In txm676/mmSAR2: Fit and Compare Species-Area Relationship Models Using Multimodel Inference
knitr::opts_chunk$set(
out.width = "100%"
)
options(cli.unicode = FALSE)
This vignette is heavily based on the paper that accompanies the package
[@Matthews2019].To cite this vignette, please cite the corresponding paper:
Matthews, T.J., Triantis, K., Whittaker, R.J. and Guilhaumon, F. (2019) sars:
an R package for fitting, evaluating and comparing species–area relationship
models. Ecography, 42, 1446-1455.
Version 1.1.1 of the package has been archived on the Zenodo research data
repository (DOI: 10.5281/zenodo.2573067).
BACKGROUND
The species–area relationship (SAR) describes the near universally observed
pattern whereby the number of species increases with the area sampled, and it
has been described as one of ecology’s few laws (@Rosenzweig1995). The SAR is
a fundamental component of numerous ecological and biogeographical theories,
such as the equilibrium theory of island biogeography (@MacArthurWilson1967).
In addition, SAR models have been widely used in applied ecology and
conservation biogeography: for example, to predict the number of extinctions
due to habitat loss. Numerous types of SAR have been described, and one
primary dichotomy employed is the split of SARs into island SARs (ISARs),
whereby each data point is an individual island or isolated sample, and
species accumulation curves (SACs) that represent cumulative counts of
increased species number with sampling area (@Gray2004). Whilst the remainder
of the paper and the described R package are focused on ISARs, the models and
the model fitting procedure can equally be applied to SACs (see
@Matthews2016a), although it should be noted that in SACs the data points are
not independent of one another.
Over 20 SAR models have been described in the literature (@Dengler2009,
@Tjorve2003, @Tjorve2009, @Triantis2012). However, despite this wide range of
models, the majority of SAR studies are still based exclusively on the power
model (@Arrhenius1921), which if fitted in its non-linear (untransformed)
form generally takes a convex form. Often, the log–log representation of the
power model is used as it can be fitted using standard linear regression, and
its parameters are more easily interpretable (@Rosenzweig1995). However,
whilst the power model has been found to provide a reasonable fit to a wide
range of datasets (@Dengler2009, @Matthews2016b), it is not universally the
best model, and a number of studies have reported other models to provide
better fits to empirical data (e.g. @Triantis2012, @Matthews2016b). The
possibility of scale dependency of the form of the SAR has long been of
interest, with, for example, a theoretical case being made for SARs at
intermediate spatial scales being approximated by a power model, whilst at
larger spatial scales the form of the SAR has been theorised to be sigmoidal.
Additionally, it is only recently that the SAR for archipelagos as units of
analysis and not just islands has started to be studied, and thus we know
little about the form of archipelago SARs.
Due to the increased recognition of model uncertainty in SAR research, a number
of recent studies have employed a multi-model inference approach
(@BurnhamAnderson2002) in the analysis of SARs, whereby either (1) multiple SAR
models are compared using various criteria (e.g. AIC) and a best model is chosen
(e.g. @Dengler2009), or (2) multiple SAR models are fitted and a multi-model
averaged curve is calculated using, for example, AIC weights (e.g.
@Guilhaumon2008). We are not aware of any published software package that
enables users to fit, and create multi-model averaged curves using more than
eight SAR models. Considering currently available software, the BAT R package
provides functions to fit three SAR models (linear, power and logarithmic);
however, this package is focused on general biodiversity assessment and thus
does not provide any additional SAR functionality. The mmSAR R package
(@Guilhaumon2010) is focussed on SARs and while it allows users to fit eight SAR
models using an information theoretic framework, it does not include several
models that have been found to provide the best fits to several empirical
datasets. To provide a set of tools to fill these gaps, we have developed the R
package ‘sars’. The package provides functions to fit 20 SAR models using
non-linear and linear regression, calculate multi-model averaged curves using
various information criteria, and generate confidence intervals using
bootstrapping. Novel features compared with mmSAR include (i) user-friendly
functions for plotting (the user can now plot weighted multimodel SAR curves
along with the individual SAR model curves) and (ii) determining the observed
shape of the model fit (i.e. linear, convex up, convex down or sigmoidal) and
(iii) presence or not of an asymptote, and (iv) functions to fit, plot and
evaluate @Coleman1981's random placement model using a species-site abundance
matrix, and (v) to fit the general dynamic model (GDM) of island biogeography
(@Whittaker2008). In addition, the mmSAR package (which has been deprecated) no
longer complies with recognised programming good practice , is not on CRAN (the
main repository of R packages), and is not user friendly (e.g. it requires the
user to load individual models prior to fitting). There was therefore a need to
design a new package from scratch.
METHODS AND FEATURES
The ‘sars’ (species–area relationships) package has been programmed using
standard S3 methods and is available on CRAN (version 1.2.1) and the
development version on GitHub (txm676/sars), meaning researchers can easily
add in their own models and functions and integrate these into the
multi-model inference framework. Thus, the package represents a resource for
future SAR work that can be built on and expanded by workers in the field.
Fitting individual SAR models and a set of SAR models The package provides
functions to fit each of the 20 SAR models (see also
@Triantis2012). The ‘sar_models’ function can be used to bring up a list of
the 20 model names and a Table can be generated in the package using the
‘display_sars_models’ function. With the exception of the linear model (which
is fitted using standard linear regression), all models are fitted using
non-linear regression and the model parameters are estimated by minimizing
the residual sum of squares with an unconstrained Nelder-Mead optimization
algorithm and the ‘optim’ R function.
Choosing starting parameter values for non-linear regression optimisation
algorithms is not always straight forward, depending on the data at hand. The
starting values for the parameter estimates used as defaults in the package are
carefully chosen to avoid numerical problems and to speed up the convergence
process. Custom starting values can be provided for any of the 20 models using
the ‘start’ argument in the model fit functions. In addition, we also use a grid
search / brute force process (see Archontoulis & Miguez, 2014). ‘grid_start’
creates 100 starting values for each parameter within sensible bounds and then
creates a large matrix storing all possible combinations of starting values
across the n parameters in the model. For example, for a three parameter model
this matrix contains one million rows. A random number of these rows are then
selected and used as starting parameter values. There are three options for the
grid_start argument to control this process. The default (grid_start =
"partial") randomly samples 500 different sets of starting parameter values for
each model, adds these to the model's default starting values and tests all of
these. A more comprehensive set of starting parameter estimates can be used
(grid_start = "exhaustive") - this option allows the user to choose the number
of starting parameter sets to be tested (using the grid_n argument) and includes
a range of additional starting parameter estimates, e.g. very small values and
particular values we have found to be useful for individual models. A large
‘grid_n’ will ensure a more extensive search of starting parameter value space,
but will increase the time taken to fit the model. More generally, using
grid_start = "exhaustive" in combination with a large grid_n can be very time
consuming; however, we would recommend it as it makes it more likely that the
optimal model fit will be found, particularly for the more complex models. This
is particularly true if any of the model fits does not converge, returns a
singular gradient at parameter estimates, or the plot of the model fit does not
look optimum. The grid start procedure can also be turned off (grid_start =
"none"), meaning just the default starting parameter estimates are used. It
should be noted that for certain models (e.g. Weibull 3 par.) 'grid_start' has
been found to occasionally push model fits into strange parameter space (e.g.
step curves) and so for these models 'grid_start' just reverts back to our
provided parameter starting values. However, for these models the 'start'
argument can still be used. For certain models, there are also checks to ensure
that the returned parameters are logical. For example, the optimisation
procedure can sometimes return a negative z-value for the asymptotic model,
meaning in the term z^A, the base is negative. Thus, if a fitting process
returns a negative z-value we return that the model could not be fitted.
