Article type
Year
Abstract
Practically, methods of undertaking a meta-analysis may involve collecting either aggregate data, or data on each patient individually. The advantages of the latter approach, described as the "yardstick", include (i) a more complete analysis of time-to-event outcomes and (ii) a more powerful analysis of whether treatment is more or less effective in particular subgroups (Stewart and Pannar 1993).
A frequent approach to analysing individual patient data from randomised controlled trials with time-to-event endpoints has been to use stratified (by trial) logrank methods (Early Breast Cancer Trialists' 1990). Stratified regression models have been seldomly employed despite their potential to investigate sources of heterogeneity. To allow for uncertainty in the estimation of the between-trial variance component, a Bayesian approach to random effects meta-analysis with individual patient data is proposed, extending existing programs within the BUGS software for time-to-event outcomes (Gilksetal 1994).
A meta-analysis of five trials comparing two antiepileptic drugs, carbamazepine and sodium valproate, is used to illustrate and compare the above methods. For the outcome time-to-12-month-remission, the Q-statistic for homogeneity of treatment effects across trials is 9.21, p=0.06, warranting further investigation. In this example, there was strong a priori interest surrounding a possible interaction between treatment effect and type of epilepsy.
Problems in implementing BUGS for Cox regression modelling in this dataset are described and possible solutions suggested. Inevitably due to the small number of trials in this example, estimation of the between-trial variance component is poor and emphasis is placed on the results from the stratified regression modelling approach.
Stewart LA and Parmar MKB. Meta-analysis of the literature or of individual patient data: is there a difference? The Lancet 1993; 341: 418-422 Early Breast Cancer Trialists' Collaborative Group. Treatment of early breast cancer, vol 1: worldwide evidence 1985-1990. Oxford: Oxford University Press 1990 Gilks WR, Thomas A and Spiegelhalter DJ. A language and program for complex Bayesian modelling. Statistician 1994; 43: 169-177
A frequent approach to analysing individual patient data from randomised controlled trials with time-to-event endpoints has been to use stratified (by trial) logrank methods (Early Breast Cancer Trialists' 1990). Stratified regression models have been seldomly employed despite their potential to investigate sources of heterogeneity. To allow for uncertainty in the estimation of the between-trial variance component, a Bayesian approach to random effects meta-analysis with individual patient data is proposed, extending existing programs within the BUGS software for time-to-event outcomes (Gilksetal 1994).
A meta-analysis of five trials comparing two antiepileptic drugs, carbamazepine and sodium valproate, is used to illustrate and compare the above methods. For the outcome time-to-12-month-remission, the Q-statistic for homogeneity of treatment effects across trials is 9.21, p=0.06, warranting further investigation. In this example, there was strong a priori interest surrounding a possible interaction between treatment effect and type of epilepsy.
Problems in implementing BUGS for Cox regression modelling in this dataset are described and possible solutions suggested. Inevitably due to the small number of trials in this example, estimation of the between-trial variance component is poor and emphasis is placed on the results from the stratified regression modelling approach.
Stewart LA and Parmar MKB. Meta-analysis of the literature or of individual patient data: is there a difference? The Lancet 1993; 341: 418-422 Early Breast Cancer Trialists' Collaborative Group. Treatment of early breast cancer, vol 1: worldwide evidence 1985-1990. Oxford: Oxford University Press 1990 Gilks WR, Thomas A and Spiegelhalter DJ. A language and program for complex Bayesian modelling. Statistician 1994; 43: 169-177