<p>Estimating long-term causal effects from long-term observational and short-term experimental data is challenging yet crucial for applications such as marketing and policy-making. Despite the success in certain applications, existing estimators are prone to biased results and cannot be adapted to unseen domains if the data distribution shifts across the domains. To address this problem, first, we propose an invariant surrogate representation learning method to address the distribution shift, in which, the surrogate representation is disentangled into the domain-level latent surrogates, the unit-level latent surrogates, and the invariant latent surrogates. Then, we devise a long-time causal effects estimator based on the learned invariant representation. We further theoretically show that these latent covariates are identifiable under mild assumptions, which ensures the correctness of the learned invariant surrogate representation and the estimated long-term causal effects. Extensive experiments on two real-world datasets demonstrate our method’s effectiveness, and the visualization of the learned representation further verifies the soundness of the invariant surrogate representation learning approach.</p>