As statistical classifiers become integrated into real-world applications, it\nis important to consider not only their accuracy but also their robustness to\nchanges in the data distribution. In this paper, we consider the case where\nthere is an unobserved confounding variable $z$ that influences both the\nfeatures $\\mathbf{x}$ and the class variable $y$. When the influence of $z$\nchanges from training to testing data, we find that the classifier accuracy can\ndegrade rapidly. In our approach, we assume that we can predict the value of\n$z$ at training time with some error. The prediction for $z$ is then fed to\nPearl's back-door adjustment to build our model. Because of the attenuation\nbias caused by measurement error in $z$, standard approaches to controlling for\n$z$ are ineffective. In response, we propose a method to properly control for\nthe influence of $z$ by first estimating its relationship with the class\nvariable $y$, then updating predictions for $z$ to match that estimated\nrelationship. By adjusting the influence of $z$, we show that we can build a\nmodel that exceeds competing baselines on accuracy as well as on robustness\nover a range of confounding relationships.\n