Confidence Intervals in a research paper

[leafonthewind]

Hi, all. I'm writing a paper on concussion data from the NFL Surveillance System, and I want to be able to put the data analysis from the journal article in my own words, but I don't really understand what the researchers are saying. Here's the quote. Can someone please tell me basically what it means?

"Each concussion was analyzed as an independent event. Summary measures are reported with 95% confidence intervals and significant differences with P<0.05. Injury rates per 100 game positions (gp) were calculated to provide perspective on the risk of injury among the position categories." (Casson et al. 2010)

Thanks so much for any input. :wave:

 

[hlsmith]

Without reading the paper, the following is my interpretation.




1. Each concussion was analyzed as an independent event.
- They did not control for players contributing more than 1 concussive episode in the study. Example, if I was trying to find the average weight of everyone that came to my restaurant, I would find the average, though some patrons may have came more than once and would contribute more information. When modeling, you typically want to examine to see if controlling for multiple observations helps to better explain the outcome or not. So as per my bad example, if the person had unique traits or trends (say they were 330 pounds), knowing they are in the dataset 4 times can help explain the outcome better.




2. Summary measures are reported with 95% confidence intervals and significant differences with P<0.05.
They have a single sample, which may vary from the population at large, so it is typical to provide 95% CI to describe where the true measure may be.


Differences: they are saying they compared values and used a statistical test and if the p-value for the test was below 0.05 is was determined to be statistically significant beyond their acceptable level of chance.


3. Injury rates per 100 game positions (gp) were calculated to provide perspective on the risk of injury among the position categories." (Casson et al. 2010)
# concussion per 100 games, and they may have broken this out by NFL position. So over a hundred games how many centers sustained a concussion. So if there were 330 concussions in centers over 10000 game the risk of concussion for a center per 100 games would be 3.3%.



I wrote this quickly, so let me know if you need clarification.

 

[the42up]

Without reading the paper, the following is my interpretation.




1. Each concussion was analyzed as an independent event.
- They did not control for players contributing more than 1 concussive episode in the study. Example, if I was trying to find the average weight of everyone that came to my restaurant, I would find the average, though some patrons may have came more than once and would contribute more information. When modeling, you typically want to examine to see if controlling for multiple observations helps to better explain the outcome or not. So as per my bad example, if the person had unique traits or trends (say they were 330 pounds), knowing they are in the dataset 4 times can help explain the outcome better.

Or more importantly if they had previously had a concussion.

Personally, considering the recent reports about concussion in the NFL, treating the concussions as independent will probably cause you to OVERestimate the likliehood of a concussion. Especially considering that having a concussion makes you more susceptible to others in the future. Either they are working with the data they had (i.e. know way to determine dependence) or just lazy stats.

 

[hlsmith]

I would agree with "the42up", that not controlling for non-independence in the study sample will typically lead to narrower confidence bands. Not controlling for it can be due to not knowing how to do it or possibly they went to control for it, but determined it did not make a difference. Though under the latter scenario, you should inform the reader of this action and reasoning.