Scott, this is a pragmatic solution. Variance estimation methods in
Distance are remarkably robust to non-independence. In fact, if you use
the bootstrap option based on lines or points, you don't assume that
detections are independent at all. So there is nor real problem for
variance estimation. But the following aspects of Distance are NOT
robust to non-independence: goodness-of-fit (including
Kolmogorov-Smirnov and Cramer-von Mises) and AIC. Hence there is a
tendency to select models with too many parameters. To avoid this, you
can conduct model selection on the analysis of groups, where independent
detections is a more reasonable assumption. You can then fit the
selected model to an analysis of individuals.
Personally, I don't think this is the best solution to your problem
though. If you have data as clusters, why not just do the default
analysis, where you analyse clusters, and separately estimate mean
cluster size? The default regression methods for this in Distance
corrects for size bias. You could also consider using the MCDS engine,
with cluster size as a covariate. There are limitations on what is
currently implemented in Distance in this case, e.g. if you have strata,
but it may do what you need.
Steve Buckland
scrosbie wrote:
> > Hello distance samplers,
> >
> > I have collected point transect data using clusters as the recording unit and want to correct for size bias. The "Introduction to Distance Sampling" book (pg 167) states that one method is to replace clusters with individuals but then the assumption that detections are independent is violated. I beleive I understand how to conduct the solution - analyze the clustered data then refit the same model to the individual data. But, I don't understand how this solves the problem of independence of detections.
> >
> > If anyone can provide information on how this method solves the problem of independence of detections I would greatly appreciate it. I do not yet have a copy of the "Advanced Guide to DS" and a search of the email archive did not come up with anything. Also, is there is any thought that stratification by cluster size would be a better approach?
--
_____________________________________________________________________
Stephen T. Buckland
CREEM, The Observatory, Buchanan Gdns, St Andrews KY16 9LZ, Scotland
Tel.
+44-(0)1334-461841 Secretary
+44-(0)1334-461842
Fax
+44-(0)1334-461800 e-mail
st...@mcs.st-and.ac.uk
Professor of Statistics
http://www.mcs.st-and.ac.uk/
Co-Director, NCSE
http://www.ncse.org.uk/
CREEM
http://www.creem.st-and.ac.uk/
_____________________________________________________________________