Meta analysis possible if only Relative Risks and 95% CI's are available?

[thaman]

Dear members,

My knowledge of statistics is extremely limited. Since a few months I am wondering if it is possible to perform a meta analysis in SPSS if only Relative Risks (RR's) and 95% CI's are available from scientific publications.

Here is a link to a table showing a meta-analysis based on RR's from 16 different publications. Each RR is described in addition to it's 95% CI and the % weight it is given in the analysis: http://aje.oxfordjournals.org/conten...expansion.html

My question is: Can I enter this data in SPSS and perform a meta-analysis myself? If so, can somebody tell me how to I must enter the data? As far as I am concerned the example can be based on the following 3 RR's:

RR = 1.06 (0.36-3.12) % weight = 8%.
RR = 0.76 (0.49-1.18) % weight = 38%.
RR = 0.90 (0.69-1.18) % weight = 54%.

Thanx!

[trinker]

Your effect sizes should all be in the same metric. So 2 questions:

1) do you know the sample sizes or can you figure them out from the degrees of freedom?
2) what metric are all the rest of your effect sizes in? It may be possible to convert them to r for instance but this may not be meaningful.

[thaman]

Thank you for your response, trinker.

Maybe I should have asked the question in a different way: What data do I need to put into SPSS to perform a (preferably fixed effects) meta-analysis?

As an example it would be better if I use the data I actually want to enter. Using the 3 following data sets.

Cohort 1: 1,900 subjects. 215 disease cases total. RR = 1.11 (0.91-1.36).
Cohort 2: 10,059 subjects. 1,070 diseases cases total. RR = 0.86 (0.56-1.35).
Cohort 3: 7,088 subjects. 1,177 disease cases total. RR = 0.86 (0.67-1.12).

RR's are for high vs low consumption of fat.
I do not know if it is necessary to describe the amounts of cases for low vs high intake categories seperately.

And I do not know if I understand your second question correctly. Pleas tell me if other data is needed.

[trinker]

I don't know SPSS just something about meta analysis (I do my work in R) so I can't speak to SPSS. Basically you need a data set of effect sizes and and corresponding variances (these need to be calculated). As far as fixed effects I'd argue that you should test the random effects model and if the tau is not significant it reduces to the fixed effects model anyway. In social sciences we rarely can use a fixed effects model but in the medical field you may be able to use the fixed effects model. The Q and common/average effect (estimate) statistics are important (Q is chi^2 distributed and the estimate is normally distributed (or assumed to be). If Q is not significant you can reasonably assume fixed effects model. This is particularly important if Q is significant but common/average effect (estimate) is not. This may indicate you should include moderators in your study. So at minimal you need effect sizes and variances but possibly moderators as well.

[victorxstc]

Thanks a lot trinker. Is Q the q-value corrected for the familywise error? Or is it something else?

@thaman
I think SPSS can't do any analyses from the descriptive statistics. At least non of its menus let you do something like that. Maybe you can code in SPSS to make it run extra analyses, but most of the analyses it can do using its code is restricted to those performable using the menu. So I suggest you using specific programming languages like R which allow you to freely run whatever you need to. There are menu-based software designed for meta-analyses and you can search for them on the net or here (in TalkStats), but SPSS is not the choice.

[trinker]

I know SPSS does meta analysis because I'm sitting in meta analysis right now looking at slides of the output and they are the same as R's matafor package. I'm not sure if SPSS can do random effects though.

[trinker]

Q is a test of homogeneity between studies.