23782242, 2017, 3, Downloaded from https://aslopubs.onlinelibrary.wiley.com/doi/10.1002/lol2.10037 by National Sun-Yat-Sen, Wiley Online Library on [13/03/2023].

See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Limnology and Oceanography Letters 2, 2017, 63–69
C 2017 The Author. Limnology and Oceanography Letters published by Wiley Periodicals, Inc.
V
on behalf of Association for the Sciences of Limnology and Oceanography
doi: 10.1002/lol2.10037

LETTER

Extreme events in lake ecosystem time series

Ryan D. Batt ,1,a* Stephen R. Carpenter,1 Anthony R. Ives2
1
Center for Limnology, University of Wisconsin – Madison, Madison, Wisconsin; 2Department of Zoology, University of
Wisconsin – Madison, Madison, Wisconsin

Scientific Significance Statement

Extreme events are large and potentially surprising. Although they are often studied in environmental variables like precipitation,
they are far less often investigated in biological variables like population size. Are massive events common in biology? What types
of dynamics tend to produce extremes? To answer these questions, we analyzed hundreds of decades-long time series collected
from 11 lakes. We found that biological variables had the strongest tendency to produce record-breaking annual extremes, greater
than the meteorological, physical, and chemical variables that, in part, drive the biological variables. Furthermore, the biological
extremes were driven by within-year dynamics. The results imply that we should expect to be surprised by extreme fluctuations,
particularly biological ones.

Abstract
Climate change has generated growing interest in extreme events, and extremes are known to have important con-
sequences for ecosystems. Theoretical mechanisms generating frequent extremes apply to both environmental and
biological processes, yet past studies of ecological extremes have focused primarily on the abiotic environment. The
rarity or commonness of extremes in biological time series is unknown. We evaluated the statistical tendency to
produce extreme events in 595 biological, chemical, physical, and meteorological time series taken from 11 lakes.
We found that extreme events were much more likely to occur for biological variables than for other categories.
Additional analysis revealed that the tendency to produce extremes was driven primarily by within-year dynamics,
suggesting that processes occurring at short time scales underlie the high frequency of extremes in biological varia-
bles. These results should lead us to expect surprises in long-term observations of biological populations.

Extreme but uncommon events, such as floods and pest

outbreaks, pose significant risks to the environment and
*Correspondence: [email protected] society. The frequency and substantial ecosystem impacts of
a
extreme events in environmental variables like temperature
Present address: Department of Ecology, Evolution, and Natural Resources,
Rutgers University, New Brunswick, New Jersey.
and precipitation are well studied (Gaines and Denny 1993;
Peters et al. 2004; Denny et al. 2009; Fey et al. 2014). How-
Author Contribution Statement: RDB and SRC conceived the study. ever, the prevalence of extremes in ecosystem variables is
ARI designed the mechanistic model and simulations. RDB and ARI
unclear because they have not been surveyed systematically.
wrote the R code and analyzed the data; analysis was guided by feed-
back from SRC. RDB, SRC, and ARI wrote the paper. Furthermore, the historical tendency for ecosystem variables
to show extremes may depend on whether they amplify or
Data Availability Statement: Data are available in the North Temper- dampen sources of exogenous variability like weather
ate Lakes Long Term Ecological Research database (https://lter.limnology.
wisc.edu/) and at https://github.com/rBatt/FatTails. extremes, which are increasing in frequency in many places
(Mitchell et al. 2006; IPCC 2012). Thus, comparing the his-
Additional Supporting Information may be found in the online version of this torical frequency of extreme events across components of
article.
ecosystems will further our basic understanding of ecosystem
This is an open access article under the terms of the Creative Commons dynamics, particularly in the face of global change.
Attribution NonCommercial License, which permits use, distribution and
reproduction in any medium, provided the original work is properly cited Extreme value theory (Coles 2001) is commonly
and is not used for commercial purposes. employed to gauge the probabilities of large perturbations in

63
 23782242, 2017, 3, Downloaded from https://aslopubs.onlinelibrary.wiley.com/doi/10.1002/lol2.10037 by National Sun-Yat-Sen, Wiley Online Library on [13/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Batt et al. Extreme events in lake ecosystem time series

time series (Palutikof et al. 1999; Katz et al. 2002, 2005).

