Finally, the time-to-onset measure available to us was incomplete: at each wave, only drinking during the 30 days prior to the assessment was obtained. This would lead to some instances of missed onset, biasing the overall rate of onset in a downward direction. Fortunately, a proportional hazard model can still estimate relative hazard rates accurately, provided that drinking during the observed periods is not correlated with predictors. This could not be empirically addressed, given the uniform absence of data from one assessment to 30 days prior to the next, but we see no strong reason to expect such correlations.
