporal coverage (Palomares and Pauly, 2019). For example, Squires testing and comparing CPUE standardization modelling approaches, so
(1992) found that the most important sources of technical progress in that the analyst may select approaches to produce indices likely to be
the Pacific coast trawl fishery were electronics, the application of sci­ unbiased. Finally, the development of methods to explicitly model fisher
entific rather than craft principles to vessel and equipment design, and location choice to improve predictions in un-sampled areas would
to harvesting methods. constitute a promising extension to STMs.
Estimates of mean rates of fishing power increases across fisheries In Table 3, we reiterate the major recommendations introduced in
are variable. Eigaard et al. (2014) estimated a mean rate of increase of the preceding sections. Analysts are encouraged to consider the good
3.2% per year, while Palomares and Pauly (2019) reviewed 51 estimates practices identified in the 16 areas covered in this review and to apply
of about 2–4% per year and estimated an expected rate of 1.3% per year them where practical. Not all will be feasible to apply to every analysis,
for studies that cover a 100-year period, although this is likely to be an but it is important to understand the implications of not applying them.
underestimate (Scherrer and Galbraith, 2020). In Table 4 (based on ISC Albacore Working Group, 2016) we sum­
One difficulty in estimating rates of fishing power change is marize information requirements for presenting the results of CPUE
obtaining the requisite data (Scherrer and Galbraith, 2020). A major analyses. See also Hoyle et al. (2014a) and IOTC (2015). In Table 5
component of fishing power is the ability to locate fish; though tech­ (based on ISC Albacore Working Group, 2016), we list criteria that can
nology supports this in many ways, few are recorded for analysis. Ex­ be used to assess the strengths and weaknesses of candidate abundance
amples include the installation of GPS systems, provision of increasingly indices.
informative environmental data from satellites and models, upgraded Finally, we raise a note of caution about the potential reliability of
communication technology to share information between vessels and indices of abundance based on CPUE data. The relationship between
with fishing companies, and scientific progress in understanding the CPUE indices and relative biomass or abundance can be proportional,
factors that affect fish distribution, which may derive from public but there are many scenarios in which this relationship changes or

16
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

Table 3 Table 5
Summary of good practices in catch-per-unit-effort (CPUE) standardization. Criteria for evaluating the strengths and weaknesses of candidate abundance
1. Indices of abundance are often the most influential part of an assessment—invest
indices.
accordingly. Criterion Description
Data
2. Defining fleets is key. Definitions depend on stock and fishery structure, the type of Spatial distribution Proportion of stock covered by fishery; latitude and
CPUE analysis (time + space versus spatiotemporal model (STM)), and the longitude
assessment approach (biomass dynamic versus age/size-structured, conventional Size/age range Distribution of size or ages in catch
versus index fishery). Fishing ground map Show area of operations for each fishery by season/
a. Explore and understand data (catch, effort, CPUE, sizes, ages, maturity, gear types, decade
logbook types, vessel turnover, misreporting, etc.). Plot everything. Relative contribution Proportion of total catch in the fishery
b. Talk to fishers and other stakeholders. Temporal coverage Time period of data collection
3. Structural changes based on understanding the system can be very important for the Temporal consistency Change in spatial location of fishing grounds over
assessment, whereas many other issues just cause small changes in the index trends. temporal period, e.g., decadal changes/seasonal
4. Revisit data exploration when updating indices. Do not simply ‘turn the handle’. changes
5. Identify likely covariates before you start modelling. Avoid data dredging. Temporal consistency in Decadal and seasonal changes in size of fish captured
6. Differentiate between catchability and density variables. size composition
7. Always include the variables that affect catchability – this is usually more important Statistical soundness Standardization method, diagnostic plots, and CPUE
than the type of model you use. Think about potential for bias due to missing variability provided
variables. Targeting Primary target, by-catch species
8. Consider targeting and target change through time and how to address it. Drivers of catchability Time series of external factors affecting catchability
Understand the fleet well enough to know what the targeting strategies might be. change (e.g., management practices, fishing technology,
Analysis targeting changes)
9. Generalized additive models (GAMs) and STMs are better than generalized linear Socio-economic factors Time series of price, demand, technological changes
models (GLMs). GAMs are best for exploration. STMs can be better for the final (e.g., freezers), etc.
model(s). Each has unique capabilities. Source: Adapted from ALBWG (2013.
10. Model the whole stock if you can do so without dropping important covariates or
relationships due to data gaps or difficult spatial interactions.
11. Test your model by simulation. CRediT authorship contribution statement
12. Build multiple models using different approaches to develop your understanding
of how the models are working and to consider alternative hypotheses. Start simple. Simon D. Hoyle: Conceptualization, Writing – original draft,
13. Use influence plots to understand how the variables and their values affect the
indices.
Writing – review & editing, Visualization, Supervision, Funding acqui­
14. Construct the index via the “predict-then-aggregate” approach rather than the sition. Robert A. Campbell: Writing – original draft, Writing – review &
previous practice of treating the partial effect of year as the index. editing. Nicholas D. Ducharme-Barth: Writing – original draft, Writing
Assessment – review & editing, Visualization. Arnaud Grüss: Writing – original
15. Use the index fishery approach if you can.
draft, Writing – review & editing. Bradley R. Moore: Writing – original
16. Assume effort creep. There are catchability changes that your model has not
captured. Catchability increases are almost inevitable in the long term. draft, Writing – review & editing, Visualization. James T. Thorson:
17. Do not blindly split indices and assume the model will scale them correctly – this is Writing – original draft, Writing – review & editing. Laura Tremblay-
unlikely and dangerous. Boyer: Writing – original draft, Writing – review & editing. Henning
18. Do not include several conflicting indices in an assessment at the same time. Do Winker: Writing – original draft, Writing – review & editing. Shijie
include alternative indices as alternative assessment scenarios.
Zhou: Writing – original draft, Writing – review & editing. Mark N.
Maunder: Writing – original draft, Writing – review & editing.

Table 4
Declaration of Competing Interest
Information requirements to support the acceptance of abundance indices.
Fishery description Describe fishery including catch, effort, size composition of The authors declare that they have no known competing financial
catch, nominal CPUE by area, season, history of fishery
development and changes.
interests or personal relationships that could have appeared to influence
Analysis description Describe data selection, CPUE standardization model, and the work reported in this paper.
CPUE estimates. Include any data filtering, outlier removal.
Statistical Results Provide model diagnostics and goodness-of-fit criteria Data Availability
relative to alternative model configurations; tables, etc.
Nominal/ Comparison plot of nominal and standardized indices.
Standardized No data was used for the research described in the article.
Diagnostic plots QQ, residuals, etc.
Point estimate & Characterize uncertainty in estimates of standardized Acknowledgments
variability CPUE; SE or CV of standardized CPUE (generated or
assumed).
We thank Steven Holmes, Jill Coyle, Carolina Minte-Vera, and three
Source: Adapted from ALBWG (2013. anonymous reviewers for their advice. The scientific results and con­
clusions, as well as any views or opinions expressed herein, are those of
breaks down (Cooke and Beddington, 1984; Dunn et al., 2000; Harley the authors and do not necessarily reflect those of their employers.
et al., 2001; National Research Council, 2000; Ye and Dennis, 2009).
Although there are no easy answers, the recommendations presented References
here will help analysts to obtain the best information available.
Abascal, F.J., Mejuto, J., Quintans, M., Ramos-Cartelle, A., 2010. Horizontal and vertical
movements of swordfish in the Southeast Pacific. ICES J. Mar. Sci. 67, 466–474.
Funding Abbott, J.K., Haynie, A.C., Reimer, M.N., 2015. Hidden flexibility: institutions,
incentives, and the margins of selectivity in fishing. Land Econ. 91, 169–195.
This work was supported in part by NIWA Strategic Science Invest­ Aho, K., Derryberry, D., Peterson, T., 2014. Model selection for ecologists: the
worldviews of AIC and BIC. Ecology 95, 631–636.
ment Funding and by the Indian Ocean Tuna Commission. The funders
Akaike, H., 1973. Information theory and an extension of the maximum likelihood
had no role in study design, data collection and analysis, decision to principle. Springer Verlag.
publish, or preparation of the manuscript. Alglave, B., Rivot, E., Etienne, M.-P., Woillez, M., Thorson, J.T., Vermard, Y., 2022.