Remember that, as grid_start has a random component, when grid_start != "none",
you can get slightly different results each time you fit a model or run
sar_average().
Each individual model fit returns an object of class ‘sars’, which is a list
of 22 elements containing relevant model fit information, such as the model
parameter estimates, the fitted values, the residuals, model fit statistics
(e.g. AIC, R2), the observed model shape (linear, convex or sigmoidal),
whether or not the fit is asymptotic, and convergence information. The
returned object can easily be plotted using the ‘plot.sars’ generic function;
as this function is based on the base R plotting framework, the plot
aesthetics can be edited using standard plotting arguments (see the
‘plot.sars’ documentation in the package). Summary and print generic
functions are also provided for class ‘sars’; these functions follow the
output of the standard ‘lm’ function in the ‘stats’ R package.
Multiple SAR models can be fitted to the same dataset using the ‘sar_multi’
function and the resultant n model fit objects stored together as a
‘fit_collection’ object. This object is a list of class ‘sars’, where each of
the n elements contains an individual SAR model fit. Using the ‘plot.sars’
generic function on a ‘fit_collection’ object generates a grid of the n
individual model fit plots (Fig. 1).
library(sars)
#load an example dataset (Preston, 1962), fit the logarithmic SAR model using
#the grid_start method of selecting starting parameter values, return a model
#fit summary and plot the model fit.
data(galap)
fit <- sar_loga(data = galap, grid_start = "partial")
summary(fit)
plot(fit)
#Create a fit_collection object containing multiple SAR model fits, and
#plot all fits.
fitC <- sar_multi(data = galap, obj = c("power", "loga", "monod"))
plot(fitC) #see Fig.1
Model fit validation
Model fits can be evaluated through tests of the normality and homoscedasticity
of the residuals. In certain cases this is advised, given that the maximum
likelihood parameter estimates and least-squares estimates are equivalent only
under the assumption of normal errors with constant variance (see the Potential
issues section, below). Any of three tests can be selected to test the normality
of the residuals: 1) the Lilliefors extension of the Kolmogorov normality test
(normaTest = “lillie”), 2) the Shapiro-Wilk test of normality (to be preferred
when sample size is small; “shapiro”), and 3) the Kolmogorov-Smirnov test
(“kolmo”). As a default, the option to omit a residuals normality test (“none”)
is used - thus it is up to the user to select a test if appropriate. Three
options are provided to check for the homogeneity of the residuals: 1) a
correlation of the squared residuals with the model fitted values
(“cor.fitted”), 2) a correlation of the squared residuals with the area values
(“cor.area”), or 3) no homogeneity test (“none”; the default). If a test is
selected and is significant at the 5% level a warning is provided in the model
summary; alternatively, the full results of the three tests can be accessed in
the model fit output. A third model validation check for negative predicted
richness values (i.e. when at least one of the fitted values is negative) can be
undertaken and a warning is provided in the model summary if negative values are
predicted. It should be noted that in earlier versions of the package (pre
Version 1.3.1) there was a bug in the test checking for homogeneity of the
residual variance - the correlation used the raw residuals rather than the
squared residuals. This has now been corrected. Finally, the convergence code
from the optim algorithm can be checked manually. Occasionally the optim
algorithm may not fully converge (e.g. due to degeneracy of the Nelder–Mead
simplex) but still return a fitted model. As such, for each model fit we provide
the optim model convergence code (anything other than 0 can indicate an issue)
and a logical value (verge) specifying whether or not the code was 0.
#load an example dataset, fit the linear SAR model whilst running residual
#normality and homogeneity tests, and return the results of the residual
#normality test
data(galap)
fit <- sar_linear(data = galap, normaTest ="lillie", homoTest = "cor.fitted")
summary(fit) #a warning is provided indicating the normality test failed
fit$normaTest
Observed model shape and identifying an asymptote
Whilst each of the 20 models has a general shape (@Triantis2012), the actual
observed shape of the model fit can be different, for some models, depending on
the parameter estimates. This is important as the shape of the curve has
significant implications for conservation applications and the testing of
macroecological theory (@Rosenzweig1995). In ‘sars’, the observed shape of a
model fit is determined using the sequential algorithm outlined in
@Triantis2012. The shape is calculated using the model fit within the observed
range of area values. Briefly, the algorithm works by first determining whether
the fit is a straight line. Then, if the fit is classified as not being linear,
the observed shape is classified as either convex (upwards or downwards) or
sigmoidal by analysis of the second derivative (with respect to area) of the
model fit (the full algorithm is detailed in @Triantis2012, p. 220). It should
be noted that a small number of models (e.g. Extended Power 2) are listed in
previous papers (e.g. @Tjorve2009) as being sigmoidal without an upper asymptote
(and we have followed these classifications). However, a recent study
(@Godeau2020) has argued that sigmoidal models must have an upper asymptote by
definition; thus, the classification of these models is currently uncertain.
There has also been considerable debate in the SAR literature as to whether or
not the SAR is asymptotic. In the ‘sars’ package, to determine whether a fit is
asymptotic, for the relevant models the fitted model parameters are analysed to
check whether the estimated asymptote is within the range of the sample data
(@Triantis2012).
Multimodel SAR curve
As well as fitting individual models, the package provides a function
(‘sar_average’) to fit up to 20 models, compare the resultant fits using
information criteria, and construct a multimodel-averaged SAR curve based on
information criteria weight (see
@Guilhaumon2010). The multimodel average curve is constructed as a linear
combination of individual model fits by multiplying the predicted richness
values of each of the successfully fitted models by the model’s information
criterion weight, and then summing the resultant values across all models
(@BurnhamAnderson2002). Three information criteria are available in the package:
AIC, AICc and BIC. Confidence intervals around the multimodel averaged curve can
be calculated using a non-parametric bootstrap algorithm described in
@Guilhaumon2010. Briefly, each of the SAR models used in the ‘sar_average’
function is fitted to the data, and the fitted values and residuals stored. The
residuals are then transformed using the approach in
@DavisonHinkley1997 (p.259). For each bootstrap sample, an individual model fit
is selected with the probability of selection being equal to that model’s
information criterion weight. The transformed residuals from this fit are then
sampled with replacement and added to the model’s fitted richness values. The
‘sar_average’ function is then used to fit all candidate SAR models to this
bootstrapped set of response values, and the multimodel averaged fitted values
stored. Percentile confidence intervals are then calculated using all
bootstrapped fitted values.
