642
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
Host–parasite distributions under changing
climate: Tsuga heterophylla and Arceuthobium
tsugense in Alaska
Tara M. Barrett, Greg Latta, Paul E. Hennon, Bianca N.I. Eskelson, and
Hailemariam Temesgen
Abstract: Dwarf mistletoes (Arceuthobium species) influence many processes within forested ecosystems, but few studies
have examined their distribution in relation to climate. An analysis of 1549 forested plots within a 14.5 million ha region of
southeast Alaska provided strong indications that climate currently limits hemlock dwarf mistletoe (Arceuthobium tsugense
(Rosendahl) G.N. Jones) to a subset of the range of its primary tree host, western hemlock (Tsuga heterophylla (Raf.)
Sarg.), with infection varying from a high of 20% of trees at sea level to only 5% by 200 m elevation. Three types of modeling approaches (logistic, most similar neighbors, and random forests) were tested for the ability to simultaneously predict
abundance and distribution of host and pathogen as a function of climate variables. Current distribution was explained well
by logistic models using growing degree-days, indirect and direct solar radiation, rainfall, snowfall, slope, and minimum
temperatures, although accuracy for predicting A. tsugense presence at a particular location was only 38%. For future climate
scenarios (A1B, A2, and B1), projected increases for A. tsugense habitat over a century ranged from a low of 374% to a
high of 757%, with differences between modeling approaches contributing more to uncertainty than differences between climate scenarios.
Résumé : Le faux-gui (Arceuthobium species) influence plusieurs processus dans les écosystèmes forestiers mais peu d’études ont porté sur sa répartition en lien avec le climat. Une analyse de 1 549 places échantillons boisées à l’intérieur d’une
région de 14,5 millions d’hectares dans le sud-est de l’Alaska a clairement montré que le climat limite présentement le fauxgui de la pruche (Arceuthobium tsugense (Rosendahl) G.N. Jones) à un sous-ensemble de l’aire de répartition de son hôte
principal, pruche de l’Ouest (Tsuga heterophylla (Raf.) Sarg.), avec un taux d’infection qui varie d’un maximum de 20 %
des arbres au niveau de la mer à seulement 5 % vers 200 m d’altitude. Trois types de méthodes de modélisation (logistique,
voisins les plus similaires et forêts aléatoires) ont été testées pour leur capacité à prédire simultanément l’abondance et la répartition de l’hôte et du pathogène en fonction des variables climatiques. La répartition actuelle est bien expliquée par des
modèles logistiques utilisant les degrés-jours de croissance, le rayonnement solaire direct et indirect, la pluie, les chutes de
neige, la pente et les températures minimum, mais la précision pour prédire la présence d’A. tsugense à un endroit donné
est de seulement 38 %. Dans le cas des scénarios climatiques futurs (A1B, A2 et B1), l’élargissement prévu de l’habitat d’A.
tsugense sur une période de cent ans varie d’un minimum de 374 % à un maximum de 757 %, les différences entre les méthodes de modélisation contribuant davantage à l’incertitude que les différences entre les scénarios climatiques.
[Traduit par la Rédaction]
Introduction
Dwarf mistletoes (Arceuthobium species) are important
components within many forest ecosystems of North America. Obligate parasites of conifers, dwarf mistletoes serve essential functions for a number of species, ranging from
providing food for the rare Johnson’s Hairstreak Butterfly
(Shields 1965) to hosting hyperparasitic fungi (Wicker and
Shaw 1968). Arceuthobium species also contribute to food
and habitat for birds and mammals (e.g., Hamer et al. 2008),
impact processes ranging from fire disturbance to water use
and carbon sequestration (Meinzer et al. 2004), and perhaps
even serve as keystone species (Watson 2001). Dwarf mistletoes are also viewed as one of the most important disease
agents affecting forests, with economic impacts from growth
loss thought to be in the order of billions of dollars annually
(Hawksworth and Wiens 1996). Infection by Arceuthobium
species, like other forest diseases, is expected to be impacted
by changing climate (Kliejunas et al. 2009). In a comprehensive review of Arceuthobium, Hawksworth and Wiens (1996)
noted that while climate factors were probably responsible for
differences in distribution between hosts and parasites, effects
of climate on distribution of Arceuthobium are little studied.
Received 27 September 2011. Accepted 20 January 2012. Published at www.nrcresearchpress.com/cjfr on 14 March 2012.
T.M. Barrett. USDA Forest Service, Pacific Northwest Research Station, 3301 C Street Suite 200, Anchorage, AK 99503, USA.
G. Latta, B.N.I. Eskelson, and H. Temesgen. Department of Forest Engineering, Resources and Management, 280 Peavy Hall, Oregon
State University, Corvallis, OR 97331, USA.
P.E. Hennon. USDA Forest Service, Pacific Northwest Research Station, 11305 Glacier Highway, Juneau, AK, 99801, USA.
Corresponding author: Tara M. Barrett (e-mail: tbarrett@fs.fed.us).
Can. J. For. Res. 42: 642–656 (2012)
doi:10.1139/X2012-016
Published by NRC Research Press
Barrett et al.
643
Fig. 1. Distribution of western hemlock (Tsuga heterophylla) in
North America.
150°0'0"W
130°0'0"W
110°0'0"W 100°0'0"W
50°0'0"N
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
60°0'0"N
40°0'0"N
50°0'0"N
40°0'0"N
160°0'0"W
150°0'0"W
140°0'0"W
130°0'0"W
120°0'0"W
Climate envelope modeling — predicting the distribution
of a species as a function of climate — is a subset of species
distribution modeling (Wiens et al. 2009). Species distribution models are founded on the niche theory concept that environmental variables can be used to describe a
multidimensional space outside of which an individual species cannot survive (Vandermeer 1972). As obligate parasites,
Arceuthobium species pose an interesting instance for climate
envelope modeling: while environmental factors can directly
influence their distribution, environment also indirectly limits
dwarf mistletoe distribution and abundance through effects
on host species. In this paper, we examine the relationships
between climate and distribution and abundance of hemlock
dwarf mistletoe (Arceuthobium tsugense) and its primary
host species, western hemlock (Tsuga heterophylla (Raf.)
Sarg.).
Understanding of relationships between individual species
of Arceuthobium and climate has primarily been a function
of observational studies; few systematic surveys have been
conducted over the large spatial regions required to understand ecoregional-level distribution (Hawksworth and Wiens
1996). Coastal Alaska is potentially an ideal location to conduct a study on the possible climate controls for a tree species and an obligate parasitic pathogen because landscape
effects by other disturbances such as fire and timber harvesting are relatively uncommon.
Arceuthobium tsugense plants are dioecious and their life
cycle includes an explosive discharge of viscous seeds in an
approximately 3 week period from early September to late
October. Seeds that are intercepted by the host tree’s needles
and successfully overwinter then germinate in the spring.
After seeds infect host trees, an incubation period of 1 or 2
years precedes shoot production. Dwarf mistletoe flowers
may develop 1–2 years after the first shoots appear and flowering occurs from June to August. Fruit maturation requires
about 15 months. (Muir and Hennon 2007)
Cold temperatures can impact Arceuthobium species’ life
cycles at several different stages. Hawksworth and Wiens
(1996) described an unpublished study where seedlings infected by A. americanum were transplanted to a site 120 m
above the natural elevation limit; the dwarf mistletoe plants
survived for over 20 years but fruits never matured before au-
tumn frosts. Arceuthobium species can also be affected by
cold temperatures in winter that reduce seed germinability
(Brandt et al. 2004) and freezing night temperatures in early
summer that reduce pollen germinability (Gilbert and Punter
1991).
Snow can affect Arceuthobium reproduction because seeds
must survive in tree canopies for one winter before germinating in spring. Muir and Hennon (2007) speculated that very
high snow levels in northern forests could be a factor responsible for the low incidence or absence of A. tsugense in some
areas. Heavy rainfall can also wash away seeds before they
germinate (Roth 1959).
The length and warmth of the growing season are factors
in Arceuthobium population dynamics. The maturation time
of A. tsugense is believed to be longer in Alaska than in
other parts of its range. In British Columbia, seed dispersal
becomes common 5 years after inoculation (Smith 1971),
but in Alaska, inoculation to seed dispersal may take longer
than 12 years (Shaw and Loopstra 1991). Slower development of A. tsugense in young stands in Alaska has led to
conclusions that A. tsugense is less of a problem for timber
production in Alaska than in British Columbia, Oregon, or
Washington (Shaw and Hennon 1991). In summary, previous
research suggests that climate could limit A. tsugense at a
number of different stages in its life cycle, influencing its
rate of spread, its effects on individual trees, and its overall
range.
