Article type
Year
Abstract
Background: Skewed data presents a particular problem when dealing with continuous outcomes in a meta-analysis. Sometimes primary researchers use a logarithmic transformation, which can lead to complications when performing a meta-analysis, as some studies may present means and standard deviations after logarithmic transformation, while other studies present means and standard deviations on the raw scale.
Objectives: To review, develop and compare methods for transformations that enable meta-analyses on either the raw or the logarithmic scale, irrespective of how results are presented.
Methods: We compared three methods, two of which have alternative standard error formulae, in two applications and in a series of simulation studies. The examples were taken from different areas: the first one combined information from clinical trials assessing the effect of gonadotrophin-releasing hormone (GnRH) analogues, compared with placebo, for improving intra-uterine operating environment and treatment outcomes after surgery, while the second one was selected from the field of genetic epidemiology. To evaluate the methods, we simulated continuous outcome data for a single, two-group study, according to various distributions and different sample sizes.
Results: For the first example there were no substantial differences between the methods. Conclusions of the meta-analyses using both scales were consistent, with the analysis on the logarithmic scale having marginally more consistent results across the studies. In the genetic epidemiology example some differences between the methods were found due to large differences in standard deviations between groups. The simulation study did not reveal a universally superior method. All methods were reasonably robust to data having distributions other than the log-normal. The most serious threat to validity from among the scenarios we simulated was when the standard deviations differed between the groups.
Conclusions: We conclude that an approach based on a log-normal assumption for the raw data is reasonably robust to different true distributions. In theory, this method offers a way of undertaking meta-analyses of skewed data on the logarithmic scale, irrespective of the scale in which means and standard deviations are presented.
Objectives: To review, develop and compare methods for transformations that enable meta-analyses on either the raw or the logarithmic scale, irrespective of how results are presented.
Methods: We compared three methods, two of which have alternative standard error formulae, in two applications and in a series of simulation studies. The examples were taken from different areas: the first one combined information from clinical trials assessing the effect of gonadotrophin-releasing hormone (GnRH) analogues, compared with placebo, for improving intra-uterine operating environment and treatment outcomes after surgery, while the second one was selected from the field of genetic epidemiology. To evaluate the methods, we simulated continuous outcome data for a single, two-group study, according to various distributions and different sample sizes.
Results: For the first example there were no substantial differences between the methods. Conclusions of the meta-analyses using both scales were consistent, with the analysis on the logarithmic scale having marginally more consistent results across the studies. In the genetic epidemiology example some differences between the methods were found due to large differences in standard deviations between groups. The simulation study did not reveal a universally superior method. All methods were reasonably robust to data having distributions other than the log-normal. The most serious threat to validity from among the scenarios we simulated was when the standard deviations differed between the groups.
Conclusions: We conclude that an approach based on a log-normal assumption for the raw data is reasonably robust to different true distributions. In theory, this method offers a way of undertaking meta-analyses of skewed data on the logarithmic scale, irrespective of the scale in which means and standard deviations are presented.