The ‘sar_average’ function can be used without specifying any models, in which
case the package attempts to fit each of the 20 models in Table 1;
alternatively, a vector of model names or a ‘fit_collection’ object (generated
using the ‘sar_multi’ function) can be provided using the ‘obj’ argument. The
three model validation tests listed above (normality and homogeneity of
residuals, and negative predicted values) can be selected (by default none are
selected, so it is up to the user to select any that are appropriate); if any
model fails one or more of the tests during the fitting process it is removed
from the resultant multimodel SAR curve. The output of the ‘sar_average’
function is a list of class ‘multi’ and class ‘sars’, with two elements. The
first element (‘mmi’) contains the multi model inference (fitted values of the
multimodel SAR curve), and the second element (‘details’) contains a range of
information regarding the fitting process, including the successfully fitted
models, the models removed due to failing any of the validation tests, and the
information criterion values, delta values and weights for the successfully
fitted models. The returned object can easily be plotted using the ‘plot.multi’
generic function, and multiple plot options are available (using the ‘type’
argument; see Fig. 2). The fits of all the successfully fitted models and the
multimodel SAR curve (with or without confidence intervals) can be plotted
together (‘type’ = “multi”), and a barplot of the information criterion weights
of each model can also be produced (‘type’ = “bar”).
#load an example dataset (Niering, 1963), run the ‘sar_average’ function
#using a vector of model names and with no model validation tests, and
#produce the plots in Figure 2 of the paper
data(niering)
#run the ‘sar_average’ function using a vector of model names, and with simply
#using the default model starting parameter estimates (grid_start = "none")
fit <- sar_average(data= niering, obj =c("power","loga","koba","logistic","monod",
"negexpo","chapman","weibull3","asymp"),
grid_start = "none", normaTest = "none", homoTest = "none", neg_check = FALSE,
confInt = TRUE, ciN = 50) #a message is provided indicating that one model
#(asymp) could not be used in the confidence interval calculation
par(mfrow = c(3,1)) #plot all model fits and the multimodel SAR curve as a separate curve on top
plot(fit, ModTitle = "a) Multimodel SAR", mmSep = TRUE)
#plot the multimodel SAR curve (with confidence intervals; see explanation
#in the main text, above) on its own
plot(fit, allCurves = FALSE, ModTitle =
"c) Multimodel SAR with confidence intervals", confInt = TRUE)
#Barplot of the information criterion weights of each model
plot(fit, type = "bar", ModTitle = "b) Model weights", cex.lab = 1.3)
One thing to note is how the sar_average() function deals with model
non-convergence. There are different types of non-convergence and these are
dealt with differently. If the optimisation algorithm fails to return any
solution, the model fit is defined as NA and is then removed, and so does not
appear in the model summary table or multi-model curve etc. However, the
optimisation algorithm (e.g. Nelder-Mead) can also return non-NA model fits but
where the solution is potentially non-optimal (e.g. degeneracy of the
Nelder–Mead simplex) - these cases are identified by any optim convergence code
that is not zero. We have decided not to remove these fits (i.e. they are kept
in the model summary table and multimodel curve), but any instances can be
checked using the returned 'details$converged' vector and then the model fitting
re-run without these models, if preferred. Increasing the starting parameters
grid search (see above) may also help avoid this issue.
Additional functions
In addition to the main functions used to fit and compare the 20 SAR models, the
sars package provides additional functions for specific SAR-based analyses.
First, a function is provided to fit the log–log version of the power model (a
function that is often fitted in SAR studies; Rosenzweig 1995) and compare
parameter values with those generated using the non-linear power model. The
log–log version of the power model is not equivalent to its non-linear
counterpart because of non-equivalence in the study of the variation in a
variable and in its transformation, and bias of back-transformed results
obtained on a logarithmic scale. Second, a function has been added that enables
the fitting of @Coleman1981's random placement model to a species/sites
abundance matrix. According to this model, the number of species occurring on an
island depends on the relative area of the island and the regional relative
species abundances. The fit of the random placement model can be determined
through use of a diagnostic plot (which can be generated from the function
output) of island area (log transformed) against species richness, alongside the
model’s predicted values (see
@Wang2010). Following @Wang2010, the model is rejected if more than a third of
the observed data points fall beyond one standard deviation from the expected
curve. Finally, a function is provided to fit the general dynamic model of
island biogeography (@Whittaker2008) using five different SAR models (linear,
logarithmic, power(area), power(area and time), linear power).
#load an example dataset, fit the log-log power model, return a model fit
#summary and plot the model fit. When ‘compare’ == TRUE, the non-linear
#power model is also fitted and the resultant parameter values compared.
#If any islands have zero species, a constant (‘con’) is added to all
#species richness values.
data(galap)
fit <- lin_pow(dat = galap, compare = TRUE, con = 1)
summary(fit)
plot(fit)
#load an example dataset, fit the random placement model and plot the
#model fit and standard deviation. The ‘data’ argument requires a species-
#site abundance matrix: rows are species and columns are sites. The area
#argument requires a vector of site (island) area values.
data(cole_sim)
fit <- coleman(data = cole_sim[[1]], area = cole_sim[[2]])
plot(fit, ModTitle = "Hetfield")
#load an example dataset, fit the GDM using the logarithmic SAR model, and
#compare the GDM with three alternative (nested) models: area and time
#(age of each island), area only, and intercept only.
data(galap)
galap$t <- rgamma(16, 5, scale = 2)#add a random time variable
gdm(data = galap, model = "loga", mod_sel = TRUE)
SAR extrapolation
Extrapolation (e.g. predicting the richness of areas too large to be sampled) is
one of the primary uses of the SAR. Using an individual SAR model fit,
extrapolation simply involves using the model function in combination with the
parameter values estimated using the observed area / richness data to predict
the richness on a larger area. An alternative extrapolation approach to simply
using an individual model l is to use multi-model inference and model averaging,
whereby a larger number of n models is fitted to a set of islands, the models
ranked according to some criterion (e.g. Akaike’s information criterion, AIC)
and the criterion values converted into model weights (i.e. the conditional
probabilities for each of the n models). The n models are then each used to
predict the richness of a larger area and these predictions are multiplied by
the respective model weights and summed to provide a multi-model averaged
prediction. Both of these options are available inside the package.
We would recommend using grid_start = "exhaustive" when fitting models for use
with sar_pred(), as this is more likely to find the optimum fit for a given
model. However, remember that, as grid_start has a random component, when
grid_start != "none", you can get slightly different results each time you fit a
model or run sar_average, and then run sar_pred().
#fit the power model and predict richness on an island of area = 5000
data(galap)
p <- sar_power(data = galap)
sar_pred(p, area = 5000)
#fit three SAR models and predict richness on islands of area = 5000 & 10000
p2 <- sar_multi(galap, obj = c("power", "loga", "koba"))
sar_pred(p2, area = c(5000, 10000))
#calculate a multi-model curve and predict richness on islands of area = 5000 & 10000
#grid_start set to "none" for speed
p3 <- sar_average(data = galap, grid_start = "none")
sar_pred(p3, area = c(5000, 10000))
SAR thresholds
An increasing number of studies have focused on identifying thresholds in the
ISAR, including studies i) focused on identifying the small island effect, ii)
determining whether the ISAR has an upper asymptote and iii) identifying
thresholds in habitat ISARs, which are often systems of conservation concern.
