Ecological Modelling 220 (2009) 11971202

Contents lists available at ScienceDirect

Ecological Modelling
journal homepage: www.elsevier.com/locate/ecolmodel

Population viability analysis for several populations using multivariate state-space models
Richard A. Hinrichsen
Hinrichsen Environmental Services, 901 NE 43rd ST #101, Seattle, WA 98105, USA

a r t i c l e

i n f o

a b s t r a c t
The International Union for the Conservation of Nature and Natural Resources (IUCN), the worlds largest and most important global conservation network, has listed approximately 16,000 species worldwide as threatened. The most important tool for recognizing and listing species as threatened is population viability analysis (PVA), which estimates the probability of extinction of a population or species over a specied time horizon. The most common PVA approach is to apply it to single time series of population abundance. This approach to population viability analysis ignores covariability of local populations. Covariability can be important because high synchrony of local populations reduces the effective number of local populations and leads to greater extinction risk. Needed is a way of extending PVA to model correlation structure among multiple local populations. Multivariate state-space modeling is applied to this problem and alternative estimation methods are compared. The multivariate state-space technique is applied to endangered populations of pacic salmon, USA. Simulations demonstrated that the correlation structure can strongly inuence population viability and is best estimated using restricted maximum likelihood instead of maximum likelihood. 2009 Elsevier B.V. All rights reserved.

Article history: Received 22 February 2008 Received in revised form 7 February 2009 Accepted 11 February 2009 Available online 16 March 2009 Keywords: IUCN Red List Population viability analysis (PVA) Stochastic growth rate Multivariate State-space models Restricted maximum likelihood Measurement error Salmon Covariance

1. Introduction The International Union for the Conservation of Nature and Natural Resources (IUCN), the worlds largest and most important global conservation network, has listed approximately 16,000 species worldwide as threatened (IUCN, 2007). Placing a species on this list, known as the Red List, brings threatened species special attention which can lead to the development of specic management plans to reduce the likelihood of extinction. The most important tool for recognizing and listing species as threatened is population viability analysis (PVA), which estimates the probability of extinction of a population or species over a specied time horizon. The most common PVA approach is to apply it to a single time series of population abundance (e.g., Dennis et al., 1991). This approach ignores correlation structure which arises when a population actually consists of several local, possibly interbreeding, populations. Correlation structure of local populations actually has important implications for the viability of a species and should not be ignored. High population synchrony reduces the effective number of local populations, leading to greater extinction risk (Harrison and Quinn, 1989; Hill et al., 2002; Hilderbrand, 2003). Most PVA

Tel.: +1 206 715 2859. E-mail address: hinrich@seanet.com. 0304-3800/$ see front matter 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.ecolmodel.2009.02.014

approaches, however, do not model correlation structure explicitly, and either lump the populations together, or run PVAs separately on each related population. In this paper, I analyze alternative methods to estimate correlation structure, while allowing the possibility of measurement error in population counts. Another important consideration for PVA is measurement error in population counts. Measurement error, which arises because of population counting or sampling error, if ignored, can create bias in population viability analysis results (Holmes, 2001; McNamara and Harding, 2004). Several methods for treating measurement error include using a robust estimator that does not rely on information about the measurement error distribution (Holmes, 2001), or, applying a maximum likelihood (Lindley, 2003) or restricted maximum likelihood (Staples et al., 2004) to a state-space model. I extend the approach of Staples et al. (2004), originally developed to handle single populations, into multivariate framework, which models correlation structure of several local populations. The main objective of this work is to develop a multivariate model to handle multiple local populations at once and make inferences on PVA model structures (e.g., population growth rate and covariance structure). I begin by demonstrating the importance of considering correlation structure in population viability using simulations. The state-space method is then applied in another simulation study that demonstrates the accuracy of estimating correlation structures and population growth rate differences between local populations. Three alternative estimation techniques for the