Combining scientific survey and commercial catch data to map fish distribution.
ICES J. Mar. Sci. 79, 1133–1149. https://doi.org/10.1093/icesjms/fsac032.

17
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

Anderson, S.C., Ward, E.J., English, P.A., Barnett, L.A., 2022. sdmTMB: an R package for the South Atlantic. Fish. Oceanogr. 29, 169–184. https://doi.org/10.1111/
fast, flexible, and user-friendly generalized linear mixed effects models with spatial fog.12462.
and spatiotemporal random fields. bioRxiv. Commander, C.J., Barnett, L.A., Ward, E.J., Anderson, S.C., Essington, T.E., 2022. The
Anon. Report of the IOTC CPUE Workshop, San Sebastian, Spain, 21–22 October, 2013. shadow model: how and why small choices in spatially explicit species distribution
Indian Ocean Tuna Commission; 2013. models affect predictions. PeerJ 10, e12783.
Anonymous. Report of the 2018 ICCAT bigeye tuna stock assessment meeting. SCRS/ Conn, P.B., 2010. Hierarchical analysis of multiple noisy abundance indices. Can. J. Fish.
2018/010. Collective Volume of Scientific Papers ICCAT. 75:1721–1855; 2019. Aquat. Sci. 67, 108–120. https://doi.org/10.1139/F09-175.
Beerkircher, L., Arocha, F., Barse, A., Prince, E., Restrepo, V., Serafy, J., Shivji, M., 2009. Conn, P.B., Thorson, J.T., Johnson, D.S., 2017. Confronting preferential sampling when
Effects of species misidentification on population assessment of overfished white analysing population distributions: diagnosis and model-based triage. Methods Ecol.
marlin Tetrapturus albidus and roundscale spearfish T. georgii. Endanger. Species Res. Evol. 8, 1535–1546. https://doi.org/10.1111/2041-210x.12803.
9, 81–90. Conn, P.B., Johnson, D.S., Williams, P.J., Melin, S.R., Hooten, M.B., 2018. A guide to
Bentley, N., Kendrick, T.H., Starr, P.J., Breen, P.A., 2012. Influence plots and metrics: Bayesian model checking for ecologists. Ecol. Monogr. 88, 526–542. https://doi.org/
tools for better understanding fisheries catch-per-unit-effort standardizations. ICES 10.1002/ecm.1314.
J. Mar. Sci. 69, 84–88. https://doi.org/10.1093/icesjms/fsr174. Cooke, J., Beddington, J., 1984. The relationship between catch rates and abundance in
Beverton, R.J.H., Holt, S.J., 1957. On the dynamics of exploited fish populations. fisheries. Math. Med. Biol.: A J. IMA 1, 391–405.
Ministry of Agriculture and Fisheries, London: UK. Cosandey-Godin, A., Krainski, E.T., Worm, B., Flemming, J.M., 2015. Applying Bayesian
Biseau, A., 1998. Definition of a directed fishing effort in a mixed-species trawl fishery, spatiotemporal models to fisheries bycatch in the Canadian Arctic. Can. J. Fish.
and its impact on stock assessments. Aquat. Living Resour. 11, 119–136. Aquat. Sci. 72, 186–197. https://doi.org/10.1139/cjfas-2014-0159.
Bishop, J., 2006. Standardizing fishery-dependent catch and effort data in complex Dambly, L.I., Isaac, N.J., Jones, K.E., Boughey, K.L., O’Hara, R.B., 2023. Integrated
fisheries with technology change. Rev. Fish. Biol. Fish. 16, 21–38. https://doi.org/ species distribution models fitted in INLA are sensitive to mesh parameterisation.
10.1007/s11160-006-0004-9. Ecography e06391.
Bolker, B. GLMM FAQ. 〈https://bbolker.github.io/mixedmodels-misc/glmmFAQ.html〉; Denis, V., Lejeune, J., Robin, J., 2002. Spatio-temporal analysis of commercial trawler
2022. data using General Additive models: patterns of Loliginid squid abundance in the
Bosch, S., Tyberghein, L., Deneudt, K., Hernandez, F., De Clerck, O., 2018. In search of north-east Atlantic. ICES J. Mar. Sci. 59, 633–648. https://doi.org/10.1006/
relevant predictors for marine species distribution modelling using the MarineSPEED jmsc.2001.1178.
benchmark dataset. Diversity and Distributions 24 (2), 144–157. https://doi.org/ Devine, J.; Ballara, S.; Hoyle, S. Fisheries characterisations for barracouta (Thyrsites
10.1111/ddi.12668. Wiley Online Library. atun) and preliminary standardised CPUE analyses in BAR 4 and BAR 5, 1990 to
Braccini, M., Denham, A., O’Neill, M.F., Lai, E., 2021. Spatial and temporal patterns in 2020. New Zealand Fisheries Assessment Report:13; 2022.
catch rates from multispecies shark fisheries in Western Australia. Ocean Coast. Dick, E.J., 2004. Beyond ‘lognormal versus gamma’: discrimination among error
Manag. 213, 105883 https://doi.org/10.1016/j.ocecoaman.2021.105883. distributions for generalized linear models. Fish. Res. 70, 351–366. https://doi.org/
Buck, E.H., 1995. Individual transferable quotas in fishery management. Congr. Res. 10.1016/j.fishres.2004.08.013.
Serv., Libr. Congr. 95, 849. Diggle, P.J., Menezes, R., Su, Tl, 2010. Geostatistical inference under preferential
Bürkner, P.-C., 2017. brms: an R package for Bayesian multilevel models using Stan. sampling. J. R. Stat. Soc.: Ser. C. (Appl. Stat. ) 59, 191–232. https://doi.org/
J. Stat. Softw. 80, 1–28. 10.1111/j.1467-9876.2009.00701.x.
Burnham, K.P., Anderson, D.R., 2004. Multimodel inference understanding AIC and BIC Ducharme-Barth, N.; Vincent, M.; Vidal, T.; Hamer, P. Analysis of Pacific-wide
in model selection. Sociol. Methods Res. 33, 261–304. https://doi.org/10.1177/ operational longline dataset for bigeye and yellowfin tuna catch-per-unit-effort
0049124104268644. (CPUE). Technical Report SC16-SA-WP-03; 2020.
Campbell, R. Aggregate and size-based standardised CPUE indices for longline target Ducharme-Barth, N.D., Grüss, A., Vincent, M.T., Kiyofuji, H., Aoki, Y., Pilling, G.,
species caught within the ETBF - 2019 Update. 24th meeting of the Tropical Tuna Hampton, J., Thorson, J.T., 2022. Impacts of fisheries-dependent spatial sampling
Resource Assessment Group, 17–18 July 2019. Mooloolaba, Australia; 2019. patterns on catch-per-unit-effort standardization: a simulation study and fishery
Campbell, R.; Williams, K.; Williams, D. Summary of size data collected from the Eastern application. Fish. Res. 246, 106169 https://doi.org/10.1016/j.fishres.2021.106169.
Tuna and Billfish Fishery and associated indicators. 24th meeting of the Tropical Dulvy, N.K., Metcalfe, J.D., Glanville, J., Pawson, M.G., Reynolds, J.D., 2000. Fishery
Tuna Resource Assessment Group, 17–18 July 2019. Mooloolaba. Australia; 2019. stability, local extinctions, and shifts in community structure in skates. Conserv. Biol.
Campbell, R.A., 2004. CPUE standardisation and the construction of indices of stock 14, 283–293. https://doi.org/10.1046/j.1523-1739.2000.98540.x.
abundance in a spatially varying fishery using general linear models. Fish. Res. 70, Dunn, A., Harley, S., Doonan, I., Bull, B., 2000. Calculation and interpretation of catch-
209–227. https://doi.org/10.1016/j.fishres.2004.08.026. per-uniteffort (CPUE) indices. N. Z. Fish. Assess. Rep. 1, 44.