The most common approach in such studies is to use piecewise regression. As
such, the package provide a set of functions for fitting six piecewise
regression models to ISAR data, calculating confidence intervals around the
breakpoint estimates (for certain models), comparing the models using various
information criteria, and plotting the resultant model fits. The six piecewise
models are a selection of 6 models out of the 14 listed by @Gao2019 and
comprise the continuous one-threshold and two-threshold, discontinuous
one-threshold and two-threshold, and left-horizontal one-threshold and
two-threshold models (see @Gao2019 for further information).
The main function is ‘sar_threshold’, which fits the six piecewise models, in
addition to a linear model and an intercept only model for comparison (using the
‘non_th_models’ argument). Summary and plot generic functions are available
which return user-friendly output. As the coefficients in the fitted breakpoint
regression models do not all represent the intercepts and slopes of the
different segments (for these it is necessary to add different coefficients
together), a separate function (‘get_coef’) can be used to calculate these. A
separate function (‘threshold_ci’) generates the confidence intervals around the
breakpoints of the one-threshold continuous and left-horizontal models.
Detailed information on fitting and plotting the threshold models, as well as
the additional functions for calculating model slopes and intercepts, and
generating confidence intervals around the threshold estimates, can be found in
@Matthews2020.
#load an example dataset, and fit the continuous two-threshold model
#to the data (with area transformed using log to the base 10), using an
#interval of 0.1 (for speed)
data(aegean2)
fit <- sar_threshold(data = aegean2, mod = c("ContTwo"), interval = 0.1,
non_th_models = FALSE, logAxes = "area", con = 1,
logT = log10, nisl = NULL)
#generate model fitting summary table (generally more useful when fitting multiple models)
summary(fit)
#Plot the resultant model fit
plot(fit, cex = 0.8, cex.main = 1, cex.lab = 0.8, pcol = "grey")
On the calculation of information criteria
There are different ways to calculate the various information criteria (IC) used
for model comparison (e.g. AIC, BIC). One difference relates to whether the
residual sum of squares (rss) or the log-likelihood (ll) is used. Under standard
assumptions (e.g. independence of data points, homoscedasticity and normality of
the residuals), the two approaches produce identical parameter estimates for
regression models. However, the formulas are different and thus can produce
different IC values for the same model. For example, historically in the ‘sars’
package we have calculated IC values using formulas based on the rss (Burnham &
Anderson, 2002; Guilhaumon et al. 2008). This meant that the IC values generated
in ‘sars’ were not comparable with values generated in packages using different
formulas. For example, in the nls (the main function for non-linear regression
in R) and lm functions in the stats R package, a ll approach is used, meaning IC
values from models fitted using nls could not be compared with IC values from
sars models. To re-iterate, the parameter estimates are comparable, it is simply
that the IC values will differ. In version of ‘sars’ (version 1.2.2) we changed
our IC formulas to match those in nls and lm. Thus, if users have built their
own models using nls (or lm) and wish to compare IC values with models fitted in
‘sars’, this is now straightforward. To re-create IC values from previous
studies (i.e. those using a version of sars pre 1.2.2), it will be necessary to
download sars Version 1.2.1 or earlier (either from CRAN or GitHub; version
1.1.1 was published as a release on GitHub). It is important to note that these
are not the only two ways of calculating ICs for regression models, and other
formulas exits. Thus, if building models using other functions and packages
(i.e. other than nls or lm), users should make sure to check how these packages
calculate IC values before comparing with models fitted in sars. In sars, as in
nls and lm, we include an additional parameter for estimation of the variance.
For example, the power model is considered to have three parameters when
calculating IC values. Finally, if users are comparing models fitted in sars
with their own models fitted using other packages, it is essential that IC
values are all calculated using the same dependent variable (i.e.
non-transformed richness values and not logged richness, as sars only uses the
former currently).
Potential issues with using information criteria in conjunction with
non-linear regression models
Before starting this section, it is worth pointing out that we are not
statisticians and the following discussion should be interpreted with this point
in mind. We have however reviewed a lot of the literature (including a number of
statistics forums) on the use of information criteria (e.g. AIC) in the context
of comparing non-linear regression models. Based on this review, we have
concluded that there are three potential issues with this approach.
First, many information criteria (including AIC) are theoretically applicable
only when a model is fitted through the approach of maximum likelihood. In the
case of linear and non-linear regression (e.g. the nls function in R, and the
approach we use in ‘sars’), a least-squares estimator (minimising the residual
sum of squares) is generally used rather than likelihood maximisation, for
various reasons, such as analytical tractability. This is not necessarily a
problem, as the least-squares parameter estimates and the maximum likelihood
parameter estimates are equivalent in the case of normal and independent errors
with a mean of zero and constant variance (@Bates1988). This holds for both
linear and non-linear regression models (@Bates1988; @Forum1). However, if these
assumptions are not met this equivalence may break down and thus the use of AIC
etc would not necessarily be appropriate. The use of the normality and
homogeneity of variance checks within the package may be useful in this regard,
but note that it is now up to the user to manually turn these checks on.
Second, most of the commonly used information criteria (e.g. AIC) include a term
that penalises for the number of parameters. In the case of linear regression
models, counting the number of parameters is straight forward. However,
non-linear models are so called because they are non-linear in regard to their
parameters (e.g. a quadratic model is a linear model but with a non-linear
shaped curve). As such, it is not always as straight forward to count parameters
in the same way, particularly in the case of very complex non-linear and machine
learning type models (i.e. not those used in ‘sars’). “That is, a single
parameter in the model may count as more or less than one parameter, in some
sense” (@Forum2). However, in the absence of any other information, or indeed
alternative model comparison approaches, all one can do is use the number of
estimable parameters (sensu Burnham & Anderson) in a model, which is what we
follow in the ‘sars’ package, and is what a large number of studies do (and
advise) in the ecological and wider literature (e.g. @Archontoulis2015;
@Banks2017; see also @Forum3’s response to @Forum2).
Finally, some have argued that you should only use AIC etc to compare models
within the same ‘model complex’, and even only in the case of nested models
(e.g. @Ripley). From this it could be concluded that we should not compare the
suite of 19 non-linear models with the linear model. However,
@BurnhamAnderson2002 and @Anderson2006 have argued that this is not the case and
AIC for example can be used to compare non-nested models. In addition, again
many studies use AIC to compare non-nested models from different ‘complexes’.
We have worked on the assumption that, as long as the response variable is the
same and models are comparable in terms of parameter number calculation (e.g.
all include or not a parameter for the variance), it is appropriate to compare
different types of model (at least within the limited selection used in the
package).
CONCLUSIONS
The SAR has been a cornerstone of ecological and biogeographical science for
almost a century and its form and fit are still of great significance in both
theoretical and applied contexts. The development of the ‘sars’ R package
should aid future SAR research by providing a comprehensive set of tools that
enable in-depth exploration of SARs and SAR-related patterns. In addition,
the package has been designed in such a way as to allow other SAR researchers
to add (e.g. via GitHub) new functions and models in the future to ensure the
package is of lasting value. For example, future additions to the package
could include the suite of countryside biogeography SAR models that have
recently been published, standard SAR functions not so far incorporated, or
functions specifically intended for analysis of species accumulation curves
or endemics–area relationships. Finally, whilst the focus of this paper has
been on classic SARs, there is no reason that the functionality in the ‘sars’
package cannot be used to analyse other diversity–area relationships (e.g.
functional or phylogenetic diversity–area relationships). Application of the
full set of 20 models, in addition to the multimodel SAR framework, included
in the ‘sars’ package to a wider range and type of data (e.g. trait and
phylogenetic data) will likely be revealing and will help in improving our
understanding of SARs, and diversity–area relationships more generally.
