Estimating causal relation using observational survey data or non-experimental data has been a challenging task. Likewise, satisfying conditional independence over unobservable is an untestable and restrictive assumption. This paper uses machine and deep learning algorithms to resolve the fundamental problem of causal effect using non-experimental data. By satisfying a “sufficient” assumption of exogeneity the model identifies instances from the database that mimics the Quasi-Experimental Designs. This helps in estimating the complete distribution of counterfactuals for causal inference. The paper exploits the asymptotic properties – weak monotonicity and identity function – to show that average treatment effect (ATE) and conditional treatment effect (CATE) converges to true value and are consistent and unbiased. The assumption of connectedness of potential outcomes in two periods is central for model selection. The model uses the intelligence of system to look for units in population database and select that subset of data which satisfies the sufficient condition for casual inference (assumption of exogeniety). The study also suggests that an availability of well-defined database with extensive features and continuous time interval in developing countries will help to estimate the impact of key welfare schemes/programs. This paper addresses the casual effect inference for classification problem. The study estimates the heterogeneous treatment effect using differential eligibility rules of national health program (NRHM) over maternal health outcomes in India. It uses scenario of panel data from Indian Human Development Survey (2004-05 & 2011-12). The findings suggest that potency of intervention for increasing institutional delivery. However, the CATE for antenatal (ANC), postnatal (PNC) and Safe-Delivery indicates that intervention failed to remove the inequalities

Heterogeneous Treatment Effect; Machine Learning; Causal Inference; QuasiExperimental Design; Maternal Health