[R-meta] Mata-analysis: combining retrospective studies and randomized trials

[Christian Mirian Larsen]

Thank you for your comment. I have replied further down,

Regards Christian

2018-03-19 15:23 GMT+01:00 Michael Dewey <lists at dewey.myzen.co.uk>:

> Dear Christian
>
> Some comments in-line
>
> On 19/03/2018 10:57, Christian Mirian Larsen wrote:
>
>> Dear all
>>
>> My main question is "can retrospective studies and randomized controlled
>> trials be combined in a meta-analysis?".
>>
>> E.g., I want to investigate the prophylactic effect of antiepileptic drugs
>> (AED) given to patients with brain tumors to prevent brain-tumor related
>> seizures.
>> I have found seven RCT that investigated this subject. Similarly, I have
>> found nine retrospective studies examining this topic by going through
>> patient records.
>>
>>
> I assume the retrospective studies are effectively non-randomised trials,
> having some people who took AED and some not?


> Can I pool the data from the RCT and RS into my matrix (AED/no AED vs.
>> seizures/no seizures) and conduct my meta-analysis based on these data?
>>
>>
> If I were doing that I would also include a moderator variable for study
> type to see if the effect differed depending on it. I would also prepare
> myself for the inevitable comment from referees that the results of the
> non-randomised studies are subject to bias and so should not be compared.
>
>

I constructed a moderator model taking study characteristics into account
(like study design, which showed for randomized controlled trials an
estimate of 0.93 (95%CI 0.63; 1.38, I2 = 11.5%, p = 0.34) and for
retrospective studies an estimate of 1.49 (95%CI 1.02; 2.17, I2 =23%,
p=0.27. The overall model estimate was 1.19 (95%CI 0.9; 1.56, I2=23%,
p=0.19).
When adjusting for confounding, it drops to 1.01 (95% CI: 0.72; 1.41, I2 =
7.9%, p =0.9).

My point is that the estimate from RCT and the retrospective studies are
very close to each other. A test for residual heterogeneity was 99.9%. Is
it completely irresponsible to combine these retrospective studies with
randomized controlled trials?

What bias could one expect in such retrospective analyzes?



> And lastly, are you familiar to any (epidemiological/biostatistic) studies
>> addressing this specific topic on combining RCT and RS?
>>
>>
> The issues of having randomised within study comparisons and
> non-randomised ones is often discussed in the context of network
> meta-analysis so that might be the place to look. There were several
> articles in the journal research Synthesis Methods in 2012 starting with a
> general introduction by Salanti starting on page 80 and all in sequence
> from there.
>
> Thank you in advance
>>
>> Regards,
>> Christian
>>
>>         [[alternative HTML version deleted]]
>>
>> _______________________________________________
>> R-sig-meta-analysis mailing list
>> R-sig-meta-analysis at r-project.org
>> https://stat.ethz.ch/mailman/listinfo/r-sig-meta-analysis
>>
>>
> --
> Michael
> http://www.dewey.myzen.co.uk/home.html
>

	[[alternative HTML version deleted]]