Campbell, R.A., 2015. Constructing stock abundance indices from catch and effort data: Dunn, P.Tweedie, 2017. Evaluation of Tweedie exponential family models. R. Package
some nuts and bolts. Fish. Res. 161, 109–130. https://doi.org/10.1016/j. Version 2.
fishres.2014.07.004. Duparc, A., Aragno, V., Depetris, M., Floch, L., Cauquil, P., Lebranchu, J., Gaertner, D.,
Campbell, R.A., 2016. A new spatial framework incorporating uncertain stock and fleet Marsac, F., Bach, P., 2020. Assessment of the species composition of major tropical
dynamics for estimating fish abundance. Fish Fish 17, 56–77. https://doi.org/ tunas in purse seine catches: a new modelling approach for the tropical tuna
10.1111/faf.12091. treatment processing (case of the French fleet in Atlantic Ocean). Collect Vol. Sci.
Campbell, R.A., Zhou, S., Hoyle, S.D., Hillary, R., Haddon, M., Auld, S., 2017. Developing Pap. ICCAT 76, 951–982.
innovative approaches to improve CPUE standardisation for Australia’s multispecies Eigaard, O.R., Marchal, P., Gislason, H., Rijnsdorp, A.D., 2014. Technological
pelagic longline fisheries. Fisheries Research and Development Corporation, development and fisheries management. Rev. Fish. Sci. Aquac. 22, 156–174. https://
Canberra, Australia. doi.org/10.1080/23308249.2014.899557.
Cao, J., Thorson, J.T., Richards, R.A., Chen, Y., 2017. Spatiotemporal index Emery, T.J., Noriega, R., Williams, A.J., Larcombe, J., 2019. Changes in logbook
standardization improves the stock assessment of northern shrimp in the Gulf of reporting by commercial fishers following the implementation of electronic
Maine. Can. J. Fish. Aquat. Sci. 74, 1781–1793. monitoring in Australian Commonwealth fisheries. Mar. Policy 104, 135–145.
Cardinale, M., Nugroho, D., Jonson, P., 2011. Serial depletion of fishing grounds in an https://doi.org/10.1016/j.marpol.2019.01.018.
unregulated, open access fishery. Fish. Res. 108, 106–111. https://doi.org/10.1016/ Forrestal, F., Goodyear, C.P., Schirripa, M., Babcock, E., Lauretta, M., Sharma, R., 2017.
j.fishres.2010.12.007. Testing robustness of CPUE standardization using simulated data: findings of initial
Carruthers, T.R., Ahrens, R.N.M., McAllister, M.K., Walters, C.J., 2011. Integrating blind trials. Collect Vol. Sci. Pap. ICCAT 74, 391–403.
imputation and standardization of catch rate data in the calculation of relative Forrestal, F.C., Schirripa, M., Goodyear, C.P., Arrizabalaga, H., Babcock, E.A., Coelho, R.,
abundance indices. Fish. Res. 109, 157–167. https://doi.org/10.1016/j. Ingram, W., Lauretta, M., Ortiz, M., Sharma, R., 2019. Testing robustness of CPUE
fishres.2011.01.033. standardization and inclusion of environmental variables with simulated longline
Carvalho, F.C., Murie, D.J., Hazin, F.H., Hazin, H.G., Leite-Mourato, B., Travassos, P., catch datasets. Fish. Res. 210, 1–13. https://doi.org/10.1016/j.fishres.2018.09.025.
Burgess, G.H., 2010. Catch rates and size composition of blue sharks (Prionace Francis, C.R.; Hurst, R.J.; Renwick, J.A. Quantifying annual variation in catchability for
glauca) caught by the Brazilian pelagic longline fleet in the southwestern Atlantic commercial and research fishing. 2003.
Ocean. Aquat. Living Resour. 23, 373–385. Francis, R.I.C.C., 2011. Data weighting in statistical fisheries stock assessment models.
Chambers, M.; Hoyle, S.D. Estimates of non-member catch of SBT in the Indian and Can. J. Fish. Aquat. Sci. 68, 1124–1138.
Pacific Oceans, CCSBT-ESC/1509/10. 20th Extended Scientific Committee of the Fu, D.; Roux, M.-J.; Clarke, S.C.; Francis, M.P.; Dunn, A.; Hoyle, S.D. Pacific-wide
CCSBT. Incheon, Republic of Korea; 2015. sustainability risk assessment of bigeye thresher shark (Alopias superciliosus). NIWA
Chang, S.-K., 2003. Analysis of Taiwanese white marlin catch data and standardization of Client Report No 2016089WN for project WCP16301. Wellington; 2016.
catch rates. ICCAT Col. Vol. Sci. Pap. 55, 453–466. Garrod, D., 1964. Effective fishing effort and the catchability coefficient q. Rapp. Et.
Chang, S.-K., Hoyle, S., Liu, H.-I., 2011. Catch rate standardization for yellowfin tuna Process verbaux Des. réunions du Cons. Int. pour l’Explor. De. la Mer. 155, 66–70.
(Thunnus albacares) in Taiwan’s distant-water longline fishery in the Western and Girardin, R., Hamon, K.G., Pinnegar, J., Poos, J.J., Thébaud, O., Tidd, A., Vermard, Y.,
Central Pacific Ocean, with consideration of target change. Fish. Res. 107, 210–220. Marchal, P., 2017. Thirty years of fleet dynamics modelling using discrete-choice
https://doi.org/10.1016/j.fishres.2010.11.004. models: What have we learned? Fish Fish. 18, 638–655.
Charsley, A.R., Sibanda, N., Hoyle, S., Crow, S., 2022. Comparing the performance of Glazer, J., Butterworth, D., 2002. GLM-based standarization of the catch per unit effort
three common species distribution modelling frameworks for freshwater series for South African west coast hake, focusing on adjustments for targeting other
environments through application to eel species in New Zealand. Can. J. Fish. Aquat. species. Afr. J. Mar. Sci. 24, 323–339.
Sci. Goodyear, C.P., 2016. Modeling the time-varying density distribution of highly
Coelho, R., Infante, P., Santos, M.N., 2020. Comparing GLM, GLMM, and GEE modeling migratory species: atlantic blue marlin as an example. Fish. Res. 183, 469–481.
approaches for catch rates of bycatch species: A case study of blue shark fisheries in

18
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

Grüss, A., Thorson, J.T., 2019. Developing spatio-temporal models using multiple data Hoyle, S.D. Stock assessment of albacore tuna in the South Pacific ocean. Scientific
types for evaluating population trends and habitat usage. ICES J. Mar. Sci. 76, Committee, Seventh Regular Session, 9–17 August 2011, Pohnpei, Federated States
1748–1761. https://doi.org/10.1093/icesjms/fsz075. of Micronesia; 2011.
Grüss, A., Walter III, J.F., Babcock, E.A., Forrestal, F.C., Thorson, J.T., Lauretta, M.V., Hoyle, S.D. Investigation of potential CPUE model improvements for the primary index
Schirripa, M.J., 2019. Evaluation of the impacts of different treatments of spatio- of Southern Bluefin Tuna abundance, CCSBT-ESC/2008/29. CCSBT Extended
temporal variation in catch-per-unit-effort standardization models. Fish. Res. 213, Scientific Committee 25. online; 2020.
75–93. https://doi.org/10.1016/j.fishres.2019.01.008. Hoyle, S.D., Langley, A.D., 2020. Scaling factors for multi-region stock assessments, with
Grüss, A., Pirtle, J.L., Thorson, J.T., Lindeberg, M.R., Neff, A.D., Lewis, S.G., Essington, T. an application to Indian Ocean tropical tunas. Fish. Res. 228, 105586 https://doi.
E., 2021. Modeling nearshore fish habitats using Alaska as a regional case study. org/10.1016/j.fishres.2020.105586.
Fish. Res. 238, 105905 https://doi.org/10.1016/j.fishres.2021.105905. Hoyle, S.D.; Okamoto, H. Analyses of Japanese longline operational catch and effort for
Grüss, A., Moore, B.R., Pinkerton, M.H., Devine, J.A., 2022. VAST (vector autoregressive bigeye and yellowfin tuna in the WCPO, WCPFC-SC7-SA-IP-01. Western and Central
spatio-temporal) modelling of macrourid relative abundance in the Ross Sea region Pacific Fisheries Commission, 7th Scientific Committee. Pohnpei, Federated States of
to support bycatch management. WG-FSA-2022/48. CCAMLR, Hobart, Australia. Micronesia; 2011.
