Transactions of the American Fisheries Society 134:68–85, 2005
q Copyright by the American Fisheries Society 2005
[Article]
Do Habitat Actions Affect Juvenile Survival? An InformationTheoretic Approach Applied to Endangered Snake River
Chinook Salmon
CHARLES M. PAULSEN*
16016 Southwest Boones Ferry Road, Suite 4,
Lake Oswego, Oregon 97035, USA
TIMOTHY R. FISHER
Fisher Fisheries, Limited, 18403 South Clear Acres Drive,
Oregon City, Oregon 97045, USA
Abstract.—We used 11 years of parr-to-smolt survival estimates from 33 Snake River sites to
demonstrate that despite a number of confounding factors higher numbers of past habitat remediation or enhancement actions are associated with higher parr-to-smolt survival of endangered
wild Snake River spring2summer (stream-type) Chinook salmon Oncorhynchus tshawytscha. Information-theoretic weights were applied to help distinguish between statistical models based on
their relative plausibility. In the models with the highest estimated weights, actions taken to improve
fish habitat showed a positive association with increased parr-to-smolt survival. However, because
the actions were not sited randomly on the landscape, and because the actions may also have
influenced other potentially important covariates, it is difficult to separate habitat action effects
from effects due to other important factors.
A recent National Marine Fisheries Service
(NMFS) Biological Opinion on Chinook salmon
Oncorhynchus tshawytscha (which are listed under
the U.S. Endangered Species Act) requires the
Bonneville Power Administration (BPA) and the
U.S. Army Corps of Engineers to increase egg-toparr survival rates for juvenile salmon rearing in
tributaries (NMFS 2000). It also requires the agencies to demonstrate, in a statistically defensible
manner, that the proportion of eggs surviving to
adulthood (i.e., egg-to-adult survival) has increased.
In light of past expenditures made on improving
anadromous salmonid and resident fish habitat
(e.g., US$310 million by the BPA from 1978 to
1999 [NPPC 2001]), one might think that many
analyses would demonstrate increases in survival
rates resulting from habitat remediation, but this
is not the case. For example, a recent literature
survey of over 2,000 references (Bayley 2002) uncovered only a few studies that used statistically
rigorous experimental designs to demonstrate the
effects of habitat modifications on salmonid survival throughout the life cycle. While there are a
few recent exceptions for coastal salmonid stocks
in northwestern North America (e.g., Solazzi et al.
2000; Ward et al. 2003), empirical studies in this
subject area are very rare. Indeed, studies by Solazzi et al. (2000) and Ward et al. (2003) were the
only ones we could locate that have addressed
changes in salmonid survival during discrete life
cycle stages, as distinct from changes in local
abundance, and we found no studies of inland
stream-type Chinook salmon stocks.
Our recent work (Paulsen and Fisher 2003) suggests that the effects of fish habitat improvement
or remediation actions (hereafter referred to as actions or habitat actions) on parr-to-smolt survival
(the proportion of tagged parr released in their
natal streams that survive to Lower Granite Dam
[LGR] as out-migrating smolts) can be detected
relatively quickly. For example, by use of five control sites and three randomly assigned treatment
sites, we showed that it is theoretically possible
to detect a 30% increase in parr-to-smolt survival
rates within 7–9 years at a power of 80% and a
significance level a of 5%. We also demonstrated
that a doubling of parr-to-smolt survival rates (a
100% increase) could be detected in as little as 1–
2 years. Our analysis, however, assumed idealized
conditions that are unlikely to hold in large-scale,
real-world habitat manipulations. These assumptions included a 10-year pretreatment period during which no actions took place at any of the eight
sites. We also assumed a uniform, multiplicative
increase in parr-to-smolt survival (e.g., a 50% in-
* Corresponding author:
[email protected]
Received December 16, 2003; accepted August 5, 2004
68
DO HABITAT ACTIONS AFFECT JUVENILE SURVIVAL?
69
FIGURE 1.—Map of the Snake River basin study area. Fish symbols indicate the approximate tagging location
for each Chinook salmon stock that received passive integrated transponder tags. Locations of dams on the Snake
River are also indicated.
crease in a 10% pretreatment survival yielded a
15% posttreatment survival, while the same increase for a 20% pretreatment case resulted in 30%
posttreatment survival) for all treatment sites simultaneously. Finally, we assumed that treatment
would not alter any of the independent variables
used in the analysis, including the size (length) of
juvenile fish at tagging and parent spawner density.
In the current work, we instead ask a related but
very different question: is it possible to detect the
effects of past actions on parr-to-smolt survival
from existing information? As one might expect,
practical problems arise when examining real habitat actions, since the assumptions noted above do
not generally hold in the real world. The problems
occur because past actions are not carried out in
an experimental framework, and because there are
no broad-scale, systematic surveys of habitat conditions for these sites. The analysis described here
is an initial attempt to discover whether these problems can be surmounted in order to disentangle
the parr-to-smolt survival effects of habitat actions
(if indeed such effects exist) from the many other
factors that may affect parr-to-smolt survival rates.
The first problem is that sites of real actions are
not chosen at random. Instead, locations for actions are selected because of perceived local habitat problems, cooperative landowners, ease of access, availability of funding, and a host of other
reasons. In fact, sites with many actions may have
lower juvenile survival than those with few or
none, not because the actions are ineffective, but
because juvenile survival is initially lower at action-intensive, anthropogenically degraded sites
(e.g., agricultural lands) compared to sites where
few or no actions have been taken (e.g., wilderness
areas). For example, agricultural and grazing
lands, which one would expect to have low parrto-smolt survival rates relative to undisturbed habitats (Paulsen and Fisher 2001), have been the setting for many actions in the Snake River subbasins.
The majority of actions in Chinook salmon spawn-
70
PAULSEN AND FISHER
TABLE 1.—Snake River basin sites at which passive integrated transponder tagged Chinook salmon parr were released
and the site-specific cumulative total number of actions taken to improve fish habitat for each stock. Site–year combinations with less than 100 tagged parr are left blank because we did not estimate survival. Sites with a survival estimate
but no cumulative actions for a given year are denoted NA.
Year of tagging
Stock
State climate division
1992
1993
1994
1995
1
9
1996
Clearwater River subbasin
American River
Clear Creek
Crooked Fork Creek
Crooked River
Legendary Bear Creek
Lolo Creek
Meadow Creek (Selway)
Newsome Creek
Red River
Central Mountains
NA
1
9
NA
1
9
4
4
NA
NA
1
9
3
4
19
19
2
19
Catherine Creek
Imnaha River
Lostine River
Minam River
Upper Grand Ronde River
Lookingglass Creek
Northeast Oregon subbasin
Northeast Oregon
3
5
9
18
3
4
NA
2
14
16
Palouse/Blue Mountains
NA
23
31
13
3
20
1
Bear Valley Creek
Big Creek
Camas Creek
Cape Horn Creek
Elk Creek
Loon Creek
Marsh Creek
Sulfur Creek
Middle Fork Salmon River subbasin
Central Mountains
2
2
NA
NA
1
1
NA
2
2
NA
NA
1
1
NA
2
NA
1
NA
2
NA
1
NA
Johnson Creek
Lake Creek
Secesh River
South Fork Salmon River
South Fork Salmon River subbasin
Central Mountains
5
1
1
NA
NA
1
1
5
1
NA
2
Upper Salmon River
Valley Creek
E. Fork Salmon River
Herd Creek
Lemhi River
Pahsimeroi River
Upper Salmon River subbasin
Central Mountains
17
18
5
6
Northeast Valleys
6
8
Northeast Valleys
2
2
10
29
1
19
6
8
2
45
4
ing and rearing areas focus on irrigation diversions
and privately owned riparian cattle pastures (see
GRMWP [2003] and USBWP [2003] for general
descriptions and locations of projects carried out
in two heavily irrigated and grazed subbasins).
