Update0.3

[shawndove]

Calculate and plot confidence intervals for all indicators

[wlangera]

Hi Shawn

We are still developing the bootstrapping framework.
My earlier functions did not bootstrap on the complete dataset (the cube) but only on the aggregated one.
Did you bootstrap on the complete dataset or on aggregated datasets?
We are now creating function that bootstrap on the complete dataset in this repo. Once we have tested the functions, we will move them to the dubicube package here.

I have tested one example:

denmark_cube <- process_cube(
  system.file("extdata", "denmark_mammals_cube_eqdgc.csv", package="b3gbi"))
denmark_observed_richness_ts <- obs_richness_ts(
  example_cube_1, first_year = 1980, level = "country", region = "Denmark") 
plot(denmark_observed_richness_ts)

I notice that the indicator value is each time lower than the lower confidence limit.
This seems strange.

I still see you plot the uncertainty around the loess smoother. This uncertainty is not correct and can be dropped se = FALSE.

Should we discuss this in december?

[shawndove]

Hi Shawn

We are still developing the bootstrapping framework. My earlier functions did not bootstrap on the complete dataset (the cube) but only on the aggregated one. Did you bootstrap on the complete dataset or on aggregated datasets?

What is bootstrapped depends on the indicator. Observed richness bootstraps all occurrences. But in some other cases it would be inefficient to do so.

We are now creating function that bootstrap on the complete dataset in this repo. Once we have tested the functions, we will move them to the dubicube package here.

I have tested one example:

denmark_cube <- process_cube(
  system.file("extdata", "denmark_mammals_cube_eqdgc.csv", package="b3gbi"))
denmark_observed_richness_ts <- obs_richness_ts(
  example_cube_1, first_year = 1980, level = "country", region = "Denmark") 
plot(denmark_observed_richness_ts)

I notice that the indicator value is each time lower than the lower confidence limit. This seems strange.

Yes, there are two indicators where the confidence intervals do not work quite right, and observed richness is one of them. Richness is tricky because when you sample with replacement the richness is reduced. So there is bias in the confidence intervals. Some methods in the boot package create CIs that are too low, others overcorrect the bias and make it too high. I am still trying to work out what to do about it.

I still see you plot the uncertainty around the loess smoother. This uncertainty is not correct and can be dropped se = FALSE.

I am not clear on why this is incorrect? It is meant to be the uncertainty around the trend.

Should we discuss this in december?

Yes, good idea.

[wlangera]

Ok, let's discuss it then

It is meant to be the uncertainty around the trend.

Yes, the intention is good, but it does not take into account the uncertainty on the indicators.
It only uses the indicator values.

[wlangera]

Great work.
As discussed online, 2 changes are requested:

1. visualising uncertainty for plot_ts(), see PR 0.3plot timeseries #30. We can discuss this further in this PR
2. remove bootstrapping for indicators where it does not work (e.g. species richness)