Premium
Measures of single arm outcome in meta‐analyses of rare events in the presence of competing risks
Author(s) -
Andreano Anita,
Rebora Paola,
Valsecchi Maria Grazia
Publication year - 2015
Publication title -
biometrical journal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.108
H-Index - 63
eISSN - 1521-4036
pISSN - 0323-3847
DOI - 10.1002/bimj.201400119
Subject(s) - statistics , cumulative incidence , censoring (clinical trials) , mathematics , generalized linear mixed model , context (archaeology) , econometrics , incidence (geometry) , mixed model , sample size determination , multivariate statistics , meta analysis , generalized linear model , medicine , biology , paleontology , cohort , geometry
When performing single arm meta‐analyses of rare events in small populations, if the outcome of interest is incidence, it is not uncommon to have at least one study with zero events, especially in the presence of competing risks. In this paper, we address the problem of how to include studies with zero events in inverse variance meta‐analyses when individual patient data are not available, going beyond the naïve approach of not including the study or the use of a continuity correction. The proposed solution is the arcsine transformation of the crude cumulative incidence as its approximate variance, which is inversely proportional to the sample size, can be calculated also for studies with a zero estimate. As an alternative, generalized linear mixed models (GLMM) can be used. Simulations were performed to compare the results from inverse variance method meta‐analyses of the arcsine transformed cumulative incidence to those obtained from meta‐analyses of the cumulative incidence itself and of the logit transformation of the cumulative incidence. The comparisons have been carried out for different scenarios of heterogeneity, incidence, and censoring and for competing and not competing risks. The arcsine transformation showed the smallest bias and the highest coverage among models assuming within study normality. At the same time, the GLMM model had the best performance at very low incidences. The proposed method was applied to the clinical context that motivated this work, i.e. a meta‐analysis of 5‐year crude cumulative incidence of central nervous system recurrences in children treated for acute lymphoblastic leukemia.