These large events can be minima or maxima, but hence-

ϕ
forth we restrict our discussion to maxima because ecosys-
tem variables are usually defined over the interval [0, 1).
The “tailedness” of a distribution refers to the shape of the
tail of a probability distribution, affecting its ability to pro-
duce frequent extremes and to generate record-breaking
events. We measured tailedness by fitting the generalized

γ
extreme value distribution (GEV) to observations. The GEV
has three parameters, location (l), scale (r), and shape (n). n
measures tailedness: negative n indicates a bounded upper
tail, n near 0 indicates a thin (exponential) tail such as dem-
onstrated by a normal distribution, and positive n indicates

X
that the maxima are fat-tailed.
Fat tails have been found in many types of data, although
there is little empirical description of the tailedness of time
series of most ecosystem variables, particularly biological varia-
bles. Anecdotally, ecologists often recall large events having
been observed in their study systems, but rigorous quantifica-
tion of the size and frequency of these events is lacking. Tailed-
ness has been evaluated for environmental time series like
flood stage (Villarini and Smith 2010), stream discharge and
precipitation (Katz et al. 2002), sediment deposition (Katz et al.
2005), water temperature (Gaines and Denny 1993), wind speed
(Gaines and Denny 1993; Palutikof et al. 1999), and wave force
(Gaines and Denny 1993; Denny et al. 2009). Biological time
series, by contrast, are rarely analyzed for tailedness. In the case
of Segura et al. (2013), tailedness was calculated for the changes ξ
in population size of microbes, but not for population size
Fig. 1. The relationship between the tailedness of a population and its
itself. Schmitt et al. (2008) calculated the tailedness of the size
environmental drivers in the simple illustrative model presented in Sup-
of a single population of marine zooplankton, finding the pop- porting Information: Extremes in a model population. (A) Time series of
ulation to be fat-tailed. We could not find more examples a hypothetical environmental variable u that influences birth rates, with
quantifying tailedness in biological time series. However, annual maxima indicated by blue circles. Data were simulated for 30 yr
extreme events produced by fat-tailed time series are known to with 20 measurements per year to represent the typical time series in
our empirical analyses. (B) A hypothetical environmental variable c influ-
have significant impacts in other settings, making it important
encing survival rates, with red circles for annual maxima. (C) Population
to identify their prevalence in biological variables. size (X), with black circles for annual maxima. (D–F) Empirical density
Despite the paucity of studies of tailedness in biological plots for the parent distributions in A–C (gray lines) and for annual max-
variables, it is plausible that biological time series could be ima (blue, red, and black lines). (G) The blue, red, and black lines show
fat-tailed. One way for a biological time series to be fat-tailed density plots of n estimated from 500 simulations of the process exhib-
ited in A–C. From theory, as the number of years becomes large, the
is for a population to amplify the variability of the surround-
estimates of nX will always exceed nu and nc, although this is not always
ing environment. As a heuristic example for how such the case for the simulated 30 yr of data.
enhancement may occur, consider a population governed by
birth and survival processes that are in turn influenced by better understand the frequency and origin of extreme
environmental (stochastic) drivers (Supporting Information: events in an ecosystem context. Our goal is to fill the knowl-
Extremes in a model population). Birth and death processes edge gap about how often biological variables may be fat-
often act multiplicatively (Pielou 1977; Gurney and Nisbet tailed, and how the tailedness of biological variables com-
1998), which causes the population to have thicker tails pares to the tailedness of their environment. We surveyed
than the environment (Fig. 1). The ubiquity of multiplicative 595 long-term lake time series to quantify the relative fre-
processes in biological systems makes this mechanism for quency of extremes in biological, chemical, physical, and
amplifying tailedness potentially widespread, although other meteorological variables. These variables were measured
mechanisms can generate fat tails such as some types of within the same ecosystems, providing a direct comparison
nonlinear dynamics (Sornette 2006; e.g., chapter 14). between biological tailedness and the tailedness of environ-
Here, we document the fat-tailedness and time series mental variables, many of which are expected to affect varia-
characteristics of biological and environmental variables to tion in the biological variables. We then analyzed these time