Grüss, A., McKenzie, J.R., Lindegren, M., Bian, R., Hoyle, S.D., Devine, J.A., 2023a. Hoyle, S.D., Okamoto, H., 2015. Descriptive analyses of the Japanese Indian Ocean
Supporting a stock assessment with spatio-temporal models fitted to fisheries- longline fishery, focusing on tropical areas. Indian Ocean Tuna Comm. Work. Part.
dependent data. Fish. Res. 262, 106649. Trop. Tunas.
Grüss, A., Moore, B.R., Pinkerton, M.H., Devine, J.A., 2023b. Understanding the spatio- Hoyle, S.D.; Chang, S.-T.; Fu, D.; Kim, D.N.; Lee, S.I.; Matsumoto, T.; Chassot, E.; Yeh, Y.-
temporal abundance patterns of the major bycatch species groups in the Ross Sea M. Collaborative study of bigeye and yellowfin tuna CPUE from multiple Indian
region Antarctic toothfish (Dissostichus mawsoni) fishery. Fish. Res. 262, 106647. Ocean longline fleets in 2019, with consideration of discarding. IOTC-2019-
Han, H., Yang, C., Zhang, H., Fang, Z., Jiang, B., Su, B., Sui, J., Yan, Y., Xiang, D., 2022. WPM10–16. 10th Working Party on Methods, 17–19 October, 2019. Donostia-San
Environment variables affect CPUE and spatial distribution of fishing grounds on the Sebastian, Spain: Indian Ocean Tuna Commission; 2019b.
light falling gear fishery in the Northwest Indian Ocean at different time scales. Hoyle, S.D.; Lauretta, M.; Lee, M.K.; Matsumoto, T.; Sant’Ana, R.; Yokoi, H.; Su, N.-J.
Front. Mar. Sci. 9, 939334 https://doi.org/10.3389/fmars.2022.939334. Collaborative study of yellowfin tuna CPUE from multiple Atlantic Ocean longline
Han, Q., Grüss, A., Shan, X., Jin, X., Thorson, J.T., 2021. Understanding patterns of fleets in 2019. 2019c.
distribution shifts and range expansion/contraction for small yellow croaker Hoyle, S.D.; Langley, A.D.; Campbell, R.A. Guidelines for presenting CPUE indices of
(Larimichthys polyactis) in the Yellow Sea. Fish. Oceanogr. 30, 69–84. abundance for WCPFC stock assessments. WCPFC-SC10–2014/ SA-IP-11. WCPFC
Han, Q., Shan, X., Jin, X., Gorfine, H., 2023. Contrasting stock status trends obtained Scientific Committee 10. Majuro, Republic of the Marshall Islands.; 2014a.
from survey and fishery CPUE, taking Larimichthys polyactis in Yellow Sea Large Hoyle, S.D.; Langley, A.D.; Campbell, R.A. Recommended approaches for standardizing
Marine Ecosystem as an example. Ecol. Indic. 147, 110032. CPUE data from pelagic fisheries, WCPFC-SC10–2014/ SA-IP-10. WCPFC Scientific
Hansell, A.C., Becker, S.L., Cadrin, S.X., Lauretta, M., Walter III, J.F., Kerr, L.A., 2022. Committee 10. Majuro, Republic of the Marshall Islands.; 2014b.
Spatio-temporal dynamics of bluefin tuna (Thunnus thynnus) in US waters of the Hoyle, S.D., Lee, S.I., Kim, Z.G., 2015a. Descriptive analyses of the Korean Indian Ocean
northwest Atlantic. Fish. Res. 255, 106460 https://doi.org/10.1016/j. longline fishery, focusing on tropical areas. Indian Ocean Tuna Comm. Work. Part.
fishres.2022.106460. Trop. Tunas.
Harley, S.J., Myers, R.A., Dunn, A., 2001. Is catch-per-unit-effort proportional to Hoyle, S.D., Okamoto, H., Yeh, Y.-m, Kim, Z.G., Lee, S.I., Sharma, R., 2015b.
abundance? Can. J. Fish. Aquat. Sci. 58, 1760–1772. https://doi.org/10.1139/cjfas- IOTC–CPUEWS02 2015: report of the 2nd CPUE Workshop on Longline Fisheries, 30
58-9-1760. April – 2 May 2015. Indian Ocean Tuna Comm.
Harrell, F.E., 2001. Regression modeling strategies: with applications to linear models, Hoyle, S.D., Yeh, Y.-M., Chang, S.-T., Wu, R.-F., 2015c. Descriptive analyses of the
logistic regression, and survival analysis. Springer. Taiwanese Indian Ocean longline fishery, focusing on tropical areas. Indian Ocean
Hartig, F. DHARMa: residual diagnostics for hierarchical (Multi-Level/Mixed) regression Tuna Comm. Work. Part. Trop. Tunas.
models. 2020. Hoyle, S.D.; Semba, Y.; Kai, M.; Okamoto, H. Development of porbeagle shark stock
He, X., Bigelow, K.A., Boggs, C.H., 1997. Cluster analysis of longline sets and fishing abundance indicators using Japanese data. New Zealand Fisheries Assessment
strategies within the Hawaii-based fishery. Fish. Res. 31, 147–158. Report 2017/07; 2017.
Helle, K., Pennington, M., Hareide, N.-R., Fossen, I., 2015. Selecting a subset of the Hoyle, S.D., Lee, S.I., Kim, D.N., 2022. CPUE standardization for southern bluefin tuna
commercial catch data for estimating catch per unit effort series for ling (Molva (Thunnus maccoyii) in the Korean tuna longline fishery, accounting for
molva L.). Fish. Res. 165, 115–120. spatiotemporal variation in targeting through data exploration and clustering. PeerJ
Henderson, H.V., Velleman, P.F., 1981. Building multiple regression models 10, e13951.
interactively. Biometrics 391–411. Hsu, J., Chang, Y.-J., Ducharme-Barth, N.D., 2022. Evaluation of the influence of spatial
Hilborn, R., Walters, C.J., 1992. Quantitative Fisheries Stock Assessment: Choice treatments on catch-per-unit-effort standardization: A fishery application and
Dynamics and Uncertainty. Chapman and Hall, New York. simulation study of Pacific saury in the Northwestern Pacific Ocean. Fish. Res. 255,
Hinton, M.G., 1996. Standardizing catch and effort statistics using physiological, 106440 https://doi.org/10.1016/j.fishres.2022.106440.
ecological, or behavioral constraints, and environmental data, with applications to Huang, H.-W., Liu, K.-M., 2010. Bycatch and discards by Taiwanese large-scale tuna
blue marlin (Makaira nigricans) and swordfish (Xiphias gladius) of the Pacific Ocean. longline fleets in the Indian Ocean. Fish. Res. 106, 261–270. https://doi.org/
Scripps Institution of Oceanography. University of California, San Diego. 10.1016/j.fishres.2010.08.005.
Hinton, M.G.; Maunder, M.N. Methods for standardizing CPUE and how to select among Huang, W.-B., Lo, N.C., Chiu, T.-S., Chen, C.-S., 2007. Geographical distribution and
them. Collective volume of scientific papersInternational Commission for the abundance of Pacific saury, Cololabis saira (Brevoort)(Scomberesocidae), fishing
Conservation of Atlantic Tunas/Recueil de documents scientifiquesCommission stocks in the Northwestern Pacific in relation to sea temperatures. Zool. Stud. 46,
internationale pour la Conservation des Thonides de l’Atlantique/Coleccion de 705–716.
documentos cientificosCo:no; 2004. Huang, W.-B.; Chang, Y.-J.; Hsieh, C. Standardized CPUE of Pacific saury (Cololabis saira)
Hinton, M.G., Nakano, H., 1996. Standardizing catch and effort statistics using caught by the Chinese Taipei stick-held dip net fishery up to 2019. NPFC-2020-SSC
physiological, ecological, or behavioral constraints and environmental data, with an PS06-WP05; 2020.
application to blue marlin (Makaira nigricans) catch and effort data from the Ichinokawa, M., Brodziak, J., 2010. Using adaptive area stratification to standardize
Japanese longline fisheries in the Pacific. Inter Am. Trop. Tuna Comm. Bull. 21, catch rates with application to North Pacific swordfish (Xiphias gladius). Fish. Res.
171–200. 106, 249–260. https://doi.org/10.1016/j.fishres.2010.08.001.