Second, in our previous power analysis (Paulsen
and Fisher 2003), we assumed that actions would
not influence covariates that were important for
explaining parr-to-smolt survival rates (e.g., juvenile size at tagging and parent spawner abundance). However, actions may have some effect on
juvenile size at tagging, since improved habitat
may contain more food and thereby produce increased growth rates. Further, one would hope that
1
4
NA
24
33
46
21
3
34
50
26
3
2
1
1
NA
2
1
NA
2
19
14
90
12
133
26
actions will eventually result in increased spawner
abundance, since increased parr-to-smolt survival
rates should lead to increases in adults as the affected fish mature and return to spawn. If parr-tosmolt survival is sensitive to redd density (Rickertype density dependence), however, this may introduce an additional complication: increased parrto-smolt survival may lead to increased adult
spawner density, eventually leading to decreased
parr-to-smolt survival for future generations of juveniles.
Third, the power analysis assumed that a beforeafter–control-impact (BACI) design (Osenberg
and Schmitt 1996) was possible such that in the
71
DO HABITAT ACTIONS AFFECT JUVENILE SURVIVAL?
TABLE 1.—Extended.
Year of tagging
Stock
State climate division
1997
1998
American River
Central Mountains
Clear Creek
Crooked Fork Creek
Crooked River
Legendary Bear Creek
Lolo Creek
Meadow Creek (Selway)
Newsome Creek
Red River
Clearwater River subbasin
NA
NA
NA
1
1
9
3
7
7
NA
2
24
24
Catherine Creek
Northeast Oregon
Imnaha River
Lostine River
Minam River
Upper Grand Ronde River
Lookingglass Creek
Palouse/Blue Mountains
Northeast Oregon subbasin
39
44
53
56
28
31
3
3
40
41
2
3
Bear Valley Creek
Big Creek
Camas Creek
Cape Horn Creek
Elk Creek
Loon Creek
Marsh Creek
Sulfur Creek
Central Mountains
Johnson Creek
Central Mountains
Lake Creek
Secesh River
South Fork Salmon River
Upper Salmon River
Valley Creek
E. Fork Salmon River
Herd Creek
Lemhi River
Pahsimeroi River
Central Mountains
1999
NA
NA
1
9
3
8
NA
2
24
48
61
31
3
43
Middle Fork Salmon River subbasin
2
2
2
NA
NA
2
NA
NA
2
3
3
NA
NA
1
3
3
NA
NA
South Fork Salmon River subbasin
7
7
3
3
3
NA
NA
NA
2
2
2
Upper Salmon River subbasin
30
35
23
Northeast Valleys
Northeast Valleys
174
30
‘‘before’’ period no actions would take place at
any of the sites, and in the ‘‘after’’ period only a
subset of sites would have actions occurring after
an experiment was initiated and the remaining sites
would serve as untreated controls. However, the
real world is not run by researchers, and actions
have occurred continuously at many sites used in
the present analysis, from prior to the time that the
first parr-to-smolt survival estimates are possible
(early 1990s) to the end of the data set in 2002.
Therefore, a BACI design is not possible with the
data at hand.
Fourth, in the near-total absence of both designed experiments and observational studies that
24
196
35
50
29
44
24
202
39
2000
2001
2002
NA
1
1
9
3
9
NA
2
24
NA
1
1
49
62
31
4
49
64
31
4
44
3
49
64
31
4
44
3
2
3
NA
3
NA
2
NA
4
NA
3
NA
1
9
NA
24
4
3
3
9
2
24
7
3
NA
2
7
3
NA
2
7
3
NA
2
52
30
55
32
54
55
34
26
211
45
219
47
29
226
47
have examined the effects of actions on survival
in any juvenile life cycle stage (Bayley 2002), it
is impossible to say with confidence how one
should try to relate action intensity to juvenile survival. For example, is a simple linear relationship
any more or less likely than a piecewise linear
function, monotonic nonlinear relationship, or
something more complex? More generally, the
form of the relationship between parr-to-smolt survival and plausible independent variables (sitespecific land use/land cover, climate, fish size, etc.)
cannot be fully specified in advance based on past
analyses.
Finally, it is unclear how one should compare,
72
PAULSEN AND FISHER
scale, or normalize measures of action intensity.
For any given type of action, such as the screening
of irrigation diversions, a number of indicators are
potentially available, such as diversion capacity
(volume of water diverted per unit time). Should
the screening of a 1-m3/s diversion on a large
stream be considered equal to a diversion of similar capacity on a smaller tributary? To add to the
problem, a wide variety of actions are often taken
simultaneously at many sites, ranging from diversion screening to riparian restoration. How one
should scale these actions relative to one another
(e.g., 1 m3/s screened versus 1 km of stream bank
restored) is not at all clear. In the end, we chose
a very simple metric: the total number of actions
that affected each site. Obviously, this is a very
crude measure of action intensity, but as we demonstrate in this analysis, it appears to have a fairly
strong association with parr-to-smolt survival.
Here we use an information-theoretic framework (Burnham and Anderson 1998) in conjunction with a careful examination of confounding or
correlation among explanatory variables to address some of these issues. Briefly, the methods
consist of estimating a number of plausible models
and comparing the results based on adjusted (corrected) Akaike information criterion (AICc)
scores. Within this framework, the models’ information-theoretic weights may be interpreted as indicators of relative plausibility, given the data and
the suite of models estimated. This allows one to
make strong inferences about the models’ relative
importance, even in the face of the challenges described above.
Because this effort is the first of its kind, interpretation of the results should be made with caution. While the problems noted above are real ones,
they only hint at the confounding and other issues
we have encountered. This analysis makes opportunistic use of data collected for many other purposes to test hypotheses that the original research
efforts did not consider. We believe this is a useful
initial step in assessing the Chinook salmon life
cycle survival effects of habitat actions. However,
although this study may provide guidance for carefully designed experiments, it cannot substitute for
them. Neither this nor any observational study
(sensu Rosenbaum 2002; see Lawson et al. [2004]
for a similar design) can completely compensate
for the problems noted above. Our intention instead is to encourage design and funding of such
experiments, since it appears that the effects are
real and should be measurable within a reasonable
time frame—well within reach of a multi-watershed field experiment of 5–10 years’ duration.
Data
Since the data and estimates of parr-to-smolt
survival for tagged parr (the dependent variable in
the models) are an extension of those we developed previously (Paulsen and Fisher 2001), we
briefly summarize the methods used in that paper
and present updated estimates of juvenile survival.
We then describe the independent variables we
used, with particular attention to habitat actions.
Since the late 1980s, BPA and NMFS have sponsored numerous passive integrated transponder
(PIT) tagging studies on stream-type Chinook
salmon populations originating in Snake River
subbasins. In late summer and early autumn, age0 wild spring2summer Chinook salmon parr
(progeny of the previous year’s spawners) are collected, PIT-tagged, and returned to their natal
streams. These rearing areas are headwater streams
and small rivers in Idaho and Oregon upstream
from LGR on the Snake River, the upstream-most
dam through which these stocks migrate as adults
and juveniles (Figure 1). Over 600,000 wild Chinook salmon parr were PIT-tagged from 1988
through 2002. Note, however, that motivations for
the PIT tagging of juveniles vary widely, from
estimating arrival timing at LGR (Achord et al.
1997) to comparing parr-to-smolt survival rates of
hatchery and wild juveniles (Berggren et al. 2002).
To our knowledge, no parr were tagged for the
purpose of estimating the effects of specific habitat
actions.
