Beruflich Dokumente
Kultur Dokumente
CSIRO Division of Marine Research, GPO Box 1538, Hobart, TAS 7001, Australia
Abstract
Constructionof annual indices of stockabundance basedoncatchandeffort data remains central tomanysheries assessments.
While the use of more advanced statistical methods has helped catch rates to be standardised against many explanatory variables,
the changing spatial characteristics of most sheries data sets provide additional challenges for constructing reliable indices of
stock abundance. After reviewing the use of general linear models to construct indices of annual stock abundance, potential
biases which can arise due to the unequal and changing nature of the spatial distribution of shing effort are examined and
illustrated through the analysis of simulated data. Finally, some options are suggested for modelling catch rates in unshed strata
and for accounting for the uncertainties in the stock and shery dynamics which arise in the interpretation of spatially varying
catch rate data.
2004 Elsevier B.V. All rights reserved.
Keywords: Standardisation of catch rates; General linear models; Stock abundance indices; Spatial distribution of shing effort; Modelling
uncertainty
1. Introduction
Despite an ongoing debate about the nature of the re-
lationship between catch rates and underlying resource
abundance (e.g. Harley et al., 2001), the interpretation
of catch and effort data, and the construction of in-
dices of resource abundance based on these data, re-
mains an integral part of the stock assessment process
2
)
= exp(
o
+y
j
+q
j
+r
k
+(yr)
ik
) exp(
1
2
2
)
= a
o
q
o
D
o
exp(y
j
+q
j
+r
k
+(yr)
ik
) exp(
1
2
2
),
E(CPUE
ijkoo
) = CPUE
o
exp(y
j
+q
j
+r
k
+(yr)
ik
+
1
2
2
) (8)
where CPUE
o
= a
o
q
o
D
o
is the standardised catch rate
in the reference spatio-temporal stratum. Note that
inclusion of in Eq. (8) can result in the expected
catch rate being less than zero. In these cases, the
expected catch rate should be set to zero.
From Eq. (4a), a relative index of abundance, B
ijk
,
for the size of the sh population in the ith year, jth
quarter and kth region can be obtained by multiplying
the standardised catch rate by the size, A
k
, of the region
shed, i.e. B
ijkoo
= A
k
E(CPUE
ijkoo
). The total index of
abundance for a season is then obtained by summing
over all regions of the shery. An annual index of abun-
dance for the ith year, I
i
, can then be obtained by taking
the average over all seasons inthat year. Either the arith-
metic or geometric mean can be used, the latter being
scale invariant:
I
i
=
1
NS
NS
j=1
_
NR
k=1
A
k
E(CPUE
ijkoo
)
_
(9a)
I
i
=
NS
_
NS
j=1
_
NR
k=1
A
k
E(CPUE
ijkoo
)
_
(9b)
R.A. Campbell / Fisheries Research 70 (2004) 209227 213
where NS and NR are the number of seasons (quarters)
and regions used in the analysis, respectively. Seasonal
weightings can be used if desired. Finally, a relative
index which relates the average abundance in year i to
some reference year can be calculated as:
I
i,ref
=
I
i
I
ref
For an analysis based on Eq. (8), the annual index of
abundance (using the arithmetic mean) is given by:
I
i
= a
o
q
o
D
o
exp
_
y
i
+
2
2
_
_
_
1
NS
NS
j=1
NR
k=1
A
k
exp(q
j
+r
k
+(yr)
ik
)
_
_
NR
For those models where the interaction termis not used
and the sizes of the areas shed remain constant across
years, the term in the square brackets is constant for
all years. If one also ignores the term involving the
constant , the relative index is then reduced to:
I
i,r
= exp(y
i
y
r
) (10)
In this situation, the year effects alone are the indices
of abundance. However, for models which incorporate
interaction terms incorporating the year effect, the ex-
pression for the relative index is more complicated,
involving the sum over a number of other explanatory
variables.
The formulation of the annual abundance index (Eq.
