This study tested predictions from W. Ickes and J. A. Simpson's (1997, 2001) empathic accuracy model. Married couples were videotaped as they tried to resolve a problem in their marriage. Both spouses then viewed a videotape of the interaction, recorded the thoughts and feelings they had at specific time points, and tried to infer their partner's thoughts and feelings. Consistent with the model, when the partner's thoughts and feelings were relationship-threatening (as rated by both the partners and by trained observers), greater empathic accuracy on the part of the perceiver was associated with pre-to-posttest declines in the perceiver's feelings of subjective closeness. The reverse was true when the partner's thoughts and feelings were nonthreatening. Exploratory analyses revealed that these effects were partially mediated through observer ratings of the degree to which partners tried to avoid the discussion issue.