Adjusted Risk Difference Estimation: An Assessment of Convergence Problems with Application to Malaria Efficacy Studies

A common measure of treatment effect in malaria efficacy studies is the risk difference, which can be estimated using binomial regression models. These models can fail to provide estimates, however, due to model failure or model convergence problems. Such failure most commonly occurs when the rate is close to 0% or 100% (a “boundary problem”) but can also occur occasionally even when the rate is not close to a boundary. This paper reports the findings from simulation studies performed to evaluate the factors that may contribute to model failure when using binomial regression to derive risk difference estimates.

Convergence rates were found to fall:

i) As one or both efficacy rates moved towards a boundary value, irrespective of the number of covariates included in the model;

ii) As the numbers of covariates in the model increased;

iii) As the levels of correlation between covariates the covariates increased. In all circumstances, convergence was poor when the efficacy rate in either group was 90% or more.