Climate envelope analysis of species has been the focus of
much previous research (Wiens et al. 2009) including work
on tree species (e.g., Iverson and Prasad 1998), although we
are not aware of any other climate envelope research for Arceuthobium. Methods used to predict range as a function of
climate vary widely and include logistic regression, general
additive models, neural networks, simulation modeling, and
a variety of customized models. Because different modeling
techniques can result in substantially different range predictions and no single modeling method is ideal for all situations, using several different methods (and validating results
against independent data sets) helps to quantify and understand the component of uncertainty associated with model selection.
For many applications, nonparametric methods such as
nearest neighbor imputation provide good alternatives to
parametric methods such as logistic regression. Nearest
neighbor imputation approaches are donor-based methods
where missing values are imputed using values observed for
a different unit (for details, see Eskelson et al. 2009a). The
donors are typically determined based on a similarity metric.
The Most Similar Neighbor (MSN) method (Moeur and
Stage 1995), which has found widespread use in forestry applications, uses a similarity metric that is based on a canonical correlation of the response and explanatory variables.
The Random Forest (RF) imputation method (Crookston and
Finley 2008) determines similarity with a proximity matrix
that is obtained from multiple classification and regression
trees (see Breiman 2001). RF has become a popular nearest
neighbor imputation method in recent years and has been
shown to outperform more traditional nearest neighbor imputation methods such as MSN in a variety of applications
(e.g., Eskelson et al. 2009b). One major advantage of nearest neighbor imputation methods is that their multivariate
Published by NRC Research Press
644
Can. J. For. Res. Vol. 42, 2012
Fig. 2. Location of weather stations in relation to the study area.
170°0'0"E 180°0'0"
160°0'0"W
140°0'0"W
120°0'0"W 110°0'0"W
100°0'0"W
70°0'0"N
Weather Stations
Study Area: Coastal rainforest ecoregion
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
North America
70°0'0"N
60°0'0"N
Gulkana
Talkeetna
Anchorage
60°0'0"N
Yakutat
Homer
Kodiak
0
200
Juneau
Annette
400
800 km
50°0'0"N
50°0'0"N
160°0'0"W
150°0'0"W
nature allows simultaneous prediction of multiple response
variables (Eskelson et al. 2009a). Applied to prediction of
a host–pathogen climate envelope, this has the advantage of
never predicting combinations of response variables that do
not exist in nature; thus, nearest neighbor imputation will
not predict pathogen presence without a host.
While previous research on Arceuthobium species suggests
a variety of possible mechanisms for climate effects on dwarf
mistletoe range and abundance, there has been no previous
study of A. tsugense in relation to climate at the northern extent of its range. In this paper, our objectives are to (i) estimate the current range and abundance of A. tsugense and its
host species T. heterophylla in Alaska in relation to climate,
(ii) develop, compare, and validate parametric and nonparametric models to predict A. tsugense and T. heterophylla distributions in the Gulf of Alaska region as a function of
climate variables, and (iii) predict potential future distributions of A. tsugense and T. heterophylla for several possible
future climate regimes.
Methods
Data for T. heterophylla and A. tsugense
We used two data sets for initial examination of the current
range and abundance of A. tsugense and T. heterophylla. The
first data set was from a forest inventory throughout the
coastal region of Alaska. The second data set came from
Washington, Oregon, and California. Thus, together, the two
140°0'0"W
130°0'0"W
data sets span the northern and southern parts of the range of
the T. heterophylla species but do not include portions of the
range in the Rocky Mountains where A. tsugense is not found
or coastal British Columbia, which uses a different inventory
system (Fig. 1). Only the Alaska data set was used for climate envelope modeling because this paper focuses on those
climate influences (e.g., minimum temperatures, length of
growing season) that may limit A. tsugense at the northern
portion of its range rather than climatic factors (e.g., maximum temperatures, moisture deficits) that likely limit it at
the southern portion of its range.
The Alaska data set consisted of 4921 plots established between 1995 and 2003 by the Forest Inventory and Analysis
program of the US Forest Service. All public and private
land within the study area boundaries was included in the
sampled population with the exception of national forest wilderness and Glacier Bay National Park. Of the sample, 1549
plots were forested with T. heterophylla trees present on 984
plots and A. tsugense present on 149 plots. Trees on the sample plots were selected for measurement using four 7.3 m radius subplots, with trees less than 12.5 cm diameter at breast
height selected using four 2 m fixed radius areas centered on
each of the four subplots. Infection by A. tsugense was recorded for T. heterophylla trees using the Hawksworth
(1977) six-class rating system, but we collapsed this to presence–absence for this study, as accurately rating with the
Hawksworth system can be difficult from the ground (Shaw
et al. 2000). Because A. tsugense infection was just one of
Published by NRC Research Press
Barrett et al.
645
Fig. 3. Mean annual temperature, mean minimum winter (December, January, and February) temperature, and mean annual precipitation from
1954 to 2007 for the eight primary weather stations near the study area.
10
600
(15)
300
(20)
200
Temperatur e (°C)
1990
600
(15)
300
(20)
200
100
(30)
2010
Homer
600
500
(5)
400
(10)
(15)
300
(20)
200
(35)
1950
‐
1970
1990
100
(30)
2010
10
‐
1970
1990
2010
10
‐
‐
300
(15)
(20)
200
Te mperature (°C)
400
(10)
Precipitation (cm)
P
500
(5)
600
5
(25)
500
(5)
400
(10)
(15)
300
(20)
200
Precipitation (cm)
P
600
5
emperature (°C)
Te
1990
(25)
(25)
(25)
100
(30)
Kodiak
‐
1970
1990
(35)
1950
2010
100
Annette
‐
1970
1990
2010
10
10
(5)
5
500
‐
400
(10)
300
(15)
(20)
200
(25)
100
Juneau
‐
1970
1990
Mean Annual Temperature
2010
Temperature (°C)
‐
600
Precipitation (cm)
5
(35)
1950
‐
1970
‐
Temperrature (°C)
Temperrature (°C)
400
100
5
500
(10)
(30)
200
10
Anchorage
‐
(35)
1950
(20)
(35)
1950
2010
(5)
(30)
300
(30)
‐
1970
5
(35)
1950
(15)
Precip
pitation (cm)
100
Precip
pitation (cm)
10
400
(10)
( )
(25)
(30)
(35)
1950
500
(5)
500
(5)
400
(10)
(15)
300
(20)
200
(25)
(30)
(35)
1950
Mean Winter Minimum Temperature
many variables collected in the inventory, it seems likely that
infection would have been undercounted, although brooming
and deformation from other causes could also have led to error commissions.
600
Precipitation (cm)
Temperatur e (°C)
400
((10)
0)
600
‐
500
(25)
Temperature (°C)
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
‐
(5)
Talkeetna
5
Precipitati on (cm)
Gulkana
5
Precipitation (cm)
10
100
Yakutat
1970
1990
‐
2010
Mean Annual Precipitation
Arceuthobium tsugense can infect other conifers in the region such as lodgepole pine (Pinus contorta Douglas ex
Loudon), Sitka spruce (Picea sitchensis (Bong.) Carrière),
and mountain hemlock (Tsuga mertensiana (Bong.) Carrière)
Published by NRC Research Press
646
Can. J. For. Res. Vol. 42, 2012
Fig. 4. Estimated abundance of western hemlock (Tsuga heterophylla) and hemlock dwarf mistletoe (Arceuthobium tsugense) and infection
rate by elevation in southeast Alaska (longitude ≤ –138°).
80
25
70
50
15
40
10
30
Dwarf mistletoe infection (%)
60
Trees (millions)
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
20
20
5
10
0
‐
0‐30
61‐91
122‐152
183‐213
244‐274
305‐335
366‐396
427‐457
488‐518
549‐579
610‐640
671‐701
732‐762
Elevation (m)
Western hemlock trees with dwarf mistletoe
Western hemlock trees
(Muir and Hennon 2007), but this is uncommon and over
96% of A. tsugense observations in our data set occurred in
T. heterophylla. For our analysis, only observations of A. tsugense on T. heterophylla were used. Plots where A. tsugense
infected T. heterophylla trees ranged from the Canadian border in southeast Alaska to –137° longitude, while plots with
T. heterophylla trees extended more than 600 km further west
to –149° longitude.
