Background: Previous studies found inconsistencies between available between-trial variance estimators in meta-analyses, but implications to meta-analytical inferences in practice are limited since these studies only investigated bias and mean square error of the between-trial variance parameter. Meta-analytical inferences in practice are, however, chiefly based on pooled summary effect estimates and their conficende intervals. Objectives: To empirically compare meta-analytical inferences in practice from three selected between-trial variance estimators: the traditional Dersimonial-Laird (DL) estimator and two other estimators previously shown to perform more optimal than the DL estimator; a model error variance (MV) based estimator; and a variance component (VC) based estimator. Methods: We searched The Cochrane Library Issue 2, 2007 for heterogenous meta-analyses (I² > 25%) that reported a binary outcome among the first three outcome measures in the first comparison between the experimental intervention and the control intervention. A total of 296 meta-analyses were eligible. We recorded meta-analysis specific variables: heterogeneity (I²), control group event proportion, estimated pooled summary effect, number of trials, and number of participants. We metaanalysed the data using the random-effects model and obtained metaanalytical inferences following from use of the three different between trial variance estimators. We assessed disagreements by measuring percental deviation in summary effect estimates and width of confidence intervals pairwise and calculated the proportion of disagreements with respect to a 25%, 50%, and 100% threshold. We also performed logistic regression based on the recorded variables. Results: We found large pair wise disagreements in the width of confidence intervals and moderate differences in pooled summary effect estimates. The number of trials and the number of participants predicted the magnitude of disagreements. In general, disagreements in confidence intervals is an increasing function of statistical information (i.e. number of trials an participants), and disagreements in pooled summary effect estimates is a decreasing function hereof. Conclusions: Meta-analytical random-effects inferences in practice are highly affected by the choice of between-study variance estimator. Our findings provide useful guidelines for interpretations and sensitivity analysis.