LAUSR

Kassahun et al. [1] proposed a two-level marginalized hurdle combined model for analysis of zeroinflated overdispersed correlated count data. Specifically, the zero-inflation in the count data is accounted for by utilizing a two-part hurdle Poisson regression model, where the first part models a probability function for the zero state and the second part defines a probability function for the non-zero counts based on a zero-truncated Poisson distribution [2]. The extra-Poisson variation in the data is represented by including an overdispersion parameter and the correlation structure arising from the repeated measurements of subjects is taken into account via including random effects into the model. Following Heagerty [3], the whole regression model is built up by a two-level marginalized regression modeling approach, where the first level refers to the marginal mean model, whereas the second level is the conditional mean model depending on random effects. As a consequence of this model structure, the regression parameters in the first level of the model have population-averaged interpretations [1, 3], which is a desirable advantage over the traditional random effects models with subject-specific interpretations [4]. One of the main building blocks of marginalized multilevel models is that the levels are connected to each other via connector functions, which may have closed-form expressions or not. This particularly depends how the link function used in the second level of the model is conjugate with the distribution imposed on the random effects [5]. In this study, we would like to bring to the authors’ attention that the second connector function presented in their model cannot be analytically derived and needs to be solved via numerical optimization methods. To illustrate, let Yij denote the response of ith subject at jth time. In the two-level marginalized hurdle combined model of Kassahun et al. [1], the second level takes the following form: