I fear that the risk ratio analysis shown in Fig 4 and 5 is misleading. It is contraintuitiv that you got excess death, but in this analysis you got significantly lowered risk in many cohorts.
The main problem for me is that the people within each age cohort are "rolling" in a 5 year period: those who were in, e.g., the cohort 40-44 in 2020 are almost completely in cohort 45-49 in 2023. And so on.
My assumption is that the RR analysis works reasonable when comparing 2 neighboring years, but not when comparing more distant or lumped together years.
Maybe: you run an analysis as in Fig. 3, but only for specific years of birth (YoB), maybe YoB 1070-1974 or 1970-79. Don't know if the data structure allows auch a look.
Misleading, oh yes, if we take the bare results without paying attention to the model assumptions.
Model 2 is a raw comparison of determined risks. So we see trends in the RR diagrams, if they exist.
Model 1 compares "observed vs. expected". It includes the trends and extrapolates them in the prediction. We wouldn't therefore see them in RRs, but can obtain them from the model.
I've been thinking about that for a long time. RR diagrams comparing years directly showed a saddle pattern that became more obvious the longer the years compared were apart.
Perhaps you are right, and it is related to the fact that a 5-year cohort contains different individuals 5 years later, who of course have a different history. Unfortunately, we cannot conclude from the risk changes whether people have become more robust or whether prevention and treatment of diseases have improved.
This does not contradict positive excess mortality. We could perhaps view the matter as a decision dilemma. If an additive risk meets a negative trend, these two effects cancel each other out. If we postulate a continued trend, the decision is "excess mortality", otherwise the opposite may be true.
In any case, thank you very much for your constructive comment. I'm always happy when someone is interested.
I agree with Fakten O., it would be interesting to re-run your analysis with Single Year of Age Deaths/Populations and see what difference the 5-year groupings has made.
Also interesting to do single year cohort analyses by Year of Birth - take a look at Joel Smalley's recent substack posts using this approach.
In the article I was only able to show a fraction of what else I evaluated. It's important to me to see results as quickly as possible and to show an age gradation. That's why I mainly use the "Sonderauswertung Sterbefälle" from Destatis with the 14 cohorts and weekly data.
With regard to the annual year-of-birth assessment, I see little need for work. The statistical offices do that in their period life tables. At the request of "Erbsenzähler", Destatis has provided a table with such mortality risks, but unfortunately it ends in 2023. Normally, they offer 3-year-averages.
Excellent job!
I fear that the risk ratio analysis shown in Fig 4 and 5 is misleading. It is contraintuitiv that you got excess death, but in this analysis you got significantly lowered risk in many cohorts.
The main problem for me is that the people within each age cohort are "rolling" in a 5 year period: those who were in, e.g., the cohort 40-44 in 2020 are almost completely in cohort 45-49 in 2023. And so on.
My assumption is that the RR analysis works reasonable when comparing 2 neighboring years, but not when comparing more distant or lumped together years.
Maybe: you run an analysis as in Fig. 3, but only for specific years of birth (YoB), maybe YoB 1070-1974 or 1970-79. Don't know if the data structure allows auch a look.
Misleading, oh yes, if we take the bare results without paying attention to the model assumptions.
Model 2 is a raw comparison of determined risks. So we see trends in the RR diagrams, if they exist.
Model 1 compares "observed vs. expected". It includes the trends and extrapolates them in the prediction. We wouldn't therefore see them in RRs, but can obtain them from the model.
I've been thinking about that for a long time. RR diagrams comparing years directly showed a saddle pattern that became more obvious the longer the years compared were apart.
Perhaps you are right, and it is related to the fact that a 5-year cohort contains different individuals 5 years later, who of course have a different history. Unfortunately, we cannot conclude from the risk changes whether people have become more robust or whether prevention and treatment of diseases have improved.
This does not contradict positive excess mortality. We could perhaps view the matter as a decision dilemma. If an additive risk meets a negative trend, these two effects cancel each other out. If we postulate a continued trend, the decision is "excess mortality", otherwise the opposite may be true.
In any case, thank you very much for your constructive comment. I'm always happy when someone is interested.
Nice work, Ulf!
I agree with Fakten O., it would be interesting to re-run your analysis with Single Year of Age Deaths/Populations and see what difference the 5-year groupings has made.
Also interesting to do single year cohort analyses by Year of Birth - take a look at Joel Smalley's recent substack posts using this approach.
https://metatron.substack.com/p/a-history-of-life-expectancy-1974
https://metatron.substack.com/p/breaking-uk-immigrants-dying-at-substantially
Thanks for the link. I'll take a look at it.
In the article I was only able to show a fraction of what else I evaluated. It's important to me to see results as quickly as possible and to show an age gradation. That's why I mainly use the "Sonderauswertung Sterbefälle" from Destatis with the 14 cohorts and weekly data.
With regard to the annual year-of-birth assessment, I see little need for work. The statistical offices do that in their period life tables. At the request of "Erbsenzähler", Destatis has provided a table with such mortality risks, but unfortunately it ends in 2023. Normally, they offer 3-year-averages.
And: Yes, it's scripted in R.
P.s. what are you running your models in, is it R?