1198

R.A. Hinrichsen / Ecological Modelling 220 (2009) 11971202

multivariate PVA model are developed and compared: a multivariate Dennis approach (MVD), which ignores measurement error; a maximum likelihood (ML), and restricted maximum likelihood (REML). These estimation techniques are applied to two groups of local populations of chinook salmon (Oncorhynchus tshawytscha): one group in the Snake River basin, USA, and the other in the Lake Washington basin, USA. 2. Methodology 2.1. State-space models When detailed models of population structure are not possible because only population counts are available, the most common approach is to use a diffusion approximation approach (Holmes et al., 2007). The typical diffusion approximation approach, often called the Dennis Model, for the growth of a single population is a stochastic, discrete-time model of exponential growth that is described by Nt = Nt1 exp ( + et ), (1)

unequal diagonal entries; a diagonal covariance matrix with equal variances; and a matrix with equal variances and equal covariances. The unrestricted covariance matrix contains no restrictions on its elements, such as common variances, covariances, or zeroes, while the other forms do contain such restrictions. These four alternative covariance matrix forms were also used for the measurement error covariance matrix H To determine which state-space model was best supported by the data, AIC was used (Akaike, 1973). 2.3. Estimation of state-space models To determine which approach is best for estimating alternative stochastic growth rate and covariance structures, I evaluated three alternative estimation procedures: ML, MVD, and restricted maximum likelihood (REML). All of the methods work by forming likelihood functions based on differences of the observed log abundances, which have multivariate normal distributions. The rst differences series of log-transformed counts can be written w t = y t+1 y t = c + t+1 + t+1 t . (5)

where t = 1, 2,. . .,T; Nt represents population abundance at time t; and exp( ) is the deterministic per-unit-abundance growth rate in population density (Dennis et al., 1991). The error term, et , represents the stochastic deviations from long-term trend due to environmental effects. These errors are assumed to have a normal distribution with mean zero and constant variance. Dening xt = log(Nt ), the growth process can be expressed as xt = xt1 + + et , (2)

Now form the n(T 1) 1 vector W by stacking all of the wt vectors into a single column. The distribution of W is multivariate normal with mean Ec and variancecovariance matrix , where E is a n(T 1) n matrix formed by stacking (T 1) copies of the n n identity. Using the fact that E[wt c)(wt c) ] = 2H + Q and E[wt+1 c)(wt c) ] = H, we can express the variancecovariance matrix for W as

which represents a discrete-time random walk with drift process. This approach, although useful for single populations, must be extended to model several local populations simultaneously. Suppose that instead of a single population, there are n related local populations to model, which may have correlated growth rates. Further suppose that xt is observed with error. The Dennis et al. (1991) model can be extended to this situation through a multivariate state-space approach (Dennis et al., 1998). The true (but unknown) log population abundances, denoted by the state variable vector t , are governed by the dynamic equation t = t1 + c + t , (3)

H 0 0 = . . .
0

2H + Q

H 2H + Q H 0

0 H 2H + Q H 0

0 0 H . . .

. . 0 . . . .

. .

. .

0 . .

. . . . 0

. . . . H

. . 0 H 2H + Q

.
(6)

where c is a n 1 vector that represents the stochastic growth rates, and t is a n 1 vector of serially uncorrelated disturbances with mean zero and n n covariance matrix Q. The observed values of log abundances, yt are related to the true log abundances, t , through the measurement equation y t = t + t , (4)

