*Bayesian Additive Regression Trees (BART) have been shown to be a flexible, non-parametric regression approach that captures non-linear interactions between covariates and response variables. Building on a multivariate extension of the BART framework, we propose a non-linear vector autoregressive model. To induce sparsity in the model, we incorporate a Dirichlet prior over the splitting variables, which effectively shrinks the predictor space by selecting only the most relevant (lagged) variables for building the trees. The resulting model is called SUDART. Through a comprehensive Monte Carlo study, we demonstrate that the estimated generalized impulse responses converge to their underlying true values for linear as well as non-linear data generating processes. This indicates that the SUDART model correctly recovers both the dynamic structure and the error covariance matrix of the true model, highlighting its flexibility with respect to the complexity of the data. The multivariate extension enables us to incorporate various (structural) identification methods beyond the order-dependent recursive Cholesky decomposition to identify macroeconomic shocks, including for instance identification based on external instruments. We illustrate the usefulness of the model in an empirical application on monetary policy shocks.

**To delay regulation, the tobacco industry pioneered scientific lobbying to encourage favorable findings, direct research agendas, and repress unfavorable publications. I aim to provide the first empirical analysis of how its $355 million in research funding between 1954 and 1998 affected research output, content, quality, and co-authorship networks. By linking internal tobacco scientific reports with bibliometric data, I built a novel database of researchers funded by tobacco companies during that time, including their publications, career trajectories, and collaboration links.