re: poor precision for one strata
47 views
Skip to first unread message
Asia Murphy
Jan 28, 2016, 10:00:19 AM1/28/16
to distance-sampling
Hello all,
I'm attempting to examine trends in lemur density, both how they are influenced by habitat degradation, and what their trend is over time. For the habitat degradation, I am stratifying by discrete classes (A = intact, B = intermediate, C = degraded), and for the annual trends, I have one site that has 4 years' worth of data, and another site with 3 years' worth of data.
My problem is that it always seems to be the last stratum that has an issue with precision. For example, for the habitat degradation analysis, the degraded site has the fewest lemur observations (around 13). I would assume that estimating density using a strata-specific detection probability would decrease precision, and that's what happens.
However, even when running a global detection probability, where presumably detection is being pooled from all the observations, I still have terrible precision for the degraded site (see attached bar chart).
I could understand that since there are so few observations, but with my examination of annual trends, I'm facing the same problem, when the last year actually has the most observations (32), compared to the other years (see screenshot attached). The first three years, even with less than 30 observations for each year, have perfectly fine precision and confidence intervals, but the last year, with the most observations, has an absolutely horrendous precision estimate.
I don't know what's going on, and I would very much appreciate any help that anyone can give me.
Eric Rexstad
Jan 28, 2016, 10:32:07 AM1/28/16
to Asia Murphy, distance-sampling
A handful of issues here Asia.
First, precision of density estimates from distance sampling is comprised of two bits: how uncertain you are about the shape of the detection function, which enlightens you about the confidence you have in the proportion of animals you missed within the region covered by your transects; and variability in animals encountered on you replicate transects. This second component informs you about confidence you can have extrapolating from transects you visited to the entire study area.
Imagine your replicate transects as boxes of chocolates. If you open a box 1 and find 12 chocolates, then open box 2 and find 12 and box 3 and find 12, how many chocolates would you guess there to be in each of a stack of boxes you haven't opened? In contrast if your sample of 3 boxes possessed 25, 3 and 15 chocolates, respectively, how confident are you about the contents of unopened boxes.
I speculate it is this encounter rate variance that is responsible for the poor precision in the degraded habitat and the last year.
I know nothing about lemurs, but if the degradation you describe is patchy, with some of your replicate transects in remnants that have not yet been degraded and other replicates degraded, that may heighten the difference in number of encounters between transects.
Look through the Distance output and find the page labelled "Density Estimates/Global" and at the bottom you will see a table labelled "Component percentages of Var(D)". This will tell you what piece (detection function/encounter rate/(maybe)cluster size) contributes to the uncertainty in your overall estimates.
Other comments: using a global detection function to apply to make stratum-specific estimates results in an inappropriate detection function being applied to each stratum, leading to some bias. If the strata are spatial, that means you are using detections from, say, degraded habitats in your detection function for your intact habitats (and vice versa). In addition, if your strata are temporal, then applying a single global detection function to produce annual estimates induces dependence between the annual estimates mucking up the precision estimates of the individual estimates.
I suspect you are on the edge with respect to number of detections for fitting your detection function as well as the amount of replicate transects you have. Hence the uncertainty you have in your estimates are challenging your ability to see clear patterns.
First, precision of density estimates from distance sampling is comprised of two bits: how uncertain you are about the shape of the detection function, which enlightens you about the confidence you have in the proportion of animals you missed within the region covered by your transects; and variability in animals encountered on you replicate transects. This second component informs you about confidence you can have extrapolating from transects you visited to the entire study area.
Imagine your replicate transects as boxes of chocolates. If you open a box 1 and find 12 chocolates, then open box 2 and find 12 and box 3 and find 12, how many chocolates would you guess there to be in each of a stack of boxes you haven't opened? In contrast if your sample of 3 boxes possessed 25, 3 and 15 chocolates, respectively, how confident are you about the contents of unopened boxes.
I speculate it is this encounter rate variance that is responsible for the poor precision in the degraded habitat and the last year.
I know nothing about lemurs, but if the degradation you describe is patchy, with some of your replicate transects in remnants that have not yet been degraded and other replicates degraded, that may heighten the difference in number of encounters between transects.
Look through the Distance output and find the page labelled "Density Estimates/Global" and at the bottom you will see a table labelled "Component percentages of Var(D)". This will tell you what piece (detection function/encounter rate/(maybe)cluster size) contributes to the uncertainty in your overall estimates.
Other comments: using a global detection function to apply to make stratum-specific estimates results in an inappropriate detection function being applied to each stratum, leading to some bias. If the strata are spatial, that means you are using detections from, say, degraded habitats in your detection function for your intact habitats (and vice versa). In addition, if your strata are temporal, then applying a single global detection function to produce annual estimates induces dependence between the annual estimates mucking up the precision estimates of the individual estimates.
I suspect you are on the edge with respect to number of detections for fitting your detection function as well as the amount of replicate transects you have. Hence the uncertainty you have in your estimates are challenging your ability to see clear patterns.
--
You received this message because you are subscribed to the Google Groups "distance-sampling" group.
To unsubscribe from this group and stop receiving emails from it, send an email to distance-sampl...@googlegroups.com.
To post to this group, send email to distance...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/8237e504-ecca-4911-9a2b-57f2d6aa5fd2%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
-- Eric Rexstad Research Unit for Wildlife Population Assessment Centre for Research into Ecological and Environmental Modelling University of St. Andrews St. Andrews Scotland KY16 9LZ +44 (0)1334 461833 The University of St Andrews is a charity registered in Scotland : No SC013532
Reply all
Reply to author
Forward
0 new messages