References
txm676/mmSAR2 documentation built on Nov. 21, 2024, 5:03 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
knitr::opts_chunk$set( out.width = "100%" ) options(cli.unicode = FALSE)
This vignette is heavily based on the paper that accompanies the package [@Matthews2019].To cite this vignette, please cite the corresponding paper:
Matthews, T.J., Triantis, K., Whittaker, R.J. and Guilhaumon, F. (2019) sars: an R package for fitting, evaluating and comparing species–area relationship models. Ecography, 42, 1446-1455.
Version 1.1.1 of the package has been archived on the Zenodo research data repository (DOI: 10.5281/zenodo.2573067).
BACKGROUND
The species–area relationship (SAR) describes the near universally observed pattern whereby the number of species increases with the area sampled, and it has been described as one of ecology’s few laws (@Rosenzweig1995). The SAR is a fundamental component of numerous ecological and biogeographical theories, such as the equilibrium theory of island biogeography (@MacArthurWilson1967). In addition, SAR models have been widely used in applied ecology and conservation biogeography: for example, to predict the number of extinctions due to habitat loss. Numerous types of SAR have been described, and one primary dichotomy employed is the split of SARs into island SARs (ISARs), whereby each data point is an individual island or isolated sample, and species accumulation curves (SACs) that represent cumulative counts of increased species number with sampling area (@Gray2004). Whilst the remainder of the paper and the described R package are focused on ISARs, the models and the model fitting procedure can equally be applied to SACs (see @Matthews2016a), although it should be noted that in SACs the data points are not independent of one another.
Over 20 SAR models have been described in the literature (@Dengler2009, @Tjorve2003, @Tjorve2009, @Triantis2012). However, despite this wide range of models, the majority of SAR studies are still based exclusively on the power model (@Arrhenius1921), which if fitted in its non-linear (untransformed) form generally takes a convex form. Often, the log–log representation of the power model is used as it can be fitted using standard linear regression, and its parameters are more easily interpretable (@Rosenzweig1995). However, whilst the power model has been found to provide a reasonable fit to a wide range of datasets (@Dengler2009, @Matthews2016b), it is not universally the best model, and a number of studies have reported other models to provide better fits to empirical data (e.g. @Triantis2012, @Matthews2016b). The possibility of scale dependency of the form of the SAR has long been of interest, with, for example, a theoretical case being made for SARs at intermediate spatial scales being approximated by a power model, whilst at larger spatial scales the form of the SAR has been theorised to be sigmoidal. Additionally, it is only recently that the SAR for archipelagos as units of analysis and not just islands has started to be studied, and thus we know little about the form of archipelago SARs.
Due to the increased recognition of model uncertainty in SAR research, a number of recent studies have employed a multi-model inference approach (@BurnhamAnderson2002) in the analysis of SARs, whereby either (1) multiple SAR models are compared using various criteria (e.g. AIC) and a best model is chosen (e.g. @Dengler2009), or (2) multiple SAR models are fitted and a multi-model averaged curve is calculated using, for example, AIC weights (e.g. @Guilhaumon2008). We are not aware of any published software package that enables users to fit, and create multi-model averaged curves using more than eight SAR models. Considering currently available software, the BAT R package provides functions to fit three SAR models (linear, power and logarithmic); however, this package is focused on general biodiversity assessment and thus does not provide any additional SAR functionality. The mmSAR R package (@Guilhaumon2010) is focussed on SARs and while it allows users to fit eight SAR models using an information theoretic framework, it does not include several models that have been found to provide the best fits to several empirical datasets. To provide a set of tools to fill these gaps, we have developed the R package ‘sars’. The package provides functions to fit 20 SAR models using non-linear and linear regression, calculate multi-model averaged curves using various information criteria, and generate confidence intervals using bootstrapping. Novel features compared with mmSAR include (i) user-friendly functions for plotting (the user can now plot weighted multimodel SAR curves along with the individual SAR model curves) and (ii) determining the observed shape of the model fit (i.e. linear, convex up, convex down or sigmoidal) and (iii) presence or not of an asymptote, and (iv) functions to fit, plot and evaluate @Coleman1981's random placement model using a species-site abundance matrix, and (v) to fit the general dynamic model (GDM) of island biogeography (@Whittaker2008). In addition, the mmSAR package (which has been deprecated) no longer complies with recognised programming good practice , is not on CRAN (the main repository of R packages), and is not user friendly (e.g. it requires the user to load individual models prior to fitting). There was therefore a need to design a new package from scratch.
METHODS AND FEATURES
The ‘sars’ (species–area relationships) package has been programmed using standard S3 methods and is available on CRAN (version 1.2.1) and the development version on GitHub (txm676/sars), meaning researchers can easily add in their own models and functions and integrate these into the multi-model inference framework. Thus, the package represents a resource for future SAR work that can be built on and expanded by workers in the field.
Fitting individual SAR models and a set of SAR models The package provides functions to fit each of the 20 SAR models (see also @Triantis2012). The ‘sar_models’ function can be used to bring up a list of the 20 model names and a Table can be generated in the package using the ‘display_sars_models’ function. With the exception of the linear model (which is fitted using standard linear regression), all models are fitted using non-linear regression and the model parameters are estimated by minimizing the residual sum of squares with an unconstrained Nelder-Mead optimization algorithm and the ‘optim’ R function.
Choosing starting parameter values for non-linear regression optimisation algorithms is not always straight forward, depending on the data at hand. The starting values for the parameter estimates used as defaults in the package are carefully chosen to avoid numerical problems and to speed up the convergence process. Custom starting values can be provided for any of the 20 models using the ‘start’ argument in the model fit functions. In addition, we also use a grid search / brute force process (see Archontoulis & Miguez, 2014). ‘grid_start’ creates 100 starting values for each parameter within sensible bounds and then creates a large matrix storing all possible combinations of starting values across the n parameters in the model. For example, for a three parameter model this matrix contains one million rows. A random number of these rows are then selected and used as starting parameter values. There are three options for the grid_start argument to control this process. The default (grid_start = "partial") randomly samples 500 different sets of starting parameter values for each model, adds these to the model's default starting values and tests all of these. A more comprehensive set of starting parameter estimates can be used (grid_start = "exhaustive") - this option allows the user to choose the number of starting parameter sets to be tested (using the grid_n argument) and includes a range of additional starting parameter estimates, e.g. very small values and particular values we have found to be useful for individual models. A large ‘grid_n’ will ensure a more extensive search of starting parameter value space, but will increase the time taken to fit the model. More generally, using grid_start = "exhaustive" in combination with a large grid_n can be very time consuming; however, we would recommend it as it makes it more likely that the optimal model fit will be found, particularly for the more complex models. This is particularly true if any of the model fits does not converge, returns a singular gradient at parameter estimates, or the plot of the model fit does not look optimum. The grid start procedure can also be turned off (grid_start = "none"), meaning just the default starting parameter estimates are used. It should be noted that for certain models (e.g. Weibull 3 par.) 'grid_start' has been found to occasionally push model fits into strange parameter space (e.g. step curves) and so for these models 'grid_start' just reverts back to our provided parameter starting values. However, for these models the 'start' argument can still be used. For certain models, there are also checks to ensure that the returned parameters are logical. For example, the optimisation procedure can sometimes return a negative z-value for the asymptotic model, meaning in the term z^A, the base is negative. Thus, if a fitting process returns a negative z-value we return that the model could not be fitted. Remember that, as grid_start has a random component, when grid_start != "none", you can get slightly different results each time you fit a model or run sar_average().