(9)) highlights the fact that the index is the product of
the density of sh within several spatial areas and the
sizes of those areas. The common practice of reducing
the index to a function of the year effect alone runs the
risk of ignoring information on the spatial dynamics
of the shery which may be relevant to the underlying
dynamics of the stock and hence the correct interpreta-
tion of the catch and effort data. Additionally, Eq. (9)
will give indices of total stock abundance only if the
spatial extent of the shery coincides with, or is greater
than, the spatial extent of the stock. Otherwise, the in-
dex of abundance pertains only to that portion of the
stock that is found on the shing grounds. Uncertainty
will remain as to the size of the stock beyond the area
shed.
3. Inuence of an unequal spatial distribution
of shing effort
The equations dened in the previous section made
no assumptions about the spatial characteristics of
the shery. However, a sh population is usually dis-
tributed with variable densities across its stock range
and consequently the spatial distribution of shing ef-
fort is also usually highly variable. The shery is usu-
ally divided into a number of regions, and estimates of
stock density are obtained for each region to account
for this spatial heterogeneity. For example, consider a
regulatory area of total size A consisting of R sepa-
rate regions. The total abundance, N, across the entire
area can be related to the regional abundances, N
r
, and
densities, D
r
, by:
N =
R
r=1
N
r
=
R
r=1
A
r
D
r
where A
r
is the size of region r. Ideally the spa-
tial distribution of the resource within each region
should be reasonably homogeneous and the corre-
sponding estimates of density should be based on a
random sample from across each region. Assuming
that the catchability coefcient q is constant across
all regions (and availability in each region is unity),
and using the relationship between the mean den-
sity and the mean of the individual observed catch
rates CPUE
ri
in each region given by Eq. (2), we
have:
N =
R
r=1
N
r
=
R
r=1
A
r
CPUE
r
q
=
1
q
R
r=1
_
A
r
n
r
n
r
i=1
CPUE
ri
_
(11)
where n
r
is the number of CPUEobservations in the rth
region. Hence, the average of the regional catch rates
weighed by the size of each region gives an unbiased
estimate of the total stock abundance across the entire
regulatory area (Quinn and Hoag, 1982).
In terms of the structure of a GLM, the above situ-
ation can be described by:
P
r
= E(CPUE
r
) = +R
r
214 R.A. Campbell / Fisheries Research 70 (2004) 209227
where is a constant, R
r
and P
r
are the region-effect
and the predicted catch rate in the rth region respec-
tively, and where, for illustrative purposes, we have
used an additive model for CPUE instead of a mul-
tiplicative model. Values of the parameters and R
r
can be found from a least-squares t, which minimises
the sum of squares of the differences between the n
r
observed (Z
ri
) and predicted catch rates across all re-
gions. There is no loss of generality by setting R
1
= 0
(in which case parameter gives the expected catch
rate in region 1), so that:
=
Z
1
and
R
r
=
Z
r
Z
1
where
Z
r
=
1
n
r
n
r
i=1
Z
ri
Using the predicted value of the catch rate within each
region, the population index for the total area is found
to be:
I =
R
r=1
A
r
P
r
q
=
R
r=1
A
r
( +R
r
)
q
=
R
r=1
A
r
Z
r
q
That is, the index is proportional to the weighted mean
of the average observed catch rate in each region, and
according to Eq. (11), this gives an unbiased estimate
of the total population abundance over the entire regu-
latory area.