Hinz, H., Murray, L.G., Lambert, G.I., Hiddink, J.G., Kaiser, M., 2013. Confidentiality Ichinokawa, M.; Takeuchi, Y. Estimation of coefficient of variances in standardized CPUE
over fishing effort data threatens science and management progress. Fish Fish. 14, of Pacific bluefin tuna caught by Japanese coastal longline with a nonparametric
110–117. https://doi.org/10.1111/j.1467-2979.2012.00475.x. method. Working document submitted for publication to the ISC Pacific bluefin tuna
Hiraoka, Y.; Kanaiwa, M.; Yokawa, K. Estimation of abundance indices for blue shark in Working Group, International Scientific Committee for Tuna and Tuna-Like Species
the North Pacific. ISC/12/SHARKWG-1/09. Working document submitted for in the North Pacific Ocean (ISC); 2012.
publication to the ISC Shark Working Group Workshop, 28 May – 4 June, 2012; IOTC. Guidelines for the Presentation of CPUE Standardisations and Stock Assessment
2012. Models. Extract from the 17th Session of the Scientific Committee. Appendix IV.
Højsgaard, S., Halekoh, U., Yan, J., 2006. The R package geepack for generalized Indian Ocean Tuna Commission; 2015.
estimating equations. J. Stat. Softw. 15, 1–11. ISC. Stock Assessment Report for Striped Marlin (Kajikia audax) in the Western and
Hoyle, S. Potential CPUE model improvements for the primary index of Southern Bluefin Central North Pacific Ocean through 2017. WCPFC-SC15–2019/SA-WP-09. WCPFC
Tuna abundance. CCSBT Extended Scientific Committee for the 26th Meeting of the Scientific Committee, Fifteenth Regular Session, 12–20 August 2019, Pohnpei,
Scientific Committee. Online; 2021. Federated States of Micronesia 2019.
Hoyle, S.; Huang, H.; Kim, D.N.; Lee, M.K.; Matsumoto, T.; Walter, J. Collaborative study ISC Albacore Working Group. Annex 4: Report of the Albacore Working Group
of bigeye tuna CPUE from multiple Atlantic Ocean longline fleets in 2018. Collect Workshop. 8–14 November 2016. Nanaimo, British Columbia, Canada. International
Vol Sci Pap ICCAT. 75:2033–2080; 2019a. Scientific Committee for Tuna and Tuna-like Species In the North Pacific Ocean;
Hoyle, S.; Kleiber, P.; Davies, N.; Langley, A.D.; Hampton, J.. Stock assessment of 2016.
skipjack tuna in the western and central Pacific Ocean, WCPFC-SC-7-SA/WP-04. ISC Shark Working Group. Stock assessment and future projections of blue sharks in the
WCPFC Scientific Committee Regular Session, Pohnpei, FSM, 9–17 August 2011, 7th; North Pacific Ocean through 2020, WCPFC-SC9–2013/SA-WP-11. 18th WCPFC
2011. Scientific Committee, 10–18 August 2022. Online; 2022.
Hoyle, S.; Hampton, J.; Davies, N. Stock assessment of albacore tuna in the South Pacific Jones, C.M., 2004. The use of recreational CPUE statistics in groundfish stock
ocean. Scientific Committee, Eighth Regular Session, 7–15 August 2012, Busan, assessments. Rep. CIE Recreat. CPUE Indices.
Republic of Korea; 2012.

19
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

Kai, M., 2019. Spatio-temporal changes in catch rates of pelagic sharks caught by and associated composition data for inclusion in stock assessment models. Fish. Res.
Japanese research and training vessels in the western and central North Pacific. Fish. 229, 105594 https://doi.org/10.1016/j.fishres.2020.105594.
Res. 216, 177–195. https://doi.org/10.1016/j.fishres.2019.02.015. Maureaud, A.A., Frelat, R., Pécuchet, L., Shackell, N., Mérigot, B., Pinsky, M.L.,
Kai, M., Thorson, J.T., Piner, K.R., Maunder, M.N., 2017. Spatiotemporal variation in Amador, K., Anderson, S.C., Arkhipkin, A., Auber, A., 2021. Are we ready to track
size-structured populations using fishery data: an application to shortfin mako climate-driven shifts in marine species across international boundaries?-A global
(Isurus oxyrinchus) in the Pacific Ocean. Can. J. Fish. Aquat. Sci. 74, 1765–1780. survey of scientific bottom trawl data. Glob. Change Biol. 27, 220–236. https://doi.
https://doi.org/10.1139/cjfas-2016-0327. org/10.1111/gcb.15404.
Kanamori, Y., Nishijima, S., Okamura, H., Yukami, R., Watai, M., Takasuka, A., 2021. McKechnie, S.; Hoyle, S.; Harley, S. Longline CPUE series that account for changes in the
Spatio-temporal model reduces species misidentification bias of spawning eggs in spatial extent of fisheries. WCPFC-SC9–2013/SA-IP-05. WCPFC 9th Scientific
stock assessment of spotted mackerel in the western North Pacific. Fish. Res. 236, Committee. Pohnpei, Federated States of Micronesia: Western and Central Pacific
105825. Fisheries Commission; 2013.
Kendrick, T.H., Bentley, N., 2011. Fishery characterisations and catch-per-unit-effort McKenzie, J.R.; Parsons, D.M. Fishery Characterisations and Catch-per-unit-effort Indices
indices for three sub-stocks of red gurnard in GUR 1, 1989–90 to 2008–09. Ministry for Three Sub-stocks of Snapper SNA 1, 1989–90 to 2009–10: Ministry for Primary
of Fisheries, Wellington, New Zealand. Industries; 2012.
Kerwath, S.E., Parker, D., Winker, H., Potts, W., Mann, B., Wilke, C., Attwood, C., 2019. Mejuto, J., García-Cortés, B., Ramos-Cartelle, A., De la Serna, J., 2009. Standardized
Tracking the decline of the world’s largest seabream against policy adjustments. catch rates for the blue shark (Prionace glauca) and shortfin mako (Isurus oxyrinchus)
Mar. Ecol. Prog. Ser. 610, 163–173. caught by the Spanish surface longline fleet in the Atlantic Ocean during the period
Klaer, N.L., Smith, D.C., 2012. Determining primary and companion species in a multi- 1990-2007. Collect Vol. Sci. Pap. ICCAT 64, 1509–1521.
species fishery: implications for TAC setting. Mar. Policy 36, 606–612. Minami, M., Lennert-Cody, C.E., Gao, W., Román-Verdesoto, M., 2007. Modeling shark
Kleiber, P., Maunder, M.N., 2008. Inherent bias in using aggregate CPUE to characterize bycatch: the zero-inflated negative binomial regression model with smoothing. Fish.
abundance of fish species assemblages. Fish. Res. 93, 140–145. https://doi.org/ Res. 84, 210–221. https://doi.org/10.1016/j.fishres.2006.10.019.
10.1016/j.fishres.2008.03.013. Miyabe, N., Takeuchi, Y., 2003. Standardized bluefin CPUE from the Japanese longline
Kleiven, A.R., Espeland, S.H., Stiansen, S., Ono, K., Zimmermann, F., Olsen, E.M., 2022. fishery in the Atlantic including those for mixing studies. ICCAT Col. Vol. Sci. Pap.
Technological creep masks continued decline in a lobster (Homarus gammarus) 55, 1190–1207.
fishery over a century. Sci. Rep. 12, 1–10. https://doi.org/10.1038/s41598-022- Monnahan, C.C., Thorson, J.T., Kotwicki, S., Lauffenburger, N., Ianelli, J.N., Punt, A.E.,
07293-2. 2021. Incorporating vertical distribution in index standardization accounts for
Kristensen, K., Nielsen, A., Berg, C.W., Skaug, H.J., Bell, B., 2016. TMB: automatic spatiotemporal availability to acoustic and bottom trawl gear for semi-pelagic
differentiation and laplace approximation. J. Stat. Softw. 70, 1–21. species. ICES J. Mar. Sci. 78, 1826–1839. https://doi.org/10.1093/icesjms/fsab085.
Lambert, D.Zero-inflated, 1992. Poisson regression, with an application to defects in Munoz, B., Lesser, V.M., Smith, R.A., 2010. Applying multiple imputation with
manufacturing. Technometrics 34, 1–14. https://doi.org/10.2307/1269547. geostatistical models to account for item nonresponse in environmental data. J. Mod.