Each individual model fit returns an object of class ‘sars’, which is a list of 22 elements containing relevant model fit information, such as the model parameter estimates, the fitted values, the residuals, model fit statistics (e.g. AIC, R2), the observed model shape (linear, convex or sigmoidal), whether or not the fit is asymptotic, and convergence information. The returned object can easily be plotted using the ‘plot.sars’ generic function; as this function is based on the base R plotting framework, the plot aesthetics can be edited using standard plotting arguments (see the ‘plot.sars’ documentation in the package). Summary and print generic functions are also provided for class ‘sars’; these functions follow the output of the standard ‘lm’ function in the ‘stats’ R package. Multiple SAR models can be fitted to the same dataset using the ‘sar_multi’ function and the resultant n model fit objects stored together as a ‘fit_collection’ object. This object is a list of class ‘sars’, where each of the n elements contains an individual SAR model fit. Using the ‘plot.sars’ generic function on a ‘fit_collection’ object generates a grid of the n individual model fit plots (Fig. 1).
library(sars)
#load an example dataset (Preston, 1962), fit the logarithmic SAR model using #the grid_start method of selecting starting parameter values, return a model #fit summary and plot the model fit. data(galap) fit <- sar_loga(data = galap, grid_start = "partial") summary(fit) plot(fit)
#Create a fit_collection object containing multiple SAR model fits, and #plot all fits. fitC <- sar_multi(data = galap, obj = c("power", "loga", "monod")) plot(fitC) #see Fig.1
Model fit validation
Model fits can be evaluated through tests of the normality and homoscedasticity of the residuals. In certain cases this is advised, given that the maximum likelihood parameter estimates and least-squares estimates are equivalent only under the assumption of normal errors with constant variance (see the Potential issues section, below). Any of three tests can be selected to test the normality of the residuals: 1) the Lilliefors extension of the Kolmogorov normality test (normaTest = “lillie”), 2) the Shapiro-Wilk test of normality (to be preferred when sample size is small; “shapiro”), and 3) the Kolmogorov-Smirnov test (“kolmo”). As a default, the option to omit a residuals normality test (“none”) is used - thus it is up to the user to select a test if appropriate. Three options are provided to check for the homogeneity of the residuals: 1) a correlation of the squared residuals with the model fitted values (“cor.fitted”), 2) a correlation of the squared residuals with the area values (“cor.area”), or 3) no homogeneity test (“none”; the default). If a test is selected and is significant at the 5% level a warning is provided in the model summary; alternatively, the full results of the three tests can be accessed in the model fit output. A third model validation check for negative predicted richness values (i.e. when at least one of the fitted values is negative) can be undertaken and a warning is provided in the model summary if negative values are predicted. It should be noted that in earlier versions of the package (pre Version 1.3.1) there was a bug in the test checking for homogeneity of the residual variance - the correlation used the raw residuals rather than the squared residuals. This has now been corrected. Finally, the convergence code from the optim algorithm can be checked manually. Occasionally the optim algorithm may not fully converge (e.g. due to degeneracy of the Nelder–Mead simplex) but still return a fitted model. As such, for each model fit we provide the optim model convergence code (anything other than 0 can indicate an issue) and a logical value (verge) specifying whether or not the code was 0.
#load an example dataset, fit the linear SAR model whilst running residual #normality and homogeneity tests, and return the results of the residual #normality test data(galap) fit <- sar_linear(data = galap, normaTest ="lillie", homoTest = "cor.fitted") summary(fit) #a warning is provided indicating the normality test failed fit$normaTest
Observed model shape and identifying an asymptote
Whilst each of the 20 models has a general shape (@Triantis2012), the actual observed shape of the model fit can be different, for some models, depending on the parameter estimates. This is important as the shape of the curve has significant implications for conservation applications and the testing of macroecological theory (@Rosenzweig1995). In ‘sars’, the observed shape of a model fit is determined using the sequential algorithm outlined in @Triantis2012. The shape is calculated using the model fit within the observed range of area values. Briefly, the algorithm works by first determining whether the fit is a straight line. Then, if the fit is classified as not being linear, the observed shape is classified as either convex (upwards or downwards) or sigmoidal by analysis of the second derivative (with respect to area) of the model fit (the full algorithm is detailed in @Triantis2012, p. 220). It should be noted that a small number of models (e.g. Extended Power 2) are listed in previous papers (e.g. @Tjorve2009) as being sigmoidal without an upper asymptote (and we have followed these classifications). However, a recent study (@Godeau2020) has argued that sigmoidal models must have an upper asymptote by definition; thus, the classification of these models is currently uncertain.
There has also been considerable debate in the SAR literature as to whether or not the SAR is asymptotic. In the ‘sars’ package, to determine whether a fit is asymptotic, for the relevant models the fitted model parameters are analysed to check whether the estimated asymptote is within the range of the sample data (@Triantis2012).
Multimodel SAR curve
As well as fitting individual models, the package provides a function (‘sar_average’) to fit up to 20 models, compare the resultant fits using information criteria, and construct a multimodel-averaged SAR curve based on information criteria weight (see @Guilhaumon2010). The multimodel average curve is constructed as a linear combination of individual model fits by multiplying the predicted richness values of each of the successfully fitted models by the model’s information criterion weight, and then summing the resultant values across all models (@BurnhamAnderson2002). Three information criteria are available in the package: AIC, AICc and BIC. Confidence intervals around the multimodel averaged curve can be calculated using a non-parametric bootstrap algorithm described in @Guilhaumon2010. Briefly, each of the SAR models used in the ‘sar_average’ function is fitted to the data, and the fitted values and residuals stored. The residuals are then transformed using the approach in @DavisonHinkley1997 (p.259). For each bootstrap sample, an individual model fit is selected with the probability of selection being equal to that model’s information criterion weight. The transformed residuals from this fit are then sampled with replacement and added to the model’s fitted richness values. The ‘sar_average’ function is then used to fit all candidate SAR models to this bootstrapped set of response values, and the multimodel averaged fitted values stored. Percentile confidence intervals are then calculated using all bootstrapped fitted values.
