Article type
Year
Abstract
Introduction: Individual patient data meta-analyses are nowadays the golden standard for conducting meta-analyses. When individual patient data cannot be retrieved for a trial, omitting the trial from the analysis should be avoided.
Objective: This paper investigates the possibility of including in an individual patient data meta-analysis the results of trials for which only published results are available.
Methods: Assuming the sample size of each treatment arm (N1, N2), the observed number of events per treatment arm (O1, O2) and the value of the logrank test statistic (LR=(O1-E1)^2/ Var(O1-E1)), or even only its associated p-value are available in the trial's publication, the following estimates of the quantity Var(O1-E1) are considered:
1) (O1+O2)/4
2) (O1+O2)N1N2/(N1+N2)^2
3) O1O2/(O1+O2)
The estimated value of |O1-E1| is SqrRoot(LR x V) where V is one of the above variance estimates.
Results: Based on simulations using these approximate estimates of the variance and the logrank statistic for each trial, it has been found that reliable estimates of the common hazard ratio and of the overall logrank p-value are obtained. These estimates of the common hazard ratio and of the overall logrank chi-square test statistic have the same limitations as the individual patient data estimates of the same quantities with regards to the proportional hazards assumption, to differences in follow-up between the treatment arms and to the presence of heterogeneous treatment effects.
Discussion: When other biases can be considered to be negligible, this method provides a reliable way of introducing literature based data in an individual patient data meta-analysis. Although these additional data contribute to the estimate of the treatment effect, they cannot be used in more refined analyses, for instance computation of overall Kaplan-Meier curve and subgroup analyses.
Objective: This paper investigates the possibility of including in an individual patient data meta-analysis the results of trials for which only published results are available.
Methods: Assuming the sample size of each treatment arm (N1, N2), the observed number of events per treatment arm (O1, O2) and the value of the logrank test statistic (LR=(O1-E1)^2/ Var(O1-E1)), or even only its associated p-value are available in the trial's publication, the following estimates of the quantity Var(O1-E1) are considered:
1) (O1+O2)/4
2) (O1+O2)N1N2/(N1+N2)^2
3) O1O2/(O1+O2)
The estimated value of |O1-E1| is SqrRoot(LR x V) where V is one of the above variance estimates.
Results: Based on simulations using these approximate estimates of the variance and the logrank statistic for each trial, it has been found that reliable estimates of the common hazard ratio and of the overall logrank p-value are obtained. These estimates of the common hazard ratio and of the overall logrank chi-square test statistic have the same limitations as the individual patient data estimates of the same quantities with regards to the proportional hazards assumption, to differences in follow-up between the treatment arms and to the presence of heterogeneous treatment effects.
Discussion: When other biases can be considered to be negligible, this method provides a reliable way of introducing literature based data in an individual patient data meta-analysis. Although these additional data contribute to the estimate of the treatment effect, they cannot be used in more refined analyses, for instance computation of overall Kaplan-Meier curve and subgroup analyses.