poptrend (version 0.2.0)

ptrend: Fit a smooth or linear trend to count survey data.

Description

The function estimates a trend from count survey data.

Usage

ptrend(
  formula,
  data = list(),
  family = quasipoisson(),
  nGrid = 500,
  nBoot = 500,
  bootType = "hessian",
  gamModel = TRUE,
  engine = "gam",
  ...
)

Value

An object of class trend.

Arguments

formula

A trend formula. This is a GAM formula with an extra term trend describing the time variable and properties of the trend. All terms except the trend term are treated as covariates. Effect of temporal variation in these covariates are not included in the calculation of the trend.

data

A data frame containing response variables and covariates.

family

The distributional family of the response. The family most use a log-link, it defaults to a quasi-Poisson.

nGrid

The number of grid points over which to compute the trend. If the length of the argument is one, an equally spaced grid over the survey period of length nGrid is set up. nGrid can also be a vector of length 3, in which case the first element is the number of grid points and the second and third elements give, respectively, the start and endpoints of the grid.

nBoot

The number of bootstrap samples to draw.

bootType

Only one method, "hessian", currently implemented. Type "hessian", draws bootstrap samples using the Bayesian covariance matrix of the parameters (see vcov.gam).

gamModel

If true, the fit of the underlying gam model is saved in the output. May be set to FALSE to save memory, but with the side effect that the fit of the gam model cannot be checked.

engine

If 'gam', the default, model fitting is done via gam. If 'bam', model fitting is done via bam, which is less memory hungry and can be faster for large data sets.

...

Further arguments passed to gam.

Author

Jonas Knape

Details

The function estimates smooth or loglinear population trends, or indexes from simple design count survey data. It is essentially a wrapper around a call to gam, processing its output using predict.gam to produce a trend estimate. For smooth trends, cubic regression splines for the temporal variable are set up by the term s(var, k = k, fx = fx , bs = "cr") where var is the first argument to trend in the formula. For loglinear trends, the identity of var is used, and for index models a factor variable is constructed from var.

Temporal random effects are set up by converting the temporal variable supplied to trend to a factor variable and adding this factor variable to the data supplied to gam.

Bootstrap confidence intervals are computed by drawing normally distributed random variable with means equal to the estimated coefficients and covariance matrix equal to the Bayesian posterior covariance matrix (see vcov.gam).

Examples

## Simulate a data set with 15 sites and 25 years
data = simTrend(15, 25)
## Fit a smooth trend with fixed site effects, random time effects,
## and automatic selection of degrees of freedom
trFit = ptrend(count ~ trend(year, tempRE = TRUE, type = "smooth") + site, data = data)
## Check the model fit
checkFit(trFit)
## Plot the trend
plot(trFit)
summary(trFit)
## Check the estimated percent change from year 8 to 25
change(trFit, 8, 25)

## Fit a loglinear trend model with random site effects and random time effects 
## to the same data set.
trLin = ptrend(count ~ trend(year, tempRE = TRUE, type = "loglinear") +
                 s(site, bs = "re"), data = data)
plot(trLin)
summary(trLin)

## Fit an index model with fixed site effects and an (unrelated) continous covariate 
## as a smooth effect.
# Simulate mock covariate unrelated to data.
cov = rnorm(nrow(data))
trInd = ptrend(count ~ trend(year, type = "index") + site + s(cov), data = data)
plot(trInd)
summary(trInd)