The sites used in this analysis consisted of 33
locations above LGR (Figure 1). At least 100 age0 spring2summer Chinook salmon parr were PITtagged at each site during at least 5 of the 11 years
between 1992 and 2002, inclusive (Table 1). Data
were drawn from the PIT Tag Information System
(PTAGIS; PSMFC 2002). Over 400,000 parr were
PIT-tagged at the study sites during the 11-year
period. Many site–year combinations were missing, because few or no fish were tagged. As a
result, there were only 271 observations (out of
33 sites 3 11 years 5 363 possible observations);
an observation consisted of estimated survival to
the smolt stage for all parr tagged at a given site
in a given year and the associated independent
variables. Based on detections of surviving, PIT
tagged smolts at dams the following spring, the
parr-to-smolt survival rates for fish released at
each site and year and the associated measurement
error were estimated (see Methods). The natural
DO HABITAT ACTIONS AFFECT JUVENILE SURVIVAL?
logarithm of juvenile survival rate was the dependent variable in the regression models.
The number of fish tagged annually at each site
varied from just over 100 to almost 9,000 (mean
5 1,012; Figure 2a). Although the median natural
logarithm of parr-to-smolt survival was relatively
constant from year to year, the minimum value was
24.2 and the maximum value was 20.5, depending on the year; the lowest parr-to-smolt survival
rates overall occurred in tagging year 2002 (outmigration/detection year 2003; Figure 2b). When
expressed in untransformed units, parr-to-smolt
survival ranged from just over 1% to almost 50%.
The coefficient of variation (CV) of estimated
loge(survival) was also highly variable due to measurement error or variability, ranging between 0.02
and 0.4 (median 5 0.06; Figure 2c).
With the exception of habitat actions, the selection of independent variables was based on our
previous publications (Paulsen and Fisher 2001,
2003; Paulsen and Hinrichsen 2002). We used
three groups of independent variables. The first
group consisted of physical variables used to characterize each site, and did not change over time.
For these, we used a suite of variables developed
by the Interior Columbia Basin Ecosystem Management Project (ICBEMP; Quigley and Arbelbide
1997) that estimated basic geographic information
(e.g., elevation, stream density), average climatic
conditions, land use, and vegetation cover (Table
2). The ICBEMP estimated these variables for
small watersheds (sixth field hydrologic units
[HUCs], which average 80 km2; ICBEMP 1998).
The ICBEMP variables were used in our previous
work (Paulsen and Fisher 2001) as land use/land
cover clusters to help explain variation in parr-tosmolt survival rates. In the present study, we used
them directly as averages for the ICBEMP sixth
field HUCs that contained Chinook salmon spawning and parr tagging sites.
The second group of independent variables included biological information on the fish at each
site. This group consisted of average parr size at
tagging (total length, mm) and adult spawner redd
density in the year of tagging. As with the number
tagged and juvenile survival rates, average parr
size varied widely among sites and over time, from
about 60 mm to over 110 mm (Figure 3). Length
was recorded when the parr were tagged. Redd
survey information (Chinook salmon redds counted, stream length surveyed) was obtained from
various sources for the spawning streams (P. Keniry, Oregon Department of Fish and Wildlife
[ODFW], personal communication [Grande Ronde
73
River and Imnaha River basins]) (Hassemer 1993;
Elms-Cockrum 2001; updated by S. Keifer and E.
Brown, Idaho Department of Fish and Game
[IDFG], personal communication [Clearwater River and Salmon River basins]). Redd density was
also highly variable, from near zero to over 60
redds per mile of stream (Figure 4; 1 mi 5 1.61
km). Note that Figure 4 displays redd densities in
the year of tagging; brood-year (parent stock) redd
densities were always greater than zero.
The third group of independent variables consisted of abiotic factors that varied among sites
and over time. The first variable in this group is
the Palmer drought severity index (PDSI; NOAA
2002a). The PDSI is calculated for state climate
divisions (NOAA 2002b) and uses air temperature
and rainfall in a formula to determine relative dryness. It uses a zero as normal (mean), and drought
is indicated by negative numbers (20.5 to 21.0
5 incipient drought; 21.0 to 22.0 5 mild drought;
22.0 to 23.0 5 moderate drought; 23.0 to 24.0
5 severe drought; less than 24.0 5 extreme
drought). Positive values of the PDSI correspond
to wetter-than-average conditions. Thus, 2001 and
2002 were severe drought years on average, whereas 1995 was moderately wet on average (Figure
5). Because the PDSI was calculated based on climate divisions, it varied among groups of sites in
each climate division for any given year (Table 1),
and also varied over time (Figure 5).
The second variable in the third group is the
cumulative number of habitat actions taken at each
site. Note that the cumulative total for tagging year
1992 included all projects we could locate that
took place prior to and during 1992. The number
of actions varied among sites and over time, as
one would expect (Figure 6). Actions were represented in the regression models either as totals
or as approximate quartiles or ‘‘bins’’ (Figure 7),
where each bin was represented as a dummy variable in the model. Actions were remediation, mitigation, and other types of actions undertaken with
the intent of improving habitat for anadromous
salmonids. These actions were carried out over a
span of at least 25 years by various federal, state,
and private entities. Our sources were primarily
federal agencies and federally sponsored watershed groups (BPA 2002; GRMWP 2003; NPPC
2002, 2003; USBWP 2003), the U.S. Department
of Agriculture extension service, the U.S. Bureau
of Land Management (BLM), the U.S. Forest Service (USFS), and state governments (B. Riggers,
Oregon Watershed Enhancement Board, personal
communication).
74
PAULSEN AND FISHER
FIGURE 2.—Box plots of (a) the number of Chinook salmon parr tagged at Snake River basin sites from 1992
to 2002, (b) the natural logarithm of parr-to-smolt survival (from tagging to passage at Lower Granite Dam) for
all stocks by year of tagging, and (c) the coefficient of variation of survival. Vertical lines with horizontal dashes
indicate minima and maxima. The boxes represent the first and third quartiles (25th and 75th percentiles) of the
distributions, and the horizontal lines in the boxes indicate the medians.
75
DO HABITAT ACTIONS AFFECT JUVENILE SURVIVAL?
TABLE 2.—Minima, means, and maxima for variables used in models of Chinook salmon parr-to-smolt survival in
the Snake River basin.
Description
Minimum
Natural logarithm of Chinook salmon parr survival from tagging to the first
main-stem dam
Drainage density (km stream/km 2 )
Number of sixth field hydrologic units upstream
Total 1:100,000-scale streams upstream
Mean elevation of drainage (ft)a
Geometric mean road density (km road/km 2 )
Annual average temperature (8C)
Prism precipitation (mm)
Solar radiation (W/m 2 )
Private and Bureau of Land Management (BLM)
U.S. Forest Service (USFS) forest and range, moderate impact; grazed
Private land and USFS forest land
USFS forest, high–moderate impact; no grazing
BLM rangeland
USFS managed wilderness
Moist forest, understory reinitiation
Desert shrub
Transition forest
Young, dry forest
Young spruce-fir-lodgepole pine forest
Old spruce-fir-lodgepole pine forest
Moist forest, stem exclusion
Mean Chinook salmon total length at tagging (mm)
Redd density in the year of tagging (redds/mi)b
Palmer drought severity index
Cumulative number of habitat actions taken
a
b
24.1
0.53
0.00
14.10
264.4
0.01
1.1
304.6
255.2
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
60.7
0.0
25.9
0.0
Mean
21.6
1.27
4.10
38.44
671.5
0.86
3.4
926.2
316.0
0.032
0.237
0.121
0.173
0.072
0.210
0.117
0.028
0.274
0.063
0.381
0.005
0.081
77.2
6.2
21.0
16.9
Maximum
20.5
1.73
15.76
55.43
903.6
2.95
8.5
1,309.3
355.9
0.372
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
1.0
0.278
1.0
114.1
62.8
5.6
226
1 ft 5 30.5 cm.