An interesting feature of the above result is that
an unbiased estimate of abundance is obtained even
though the number of observations in each region may
not be equal. However, a further example will show
that this is not the case when one considers more than
one year of observations. Again, consider observations
from NY years for a shery divided into NR regions
with n
yr
catch rate observations in the rth region dur-
ing the yth year. We t the following model to these
observations:
P
yr
= E(CPUE
yr
) = +Y
y
+R
r
For simplicity, we consider the situation where NY =
NR = 2 and again, with no loss of generality, we set
Y
2
= R
2
= 0. The following least-squares solution is
found:
=
_
1
n
21
+
1
n
12
+
1
n
11
_
Z
22
+
1
n
11
(
Z
21
+
Z
12
Z
11
),
Y
1
=
_
1
n
22
+
1
n
12
_
(
Z
11
Z
21
)
+
_
1
n
11
+
1
n
21
_
(
Z
12
Z
22
),
R
1
=
_
1
n
22
+
1
n
21
_
(
Z
11
Z
12
)
+
_
1
n
11
+
1
n
12
_
(
Z
21
Z
22
)
where
=
_
1
n
11
+
1
n
12
+
1
n
21
+
1
n
22
_
and
Z
yr
=
1
n
yr
n
yr
i=1
Z
yri
Based on the above model, the population index in the
rst year, given that each region is of area A, is given
by
I
1
=
A( +Y
1
+R
1
)
q
+
A( +Y
1
+R
2
)
q
=
A(2 +2Y
1
+R
1
)
q
=
A[
Z
11
(2/n
12
+1/n
21
+1/n
22
)
+
Z
12
(2/n
11
+1/n
21
+1/n
22
)
+
Z
21
(1/n
11
1/n
12
) +
Z
22
(1/n
12
1/n
11
)]
q
Unlike the previous example, the predicted abundance
index is no longer proportional to the weighted mean
of the observed catch rates across each region, but is
instead dependent in a complex manner on the number
of observations in each region. Only in the special sit-
uation where the n
yr
are all equal does the total index
equal the weighted mean of the regional estimates. A
consequence of this result is that misleadingdifferences
between the indices of annual population size can be
obtained. To illustrate this point, consider the example
where
Z
11
=
Z
22
= 5,
Z
12
=
Z
21
= 10, n
11
= n
21
= n
22
R.A. Campbell / Fisheries Research 70 (2004) 209227 215
= 50, n
12
= 200 and A = q = 1. Based on these values,
the sum of the catch rates across the two regions is the
same for both years (15) but based on the results of the
least-squares t we obtain I
1
= 17.3 and I
2
= 15.
The above examples illustrate that when the number
of observations in each spatio-temporal stratum varies
(i.e. the data set is unbalanced) the relative indices of
annual abundance based on the parameter estimates
obtained from a GLM may be biased. This is due to
the fact that equal weight is given in the estimation
procedure to each observation instead of giving equal
weight to each region, as demanded by Eq. (11). Thus,
the annual indices based on the least-squares t will be
biased to favour those regions with the most number of
observations.
Aweighted least-squares approach needs to be used
to obtain an unbiased index of population abundance.
For example, if one weights each observation by the in-
verse of the number of observations in the correspond-
ing stratum, one obtains the corresponding solution:
=
1
4
(3
Z
22
+
Z
21
+
Z
12
Z
11
),
Y
1
=
1
2
(
Z
11
+
Z
12
Z
21
Z
22
),
R
1
=
1
2
(
Z
11
+
Z
21
Z
12
Z
22
)
From this result we obtain I
1
= A(
Z
11
+
Z
12
) and
I
2
= A(
Z
21
+
Z
22
) as desired. Note that a similar re-
sult is also obtained when the full model is tted to the
data (i.e. when a year region term is included). In
this case, the number of parameters equals the number
of spatio-temporal strata and parameter estimates can
be found which are independent of the relative number
of observations in each region. The single year model
described previously was a simple (if trivial) example
of this situation. The annual indices are not equal to the
weighted means of the catch rates in each region when
the regions are of different sizes. This indicates that
when constructing indices of abundance based on the
results of a GLM analysis, regions of equal size should
be used.
The weighting to be given to each observation to
achieve an unbiased annual abundance index is not
unique; indeed, any weights that satisfy the condition
that the sum of the individual weights given to each
observation in each region is the same for all regions
will ensure that all regions are treated equally. For the
observations within each region, the weight assigned
to each observation will itself ensure the importance of
that observation. A suitable weighting factor for each
observation may be based on the corresponding effort
of that observation divided by the total effort in that
region. This would be most appropriate in situations
where aggregated catch rate observations are used in
the analysis. An example of such a weighting factor,
which also takes into consideration the fact that the
logarithm of the catch rates is taken, is described by
Punsley (1987). A consequence of a weighted estima-
tion procedure is that the squared residuals from the
model are no longer
2
-distributed, thus invalidating
the F-test used to determine signicance among differ-
ent models. Again, Punsley (1987) suggests an alter-
native approach.