For initial analysis of distributions in the larger west coast
region, the second data set used consisted of a sample collected by the Forest Inventory and Analysis program between
2001 and 2007 throughout California, Oregon, and Washington. These data included 1503 forested plots with T. heterophylla trees, with A. tsugense observed on 212 plots. Plots
with T. heterophylla ranged in elevation from sea level to
1860 m and from the Canadian border south to 39° latitude
in California. Estimates of area with A. tsugense present defined presence as at least one occurrence of A. tsugense on a
T. heterophylla tree within the plot. Estimates of abundance
of A. tsugense and distribution by elevation and climate in
Alaska, Washington, Oregon, and California were made using combined ratio estimation with post-stratification (Scott
et al. 2005, pp. 55–56).
Climate of the study area
The study area for the climate envelope analysis was the
Percentage of dwarf mistletoe infection
14.5 million ha area encompassing the temperate rainforest
ecoregion in Alaska. Weather data from all eight primary stations near our study area (Fig. 2) show a trend of increasing
average annual temperature from 1954 to 2008 (Fig. 3). Assuming a linear model, mean annual temperature in the region increased at a rate of 1.5 °C per 50 years (annual
increase estimated as 0.029° with a standard error of 0.009°;
for a null hypothesis that slope = 0, the p value was 0.0015).
Assuming a linear model, mean minimum winter (December,
January, and February) temperatures increased at a rate of
3.0 °C per 50 years (annual increase estimated as 0.061°
with a standard error of 0.022°; for a null hypothesis that
slope = 0, the p value was 0.007). This 50 year historical
trend of increasing temperatures in the Gulf of Alaska region
is consistent with most General Circulation Model (GCM)
projections. For example, the Hadley model (Pope et al.
2000) predicts 50 year increases of 1.4, 1.9, and 2.2 °C at
Juneau and increases of 1.6, 2.5, and 2.9 °C at Homer for
the B1, A2, and A1B scenarios, respectively, using a linear
model from 2000 to 2100. Local climate variations, however,
may not be captured by GCMs. The Pacific Decadal Oscillation, a shift in North Pacific sea temperature and currents that
had a cool phase from 1947 to 1976 and a warm phase after
1977 (Mantua and Hare 2002), is a primary contributing factor to the climate in the study area.
Most GCMs predict increased precipitation for the region.
Published by NRC Research Press
Variable
Growing season (days >0 °C)
Growing days >10 °C
Growing degree-days (°C)
Summer evapotranspiration index
Summer climate moisture index
Growing season radians
Lowest average minimum monthly
temperature (°C)
Standard deviation of average minimum
monthly temperature (°C)
Precipitation as snow (mm)
Precipitation as rain (mm)
Lowest average minimum monthly
temperature in August–September (°C)
Lowest average minimum monthly
temperature in May–June (°C)
Slope (%)
Elevation (m)
Minimum
144
0
773
1.8
–8.7
10.5
–26.1
T. heterophylla present
Maximum
365
139
2854
4.4
28.2
25.8
0.1
Mean
282
100.4
2068
3.2
3
17.5
–6.6
Minimum
190
3
1214
2.1
–8.7
10.5
–16.8
Maximum
365
139
2854
4.4
20.3
25.8
–0.8
A. tsugense present
Mean
313
108
2248
3.3
4.4
17.8
–4.7
Minimum
249
71
1730
2.3
–4.5
10.5
–8.3
Maximum
365
138
2854
4.4
10.6
25.8
–1.7
Mean
335
119
2446
3.2
3.2
17.1
–3.6
4.1
12.3
5.8
4.1
8.6
5.1
4.3
6.1
4.9
174
156
–1.6
5041
1829
9.4
1663
733
6.2
598
227
2.1
5041
1551
9.3
2009
832
7.1
1100
436
5.7
3089
1069
9.2
1849
732
7.7
3.3
–0.2
6.1
4.1
2.7
–3.5
0
0
7
160
884
29
208
0
0
160
823
34
199
0
0
6
160
396
4.7
32
101
Table 2. Pearson correlation coefficients for modeled climate variables in southeast and southcentral Alaska.
SLOPE
GDD
ET
CMI
RADIANS
MINTEMP
MINTEMPSD
RAIN
SNOW
MINFALLTEMP
MINSPRINGTEMP
SLOPE
1.00
0.06
0.25
0.22
0.26
0.15
–0.22
0.32
0.18
0.10
0.04
GDD
0.06
1.00
0.30
0.15
0.12
0.80
–0.70
0.56
–0.53
0.90
0.93
ET
0.25
0.30
1.00
–0.11
0.97
0.25
–0.26
0.26
–0.07
0.25
0.24
CMI
0.22
0.15
–0.11
1.00
–0.07
0.40
–0.45
0.80
0.47
0.33
0.23
RADIANS
0.26
0.12
0.97
–0.07
1.00
0.18
–0.23
0.24
0.06
0.12
0.08
MINTEMP
0.15
0.80
0.25
0.40
0.18
1.00
–0.96
0.72
–0.34
0.91
0.82
MINTEMPSD
–0.22
–0.70
–0.26
–0.45
–0.23
–0.96
1.00
–0.77
0.20
–0.83
–0.70
RAIN
0.32
0.56
0.26
0.80
0.24
0.72
–0.77
1.00
0.16
0.66
0.54
SNOW
0.18
–0.53
–0.07
0.47
0.06
–0.34
0.20
0.16
1.00
–0.41
–0.48
MINFALLTEMP
0.10
0.90
0.25
0.33
0.12
0.91
–0.83
0.66
–0.41
1.00
0.93
MINSPRINGTEMP
0.04
0.93
0.24
0.23
0.08
0.82
–0.70
0.54
–0.48
0.93
1.00
647
Published by NRC Research Press
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
All forested plots
Barrett et al.
Table 1. Range of modeled climate variables in southeast and southcentral Alaska for forested plots, forested plots with western hemlock (Tsuga heterophylla), and
forested plots with hemlock dwarf mistletoe (Arceuthobium tsugense).
648
Can. J. For. Res. Vol. 42, 2012
Table 3. Coefficient estimates for fitting models with a single predictor
variable and a forward-selection logistic model for prediction of hemlock
dwarf mistletoe (Arceuthobium tsugense) and western hemlock (Tsuga heterophylla) occurrence (n = 1033).
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
Univariate models
Variable
Arceuthobium tsugense
b0
SLOPE
GDD
RADIANS
MINTEMP
MINTEMPSD
RAIN
RAIN2
SNOW
MINFALLTEMP
MINSPRINGTEMP
Tsuga heterophylla
b0
SLOPE
GDD
RADIANS
CMI
ET
MINTEMP
MINTEMPSD
RAIN
SNOW
MINFALLTEMP
MINSPRINGTEMP
Estimate
SE
p
0.0029
0.0039
–0.0738
0.0412
–0.1097
0.0004
4.6×10–8
–0.0074
0.0959
0.1370
0.0035
0.0004
0.0415
0.0059
0.0190
0.0001
2.4×10–7
0.0012
0.0114
0.0155
0.4088
0.0001
0.0754
0.0001
0.0001
0.0006
0.0534
0.0001
0.0001
0.0001
672
674
545
671
584
609
672
671
605
556
528
0.0228
0.0054
0.1532
0.1999
1.4332
0.0422
–0.1378
0.0017
–0.0008
0.1072
0.1311
0.0029
0.0003
0.0273
0.0169
0.1854
0.0027
0.0090
0.0001
0.0002
0.0066
0.0078
0.0001
0.0001
0.0001
0.0001
0.0001
0.0001
0.0001
0.0001
0.0013
0.0001
0.0001
1336
1262
858
1305
1156
1271
914
887
902
1328
855
848
For example, the Hadley model predicts 50 year increases in
average annual precipitation of 121, 135, and 168 mm at Juneau and increases in average annual precipitation of 42, 140,
and 204 mm at Homer for the B1, A2, and A1B scenarios,
respectively. In contrast with these predictions, historical precipitation patterns are more complex, with no single trend
shown for all climate stations in the region (Fig. 3). In addition to the climate station information at Anchorage and
Homer that shows relatively constant annual precipitation
over the past 50 years (Fig. 3), recent research by Berg et al.
(2009) documenting drying lakes and wetlands also indicates
a possible ongoing change in the precipitation to evaporation
balance along the western edge of our study area.
