sar_average: Fit a multimodel averaged SAR curve

In txm676/mmSAR2: Fit and Compare Species-Area Relationship Models Using Multimodel Inference

Fit a multimodel averaged SAR curve

Description

Construct a multimodel averaged species-area relationship curve using information criterion weights and up to twenty SAR models.

Usage

sar_average(obj = c("power",
  "powerR","epm1","epm2","p1","p2","loga","koba",
  "monod","negexpo","chapman","weibull3","asymp",
  "ratio","gompertz","weibull4","betap","logistic", "heleg", "linear"), data =
  NULL, crit = "Info", normaTest = "none", homoTest = "none", homoCor =
  "spearman", neg_check = FALSE, alpha_normtest = 0.05, alpha_homotest =
  0.05, grid_start = "partial", grid_n = NULL, confInt = FALSE, ciN = 100,
  verb = TRUE, display = TRUE)

Arguments

obj	Either a vector of model names or a fit_collection object created using sar_multi. If a vector of names is provided, sar_average first calls sar_multi before generating the averaged multimodel curve.
data	A dataset in the form of a dataframe with two columns: the first with island/site areas, and the second with the species richness of each island/site. If obj is a fit_collection object, data should be NULL.
crit	The criterion used to compare models and compute the model weights. The default crit = "Info" switches to AIC or AICc depending on the number of data points in the dataset. AIC (crit = "AIC") or AICc (crit = "AICc") can be chosen regardless of the sample size. For BIC, use crit ="Bayes".
normaTest	The test used to test the normality of the residuals of each model. Can be any of "lillie" (Lilliefors Kolmogorov-Smirnov test), "shapiro" (Shapiro-Wilk test of normality), "kolmo" (Kolmogorov-Smirnov test), or "none" (no residuals normality test is undertaken; the default).
homoTest	The test used to check for homogeneity of the residuals of each model. Can be any of "cor.fitted" (a correlation of the squared residuals with the model fitted values), "cor.area" (a correlation of the squared residuals with the area values), or "none" (no residuals homogeneity test is undertaken; the default).
homoCor	The correlation test to be used when homoTest != "none". Can be any of "spearman" (the default), "pearson", or "kendall".
neg_check	Whether or not a check should be undertaken to flag any models that predict negative richness values.
alpha_normtest	The alpha value used in the residual normality test (default = 0.05, i.e. any test with a P value < 0.05 is flagged as failing the test).
alpha_homotest	The alpha value used in the residual homogeneity test (default = 0.05, i.e. any test with a P value < 0.05 is flagged as failing the test).
grid_start	Should a grid search procedure be implemented to test multiple starting parameter values. Can be one of 'none', 'partial' or 'exhaustive' The default is set to 'partial'.
grid_n	If grid_start = exhaustive, the number of points sampled in the starting parameter space (see details).
confInt	A logical argument specifying whether confidence intervals should be calculated for the multimodel curve using bootstrapping.
ciN	The number of bootstrap samples to be drawn to calculate the confidence intervals (if confInt == TRUE).
verb	verbose - Whether or not to print certain warnings (default: verb == TRUE)
display	Show the model fitting output and related messages. (default: display == TRUE).

Details

The multimodel SAR curve is constructed using information criterion weights (see Burnham & Anderson, 2002; Guilhaumon et al. 2010). If obj is a vector of n model names the function fits the n models to the dataset provided using the sar_multi function. A dataset must have four or more datapoints to fit the multimodel curve. If any models cannot be fitted they are removed from the multimodel SAR. If obj is a fit_collection object (created using the sar_multi function), any model fits in the collection which are NA are removed. In addition, if any other model checks have been selected (i.e. residual normality and heterogeneity tests, and checks for negative predicted richness values), these are undertaken and any model that fails the selected test(s) is removed from the multimodel SAR. The order of the additional checks inside the function is (if all are turned on): normality of residuals, homogeneity of residuals, and a check for negative fitted values. Once a model fails one test it is removed and thus is not available for further tests. Thus, a model may fail multiple tests but the returned warning will only provide information on a single test. We have now changed the defaults so that no checks are undertaken, so it is up to the user to select any checks if appropriate.

The resultant models are then used to construct the multimodel SAR curve. For each model in turn, the model fitted values are multiplied by the information criterion weight of that model, and the resultant values are summed across all models (Burnham & Anderson, 2002). Confidence intervals can be calculated (using confInt) around the multimodel averaged curve using the bootstrap procedure outlined in Guilhaumon et al (2010).The procedure transforms the residuals from the individual model fits and occasionally NAs / Inf values can be produced - in these cases, the model is removed from the confidence interval calculation (but not the multimodel curve itself). There is also a constraint within the procedure to remove any transformed residuals that result in negative richness values. When several SAR models are used, when grid_start is turned on and when the number of bootstraps (ciN) is large, generating the confidence intervals can take a (very) long time. Parallel processing will be added to future versions.

