This function calculates generalized linear models for a set of (species) presence/absence records in a data frame, with a wide set of options for data partition, variable selection, and output form.
multGLM(data, sp.cols, var.cols, id.col = NULL, family = "binomial",
test.sample = 0, FDR = FALSE, correction = "fdr", corSelect = FALSE,
cor.thresh = 0.8, step = TRUE, trace = 0, start = "null.model",
direction = "both", select = "AIC", trim = TRUE, Y.prediction = FALSE,
P.prediction = TRUE, Favourability = TRUE, group.preds = TRUE,
TSA = FALSE, coord.cols = NULL, degree = 3, verbosity = 2, ...)a data frame in wide format (see splist2presabs) containing, in separate columns, your species' binary (0/1) occurrence data and the predictor variables.
index numbers of the columns containing the species data to be modelled.
index numbers of the columns containing the predictor variables to be used for modelling.
(optional) index number of column containing the row identifiers (if defined, it will be included in the output predictions data frame).
argument to be passed to the glm function; only 'binomial' is implemented in multGLM so far.
a subset of data to set aside for subsequent model testing. Can be a value between 0 and 1 for a proportion of the data to choose randomly (e.g. 0.2 for 20%), or an integer number for a particular number of cases to choose randomly among the records in data, or a vector of integers for the index numbers of the particular rows to set aside, or "Huberty" for his rule of thumb based on the number of variables (Huberty 1994, Fielding & Bell 1997).
logical value indicating whether to do a preliminary exclusion of variables based on the false discovery rate (see FDR). The default is FALSE.
logical value indicating whether to do a preliminary exclusion of highly correlated variables (see corSelect). The default is FALSE.
numerical value indicating the correlation threshold to pass to corSelect (used only if corSelect = TRUE).
logical, whether to use the step function to perform a stepwise variable selection (based on AIC or BIC).
if positive, information is printed during the running of step. Larger values may give more detailed information.
character, whether to start with the 'null.model' (so that variable selection starts forward) or with the 'full.model' (so selection starts backward). Used only if step = TRUE.
argument to be passed to step specifying the direction of variable selection ('forward', 'backward' or 'both'). Used only if step = TRUE.
character string specifying the criterion for stepwise selection of variables. Options are "AIC" (Akaike's Information Criterion; Akaike, 1973), the default; or BIC (Bayesian Information Criterion, also known as Schwarz criterion, SBC or SBIC; Schwarz, 1978). Used only if step = TRUE.
logical, whether to trim non-significant variables off the models using the modelTrim function; can be used whether or not step is TRUE; works as a backward variable elimination procedure based on significance.
logical, whether to include output predictions in the scale of the predictor variables (type = "link" in predict.glm).
logical, whether to include output predictions in the scale of the response variable, i.e. probability (type = "response" in predict.glm).
logical, whether to apply the Favourability function to extract the effect of prevalence on probability (Real et al. 2006) and include its results in the output.
logical, whether to group together predictions of similar type (Y, P or F) in the output predictions table (e.g. if FALSE: sp1_Y, sp1_P, sp1_F, sp2_Y, sp2_P, sp2_F; if TRUE: sp1_Y, , sp2_Y, sp1_P, sp2_P, sp1_F, sp2_F).
logical, whether to add a trend surface analysis (calculated individually for each species) as a spatial variable in each model. See multTSA for more details. The default is FALSE.
argument to pass to multTSA (if TSA=TRUE).
argument to pass to multTSA (if TSA=TRUE).
integer value indicating the amount of messages to display; currently implemented values are from 0 to 2 (the default).
additional arguments to be passed to modelTrim.
This function returns a list with the following components:
a data frame with the model predictions (if either of Y.prediction, P.prediction or Favourability are TRUE).
a list of the resulting model objects.
This function automatically calculates binomial GLMs for one or more species (or other binary variables) in a data frame. The function can optionally perform stepwise variable selection (and it does so by default) instead of forcing all variables into the models, starting from either the null model (the default, so selection starts forward) or from the full model (so selection starts backward) and using Akaike's information criterion (AIC) as a variable selection criterion. Instead or subsequently, it can also perform stepwise removal of non-significant variables from the models using the modelTrim function.
There is also an optional preliminary selection of non-correlated variables, and/or of variables with a significant bivariate relationship with the response, based on the false discovery rate (FDR). Note, however, that some variables can be significant in a multivariate model even if they would not have been selected by FDR.
Favourability is also calculated, removing the effect of species prevalence from occurrence probability and thus allowing direct comparisons between models (Real et al. 2006).
By default, all data are used in model training, but you can define an optional test.sample to be reserved for model testing afterwards. You may also want to do a previous check for multicollinearity among variables, e.g. the variance inflation factor (VIF).
The multGLM function will create a list of the resulting models (each with the name of the corresponding species column) and a data frame with their predictions (Y, P and/or F, all of which are optional). If you plan on representing these predictions in a GIS based on .dbf tables, remember that dbf only allows up to 10 characters in column names; multGLM predictions will add 2 characters (_Y, _P and/or _F) to each of your species column names, so use species names/codes with up to 8 characters in the data set that you are modelling. You can create (sub)species name abbreviations with the spCodes function.
Akaike, H. (1973) Information theory and an extension of the maximum likelihood principle. In: Petrov B.N. & Csaki F., 2nd International Symposium on Information Theory, Tsahkadsor, Armenia, USSR, September 2-8, 1971, Budapest: Akademiai Kiado, p. 267-281.
Fielding A.H. & Bell J.F. (1997) A review of methods for the assessment of prediction errors in conservation presence/absence models. Environmental Conservation 24: 38-49
Huberty C.J. (1994) Applied Discriminant Analysis. Wiley, New York, 466 pp. Schaafsma W. & van Vark G.N. (1979) Classification and discrimination problems with applications. Part IIa. Statistica Neerlandica 33: 91-126
Real R., Barbosa A.M. & Vargas J.M. (2006) Obtaining environmental favourability functions from logistic regression. Environmental and Ecological Statistics 13: 237-245.
Schwarz, G.E. (1978) Estimating the dimension of a model. Annals of Statistics, 6 (2): 461-464.
# NOT RUN {
data(rotif.env)
names(rotif.env)
# make models for 3 of the species in rotif.env:
mods <- multGLM(rotif.env, sp.cols = 45:47, var.cols = 5:17, id.col = 1,
step = TRUE, FDR = TRUE, trim = TRUE)
names(mods)
head(mods$predictions)
names(mods$models)
mods$models[[1]]
mods$models[["Ttetra"]]
# }
Run the code above in your browser using DataLab