64
 23782242, 2017, 3, Downloaded from https://aslopubs.onlinelibrary.wiley.com/doi/10.1002/lol2.10037 by National Sun-Yat-Sen, Wiley Online Library on [13/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Batt et al. Extreme events in lake ecosystem time series

series to determine if they showed evidence for generating Distribution fitting and waiting times
fat tails at time scales shorter or longer than the sampling We fit a generalized extreme value distribution (GEV) to
interval. This comparison constitutes a direct empirical the time series of annual maxima for each variable; this is
assessment of the tailedness of a large number of biological the theoretical limiting distribution of maxima that is
time series, and may provide a clue as to if and how biology approached for many statistical processes with large enough
amplifies variability in exogenous drivers to produce extreme sample sizes, much like the normal distribution is the limit-
events. ing distribution of sums of random variables (Coles 2001).
The cumulative distribution function of the GEV is
8 n o
Materials and methods < exp 2½11nðx2lÞ=r21=n n 6¼ 0
Fðx; l; r; nÞ5
Data collection :
exp f2exp ½2ðx2lÞ=rg n50
Raw data were downloaded from the online databases of
the North Temperate Lakes Long-Term Ecological Research
where the case of n 6¼ 0 is subject to the constraint that
(NTL-LTER) program (lter.limnology.wisc.edu/; Supporting
11nðx2lÞ=r > 0. Variables that follow very different distribu-
Information). Data were downloaded on 18 April 2013,
tions can nonetheless have their tails compared by the GEV
except wind data, which were downloaded on 30 March because it is the limiting distribution for maxima. The
2014. Most time series from the northern Wisconsin lakes extremes to which we fit the GEV were defined via the
begin in 1981. Most time series from the southern lakes block-maximum method (Palutikof et al. 1999) with blocks
begin in 1996, except fish time series that begin in 1995, corresponding to years to yield annual maxima. Each time
zooplankton that begin in 1976, and three ice cover time series contained between 15 and 158 annual maxima
series that begin in 1851, 1852, and 1877. Most time series (median 5 29) and was categorized as biological (n 5 283 lake
end between 2010 and 2012. Measurement frequencies var- time series, including chlorophyll, 28 fish genera, and 34
ied among time series; meteorological time series were typi- zooplankton genera), chemical (n 5 220, e.g., ion concentra-
cally daily, fish time series were annual, and other lake time tion), physical (n 5 80, e.g., water temperature), or meteoro-
series were quarterly or every 2–6 weeks, depending on the logical (12 regional time series, e.g., mean daily wind speed).
study site and time of year. Additionally, some lake time We used the GEV to assess tailedness and estimate waiting
series were only sampled once during the winter, depending time between record-breaking extremes. The GEV parameters
on ice conditions. Empirically, observation frequency did were fit using maximum likelihood after removing a linear
not have an influence on tailedness (Supporting Information temporal trend from time series of maxima (Katz et al. 2005).
Fig. S1). Northern Wisconsin lakes include Allequash Lake, We calculated the unconditional expected waiting time as the
Big Muskellunge Lake, Crystal Bog, Crystal Lake, Little Rock inverse of the probability p (calculated from the cumulative
Lake, Sparkling Lake, Trout Bog, and Trout Lake. The only distribution function) of observing a value  10% over the cur-
time series from Little Rock Lake was ice cover. Southern rent record, multiplied by the observation frequency (F, yr21)
Wisconsin lakes include Fish Lake, Lake Mendota, Lake of maxima in the time series (F was usually 1); thus, waiting
Monona, and Lake Wingra. time 5 1/p*F. This calculation does not account for possible
Meteorological variables included air temperature (daily autocorrelation in the annual maxima in the time series; if
minimum, maximum, range), sum daily precipitation, snow there is autocorrelation, then the expected conditional wait-
depth, and daily average wind speed. Physical variables ing time will depend on the most recent value of the time
included dissolved oxygen concentration, dissolved oxygen series. Statistical analyses were performed in the statistical pro-
percent saturation, water temperature, the fraction of light gramming language R (R Core Team 2014). In some cases (79
at depth relative to light at surface, light extinction coeffi- of the 595 time series; 0 biological, 44 chemical, 32 physical, 3
cient, Secchi depth, lake level, duration of the ice-free sea- meteorological), waiting times were undefined because the
son, and the depth of the epilimnion. Chemical variables value 10% over the record (x) was outside the support of the
included alkalinity, bicarbonate-reactive silica (unfiltered), generalized extreme value distribution (GEV):
calcium, chloride, conductivity, dissolved inorganic carbon,
x 2 ½l2r=n; 11Þ when n > 0;
dissolved organic carbon, dissolved reactive silica (filtered),
iron, potassium, magnesium, manganese, sodium, ammo- x 2 ð21; 11Þ when n50;
nium, nitrate plus nitrite, pH, sulfate, total inorganic carbon, x 2 ð21; l2r=n when n < 0
total organic carbon, total nitrogen (unfiltered), total phos-
phorus (unfiltered), and total particulate matter. Biological For example, the return level could be outside the support if
variables included the abundance of fishes and zooplankton, the estimate of n is less than zero, and the return level is
and epilimnetic chlorophyll concentration. Each genus of greater than the limit of the upper tail (l–r/n). Undefined
fish or zooplankton comprised its own time series. waiting times were excluded from comparisons of waiting

65
 23782242, 2017, 3, Downloaded from https://aslopubs.onlinelibrary.wiley.com/doi/10.1002/lol2.10037 by National Sun-Yat-Sen, Wiley Online Library on [13/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Batt et al. Extreme events in lake ecosystem time series

time between variable types. Note that removing the 79 time

1.5
A
series with undefined (infinite) waiting times makes Fig. 2
conservative with respect to showing that biological time

1.0
series have shorter waiting times than the other categories.

ARMA models

0.5
Tailedness (ξ)
We used autoregressive moving average (ARMA) models
to assess the relative contributions of processes occurring at

0.0
time scales shorter (within-sample) or longer (between-sam-
ple) than a sampling interval (Supporting Information: Using

−0.5
ARMA to assess the time scale of sources of fat tails). ARMA
models are a generic class of time series model capable of
approximating many processes by predicting the current

−1.0
value of a variable from its past values and with past values
of error terms. The parameter values of an ARMA model can
be used to make general inferences about a variable’s tempo-
Bio (283) Chem (220) Phys (80) Met (12)
ral dynamics, such as if it has long-term “memory,” which is
the autocorrelation of current values with values from many
time steps ago. Furthermore, the residuals from the fitted B

8
ARMA models give information about the dynamics after

10
memory has been removed. We fit ARMA models to parent
time series, with the number of AR (p) and MA (q) parame-
Waiting Time (years)
6
10
ters being chosen by AIC, subject to the constraint that
1  p  3 and q < p (Supporting Information: ARMA models).
If the set of block maxima for a time series is fat-tailed and
4
10

that for the ARMA residuals is not, this indicates that

between-sample dynamics are responsible for generating the
fat tail. Conversely, if both the time series and the ARMA
2
10

residuals are fat-tailed, then within-sample dynamics are

responsible at least in part for the generating the fat tail
(Ives et al. 2003, 2010). ARMA is only an approximation of
0
10

