I am working on a systematic review about interventions to prevent obesity. Some of the included studies have before-after design. What is the best effect size for these studies?
Can you tell me any more about these "before-after" designs? Are you referring to repeated measure designs? This article by Morris and DeShon (2002) may be helpful...
If the ES are based on paired and not independent data, you need to consider this properly. In a systematic review (I assume it is a meta-analysis), you will have to make sure the reports in the original reports also report mean change statistics. That is, the formula that they have used to compute a Cohen's d (or, even better, a Hegde's g) should have adjusted the standard error and CI for the within-pair correlation between the two measurements. If you code raw data rather than pick reported ES and SE, you will need to also retrieve the within-pair correlation. Surprisingly often authors forget to report it.
Check this reference for more details on how to do that.
Becker, B. J. (1988). Synthesizing standardized mean‐change measures. British Journal of Mathematical and Statistical Psychology, 41(2), 257-278.
Here is an example of how we dealt with a similiar issue in a recent meta-analysis on pre-post designs:
Article Working memory training revisited: A multi-level meta-analys...
A before-after study measures an outcome at two time points. The first time point is before the initiation of intervention. The second time point is after the intervention has begun.
In principle I don't see why Cohen's d cannot be used for the type of primary studies you are dealing with (comparing means of the outcome measured in pre- and post- treatments). I think I've found an interesting discussion on stackexchange; hopefully, this matches with what you want to do (https://stats.stackexchange.com/questions/282270/how-to-perform-a-meta-analysis-on-change-from-baseline-after-a-treatment). Apparently there is the need to account for the pre-post correlation of within-group change (if reported by the study) when calculating the sampling variance (something that Jan Antfolk probably already mentioned above).
Hope this help (medical/epidemiological meta-analysis are not my expertise).
You can use Cohen's d (or Hedge's g) for before-after comparisons as well, but you need to adjust the SE for the within-pair correlation of measurements and the two time points.