Are Neural Representation Learning Methods a Viable Alternative to TMLE for Causal Estimation?

{Simulation is used to evaluate the performance of deep learning and semiparametric causal estimators under realistic high- and low-dimensional data-generating mechanisms from epidemiologic studies.} Deep learning models that leverage representation learning have recently gained attention for estimating causal effects in observational studies, offering the ability to learn latent structures that adjust for confounding. However, their performance relative to established semiparametric estimators—such as Targeted Maximum Likelihood Estimation (TMLE)—remains unclear across different data complexities. We used plasmode simulations based on two real-world epidemiologic datasets (low- and high-dimensional) to compare three neural representation learning models (TARNET, Dragonnet, NEDnet) with benchmark methods, including inverse probability weighting, TMLE with Super Learner, and double cross-fit TMLE. Performance was assessed in terms of bias, variance, and confidence interval coverage. Results showed that representation learning methods like TARNET excelled in low-dimensional settings, while NEDnet achieved robust coverage in high-dimensional contexts but with higher variance. TMLE with smooth learners consistently yielded stable variance estimates. These findings inform when representation learning may offer advantages over classical semiparametric methods for causal inference.