these between-sample processes, and could fail in some

cases. Nonetheless, this approach can be informative and Bio (283) Chem (175) Phys (48) Met (9)
provides a common analysis for comparisons across our large
set of diverse time series. Fig. 2. (A) Bean plots of tailedness (n) estimated from the GEV, and
(B) waiting times to the next record-breaking event estimated from the
GEV fits. Waiting times are the time until the maximum observation in a
Results time series is broken by at least 10%. Each “bean” consists of a mirror
Frequency of extremes image of the density plot of n or of the waiting times in each category
We found a wide range of tailedness (n) in each variable (red 5 biological, yellow 5 chemical, green 5 physical, blue 5 meteoro-
category, but on average the biological variables had higher logical). The values for individual time series are indicated by the small
black dashes inside each bean, and the number of time series in each
values of n than other categories of variables (Fig. 2A; Sup- category is in parentheses. Horizontal dotted line is the mean across all
porting Information Table S1), even after controlling for categories. Solid horizontal lines are the means within each category.
length of time series and the identity of the lake (pairwise
comparisons with biological time series: chemical, z 5 210.1, To further compare the different categories while account-
p  0.0001; physical, z 5 211.7, p  0.0001; meteorological, ing for uncertainty in the estimates of n, we tested whether
z 5 22.1, p 5 0.0234 (Supporting Information Table S2). For the time series were fat-tailed by, for each time series, com-
waiting times to break a current record, the pattern across paring the null hypothesis that n 5 0 against the alternative
categories was similar to that for n: biological variables were n > 0 for positive estimates or n < 0 for negative estimates
likely to break their current records sooner on average than (1-tailed test with a 5 0.05, not correcting for multiple com-
other categories of variables (Fig. 2B). These results indicate parisons). We present results that are not corrected for multi-
that the biological variables tended to produce extremes ple comparisons, and subsequently those with p-values that
more frequently than other categories of ecosystem are corrected to maintain the false discovery rate (Benjamini
variables. and Hochberg 1995). Of the 283 biological time series, 38%

66
 23782242, 2017, 3, Downloaded from https://aslopubs.onlinelibrary.wiley.com/doi/10.1002/lol2.10037 by National Sun-Yat-Sen, Wiley Online Library on [13/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Batt et al. Extreme events in lake ecosystem time series

ξ
ξ

ξ λ
Fig. 3. Scatter plot of time series n vs. ARMA residual n. Density plots Fig. 4. Scatter plot of time series n vs. the dominant eigenvalue (jjkjj)
in the margins reflect the distribution of the parallel axis. The color of of the autoregressive parameters from the ARMA fits. Marginal plots and
the points in the scatter plot is proportional to the number of points in colors are as in Fig. 3.
that coordinate region, with yellow colors in regions overlain with many
points.
calculated for the ARMA residuals (slope 5 0.40, p  0.0001;
had an estimate of n statistically significantly greater than 0 Supporting Information Fig. S2, Fig. 3). Moreover, estimates
(corrected 5 12%), by far the highest frequency of statisti- of n for the times series were negatively related with the
cally significant fat-tailedness among categories (13% for eigenvalues (slope 5 20.23, p  0.0001; Fig. 4). These corre-
chemical, 3.8% for physical, 0% for meteorological; lations are consistent with the inference that extreme events
corrected 5 1.8%, 0%, and 0%, respectively). Using the same were generated primarily by processes that occur within a
procedure to investigate distributions with bounded upper time step.
limits (n < 0), only 0.7% of biological time series had statisti-
cally significant estimates of n < 0 (corrected 5 0.4%), which Discussion
was the lowest frequency among variable categories (22% for
Biological time series had distributions with fatter tails
chemical, 51% for physical, 42% for meteorological;
than time series of other ecosystem variables. As tailedness
corrected 5 8.2%, 25%, 25%, respectively). Among the bio-
(n) increases, the tail probabilities become large or “fat” rela-
logical variables, the fish, zooplankton and chlorophyll time
tive to distributions like the normal, implying that extreme
series with the largest estimates of n were, respectively, Lepo-
events will be more common. A fatter tail also decreases the
mis spp. in Lake Mendota (n 5 1.2; 90% CI: 0.7 < n < 1.8), Kel-
time that passes until an event of arbitrarily large size is
licottia spp. in Crystal Bog (n 5 1.2; 90% CI: 0.7 < n < 1.7),
observed, increases the largest value observed after an arbi-
and chlorophyll in Crystal Lake (n 5 0.5; 90% CI:
trary amount of time, and makes the chances of observing
0.2 < n < 0.8).
“massive” and “big” events more similar. Thus, in our analy-
ARMA dynamics ses biological variables have larger and more frequent
If extreme events were generated largely by long-term extremes. As values of n increase for a distribution, the statis-
dynamics created by correlations across multiple time steps tical moments of the distribution  1/n become undefined
in the parent time series, then estimates of n for the original (Coles 2001); for example, at n  1 the mean (first moment)
time series should be uncorrelated with estimates of n for the is infinite, and at n  0.5 the variance (second moment) is
ARMA residuals. Additionally, the magnitude of correlation infinite. By contrast, biological variables are often character-
across time steps, measured by the eigenvalue of the ARMA ized by distributions that have defined means and variances
process, should be positively correlated with the tailedness. (e.g., lognormal). However, the estimates of n for some of
Both of these expectations were false. Estimates of n for the the fish and zooplankton time series were greater than 1,
time series were positively related with the estimates of n implying that their annual maxima will not converge to a

67
 23782242, 2017, 3, Downloaded from https://aslopubs.onlinelibrary.wiley.com/doi/10.1002/lol2.10037 by National Sun-Yat-Sen, Wiley Online Library on [13/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Batt et al. Extreme events in lake ecosystem time series

long-term mean as more data are collected. Therefore, even time series may serve as an explanation for the relative pau-
distributions like the lognormal would underestimate the city of empirical assessments of biological tailedness—time
frequency and intensity of extremes in these fat-tailed bio- series like those used in this study have only recently
logical variables. This result underscores the challenge that matured to a length suitable for extreme value analysis,
fat-tailed distributions pose for management: prevalent whereas long environmental time series are more common.
extreme events can make it impossible statistically to charac- Therefore, continuing long-term monitoring of ecological
terize even the mean abundance of a species. variables and characterizing the tailedness of additional time
Compared to meteorological, physical, and chemical time series are critical for understanding ecological extremes.
series, the biological time series we analyzed had larger val- Because fat tails dictate the frequency and intensity of
ues for tailedness (n) and shorter expected waiting times massive events produced by an ecosystem, biological fat tails
until record-breaking events, even though the variables were stand to greatly impact the management of populations that
measured in the same ecosystems. This result provides evi- are critical to human well-being, such as fish stocks, invasive
dence to suggest that biological variables should be expected species, pests and diseases, or phytoplankton that form
to produce larger and more frequent surprises than the other harmful blooms. Although there are several classes of general
categories of variables. Overall, these analyses provide evi- causal mechanisms for fat tails (literature cited above), more
dence that biological variables are more likely to be fat-tailed research is needed to identify causes of fat tails for specific
than other types of variables in the same ecosystems. populations. Our study focused on 11 lakes, but similar anal-
Mechanisms yses of additional ecosystems are needed to test the general-
Mechanisms for promoting or diminishing fat tails are ity of the patterns found here. Tests in additional ecosystems
common in biological processes (Sornette 2006). Negative are also crucial for determining whether certain taxa consis-
density dependence can lead to thinner tails and is common tently have higher tailedness, or if certain ecosystem charac-
in population dynamics (May 1976; Ziebarth et al. 2010), teristics can affect the tailedness of its denizens. Such
giving reason to expect less fat-tailedness in biological varia- patterns could help make predictions of extremes for ecosys-
bles relative to the environmental variables that drive them. tems without long-term records, and would provide further
However, nonlinear models fit to time series of disease (Sugi- insight into the causes of biological tailedness, its impacts
hara et al. 1990), flour beetles (Dennis et al. 2001), spruce on ecosystem functioning, and its implications for human
budworms (Ludwig et al. 1978), and phytoplankton (Beninca  use and management of ecosystems.
et al. 2008) can produce fat tails under certain parameter It is dangerous to consider the future as a set of norms
combinations, even deterministically. Nonlinear processes from the past (Oppenheimer et al. 2007; Burgman et al.
are common in biology (Hsieh et al. 2005), suggesting that 2012), and forecasting future events, especially extreme
they may produce fat tails in biology just as they are known events, is difficult. Our empirical results indicate that biolog-
to do in other settings (Baraba si 2005; Newman 2005). ical time series present frequent surprises. If this pattern was
Our ARMA analysis suggests that mechanisms within a caused by directional environmental change in climate,
time step ( 1 year) often generate observed fat tails, and nutrient inputs, habitat loss, or other factors, then extremes
that increasingly long memory cannot explain the observed may become more common in ecosystem time series.
fat tails. A simple process for generating such fat tails is for a
population to amplify the variability of its environment. References
Multiplicative birth and survival processes influenced by sto- Barabasi, A.-L. 2005. The origin of bursts and heavy tails in
chastic environmental drivers are one way that fat tails could human dynamics. Nature 435: 207–211. doi:10.1038/
be generated within time steps (Fig. 1 and Supporting Infor- nature03459
mation: Extremes in a model population). Comparing alter- Beninca , E., J. Huisman, R. Heerkloss, K. D. Jo € hnk, P.
native mechanisms using models tailored for specific Branco, E. H. Van Nes, M. Scheffer, and S. P. Ellner. 2008.
biological systems is needed if we are to better understand Chaos in a long-term experiment with a plankton com-
how fat tails are generated.
munity. Nature 451: 822–825. doi:10.1038/nature06512
Concluding remarks Benjamini, Y., and Y. Hochberg. 1995. Controlling the false
Our analysis included all of the variables typically col- discovery rate: A practical and powerful approach to mul-
lected by limnologists (Magnuson et al. 2006), and because tiple testing. J. R. Stat. Soc. Ser. B 57: 289–300.
our results might have differed if we had investigated differ- Burgman, M., J. Franklin, K. R. Hayes, G. R. Hosack, G. W.
ent subsets of the variables, we analyzed all of the data that Peters, and S. A. Sisson. 2012. Modeling extreme risks in
were available in a comprehensive analysis. Estimating a sta- ecology. Risk Anal. 32: 1956–1966. doi:10.1111/j.1539-
tistic based on rare events requires long time series (Palutikof 6924.2012.01871.x
et al. 1999; Katz et al. 2005), and estimates of n will improve Coles, S. 2001. An introduction to statistical modeling of
with continued data collection. In fact, this need for long extreme values. Springer.