4. Potential biases in a spatially contracting
shery
The previous section considered the inuence on pa-
rameter estimation of an unequal distribution of shing
effort across regions. However, for many sheries the
number of regions shed each year can also vary. Such
changes may be inuenced by perceived shifts in the
distribution of the sh and/or changes in spatial man-
agement arrangements. Increased knowledge about the
spatial distribution of the resource and technical im-
provements in the ability to nd and target those areas
having greater abundance will also inuence the dis-
tribution of shing effort. These changes may occur
on both large- and ne-spatial scales and, as shall be
demonstrated, each has ramications for the manner in
which catch and effort data need to be interpreted. The
importance of any contraction in the spatial distribution
of effort will be magnied if shing effort concentrates
into regions with generally high catch rates. An exam-
ple of shery in which this has occurred is that for
southern bluen tuna.
4.1. Spatial changes in the shery for southern
bluen tuna
Southern bluen tuna (SBT) have a widespread but
patchy distribution, which is reected in the spatial dis-
tribution of the shing effort for this species. Like other
tunas, SBT also tend to form transient aggregations in
areas where oceanic thermal features favour local en-
216 R.A. Campbell / Fisheries Research 70 (2004) 209227
Fig. 1. Statistical areas used to provide the coverage of the shery for southern bluen tuna.
richment. The Japanese longline shery for southern
bluen tuna has undergone remarkable changes since
its inception as a major shery in the early 1950s.
Shingu and Hisada (1971) outline the changes between
1957 and 1969, during which the area exploited by this
shery expanded by about nine-fold. The statistical ar-
eas for the shery are shown in Fig. 1.
No major new shing grounds for SBT have been
discovered since 1971. However, possible shifts in sh-
ing effort to favourable areas and other changes in the
spatial distribution of effort since 1971 have created
problems for the interpretation of catch and effort data.
The number of 1
-squares
shed each year were ordered by the amount of shing
effort (number of hooks), and the cumulative percent-
age of the total annual effort expended in each decile of
the squares shed each year was then calculated. These
results were then averaged over each 5-year period
between 1970 and 1994, and used to plot cumulative
effort against cumulative area shed for each 5-year
period (Fig. 3a). During 19701974, on average 94%
of shing effort occurred in only 50%of the 1
-squares
shed, with 44%in the top 10%. This pattern of spatial
aggregation is repeated in all subsequent periods.
There appears to be little change during 19701984,
but the proportion of the effort expended in the top
10% of squares increased substantially after this time,
reaching 68% during 19901994. This increase in the
level of aggregation appears to have been a feature of
the shery since 1986 and may be a response to the
introduction of total catch quotas in the mid-1980s.
The relationship between the distribution of effort
and stock density (using catch rates as a proxy) was
analysed to investigate the extent of targeting of areas
of higher stock densities; cumulative effort was plotted
against area shed after ranking the 1
-squares by catch
rate (Fig. 3b). For 19711974, on average 56% of the
effort was expended in the top 50% of squares shed
and only 8% in the top 10% of 1
-squares indicating
that there was not very effective targeting of the areas
with higher nominal catch rates. However, the extent
of spatial targeting increased substantially over time.
This reached its greatest extent during the 19851989
when 78% of the effort was targeted in the top 50%
of 1
-squares shed by Japanese longliners in each of the SBT statistical areas 49 (19691995).
of strata with no data on the estimation of annual abun-
dance indices using GLMs. As a reference case, the
simulated shery was divided into a number of spatial
areas and catch and effort data were generated uni-
formly across these areas over ve years. These data
were then standardized to obtain an index of relative
abundance for the shery. Variations on the reference
case were then explored to ascertain the impact on the
calculated abundance index of spatial aggregation and
contraction of the shery over time. The sensitivity of
the resulting abundance index to differences between
the underlying spatial scale used to generate the catch
and effort data for the shery and the spatial scale
assumed in the standardising model was also inves-
tigated.
