Article type
Year
Abstract
Introduction: In order to make the results of meta-analyses interpretable by a wide audience the summary plot has to be as self explanatory as possible. The graphical displays summarizing the results of meta-analyses are not always easy to interpret.
Objective: The objective of this paper is to present some way of simplifying the plots that summarize the results of meta-analyses.
Methods: The following modifications of the forest plot (graphical summary) are suggested: 1. deletion of the (O-E) and Var(O-E) columns; 2. addition of the survival benefit; 3. systematic use of the log scale; 4. use of 95% confidence intervals for the hazard ratio.
Results:
1. The tabular part of a forest plot provides O-E and Var(O-E) for each study but the graphical part contains similar information. Therefore the forest plot is overloaded by statistical results (which belong in an appendix) that are difficult for non-statisticians to interpret.
2. The survival benefit is the difference between the survival estimates at a given point in time. This gives medically relevant information regarding the comparison of two treatments. In meta-analyses of individual patient data the survival benefit should be presented in the forest plot.
3. As Gailbrath (Stats. in. Med. 1988; 7: 889-894) pointed out the x-axis of the graphical part of the forest plot should be on a logarithmic scale because an arithmetic scale visually amplifies the right hand side of the plot.
4. Another confusing issue in forest plots is the confidence interval about the hazard ratio. To control the type I error for multiple testing some people give 99% confidence intervals for the individual studies while they give 95% confidence intervals for the overall results. From a practical point of view it should be the opposite. For a meta-analysis to be meaningful the overall p-value should be convincingly small.
Objective: The objective of this paper is to present some way of simplifying the plots that summarize the results of meta-analyses.
Methods: The following modifications of the forest plot (graphical summary) are suggested: 1. deletion of the (O-E) and Var(O-E) columns; 2. addition of the survival benefit; 3. systematic use of the log scale; 4. use of 95% confidence intervals for the hazard ratio.
Results:
1. The tabular part of a forest plot provides O-E and Var(O-E) for each study but the graphical part contains similar information. Therefore the forest plot is overloaded by statistical results (which belong in an appendix) that are difficult for non-statisticians to interpret.
2. The survival benefit is the difference between the survival estimates at a given point in time. This gives medically relevant information regarding the comparison of two treatments. In meta-analyses of individual patient data the survival benefit should be presented in the forest plot.
3. As Gailbrath (Stats. in. Med. 1988; 7: 889-894) pointed out the x-axis of the graphical part of the forest plot should be on a logarithmic scale because an arithmetic scale visually amplifies the right hand side of the plot.
4. Another confusing issue in forest plots is the confidence interval about the hazard ratio. To control the type I error for multiple testing some people give 99% confidence intervals for the individual studies while they give 95% confidence intervals for the overall results. From a practical point of view it should be the opposite. For a meta-analysis to be meaningful the overall p-value should be convincingly small.