We appreciate the interest of Wu et al. in our systematic review and we wish to discuss some aspects raised in their letter.

The method we used in the meta-analysis, which compared pre- and postoperative data with the assumption of preoperative data as a control group, was rendered necessary by the lack of selected studies presenting a control group. In such a case, control must be perceived as a self controlled situation before treatment, not as a non-treated sample.

We would like to thank Wu et al. for their suggested approach to obtaining forest plots. We performed the analyses as proposed and observed that the results obtained further validate and corroborate our findings. There was practically no change when these newly calculated data were compared to the outcomes presented in our article and significant and non-significant differences in our published results were confirmed with this approach. The summary effects calculated assuming a correlation coefficient (Corr) of 0.5 showed a difference of less than 0.09 mm and 0.41 mm ² and this difference can be represented as less than 1% in 10 subgroups studied, 1.5% in 1 subgroup and 4.4% in 1 subgroup when compared with the published summary effects. The stability of the outcomes was confirmed by the sensitivity analysis by assuming Corr as 0.7 and 0.3, as the most discrepant summary effects showed a difference of 0.21 mm and 0.95 mm ² from the published ones. CIs and heterogeneity were fairly different, but as the summary effects were basically the same, the results presented and discussed are valid.

Although the summary effects calculated did not really differ from the ones previously presented when the method suggested was applied to the data summarized in our article, we were interested in the method and we think it should be used in future meta-analyses that compare data in a paired way.

Another point raised was the use of a fixed-effects model when I ² < 75% in our meta-analysis. Authors tend to diverge when it comes to the appropriate method for selecting between fixed- and random-effects models. We followed the method used by Chen et al., and the use of random-effects model when I ² > 75% is also recommended by Ried. In our published analyses, from the 12 subgroups studied, in only two subgroups I ² was not lower than 50% (50% and 69%). The most recently updated Cochrane Handbook says that thresholds for the interpretation of I ² can be misleading, since the importance of inconsistency depends on several factors. Looking in the Handbook, we could not find the suggestion that a fixed-effects model should be used when I ² < 50% or the suggestion that I ² = 75% is the threshold for whether to conduct a meta-analysis. The Handbook does recommend that when review authors are concerned about the influence of small-study effects on the results of a meta-analysis in which there is evidence of between-study heterogeneity ( I ² > 0), they compare the fixed- and random-effects estimates of the intervention effect. Borenstein et al. suggest that the selection between fixed- and random-effects models be based on the authors’ expectation about whether the studies share a common effect size and on the goals in performing the analysis – either computing the common effect size for the identified population or generalizing to other populations. Recent literature leads us to think that in future analyses a model must be chosen and heterogeneity should not be a determinant point in this choice and that high heterogeneity must be further investigated through subgroup analyses and meta-regression.

We are pleased for the opportunity to discuss these points as this discussion may enhance our knowledge and help other authors when performing their meta-analyses.