Article type
Year
Abstract
Background: Sometimes multiple-arm trials only report the pairwise difference between each arm and the control (with a 95% confidence interval (CI)).
These paired differences can be individually entered into Review Manager 5 using the Generic Inverse Variance (GIV) method, but they cannot be combined in the same forest plot, as the control arm would be counted more than once.
Objectives: To find a simple adjustment that allows the paired comparisons to be combined in the same forest plot by adjusting the standard error (SE).
Methods: From first principles it is possible to derive a formula for adjusting the SE in each paired comparison in order to allow pooling in a forest plot. The adjustment is based on the assumption of equal variance in each of the trial arms. The method should not be used if the size of each trial arm is clearly unequal.
Results:
- If there are N active arms in the study, the SE for each paired comparison should be multiplied by the square root (SQRT) of (N+1)/2)). So for a three-arm trial with two active arms and one control, the SE for each of the two paired comparisons should be multiplied by SQRT(3/2) = 1.225.
- Where it is possible to calculate the average of the mean differences (for example) across the paired comparisons, and each paired comparison has a similar 95% CI, then the appropriate SE for the combined comparison groups compared with control can be obtained (under the same assumptions of equal variance across the arms) by dividing the paired SE by SQRT ((N+1)/2N).
Conclusions: A simple method is available to allow pooling of data from multi-arm trials in the same forest plot using GIV analysis. However, if a random-effects meta-analysis is being used, combining the active arms will give more weight to the study data than separate paired comparisons.
These paired differences can be individually entered into Review Manager 5 using the Generic Inverse Variance (GIV) method, but they cannot be combined in the same forest plot, as the control arm would be counted more than once.
Objectives: To find a simple adjustment that allows the paired comparisons to be combined in the same forest plot by adjusting the standard error (SE).
Methods: From first principles it is possible to derive a formula for adjusting the SE in each paired comparison in order to allow pooling in a forest plot. The adjustment is based on the assumption of equal variance in each of the trial arms. The method should not be used if the size of each trial arm is clearly unequal.
Results:
- If there are N active arms in the study, the SE for each paired comparison should be multiplied by the square root (SQRT) of (N+1)/2)). So for a three-arm trial with two active arms and one control, the SE for each of the two paired comparisons should be multiplied by SQRT(3/2) = 1.225.
- Where it is possible to calculate the average of the mean differences (for example) across the paired comparisons, and each paired comparison has a similar 95% CI, then the appropriate SE for the combined comparison groups compared with control can be obtained (under the same assumptions of equal variance across the arms) by dividing the paired SE by SQRT ((N+1)/2N).
Conclusions: A simple method is available to allow pooling of data from multi-arm trials in the same forest plot using GIV analysis. However, if a random-effects meta-analysis is being used, combining the active arms will give more weight to the study data than separate paired comparisons.