Climate variables
Potential independent variables for climate were chosen
based on the review of Arceuthobium research and possible
climate mechanisms that limit A. tsugense abundance and
distribution in Alaska. The possible mechanisms considered
were as follows: (i) the short growing season and cool temperatures during the growing season do not allow A. tsugense
fruits to fully mature, (ii) low minimum winter temperatures
reduce overwintering seed viability, (iii) precipitation (either
rain or snow) reduces the number of seeds that can establish
by causing them to be sloughed or flushed from foliage before germination, (iv) early autumn freezes damage fruit, and
(v) low spring–summer temperatures damage pollen.
These mechanisms are not mutually exclusive, as climate
AIC
impacts on A. tsugense abundance and distribution could occur through multiple causes. The variables that were chosen
to correspond to these proposed mechanisms were (i) “growing season variables”: growing degree-days (GDD) and direct
and indirect solar radiation during the growing season (RADIANS), (ii) “low winter temperatures”: coldest mean minimum temperature over the year (MINTEMP) and variation
in mean minimum temperature over the year (MINTEMPSD), (iii) “precipitation”: precipitation as snow
(SNOW) and precipitation as rain (RAIN), (iv) “autumn
freezes”: lowest mean minimum temperature in August or
September (MINFALLTEMP), and (v) “low temperature during pollination”: lowest mean minimum temperature in May
or June (MINSPRINGTEMP).
Methods for deriving these climate variables are described
below. Slope (SLOPE) was also added as an independent variable because Arceuthobium’s explosive seed dispersal results
in differential ability to spread upward compared with downward (Bloomberg et al. 1980). For prediction of T. heterophylla distribution, two additional climate variables were
selected to help explain the western boundary for T. heterophylla on the Kenai Peninsula. These variables were an evapotranspiration index for the growing season (ET) and a
climate moisture index (CMI) representing precipitation in
excess of evapotranspiration over the growing season.
Spatial precipitation and temperature information came
from the Parameter-elevation Regressions on Independent
Slopes Model (PRISM Climate Group 2002), representing
Published by NRC Research Press
0.004
0.002
0.011
<0.001
0.028
0.173
<0.001
0.016
0.5×10–7
–1.2×10–7
0.010
0.033
1.7×10–3
–3.0×10–3
0.046
0.023
0.061
–0.096
–0.067
–0.176
Value
–1.396
0.012
0.024
0.005
0.007
0.042
0.004
0.003
p
Value
–6.492
0.013
0.003
–0.095
–0.077
–0.198
0.120
0.083
0.130
–0.076
–0.205
–6.3×10–4
0.037
0.003
0.004
Value
–6.943
0.013
0.003
p
0.551
0.006
0.033
0.027
0.066
2.5×10–4
p
6
131.97
515.09
2.83
0.180
0.83
SE
2.811
0.004
0.001
6
113.83
517.92
5.65
0.044
0.85
SE
2.341
0.005
Model C
7
146.20
512.26
0.00
0.742
0.85
SE
2.879
0.005
0.001
0.047
0.027
0.066
Model B
p
<0.001
0.005
0.027
0.062
0.003
0.030
where PPASm is the percentage of precipitation that falls as
snow in month m, T is mean monthly temperature (degrees
Celsius) at each location, and a and b are parameters that
vary by month. The form of the equation is logistic, and
PPASm is bound between the values of 0 to 1. Parameters a
and b were calculated as a = –3.39 and b = –2.53 for January, a = –1.80 and b = –2.10 for February, a = –0.33 and
b = –1.37 for March, a = 0.90 and b = –1.29 for April, a =
0.27 and b = –1.21 for October, a = –2.18 and b = –1.88 for
November, and a = –3.22 and b = –2.36 for December, developed from fitting eq. 2 to 1954–2008 snowfall data from
the eight weather stations in the region (Fig. 2). Because
weather stations recorded snow depth but not snow as precipitation, snow was converted based on a ratio of 8 for snow
depth to equivalent precipitation. The percentage of snow as
precipitation was assumed to be zero for May through Sep-
0.090
Þ
0.5×10–7
ðT aÞ=b
–1.2×10–7
PPASm ¼ 1=ð1 þ e
Value
–8.965
0.012
0.002
–0.087
½2
5
136.37
518.47
6.20
0.033
0.85
SE
1.475
0.004
0.010
0.046
where d is the distance in kilometres from the sample point
to PRISM pixel centers and calculations were limited to be
within a 14 km window. To avoid issues related to excessively widening or narrowing temperature differences related
to independent rescaling of monthly minimum and maximum
temperatures, we used geographically weighted regression to
determine the average monthly temperature as well as the difference between monthly averages of minimum and maximum temperatures at each plot. We then used those rescaled
data to recalculate the minimum and maximum monthly temperatures. GDD are the degrees Celsius greater than zero for
each day summed over all of the days of the growing season,
where growing season is defined as the calculated interval
where mean temperature is above 0 °C.
MINSPRINGTEMP and MINFALLTEMP represent minimum daily temperature averaged over a month and then averaged over the years 1961–1990, with the lowest value chosen
from May or June (for MINSPRINGTEMP) and August or
September (for MINFALLTEMP). MINTEMP is derived by
averaging minimum daily temperature over a month, then
averaging minimum daily temperature over the years 1961–
1990, and then selecting the lowest value from the 12
months. MINTEMPSD is the standard deviation of the minimums over the 12 months of the calendar year. ET uses the
Hargreaves method described in Narongrit and Yasuoka
(2003), with CMI calculated as shown in Latta et al. (2009).
The solar radiation variable (RADIANS) is calculated as the
sum of direct radiation (Swift 1976) and indirect solar radiation (Coops et al. 2000).
Precipitation was divided into rain and snow, with snow
models adapted from Wang et al. (2006):
Model D
0:5ðd=5Þ2
Model A
W ¼e
Number of variables
c2 score
AIC
AIC – AICmin
Akaike weight
Area under ROC curve
Parameter
b0
SLOPE
GDD
RADIANS
MINTEMP
MINTEMPSD
RAIN
RAIN2
SNOW
MINFALLTEMP
MINSPRINGTEMP
½1
0.014
0.004
0.010
mean values from 1961 to 1990 at a 2 km spatial resolution.
PRISM’s 2 km resolution is large relative to the distance between sea level and tree line in coastal Alaska, and therefore,
location and elevation of each plot were used to rescale the
climate data. Like Hamann and Wang (2005), we rescaled
precipitation data using two-dimensional linear interpolation;
however, we used geographically weighted regression to rescale temperature data. Weights (W) used for the latter were
calculated as
<0.001
649
Table 4. Results for prediction of hemlock dwarf mistletoe (Arceuthobium tsugense) in coastal Alaska for four candidate logistic models.
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
Barrett et al.
Published by NRC Research Press
650
Can. J. For. Res. Vol. 42, 2012
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
tember; while the records between 1954 and 2008 did show
occasional snow during those months in some years at some
locations, it was a very small portion of annual snowfall. Total annual snowfall as precipitation (SNOW) was calculated
as PPASm multiplied by precipitation for each month and
summed over the year. Rainfall (RAIN) was calculated as
precipitation minus SNOW.
Model development
For developing a predictive logistic model of A. tsugense
presence, we used the probability function
pi ¼
efi ðxÞ
1 þ efi ðxÞ
where pi is the probability of A. tsugense being present at location i and f(x) is a linear combination of the climate variables:
f ðxÞ ¼ b0 þ b1 x1 þ b2 x2 þ ::: þ bn xn
and the log odds (logit) of A. tsugense presence was thus
pi
ln
1 pi
Logistic regression models and plots of the smoothed loess
were used to examine the relationship between each climate
variable and the logit for A. tsugense. The smoothed loess
graphs suggested that there was a possible nonlinear relationship between A. tsugense probability and precipitation: a decline in probability after 2000 mm of total precipitation in
southeast Alaska but an increase in probability at lower levels
of precipitation. Thus, RAIN2 was included as a possible
term, leaving the set of candidate variables as SLOPE, GDD,
RADIANS, MINTEMP, MINTEMPSD, RAIN, RAIN2,
SNOW, MINFALLTEMP, and MINSPRINGTEMP. Data for
individual plots were randomly split into development (two
thirds of plots) and validation (one third of plots) sets. The
SAS logistic procedure (SAS Institute Inc. 2009) was used
to select candidate models, with the significance level for
variables to enter and leave the model set at 0.15 (Lee and
Koval 1997). Because of model selection uncertainty, caused
by high correlation among climate variables and limited
knowledge of causal relationships, a model averaging approach (Burnham and Anderson 2002) was applied to four
candidate models. The first two candidate models were selected using forward selection and backward elimination.
