Article type
Year
Abstract
Background:
Some clinical study designs require that interdependency of data (IoD) be addressed in the reporting of study results. In self-controlled and cross-over trials, patients form part of both the intervention and control groups. In other studies, multiple disease locations are measured in the same patient. Because clinical studies often fail to account for IoD, one must apply specific statistical methods to the data of those studies when conducting a systematic review (SR) with a meta-analysis (MA). Otherwise, the results may be misleading.
Objectives:
This presentation describes different options for addressing and solving problems with IoD in MAs.
Methods:
We performed 3 SRs with MAs in medical fields where data clustering is common: ophthalmology (paired data on both eyes), dentistry (multiple data on single teeth), and sleep medicine (paired cross-over data in patients with chronic diseases). Before conducting the MAs, we assessed the studies included for IoD and corrected the data when IoD adjustment was necessary. If possible, we adjusted the results by using a correlation coefficient (CC) for the standard deviation of the effect estimates of the studies included.
Results:
In the ophthalmology review (on keratokonus therapy), only 3 of 7 randomized controlled trials (RCTs) adjusted for IoD. Obtaining individual patient data (IPD) from a fourth RCT with a large data set allowed for IoD adjustment, so the MA included 4 rather than only 3 RCTs.
In the dentistry review (on periodontitis), several RCTs erroneously inflated the sample size by entering up to more than 100 values per patient when calculating group means, instead of aggregating the measurements and entering a single value per patient. Using a CC based on data from 2 epidemiological studies, we adjusted the results for the standard deviation of the effect estimates.
In the sleep medicine review (on therapy for obstructive sleep apnea), correction for IoD was made for studies with a cross-over design. Only a few studies had properly accounted for IoD. Using their data, we estimated the CCs for the MAs and conducted sensitivity analyses with the set of CCs to test the robustness of our results.
Conclusions:
There are several ways to solve problems with IoD when performing MAs. Ideally, IPD are available for re-analysis of study results and estimation of CCs. Alternatively, CCs from other studies included in the MA can be used to adjust the results of studies with IoD problems. Finally, study results based on multiple measurements per patient can be adjusted by applying CCs from other studies, even if these studies are not included in the SR.
Patient or healthcare consumer involvement:
For the described methodological procedures regarding meta-analyses, patient or healthcare consumer involvement is not applicable.
Some clinical study designs require that interdependency of data (IoD) be addressed in the reporting of study results. In self-controlled and cross-over trials, patients form part of both the intervention and control groups. In other studies, multiple disease locations are measured in the same patient. Because clinical studies often fail to account for IoD, one must apply specific statistical methods to the data of those studies when conducting a systematic review (SR) with a meta-analysis (MA). Otherwise, the results may be misleading.
Objectives:
This presentation describes different options for addressing and solving problems with IoD in MAs.
Methods:
We performed 3 SRs with MAs in medical fields where data clustering is common: ophthalmology (paired data on both eyes), dentistry (multiple data on single teeth), and sleep medicine (paired cross-over data in patients with chronic diseases). Before conducting the MAs, we assessed the studies included for IoD and corrected the data when IoD adjustment was necessary. If possible, we adjusted the results by using a correlation coefficient (CC) for the standard deviation of the effect estimates of the studies included.
Results:
In the ophthalmology review (on keratokonus therapy), only 3 of 7 randomized controlled trials (RCTs) adjusted for IoD. Obtaining individual patient data (IPD) from a fourth RCT with a large data set allowed for IoD adjustment, so the MA included 4 rather than only 3 RCTs.
In the dentistry review (on periodontitis), several RCTs erroneously inflated the sample size by entering up to more than 100 values per patient when calculating group means, instead of aggregating the measurements and entering a single value per patient. Using a CC based on data from 2 epidemiological studies, we adjusted the results for the standard deviation of the effect estimates.
In the sleep medicine review (on therapy for obstructive sleep apnea), correction for IoD was made for studies with a cross-over design. Only a few studies had properly accounted for IoD. Using their data, we estimated the CCs for the MAs and conducted sensitivity analyses with the set of CCs to test the robustness of our results.
Conclusions:
There are several ways to solve problems with IoD when performing MAs. Ideally, IPD are available for re-analysis of study results and estimation of CCs. Alternatively, CCs from other studies included in the MA can be used to adjust the results of studies with IoD problems. Finally, study results based on multiple measurements per patient can be adjusted by applying CCs from other studies, even if these studies are not included in the SR.
Patient or healthcare consumer involvement:
For the described methodological procedures regarding meta-analyses, patient or healthcare consumer involvement is not applicable.