This begs a simple question: can't we use the same analysis to test the hypothesis whether insulin is indeed linked to Alzheimer - simply by taking the same approach and replacing "cancer" by "alzheimer" in the database query?
After all, if we know these same patients have had a lower rate of cancer, maybe we know also about which other diseases they may have had at the same time.
If the data had been openly available, it could have been done, maybe easily if the data is properly coded (using ICD codes, replacing cancer by alzheimer or even broadening the query to other neurodegeneratives diseases is just a "or" away in SQL)
Yet with the current closed-garden-behind-a-paywall approach of science, we can't.
There are many wild hypothesis about nutrition thrown around.
But until I see actual data, I will not call any of them "hypothesis".
EDIT: added HN URL for the discussion about metformin
The problem with your idea is that Alzheimer's disease is not easy to study.
1) There is no good diagnostic test for Alzheimer's outside of a brain biopsy (although Lilly just got a diagnostic approved). Today, it's usually diagnosed by a process of elimination (i.e. you don't have any signs that it's another organic brain disease, thus it must be Alzheimer's).
2) The time of onset to the time of death for Alzheimer's is often measured in decades. You'd have to run a REALLY long study to be able to capture that data.
The other issue is that analyzing data from studies retrospectively (studies that weren't designed to test your given hypotheses) are generally regarded as "lower quality". It may support your hypotheses, but it's pretty weak support.
1) Lilly's diagnostic is about beta amyloid plaques, not Alzheimer directly. There is some questions whether amyloid plaques are good indicators of Al Zheimer or not, and this will need to be validated in the future.