I'm looking for good (freely-available) resources that guide statistical approaches for conducting meta-analyses; specifically, those addressing (1) which effect size statistic to use in different situations, (2) if and when different effect size statistics (e.g., mean difference, correlation coefficient, etc.) can be combined in a meta-analysis and when they should not be combined
Thanks!
Hi Mr. Tomkins,
You can find the answers to your questions in the Cochrane Handbook. It's a comprehensive guide for systematic reviews and meta-analysis, ensuring a strong methodology for your paper. Specifically, the answers to questions 1 and 2 can be found in Chapter 6.
The first step in determining the effect size is to decide whether your question is better answered by a binary outcome or a continuous outcome. Binary outcomes can be presented as risk ratios, which are easily interpreted but cannot be used with case-control studies. In such cases, odds ratios are required. Hazard ratios are also an option and generally have a greater impact as they relate both risk and time in their analysis.
For continuous outcomes, you may choose mean difference if the data present equivalent measurement units, or standardized mean difference if they are different. Studies may present this data as changes from baseline or post-intervention values. If the studies are randomized, you may use both interchangeably, but not if they are observational due to possible differences in baseline values.
All of this information and more can be found in the Cochrane Handbook. I wish you good luck on your journey.
https://training.cochrane.org/handbook/current/chapter-06
For resources on statistical approaches for conducting meta-analyses, especially regarding effect size statistics and their combination, you can explore the following:
Regarding 1):
- Odds Ratio generally in case control studies. However, it’s the most universal for binary outcomes
- Risk Ratio is more popular and easier in terms of interpretation but cannot be used for case control studies
- Mean difference for continuous outcomes.
- Std. Mean difference for continuous outcomes, WHEN metric of measurement is different. For example, you are measuring level of happiness with different scales (i.e. one reports from 0 to 5, another one from 1 to 20, while third one from 0 to 200).
- Hazard Ratio used in case of time involvement (to the best of my knowledge, it’s the only measurement of this type commonly used)
- Sensitivity, Specificity, Diagnostic Odds Ratio, Accuracy, Area under summary ROC are used specifically for diagnostic meta-analyses