To prevent diseases and reduce medical spending, many countries conduct general health checks while their effectiveness remains controversial. People voluntarily decide whether to participate in general health checks and, if ignored, this generates an endogeneity problem in the estimation of the effect of general health checks. In econometrics and statistics, instrumental variables are used to tackle this problem but classical approaches are oftentimes limited to capturing simple linear associations and overlook more complex nonlinear relationships among variables. To overcome this problem, we propose deep neural networks with two-stage structure that contains uncertainty aware attention. By using the proposed approach, we can fully leverage meaningful relationships among variables and handle the endogeneity problem while maintaining interpretable characteristics of the model. After the bias correction, we show that the effect of health checks on medical expenses turns out to be small. We also explore a pruning approach where the uncertainty of attention is used as a pruning criterion, which is analogous to the statistical significance in classical statistics. We find that the model performance improves without retraining when the proposed pruning is applied.