Notice that the entries in the variancecovariance matrix consist entirely of the parameters contained in the process error and measurement error matrices. The maximum likelihood approach, ML, proceeds by maximizing the likelihood formed by the distribution of W with respect to the elements of c, Q, and H. This log likelihood function is given by LW = 1 1 1 log |2 | W W . 2 2 (7)

for t = 1, 2,. . .,T, where t is a n 1 vector of serially uncorrelated disturbances with mean zero and n n covariance matrix H. 2.2. Alternative model structures Covariance and growth rate structures are inferred by tting alternative forms of the state-space model to population data, then determining which is best supported by the data. This represents a distinct advantage over univariate approaches which do not allow for relationships among the viability parameters of local populations. Several alternative forms of the stochastic growth rate and variance structures are possible. In this paper, the stochastic growth rate structures I explore are: population-specic stochastic growth rates, and a stochastic growth rate common to all local populations. The four alternative forms for the process error, Q, I explore are: an unrestricted covariance matrix; a diagonal covariance matrix with

The multivariate Dennis approach (MVD) proceeds by maximizing this likelihood after xing H = 0. I call this the MVD because when applied to a single population, it yields precisely the Dennis model estimates. Staples et al. (2004) observed that, in the univariate case, maximum likelihood (ML) estimators obtained by maximizing LW directly were strongly biased. They suggested using the restricted maximum likelihood is used to overcome this bias (Searle et al., 1992). In order to use REML for the estimation problem in this paper, I generalized the univariate approach of Staples et al. (2004) to a multivariate framework. The REML proceeds by using another difference of the log count data: j t = w t+1 w t , (8)

where jt is a n 1 vector and t = 1, 2,. . .,T2. Next construct the n(T 2) 1 vector J by stacking all of the vectors jt into a single

R.A. Hinrichsen / Ecological Modelling 220 (2009) 11971202

1199

column. The distribution of J is multivariate normal with common mean zero and variancecovariance matrix J = U U , where U is the n(T 2) n(T 1) matrix formed by placing n n identity matrix on the main diagonal and the negative of the n n identity matrix on the diagonal above the main diagonal. Notice that the likelihood function formed by the distribution of J does not contain the vector c. This occurred because successive differences of the log count data removed c. The REML estimates of the variancecovariance matrices H and Q are obtained by maximizing the log likelihood formed by the distribution J, which is Lj = 1 1 1 log |2 J | J J J. 2 2 (9)

algorithm implemented in the R statistical package (Nelder and Mead, 1965). The state-space methods were applied to counts of spawning salmon from two Snake River stream-type chinook (O. tshawytscha) populations (19802001) listed in 1992 under the U.S. Endangered Species Act: Bear Valley Creek/Elk Creek and Big Creek. I also applied the methods to two salmon populations in the Lake Washington basin (19812004), the Cedar River and North Lake Washington chinook salmon populations, listed in 1999. I applied alternative state-space models to these data and evaluated each in terms of its t to the data using AIC. 3. Results Extinction probabilities rose sharply as synchronicity between populations increased (Fig. 1.) As process error correlation increased from 0 to 1, the extinction probability increased by about 0.20, regardless of the assumptions about stochastic growth rates. This suggests that ignoring correlation in a population viability analysis of multiple populations will underestimate extinction risk. The simulations showed that the three methods yielded quite similar bias and precision of the difference in stochastic growth rate (Fig. 2). In contrast, the accuracy of the process error correlation differed greatly among the three estimation techniques (Fig. 3). The REML method was by far the most accurate in terms of bias and dispersion of the process error correlation estimate. Both the ML and REML methods showed positive bias in the process error correlation, while the MVD estimate showed a negative bias. Bias increased with measurement error for all methods. Out of the alternative models considered for the two Snake River salmon populations, the best (the one with lowest AIC) had a common stochastic growth rate, population-specic process error variances, highly correlated population uctuations, and uncorrelated measurement errors with equal variances (Table 1). The best model for the Lake Washington salmon populations also had a common stochastic growth rate, but in contrast to the Snake River stocks, showed no correlation in process errors. The ML and REML parameter estimates were similar and the model with lowest AIC

To obtain the stochastic growth rate estimates, the REML estimates of H and Q are placed in the likelihood function LW , which is then maximized with respect to c. (Note that the likelihood function LW contains the stochastic growth rates, c, while the likelihood function LJ does not.) This yields the stochastic growth rate estimate = (X V1 X)
1

X V1 W ,

(10)

where V is the matrix formed by replacing the variance matrices H and Q by their estimates in the matrix . The matrix X is a vector of 1s in the case where there is a common stochastic growth rate. When there are n separate stochastic growth rates, X is the matrix formed by stacking T 1 copies of the n n identity matrix. The estimate of the variance of is given by var ( ) = (X V 1 X)
1

(11) is unknown and the matrix V is used in

which is biased because its place.

