Processing vulnerability data for evidence & agreement

Hey Cee, I went ahead and added the "uncertainty agreegated" sheet that Oronde shared with us. That looks to match the SB data after you aggregate. The evidence categories of small, medium and large make sense here and follow what Oronde shared:

Amt of Evidence:
Small if less than 5 studies measured the indicator. Medium if 5-9 studies measured the indicator.
Large if more than 9 studies measured the indicator.

Still a bit stuck confirming the level of agreement (miss matches with agreement ordinal vs cardinal values) - I thought I understood but could still be doing the math wrong?

Level of Agreement: NA if the level of agreement could not be calculated as indicator was measured only once. Low if the most recorded direction of influence reflects less than 50% of models measuring the determinant. Medium if the most recorded direction of influence reflects more than 50% but less than 74% of models measuring the determinant. High if more than 74% of models measuring the determinant recorded the same direction of influence on conditions of water insecurity.

requested review from @cnell

This could also be written as

p1_unc_agg |>
  group_by(dimension, determinant, indicator) |>
  summarize(across(c(contains('related'),
                     contains('unknown'),
                     contains('significant')),
                   list(total = ~sum(.x, na.rm=TRUE))))

changed this line in version 2 of the diff

Instead of using pos_related here, I think it should be whichever value out of pos_related, neg_related, unrelated, or unknown_direction is the highest.

You could do something like this to find the max value in each of those categories:

top_trend <- p2_unc_agg_summary |>
  select(dimension, determinant, indicator, 
         positively_related_total, negatively_related_total, unrelated_total, 
         unknown_direction_total) |>
  pivot_longer(!c(dimension,determinant,indicator)) |>
  group_by(dimension, determinant, indicator) |>
  slice_max(value) |>
  rename(sig_name = name, sig_value = value)

and then use a join to add is back to p2_unc_agg_summary and calculate the %:

p2_unc_agg_summary |>
  left_join(top_trend) |>
  mutate(level_agreement = 100*(sig_value/evidence_val))

Try that out and see if it provides the expected results

changed this line in version 2 of the diff

This looks right to me as well. So to recreate their table of indicators visual, I think we would just do the same thing but group_by(dimension, determinant) to get the same selection of indicators they show.

added 1 commit

• a4efc328 - suggested edits and quick viz view

Compare with previous version

Okay applied comments and explored using ggplotly to look at avg level of agreement across determinants

In p1_unc_stats there are 106 rows, but p2_unc_agg_summary has 107 rows. I expected that these would be the same length, since p2_unc_agg_summary is essentially recreating the stats that are published in p1_unc_stats from the raw data. Do you know if there are any indicators missing between the two?

good catch - yeah after looking at the spreadsheets (case differences and slight differences in indicators names made it hard to compare indicators in R) , its looks like p1_unc_stats contains Other non-white while p2_unc_agg_summary does not.

p2_unc_agg_summary has amenities & hazard type while p1_unc_stats does not.

compare_indicators.xlsx

hmmm ok. Can you add an issue here to document that? Then you're good to merge, and let's follow up with the team to understand why these may be different. Ultimately we'll need to work form the published data, so we may need to replicate these stats onto p1_unc_stats to develop any viz

approved this merge request

mentioned in issue #2 (closed)

merged

mentioned in commit d53d6557

mentioned in issue #46