The ‘sar_average’ function can be used without specifying any models, in which case the package attempts to fit each of the 20 models in Table 1; alternatively, a vector of model names or a ‘fit_collection’ object (generated using the ‘sar_multi’ function) can be provided using the ‘obj’ argument. The three model validation tests listed above (normality and homogeneity of residuals, and negative predicted values) can be selected (by default none are selected, so it is up to the user to select any that are appropriate); if any model fails one or more of the tests during the fitting process it is removed from the resultant multimodel SAR curve. The output of the ‘sar_average’ function is a list of class ‘multi’ and class ‘sars’, with two elements. The first element (‘mmi’) contains the multi model inference (fitted values of the multimodel SAR curve), and the second element (‘details’) contains a range of information regarding the fitting process, including the successfully fitted models, the models removed due to failing any of the validation tests, and the information criterion values, delta values and weights for the successfully fitted models. The returned object can easily be plotted using the ‘plot.multi’ generic function, and multiple plot options are available (using the ‘type’ argument; see Fig. 2). The fits of all the successfully fitted models and the multimodel SAR curve (with or without confidence intervals) can be plotted together (‘type’ = “multi”), and a barplot of the information criterion weights of each model can also be produced (‘type’ = “bar”).
#load an example dataset (Niering, 1963), run the ‘sar_average’ function #using a vector of model names and with no model validation tests, and #produce the plots in Figure 2 of the paper data(niering) #run the ‘sar_average’ function using a vector of model names, and with simply #using the default model starting parameter estimates (grid_start = "none") fit <- sar_average(data= niering, obj =c("power","loga","koba","logistic","monod", "negexpo","chapman","weibull3","asymp"), grid_start = "none", normaTest = "none", homoTest = "none", neg_check = FALSE, confInt = TRUE, ciN = 50) #a message is provided indicating that one model #(asymp) could not be used in the confidence interval calculation par(mfrow = c(3,1)) #plot all model fits and the multimodel SAR curve as a separate curve on top plot(fit, ModTitle = "a) Multimodel SAR", mmSep = TRUE) #plot the multimodel SAR curve (with confidence intervals; see explanation #in the main text, above) on its own plot(fit, allCurves = FALSE, ModTitle = "c) Multimodel SAR with confidence intervals", confInt = TRUE) #Barplot of the information criterion weights of each model plot(fit, type = "bar", ModTitle = "b) Model weights", cex.lab = 1.3)
One thing to note is how the sar_average() function deals with model non-convergence. There are different types of non-convergence and these are dealt with differently. If the optimisation algorithm fails to return any solution, the model fit is defined as NA and is then removed, and so does not appear in the model summary table or multi-model curve etc. However, the optimisation algorithm (e.g. Nelder-Mead) can also return non-NA model fits but where the solution is potentially non-optimal (e.g. degeneracy of the Nelder–Mead simplex) - these cases are identified by any optim convergence code that is not zero. We have decided not to remove these fits (i.e. they are kept in the model summary table and multimodel curve), but any instances can be checked using the returned 'details$converged' vector and then the model fitting re-run without these models, if preferred. Increasing the starting parameters grid search (see above) may also help avoid this issue.
Additional functions
In addition to the main functions used to fit and compare the 20 SAR models, the sars package provides additional functions for specific SAR-based analyses. First, a function is provided to fit the log–log version of the power model (a function that is often fitted in SAR studies; Rosenzweig 1995) and compare parameter values with those generated using the non-linear power model. The log–log version of the power model is not equivalent to its non-linear counterpart because of non-equivalence in the study of the variation in a variable and in its transformation, and bias of back-transformed results obtained on a logarithmic scale. Second, a function has been added that enables the fitting of @Coleman1981's random placement model to a species/sites abundance matrix. According to this model, the number of species occurring on an island depends on the relative area of the island and the regional relative species abundances. The fit of the random placement model can be determined through use of a diagnostic plot (which can be generated from the function output) of island area (log transformed) against species richness, alongside the model’s predicted values (see @Wang2010). Following @Wang2010, the model is rejected if more than a third of the observed data points fall beyond one standard deviation from the expected curve. Finally, a function is provided to fit the general dynamic model of island biogeography (@Whittaker2008) using five different SAR models (linear, logarithmic, power(area), power(area and time), linear power).
#load an example dataset, fit the log-log power model, return a model fit #summary and plot the model fit. When ‘compare’ == TRUE, the non-linear #power model is also fitted and the resultant parameter values compared. #If any islands have zero species, a constant (‘con’) is added to all #species richness values. data(galap) fit <- lin_pow(dat = galap, compare = TRUE, con = 1) summary(fit) plot(fit) #load an example dataset, fit the random placement model and plot the #model fit and standard deviation. The ‘data’ argument requires a species- #site abundance matrix: rows are species and columns are sites. The area #argument requires a vector of site (island) area values. data(cole_sim) fit <- coleman(data = cole_sim[[1]], area = cole_sim[[2]]) plot(fit, ModTitle = "Hetfield") #load an example dataset, fit the GDM using the logarithmic SAR model, and #compare the GDM with three alternative (nested) models: area and time #(age of each island), area only, and intercept only. data(galap) galap$t <- rgamma(16, 5, scale = 2)#add a random time variable gdm(data = galap, model = "loga", mod_sel = TRUE)
SAR extrapolation
Extrapolation (e.g. predicting the richness of areas too large to be sampled) is one of the primary uses of the SAR. Using an individual SAR model fit, extrapolation simply involves using the model function in combination with the parameter values estimated using the observed area / richness data to predict the richness on a larger area. An alternative extrapolation approach to simply using an individual model l is to use multi-model inference and model averaging, whereby a larger number of n models is fitted to a set of islands, the models ranked according to some criterion (e.g. Akaike’s information criterion, AIC) and the criterion values converted into model weights (i.e. the conditional probabilities for each of the n models). The n models are then each used to predict the richness of a larger area and these predictions are multiplied by the respective model weights and summed to provide a multi-model averaged prediction. Both of these options are available inside the package.
We would recommend using grid_start = "exhaustive" when fitting models for use with sar_pred(), as this is more likely to find the optimum fit for a given model. However, remember that, as grid_start has a random component, when grid_start != "none", you can get slightly different results each time you fit a model or run sar_average, and then run sar_pred().
#fit the power model and predict richness on an island of area = 5000 data(galap) p <- sar_power(data = galap) sar_pred(p, area = 5000) #fit three SAR models and predict richness on islands of area = 5000 & 10000 p2 <- sar_multi(galap, obj = c("power", "loga", "koba")) sar_pred(p2, area = c(5000, 10000)) #calculate a multi-model curve and predict richness on islands of area = 5000 & 10000 #grid_start set to "none" for speed p3 <- sar_average(data = galap, grid_start = "none") sar_pred(p3, area = c(5000, 10000))
SAR thresholds
An increasing number of studies have focused on identifying thresholds in the ISAR, including studies i) focused on identifying the small island effect, ii) determining whether the ISAR has an upper asymptote and iii) identifying thresholds in habitat ISARs, which are often systems of conservation concern. The most common approach in such studies is to use piecewise regression. As such, the package provide a set of functions for fitting six piecewise regression models to ISAR data, calculating confidence intervals around the breakpoint estimates (for certain models), comparing the models using various information criteria, and plotting the resultant model fits. The six piecewise models are a selection of 6 models out of the 14 listed by @Gao2019 and comprise the continuous one-threshold and two-threshold, discontinuous one-threshold and two-threshold, and left-horizontal one-threshold and two-threshold models (see @Gao2019 for further information).
