Title: Exact finite sample inference for studies with a small number of clusters
Authors: Stephane Heritier - Monash University (Australia) [presenting]
Maria-Pia Victoria-Feser - University of Geneva (Switzerland)
Stephane Guerrier - University of Geneva (Switzerland)
Abstract: Cluster randomised trials (CRTs), particularly longitudinal CRTs, often generate data with a small number of clusters typically analysed using generalised estimating equations (GEE) or generalised mixed models. A major drawback of standard techniques is that they are asymptotic in nature and rely on a large number of clusters to be valid. Ignoring this leads to: 1) a grossly-inflated type 1 error or, in general, confidence intervals (CIs) that are too short; 2) biased variance or intra-cluster correlation (ICC) estimates. We propose a new simulation-based approach allowing exact finite-sample inference for such problems. The idea is to compute first an initial simple(r) estimator, possibly biased, of the parameter of interest. In a second step, simulations are used to correct the bias via a novel algorithm called the Iterative Bootstrap (IB). The finite sample distribution can be generated under weak regularity conditions. We study the performance of this approach by simulations in various settings and show that it outperforms standard methods based on asymptotic sandwich variance formula with/without small sample correction. This indirect approach shows extremely promising results both theoretically and empirically.