NONPARAMETRIC CLASSES OF WEIGHT-FUNCTIONS TO MODEL PUBLICATION BIAS

This paper addresses the use of weight functions to model publication bias in metaanalysis. Since publication bias is hard to gauge, a nonpa rametric epsilon-contamination class of weight functions is introduced . Sensitivity of conclusions to the specification of the weight functi on is explored by examining the range of results for the entire E-cont amination class. First, lower bounds are found on the coverage of conf idence intervals. If little publication bias is suspected, results are robust even when considered over the entire epsilon-contamination cla ss. However, if more substantial publication bias is suspected, then t he coverage provided by the usual interval estimator is not robust. In this case, an alternative interval estimator is suggested. Secondly, for the case in which prior information is available, upper and lower bounds are found on posterior quantities of interest.