Re:Study: Chances of serious side effects from coronavirus v...
by ponow on Monday September 05, 2022 @10:50AM
(#32724937)
Then if we throw a logarithm in there we should be seeking zero-centered intervals, I'm thinking, as opposed to non-negative intervals containing 1 (no effect). Any why normal distributions are even part of such discussions of ratios seems crazy, given that ratios can't be negative, but log-ratios can. Not my area; could it be that a log or similar non-linearity was already effectively included? If the confidence interval method was only using ratios, then the lower bound should have been zero, and negatives impossible.
Sanity checking results shouldn't be so difficult.
The RR CI must be >= 0. The calculation is based on the incidence in a population and the incidence in a sub-population, which are themselves fractions of positive integers (e.g., 1/1000 and 1/100). I don't think RR can ever be negative, or at least I'm not sure how you would get that without messing up your data or calculations?
RR of >1 means there is excess risk, and <1 means anti-excess risk which is a word I cannot think of right now (just, less risk than the baseline anyway). But RR is also just a measure of risk over baseline to begin with, so a 1.5 RR on something that happens 1/10,000 times is ... not that concerning? I wouldn't be concerned, anyway. Depending on what the risk is, I guess.
So, no, I don't think we would expect RR CI to be centered around zero at all.
The fact that the main effect is so small and crosses from neg to pos is by far more concerning. I would not submit results like that for publication.