We develop a novel dose-finding method for two-agent combination phase I trials on the basis of the shrunken predictive probability of toxicity. In this method, a shrinkage logistic regression model that allows distinct shrinkage multipliers for the coefficients of the main effects of two agents and their interaction on the probability of toxicity constructs the toxicity outcome. We also propose dose-escalation/de-escalation decision rules on the basis of the shrunken predictive probability of toxicity. Simulation studies under various patterns of monotonic dose-response relationships for combinations of two agents demonstrated that the proposed method performed no worse than the existing two dose-finding methods we selected. Copyright © 2013 John Wiley & Sons, Ltd.