Thanks for your article, please let me state some comments:
1) The ranking page has been up for over a year now, hence it's not new.
2) The ranking page, allows many different methodological model choices. I don't know which models are the best, but I tried to provide the user with the most flexible.
3) Your choice of a 2015-2019 baseline may be statistically limited for robust linear regression analysis. Additionally, since all baseline selections and their resulting predictions carry inherent uncertainty, I would recommend displaying prediction intervals alongside the forecasts to better communicate this uncertainty to users.
4) The underlying model calculations utilize the R package fabletools, which should minimize the likelihood of statistical errors in the computations.
5) We have different underlying data sources, while you are using Eurostat, Mortality.Watch uses Mortality.org for this specific chart - which may explain the discrepancy. I will double check why the algorithm chose STMF over Eurostat.
2/ No objection to having several settings. I would however object to having no warning when data is missing ; like Serbia. You could add a completion score & exclude countries with too much data missing on a given interval maybe.
3/ 2015-2019 is relevant for short term trends ; it won't change much the picture regarding the births disaster - which I covered at length (see here for example for longer trends : https://openvaet.substack.com/p/sweden-births-in-2023)
4/ I'll check but I doubt a difference comes from there.
Thanks, the issue was indeed the usage of mortality.org data (which is interesting, that it yields such a different result) - now that I've fixed the source selection bug, it defaults to using the higher resolution Eurostat data, which shows similar but maybe even higher excess than what you have:
Excellent, my friend. I've added this article to the OpenVAET page I created on the White Rose Wiki: https://www.whiteroseintelligence.com/openvaet/
Thanks to you (twice) my good Sir, it's good to see that you're as efficient as usual, some readings ahead there !
Thanks for your article, please let me state some comments:
1) The ranking page has been up for over a year now, hence it's not new.
2) The ranking page, allows many different methodological model choices. I don't know which models are the best, but I tried to provide the user with the most flexible.
3) Your choice of a 2015-2019 baseline may be statistically limited for robust linear regression analysis. Additionally, since all baseline selections and their resulting predictions carry inherent uncertainty, I would recommend displaying prediction intervals alongside the forecasts to better communicate this uncertainty to users.
4) The underlying model calculations utilize the R package fabletools, which should minimize the likelihood of statistical errors in the computations.
5) We have different underlying data sources, while you are using Eurostat, Mortality.Watch uses Mortality.org for this specific chart - which may explain the discrepancy. I will double check why the algorithm chose STMF over Eurostat.
1/ Thanks, wasn't aware.
2/ No objection to having several settings. I would however object to having no warning when data is missing ; like Serbia. You could add a completion score & exclude countries with too much data missing on a given interval maybe.
3/ 2015-2019 is relevant for short term trends ; it won't change much the picture regarding the births disaster - which I covered at length (see here for example for longer trends : https://openvaet.substack.com/p/sweden-births-in-2023)
4/ I'll check but I doubt a difference comes from there.
5/ Let me know if you find the cause of these.
2) Good point, I'll need to add some detection mechanism and flagging. Noted.
3) For France, 2014 ASMR is sign. lower than 2015, which is likely due to different timings of winter peaks. Hence, when you use a yearly seasonal (July-June) time frame, 2024 is right on the linear trend: https://www.mortality.watch/explorer/?c=FRA&ct=midyear&df=2000%252F01&bf=2014%252F15&bm=lin_reg
5) Turns out it was missing 'eurostat' from the data resolution mechanism, hence STMF was preferred. Updated & rerunning: https://github.com/MortalityWatch/charts/commit/81708e4afdb1a61758d161e68b122f0b5efe965e#diff-19c2685a1c17298177083d952dbe97e27d98c85857452d8a88ecb08445e85d7aR44
Edited the intro & the mention of the issue to be more accurate.
Thanks, the issue was indeed the usage of mortality.org data (which is interesting, that it yields such a different result) - now that I've fixed the source selection bug, it defaults to using the higher resolution Eurostat data, which shows similar but maybe even higher excess than what you have:
https://www.mortality.watch/explorer/?c=FRA&ct=yearly&e=1&df=2015&bf=2015&sb=1&bm=lin_reg&sl=1
Wow, in the first two charts, it appears Slovakia really got its ass kicked in 2021.
2021 is also the census year in Slovakia, so adjustments could have impacted the baseline ; it'll have to be investigated further.
https://www.scitanie.sk/en/census-history-in-the-sr-1918-2021
Thanks for double checking. I'm going to check why the logic uses STMF over Eurostat here.