1 mi 5 1.61 km.
We used our judgment to narrow the list of actions to those that would most likely affect parrto-smolt survival of spring2summer Chinook
salmon. These actions occurred near the principal
spawning and rearing areas of the stocks and generally targeted four problem areas: (1) restoring
riparian areas or controlling grazing (54% of all
actions), (2) improving instream habitat (25%), (3)
improving parr passage conditions (13%), and (4)
improving water quantity or quality (8%). Many
site2year combinations had more than one type
of action, and some sites and years had all four
types of actions (Figure 8). This poses an additional complication for researchers trying to dis-
FIGURE 3.—Box plot of the average length of Snake River basin Chinook salmon parr at tagging. Vertical lines
with horizontal dashes indicate minima and maxima. The boxes represent the first and third quartiles (25th and
75th percentiles) of the distributions, and the horizontal lines in the boxes indicate the medians.
76
PAULSEN AND FISHER
FIGURE 4.—Box plot of Snake River basin Chinook salmon redd density (redds/mi; 1 mi 5 1.61 km) for the
year of spawning (i.e., tagging year 2 1), presented by year of tagging. Vertical lines with horizontal dashes indicate
minima and maxima. The boxes represent the first and third quartiles (25th and 75th percentiles) of the distributions,
and the horizontal lines in the boxes indicate the medians.
cover the effects of actions at levels finer than the
totals or quartiles used here. It will be very difficult
to disentangle the effects of many types of actions
occurring simultaneously in the same general location. In calculating the number of actions, we
assumed that any action, once taken, would be
effective from the year in which it was implemented through the end of the survival data set.
Of course, this assumption may not be correct, as
we will discuss later.
Methods
In this section, we first outline how the parr-tosmolt survival estimates were obtained and then
discuss at length the model selection methods employed. Each site2year combination contained
data from at least 100 wild spring2summer Chinook salmon parr that were PIT-tagged in late summer to early autumn. Each PIT tag has a unique
serial number (Achord et al. 1997). Therefore, subsequent capture histories for each fish can be recorded at detectors installed in the juvenile bypass
systems at main-stem Snake River and Columbia
River hydroelectric dams. Tagging and detection
data are available for download from PTAGIS
(PSMFC 2002). The following spring after tagging
(approximately April through June), tagged smolts
are detected at LGR, Lower Monumental, and Little Goose dams on the lower Snake River, and
McNary, John Day, and Bonneville dams on the
lower Columbia River, as they migrate to the Pacific Ocean. Survival from tagging to detection at
FIGURE 5.—Palmer drought severity index (PDSI)for September2December of years in which Snake River basin
Chinook salmon parr were tagged. Vertical lines indicate ranges, and horizontal lines indicate medians.
DO HABITAT ACTIONS AFFECT JUVENILE SURVIVAL?
77
FIGURE 6.—Box plot of the total number of habitat improvement or remediation actions at sites where Snake
River basin Chinook salmon parr were tagged, by year of tagging. Vertical lines with horizontal dashes indicate
minima and maxima. The boxes represent the first and third quartiles (25th and 75th percentiles) of the distributions,
and the horizontal lines in the boxes indicate the medians.
LGR can be estimated from the numbers of fish
that were released at upstream locations and later
recaptured (detected) at the dams (Smith et al.
1994; Paulsen and Fisher 2001).
The fish from each site2year combination were
then placed into five mutually exclusive categories: (1) never seen after release, (2) seen only at
LGR and returned to the river, (3) seen only at
dams downstream from LGR, (4) seen at both LGR
and one or more downstream dams, and (5) transported at LGR to below Bonneville Dam. The
counts for fish released from each site and year
were then used to estimate the proportion of tagged
fish that survived from tagging in summer to LGR
passage the following spring. Essentially, the
method consisted of estimating the gross proportion of tagged fish that were detected at LGR, then
correcting for the fact that the bypass system diverts considerably less than all of the fish passing
the dam. The correction was made by dividing the
gross proportion of tagged fish detected by the
detection rate at LGR. Estimates of parr-to-smolt
survival rates and the proportion of survivors detected at LGR vary across sites and years. Therefore, survival rates and detection proportions were
calculated independently for each site2year combination. Details on maximum likelihood estimates
of survival rates can be found in Paulsen and Fisher (2001).
The general statistical log-linear model is
FIGURE 7.—Frequency distribution of the quartiles
(bins) used to describe the cumulative percentage of habitat improvement or remediation actions in selected
Snake River basin Chinook salmon parr-to-smolt survival models.
FIGURE 8.—Proportions of Snake River basin Chinook
salmon tagging sites with up to four types of habitat
improvement or remediation actions in 2002. See the
Methods section for a description of the four types of
actions.
78
PAULSEN AND FISHER
TABLE 3.—Pearson’s correlations between the natural logarithm of Chinook salmon parr-to-smolt survival (from the
time of tagging in the Snake River basin to passage at Lower Granite Dam [LGR]), total actions taken to improve fish
habitat at tagging sites, and continuous independent variables (PDSI 5 Palmer drought severity index). Correlations
with absolute values over 0.12 were significant (P , 0.05).
Log e
(survival),
tagging to
LGR
Total
actions
Mean length
at tagging
(mm)
Average
Sep–Dec
PDSI
Redd
density, year
of spawning
Time-varying variables
Log e (survival), tagging to LGR
Total habitat actions
Mean length at tagging (mm)
Average Sep–Dec PDSI
Redd density year of spawning (redds/mi)a
1.000
0.227
0.540
0.429
20.356
0.227
1.000
0.499
0.262
20.130
0.540
0.499
1.000
0.432
20.344
0.429
0.262
0.432
1.000
20.269
20.356
20.130
20.344
20.269
1.000
Geographic variables
Drainage density (km stream/km 2 )
Number of sixth field hydrologic units upstream
Total 1:100,000-scale streams upstream
Mean elevation of drainage (ft)b
Geometric mean road density
20.311
0.342
20.184
20.084
20.036
20.175
0.535
0.014
0.007
20.017
20.446
0.565
20.144
20.162
0.102
20.239
0.274
20.111
0.060
20.008
0.007
20.070
0.063
0.074
20.084
Climate variables (long-term averages)
Annual average temperature (8C)
Prism precipitation (mm)
Solar radiation (W/m 2 )
0.130
20.144
20.064
0.164
20.515
0.105
0.368
20.466
20.100
0.092
20.269
0.053
20.111
0.232
0.001
Variable
Land use variables
Private and Bureau of Land Management (BLM) rangeland
U.S. Forest Service (USFS) forest and range, moderate impact;
grazed
Private land and USFS forest land
USFS forest, high-moderate impact: no grazing
BLM rangeland
USFS managed wilderness
0.294
0.553
0.518
0.295
20.098
20.136
0.055
20.047
0.079
0.131
20.003
0.194
20.103
0.137
20.144
20.195
0.227
20.105
0.374
20.109
20.059
0.001
20.090
0.381
20.135
0.014
20.144
0.213
20.134
20.060
Vegetation cover variables
Moist forest, understory reinitiation
Desert shrub
Transition forest
Young, dry forest
Young spruce–fir–lodgepole pine forest
Old spruce–fir–lodgepole pine forest
Moist forest, stem exclusion
20.095
20.028
0.245
0.082
20.084
0.191
20.201
20.163
0.001
0.349
20.043
20.218
20.060
20.115
20.224
0.007
0.487
0.053
20.303
0.096
20.096
20.027
0.197
0.107
0.049
20.172
0.006
20.074
0.135
20.090
20.193
20.060
0.116
20.076
0.114
a
b
1 mi 5 1.61 km.
