This vignette describes the simple workflow of fitting model to first records data. We’ll go over two examples: fitting the Solow & Costello (2004), and fitting the modified sampling-proxy model.
For the most basic demonstration, let’s look at the data provided in Solow and Costello (2004) which describes discoveries of introduced species in the San Francisco estuary (California, USA) between the years 1850–1995 (Cohen, 1995). The data in this case are simply first records of aliens:
data("sfestuary")
print(sfestuary)
#> [1] 0 0 1 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 0 0 2 0 1 5 1 0 0 1 1 0 0 0 0 0 0 0 0
#> [38] 0 0 1 1 0 2 2 2 1 1 1 0 0 2 0 0 3 0 1 1 2 0 0 0 1 1 0 0 0 0 0 0 2 0 0 0 0
#> [75] 0 0 1 0 1 1 1 1 0 0 1 1 2 4 0 0 0 0 2 0 4 2 1 1 1 0 3 0 1 1 4 0 1 1 0 0 1
#> [112] 2 4 0 1 1 0 1 1 1 2 3 4 1 0 3 5 4 5 1 0 4 2 0 1 4 1 1 2 0 1 7 4 0 0
We’ll plot it in a cumulative form, replicating the plot from Solow and Costello (2004):
library(alien)
library(ggplot2)
years <- seq_along(sfestuary) + 1850 # set starting year for the figure
ggplot()+
aes(x = years, y = cumsum(sfestuary))+
geom_line() +
coord_cartesian(ylim = c(0,150))+
scale_x_continuous(breaks = seq(1860, 1980, 20)) +
scale_y_continuous(breaks = seq(0, 150, 50)) +
ylab("Cumulative discoveries") + theme(axis.title.x = element_blank())
As described thoroughly, these discoveries also entail trends in the
probability of detecting new alien species. To estimate the introduction
rate, \({\beta_1}\), from these data,
we will fit the Solow and Costello model using the snc
function. We can use the control
argument to pass a list of
options to optim
which does the Maximum-Likelihood
Estimation1:
model <- snc(y = sfestuary, control = list(maxit = 1e4))
#> ! no data supplied, using time as independent variable
When only a vector describing discoveries is supplied,
snc
warns users that it uses the time as the independent
variable, similar to the original S&C model.
The result is a list containing several objects:
names(model)
#> [1] "records" "convergence" "log-likelihood" "coefficients"
#> [5] "type" "fitted.values" "predict"
We’ll go over each.
Shows the supplied records data.
model$records
#> [1] 0 0 1 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 0 0 2 0 1 5 1 0 0 1 1 0 0 0 0 0 0 0 0
#> [38] 0 0 1 1 0 2 2 2 1 1 1 0 0 2 0 0 3 0 1 1 2 0 0 0 1 1 0 0 0 0 0 0 2 0 0 0 0
#> [75] 0 0 1 0 1 1 1 1 0 0 1 1 2 4 0 0 0 0 2 0 4 2 1 1 1 0 3 0 1 1 4 0 1 1 0 0 1
#> [112] 2 4 0 1 1 0 1 1 1 2 3 4 1 0 3 5 4 5 1 0 4 2 0 1 4 1 1 2 0 1 7 4 0 0
Did the optimization algorithm converge? This prints out the
convergence code from optim
:
Code | Meaning/Troubleshooting |
---|---|
0 | Successful convergence |
1 | Iteration limit maxit had been reached (increase
maxit using
control = list(maxit = number)) |
10 | Degeneracy of the Nelder-Mead simplex |
51 | Warning from the "L-BFGS-B" method; Use
debug(snc) and check the optim component
message for further details. |
52 | Error from the "L-BFGS-B" method; Use
debug(snc) and check the optim component
message for further details. |
The log-likelihood at the end point of the algorithm (preferably at convergence). Can be used for model selection if needed:
The parameter estimates.
beta0
signifies \({\beta_0}\) - the intercept for \({\mu}\).gamma0
signifies \({\gamma_0}\) - the intercept for \({\Pi}\).gamma2
signifies \({\gamma_2}\) - and will only appear when
the snc
argument growth
is set to
TRUE
(the default).The fitted \({\lambda_t}\) values of the model. The mean of the Poisson distribution from which the records are assumed to derive.
