This thesis studies the identification of structural or causal relationships in the presence of unob- served heterogeneity or/and endogeneity using two different approaches. In the first approach, multi-dimensional panel data is exploited to control for unobserved heterogeneity that is not only individual specific. This can be complemented with the use of instrumental variables. In the second approach, machine learning techniques in combination with orthogonal scores are exploited to identify parameters of interest in models without exclusion restrictions for cross- sectional data. This is useful when conditioning on unobserved specific heterogeneity is not enough or possible and when instrumental variables are not available. More specifically, the first chapter focuses on random coefficients model for multi-dimensional panel data, the second chapter investigates dynamic heterogeneous panel data models and the third chapter focuses on triangular models without exclusion restrictions for cross-sectional data.