Table S2 of Ganley et al. (2019): the goodness-of-fit and average detection probability of a separately fitted detection function for each year from 1998 to 2017, from aerial line-transect surveys of Cape Cod Bay.
Format
A tibble with 20 rows and 5 columns:
- year
Survey year, 1998 to 2017.
- cvm_p
Cramér-von Mises p-value for the year's detection function.
- ks_p
Kolmogorov-Smirnov p-value.
- p
Average detection probability, \(\hat{P}_a\).
- p_se
Standard error of
p.
Source
Ganley, L.C., Brault, S. and Mayo, C.A. (2019) What we see is not what there is: estimating North Atlantic right whale Eubalaena glacialis local abundance. Endangered Species Research 38:101-113. doi:10.3354/esr00938 , Table S2. Open access under CC-BY.
This is p, not perception bias
p here is the average detection probability of the fitted detection
function — the same quantity selection_table() reports as p, with its
standard error as p_se. It is not a g(0) component. Ganley et al.
say so directly: perception bias "was not addressed directly in this study",
because estimating it needs a second observer team they did not have.
The distinction matters because the two are easy to conflate and correcting
for one while believing you have corrected for the other is how these
estimates go wrong by a factor. See g0(), which will name any component
you have not supplied.
What it is good for
It is a long, real example of the thing sweep_models() is built around:
detection probability is not a constant of the survey. Twenty years of
the same programme, the same aircraft and the same bay give p between
0.431 and 0.866 — a twofold range — with standard errors spanning an order
of magnitude, from 0.028 to 0.333.
The goodness-of-fit columns are worth reading next to it. Ganley et al. took
p > 0.05 as adequate fit, and 2003 fails on Kolmogorov-Smirnov (0.048) while
passing Cramér-von Mises (0.103); it also carries much the largest standard
error on p. A model can top a ranking and still fail a fit test, which is
why selection_table() carries cvm_p alongside the AIC.
Examples
ganley_detection
#> # A tibble: 20 × 5
#> year cvm_p ks_p p p_se
#> <int> <dbl> <dbl> <dbl> <dbl>
#> 1 1998 0.751 0.564 0.663 0.156
#> 2 1999 0.521 0.426 0.601 0.094
#> 3 2000 0.531 0.248 0.75 0.041
#> 4 2001 0.867 0.802 0.64 0.07
#> 5 2002 0.97 0.965 0.708 0.104
#> 6 2003 0.103 0.048 0.76 0.333
#> 7 2004 0.969 0.965 0.655 0.142
#> 8 2005 0.985 0.995 0.866 0.059
#> 9 2006 0.431 0.392 0.703 0.163
#> 10 2007 0.866 0.777 0.712 0.065
#> 11 2008 0.97 0.977 0.595 0.05
#> 12 2009 0.581 0.507 0.591 0.038
#> 13 2010 0.668 0.785 0.599 0.046
#> 14 2011 0.847 0.737 0.604 0.043
#> 15 2012 0.991 0.982 0.611 0.062
#> 16 2013 0.98 0.993 0.576 0.03
#> 17 2014 0.93 0.858 0.507 0.035
#> 18 2015 0.583 0.815 0.541 0.028
#> 19 2016 0.115 0.065 0.431 0.038
#> 20 2017 0.341 0.129 0.518 0.028
# Detection probability is not a constant of a survey programme
range(ganley_detection$p)
#> [1] 0.431 0.866
# The year that fails Kolmogorov-Smirnov also has the worst precision on p
ganley_detection[ganley_detection$ks_p < 0.05, ]
#> # A tibble: 1 × 5
#> year cvm_p ks_p p p_se
#> <int> <dbl> <dbl> <dbl> <dbl>
#> 1 2003 0.103 0.048 0.76 0.333