1 ft 5 30.5 cm.
log e (ŝ i,t ) 5 b0 1 Y t 1 gL i,t 1 lD i,t21 1 y C i,t
Orh
22
1
j51
j
i,j
1 uH i,t 1 « i,t ,
where i indexes tagging site, t denotes year of tagging, and j indexes the ICBEMP variables and parameter estimates. Loge(ŝi,t) is the natural logarithm of survival to LGR, the Yt are year-specific
classification variables common to all sites, Li,t is
the average length (mm) of each group of parr at
tagging, Di,t21 are redd densities in the previous
(brood) year, Ci,t is the climate index (PDSI), the
hi,j are the 22 ICBEMP physical variables (specific
to each site), and the Hi,t are habitat actions expressed either as the cumulative total number of
actions or as quartiles of the cumulative total. The
error terms (ei,t), a combination of process and
measurement error, are assumed to be independently and normally distributed (0, s2). The terms
b0 (intercept), Yt, g, y, rj, and u are estimated parameters. Where quartiles are used (as dummy variables) for the actions, each quartile will have its
own parameter estimate (uk) in the model. The circumflex, or hat symbol, above the survival term
is retained to emphasize the fact that survival is
estimated with measurement error. Each observation is weighted by the inverse of the CV of
loge(ŝi,t), giving more weight to those observations
for which the juvenile survival estimate has lower
measurement variability or error. We applied a variety of common diagnostic techniques to the mod-
79
DO HABITAT ACTIONS AFFECT JUVENILE SURVIVAL?
TABLE 4.—Summary of results from 18 estimated models of Snake River Chinook salmon parr-to-smolt survival.
A hyphen (-) indicates that the variable(s) were not used
in the model. An ‘‘x’’ indicates that the variable was included in the model but that the estimated parameters were
not significant. A bold ‘‘X’’ denotes a significant result
(two-tailed t-test; P , 0.05). Time-varying variables include the Palmer drought severity index (PDSI). Timeinvariant variables were estimated by the Interior Columbia Basin Ecosystem Management Project (ICBEMP).
Time-varying
variables
Timeinvariant
Model (ICBEMP)
number variables
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
a
Number of habitat
actions
Cumulative
total
Quartiles
Year
effects
PDSI,
length
at tagging,
redd
density a
X
x
X
X
x
X
-
X
x
X
x
X
X
-
X
X
X
X
X
X
-
X
X
X
X
X
X
-
X
X
X
X
X
X
X
X
X
-
Redd density is measured in redds/mi; 1 mi 5 1.61 km.
els with the highest information-theoretic weights
(see Results).
As noted in the previous section, we selected
the independent variables based on recently published analyses for the PIT-tagged Chinook salmon
stocks originating in the Snake River basin. Although stepwise regression is often used in similar
circumstances, we instead used an informationtheoretic approach (Burnham and Anderson 1998)
to address this issue. We did so because the approach gives a formal accounting for the relative
plausibility of the models estimated. Thompson
and Lee (2002) applied similar information-theoretic approaches to Snake River spring2summer
Chinook salmon spawner–recruit models, whereas
we (Paulsen and Fisher 2003) applied the approach
to models of parr-to-smolt survival.
The information-theoretic approach is described
at length by Burnham and Anderson (1998). Briefly, the method consists of the following steps: (1)
identification of a candidate set of models a priori,
based on information about scientifically plausible
relationships between candidate independent variables and the dependent variable of interest; (2)
estimation of the regression models based on the
same data set (the 271 observations described
above) and the same dependent variable (log e[ŝi,t]);
(3) calculation of the AICc for each model (the
AICc is adjusted for having a small number of
observations relative to the number of parameters);
(4) selection of the model with the lowest AICc
from among the candidate models, and subtraction
TABLE 5.—Corrected Akaike information criterion (AIC c ) scores, model (D AIC c ), AIC c weights and ranks, R 2 , and
adjusted R 2 for small sample size for each model describing Snake River basin Chinook salmon parr-to-smolt survival.
The three models with the highest weights are presented in bold italics.
Model Number of
number parameters
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
36
29
26
34
27
24
33
26
23
15
8
5
13
6
3
12
5
2
AIC c
D AIC c
w(i)
AIC c
rank
R2
Adjusted
R2
175.372
204.364
329.361
176.507
198.002
328.985
178.426
196.988
331.984
279.734
269.500
384.051
265.521
272.463
374.817
278.427
271.561
387.074
0.000
28.992
153.990
1.135
22.631
153.614
3.054
21.616
156.612
104.363
94.129
208.679
90.150
97.091
199.446
103.055
96.189
211.703
0.561
,0.0001
,0.0001
0.318
,0.0001
,0.0001
0.122
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
1
6
14
2
5
13
3
4
15
12
8
17
7
10
16
11
9
18
0.66
0.60
0.34
0.65
0.60
0.33
0.65
0.60
0.31
0.39
0.38
0.03
0.41
0.36
0.05
0.38
0.36
0.00
0.61
0.55
0.27
0.60
0.55
0.26
0.60
0.55
0.25
0.36
0.36
0.02
0.39
0.35
0.04
0.35
0.35
0.00
80
PAULSEN AND FISHER
TABLE 6.—Parameter estimates for the top-three models (see Table 5) used for estimating Snake River basin Chinook
salmon parr-to-smolt survival (models 1, 4, and 7). Significant parameters (P , 0.05) are in bold type.
Model 1 (ranked 1)
Parameter
Intercept
Drainage density (km stream/km 2 )
Number of sixth field hydrologic units upstream
Total 1:100,000-scale streams upstream
Mean elevation of drainage (ft)a
Geometric mean road density (km road/km 2 )
Annual average temperature (8C)
Prism precipitation (mm)
Solar radiation (W/m 2 )
Private and Bureau of Land Management (BLM) rangeland
U.S. Forest Service (USFS) forest and range, moderate impact; grazed
Private land and USFS forest land
USFS forest, high–moderate impact; no grazing
BLM rangeland
USFS managed wilderness
Moist forest, understory reinitiation
Desert shrub
Transition forest
Young, dry forest
Young spruce–fir-lodgepole pine forest
Old spruce–fir–lodgepole pine forest
Moist forest, stem exclusion
Total habitat actions
Quartile (0 actions)
Quartile 2 (1–2 actions)
Quartile 3 (3–23 actions)
Quartile 4 (.23 actions)
Year effects
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
a
Estimate
SE
P.t
22.99
0.08
0.02
0.00
0.00
20.08
0.07
0.00
0.00
1.25
20.09
20.18
0.03
20.23
0.16
1.09
1.15
1.16
1.31
1.06
4.60
0.63
1.14
0.15
0.01
0.00
0.00
0.06
0.06
0.00
0.00
0.69
0.11
0.20
0.12
0.32
0.13
0.26
0.31
0.27
0.36
0.23
0.68
0.33
0.009
0.577
0.083
0.180
0.568
0.193
0.224
0.788
0.505
0.071
0.424
0.349
0.819
0.468
0.209
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
0.057
20.20
20.19
20.04
0.09
0.08
0.07
0.021
0.019
0.585
0.44
0.64
0.33
0.67
0.70
1.11
0.72
0.71
0.76
0.43
0.09
0.08
0.08
0.10
0.12
0.09
0.07
0.08
0.08
0.08
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
1 ft 5 30.5 cm.
of the lowest AICc from each model, thus yielding
a delta (D) equal to zero for the model with the
lowest AICc; and (5) calculation of AICc weights
for each model by use of a simple exponential
function of the delta terms. The weights are then
normalized to sum to 1.0, and their values may be
interpreted as the relative plausibility of each model, given the data and the set of candidate models.