Three variables that were significant (SNOW, RAIN, and
MINFALLTEMP) as single-variable predictors were not included in either of those candidate models. To vary the models considered, two additional candidate models were chosen
by selecting the best (by the Akaike information criterion
(AIC)) six-variable models that included the SNOW and
RAIN terms, respectively. Models with low AIC scores either
did not contain the MINFALLTEMP term or, if included, the
variable was not significant (p ≥ 0.15). Accuracy of the logistic models was also assessed with receiver operating characteristic (ROC) curves, confusion matrices, and comparison
of the predicted geographic range against observed locations.
A similar process was used for developing logistic models for
T. heterophylla.
In addition to the logistic models, we used two nonparametric methods for predicting A. tsugense distribution as a
function of climate: RF imputation (Crookston and Finley
2008) and MSN imputation (Moeur and Stage 1995), implemented with R 2.10.0 (R Development Core Team 2009) and
the R packages randomForest 4.5 and yaImpute (Crookston
and Finley 2008). Explanatory variables used were SLOPE,
GDD, RADIANS, ET, CMI, RADIANS, MINTEMP, MINTEMPSD, RAIN, RAIN2, SNOW, MINFALLTEMP, and
MINSPRINGTEMP. As both RF and MSN are well suited
for problems with multiple response variables (Cutler et al.
2007; Eskelson et al. 2009a), presence of host and pathogen
was predicted together.
To examine how A. tsugense and T. heterophylla distributions could change as a function of future climate, final models were also used to predict probability of occurrence with
climate variables updated to 25, 50, and 100 years in the future. For climate projection data, we used downscaled GCM
composites created by the Scenarios Network for Alaska
Planning (SNAP) (2011). The composite models were made
from the MPI ECHAM5, GFDL CM2.1, MIROC 3.2 (medres), UKMO HADCM3, and CCCma CGCM3.1 models,
which had been chosen based on relatively good performance
in an evaluation of GCMs for Alaska and Greenland by
Walsh et al. (2008). The A1B (rapid economic growth, globalization, balanced energy), A2 (less economic growth but
more population growth than A1, regionalization), and B1
(same population growth as A1 but more sustainable development) scenarios were used; see Nakicenovic et al. (2000)
for full descriptions of scenarios. Temperature and precipitation values were projected by calculating the difference between the GCM projected value at 2011–2020, 2041–2050,
and 2090–2099 and the GCM 1990–1999 base period and
adding this to the PRISM values for current (1960–1990) climate. Derived climate variables such as GDD and MINSPRINGTEMP were then recalculated from the adjusted
precipitation and temperature values.
Results
Current abundance and distribution of A. tsugense
Using the Pacific Northwest data set, an estimated 4.8% of
the 2.259 billion T. heterophylla trees in Oregon, Washington, and California are infected with A. tsugense. Within the
temperate rainforest ecoregion of Alaska (excluding national
forest wilderness), an estimated 3.1% of 1.686 billion T. heterophylla trees are infected with A. tsugense. The estimated
forest area with A. tsugense present is 752 000 ha in the Pacific Northwest and 376 000 ha in Alaska.
Combining the Alaska data set with the Pacific Northwest
data set shows that the infection rate of individual T. heterophylla trees by A. tsugense varies by an interaction of latitude
and elevation. At latitudes from 42.0 to 46.4°, the midpoint
of which is near Eugene, Oregon, the mean infection rate
(percentage of T. heterophylla trees with A. tsugense) was estimated as 7.0% (standard error (SE) = 1.2%). For 46.5–49.0°
latitude, centered near Seattle, Washington, the mean infection rate was estimated as 3.3% (SE = 0.6%). Mean infection
rate in Alaska, where latitudes are greater than 54.7°, was
3.1% (SE = 0.4%). The highest levels of infection rates
within each latitudinal range were from 600 to 1400 m elevaPublished by NRC Research Press
Barrett et al.
651
Table 5. Results for prediction of western hemlock (Tsuga heterophylla) in coastal Alaska for three candidate logistic models.
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
Model A
Number of variables
c2 score
AIC
AIC – AICmin
Akaike weight
Area under ROC curve
Parameter
b0
SLOPE
GDD
ET
CMI
RADIANS
MINTEMP
MINTEMPSD
MINFALLTEMP
MINSPRINGTEMP
SNOW
RAIN
Value
0.848
0.015
Model B
5
610.35
624.47
5.66
0.055
0.92
SE
1.395
0.004
p
0.543
<0.001
Value
–1.161
0.018
2.8×10–3
Model C
6
613.66
618.81
0.00
0.937
0.92
SE
p
1.851
0.531
0.004
<0.001
0.8×10–3 <0.001
Value
3.877
0.018
5
601.18
628.32
9.51
0.008
0.92
SE
1.181
0.004
p
0.001
<0.001
0.129
0.024
<0.001
0.097
0.021
<0.001
–0.126
–0.307
0.015
0.037
<0.001
<0.001
–0.133
–0.334
0.015
0.037
<0.001
<0.001
–0.130
–0.348
0.016
0.038
<0.001
<0.001
0.214
0.019
<0.001
0.171
0.024
<0.001
0.224
0.020
<0.001
8.5×10–4
1.7×10–4 <0.001
tion in Oregon, from 200 to 800 m in Washington, and from
0 to 200 m in Alaska. Because T. heterophylla is not common in California, only 28 plots with the species were in the
inventory so that reliable estimates of A. tsugense infection
rates by elevation could not be made for California.
Within the existing broad range of A. tsugense in southeast
Alaska (longitude ≤ –138°), both host and pathogen abundance decreased with elevation, but the pathogen reached an
elevation threshold well before that of the host; T. heterophylla trees were found on plots that ranged in elevation
from sea level to 825 m, while A. tsugense was found on
plots from sea level to 396 m. The infection rate also decreased with elevation; infection by A. tsugense varied from
a high of 20% at sea level to 5% by 200 m elevation in
southeast Alaska (Fig. 4). Using the modeled climate variables, plots with A. tsugense had an average of an additional
22 days per year above freezing, had an additional 11 days
above 10 °C, had warmer minimum spring and fall temperatures, had less solar radiation, and had less snow and rain than
plots with T. heterophylla and without A. tsugense (Table 1).
Modeling results, logistic model
Most climate terms were correlated with each other (Table 2), with high correlations (>0.80) among those variables
related to overall warmer temperatures (GDD, MINTEMP,
MINFALLTEMP, and MINSPRINGTEMP). Fitting models
for each independent variable showed that all variables except
SLOPE were individually significant in predicting A. tsugense presence (Table 3). The sign of coefficients for most
univariate (single independent variable) models corresponded
to expectations, with increases in A. tsugense probability related to increases in the temperature variables (GDD, MINTEMP,
MINFALLTEMP,
and
MINSPRINGTEMP).
MINSPRINGTEMP created the greatest improvement in the
deviance score (145) followed by GDD (129) and MINFALLTEMP (118).
Sign and magnitude of climate term coefficients were similar among the four candidate A. tsugense models (Table 4).
The seven-variable model (SLOPE, GDD, RADIANS, MINTEMP, MINTEMPSD, RAIN2, and MINSPRINGTEMP;
model B in Table 4) had the lowest AIC score and highest
c2 score, with an Akaike weight of 0.742 (Table 4) in relation to the four candidate models. Area under the ROC curve
was more than 0.83 for all four models, a value that Hosmer
and Lemeshow (2000, p. 162) characterized as excellent discrimination.
For the averaged logistic model, we chose a classification
cutoff value for the estimated probability of 0.276 because it
resulted in a predicted 10% infection rate, which was the rate
of infection observed in the developmental data set (103/
1033 = 0.10). The confusion matrix showed sensitivity (A.
tsugense presence that was classified correctly) of 38% for
the development data set and 37% for the validation data set.
Specificity (correctly classified absence of A. tsugense) was
93% for the development data set and 93% for the validation
data set. Predicted values for both data sets were visually
compared with the current spatial range in a GIS, and all
plots that were predicted to have A. tsugense were within the
current geographic range in southeast Alaska.