2.4. Simulation and application To illustrate the importance of synchrony in population viability analysis of several local populations, extinction probabilities were graphed against process error correlation between two hypothetical populations. The process error variances for the two populations were 1.0 and 1.5, respectively. The initial number of spawning adults in each population was 5. Three different stochastic growth rate pairs were used for the two populations: 1 = 0 and 2 = 0.05, 1 = 0 and 2 = 0, and 1 = 0.05 and 2 = 0.05. Extinction probability was dened as the probability that there was less than one spawning adult returning to the spawning grounds for 5 consecutive years over 100 years. Process error correlation was varied from 0 to 1.0. For each level of process error correlation, and each alternative pair of stochastic growth rates, a single extinction probability was calculated by simulating 2000 population trajectories. The trajectories were formed by using the state Eq. (3) prospectively. Extinction probability was calculated as the number of the trajectories that resulted in extinction divided by 2000. The accuracy of alternative estimates of stochastic growth rate differences and process error correlations were evaluated using simulations of two hypothetical related populations. The stochastic growth rates were 1 = 0 and 2 = 0.05 for the rst and second populations, respectively. The process errors variances were 1.0 and 1.5, respectively, and the process error correlation was set at 0.5. Three levels of common measurement error variance were used: 0.2 (small), 1.0 (medium), 1.5 (large). I generated 1000 synthetic time series of length 50 and used the multivariate Dennis, maximum likelihood, and restricted maximum likelihood estimation techniques on each. To assess the accuracy of the alternative estimation approaches, boxplots of the stochastic growth rate differences and process error correlations were used. R code for the simulations and tting procedures is available as supplemental material. The likelihoods were maximized using the NelderMead

Fig. 1. The probability of extinction of two combined populations was plotted as a function of the process error correlation coefcient for three different stochastic growth rate pairs: (a) 1 = 0 and 2 = 0.05 (diamonds), (b) 1 = 0 and 2 = 0 (squares), and 1 = 0.05 and 2 = 0.05 (triangles). The process error variances for the two populations were 1.0 and 1.5, respectively. The initial number of spawning adults in each population was 5. Extinction probability was dened as the probability that there was less than one spawning adult 5 consecutive years over 100 years.

1200

R.A. Hinrichsen / Ecological Modelling 220 (2009) 11971202

Fig. 2. Comparison of the multivariate Dennis, maximum likelihood, and restricted maximum likelihood estimates of the stochastic growth rate difference between two local populations when measurement error variance is (a) 0.20, (b) 1.0, and (c) 1.5. There were 1000 simulations of time series of length 50 with a true stochastic growth rate difference of 0.05. The dashed horizontal line represents the true stochastic growth rate difference. In the boxplots, the solid horizontal lie within the box indicates the median; the box encompasses 75% of the 1000 estimates, and the whiskers are drawn to the nearest value not beyond 1.5 (interquartile range) from the quartiles. Open circles represent extreme values.

Fig. 3. Comparison of the multivariate Dennis, maximum likelihood, and restricted maximum likelihood estimates of process error correlation between two local populations when measurement error variance is (a) 0.20, (b) 1.0, and (c) 1.5. There were 1000 simulations of time series of length 50 with a true process error correlation of 0.5. The dashed horizontal line represents the true correlation. The boxplots are as dened in Fig. 2.

