This study examines the effect of point of reference on the predictive validity of self-rated health for mortality in a 5-year follow-up period. Two self-rated health measures are examined: an age group comparative question and a global question with no explicit point of reference. The baseline data (SweOld) is a nationally representative interview survey among Swedish people aged 77+ in 1992. Mortality for the 1992-1996 period was analysed using Cox proportional hazards regression models. Age-referential self-rated health was found to be a better predictor of elderly men's mortality both in non-adjusted models and in models adjusting for age and both self-rated health measures. In separate analyses, both measures were found to be equally strong predictors of women's mortality. When adding both measures into the model simultaneously, the age-referential question lost much of its predictive power. The findings suggest that self-rated health measures are not insensitive to differences in question wording.