## Probability of population extinction

In the last section we considered extinction events with an evolutionary perspective, which generally includes long timescales and large-scale extinc-

Time interval 1

Time interval 2

Overall probability assuming events are independent

Number of species in genus qp pqq ppq

p ppq

Fig. 3.5 All possible outcomes after two time steps in the pure birth process formulated by Yule (1924). Overleaf the outcomes are expressed as a tree diagram.

Continued tion events (not necessarily 'mass' extinctions - see the review in Bambach 2006 - but perhaps of many species). The corollary of these activities is a focus on extinction today, where activity may be aimed at smaller taxonomic units. Conservation biologists have addressed both the mechanisms underlying population extinctions and ways of assessing the likelihood of extinction (e.g. Pimm et al. 1988, Foley 1994, Holmes et al. 2007). A recurring problem in conservation biology is how to use population abundance data and population dynamics theory to evaluate which populations of which species should be protected or managed. International conservation organizations, such as the International Union for Conservation of Nature and Natural Resources (IUCN), have become increasingly interested in the application of population biology theory to conservation problems, beginning with Mace and Lande (1991). This process and its outcomes are referred to as population viability analysis (PVA). Here we will examine how modelling techniques can be used to assess the likelihood that a population will become extinct over a given time period.

Extinction may be caused by both stochastic and deterministic processes. Deterministic processes might include habitat loss or hunting (although these will be stochastic in the short term), whereas stochastic processes may include extreme weather events. To model extinction we may think of stochastic processes as reducing population size in the short term but not affecting mean population size, whereas deterministic events reduce the mean population size in a predictable manner. A combination of these two types of process can also be envisaged; that is, a population with a declining mean size that also fluctuates about the mean (Fig. 3.6). 01 23456789 10 Time (years) 01 23456789 10 Time (years) ^

Fig. 3.6 Types of population extinction. (a) Extinction due to fluctuations around a constant mean population size (a stochastic process). (b) Extinction due to a decreasing mean population size (a deterministic process). (c) A combination of (a) and (b).

01 23456789 10 Time (years) ^

Fig. 3.6 Types of population extinction. (a) Extinction due to fluctuations around a constant mean population size (a stochastic process). (b) Extinction due to a decreasing mean population size (a deterministic process). (c) A combination of (a) and (b).

Fluctuations in population size over time may be large and unpredictable. Populations can be modelled by equation 3.1 in a similar manner to species or other taxonomic diversity through time. In a given period of time there is the possibility that a sufficiently large fluctuation will cause the population to become extinct. An estimate of the probability of extinction allows us to quantify that possibility. The random-walk model described above gives one method of modelling these processes. We need to define probabilities of extinction for populations over a given time period; for example, a population may have a 1 in 10 chance of becoming extinct in any one year.

We will apply these ideas to the density-independent model Nt+i = XNt. In Chapter 2 X was assumed to be held constant and in equation 3.1 it was multiplied by a value drawn at random from a probability distribution, £. In this example we will retain the stochastic element but simplify the probability distribution. As before, we are going to assume that the changes in X are random in their operation so there are good and bad years for populations and these are entirely unpredictable in their occurrence over time. These unpredictable factors could affect X through either the survival of different stages and/or the fecundity of individuals. If X can take a range of values, each with a certain probability, it is no longer described by its mean, but by a pdf. In this case we will assume a discrete pdf of six values, each of which has the same probability (an example of a uniform distribution; Fig. 3.7). If we take a value of X at random from this distribution and multiply it by an initial size at time 1 (N1), this will give the value of N2. N2 is then multiplied by a new value of X, plucked again at random from the pdf, to give N3 and so on. The six possible values of X can be matched into three pairs: 1/10, 10; 1/2, 2 and 3/4, 4/3. So, for example, the population has the same chance of being halved as being doubled. We might therefore expect the net change in population size to be zero; that is, that the population will fluctuate around its original level. However, there is a chance of a number of bad years in a row, which might lead to extinction. For example, five very bad years in a row, starting from an initial population size of 1, would give a population size of 1 x (1/10)5.

It is necessary to set an arbitrary extinction density greater than zero because the population will not reach zero given the assumptions of the model. In fact, the population moves asymptotically towards zero. Once the extinction size (or density) is set a proportion of the population following the dynamics described by Nt+1 = XNt and the pdf in Fig. 3.7 may become extinct over a given time period. A simulation using initial population sizes of 1, 5 and 10 is shown in Table 3.1. The simulation was repeated 10 times for each initial population size and stopped after 20 time periods if the population had not become extinct. An arbitrary extinction density of 0.5 was assumed. It can be seen that increasing the initial population size from 1 through to 10 decreased the probability of population extinction and increased the mean persistence time, which is what we would expect. It may seem surprising that, with an initial population size of 1, any of these populations persisted at all. The fact that three out of 10 did persist is a consequence of the high variance in the distribution of X.  Fig. 3.7 Probability distribution for X used in the extinction exercise: (a) raw values and (b) log-transformed values.

Table 3.1 Simulation results for a density-independent model Nt+i = XNt. Each simulation was for a maximum of 20 time periods and was replicated 10 times with the pdf illustrated in Fig. 3.7.

Initial population size Probability of extinction Mean persistence time (generations)

X was made to follow a uniform distribution, the arithmetic mean of which is 2.45. Despite this it is not predicted that the population would increase in size if it persists. This is because we need to consider the mean effect of multiplying Nt by X, not the mean effect of adding X. In this example, the pdf was so simple that we could see that, on average, the mean effect would be as if X was 1 (indeed, the values were chosen for that reason!). So, the net effect of multiplying Nt by the six X values was to multiply by 1 (as with equation 3.1); that is, not to change the original value of N when averaged over many years. Rather than the arithmetic mean of N it is the geometric mean that is relevant here. This applies to all models of the type Nt+i = XNt. The geometric mean of a set of r numbers is found by multiplying them together and taking the rth root (recall the method of determining the diversification rate in Chapter 2). This is equivalent to taking the logarithm (to a given base) of the raw values, taking the arithmetic mean of those values and then back-transforming to get the geometric mean. As we saw with equations 3.1 and 3.2, using the logarithms of the numbers gives us an additive rather than a multiplicative model. Thus, Nt+1 = XNt is transformed to log Nt+1 = log X + log Nt. If we look at the pdf of log10 X we see that the arithmetic mean is zero (see Fig. 3.7, where the bars are symmetrical about zero) and therefore, on average, log Nt+1 = log Nt. Back-transforming (taking the antilog) gives the geometric mean of 1.

Models of population extinction have used these ideas to estimate probabilities of extinction and mean times to extinction (or, conversely, persistence time) for a wide range of species. Foley (1994) used the density-independent model ln Nt+1 = ln X + ln Nt, assuming that ln X was normally distributed with a variance vr and mean 0, and that the population started at size N0 and proceeded on a random walk with a maximum value of k and a minimum value of 0: extinction. The mean time to extinction, Te, is then given as:

This model was applied to different species, showing times to extinction of 19-237 years for five populations of the wolf (Canis lupus) and 1378-6107 years for six populations of the mountain lion (Felis concolor). It would be unwise to accept these and other estimates of Te or probability of extinction as absolute estimates. They are based on simple models with a series of assumptions. Perhaps their most useful function is to provide an estimate of the relative likelihood of extinction, as a contribution to a population or species viability assessment, such as that of the IUCN Red List.