68
 23782242, 2017, 3, Downloaded from https://aslopubs.onlinelibrary.wiley.com/doi/10.1002/lol2.10037 by National Sun-Yat-Sen, Wiley Online Library on [13/03/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Batt et al. Extreme events in lake ecosystem time series

Dennis, B., R. A. Desharnais, J. M. Cushing, S. M. Henson, Newman, M. 2005. Power laws, Pareto distributions and
and R. F. Costantino. 2001. Estimating chaos and com- Zipf’s law. Contemp. Phys. 46: 323–351. doi:10.1080/
plex dynamics in an insect population. Ecol. Monogr. 71: 00107510500052444
277–303. doi:10.1890/0012-9615(2001)071[0277:ECAC- Oppenheimer, M., B. C. O’Neill, M. Webster, and S.
DI]2.0.CO;2 Agrawala. 2007. The limits of consensus. Science 317:
Denny, M. W., L. J. H. Hunt, L. P. Miller, and C. D. G. 1505–1506. doi:10.1126/science.1144831
Harley. 2009. On the prediction of extreme ecological Palutikof, J. P., B. B. Brabson, D. H. Lister, and S. T. Adcock.
events. Ecol. Monogr. 79: 397–421. doi:10.1890/08-0579.1 1999. A review of methods to calculate extreme wind
Fey, S. B., and others. 2014. Recent shifts in the occurrence, speeds. Meteorol. Appl. 6: 119–132. doi:10.1017/
cause, and magnitude of animal mass mortality events. S1350482799001103
Proc. Natl. Acad. Sci. USA 112: 1083–1088. doi:10.1073/ Peters, D. P. C., R. A. Pielke, B. T. Bestelmeyer, C. D. Allen,
pnas.1414894112 S. Munson-McGee, and K. M. Havstad. 2004. Cross-scale
Gaines, S. D., and M. W. Denny. 1993. The largest, smallest, interactions, nonlinearities, and forecasting catastrophic
highest, lowest, longest, and shortest: Extremes in ecol- events. Proc. Natl. Acad. Sci. USA 101: 15130–15135. doi:
ogy. Ecology 74: 1677–1692. doi:10.2307/1939926 10.1073/pnas.0403822101
Gurney, W. S. C., and R. Nisbet. 1998. Ecological dynamics. Pielou, E. C. 1977. Mathematical ecology, 2nd. John Wiley
Oxford Univ. Press. & Sons.
Hsieh, C., S. M. Glaser, A. J. Lucas, and G. Sugihara. 2005. R Core Team. 2014. R: A language and environment for statis-
Distinguishing random environmental fluctuations from tical computing. R Foundation for Statistical Computing.
ecological catastrophes for the North Pacific Ocean. Schmitt, F. G., J. C. Molinero, and S. Z. Brizard. 2008. Non-
Nature 435: 336–340. doi:10.1038/nature03553 linear dynamics and intermittency in a long-term cope-
IPCC. 2012. Changes in climate extremes and their impacts
pod time series. Commun. Nonlinear Sci. Numer. Simul.
on the natural physical environment. In C. B. Field and
13: 407–415. doi:10.1016/j.cnsns.2006.04.005
others [eds.], Managing the risks of extreme events and
Segura, A. M., D. Calliari, H. Fort, and B. L. Lan. 2013. Fat
disasters to advance climate change adaptation. Cam-
tails in marine microbial population fluctuations. Oikos