R.A. Hinrichsen / Ecological Modelling 220 (2009) 11971202

1201

Table 1 Parameter estimates using the multivariate Dennis (MVD), maximum likelihood (ML), and restricted maximum likelihood (REML), estimation methods for chinook salmon population groups. These estimates are based on models with lowest AIC score. Population group and analysis method Snake River chinook salmon (19802001) MVD ML REML Population Trend (1 SE) Process variance 1.6834 1.6834 0.9076 1.476 0.9638 1.5608 Process correlation 0.6921 1 1 0.1854 0.1854 0.1933 0.1933 0 0 Measurement variance Measurement correlation

Bear Valley/Elk Creek Big Creek Bear Valley/Elk Creek Big Creek Bear Valley/Elk Creek Big Creek

0.1526 (0.2604) 0.1526 (0.2604) 0.1558 (0.1065) 0.1558 (0.1065) 0.1571 (0.0827) 0.1571 (0.0827)

Lake Washington chinook salmon (19812004) MVD North Lake Washington Cedar River ML REML North Lake Washington Cedar River North Lake Washington Cedar River

0.0016 (0.1395) 0.0016 (0.1395) 0.0244 (0.0426) 0.0244 (0.0426) 0.0148 (0.0515) 0.0148 (0.0515)

0.8717 0.4618 0.4755 0.0262 0.5319 0.0595

0.5781 0 0 0.2051 0.2051 0.1834 0.1834 0.8847 0.9778

was the same for both estimation methods. The multivariate Dennis method, which ignored measurement error and placed all variance in the process error term, yielded relatively large standard errors for the stochastic growth rate estimate. Greater accuracy was provided by using either the maximum likelihood or restricted maximum likelihood estimates of stochastic growth rate. 4. Discussion A multivariate state-space approach can reveal important relationships among local populations that cannot be detected by a univariate approach. Specically, it allows one to objectively choose between alternative stochastic growth rate and covariance structures. The Snake River salmon application, for example, revealed synchronous uctuations in growth and a common stochastic growth rate. This implied that the two populations tended to act as a single population. Understanding covariance structure is useful for making inferences about extinction risk. Goodman (1987) highlighted the importance of independently uctuating local populations for persistence. In the multivariate state-space framework, the degree of independence is easily characterized by process error correlations. The usual viability parameters of local population growth rate and variances are also estimated in the multivariate approach, but with the added feature of allowing common parameters. This serves to strengthen inferences because extra parameters come at the cost of reduced statistical accuracy. When population-specic stochastic growth rates and variances are justied by the data, these can be used to characterize extinction risk of individual local populations and help focus recovery effort. With limited resources, for example, it may be best to focus recovery effort on maintaining a local population that tended to vary independently of the others and showed a negative growth rate. Maintaining this population might be a priority because its loss would reduce the effective number of populations more than the loss of populations more synchronous with the remainder of the related populations. There are some future directions of research that would make multivariate state-space modeling more useful. Populations can sometimes undergo large shifts in population trend, so that the underlying stochastic process is nonstationary (Coulson et al., 2001). Such shifts can have a profound effect on the population dynamics and on estimates of extinction probability. For example, recent research suggests that large shifts in salmon dynamics associated with climate regime shifts (Peterson and Schwing, 2003). Lindley (2003) showed how such a shift could be modeled with