Langley, A., 2019. An investigation of the performance of CPUE modelling approaches–a Appl. Stat. Methods 9, 27.
simulation study. N. Z. Fish. Assess. Rep. 57, 50. Nakano, H., 1998. Stock status of Pacific swordfish, Xiphias gladius, inferred from CPUE
Lee, L.A.; Terrell, J.W. Habitat suitability index models: Flathead catfish. Biological of the Japanese longline fleet standardized using general linear models. US Nat. Mar.
Report 82: US Department of the Interior, Fish and Wildlife Service, Research and Fish. Serv., NOAA Tech. Rep. NMFS 142, 195–209.
Development; 1988. National Research Council, 2000. Improving the collection, management, and use of
Lehodey, P., Bertignac, M., Hampton, J., Lewis, A., Picaut, J., 1997. El Nino Southern marine fisheries data. National Academies Press.
Oscillation and tuna in the western Pacific. Nature 389, 715–718. https://doi.org/ Noriega, R., Werry, J.M., Sumpton, W., Mayer, D., Lee, S.Y., 2011. Trends in annual
10.1038/39575. CPUE and evidence of sex and size segregation of Sphyrna lewini: management
Lennert-Cody, C.E., Maunder, M.N., Aires-da-Silva, A., Minami, M., 2013. Defining implications in coastal waters of northeastern Australia. Fish. Res. 110, 472–477.
population spatial units: Simultaneous analysis of frequency distributions and time https://doi.org/10.1016/j.fishres.2011.06.005.
series. Fish. Res. 139, 85–92. https://doi.org/10.1016/j.fishres.2012.10.001. O’Leary, C.A., Thorson, J.T., Ianelli, J.N., Kotwicki, S., 2020. Adapting to climate-driven
Li, Z., Ye, Z., Wan, R., Zhang, C., 2015. Model selection between traditional and popular distribution shifts using model-based indices and age composition from multiple
methods for standardizing catch rates of target species: a case study of Japanese surveys in the walleye pollock (Gadus chalcogrammus) stock assessment. Fish.
Spanish mackerel in the gillnet fishery. Fish. Res. 161, 312–319. Oceanogr. 29, 541–557.
Lindegren, M., van Deurs, M., Maureaud, A., Thorson, J.T., Bekkevold, D., 2022. Okamura, H., Morita, S.H., Funamoto, T., Ichinokawa, M., Eguchi, S., 2017. Target-based
A spatial statistical approach for identifying population structuring of marine fish catch-per-unit-effort standardization in multispecies fisheries. Can. J. Fish. Aquat.
species: European sprat as a case study. ICES J. Mar. Sci. 79, 423–434. https://doi. Sci. 75, 452–463.
org/10.1093/icesjms/fsac007. Ono, K., Punt, A.E., Hilborn, R., 2015. Think outside the grids: an objective approach to
Lindgren, F., Rue, H., 2015. Bayesian spatial modelling with R-INLA. J. Stat. Softw. 63, define spatial strata for catch and effort analysis. Fish. Res. 170, 89–101. https://doi.
1–25. https://doi.org/10.18637/jss.v063.i19. org/10.1016/j.fishres.2015.05.021.
Liu, S., Liu, Y., Li, J., Cao, C., Tian, H., Li, W., Tian, Y., Watanabe, Y., Lin, L., Li, Y., 2022. Ono, K., Ianelli, J.N., McGilliard, C.R., Punt, A.E., 2018. Integrating data from multiple
Effects of oceanographic environment on the distribution and migration of Pacific surveys and accounting for spatio-temporal correlation to index the abundance of
saury (Cololabis saira) during main fishing season. Sci. Rep. 12, 1–13. https://doi. juvenile Pacific halibut in Alaska. ICES J. Mar. Sci. 75, 572–584. https://doi.org/
org/10.1038/s41598-022-17786-9. 10.1093/icesjms/fsx174.
Lo, N.C.-h, Jacobson, L.D., Squire, J.L., 1992. Indices of relative abundance from fish Palmer, M., Quetglas, A., Guijarro, B., Moranta, J., Ordines, F., Massutí, E., 2009.
spotter data based on delta-lognornial models. Can. J. Fish. Aquat. Sci. 49, Performance of artificial neural networks and discriminant analysis in predicting
2515–2526. fishing tactics from multispecific fisheries. Can. J. Fish. Aquat. Sci. 66, 224–237.
Lynch, H.J., Thorson, J.T., Shelton, A.O., 2014. Dealing with under-and over-dispersed Palomares, M.L., Pauly, D., 2019. On the creeping increase of vessels’ fishing power.
count data in life history, spatial, and community ecology. Ecology 95, 3173–3180. Ecol. Soc. 24 https://doi.org/10.5751/Es-11136-240331.
https://doi.org/10.1890/13-1912.1. Parker, D., Winker, H., West, W., Kerwath, S.E., 2017. Standardization of the catch per
Mac Nally, R., Duncan, R.P., Thomson, J.R., Yen, J.D., 2018. Model selection using unit effort for swordfish (Xiphias gladius) for the South African longline fishery.
information criteria, but is the “best” model any good? J. Appl. Ecol. 55, 1441–1444. Collect Vol. Sci. Pap. ICCAT 74, 1295–1305.
https://doi.org/10.1111/1365-2664.13060. Pearl, J., 2009. Causal inference in statistics: an overview. Stat. Surv. 3, 96–146.
Magee, L., 1998. Nonlocal behavior in polynomial regressions. Am. Stat. 52, 20–22. Peatman, T.; Nicol, S. Updated purse seine bycatch estimates in the WCPO. WCPFC-
Marr, J.C., 1951. On the use of the terms abundance, availability and apparent SC17–2021/ST-IP-06. WCPFC Scientific Committee, Seventeenth Regular Session.
abundance in fishery biology. Copeia 1951, 163–169. Online Meeting: Western and Central Pacific Fisheries Commission; 2021.
Marra, G., Wood, S.N., 2011. Practical variable selection for generalized additive models. Peatman, T., Fukofuka, S., Park, T., Williams, P., Hampton, J., Smith, N., 2019. Better
Comput. Stat. Data Anal. 55, 2372–2387. purse seine catch composition estimates: progress on the Project 60 work plan.
Marshall, E., Spiegelhalter, D., 2003. Approximate cross-validatory predictive checks in WCPFC-SC15-2019/ST-WP 02.
disease mapping models. Stat. Med. 22, 1649–1660. https://doi.org/10.1002/ Pelletier, D., Ferraris, J., 2000. A multivariate approach for defining fishing tactics from
sim.1403. commercial catch and effort data. Can. J. Fish. Aquat. Sci. 57, 51–65.
Martell, S., Stewart, I., 2014. Towards defining good practices for modeling time-varying Pennino, M.G., Paradinas, I., Illian, J.B., Muñoz, F., Bellido, J.M., López-Quílez, A.,
selectivity. Fish. Res. 158, 84–95. https://doi.org/10.1016/j.fishres.2013.11.001. Conesa, D., 2019. Accounting for preferential sampling in species distribution
Maunder, M.N.; Hinton, M.G. Estimating relative abundance from catch and effort data, models. Ecol. Evol. 9, 653–663. https://doi.org/10.1002/ece3.4789.
using neural networks. 2006. Perretti, C.T., Thorson, J.T., 2019. Spatio-temporal dynamics of summer flounder
Maunder, M.N., Piner, K.R., 2017. Dealing with data conflicts in statistical inference of (Paralichthys dentatus) on the Northeast US shelf. Fish. Res. 215, 62–68. https://doi.
population assessment models that integrate information from multiple diverse data org/10.1016/j.fishres.2019.03.006.
sets. Fish. Res. 192, 16–27. https://doi.org/10.1016/j.fishres.2016.04.022. Peterson, C.D., Belcher, C.N., Bethea, D.M., Driggers III, W.B., Frazier, B.S., Latour, R.J.,
Maunder, M.N., Punt, A.E., 2004. Standardizing catch and effort data: a review of recent 2017. Preliminary recovery of coastal sharks in the south-east United States. Fish
approaches. Fish. Res. 70, 141–159. https://doi.org/10.1016/j.fishres.2004.08.002. Fish. 18, 845–859.
