lassoGaussian: Perform LASSO regularized gaussian regressions to model the...

In nicholasjclark/BBS.occurrences: BBS community composition and co-occurrence analysis

Description

lassoGaussian runs cross-validated regularized regressions for a single species and returns important coefficients

Usage

lassoGaussian(response, covariates, cutoff, n_reps)

lassoAbund_comm(outcome_data, binary_data, outcome_indices, covariates,
  cutoff, n_reps, n_cores)

Arguments

response	A vector of count observations for the focal species
covariates	A matrix of covariates, with nrow(covariates) == nrow(response)
cutoff	Positive numeric value representing the proportion of models in which a predictor must be retained in order to be treated as meaningful. If the predictor is retained in fewer than cutoff proportion of cv.glmnet regularized models, its coefficient is forced to be zero. If the predictor is retained in at least cutoff proportion of models, its mean coefficient is returned. Default is 0.80
n_reps	Positive integer representing the number of times to repeat 10-fold cv.glmnet regularized regressions (default is 10)
outcome_data	A dataframe containing count observations for species (each column representing a different species)
binary_data	A dataframe containing binary presence-absence observations for species (each column representing a different species)
outcome_indices	A sequence of positive integers representing the column indices in outcome_data that are to be modelled as poisson outcome variables (i.e. species counts). Each one of these columns will be treated as a separate species whose abundance is to be modelled using lassoGaussian. Default is to run models for all columns in outcome_data
n_cores	Positive integer stating the number of processing cores to split the job across. Default is parallel::detect_cores() - 1

Details

Regularized regressions are performed to identify meaningful predictors of the species' scaled abundance using cv.glmnet. These models use coordinated gradient descent, applied to training sets of the data, to identify regression parameters. These parameters are predicted on the remaining subset of the data (the test set) to assess model fit. The process is repeated until a best-fitting model is identified (minimising the loss function, which is cross-validated deviance in this case). By replicating the process n_reps times, we account for uncertainty in the fold generating process and can more confidently identify meaningful predictors (i.e. those that are retained in at least cutoff proportion of n_reps models)

Value

lassoGaussian returns a single vector of coefficients for predictors in covariates.

lassoAbund_comm binds these coefficient vectors into a dataframe with rownames matching species names in outcome_data. It then returns a list containing coefficients and scaling factors, which are used in predictive functions

See Also

cv.glmnet

nicholasjclark/BBS.occurrences documentation built on July 19, 2020, 8:31 p.m.