The simulated shery was deemed to consist of 200
grids of equal size. Multiple catch rate observations
were generated for each grid using the model:
log(CPUE
ik
) = Y
i
+G
ik
(12)
218 R.A. Campbell / Fisheries Research 70 (2004) 209227
Fig. 3. Cumulative effort vs. cumulative area shed (both expressed as a percentage of the respective annual totals) for the SBTshery in statistical
area 7 during each 5-year period between 1970 and 1994, after ordering the 1
k=1
CPUE
std
ik
,
B-avg
ij
=
N
max
j
N
obs
ij
N
obs
ij
k=1
CPUE
std
ik
Given the ner spatial scale of the analysis, several
other indices may also be dened. First, one can dene
the B-min index, which is similar to the B-avg index
but assumes that the catch rates in those grids, which
are not shed are, on average, equal to the minimum of
the catch rates in the shed grids. Alternatively, one can
dene the B-target index which assumes that the spatial
extent of the stock remains the same for all years and
is equivalent to the maximum number of grids shed
in any year, but assumes that the grids shed in any
year coincide with those grids with the highest catch
rates (i.e. there is prefect targeting). The catch rates in
those grids not shed each year are then modelled by
the tail of the average distribution of catch rates across
the maximal extent of grids shed, i.e. for each region:
(a) For each year, sort the standardised catch rates in
the grids shed in descending order and nd the
maximum standardised catch rate, CPUE
max
ij
.
(b) Calculate the relative index for each grid:
CPUE
rel
ir
=
CPUE
std
ir
CPUE
max
ij
r = 1, . . . , N
obs
ij
(c) Calculate the mean relative index for each grid
CPUErel
r
across those years when all grids are
shed in that region (i.e. when N
obs
ij
= N
max
j
).
(d) The expected standardised catch rates for those
grids with no observations are then modelled as:
CPUE
mod
ir
= CPUE
max
ij
CPUE
rel
r
r = N
obs
ij
+1, . . . , N
max
j
An abundance index for the region can then be
dened as:
B-target
ij
=
N
obs
ij
k=1
CPUE
std
ik
+
N
max
j
k=N
obs
ij
+1
CPUE
mod
ik
Finally, the annual abundance indices are calcu-
lated by summing across the regional indices for
each year. An index for a region with no observa-
tions can be modelled as in Section 5.1.
5.3. Results
Annual abundance indices were calculated for each
of the 30 data sets for scenarios 3 and 4, using both the
region- and grid-scale analyses described above. The
224 R.A. Campbell / Fisheries Research 70 (2004) 209227
Fig. 6. Comparison of the relative indices of abundance based on modelling of catch rates in unshed spatial strata for various hypotheses
concerning the spatial distribution of the stock and shing effort.
mean of the calculated indices across all data sets for
each scenario, relative to the value of the corresponding
index for the last year, were then calculated (Fig. 6).
For the region-scale analyses, the B-avg index was
similar to, if slightly worse, than the indices calculated
using the GLM analyses in Fig. 5. However, for sce-
nario 3 the region-scale annual B-zero and B-avg in-
dices bracketed the true annual index. While it is useful
to obtain a set of indices which bound the true index,
the difference between the B-zero and B-avg indices is
so large that considerable uncertainty remains as to the
true state of the stock over time. On the other hand, for
scenario 4 both indices under-estimated the true state
of the annual index.
The grid-scale annual B-zero and B-avg indices
bracketed the true annual index for scenarios 3 and 4.
However, as before, considerable uncertainty remained
regarding the true value of the index in any year. On the
other hand, both the B-min and B-target indices were
considerably closer to the true annual index, with the
B-target index being more accurate. This result is due to
the fact that the assumptions used in constructing these
latter indices more closely represent the true dynamics
of the stock and the shery.
The results presented here are limited to ve ways
of constructing indices. This should, however, in no
way limit the nature or the number of indices which
can be constructed. Indeed, the nature of the stock and
effort assumptions used to construct the various indices
should be based on an understanding of the actual stock
and shery dynamics of the shery being analysed
(Campbell and Tuck, 1996). For example, one could
use the results concerning the levels of targeting of high
catch grids each year to weight the effort assumption
in the B-target index for the SBT shery. Alternatively,
the existing individual indices could be combined us-
R.A. Campbell / Fisheries Research 70 (2004) 209227 225
ing different weights in different years. For example,
Hearn and Polacheck (1996) used the concept behind
the density-dependent habitat model of MacCall (1990)
as a basis for constructing an index that assigns differ-
ent annual weights to two indices based on the B-zero
and B-avg indices.