The main function is ‘sar_threshold’, which fits the six piecewise models, in addition to a linear model and an intercept only model for comparison (using the ‘non_th_models’ argument). Summary and plot generic functions are available which return user-friendly output. As the coefficients in the fitted breakpoint regression models do not all represent the intercepts and slopes of the different segments (for these it is necessary to add different coefficients together), a separate function (‘get_coef’) can be used to calculate these. A separate function (‘threshold_ci’) generates the confidence intervals around the breakpoints of the one-threshold continuous and left-horizontal models.
Detailed information on fitting and plotting the threshold models, as well as the additional functions for calculating model slopes and intercepts, and generating confidence intervals around the threshold estimates, can be found in @Matthews2020.
#load an example dataset, and fit the continuous two-threshold model #to the data (with area transformed using log to the base 10), using an #interval of 0.1 (for speed) data(aegean2) fit <- sar_threshold(data = aegean2, mod = c("ContTwo"), interval = 0.1, non_th_models = FALSE, logAxes = "area", con = 1, logT = log10, nisl = NULL) #generate model fitting summary table (generally more useful when fitting multiple models) summary(fit) #Plot the resultant model fit plot(fit, cex = 0.8, cex.main = 1, cex.lab = 0.8, pcol = "grey")
On the calculation of information criteria
There are different ways to calculate the various information criteria (IC) used for model comparison (e.g. AIC, BIC). One difference relates to whether the residual sum of squares (rss) or the log-likelihood (ll) is used. Under standard assumptions (e.g. independence of data points, homoscedasticity and normality of the residuals), the two approaches produce identical parameter estimates for regression models. However, the formulas are different and thus can produce different IC values for the same model. For example, historically in the ‘sars’ package we have calculated IC values using formulas based on the rss (Burnham & Anderson, 2002; Guilhaumon et al. 2008). This meant that the IC values generated in ‘sars’ were not comparable with values generated in packages using different formulas. For example, in the nls (the main function for non-linear regression in R) and lm functions in the stats R package, a ll approach is used, meaning IC values from models fitted using nls could not be compared with IC values from sars models. To re-iterate, the parameter estimates are comparable, it is simply that the IC values will differ. In version of ‘sars’ (version 1.2.2) we changed our IC formulas to match those in nls and lm. Thus, if users have built their own models using nls (or lm) and wish to compare IC values with models fitted in ‘sars’, this is now straightforward. To re-create IC values from previous studies (i.e. those using a version of sars pre 1.2.2), it will be necessary to download sars Version 1.2.1 or earlier (either from CRAN or GitHub; version 1.1.1 was published as a release on GitHub). It is important to note that these are not the only two ways of calculating ICs for regression models, and other formulas exits. Thus, if building models using other functions and packages (i.e. other than nls or lm), users should make sure to check how these packages calculate IC values before comparing with models fitted in sars. In sars, as in nls and lm, we include an additional parameter for estimation of the variance. For example, the power model is considered to have three parameters when calculating IC values. Finally, if users are comparing models fitted in sars with their own models fitted using other packages, it is essential that IC values are all calculated using the same dependent variable (i.e. non-transformed richness values and not logged richness, as sars only uses the former currently).
Potential issues with using information criteria in conjunction with
non-linear regression models
Before starting this section, it is worth pointing out that we are not statisticians and the following discussion should be interpreted with this point in mind. We have however reviewed a lot of the literature (including a number of statistics forums) on the use of information criteria (e.g. AIC) in the context of comparing non-linear regression models. Based on this review, we have concluded that there are three potential issues with this approach.
First, many information criteria (including AIC) are theoretically applicable only when a model is fitted through the approach of maximum likelihood. In the case of linear and non-linear regression (e.g. the nls function in R, and the approach we use in ‘sars’), a least-squares estimator (minimising the residual sum of squares) is generally used rather than likelihood maximisation, for various reasons, such as analytical tractability. This is not necessarily a problem, as the least-squares parameter estimates and the maximum likelihood parameter estimates are equivalent in the case of normal and independent errors with a mean of zero and constant variance (@Bates1988). This holds for both linear and non-linear regression models (@Bates1988; @Forum1). However, if these assumptions are not met this equivalence may break down and thus the use of AIC etc would not necessarily be appropriate. The use of the normality and homogeneity of variance checks within the package may be useful in this regard, but note that it is now up to the user to manually turn these checks on.
Second, most of the commonly used information criteria (e.g. AIC) include a term that penalises for the number of parameters. In the case of linear regression models, counting the number of parameters is straight forward. However, non-linear models are so called because they are non-linear in regard to their parameters (e.g. a quadratic model is a linear model but with a non-linear shaped curve). As such, it is not always as straight forward to count parameters in the same way, particularly in the case of very complex non-linear and machine learning type models (i.e. not those used in ‘sars’). “That is, a single parameter in the model may count as more or less than one parameter, in some sense” (@Forum2). However, in the absence of any other information, or indeed alternative model comparison approaches, all one can do is use the number of estimable parameters (sensu Burnham & Anderson) in a model, which is what we follow in the ‘sars’ package, and is what a large number of studies do (and advise) in the ecological and wider literature (e.g. @Archontoulis2015; @Banks2017; see also @Forum3’s response to @Forum2).
Finally, some have argued that you should only use AIC etc to compare models within the same ‘model complex’, and even only in the case of nested models (e.g. @Ripley). From this it could be concluded that we should not compare the suite of 19 non-linear models with the linear model. However, @BurnhamAnderson2002 and @Anderson2006 have argued that this is not the case and AIC for example can be used to compare non-nested models. In addition, again many studies use AIC to compare non-nested models from different ‘complexes’. We have worked on the assumption that, as long as the response variable is the same and models are comparable in terms of parameter number calculation (e.g. all include or not a parameter for the variance), it is appropriate to compare different types of model (at least within the limited selection used in the package).
CONCLUSIONS
The SAR has been a cornerstone of ecological and biogeographical science for almost a century and its form and fit are still of great significance in both theoretical and applied contexts. The development of the ‘sars’ R package should aid future SAR research by providing a comprehensive set of tools that enable in-depth exploration of SARs and SAR-related patterns. In addition, the package has been designed in such a way as to allow other SAR researchers to add (e.g. via GitHub) new functions and models in the future to ensure the package is of lasting value. For example, future additions to the package could include the suite of countryside biogeography SAR models that have recently been published, standard SAR functions not so far incorporated, or functions specifically intended for analysis of species accumulation curves or endemics–area relationships. Finally, whilst the focus of this paper has been on classic SARs, there is no reason that the functionality in the ‘sars’ package cannot be used to analyse other diversity–area relationships (e.g. functional or phylogenetic diversity–area relationships). Application of the full set of 20 models, in addition to the multimodel SAR framework, included in the ‘sars’ package to a wider range and type of data (e.g. trait and phylogenetic data) will likely be revealing and will help in improving our understanding of SARs, and diversity–area relationships more generally.
References
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.