Once we’ve fitted the model, we can use its fit to easily plot \({\lambda_t}\) along with the first records
using the function plot_snc
. Users can choose either
annual
or cumulative
plots. Because the output
is a ggplot
object, it can easily be customized
further:
We can use the function to specify model parameters for either the introduction or detection. Such changes are also possible without supplementation of external data, by constraining either or both of the introduction and detection processes. Next, we’ll set detection to be constant with time:
constant_detection <- snc(sfestuary, pi = ~ 1, growth = FALSE)
#> ! No formula defined for mu, using time as independent variable
Here, the model constrain the \({\gamma_1}\) to 0 by containing pi to an
intercept-only model, and constrain \({\gamma_2}\) to 0 by setting
growth
to FALSE
.
We can examine the likelihoods of the new model:
We can also constrain the introduction rate, as we did with the detection probability:]
constant_introduction <- snc(sfestuary, mu = ~1)
#> ! No formula defined for pi, using time as independent variable
Checking the likelihood of this model shows that there is weak statistical support for the introduction to be constant in this example:
Now we’ll look at a more elaborated model, which uses external data
to control for changes in sampling intensity (Buba et al, 2024). We’ll
demonstrate that using data used in Belmaker et al (2009) which describe
discoveries of native and alien species in the Mediterranean Sea between
the years 1927–2017 (Golani, 2021). We will the medfish
data included in the alien package:
data("medfish")
head(medfish)
#> year time natives aliens
#> 1 1927 0 52 5
#> 2 1930 3 0 1
#> 3 1931 4 12 0
#> 4 1934 7 18 0
#> 5 1935 8 17 1
#> 6 1937 10 1 1
The data has several columns: 1. year
- The year of the
observations. 2. time
- how much time has passed from the
first observation in the data until this point. 3. natives
- how many natives were newly described in this year. 4.
alien
- how many aliens were newly described in this
year.
Here, for demonstration sake only, we will use the trend in native species discovery as a proxy for the sampling throughout the time series time span. Note that using native discoveries as a proxy for sampling has several limitations as described in Buba et al (2024). We will begin by adding a column to the data where we scale the native species discoveries:
Now, we can use these scaled values in our model:
sampling_proxy_model <- snc(aliens, pi = ~ natives_scaled, data = medfish_for_model, control = list(maxit = 1000))
#> ! No formula defined for mu, using time as independent variable
In case we want to supply variables to \(\mu\), the introduction rate, this can be
done by the argument mu
of the snc
function,
in the same manner.
Belmaker, J., Brokovich, E., China, V., Golani, D., and Kiflawi, M. 2009. Estimating the rate of biological introductions: Lessepsian fishes in the Mediterranean. Ecology, 90(4), 1134–1141. https://esajournals.onlinelibrary.wiley.com/doi/10.1890/07-1904.1
Buba, Y., Kiflwai, M., McGeoch, M. A., Belmaker, J. (2024) Evaluating models for estimating introduction rates of alien species from discovery records. https://doi.org/10.1111/geb.13859
Cohen, A. N., and J. T. Carlton. 1995. Nonindigenous aquatic species in a United States estuary: a case study of the biological invasions of the San Francisco Bay and Delta. U.S. Fish and Wildlife Service, Washington, D.C., USA. https://repository.library.noaa.gov/view/noaa/40918
Golani, D. 2021. An updated Checklist of the Mediterranean fishes of Israel, with illustrations of recently recorded species and delineation of Lessepsian migrants. Zootaxa, 4956, 1-108. https://www.mapress.com/zt/article/view/zootaxa.4956.1.1
Solow, A. R., & Costello, C. J. (2004). Estimating the rate of species introductions from the discovery record. Ecology, 85(7), 1822–1825. https://doi.org/10.1890/03-3102
In this case we increase maxiter
so the
algorithm will converge↩︎