OccurrenceChange - 0.0001 and 0.99999

[AugustT]

In the occurrence change function the occupancy estimates are modified so that 0's and 1's are changed to 0.0001 and 0.9999

https://github.com/BiologicalRecordsCentre/sparta/blob/e1a6443dc9e8e859422b0ec129e05b67f1c9fbe3/R/occurrenceChange.r#L108

This almost certainly isn't needed for all of the change metrics used, for example the raw change in occupancy. I suggest these extreme value truncation are done INSIDE the different methods, and only where needed. I don't think the change of 1's is always needed as I think percentage change issues only arise from 0's.

I also suggest that the precision of these conversions it increased, so that instead of 0.0001 it is 0.00000001, for example. I don't think there is a good reason why this isn't done.

Spotted by @JKSheard

• [x] Identify in which of the change metrics truncation is needed
• [x] Move truncation into the specific metric calculation chunks
• [x] Change the precision of truncation

[drnickisaac]

First I looked at what the options are, and their sensitivity to boundary values (0 and 1) in the posterior.

• lineargrowth is a binomial glm, so boundary values are expected.
• difference is simply first minus last, so again no issue with boundary values
• percentdif is the difference divided by the first year. So zeros in first year (denominator) are going to cause apparent infinite increases.
• growthrate has the first year as the denominator, so Infinite values can emerge when first year = 0.

In summary, the problems arise when dividing by zero. So the issue is restricted to:

• zeros, not ones.
• the first year, not the last (because the last year is never in the denominator)
• only two out of four of the change metrics.

My solution is therefore to: 1) delete occ_it[occ_it == 1] <- 0.9999 2) replace occ_it[occ_it == 0] <- 0.0001 with res_tab[,1][res_tab[,1] == 0] <- 0.0000001 inside the loop for percentdif and growthrate.