Average percentage across studies

[Aris]

Hi!
I am doing a study in sudden death in athletes. So for every study I get a percentage that represents the number of athletes that died from a particular disease. For example:
Study Duration / Total Number of deaths / Deaths from Myocardial Infarction (MI) / MI % / Deaths from Cardiomyopathy (CM) / CM %
1990-2000____ / 100 _______________________ / 50__________________________________________ / 50% / 25 _____________________________________/ 25%
2000-2002____ / 1500 ______________________ / 250________________________________________ / 16.6%/ 120 ____________________________________/ 8%
1980-1985____ / 20 _________________________ / 13_________________________________________ / 65% / 7 ______________________________________/ 35%
....
How can I calculate the average of the percentages in order to get an overall "average percentage" across the studies included? For this example lets say "Deaths from Myocardial Infarction" for which the percentages are 50%, 16.6%, and 65%. Obviously just to sum and divide them is not enough. Maybe for every study percentage should be a weight according to the population?
Thank you

 

[hlsmith]

So you are only trying to pool 3 studies?

 

[fed2]

i think we can use t-test here?

 

[Aris]

Thanks for the reply. No there are many studies. I just gave this example so we can have an idea what am I trying to do. There are actually 31 studies.

 

[Aris]

So this is a part of the table. As you can see for every study there is a total number of deaths (TotalSrSCD/SCA). From this total, HCM attributed deaths for example, were 7. Which is represented as 4.7%. For the next study is 11%, the next 7.1%, 3%, 1.9%, 6.6%, and 36%. If I will do this for every different disease (CAD, ARVC, ILVH, ...etc), I can find the prevalence for each disease and compare them (which disease has the highest prevalence across all studies in sudden death in athletes?). Obviously, the 4.7% from a sample of 147 deaths, has a different weight from a 6.6% from 15 deaths.

 

[hlsmith]

Are you familiar with metal analyses?

 

[Aris]

I know what is a meta-analysis and I have an idea from R. But I don't know the steps to do it. In some other similar meta-analysis they did DerSimonian and Laird's methodology. I only know it as a name and not how it's done.

 

[hlsmith]

Do you know about multilevel regression models?

 

[Aris]

Do you know about multilevel regression models?

Not really. Let me ask a dumb question. If I just sum the number of deaths from a particular disease and not the percentages and then divide it with the total deaths, would that be valid? For example, let's take HCM (HCM SCD No). We have 7+13+12+2+1+1+302 = 338. Then divide it by the total number of deaths across the studies (Total SrSCD/SCA) 147+118+168+64+51+15+842= 1405; So 338/1405=0.24 or 24%. Is this correct for estimating an average proportion of deaths attributed to HCM?

 

[hlsmith]

My response would be based on what you plan to do with these results?

 

[Aris]

I just want to compare the prevalence of each disease on sudden death (e.g. say CAD has the highest prevalence with 50% in the included studies, HCM the second place with 12%, etc), make a pie chart, and some subgroup comparison (e.g. according to mean age).

 

[hlsmith]

If you are not looking to share this at a national conference or try to publish it - I would just look up a formula for weighted means and use that. It should get you a ballpark value. However, if not all study were designed to answer the same question or have the same methodology, follow-up, etc. there will be some bias in the estimate.

 

[Aris]

Except the method I mentioned before (just sum up the total deaths from each disease separately and then divide each one of them with the total number of deaths), can you suggest any other formula?

 

[hlsmith]

Well, I guess the formula is as simple as that. The issue with not using multilevel modeling is that if you calculated confidence interval on your estimate, it would be wrong, because you didn't have a big study, you had a bunch of small studies.

P^=((P*n1)+(P2*n2))/(n1+n2)

 

[Aris]

OK, thanks