2
$\begingroup$

I am currently writing a protocol for a systematic review and meta analysis to analyze the available evidence regarding epidemiologic studies on risk factors for pneumonia. The end goal of the project is to use Bayes Theorem and serialized LR testing to determine the probability that a particular patient will develop pneumonia. I hope to identify studies that are sufficiently homogeneous to combine the raw data for a pooled LR. Based on review of similar articles my criteria for pooling would be as follows:

  1. Risk factor must be included in at least 3 studies

  2. Isquared <30%

  3. 95% CI does not include the value 1.

All risk factors that do not meet that criteria will be included in a narrative style summary.

I was hoping to have a statistician on the team but that has fallen through. Given my limited background (1 graduate level course in research design and statistics completed in 1992) could someone weigh in on whether my plan will result in valid information? If not, could you please offer alternatives?

$\endgroup$
1
  • $\begingroup$ Consider the point of homogeneity testing in meta-analysis is just to say whether they are homogeneous, not to exclude the heterogeneous subset as fallacious. Consider publication bias. Say a few significant studies are published: most with positive effect, a few with negative effect. Then you take only the positive ones? You create a stronger bias than there is in the already biased literature! $\endgroup$
    – AdamO
    Jan 27, 2020 at 15:39

1 Answer 1

2
$\begingroup$

There is no need to reject the results of your meta-analysis based on any of these criteria. Heterogeneity is to be expected with anything other than very tightly specified situations. Your estimate will still be the best available even if the confidence interval about it includes the null although depending on the costs and benefits of the decisions about diagnosing pneumonia you may decide that if is not worth testing using the test available. It is possible to do meta-analysis with as few as two studies although confidence in the generalisability of the results may be improved by having many more studies.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.