Skip to main content

Languages

Calculating confidence intervals for the absolute risk reduction taking uncertainties of relative risk and baseline risk estimates into account

Date and Location

Session: 

P4.034

Date

Monday 23 September 2013 - 10:30 - 12:00

Location

Presenting author and contact person

Presenting author

Ralf Bender

Contact person

Ralf Bender
Abstract text
Background: In Cochrane reviews as well as in the GRADE system absolute estimates of treatment effect are frequently calculated by using relative risk (RR) estimates based on a meta-analysis in combination with an independent baseline risk (BR) estimate. Spencer et al. (BMJ 2012;345:e7401) pointed out that GRADE and all other systems for rating confidence in absolute treatment effect estimates do not fully address uncertainties in BR estimates. Calculations of confidence intervals for the absolute risk reduction (ARR) currently performed under the GRADE framework take into account the imprecision of the RR estimate, but not that of the BR estimate. Objectives: The aims of this paper are firstly, to show that a method for interval estimation of ARR is available that takes the uncertainties of RR and BR estimates into account and secondly, to discuss the impact of accounting for both uncertainties on the confidence in absolute estimates of the treatment effect. Methods: If BR and RR are estimated from different independent sources, confidence limits for ARR can be calculated from those for BR and RR by a procedure called method of variance estimates recovery (MOVER-R). This method is explained and applied to examples. The resulting confidence intervals are compared to those obtained by the method currently used by GRADE, and to those obtained by the naive method of directly combining the confidence limits for RR and BR. Results: Neglecting the uncertainty of BR estimates leads to confidence intervals which are too narrow whereas the naive method of directly combining the confidence limits for RR and BR results in confidence intervals which are unnecessarily too wide. Conclusions: A simple and effective method is available to calculate confidence intervals for ARR from independent interval estimates of BR and RR. This method should be applied in practice.