Maunder, M.N., Hinton, M.G., Bigelow, K.A., Langley, A.D., 2006. Developing indices of Peterson, C.D., Wilberg, M.J., Cortés, E., Latour, R.J., 2021. Dynamic factor analysis to
abundance using habitat data in a statistical framework. Bull. Mar. Sci. 79, 545–559. reconcile conflicting survey indices of abundance. ICES J. Mar. Sci. 78, 1711–1729.
Maunder, M.N., Thorson, J.T., Xu, H., Oliveros-Ramos, R., Hoyle, S.D., Tremblay- https://doi.org/10.1093/icesjms/fsab051.
Boyer, L., Lee, H.H., Kai, M., Chang, S.-K., Kitakado, T., 2020. The need for spatio- Peterson, M.J., Mueter, F., Criddle, K., Haynie, A.C., 2014. Killer whale depredation and
temporal modeling to determine catch-per-unit effort based indices of abundance associated costs to Alaskan sablefish, Pacific halibut and Greenland turbot
longliners. PLoS One 9, e88906.

20
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

Phillips, A.J., Ciannelli, L., Brodeur, R.D., Pearcy, W.G., Childers, J., 2014. Spatio- Thorson, J.T., 2015. Spatio-temporal variation in fish condition is not consistently
temporal associations of albacore CPUEs in the Northeastern Pacific with regional explained by density, temperature, or season for California Current groundfishes.
SST and climate environmental variables. ICES J. Mar. Sci. 71, 1717–1727. https:// Mar. Ecol. Prog. Ser. 526, 101–112.
doi.org/10.1093/icesjms/fst238. Thorson, J.T., 2018. Three problems with the conventional delta-model for biomass
Pickens, B.A., Carroll, R., Schirripa, M.J., Forrestal, F., Friedland, K.D., Taylor, J.C., sampling data, and a computationally efficient alternative. Can. J. Fish. Aquat. Sci.
2021. A systematic review of spatial habitat associations and modeling of marine fish 75, 1369–1382. https://doi.org/10.1139/cjfas-2017-0266.
distribution: a guide to predictors, methods, and knowledge gaps. PloS One 16, Thorson, J.T., 2019a. Guidance for decisions using the Vector Autoregressive Spatio-
e0251818. https://doi.org/10.1371/journal.pone.0251818. Temporal (VAST) package in stock, ecosystem, habitat and climate assessments.
Pilling, G.; Brouwer, S. Report from the SPC pre-assessment workshop, Noumea, April Fish. Res. 210, 143–161. https://doi.org/10.1016/j.fishres.2018.10.013.
2017. WCPFC-SC13–2017/IP-02. 13th Regular Session of the Scientific Committee, Thorson, J.T., 2019b. Measuring the impact of oceanographic indices on species
9–17 August 2017. Rarotonga, Cook Islands; 2017. distribution shifts: The spatially varying effect of cold-pool extent in the eastern
Pinto, C., Travers-Trolet, M., Macdonald, J.I., Rivot, E., Vermard, Y., 2019. Combining Bering Sea. Limnol. Oceanogr. 64, 2632–2645.
multiple data sets to unravel the spatiotemporal dynamics of a data-limited fish Thorson, J.T., Barnett, L.A., 2017. Comparing estimates of abundance trends and
stock. Can. J. Fish. Aquat. Sci. 76, 1338–1349. https://doi.org/10.1139/cjfas-2018- distribution shifts using single-and multispecies models of fishes and biogenic
0149. habitat. ICES J. Mar. Sci. 74, 1311–1321. https://doi.org/10.1093/icesjms/fsw193.
Pitcher, T.J., Watson, R., Forrest, R., Valtýsson, H.Þ, Guénette, S., 2002. Estimating Thorson, J.T., Haltuch, M.A., 2019. Spatiotemporal analysis of compositional data:
illegal and unreported catches from marine ecosystems: a basis for change. Fish Fish increased precision and improved workflow using model-based inputs to stock
3, 317–339. https://doi.org/10.1046/j.1467-2979.2002.00093.x. assessment. Can. J. Fish. Aquat. Sci. 76, 401–414. https://doi.org/10.1093/icesjms/
Punt, A.E., Walker, T.I., Taylor, B.L., Pribac, F., 2000. Standardization of catch and effort fsw193.
data in a spatially-structured shark fishery. Fish. Res. 45, 129–145. Thorson, J.T., Minto, C., 2015. Mixed effects: a unifying framework for statistical
Quinn, T.J., Deriso, R.B., 1999. Quantitative Fish Dynamics. Oxford University Press. modelling in fisheries biology. ICES J. Mar. Sci. 72, 1245–1256. https://doi.org/
R Core Team, 2022. A language and environment for statistical computing. R Foundation 10.1093/icesjms/fsu213.
for Statistical Computing, Vienna, Austria. Thorson, J.T., Shelton, A.O., Ward, E.J., Skaug, H.J., 2015. Geostatistical delta-
Ralston, S. Trends in standardized catch rate of some rockfishes (Sebastes spp.) from the generalized linear mixed models improve precision for estimated abundance indices
California trawl logbook database. 1999. for West Coast groundfishes. ICES J. Mar. Sci. 72, 1297–1310. https://doi.org/
Rigby, R.A., Stasinopoulos, D.M., 2005. Generalized additive models for location, scale 10.1093/icesjms/fsw193.
and shape. J. R. Stat. Soc.: Ser. C. (Appl. Stat. ) 54, 507–554. https://doi.org/ Thorson, J.T., Fonner, R., Haltuch, M.A., Ono, K., 2016a. Winker, H. Accounting for
10.1111/j.1467-9876.2005.00510.x. spatiotemporal variation and fisher targeting when estimating abundance from
Roberts, J.J., Best, B.D., Mannocci, L., Fujioka, E., Halpin, P.N., Palka, D.L., Garrison, L. multispecies fishery data. Can. J. Fish. Aquat. Sci. 73, 1–14.
P., Mullin, K.D., Cole, T.V., Khan, C.B., 2016. Habitat-based cetacean density models Thorson, J.T., Ianelli, J.N., Larsen, E.A., Ries, L., Scheuerell, M.D., Szuwalski, C.,
for the US Atlantic and Gulf of Mexico. Sci. Rep. 6, 1–12. Zipkin, E.F., 2016b. Joint dynamic species distribution models: a tool for community
Roche, C., Guinet, C., Gasco, N., Duhamel, G., 2007. Marine mammals and demersal ordination and spatio-temporal monitoring. Glob. Ecol. Biogeogr. 25, 1144–1158.
longline fishery interactions in Crozet and Kerguelen Exclusive Economic Zones: an https://doi.org/10.1093/icesjms/fsw193.
assessment of depredation levels. CCAMLR Sci. 14, 67–82. Thorson, J.T., Pinsky, M.L., Ward, E.J., 2016c. Model-based inference for estimating
Rowden, A.A., Anderson, O.F., Georgian, S.E., Bowden, D.A., Clark, M.R., Pallentin, A., shifts in species distribution, area occupied and centre of gravity. Methods Ecol.
Miller, A., 2017. High-resolution habitat suitability models for the conservation and Evol. 7, 990–1002. https://doi.org/10.1093/icesjms/fsw193.
management of vulnerable marine ecosystems on the Louisville Seamount Chain, Thorson, J.T., Cunningham, C.J., Jorgensen, E., Havron, A., Hulson, P.-J.F.,
South Pacific Ocean. Front. Mar. Sci. 4, 335. https://doi.org/10.3389/ Monnahan, C.C., von Szalay, P., 2021. The surprising sensitivity of index scale to
fmars.2017.00335. delta-model assumptions: recommendations for model-based index standardization.
Rudd, M.B., Branch, T.A., 2017. Does unreported catch lead to overfishing? Fish Fish. 18, Fish. Res. 233, 105745 https://doi.org/10.1016/j.fishres.2020.105745.
