i don't understand "Relative Risk" in this study

gene

New Member
#1
hi, how can this be, that they say that the RR of 1.19 is an increase, and the RR 1.26 and RR 1.68 in not an increase?
"elevated serum PTH concentration increased the risk of all-cause mortality (RR 1.19; 95% CI 1.08–1.30) but not for cardiovascular mortality (RR 1.26; 95% CI 0.96–1.66). Subgroup analyses indicated that cardiovascular mortality risk appeared to be more pronounced among men (RR 1.68; 95% CI 1.05–2.67)."
https://www.sciencedirect.com/science/article/abs/pii/S0009898116300419#as0015
 

hlsmith

Less is more. Stay pure. Stay poor.
#2
I did not review your link, but traditionally what you are looking for is the 95% CI to exclude "1" to rule out chance , since "1" would mean the rates are the same. So, their results seem to a mixed bag.
 

gene

New Member
#3
hi, but are not all 3 results 95% CI? so as far as CI, they are all even? and yet the higher ones are somehow less of an increase. i must be missing some basic knowledge
 

gene

New Member
#5
i think that the range of the intervals is summed up in the means, right? but i'm trying to figure out how can RR of 1.19 be an increase, and the RR 1.26 and RR 1.68 is not an increase when the latter are larger?
 

spunky

Doesn't actually exist
#6
i think that the range of the intervals is summed up in the means, right? but i'm trying to figure out how can RR of 1.19 be an increase, and the RR 1.26 and RR 1.68 is not an increase when the latter are larger?
Because of what @hlsmith said. The Confidence Interval for 1.19 does NOT include 1 but the CI for 1.26 DOES include 1 (hence fail to reject the null hypothesis). Also, notice that although the RR for cardiovascular mortality is not statistically significant for the whole sample, it is for the subgroup of men. So your claim of "1.68 is not an increase" is not quite true within the context of the analyses done by the author.

I mean, it's still 'wrong' to go around doing sub-group analyses to fish for statistical significance (which is what the authors did), but at least that's how you get the interpretation they're using.
 

gene

New Member
#7
I think i understand now; so if a CI contains 1 it will not reject the null hypothesis and therefore will not be significantly different from 1? is that even if it was a CI width of (RR 4.9; 95% CI 0.99-9.00)?
 

hlsmith

Less is more. Stay pure. Stay poor.
#8
Yes you are correct. If I have a numerator and denominator that are the same then there is no difference. Sample size for the subgroups will impact the width of CI's. As for, "CI width of (RR 4.9; 95% CI 0.99-9.00)" if you were taking the hard rule that if "1" was in the CI you would fail to reject the null then it would be not be 'significant'. Of note, this was considered a hard rule in the past, but contemporaneously people are starting to acknowledge that the study would provide some evidence that there could be some relationship going on.

So if you are dealing with a ratio the CI cant include ''1", if that is your null value. If you are dealing with a difference (e.g. rate difference) then the CI cant include "0", if that is you null value. And in some cases, if desired, you can set your null to whatever value your want to test, like in a one-sample t-test.
 

gene

New Member
#9
right, thank you, so that just leaves me a bit perplexed as to scenarios that have the value close to (CI 1) being so different as to be significant and not, for example that (RR 4.9; 95% CI 0.99-9.00) would be not considered significant and (RR 1.02; 95% CI 1.01-1.02) would be significant, but then i suppose one has to look at the width and take a nuanced, logical view, is that correct?
and another thing i now don't understand then is why in another study they say that a CI including 1 "is" an increase: "both of the higher annual bolus dose trials increased fracture risks (RR = 1.09 (95% CI 0.93–1.28)"
now i'm really confused; i thought that since it included CL1 it would not be considered an increase??
 

spunky

Doesn't actually exist
#10
right, thank you, so that just leaves me a bit perplexed as to scenarios that have the value close to (CI 1) being so different as to be significant and not, for example that (RR 4.9; 95% CI 0.99-9.00) would be not considered significant and (RR 1.02; 95% CI 1.01-1.02) would be significant, but then i suppose one has to look at the width and take a nuanced, logical view, is that correct?
and another thing i now don't understand then is why in another study they say that a CI including 1 "is" an increase: "both of the higher annual bolus dose trials increased fracture risks (RR = 1.09 (95% CI 0.93–1.28)"
now i'm really confused; i thought that since it included CL1 it would not be considered an increase??
I wish I had better news, but it's common knowledge that many, many, MANY medical researchers are very, very, VERY sloppy data analysts. I honestly think that just escaped peer review... but a lot escapes peer review these days.
 

hlsmith

Less is more. Stay pure. Stay poor.
#11
common knowledge that many, many, MANY medical researchers are very, very, VERY sloppy data analysts.
This is a pretty bold statement. Sounds like you wont be lining up for a vaccine anytime soon!

As an author you can say whatever you want. It is up to the reader to interpret the results in the Results section as they see fit. If I reviewed a study and it was well designed and had a lower bound CI at say 0.99, I would imagine with a greater sample size it would be 'significant' and I would interpret it accordingly. So beyond just staring at whether the CIs cross the null value, one must digest the totality of the study. Conversely, there could be a low quality study reporting a significant effect, but the results may not be reproducible or repeatable.
 

spunky

Doesn't actually exist
#12
This is a pretty bold statement. Sounds like you wont be lining up for a vaccine anytime soon!

As an author you can say whatever you want. It is up to the reader to interpret the results in the Results section as they see fit. If I reviewed a study and it was well designed and had a lower bound CI at say 0.99, I would imagine with a greater sample size it would be 'significant' and I would interpret it accordingly. So beyond just staring at whether the CIs cross the null value, one must digest the totality of the study. Conversely, there could be a low quality study reporting a significant effect, but the results may not be reproducible or repeatable.
That's kind of what I got from #epitwitter. I've learned the chain goes like this:

If you're a medical researcher then EVERYBODY dunks on you.

If you're in epi then you get dunked on by biostatisticians and statisticians.

If you're a biostatistician then only actual statisticians dunk on you.
 

hlsmith

Less is more. Stay pure. Stay poor.
#13
Yeah pretty close, but the stats folks have no underlying content knowledge about the projects, so they end up getting a foul for charging!

The ideal scenario is when the epi person gets an assist or when the cross-trained all-stars check into the game (e.g., Jamie Robins, Miguel Hernan, etc.) and hit up an alley-oop self-assisted dunk.