Choosing starting parameter values for non-linear regression optimisation algorithms is not always straight forward, depending on the data at hand. In the package, we use various approaches to choose default starting parameters. However, we also use a grid search process which creates a large array of different possible starting parameter values (within certain bounds) and then randomly selects a proportion of these to test. 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. 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. Note that grid_start has been disabled for a small number of models (e.g. Weibull 3 par.). See the vignette for more information. 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.

Even with grid_start, occasionally a model fit will be able to be fitted and pass the model fitting checks (e.g. residual normality) but the resulting fit is nonsensical (e.g. a horizontal line with intercept at zero). Thus, it can be useful to plot the resultant 'multi' object to check the individual model fits. To re-run the sar_average function without a particular model, simply remove it from the obj argument.

The sar_models() function can be used to bring up a list of the 20 model names. display_sars_models() generates a table of the 20 models with model information.

Value

If no models have been successfully fitted and passed the model checks, an error is returned. If only a single model is successfully fitted, this individual model fit object (of class 'sars') is returned, given no model averaging can be undertaken. If more than two models have been successfully fitted and passed the model checks, a list of class "multi" and class "sars" with two elements. The first element ('mmi') contains the fitted values of the multimodel sar curve. The second element ('details') is a list with the following components:

• mod_names Names of the models that were successfully fitted and passed any model check

• fits A fit_collection object containing the successful model fits

• ic The information criterion selected

• norm_test The residual normality test selected

• homo_test The residual homogeneity test selected

• alpha_norm_test The alpha value used in the residual normality test

• alpha_homo_test The alpha value used in the residual homogeneity test

• ics The information criterion values (e.g. AIC values) of the model fits

• delta_ics The delta information criterion values

• weights_ics The information criterion weights of each model fit

• n_points Number of data points

• n_mods The number of successfully fitted models

• no_fit Names of the models which could not be fitted or did not pass model checks

• convergence Logical value indicating whether optim model convergence code = 0, for each model

The summary.sars function returns a more useful summary of the model fit results, and the plot.multi plots the multimodel curve.

Note

There are different types of non-convergence and these are dealt with differently in the package. 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) - as arguably a non-optimal fit is still better than no fit - 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.

The generation of confidence intervals around the multimodel curve (using confInt == TRUE), may throw up errors that we have yet to come across. Please report any issues to the package maintainer.

There are different formulas for calculating the various information criteria (IC) used for model comparison (e.g. AIC, BIC). For example, some formulas use the residual sum of squares (rss) and others the log-likelihood (ll). Both are valid approaches and will give the same parameter estimates, but it is important to only compare IC values that have been calculated using the same approach. For example, the 'sars' package used to use formulas based on the rss, while the nls function function in the stats package uses formulas based on the ll. To increase the compatibility between nls and sars, we have changed our formulas such that now our IC formulas are the same as those used in the nls function. See the "On the calculation of information criteria" section in the package vignette for more information.

The mmf model was found to be equivalent to the He & Legendre logistic, and so the former has been deprecated (as of Feb 2021). We have removed it from the default models in sar_average, although it is still available to be used for the time being (using the obj argument). The standard logistic model has been added in its place, and is now used as default within sar_average.

References

Burnham, K. P., & Anderson, D. R. (2002). Model selection and multi-model inference: a practical information-theoretic approach (2nd ed.). New-York: Springer.

Guilhaumon, F., Mouillot, D., & Gimenez, O. (2010). mmSAR: an R-package for multimodel species-area relationship inference. Ecography, 33, 420-424.

Matthews, T. J., K. A. Triantis, R. J. Whittaker, & F. Guilhaumon. (2019). sars: an R package for fitting, evaluating and comparing species–area relationship models. Ecography, 42, 1446–55.

Examples

data(galap)
#attempt to construct a multimodel SAR curve using all twenty sar models
#using no grid_start just for speed here (not recommended generally)
fit <- sar_average(data = galap, grid_start = "none")
summary(fit)
plot(fit)

# construct a multimodel SAR curve using a fit_collection object
ff <- sar_multi(galap, obj = c("power", "loga", "monod", "weibull3"))
fit2 <- sar_average(obj = ff, data = NULL)
summary(fit2)

## Not run:  
# construct a multimodel SAR curve using a more exhaustive set of starting 
# parameter values 
fit3 <- sar_average(data = galap, grid_start = "exhaustive", grid_n = 1000) 

## End(Not run)

txm676/mmSAR2 documentation built on Nov. 21, 2024, 5:03 a.m.