A similar process for developing logistic models for T. heterophylla presence resulted in three identified candidate models, with all three models having an ROC curve area of 0.92
and high significance for individual terms (Table 5). Model
averaging of these three models using the Akaike weights
(Burnham and Anderson 2002) and a cutoff value of 0.69 resulted in predictions with sensitivity (correct prediction of
presence for T. heterophylla) of 90% for the development
data set and 91% for the validation data set. Specificity for
T. heterophylla predictions was 81% for the development
data set and 80% for the validation data set (Table 6). A visual comparison using GIS indicated that in general the predicted range matched the observed range, with correct
prediction of an absence of T. heterophylla on Kodiak Island
and the western Kenai Peninsula. One predicted location for
presence was 33 km farther northwest than the current range
on the Kenai Peninsula shown by Viereck and Little (2007),
Published by NRC Research Press
652
Can. J. For. Res. Vol. 42, 2012
Table 6. Confusion matrixa showing actual versus predicted number of observations of hemlock dwarf mistletoe (Arceuthobium tsugense) and western hemlock (Tsuga heterophylla) presence in southeast and southcentral Alaska using logistic, Random Forest (RF), and Most Similar Neighbors (MSN) imputation.
Actual
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
Logisticb
Predicted
Arceuthobium tsugense
Development data set (n = 1033)
Present
Absent
Validation data set (n = 516)
Present
Absent
Tsuga heterophylla
Development data set (n = 1033)
Present
Absent
Validation data set (n = 516)
Present
Absent
Present
RF
Absent
Present
MSN
Absent
Present
Absent
39
64
64
866
38
65
19
911
30
73
78
852
17
29
35
435
9
37
20
450
11
35
46
424
607
67
67
292
650
24
61
298
579
95
94
265
283
27
42
164
292
18
46
160
273
37
63
143
a
Sensitivity, a metric for accuracy of predicting presence, is calculated from the confusion matrix as the percentage of
observed presences that were correctly predicted. Specificity, a metric for accuracy of predicting absence, is calculated as
the percentage of observed absences that were correctly predicted. For example, for the development data set and A. tsugense, sensitivity of the RF model is 38/(38 + 66) = 37%, specificity is 911/(19 + 911) = 98%, and bias can be understood by comparing predicted presence ((38 + 19)/1033 = 5.5%) with observed presence ((38 + 65)/1033 = 10.0%).
b
Logistic model classification used a cutpoint of 0.276 for A. tsugense and 0.690 for T. heterophylla.
and overall, there were fewer predictions of T. heterophylla
presence in forested areas bordering the Gulf of Alaska than
were found in the inventory. While the two logistic models
for host and pathogen had been developed separately, there
were no instances in either the development or validation
data set of locations where A. tsugense was predicted and T.
heterophylla was not.
Modeling results, most similar neighbor and random
forest models
Sensitivity of predicting A. tsugense presence was better
for RF (37%) than for MSN (29%) for the development data
set (Table 6). For the validation data set, sensitivity was better for MSN (24%) than for RF (20%). Specificity for both
the development and validation sets was better for RF (98%
and 96%) than for MSN (92% and 90%). MSN had a slight
bias toward overpredicting A. tsugense presence, with 10.5%
of development plots and 11.0% of validation plots predicted
as having A. tsugense infection compared with actual rates of
10.0% and 8.9%. RF underpredicted rates of A. tsugense infection, with predictions of 5.6% and 5.5% for the development and validation data sets, respectively. Spatially, all plots
in both data sets that were predicted to have A. tsugense presence were within the current geographic range of A. tsugense
for both RF and MSN.
RF slightly overpredicted T. heterophylla presence, with
68.8% of development plots and 65.5% of validation plots
predicted as having T. heterophylla compared with actual
rates of 65.2% and 60.1% (Table 6). MSN was unbiased in
prediction of T. heterophylla presence for development plots
(65.2%) and slightly overpredicted T. heterophylla presence
for the validation data set (65.1%). For MSN, sensitivity of
T. heterophylla prediction was 85.9% for the development
data set and 88.1% for the validation data set, and specificity
was 73.8% for the development data set and 69.4% for the
validation set. For RF, sensitivity of T. heterophylla prediction was 96.4% for the development data set and 94.2% for
the validation data set, and specificity was 83.0% for the development data set and 77.7% for the validation data set.
Spatially, RF and MSN both correctly predicted an absence
of T. heterophylla on Kodiak Island and the western Kenai
Peninsula. While the RF-predicted range matched the actual
inventory observations closely, MSN made four out-of-range
predictions of T. heterophylla presence in the Anchorage and
Kenai Fjords regions, all within 75 km from the current
range.
Modeling results, prediction of future A. tsugense for
altered climate
For T. heterophylla, all three models (RF, MSN, and logistic) predicted increases of 5%–18% by the fifth decade for the
three climate scenarios (A1B, A2, and B1). The models
showed greater differentiation by the 10th decade, with RF
showing less increase than either the MSN or logistic model
(Table 7). At the 10th decade, there was greater variation between modeling methods than between climate scenarios.
For A. tsugense, the models predicted increases of 166%–
509% by the fifth decade and increases of 242%–856% by
the 10th decade (Table 7). Adjusting by the amount of bias
shown for the validation data set, the predicted increases are
186%–450% by the fifth decade and 374%–757% by the 10th
decade. The logistic model predicted the greatest amount of
increase for all three climate scenarios. Standard deviation
between models (MSN, RF, and logistic) was much greater
than standard deviation between scenarios (A1B, A2, and
B1). While the majority of predicted potential increase of A.
Published by NRC Research Press
Barrett et al.
653
Table 7. Projection of hemlock dwarf mistletoe (Arceuthobium tsugense) and western hemlock
(Tsuga heterophylla) frequency as a percentage of modeled 1990–1999 values, by model and
climate scenario, for southeast and southcentral Alaska.
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
Decade
Model and scenario
Arceuthobium tsugense
Most Similar Neighbors
A1B
A2
B1
Random Forests
A1B
A2
B1
Logistic
A1B
A2
B1
Tsuga heterophylla
Most Similar Neighbors
A1B
A2
B1
Random Forests
A1B
A2
B1
Logistic
A1B
A2
B1
2011–2020
2041–2050
2091–2100
2091–2100a
115
171
158
357
230
236
715
739
515
577
596
416
74
99
93
166
197
147
242
283
236
384
449
374
135
248
305
464
509
402
818
856
645
724
757
571
100
102
106
118
117
109
149
153
128
137
141
118
98
103
105
115
114
110
122
123
117
112
113
107
100
102
105
112
116
105
141
143
121
134
136
115
a
Adjusted by the ratio of observed to predicted presence from the validation data set.
tsugense occurs in southeast Alaska within the current geographic range, there is also a westward expansion of potential
infection along the Gulf of Alaska (Fig. 5).
The consistency of the logistic models for predicting
pathogen presence only where the host occurred did not hold
when applied to future climates. For the 2090–2100 decade,
about 0.6% of predictions of A. tsugense presence occurred
for locations where T. heterophylla was not predicted to occur.
Discussion
Predictions over the next century from all three climate envelope models indicated a substantial difference between potential impacts on the host tree (T. heterophylla) and potential
impacts on the pathogen (A. tsugense). Results for the host
tree indicated modest potential increases in abundance and
range expansion. In contrast, the models predicted extremely
large potential increases in abundance of A. tsugense, mostly
within its current range but also from an 800 km westward
expansion in potential habitat.
However, predictions are only for potential habitat and do
not incorporate migration rates. Landscape-level dispersal of
Arceuthobium can be viewed as a combination of short-distance explosive seed dispersal each autumn and less frequent
long-distance transmission by birds and other animals (Lundquist 2005). The ability of A. tsugense to spread over short
distances through explosive seed dispersal has been well es-
tablished. In contrast, dispersal over longer distances, while
known to occur, has much more uncertainty. The presence of
A. tsugense on many of the islands within the Alexander Archipelago of Alaska that were covered by glaciers until the
end of the late Wisconsin glaciation 13 000 B.P. indicates
that long-distance transmission does occur. While dispersal
rates of T. heterophylla are also uncertain, pollen records
suggest that it recolonized southeast Alaska as early as 10
000 B.P. (Ager et al. 2010) but did not reach the Prince William Sound area until about 2000 B.P. (Heusser 1983).
As dispersal is not addressed in this paper, predictions
from the climate envelope models should be viewed as predictions of potential rather than actual occupancy. A related
difficulty posed for this analysis is that the degree to which
both species have occupied their potential climate niches is
unknown. Tidewater glaciers, fjords, and mountains provide
barriers to dispersal in this region, and it may be that T. heterophylla is still migrating westward, as is known to be the
case for P. sitchensis (Griggs 1934).