The models may be nonnested, as is the case here,
without influencing the comparisons.
Because the concept may be unfamiliar to some,
we provide a brief overview of the philosophy
behind this method of model selection. Consider
first how one can interpret the confidence interval
on a regression parameter, setting aside the mechanics of how confidence bounds are calculated.
Say that from a given set of data, one estimates a
bivariate regression model (e.g., y 5 a 1 bx1). The
point estimate of the intercept a is 2.0 in this example, with 5% and 95% confidence bounds of 1.8
and 2.2, respectively. What does this mean? One
interpretation is that if one had in hand a large
number of similar data sets and estimated the same
regression model for each set, the estimated intercept would be between 1.8 and 2.2 for 90% of
the models estimated.
The AICc weights can be thought of in a similar
way. They are used to rank the models that one
has estimated with a given set of data, to select
the best model (i.e., the model with the lowest
AICc) or the best set of m models from among the
n models estimated (m , n). To continue the sim-
81
DO HABITAT ACTIONS AFFECT JUVENILE SURVIVAL?
TABLE 6.—Extended.
Model 4 (ranked 2)
Parameter
Intercept
Drainage density (km stream/km2)
Number of sixth field hydrologic units upstream
Total 1:100,000-scale streams upstream
Mean elevation of drainage (ft)a
Geometric mean road density (km road/km2)
Annual average temperature (8C)
Prism precipitation (mm)
Solar radiation (W/m2)
Private and Bureau of Land Management (BLM) rangeland
U.S. Forest Service (USFS) forest and range, moderate impact; grazed
Private land and USFS forest land
USFS forest, high–moderate impact; no grazing
BLM rangeland
USFS managed wilderness
Moist forest, understory reinitiation
Desert shrub
Transition forest
Young, dry forest
Young spruce–fir–lodgepole pine forest
Old spruce–fir–lodgepole pine forest
Moist forest, stem exclusion
Total habitat actions
Quartile actions (0 actions)
Quartile 2 (1–2 actions)
Quartile 3 (3–23 actions)
Quartile 4 (.23 actions)
Year effects
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
ple example noted above, say that one had estimated a series of three models based on three different independent variables, x1, x2, and x3, one
independent variable per model. Assume further
that the models’ AICc weights were 0.6, 0.3, and
0.1, respectively. Analogously to the above example, the weights can be interpreted as meaning
that if one had many similar data sets and estimated the same regression models for each data
set, then the model based on x1 would be the topranked model 60% of the time—about twice as
often as the second-ranked model, and six times
as often as the third-ranked model. In this sense,
the model with the AICc weight of 0.6 is six times
more plausible that the model with the weight of
0.1.
Model 7 (ranked 3)
Estimate
SE
P.t
Estimate
SE
P.t
23.65
0.07
0.02
0.00
0.00
20.08
0.05
0.00
0.00
1.13
20.07
20.11
0.14
0.06
0.20
0.98
0.93
1.11
0.26
0.96
4.13
0.55
0.0019
1.12
0.14
0.01
0.00
0.00
0.06
0.06
0.00
0.00
0.74
0.11
0.18
0.11
0.31
0.13
0.25
0.29
0.26
0.35
0.23
0.65
0.33
0.00
0.001
0.614
0.125
0.435
0.944
0.157
0.379
0.277
0.999
0.128
0.495
0.550
0.213
0.858
0.125
0.001
0.001
,0.0001
0.000
,0.0001
,0.0001
0.090
0.042
24.00
0.03
0.02
0.00
0.00
20.08
0.05
0.00
0.00
1.71
20.02
20.07
0.16
0.02
0.26
0.93
0.93
1.04
1.23
0.88
4.02
0.57
1.11
0.14
0.01
0.00
0.00
0.06
0.06
0.00
0.00
0.69
0.11
0.18
0.11
0.31
0.13
0.25
0.29
0.26
0.35
0.23
0.65
0.33
0.000
0.823
0.184
0.955
0.895
0.152
0.356
0.426
0.625
0.013
0.824
0.707
0.164
0.942
0.046
0.000
0.002
,0.0001
0.000
0.000
,0.0001
0.081
0.43
0.62
0.31
0.66
0.70
1.10
0.72
0.70
0.76
0.43
0.09
0.08
0.07
0.10
0.12
0.09
0.08
0.08
0.08
0.08
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
0.40
0.58
0.29
0.65
0.69
1.10
0.71
0.70
0.77
0.44
0.09
0.08
0.07
0.10
0.12
0.09
0.08
0.08
0.09
0.08
,0.0001
,0.0001
0.000
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
,0.0001
Results
Correlations between the natural logarithm of
juvenile survival, total (cumulative) number of actions, and the continuous independent variables are
shown in Table 3. There was a positive, significant
correlation (a 5 0.05) between actions and parrto-smolt survival. However, there were also numerous significant correlations between survival,
actions, and many of the potential independent
variables. For example, both survival and total actions were positively correlated with length at tagging, proportion of private/BLM land, and transitional forest vegetation, whereas brood-year redd
density was inversely correlated with survival.
These results are symptomatic of the confounding
82
PAULSEN AND FISHER
noted in the introduction: actions may affect other
independent variables in the models, such as parr
size at tagging (R 5 0.499). Furthermore, the actions were not scattered randomly across the landscape. Instead, they were spatially concentrated on
private and BLM rangeland (R 5 0.553), private/
USFS land (R 5 0.194), and BLM rangeland (R
5 0.137), and were much less common in wilderness areas (R 5 20.144).
The 18 models estimated are shown in Table 4.
Site-specific information that does not change over
time was used in models 1–9 (the 22 ICBEMP land
use/land cover variables), whereas models 10218
did not include any location variables.
The effects of actions were treated as total actions (models 4–6 and 13–15) or as quartiles (models 1–3 and 10–12) or were excluded from the
models (models 7–9 and 16–18). Other time-varying factors were estimated based on year effects
(classification or dummy variables) common to all
sites (models 1, 4, 7, etc.) or on the PDSI, length
at tagging, and redd density (models 2, 5, 8, etc.),
or were excluded (models 3, 6, 9, etc.). Ignoring
interaction terms (e.g., PDSI 3 year effects), we
thus estimated models based on all combinations
of location, habitat, and time-varying effects.
Models ranged from extremely simple models
(e.g., model 18 had only intercept and variance
terms) to high-parameter models (e.g., model 1 had
36 parameters, including the variance term s). Table 4 also indicates whether the estimated parameters were significantly different from zero at an
a level of 0.05. The time-invariant (ICBEMP) variables (as a group) were always significant, as were
the common year effects and the PDSI and other
time-varying effects.
The patterns of significance for the total habitat
action terms were more complex and more intriguing, due in part to the correlations and confounding
already mentioned. First, they were almost never
significant in models that included the PDSI,
length at tagging, and redd density (models 2, 5,
8, etc.); the one exception was model 11, for which
no site-specific information was used. Among the
nine models that included the ICBEMP variables,
however, habitat was important in four (models 1,
3, 4, and 6), all models where habitat parameters
were estimated and the PDSI and similar variables
were excluded. For models 10–15, which did not
use site-specific information, habitat actions were
significant in four of the six models.
Of perhaps more interest than statistical significance is the fact that for nearly all models where
habitat was important, the signs on the estimated
coefficients were what proponents of habitat enhancement would hope for: actions were almost
always positively related to juvenile survival. The
one exception was model 11, which included the
PDSI, parr size at tagging, and redd density but
excluded the ICBEMP variables. In this model,
there was a negative relationship between actions
(expressed as a series of dummy variables for the
four quartiles) and survival. In the other eight
models where the estimated coefficients were significant, increased numbers of actions were associated with increased juvenile survival.
