We consider how fear of model misspecification on the part of the planner and/or the households affects welfare gains from optimal macroprudential taxes in an economy with occasionally binding collateral constraints as in Bianchi (2011). On the one hand, there exist welfare gains from internalizing how borrowing decisions in good times affect the value of collateral during a crisis. On the other hand, interventions by a robust planner that has in mind a model far from the true underlying distribution of shocks, can result in negligible welfare gains, or even losses. This is because a policy that is robust to misspecification, as in Hansen and Sargent (2011), is optimal under a "worst-case'' scenario but not under alternative distributions of the state. A robust planner introduces taxes that are 5 percentage points higher but does not achieve a significant increase in welfare gains compared to a non-robust planner when the true underlying model is not the worst-case. If households also make choices that are robust to model misspecification, the gains are significantly reduced and a highly-robust planner "underborrows" and induces welfare losses. If, however, the worst-case scenario is indeed realized, then welfare gains are the largest possible.