A comparison of estimation methods for multilevel logistic models (Q1424625)

A multilevel logistic model is considered of the form \(y_{ij}=\pi_{ij}+e_{ij}\), where \(y_{ij}\) is the binary outcome for the \(j\)-th person in the \(i\)-th cluster, with conditional Bernoulli distribution, \(\pi_{ij}=(1+\exp(-\beta_0-\beta_1X_{ij}+ u_{0,j}+u_{1,j} X_{ij}))^{-1}\), \(x_{ij}\) are the treatment conditions, \(u_{0,j}\), \(u_{1,j}\) are unobserved normally distributed random terms, and \(\beta_i\) are unknown coefficients. Following methods of maximum likelihood estimation are considered: maximum likelihood estimation, approximate maximum likelihood which uses the first and the second order terms in the Taylor series expansion of \(\pi_{ij}\) (penalized quasi likelihood estimation, PQL). In the simulation studies the bias, variance and power of the corresponding heterogeneity tests are compared for these estimators. The authors conclusion is that the first order PQL is the best.