The AICc scores can be helpful in selecting the
most plausible models when confounding occurs,
subject to the caveats noted in the methods section.
Of the 18 estimated models, 15 were highly implausible and had AICc weights less than 0.0001
(Table 5). As can be seen in Table 5, three models
(1, 4, and 7; in bold type in the table) had the
overwhelming majority of the AICc weighting.
The weights (w[i] values) were 0.561 for model 1,
0.318 for model 4, and 0.122 for model 7; these
three weights accounted for over 99% of the plausibility among the models estimated. All three
models used the ICBEMP data on land use/land
cover and the common year effects. Model 1, the
top-ranked model, used habitat action quartiles,
model 4 used total habitat actions, and model 7
did not use actions (or, equivalently, assumed that
their coefficients equaled zero). Model 1 was over
four times more plausible than model 7 (0.561/
0.122), whereas model 4 was over twice as plausible (0.318/0.122). Taken together, models 1 and
4 accounted for about 88% (0.56 1 0.32) of the
plausibility among models. In other words, if one
were to perform a similar exercise many times with
similar data, models that incorporate habitat actions, land use/land cover, and common year effects would have the lowest AICc scores about 88%
of the time.
Details of the parameter estimates for the three
most plausible models are shown in Table 6. The
vegetation cover coefficients and the common year
effects were similar for all three models, and were
always significant. The habitat quartile coefficients
were significant for model 1, and the total actions
coefficient was likewise significant for model 4.
Although the AICc weights are helpful in choosing plausible models, they cannot eliminate the
confounding among land use, actions, and other
variables. This is illustrated by examining two parameters for model 7 (which did not use actions)
in comparison to the parameter estimates for models 1 and 4 (which did include actions). For model
DO HABITAT ACTIONS AFFECT JUVENILE SURVIVAL?
7, the proportions of private/BLM rangeland and
USFS managed wilderness both had large, statistically significant parameters, whereas neither variable was significant for models 1 and 4. As noted
above, the correlations between these land-use
proportions and total actions were very strong. It
appears that some of the variability in juvenile
survival that was accounted for by actions in the
two top-weighted models was instead accounted
for by land use in the third-weighted model, again
as a result of correlation and confounding among
the variables. All three models had adjusted R2
values between 0.60 and 0.61, but the AICc weight
for model 7 was only 0.122, about 20% of the
weight calculated for model 1 (Table 5).
Influence diagnostics (Belsley et al. 1980) revealed between 5 and 10 moderately influential
observations and absolute values of studentized
residuals greater than 2.1 for models 1 and 4. Elimination of these observations had no appreciable
effect on the parameter estimates; the estimated
coefficients did not change by more than one standard deviation, and statistical significance did not
change for any estimated coefficients. While there
were small departures from the assumption of normality for the residuals of both models, elimination of the suspect observations made for very
modest changes in the parameter estimates and associated standard errors, and the significance of
the parameter estimates remained unchanged. We
also dropped each year of data in sequence and
determined that there was little change in the parameter estimates. Dropping each site in sequence
also made little difference, but there was one curious exception. Elimination of the Lemhi River,
which had the largest number of actions, roughly
doubled the total actions parameter estimate for
model 4, from 0.002 to 0.004.
Discussion
An obvious question, in light of the apparent
statistical significance of habitat actions, is whether or not they are biologically important. Do habitat actions make an important difference in juvenile survival rates? The overall average survival
rate (in untransformed units) was about 20–25%.
According to model 1, stocks with zero (first quartile) or 1–3 (second quartile) actions at the tagging
site had juvenile survival rates (in loge-transformed units) of about 0.2 less than stocks with
more than 24 actions (fourth quartile; Table 6). In
untransformed units, this is e0.2 or about 22%, so
a large number of actions at the tagging site resulted in about a 1.22 multiplicative increase in
83
juvenile survival rates. Based on similar logic and
the results for model 4 (total actions), the difference in survival for a stock with 100 actions at
the tagging site versus one with no actions was
also about 20% (i.e., 100 3 0.002). This may not
seem to be a very large increase in juvenile survival, but according to the NMFS Biological Opinion (NMFS 2000), changes to the hydrosystem,
costing millions of dollars per year ($225 million
was spent on the Columbia River hydrosystem by
BPA from 1978 to 1999 [NPPC 2001]), are only
expected to increase out-migration survival of
spring2summer Chinook salmon smolts migrating
in-river (i.e., not transported in barges) by about
10%. In light of the problems highlighted above,
one should not push this result too far, but it at
least suggests that if the regression relationships
have a causal component, then substantial increases in juvenile survival rates may be feasible for
many of the stocks in this analysis.
Confounding aside, several additional caveats
should be noted. First, logistical and legal constraints may well preclude any significant habitat
manipulation in wilderness areas (e.g., much of
the Middle Fork Salmon River). Therefore, even
if actions are indeed quite effective, many sites
may never benefit from them. Second, it is possible
that sites with many actions, like the Lemhi River,
may be reaching a point of declining marginal returns, as suggested by the doubling of the model
1 coefficient when the Lemhi River was excluded
from the analysis. Finally, of course, the analysis
focused exclusively on parr-to-smolt survival, and
many types of actions are aimed at egg-to-fry, fryto-parr, or the adult migration and spawning life
stages. The possible benefits of these actions to
other life cycle stages were not detectable with our
methods.
How, then, might this analysis be improved? An
obvious starting point would be a series of on-theground inspections to test our assumption that habitat actions, once taken, remain effective indefinitely. Streams are dynamic systems, and it is highly unlikely that all actions remained effective for
11 years. A second step would be a systematic
assessment of the habitat where the actions occurred. That is, it would be useful to obtain a measure of the proportion of problematic habitat that
has been improved by past actions, and how much
poor-quality habitat remains. For example, it is at
least possible that, despite the 226 actions undertaken in the Lemhi River subbasin to date (Table
1), the subbasin still has hundreds of other problematic locations. Similarly, the fact that many
84
PAULSEN AND FISHER
sites had no actions undertaken does not necessarily mean that they have no problems that are
causing decreased juvenile survival of Chinook
salmon. Systematic habitat assessments of the sort
undertaken by the Northwest Power Planning
Council (NPPC 2003) would be useful in this regard.
An assumption implicit up to this point has been
that actions do, in fact, result in habitat improvements, as distinct from increases in parr-to-smolt
survival. Plans are underway to begin broad-scale
systematic habitat monitoring both at action sites
and at comparable, untreated control sites (BPA
2003). These studies should help resolve this issue
and may lead to more direct assessments in which
parr-to-smolt survival can be modeled as a function of habitat conditions rather than the number
of actions that have occurred. However, substantial
confounding will probably still remain. For example, sites in wilderness areas, which are generally thought to have high-quality habitat, also
tend to have few actions compared to sites with
intensive agriculture and/or grazing (and therefore
many actions).
Finally, as noted in the introduction, any analysis that examines the effects of past habitat actions
is limited by the fact that the actions are not scattered randomly across the landscape. If habitat
managers and researchers could coordinate their
efforts so that actions were sited in a stratified2random fashion with simultaneous monitoring of similar control sites, this would greatly ease
the attempt to disentangle the effects of the actions
from the effects of the many other potential covariates.
Acknowledgments
The work was supported by contracts with the
BPA. The views are those of the authors. W.
Thompson of the USFS generously shared Statistical Analysis System code for estimating the AIC c
weights. We received helpful comments from I.
Parnell, R. Hinrichsen, and R. Vadas on earlier
versions of the manuscript. We also thank the biologists who conducted the PIT tagging studies
and the tagging crews from IDFG, ODFW, NMFS,
the U.S. Fish and Wildlife Service, and the Nez
Perce and Shoshone-Bannock Tribal Fisheries departments. C. Stein of PTAGIS provided assistance
with access to the tagging data.