For land managers, the prediction for large potential increases in southeast Alaska, where A. tsugense is already
present, is of more immediate applicability than the prediction for potential range expansion westward along the Gulf
of Alaska coastline. Reduced snow and longer growing seasons in southeast Alaska could allow A. tsugense to migrate
upslope above the 200 m threshold where it is currently rare.
In southeast Alaska, a change to a more favorable climate for
Published by NRC Research Press
654
Can. J. For. Res. Vol. 42, 2012
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
Fig. 5. Current distribution of western hemlock (Tsuga heterophylla) and hemlock dwarf mistletoe (Arceuthobium tsugense) from inventory
data and Most Similar Neighbor current, fifth decade, and 10th decade predictions.
A. tsugense would allow it to reproduce more effectively with
a resulting increase in disease effects (growth loss and mortality) for T. heterophylla. It may be that in the future, the
faster rate of spread and completion of the A. tsugense life
cycle that is currently seen in British Columbia (Muir and
Hennon 2007) becomes the standard for how the pathogen
behaves in southeast Alaska.
The current abundance and distribution of A. tsugense in
the temperate rainforests of Alaska appear to be strongly related to climate. The parsimonious logistic models that we
selected included explanatory climate variables of growing
degree-days, rainfall and snowfall, direct and indirect solar
radiation, minimum winter and spring temperatures, and variation in minimum temperatures. Most variables were correlated, and when extended to future climate scenarios, there is
an assumption made that the correlation between climate variables will be the same. This seems more likely to hold true
between temperature variables than for interactions between
temperature and precipitation. Correlation among climate variables also means that modeling results cannot be used to reject any particular hypothesis about causal mechanisms, e.g.,
that limitation of range is caused by a short cool growing
season rather than minimum spring temperatures or snow accumulation. In our literature review, we found that much of
the previous research on A. tsugense has been focused on
management, and some surprisingly basic information, such
as the impact of precipitation on seed establishment, or
whether pollination occurs from insects or wind, is unknown.
Some additional field research on causal mechanisms influencing A. tsugense establishment and reproduction could be
helpful in refining understanding of the relationship between
the phenology of A. tsugense and climate. An additional limitation of the models may be the use of average monthly minimum temperatures, as temperature extremes are more likely
to be injurous to both A.tsugense and T. heterophylla.
None of the three modeling approaches used here (logistic
models, RF imputation, and MSN imputation) was consistently best by all measures. RF showed some evidence of
overfitting (through a drop in accuracy between the development and validation data sets) and underpredicted A. tsugense
presence, although it also performed the best for prediction of
the development data set. The logistic modeling approach has
the advantage of having a number of well-understood diagnostics (ROC curves, AIC scores, probability plots) that can
be used for comparing models and developing understanding
of relationships, in contrast with the two imputation methods.
Climate envelope modeling is sometimes criticized for
being too simplistic to help predict actual changes in species’
distribution and abundance in that disturbance, competition,
and the influence of biotic factors such as insects and pathoPublished by NRC Research Press
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
Barrett et al.
gens are not considered. The methods presented in this paper
help to address the last concern by modeling climate effects
on both a host tree species and an infecting pathogenic species. While our expectations were that the imputation methods of MSN and RF would have an advantage by
simultaneously predicting host and pathogen, this only
proved to be an advantage for prediction of future climates
because the logistic models did not predict the presence of
A. tsugense without T. heterophylla for the current climate. It
may be that A. tsugense’s climate range is well within T. heterophylla’s climate range as a result of evolution with a single primary host; multivariate prediction methods might have
a stronger advantage in situations where this was not the
case, e.g., where a nonnative pathogen has been recently introduced to a new host species.
Conclusion
There are strong indications that climate is currently limiting A. tsugense to a subset of the range of its primary host, T.
heterophylla, in the northern temperate rainforest. Predictions
from models used here suggest large increases in potential
habitat for A. tsugense in Alaska and modest increases possible for T. heterophylla. While increases within the current
range in southeast Alaska can occur from normal explosive
seed discharges over short distances, expansion of range
westward along the Gulf of Alaska is dependent on dispersal
by birds and mammals, the rate of which is unknown.
Implications for southeast Alaska include reduced capacity
for carbon sequestration, potential beneficial impacts for
wildlife, and higher risk of negative impacts for those portions of the forest managed for timber production, which
may alter silvicultural choices. The opposing influences on
T. heterophylla, from an increase in habitat accompanied by
an increase in a major pathogen, illustrate the complexity of
prediction of climate impacts on plant species.
Substantial uncertainty as to the amount of potential increase in A. tsugense habitat exists. The use of different modeling approaches contributed more to the range of possible
outcomes than did the use of different climate scenarios, suggesting that using multiple modeling approaches is beneficial
to understanding uncertainty of predictions. In the future, collection and analysis of inventory data, similar to the approach
used in this paper, could be helpful in monitoring the predicted increase in disease behavior.
References
Ager, T.A., Carrara, P.E., Smith, J.L., Anne, V., and Johnson, J. 2010.
Postglacial vegetation history of Mitkof Island, Alexander
Archipelago, southeastern Alaska. Quat. Res. 73(2): 259–268.
doi:10.1016/j.yqres.2009.12.005.
Berg, E.E., Hillman, K.M., Dial, R., and DeRuwe, A. 2009. Recent
woody invasion of wetlands on the Kenai Peninsula Lowlands,
south-central Alaska: a major regime shift after 18 000 years of
wet Sphagnum–sedge peat recruitment. Can. J. For. Res. 39(11):
2033–2046. doi:10.1139/X09-121.
Bloomberg, W.J., Smith, R.B., and Van Der Wereld, A. 1980. A
model of spread and intensification of dwarf mistletoe infection in
young western hemlock stands. Can. J. For. Res. 10(1): 42–52.
doi:10.1139/x80-008.
Brandt, J.P., Hiratsuka, Y., and Pluth, D.J. 2004. Extreme cold
temperatures and survival of overwintering and germinated
655
Arceuthobium americanum seeds. Can. J. For. Res. 34(1): 174–
183. doi:10.1139/x03-200.
Breiman, L. 2001. Random forests. Mach. Learn. 45(1): 5–32. doi:10.
1023/A:1010933404324.
Burnham, K.P., and Anderson, D.R. 2002. Model selection and
multimodel inference. 2nd ed. Springer-Verlag, New York.
Coops, N.C., Waring, R.H., and Moncrieff, J.B. 2000. Estimating
mean monthly incident solar radiation on horizontal and inclined
slopes from mean monthly temperatures extremes. Int. J.
Biometeorol. 44(4): 204–211. doi:10.1007/s004840000073.
PMID:11131293.
Crookston, N.L., and Finley, A.O. 2008. YaImpute: an R package for
kNN imputation. J. Stat. Softw. 23(10): 1–16.
Cutler, D.R., Edwards, T.C., Jr, Beard, K.H., Cutler, A., Hess, K.T.,
Gibson, J., and Lawler, J.J. 2007. Random forests for classification
in ecology. Ecology, 88(11): 2783–2792. doi:10.1890/07-0539.1.
PMID:18051647.
Eskelson, B.N.I., Temesgen, H., Lemay, V., Barrett, T.M., Crookston,
N.L., and Hudak, A.T. 2009a. The role of nearest neighbor
methods in imputing missing data in forest inventory and
monitoring databases. Scand. J. For. Res. 24(3): 235–246.
doi:10.1080/02827580902870490.
Eskelson, B.N.I., Temesgen, H., and Barrett, T.M. 2009b. Estimating
cavity tree and snag abundance using negative binomial regression
models and nearest neighbor imputation methods. Can. J. For. Res.
39(9): 1749–1765. doi:10.1139/X09-086.
Gilbert, J., and Punter, D. 1991. Germination of pollen of the dwarf
mistletoe Arceuthobium americanum. Can. J. Bot. 69(3): 685–688.
doi:10.1139/b91-092.
Griggs, R.F. 1934. The edge of the forest in Alaska and the reasons
for its position. Ecology, 15(2): 80–96. doi:10.2307/1932778.
Hamann, A., and Wang, T.L. 2005. Models of climatic normals for
genecology and climate change studies in British Columbia. Agric.
For. Meteorol. 128(3–4): 211–221. doi:10.1016/j.agrformet.2004.
10.004.