6. Discussion
Commercial catch and effort data continue to be re-
lied upon to estimate annual indices of stock abundance
in the absence of shery-independent data. While
GLMs and other statistical techniques have improved
our ability to standardise such data, problems still per-
sist. While some of these problems relate to the choice
of the most appropriate model and error structure, and
the absence of data on factors which are likely to in-
uence catch rates, there are more general problems of
deciding whether catch rate data from a shery under-
going changes in the spatial allocation of shing effort
can, in fact, reect stock abundance.
The analyses in this paper have illustrated the
manner in which biases can enter into the estimates
of annual stock abundance due to the unbalanced and
changing spatial distribution of shing effort. While
these biases generally relate to changes in the spatial
characteristics of the shery (either for the stock or
the shing effort), biases can also arise due to a lack
of spatial detail in the analyses due to inappropriate
model structures or the use of too coarse a spatial
level of data aggregation. While the potential for such
biases is generally acknowledged, the manner in which
these biases arise in the GLM analyses commonly
used to model catch and effort data, and how they
can be dealt with, do not appear to be generally
appreciated.
The issue of an unequal spatial distribution of sh-
ing effort and the preferential targeting of areas with
higher catch rates across the spatial areas used for a
GLM analysis can be corrected for by an appropriate
weighting. However, additional biases and uncertain-
ties arise due to missing observations, i.e. the areas of
the shery which are not shed. The extent to which
the shing grounds are known to overlap the spatial ex-
tent of the stock becomes increasingly uncertain when
there is a spatial contraction of the shery over time.
The characteristics of the stock in areas not shed pre-
viously similarly remain uncertain for an expanding
shery.
Given these uncertainties, it is usually not possible
to calculate a single reliably unbiased index of stock
abundance. Instead, it may be preferable to calculate a
number of indices based on modelling the likely catch
rates in those areas not shed using various assump-
tions about the spatial distribution of both the stock and
the shing effort (i.e. concerning the presence or not of
sh in the areas not shed and the targeting practices
of the shers). Support for or rejection of the assump-
tions underlying the calculation of the various indices
can then be based on a spatial analysis of the data for
the shery itself and/or an understanding of the deci-
sion rules used by shers to allocate shing effort spa-
tially, the behaviour observed in other sheries or from
the ecological considerations. For example, changes in
the spatial range of a sh population may be consis-
tent with observations from other animal populations
and with the theory of density-dependent habitat selec-
tion (MacCall, 1990). Spatial contractions in a shery
to areas with high catch rates may also be consistent
with economic practices associated with competitive
quotas.
An advantage of constructing a number of indices
based on modelling the likely catch rates in those areas
not shed using various assumptions about the spatial
distribution of both the stock and the shing effort is
that any subsequent assessment can make use of a range
of indices which explicitly incorporate a full range of
uncertainty about the data instead of relying only on
a single CPUE-based tuning index (Polacheck et al.,
1996). Indeed, given the lack of independence among
catch rate observations within (and perhaps between
adjacent) strata, which results in an over-estimation of
the number of degrees of freedom in a GLM analysis,
the true uncertainty associated with any single abun-
dance index is usually under-estimated.
Ultimately, the interpretation of catch rates and the
construction of indices of stock abundance should be
based on an understanding of the dynamics underlying
the spatial distribution of both the stock and the sh-
ing effort, and preferably on the relationship between
them. In many sheries, this will entail the need for
surveys to understand and reduce the uncertainties in
the spatial characteristics of the stock in those areas
presently unshed. An experimental shing program
within the SBT shery was undertaken for this purpose
226 R.A. Campbell / Fisheries Research 70 (2004) 209227
(Anon, 1996). There will also be a need to more fully
understand the decision rules for the targeting prac-
tices of individual shing vessels. Furthermore, in or-
der to overcome the potential biases which can result
from using catch and effort data from a shery with a
high degree of spatial targeting, analysis of commercial
catch and effort data to obtain annual indices of relative
stock abundance should be carried out at the nest spa-
tial scale possible. For the SBT longline shery this is
likely to be at the 1