Article type
Year
Abstract
Objectives: To evaluate the usefulness of the Kaplan Meier technique to estimate the number of events to be used as summary statistics for inclusion in a literature based meta-analysis.
Methods: The theoretical relationship between the number of events estimated from the Kaplan Meier curve and the true number of events is studied. We further investigate the effect of the deviation of the estimate from the true number of events on the Mantel Haenszel (MH) test. Finally, as an example, the MH test is derived for mortality at 2 and 5 years from both true and estimated number of events for an individual patient data Meta-Analysis of Chemotherapy in Head and Neck Cancer (MACH-NC, ASCO 1998: 386a).
Results: When there is loss to follow-up, the number of events is always overestimated. Overestimation of the numbers of events increases with fewer patients at risk and as more patients are lost to follow up. This overestimation can have a large effect on the MH test. In MACH-NC, the true MH test at 2 years has a Chi-square value of 3.96 which is estimated as 5.03, whereas the true MH test at 5 years has a Chi-square value of 6.26 which is estimated as 19.11. Thus, as expected, the overestimation is much more pronounced at 5 years, explaining most of the difference between mortality at 2 and 5 years
Discussion: Based on both theoretical and practical evidence, it is therefore recommended not to estimate the number of events from the Kaplan Meier curves to derive the MH test unless there is virtually no loss to follow-up and if there is still a substantial number of patients at risk at the particular moment time at which mortality is determined.
Methods: The theoretical relationship between the number of events estimated from the Kaplan Meier curve and the true number of events is studied. We further investigate the effect of the deviation of the estimate from the true number of events on the Mantel Haenszel (MH) test. Finally, as an example, the MH test is derived for mortality at 2 and 5 years from both true and estimated number of events for an individual patient data Meta-Analysis of Chemotherapy in Head and Neck Cancer (MACH-NC, ASCO 1998: 386a).
Results: When there is loss to follow-up, the number of events is always overestimated. Overestimation of the numbers of events increases with fewer patients at risk and as more patients are lost to follow up. This overestimation can have a large effect on the MH test. In MACH-NC, the true MH test at 2 years has a Chi-square value of 3.96 which is estimated as 5.03, whereas the true MH test at 5 years has a Chi-square value of 6.26 which is estimated as 19.11. Thus, as expected, the overestimation is much more pronounced at 5 years, explaining most of the difference between mortality at 2 and 5 years
Discussion: Based on both theoretical and practical evidence, it is therefore recommended not to estimate the number of events from the Kaplan Meier curves to derive the MH test unless there is virtually no loss to follow-up and if there is still a substantial number of patients at risk at the particular moment time at which mortality is determined.