Different Methods for Modelling Severe Hypoglycaemic Events: Implications for Effectiveness and Cost-Effectiveness Analyses

ObjectivesPublished clinical trials report severe hypoglycaemic events in different ways. Some report number of patients who suffered at least one event out of total number randomised and others report number of events for a given total exposure. The different data types can be modelled in different ways; therefore, three models have been used in published Bayesian Network Meta Analyses (NMAs) of hypoglycaemic events; models with a binomial likelihood reporting odds ratios (using a logit link) or hazard ratios (using the complementary log log link) and models with a Poisson likelihood reporting hazard ratios. The objective of this paper is to establish the impact of using different models on effectiveness estimates and the outputs from cost-effectiveness models.MethodsWe analysed a dataset used in a recent NMA conducted to inform NICE guideline recommendations regarding insulin choice for patients with type 1 diabetes using the three previously used models, plus a shared parameter model combining different types of data.ResultsThe relative treatment effects are similar regardless of which model or scale is used. Differences were seen when the probability of having an event on the baseline treatment was calculated using the different models with the logit model giving a baseline probability of 0.07, the clog-log 0.17 and the Poisson 0.29. These translate into differences of up to £110 in the cost of a hypoglycaemic event and 0.004 in associated disutility when calculating the absolute probabilities of an event to use in an economic model.ConclusionsWhile choice of outcome measure may not have a significant impact on relative effects for this outcome, care should be taken to ensure that the baseline probabilities used in an economic model are realistic and accurate to avoid over or underestimating costs and effects.