Weighted Standard Error

79 views
Skip to first unread message

Mylo

unread,
Nov 12, 2019, 5:12:30 AM11/12/19
to distance-sampling

Dear distance-sampling group,

 

I have two different density estimates based on line transect sampling for a species within two different parts of a Protected Area (total area ~2,000km2), part 1 is the “outer core” (~1500km2) and part 2 is the inner core (~500km2) which is embedded in the outer core. The areas do not overlap.

 

The density estimate for the outer core was based on a line transect survey in 2010, and that of the inner core in 2011. The scientist used a pooled detection function as the encounter rate was very low for the species, thus the same detection function was used for both density estimates. The density was higher in the inner core compared to the outer core. The survey effort was also higher in the inner core as the same number of transect lines (~40) was used in a smaller area.

 

I would now like to obtain a density estimate for the entire PA by combining the two density estimates. The reason I need this single density estimate is because subsequent surveys in the PA randomly sampled the entire PA and used one density figure for the entire PA. I would like to compare this latter figure to the baseline.

 

I am considering two different approaches.

 

Approach 1: Using an area-weighted mean density (i.e. (D1*1500+D2*500)/2000 = Dtotal). This option has my preference as it is quick and easy, I need the answer fast, and I may have difficulties finding the original datafiles. However, my question is here: how to calculate the (weighted) standard error?

 

Approach 2: Re-do the original analysis. This is obviously more time-consuming and it may be hard to replicate original works exactly as some information is missing. Also, I would then have a non-random survey design where a part of the area has been purposefully sampled much more intensely + and the survey year is not-random, all seems to violate important assumptions.

 

I’d greatly appreciate recommendations J

Best wishes

Milou

Eric R

unread,
Nov 12, 2019, 11:17:46 AM11/12/19
to distance-sampling

A textbook formula under the condition of independent random variables. The variance of a sum is the sum of the variances.  When those sums are weighted, the weights need to be squared.


Note, this is all couched in terms of variance.  What you have for your stratum-specific estimates are likely the standard errors of the stratum-specific density estimates.  You'll need to convert them to variances by squaring. The result you derive from the above formula will be a variance of the weighted sum.  You'll want to square root it.

However, all of the above only holds if the X_i values are independent, which they are not in the scenario you describe (the density estimates for the two areas share a common detection function).

Variance estimation for the situation you describe is laid out in detail in Buckland et al. (2001) Section 3.7.1, pages 89-90.  Let me know if you are unable to locate a copy that provides explicit formulas for your situation.

Mylo

unread,
Nov 13, 2019, 4:47:46 AM11/13/19
to distance-sampling
Dear Eric R

Thank you so much for your fast response! Luckily I have a copy of the book which really helped me well along the way!

You are indeed correct that these are not independent and the author used a common detection function. Going through this exercise did yield a few additional questions however:

1.) The example of Section 3.7.1 p. 89 - 99 is for a point transect survey, whereas 'my' survey is a line transect survey. Does anything changes to the formulas 3.123-3.126 because of this reason? It appears very similar to the examples provided in Section 8.5 p. 360 formulas 8.12? But these are for abundance rather then density and I have an inkling it will be as simple as dividing the variance by the area..

2.) Luckily, the author reported the Effective Strip Width (ESW) and I use 1/ESW to obtain f0. SE's are also provided and so I obtain Variance by squaring. However, I Imagine 1/(Var of ESW) does not equal Var of f0? Do you have a suggestion on how to obtain Varf0 under these circumstances?

3.) For the variance of n and of cluster size I use the formula's provided on p. 108 and p. 120 respectively. For the line transect line (li) I do not use the actual length of the transect i (as it is defined on p. 425) but rather the total length walked on that transect, this is because most transects were walked multiple times and the frequency is different per line. The same for L, I now use the total length in kilometers walked during the survey in that area. Is this the correct way of approaching li and L? for k I still use the total number of line transect (I do not multiply it by the times they were walked).

Many thanks in advance!
Milou

Eric Rexstad

unread,
Nov 13, 2019, 5:53:13 AM11/13/19
to Mylo, distance-sampling
Milou

You are on the right track.  Let me address your points:

1)  For Section 3.7.1 you can treat f(0) as equivalent to h(0).
2)  Given you have an estimate of ESW and a measure of its precision; work with CV as the measure of precision.  CV is defined as SE(estimate)/estimate.   The CV of ESW is identical to CV of f(0).  Use that knowledge to solve for var(f(0)).
3) l_i is measured in terms of "effort" so you are correct l_i is the physical length of transect multiplied by the number of times walked.  Total effort (summed across all transects) is sum of effort.  You are right "k" is the physical number of transects, not multiplied by number of times walked.

Wishing you further success.

Eric Rexstad
Senior Research Fellow, CREEM, Univ. of St Andrews, Charity SC013532
--
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.

Mylo

unread,
Nov 19, 2019, 10:39:24 PM11/19/19
to distance-sampling

Dear Eric,

 

Thank you very much, that worked like a charm.

 

I have now obtained the variances for the densities of the different species and therefore the SEs (square root of Var). Then I obtained the 95%CI by applying the Jacknife method explained in formula 3.95 on page 81 (Density ± Critical T value * SE), with 37 degrees of freedom (38 transects) the T-value was 2.026192. Do you think this is an appropriate method to apply in case bootstrapping is not possible like in this specific example? Perhaps the number of line replicates is sufficient to apply 3.71 (p.77) instead?

 

Finally, I obtain the population size for the total area by summing the two estimated population sizes. For the variance I use exactly the same method as applied for the density (Formula 3.123) but substituting D for N.

 

Many thanks again!

Milou


On Wednesday, November 13, 2019 at 5:53:13 PM UTC+7, Eric Rexstad wrote:
Milou

You are on the right track.  Let me address your points:

1)  For Section 3.7.1 you can treat f(0) as equivalent to h(0).
2)  Given you have an estimate of ESW and a measure of its precision; work with CV as the measure of precision.  CV is defined as SE(estimate)/estimate.   The CV of ESW is identical to CV of f(0).  Use that knowledge to solve for var(f(0)).
3) l_i is measured in terms of "effort" so you are correct l_i is the physical length of transect multiplied by the number of times walked.  Total effort (summed across all transects) is sum of effort.  You are right "k" is the physical number of transects, not multiplied by number of times walked.

Wishing you further success.

Eric Rexstad
Senior Research Fellow, CREEM, Univ. of St Andrews, Charity SC013532
To unsubscribe from this group and stop receiving emails from it, send an email to distance...@googlegroups.com.

Eric Rexstad

unread,
Nov 20, 2019, 7:11:59 AM11/20/19
to Mylo, distance-sampling

Milou

That all seems quite thorough to me.  I don't expect there to be much difference between the confidence intervals produced by 3.95 and 3.71.  The difference will be that the confidence interval using 3.71 will be asymmetrical (point estimate not in the centre of the confidence interval) because 3.71 uses a log-normal approximation to the sampling distribution where as 3.95 uses a normal approximation.  The log-normal approximation is probably better, but you are probably working at the margins of the tolerances of interest to you.

To unsubscribe from this group and stop receiving emails from it, send an email to distance-sampl...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/e71183c8-f6e1-45b3-930b-713df6bdaf78%40googlegroups.com.
-- 
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