survey, again producing an incorrectly delayed time of onset. Although a more comprehensive measure of onset would have been ideal, this procedure would likely give rise to a bias only in estimating the absolute rate of onset. However, our proportional hazard approach (Cox, 1972) considers relative rates of onset for different values of predictors. Censoring would not bias such a model unless predictor values were correlated with observed versus unobserved periods. This seemed an unlikely concern for the present analysis.
