Going through discussion posts on the Kaggle COVID-19 challenge, I came across a poster (Savanna Reid) who
mentioned the virtues of levels of evidence and the dangers of trusting too much in peer review, journal impact factor, and publication bias.
She also introduced the Bradford Hill criteria to me, which I will outline here in the hopes that it will come to mind later if I am ever worried about causation and correlation.
Sir Austin Bradford Hill
Basically, Sir Austin Bradford Hill was a dude around when big tobacco companies had doctors in their pockets prescribing cigarettes for all kinds of diseases. Cigarette companies would gawk and feign ignorance when anyone would claim that cigarettes were anything but the miracle cure for everything.
The Wikipedia article on the subject is
here, but to summarize the 9 essential criteria are:
1. Strength
2. Consistency
3. Specificity
4. Temporality
5. Biological gradient
6. Plausibility
7. Coherence
8. Experiment
9. Analogy
Reversibility is another proposed criteria. Some argue that no set of rules should matter and just to apply common sense, others that the setting in which a study is carried out matters.
The
MAGIC criteria concern how statistical arguments should be used, and the field of
causal inference also goes into a similar subject.
The problem of correlation and causation came up last night while I was reading a submission
here that claimed that there is a definite strong relationship between air temperature and virus transmission, although he did not attempt to show statistical evidence of this aside from some nice visualizations on virus predominance in China, Europe, and North America. However, as another Kaggler humorously pointed out, there are also more Macbook Pros in Europe and China and North America versus Africa, so maybe that is the cause of the spread of the virus?? It is also likely due to a lack of testing in some countries.
