The proliferation of genomics data and broad adoption of health information systems such as electronic medical records in the past decade present opportunities for data science to elucidate important causal mechanisms in human health. Mendelian randomization (MR) is one possible approach for deducing causal effects, even in the presence of unmeasured confounding, by leveraging on genetic markers as instrumental variables (IV). However the assumptions for an IV to be valid in MR often require an unrealistic level of knowledge about the underlying biological mechanisms. As a result, possible violation of these assumptions can seldom be ruled out. In this talk, we introduce a new class of IV estimators which are robust to violation of the IV assumptions under a large collection of data generating mechanisms consistent with parametric models commonly assumed in the MR literature. For illustration, the methods are used to estimate the causal effect of type 2 diabetes mellitus on cognitive functioning based on data from the Health and Retirement Study.