bridge Univ. Press.
122: 1739–1745. doi:10.1111/j.1600-0706.2013.00493.x
Ives, A. R., B. Dennis, K. L. Cottingham, and S. R. Carpenter.
Sornette, D. 2006. Critical phenomena in natural sciences:
2003. Estimating community stability and ecological inter-
Chaos, fractals, selforganization and disorder: Concepts
actions from time-series data. Ecol. Monogr. 73: 301–330.
and tools. Taylor & Francis US.
doi:10.1890/0012-9615(2003)073[0301:ECSAEI] 2.0.CO;2
Sugihara, G., B. T. Grenfell, and R. M. May. 1990. Distin-
Ives, A. R., K. C. Abbott, and N. L. Ziebarth. 2010. Analysis
guishing error from chaos in ecological time series. Philos.
of ecological time series with ARMA(p,q) models. Ecology
Trans. Biol. Sci. 330: 235–251. doi:10.1098/rstb.1990.0195
91: 858–871. doi:10.1890/09-0442.1
Villarini, G., and J. A. Smith. 2010. Flood peak distributions
Katz, R. W., M. B. Parlange, and P. Naveau. 2002. Statistics
for the eastern United States. Water Resour. Res. 46: 1–17.
of extremes in hydrology. Adv. Water Resour. 25: 1287–
doi:10.1029/2009WR008395
1304. doi:10.1016/S0309-1708(02)00056-8
Ziebarth, N. L., K. C. Abbott, and A. R. Ives. 2010. Weak
Katz, R. W., G. S. Brush, and M. B. Parlange. 2005. Statistics
of extremes: Modeling ecological disturbances. Ecology population regulation in ecological time series. Ecol. Lett.
86: 1124–1134. doi:10.1890/04-0606 13: 21–31. doi:10.1111/j.1461-0248.2009.01393.x
Ludwig, D., D. Jones, and C. Holling. 1978. Qualitative anal-
ysis of insect outbreak systems: The spruce budworm and Acknowledgments
forest. J. Anim. Ecol. 47: 315–332. doi:10.2307/3939 We thank M. Pinsky, J. Gaeta, C. Herren, J. Kurtzweil, A. Latzka, J. Mag-
Magnuson, J. J., T. K. Kratz, and B. J. Benson [eds.]. 2006. nuson, S. Oliver, M. Turner, L. Winslow, and students in Zoology 955/6
Long-term dynamics of lakes in the landscape: Long-term (Spring 2013) and 995 (Spring 2014) at UW–Madison for discussions
ecological research on North Temperate lakes. Oxford and comments on the manuscript. We thank the North Temperate Lakes
Long-Term Ecological Research program for providing the data. RDB
Univ. Press.
was supported by a postdoctoral fellowship from the Institute of Marine
May, R. M. 1976. Simple mathematical models with very and Coastal Sciences at Rutgers University. Additional financial support
complicated dynamics. Nature 261: 459–467. doi: was provided by NSF grants DEB-1440297 (NTL-LTER), DEB-1144683
10.1038/261459a0 (SRC), and DEB-Dimensions-1240804 (ARI).
Mitchell, J. F. B., J. Lowe, R. A. Wood, and M. Vellinga.
2006. Extreme events due to human-induced climate Submitted 29 August 2016
change. Philos. Trans. R. Soc. A 364: 2117–2133. doi: Revised 20 December 2016
10.1098/rsta.2006.1816 Accepted 19 February 2017

69