a univariate state-space model that used a step in stochastic growth rate. This and other types of shifts in the stochastic growth rates can also be modeled and tested in the multivariate framework. Important is careful consideration of time series lengths needed to estimate the multivariate model parameters. Required time series lengths will depend on where inferences will be focused. If inferences are focused on stochastic growth rates, then the usual rule of thumb of 20 years of estimating a mean may be sufcient. If extinction risk is the focus, there will be greater data requirements. Fieberg and Ellner (2000) suggested that reliable estimates of extinction risk required time series to be about 510 times longer than the time horizon used in the risk calculation. When parameters describing measurement error variance and the degree of synchrony are added to the extinction risk modeling, data requirements will likely increase, because more parameters will need to be estimated. Multivariate state-space modeling has a long, rich history in the scientic literature, and remains an active area of research. Its ability to model several populations simultaneously and its exibility in allowing many different forms of the variance matrices and state process make it a valuable tool for inferring covariance and growth rate structure from population counts. From a statistical point of view, the multivariate approach represents a way to bring more information to bear upon the estimation of PVA parameters in a situation where observations are often poor. By extending PVA to multiple local populations when only simple population counts are available, PVA will become more accurate and this will improve assessments of the thousands of species worldwide that are currently on the IUCN Red List. Acknowledgments This study was funded by the U.S. Department of Energy, Bonneville Power Administration Contract No.00004031. I thank Eli Holmes, Steve Lindley, Charlie Paulsen, and Rich Zabel for their valuable reviews of the manuscript. Thanks also to Charlie Petrosky and Hans Berge for supplying the salmon population data used in the analysis. Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.ecolmodel.2009.02.014.

1202

R.A. Hinrichsen / Ecological Modelling 220 (2009) 11971202 Hill, M.F., Hastings, A., Botsford, L.W., 2002. The effects of small dispersal rates on extinction times in structured metapopulation models. Am. Nat. 160, 389 402. Holmes, E.E., 2001. Estimation risks in declining populations with poor data. Proc. Natl. Acad. Sci. U.S.A. 98, 50725077. Holmes, E.E., Sabo, J.L., Viscido, S.V., Fagan, W., 2007. A statistical approach to quasiextinction forecasting. Ecol. Lett. 10, 11821198. IUCN, 2007. 2007 IUCN Red List Threatened Species. http://www.iucnredlist.org. Lindley, S.T., 2003. Estimation of population growth and extinction parameters from noisy data. Ecol. Appl. 13, 806813. McNamara, J.M., Harding, K.C., 2004. Measurement error and estimates of population extinction risk. Ecol. Lett. 7, 1620. Nelder, J.A., Mead, R., 1965. A simplex algorithm for function minimization. Comput. J. 7, 308313. Peterson, W.T., Schwing, F.B., 2003. A new climate regime in northeast pacic ecosystems. Geophys. Res. Lett. 30, 1896, doi:1810.1029/2003GL017528. Searle, S.R., Casella, G., McCulloch, C.E., 1992. Variance Components. John Wiley, New York, NY, USA. Staples, D.F., Taper, M.L., Dennis, B., 2004. Estimating population trend and process variation for PVA in the presence of sampling error. Ecology 85, 923929.

References
Akaike, H., 1973. Information theory and an extension of the maximum likelihood principle. In: Petrov, B.N., Cski, F. (Eds.), Second International Symposium on Information Theory. Akadmia Kiad, Budapest, Hungary, pp. 267281. Coulson, T., Mace, G.M., Hudson, E., Possingham, H., 2001. The use and abuse of population viability analysis. Trends Ecol. Evol. 16, 219221. Dennis, B., Munholland, P.L., Scott, J.M., 1991. Estimation of growth and extinction parameters for endangered species. Ecol. Monogr. 61, 115143. Dennis, B., Kemp, W.P., Taper, M.L., 1998. Joint density dependence. Ecology 79, 426441. Fieberg, J., Ellner, S.P., 2000. When is it meaningful to estimate an extinction probability. Ecology 81, 20402047. Goodman, D., 1987. Considerations of stochastic demography in the design and management of reserves. Nat. Res. Model. 1, 205234. Harrison, S., Quinn, J.F., 1989. Correlated environments and the persistence of metapopulations. OIKOS 56, 293298. Hilderbrand, R.H., 2003. The roles of carrying capacity, immigration, and population synchrony on persistence of stream-resident cutthroat trout. Biol. Conserv. 110, 257266.