We examine central bank intervention in foreign exchange markets using a dynamic censored regression model. We allow the amount of purchase and sale interventions to depend nonlinearly upon lagged values of intervention and on measures of disorderly foreign exchange markets. Using data for the CBRT, we find persistence in interventions, which may suggest the presence of political costs and/or a signal of future monetary policy. We find strong evidence of nonnormality and heteroskedasticity in the Tobit model of the reaction function. Results using a robust estimator reveal the importance of considering these specification issues when modeling central bank intervention.