313–323. Thygesen, U.H., Albertsen, C.M., Berg, C.W., Kristensen, K., Nielsen, A., 2017. Validation
Rufener, M.C., Kristensen, K., Nielsen, J.R., Bastardie, F., 2021. Bridging the gap between of ecological state space models using the Laplace approximation. Environ. Ecol.
commercial fisheries and survey data to model the spatiotemporal dynamics of Stat. 24, 317–339. https://doi.org/10.1007/s10651-017-0372-4.
marine species. Ecol. Appl. 31, e02453 https://doi.org/10.1002/eap.2453. Tian, S., Chen, X., Chen, Y., Xu, L., Dai, X., 2009. Standardizing CPUE of Ommastrephes
Sadovy, Y., Domeier, M., 2005. Are aggregation-fisheries sustainable? Reef fish fisheries bartramii for Chinese squid-jigging fishery in Northwest Pacific Ocean. Chin. J.
as a case study. Coral Reefs 24, 254–262. Oceanol. Limnol. 27, 729–739. https://doi.org/10.1007/s00343-009-9199-7.
Sampson, D.B., Scott, R.D., 2012. An exploration of the shapes and stability of Tremblay-Boyer, L.; Neubauer, P. Historical catch reconstruction and CPUE
population–selection curves. Fish Fish. 13, 89–104. https://doi.org/10.1111/j.1467- standardization for the stock assessment of oceanic whitetip shark in the Western
2979.2011.00417.x. and Central Pacific Ocean. WCPFC-SC15/SA-IP-17. Scientific Committee, 15th
Scherrer, K., Galbraith, E., 2020. The risk of underestimating long-term fisheries creep. Regular Session. Pohnpei, Federated States of Micronesia: Western and Central
Ecol. Soc. 25 https://doi.org/10.5751/Es-11389-250118. Pacific Fisheries Commission; 2019.
Schnute, J.T., Hilborn, R., 1993. Analysis of contradictory data sources in fish stock Walsh, W.A., Brodziak, J., 2015. Billfish CPUE standardization in the Hawaii longline
assessment. Can. J. Fish. Aquat. Sci. 50, 1916–1923. fishery: model selection and multimodel inference. Fish. Res. 166, 151–162. https://
Shelton, A.O., Thorson, J.T., Ward, E.J., Feist, B.E., 2014. Spatial semiparametric models doi.org/10.1016/j.fishres.2014.07.015.
improve estimates of species abundance and distribution. Can. J. Fish. Aquat. Sci. Walter, J.; Urtizberea Ijurco, A.; Hiroki, Y.; Satoh, K.; Ortiz, M.; Kimoto, A.; Ailloud;
71, 1655–1666. Matsumoto, T. Stock Synthesis model for Atlantic yellowfin tuna, SCRS/2019/121.
Shemla, A., McAllister, M.K., 2006. Bayesian generalized linear models to standardize ICCAT’s Collective Volume of Scientific Papers; 2020.
and impute missing data in the Atlantic bluefin tuna (Thunnus thynnus thynnus) Walters, C., 2003. Folly and fantasy in the analysis of spatial catch rate data. Can. J. Fish.
TaskII catch and effort database. ICCAT Collect. Vol. Sci. Pap. 59, 750–768. Aquat. Sci. 60, 1433–1436.
Shono, H., 2008a. Application of the Tweedie distribution to zero-catch data in CPUE Warton, D.I., Thibaut, L., Wang, Y.A., 2017. The PIT-trap—A “model-free” bootstrap
analysis. Fish. Res. 93, 154–162. https://doi.org/10.1016/j.fishres.2008.03.006. procedure for inference about regression models with discrete, multivariate
Shono, H., 2008b. Confidence interval estimation of CPUE year trend in delta-type two- responses. PloS One 12, e0181790. https://doi.org/10.1371/journal.pone.0181790.
step model. Fish. Sci. 74, 712–717. Watters, G.; Deriso, R. Catches per unit of effort of bigeye tuna: A new analysis with
Squires, D., 1992. Productivity measurement in common property resource industries: an regression trees and simulated annealing. Bulletin Inter-American Tropical Tuna
application to the Pacific coast trawl fishery. The. Rand J. Econ. 23, 221–236. Commission/Boletin Comision Interamericana del Atun Tropical. La Jolla CA.
https://doi.org/10.2307/2555985. [BullIATTC/BolCIAT] Vol 21; 2000.
Squires, D., Vestergaard, N., 2013. Technical change in fisheries. Mar. Policy 42, Webber, D.; Starr, P. Characterisation and CPUE from 1989–90 to 2018–19 of jack
286–292. https://doi.org/10.1016/j.marpol.2013.03.019. mackerel off the west coast of New Zealand (JMA 7). New Zealand Fisheries
Sribney, B. What are some of the problems with stepwise regression? 〈https://www. Assessment Report:06; 2022.
stata.com/support/faqs/statistics/stepwise-regression-problems/〉; 1996. Wiegand, R.E., 2010. Performance of using multiple stepwise algorithms for variable
Stephens, A., MacCall, A., 2004. A multispecies approach to subsetting logbook data for selection. Stat. Med. 29, 1647–1659.
purposes of estimating CPUE. Fish. Res. 70, 299–310. https://doi.org/10.1016/j. Wilberg, M.J., Thorson, J.T., Linton, B.C., Berkson, J., 2009. Incorporating time-varying
fishres.2004.08.009. catchability into population dynamic stock assessment models. Rev. Fish. Sci. 18,
Su, N.J., Yeh, S.Z., Sun, C.L., Punt, A.E., Chen, Y., Wang, S.P., 2008. Standardizing catch 7–24.
and effort data of the Taiwanese distant-water longline fishery in the western and Winker, H., Kerwath, S.E., Attwood, C.G., 2013. Comparison of two approaches to
central Pacific Ocean for bigeye tuna, Thunnus obesus. Fish. Res. 90, 235–246. standardize catch-per-unit-effort for targeting behaviour in a multispecies hand-line
Taylor, L.R., 1961. Aggregation, variance and the mean. Nature 189, 732–735. fishery. Fish. Res. 139, 118–131.
Tesfamichael, D., Pitcher, T.J., Pauly, D., 2014. Assessing changes in fisheries using Winker, H., Kerwath, S.E., Attwood, C.G., 2014. Proof of concept for a novel procedure to
fishers’ knowledge to generate long time series of catch rates: a case study from the standardize multispecies catch and effort data. Fish. Res. 155, 149–159.
Red Sea. Ecol. Soc. 19 https://doi.org/10.5751/Es-06151-190118. Wood, S.N., 2017. Generalized additive models: an introduction with R. Chapman and
Therneau, T.M.; Lumley, T.; Atkinson, E.; Crowson, C. A Package for Survival Analysis in Hall/CRC.
R. 〈https://cran.r-project.org/web/packages/survival/index.html〉; 2022. Xu, H.; Lennert-Cody, C.E. Standardizing the purse-seine indices of abundance and
Thorson, J. VAST model structure and user interface. 〈https://github.com/James-Thors associated length compositions for skipjack tuna in the eastern Pacific Ocean.
on-NOAA/VAST〉; 2022. Document SAC-13 INF-K. Scientific Advisory Committee, 13th meeting, 16–20 May
2022. Online: Inter-American Tropical Tuna Commission; 2022.

21
S.D. Hoyle et al. Fisheries Research 269 (2024) 106860

Xu, H., Lennert-Cody, C.E., Maunder, M.N., Minte-Vera, C.V., 2019. Spatiotemporal Zhou, S., Campbell, R.A., Hoyle, S.D., 2019. Catch per unit effort standardization using
dynamics of the dolphin-associated purse-seine fishery for yellowfin tuna (Thunnus spatio-temporal models for Australia’s Eastern Tuna and Billfish Fishery. ICES J.
albacares) in the eastern Pacific Ocean. Fish. Res. 213, 121–131. Mar. Sci. 76, 1489–1504. https://doi.org/10.1093/icesjms/fsz034.
Ye, Y., Dennis, D., 2009. How reliable are the abundance indices derived from Zuur, A.F., Saveliev, A.A., Ieno, E.N., 2012. Zero inflated models and generalized linear
commercial catch–effort standardization? Can. J. Fish. Aquat. Sci. 66, 1169–1178. mixed models with R. Highland Statistics Limited.
Zhang, J.; Li, S. A Review of Machine Learning Based Species’ Distribution Modelling.
2017 International Conference on Industrial Informatics-Computing Technology,
Intelligent Technology, Industrial Information Integration (ICIICII): IEEE; 2017.

22