Dr. Giuseppe Alfonzetti presents a novel stochastic composite likelihood estimator designed to efficiently handle complex statistical models with intractable likelihoods. The method approximates the traditional composite likelihood through a sequence of stochastic updates, where lower-dimensional likelihood components are iteratively subsampled according to a defined sampling scheme. The estimator’s variance reflects both data variability and optimization noise, the latter being influenced by the sampling design. The talk highlights the advantages of using sampling schemes that respect the composite structure of the likelihood, demonstrating superior performance over standard stochastic gradient descent (SGD) methods across three models: a binary graphical model, a gamma-frailty model for count data, and a categorical factor model where standard SGD is not applicable.