1: \begin{abstract}

2: % Endogeneity in explanatory variables usually arises from omitted variables, measurement error, and simultaneity \citep{wooldridge2010econometric}. However, traditional prediction-based machine learning methods fail to deal with endogeneity and could lead to false causal discovery. 

3: Structural equation models (SEMs) are widely 

4: used in sciences, ranging from economics to psychology,

5: to uncover causal relationships underlying a complex system

6: under consideration and estimate structural parameters of interest. 

7: We study estimation in a class of generalized SEMs where the object 

8: of interest is defined as the solution to a linear operator equation.

9: We formulate the linear operator equation as a min-max game, where both 

10: players are parameterized by neural networks (NNs), and learn the

11: parameters of these neural networks using the stochastic gradient descent.

12: We consider both 2-layer and multi-layer NNs with ReLU activation 

13: functions and prove global convergence in an overparametrized regime, where

14: the number of neurons is diverging. The results are established using 

15: techniques from online learning and local linearization of NNs,

16: and improve in several aspects the current state-of-the-art. For the first 

17: time we provide a tractable estimation procedure for SEMs

18: based on NNs with provable convergence and without the need for sample

19: splitting.

20: \end{abstract}

21: