Least squares moment identification of binary regression mixture models (Q2036315)

The authors consider a mixture of \(K\) binary regression models. A first main result is that under weak assumptions cross moments up to order 3 between output and regression variables are enough to identify all parameters of the model. The second contribution of the authors is a new least squares moment method to estimate the model parameter \(\Theta\): The estimator \(\hat\Theta\) minimizes a generalized square distance of theoretical moments to empirical moments. Consistency and asymptotic normality of \(\hat\Theta\) is proved. For computation of \(\hat\Theta\) they use their R-package \textit{Morpheus}. The authors then compare experimentally their method to the maximum likelihood estimator and discuss advantages and disadvantages of both. Finally they investigate the identifiability of other mixture models and prove under certain assumptions identifiability for continuous covariates, for continuous and categorial covariates and for longitudinal observations.