Quantifying population trends from CBC counts
R M Fewster and S T Buckland, University of St Andrews
Report to the BTO
MAFF contract CSA3109: The effects of agricultural change on bird populations
Summary
Methods for the analysis of CBC counts are briefly reviewed, including the
chain
method, the Mountford method and log-linear Poisson regression using TRIM.
A method based on generalized additive models (GAMs) is preferred because:
- The TRIM method without autocorrelation is a special case, in
which no smoothing is performed.
- The need for the autocorrelation correction used by TRIM can be
overcome by means of the bootstrap.
- The user can control the degree of smoothing from none (i.e.
TRIM) at one extreme to a linear trend (on some scale) at the
other.
- GAMs allow flexibility in choice of error distribution and link
function.
- An autocovariate term can be added to identify trend in
favourability of conditions for each species.
- Further terms are easily modelled and tested, for example
a region by time interaction or habitat covariates.
- Unlike the chain method, GAMs make efficient use of the data.
- Unlike the Mountford method, GAMs yield robust estimation
over long time periods, so that no 'windowing' is required.
A procedure for identifying change points in the smoothed trend, using
second derivatives, is proposed.
- Back
to Trend analysis using generalized additive models
- Back
to Rachel Fewster's home page
Last updated: 20th February 1997