Article type
Year
Abstract
Introduction:
Objectives: To explore the reasons for differences between individual patient and literature based meta-analyses in the Meta-Analysis of Chemotherapy in Head and Neck Cancer (MACH-NC).
Methods: A meta-analysis was first performed on all individual patient data (IPD). In different cumulative steps, the EPD database was reduced to resemble more closely the literature data and at each step an analysis was done. First, the ineligible patients were removed. Second, trials for which no survival information could be obtained from the publication were removed. Third, mortality (Mantel Haenszel test (X^2MH) and Odds Ratio (OR)) rather than time to event (stratified log-rank test (X^2LR) and Hazard Ratio (HR)) was considered as this is often the only information available from the publication. Fourth, the maximum follow-up was reduced to the publication date less one year.
Results: The IPD included 10850 patients from 63 trials and resulted in HR=0.895 and X^2LR = 21.7. When excluding the 464 ineligible patients a larger difference was observed (HR=0.885, X^2LR =25.6) which represents exclusion bias. Exclusion of the 14 trials without survival information led to similar results (HR=0.889, X^2LR =19.01). When based on 5 years mortality rates, the chi-square statistic decreases substantially (OR=0.84, X^2MH =12.2). For the reduced follow-up, the chi-square statistic is closer to the X^2LR value of 19.01 (OR=0.83, X^2MH =16.3).
Discussion: Although all meta-analyses yielded a significant treatment difference, the strength of evidence varied substantially and was largest for individual patients meta-analysis which is based on the stratified log-rank test.
Objectives: To explore the reasons for differences between individual patient and literature based meta-analyses in the Meta-Analysis of Chemotherapy in Head and Neck Cancer (MACH-NC).
Methods: A meta-analysis was first performed on all individual patient data (IPD). In different cumulative steps, the EPD database was reduced to resemble more closely the literature data and at each step an analysis was done. First, the ineligible patients were removed. Second, trials for which no survival information could be obtained from the publication were removed. Third, mortality (Mantel Haenszel test (X^2MH) and Odds Ratio (OR)) rather than time to event (stratified log-rank test (X^2LR) and Hazard Ratio (HR)) was considered as this is often the only information available from the publication. Fourth, the maximum follow-up was reduced to the publication date less one year.
Results: The IPD included 10850 patients from 63 trials and resulted in HR=0.895 and X^2LR = 21.7. When excluding the 464 ineligible patients a larger difference was observed (HR=0.885, X^2LR =25.6) which represents exclusion bias. Exclusion of the 14 trials without survival information led to similar results (HR=0.889, X^2LR =19.01). When based on 5 years mortality rates, the chi-square statistic decreases substantially (OR=0.84, X^2MH =12.2). For the reduced follow-up, the chi-square statistic is closer to the X^2LR value of 19.01 (OR=0.83, X^2MH =16.3).
Discussion: Although all meta-analyses yielded a significant treatment difference, the strength of evidence varied substantially and was largest for individual patients meta-analysis which is based on the stratified log-rank test.