References
Achord, S., M. B. Eppard, B. P. Sanford, and G. M.
Matthews. 1997. Monitoring the migrations of wild
Snake River spring–summer Chinook salmon
smolts. Report by the National Marine Fisheries
Service to the Bonneville Power Administration,
Portland, Oregon.
Bayley, P. B. 2002. A review of studies on responses
of salmon and trout to habitat change, with potential
for application in the Pacific Northwest. Report to
the Washington State Independent Science Panel
Available: http://www.governor.wa.gov/gsro/science/
050802bayley.pdf. (November 2003).
Belsley, D. A., E. Kuh, and R. E. Welch. 1980. Regression diagnostics. Wiley, New York.
Berggren, T., N. Bouwes, H. Schaller, P. Wilson, C. Petrosky, E. Weber, S. Scott, and R. Boyce. 2002. Comparative survival rate study (CSS): 2002 design and
analysis. Report to the Bonneville Power Administration, Contract 00006203, Project 199602000, Portland, Oregon.
BPA (Bonneville Power Administration). 2002. Fish and
Wildlife Group technical fish and wildlife publications. BPA, Portland, Oregon. Available: http://
www.efw.bpa.gov/cgi-bin/FW/publications.cgi. (November 2003).
BPA (Bonneville Power Administration). 2003. Research,
monitoring, and evaluation (RM&E) plan for the National Marine Fisheries Service (NMFS): 2000 federal
Columbia River Power System (FCRPS) Biological
Opinion. BPA, Portland, Oregon. Available: http://
www.efw.bpa.gov/EW/FishandWildlifeDocspPost/
RME/rmepplanpdraftp03p0203.pdf. (November 2003).
Burnham, K. P., and D. R. Anderson. 1998. Model selection and inference: a practical information-theoretic approach. Springer-Verlag, New York.
Elms-Cockrum, T. J. 2001. 1999 salmon spawning
ground surveys. Idaho Department of Fish and
Game, Boise.
GRMWP (Grande Ronde Model Watershed Program). 2003.
Grande Ronde basin restoration project inventory.
Available: http://www.fs.fed.us/pnw/modelwatershed/
projectpinventory.htm. (November 2003).
Hassemer, P. 1993. Salmon spawning ground surveys,
1989–92. Idaho Department of Fish and Game, Boise.
ICBEMP (Interior Columbia Basin Ecosystem Management Project). 1998. SIMMS metadata report:subbasin/Watershed/Subwatershed–vector. Available:
http://www.icbemp.gov/spatial/metadata/hydro/437.
htm. (April 2004).
Lawson, P. W., E. A. Logerwell, N. J. Mantua, R. C.
Francis, and V. N. Agostini. 2004. Environmental
factors influencing freshwater survival and smolt
production in Pacific Northwest coho salmon (Oncorhynchus kisutch). Canadian Journal of Fisheries
and Aquatic Sciences 61:360–373.
NMFS (National Marine Fisheries Service). 2000. 2000
federal Columbia River Power System Biological
Opinion. NMFS, Seattle, Washington. Available:
http://www.nwr.noaa.gov/1hydrop/hydroweb/docs/
Final/2000Biop.html. (November 2003).
NOAA (National Oceanic and Atmospheric Administration). 2002a. Time bias corrected divisional temperature–precipitation–drought index. Available: http://
DO HABITAT ACTIONS AFFECT JUVENILE SURVIVAL?
www.ncdc.noaa.gov/oa/climate/onlineprod/drought/
ftppage.html (see also http://lwf.ncdc.noaa.gov/oa/
climate/onlineprod/drought/main.html). (November
2003).
NOAA (National Oceanic and Atmospheric Administration). 2002b. State maps with counties and climate
divisions. Available: http://www.cpc.ncep.noaa.gov/
products/analysispmonitoring/regionalpmonitoring/
CLIMpDIVS/statespcountiespclimate-divisions.html.
(November 2003).
NPPC (Northwest Power Planning Council). 2001. Inaugural annual report of the Columbia Basin Fish and
Wildlife Program, 1978–1999. NPPC, Council Document 2001-2, Portland, Oregon. Available: http://
www.nwcouncil.org/library/2001/2001–2/Default.
htm. (April 2004).
NPPC (Northwest Power Planning Council). 2002. Columbia River basin fish and wildlife program. Available:
http://www.nwcouncil.org/fw/program/Default.htm.
(November 2003).
NPPC (Northwest Power Planning Council). 2003. Subbasin planning. Available: http://www.nwcouncil.
org/fw/subbasinplanning/Default.htm. (November
2003).
Osenberg, C. W., and R. J. Schmitt. 1996. Detecting
ecological impacts caused by human activities. Pages 3–16 in C. W. Osenberg and R. J. Schmitt, editors.
Detecting ecological impacts: concepts and applications in coastal habitats. Academic Press, San Diego, California.
PSMFC (Pacific States Marine Fisheries Commission).
2002. Columbia basin PIT tag information system
(PTAGIS). Available: http://www.psmfc.org/pittag.
(November 2003).
Paulsen, C. M., and T. R. Fisher. 2001. Statistical relationship between parr-to-smolt survival of Snake
River spring–summer Chinook salmon and indices
of land use. Transactions of the American Fisheries
Society 130:347–358.
Paulsen, C. M., and T. R. Fisher. 2003. Detecting juvenile survival effects of actions taken to improve
fish habitat: power analysis applied to endangered
Snake River spring–summer Chinook (Oncorhyn-
85
chus tshawytscha). Canadian Journal of Fisheries
and Aquatic Sciences 60:1122–1132.
Paulsen, C. M., and R. A. Hinrichsen. 2002. Experimental management for Snake River spring–summer Chinook (Oncorhynchus tshawytscha): tradeoffs between conservation and learning for a threatened species. Canadian Journal of Fisheries and
Aquatic Sciences 59:717–725.
Quigley, T. M., and S. J. Arbelbide, editors. 1997. An
assessment of ecosystem components in the interior
Columbia basin and portions of the Klamath and
Great basins, volume 3, chapter 4. U.S. Forest Service and Bureau of Land Management, General
Technical Report PNW-GTR-405, Portland,
Oregon.
Rosenbaum, P. R. 2002. Observational studies. Springer-Verlag, New York.
Smith, S. G., J. R. Skalski, J. W. Schlechte, A. Hoffmann, and V. Cassen. 1994. Statistical survival
analysis of fish and wildlife tagging studies:
SURPH. 1. Center for Quantitative Science, School
of Fisheries, University of Washington, for the Bonneville Power Administration, Portland, Oregon.
Solazzi, M. F., T. E. Nickelson, S. L. Johnson, and J. D.
Rogers. 2000. Effects of increasing winter rearing
habitat on abundance of salmonids in two coastal
Oregon streams. Canadian Journal of Fisheries and
Aquatic Sciences 57:906–914.
Thompson, W. L., and D. C. Lee. 2002. A two-stage
information-theoretic approach to modeling landscape-level attributes and maximum recruitment of
Chinook salmon in the Columbia River basin. Natural Resource Modeling 15:227–257.
USBWP (Upper Salmon Basin Watershed Project). 2003.
Upper Salmon basin watershed projects. Available:
http://www.modelwatershed.org/Projects.html. (November 2003).
Ward, B. R., D. J. F. McCubbing, and P. A. Slaney. 2003.
Evaluation of the addition of inorganic nutrients and
stream habitat structures in the Keogh River watershed for steelhead trout and coho salmon. Pages
127–148 in J. Stockner, editor. Nutrients in salmonid
ecosystems: sustaining production and biodiversity.
American Fisheries Society, Symposium 34, Bethesda, Maryland.