Hamer, T.E., Varland, D.E., McDonald, T.L., and Meekins, D. 2008.
Predictive model of habitat suitability for the marbled murrelet in
western Washington. J. Wildl. Manage. 72(4): 983–993. doi:10.
2193/2006-565.
Hawksworth, F.G. 1977. The 6-class dwarf mistletoe rating system.
U.S. For. Serv. Gen. Tech. Rep. RM-48.
Hawksworth, F.G., and Wiens, D. 1996. Dwarf mistletoes: biology,
pathology, and systematics. U.S. Dep. Agric. Agric. Handb. 709.
Heusser, C.J. 1983. Holocene vegetation history of the Prince
William Sound region, south-central Alaska. Quat. Res. 19(3):
337–355. doi:10.1016/0033-5894(83)90040-6.
Hosmer, D.W., and Lemeshow, S. 2000. Applied logistic regression.
2nd ed. John Wiley & Sons, Inc., New York.
Iverson, L.R., and Prasad, A.M. 1998. Predicting abundance of 80
tree species following climate change in the eastern United States.
Ecol. Monogr. 68(4): 465–485. doi:10.1890/0012-9615(1998)068
[0465:PAOTSF]2.0.CO;2.
Kliejunas, J.T., Geils, B.W., Glaeser, J.M., Goheen, E.M., Hennon, P.,
Mee-Sook, K., Kope, H., Stone, J., Sturrock, R., and Frankel, S.J.
2009. Climate and forest diseases of western North America: a
literature review. U.S. For. Serv. Gen. Tech. Rep. PSW-GTR-225.
Latta, G., Temesgen, H., and Barrett, T.M. 2009. Mapping and
imputing potential productivity of Pacific Northwest forests using
climate variables. Can. J. For. Res. 39(6): 1197–1207. doi:10.
1139/X09-046.
Lee, K.I., and Koval, J.J. 1997. Determination of the best significance
level in forward stepwise logistic regression. Comm. Statist.
Simulation
Comput.
26(2):
559–575.
doi:10.1080/
03610919708813397.
Published by NRC Research Press
Can. J. For. Res. Downloaded from www.nrcresearchpress.com by OREGON STATE UNIVERSITY on 05/07/12
For personal use only.
656
Lundquist, J.E. 2005. Patterns in diseased landscapes: a case study of
a lodgepole pine forest infected by dwarf mistletoe. Chap. 15. In
Forest pathology: from genes to landscapes. The American
Phytopathological Society, St. Paul, Minn.
Mantua, N.J., and Hare, S.R. 2002. The Pacific decadal oscillation. J.
Oceanogr. 58(1): 35–44. doi:10.1023/A:1015820616384.
Meinzer, F.C., Woodruff, D.R., and Shaw, D.C. 2004. Integrated
responses of hydraulic architecture, water and carbon relations of
western hemlock to dwarf mistletoe infection. Plant Cell Environ.
27(7): 937–946. doi:10.1111/j.1365-3040.2004.01199.x.
Moeur, M., and Stage, A.R. 1995. Most similar neighbor: an
improved sampling inference procedure for natural resource
planning. For. Sci. 41(2): 337–359.
Muir, J.A., and Hennon, P.E. 2007. A synthesis of the literature on the
biology, ecology, and management of western hemlock dwarf
mistletoe. U.S. For. Serv. Gen. Tech. Rep. PNW-GTR-718.
Nakicenovic, N., Alcamo, J., and Davis, G., de Vries, B., et al. 2000.
Special report on emissions scenarios. Intergovernmental Panel on
Climate Change, The Hague, The Netherlands. Available from
http://www.grida.no/publications/other/ipcc_sr/?src=/climate/ipcc/
emission/
Narongrit, C., and Yasuoka, Y. 2003. The use of Terra-MODIS data
for estimating evapotranspiration and its change caused by global
warming. Environ. Informatics Arch. I: 505–511.
Pope, V.D., Gallani, M.L., Rowntree, P.R., and Stratton, R.A. 2000.
The impact of new physical parametrizations in the Hadley Centre
climate model: HadAM3. Clim. Dyn. 16(2–3): 123–146. doi:10.
1007/s003820050009.
PRISM Climate Group. 2002. Alaska average annual precipitation
and temperature 1961–1990 [online]. Oregon State University,
Corvallis, Oreg. Available from http://prism.oregonstate.edu/
[accessed 26 September 2011].
R Development Core Team. 2009. R: a language and environment for
statistical computing [online]. R Foundation for Statistical
Computing, Vienna, Austria. Available from http://www.r-project.org [accessed 26 September 2011].
Roth, L.F. 1959. Natural emplacement of dwarf-mistletoe seed on
ponderosa pine. For. Sci. 5: 365–369.
SAS Institute Inc. 2009. Base SAS 9.2 procedures guide. SAS
Institute Inc., Cary, N.C.
Scenarios Network for Alaska Planning (SNAP). 2011. Alaska
climate datasets [online]. Available from http://www.snap.uaf.edu/
data.php [accessed 5 March 2012].
Scott, C.T., Bechtold, W.A., Reams, G.A., Smith, W.D., Westfall, J.
A., Hansen, M.H., and Moisen, G.G. 2005. Sample based
estimators used by the forest inventory and analysis national
Can. J. For. Res. Vol. 42, 2012
information management system. In The enhanced forest inventory
and analysis program — national sampling design and estimation
procedures. Edited by W.A. Bechtold and P.L. Patterson. U.S. For.
Serv. Gen. Tech. Rep. GTR-SRS-80. pp. 43–67.
Shaw, C.G., III, and Hennon, P.E. 1991. Spread, intensification, and
upward advance of dwarf mistletoe in thinned, young stands of
western hemlock in southeast Alaska. Plant Dis. 75(4): 363–367.
doi:10.1094/PD-75-0363.
Shaw, C.G., III, and Loopstra, E.M. 1991. Development of dwarf
mistletoe infections on inoculated western hemlock trees in
southeast Alaska. Northwest Sci. 65(1): 48–52.
Shaw, D., Freeman, E.A., and Mathiasen, R.L. 2000. Evaluating the
accuracy of ground-based hemlock dwarf mistletoe rating: a case
study using the Wind River canopy crane. West. J. Appl. For. 15
(1): 8–14.
Shields, O. 1965. Callophrys (Mitoura) spinetorum and C. (M.)
johnsoni: their known range, habits, variation, and history. J. Res.
Lepid. 4(4): 233–250.
Smith, R.B. 1971. Development of dwarf mistletoe (Arceuthobium)
infections on western hemlock, shore pine, and western larch. Can.
J. For. Res. 1(1): 35–42. doi:10.1139/x71-005.
Swift, L.W., Jr. 1976. Algorithm for solar radiation on mountain
slopes. Water Resour. Res. 12(1): 108–112. doi:10.1029/
WR012i001p00108.
Vandermeer, J.H. 1972. Niche theory. Annu. Rev. Ecol. Syst. 3(1):
107–132. doi:10.1146/annurev.es.03.110172.000543.
Viereck, L.A., and Little, E.L., Jr. 2007. Alaska trees and shrubs. 2nd
ed. University of Alaska Press, Fairbanks, Alaska.
Walsh, J.E., Chapman, W.L., Romanovsky, V., Christensen, J.H., and
Stendel, M. 2008. Global climate model performance over Alaska
and Greenland. J. Clim. 21(23): 6156–6174. doi:10.1175/
2008JCLI2163.1.
Wang, T., Hamann, A., Spittlehouse, D.L., and Aitken, S.N. 2006.
Development of scale-free climate data for western Canada for use
in resource management. Int. J. Climatol. 26(3): 383–397. doi:10.
1002/joc.1247.
Watson, D.M. 2001. Mistletoe — a keystone resource in forests and
woodlands worldwide. Annu. Rev. Ecol. Syst. 32(1): 219–249.
doi:10.1146/annurev.ecolsys.32.081501.114024.
Wicker, E.F., and Shaw, C.G. 1968. Fungal parasites of dwarf
mistletoes. Mycologia, 60(2): 372–383. doi:10.2307/3757167.
Wiens, J.A., Stralberg, D., Jongsomjit, D., Howell, C.A., and Snyder,
M.A. 2009. Niches, models, and climate change: assessing the
assumptions and uncertainties. Proc. Natl. Acad. Sci. U.S.A. 106
(Supple. 2): 19729–19736. doi:10.1073/pnas.0901639106. PMID:
19822750.
Published by NRC Research Press