npmr: NPMR, Nonparametric multiplicative regression

In phytomosaic/ngnn: ngnn: Nonlinear Gradient Nearest Neighbors

Description

Nonparametric kernel regression with automatic interactions among multiple predictors, adapted here for in-sample estimation AND out-of-sample prediction. Also known as 'generalized product kernel' regression in econometrics.

Usage

npmr(spe, idi, ido, nm, nmulti, pa, pr, ...)

npmr_sens(obj, pick, nm, ...)

## S3 method for class 'npmr'
summary(obj, ...)

## S3 method for class 'npmr'
plot(obj, pick = NULL, nm, zlim, ...)

Arguments

spe	species dataframe, rows = sample units and columns = species
idi	in-sample predictor dataframe, rows must match 'spe'
ido	out-of-sample predictor dataframe, where rows = new sample units
nm	string vector specifying predictors to include (max 2)
nmulti	number of random starts in nonparametric regression
pa	logical, convert to presence/absence?
pr	logical, use 'beals' for probs of joint occurrence?
...	additional arguments passed to function
obj	object of class 'npmr' from call to npmr
pick	variable to query
zlim	vector of length 2, giving vertical limits for plots

Details

NPMR is nonparametric multiplicative regression (McCune 2006). This implementation uses a variant npreg from package 'np', where it is known as generalized product kernel regression (Li and Racine 2007). The current function is used in a predictive capacity for estimating environmentally constrained fitted values for NGNN, and therefore requires a predictor matrix for out-of-sample observations.

Sensitivity analysis is available, following Eqn 9 in McCune (2006:825), aka Sensitivity Formula 1 in Hyperniche software. Sensitivity (Q) is the mean absolute difference in the response resulting from nudging the predictors +/- 5 proportion of the range of the response variable. Q = 1 means that nudging a predictor results in a change in response of equal magnitude, and Q = 0 means that nudging a predictor has no detectable effect on the response.

Value

List of class 'npmr' with elements:

• spe = original species matrix

• id_i = in-sample predictors used in NPMR

• nm = which predictors were used

• nm_len = their length

• iYhat = in-sample fitted values from NPMR

• oYhat = out-of-sample fitted values from NPMR

• np_stat = fit, tolerances and results of signif tests

• np_mods = list of NPMR regression models for every species

• np_bw = list of NPMR bandwidths for every species

References

Li, Q. and J. S. Racine. 2007. Nonparametric Econometrics: Theory and Practice, Princeton University Press.

McCune, B. 2006. Non-parametric habitat models with automatic interactions. Journal of Vegetation Science 17(6):819-830.

Examples

# set up
set.seed(978)
require(vegan)
data(varespec, varechem)
spe <- varespec ; id  <- varechem
i   <- sample(1:nrow(spe), size=floor(0.75*nrow(spe))) # sample
spe <- spe[i, ]          # in-sample species
idi <- id[i, ]           # in-sample predictors
ido <- id[-i, ]          # out-of-sample predictors
nm  <- c('Al', 'K')      # select 1 or 2 gradients of interest


# NPMR basic usage
res_npmr <- npmr(spe, idi, ido, nm, nmulti=5)
summary(res_npmr)

# sensitivity analysis
Q <- npmr_sens(obj=res_npmr, pick = 'Vaccmyrt', nm)
Q

# plot NPMR species response curves
plot(res_npmr, pick=4, nm)   # alternatively: pick='Vaccmyrt'
plot(res_npmr, pick=1:9, nm)

phytomosaic/ngnn documentation built on May 9, 2019, 5:57 a.m.