Since prior evidence pointed most strongly to age at onset and recurrence as predictors of risk of illness in relatives, we began with these variables. Model 1 included only age at onset, which significantly predicted the hazard rate in co-twins (χ21=13.8; P=.002; HR, 0.85; 95% CI, 0.78-0.93) (Table 1). Model 2 added number of episodes that also significantly predicted risk for MD in the co-twin. In model 3, we added (number of episodes)2, which was significant and negative, indicating a lower risk for MD with very high and very low numbers of episodes. Adding (number of episodes)2 to the model also strengthened the linear effect of number of episodes (HR from 1.18 to 1.98). In model 3, both number of episodes (χ2l=17.5; P<.001; HR, 1.98; 95% CI, 1.43-2.73) and (number of episodes)2 (χ2l=11.4; P=.007; HR, 0.78, 95% CI, 0.67-0.90) were substantially stronger predictors of risk of MD in the co-twin than was age at onset (χ2l=6.3; P=.01; HR, 0.89; 95% CI, 0.81-0.98). Model 4 added duration of longest episode, which also strongly predicted the hazard rate for MD in the co-twin (χ2l=13.5; P=.002; HR, 1.23; 95% CI, 1.10-1.38). Adding (length of episode)2, however, produced no improvement in model fit. Model 5 added recurrent thoughts of death/suicide, which was a unique predictor (χ2l=6.0; P=.01; HR, 1.23; 95% CI, 1.04-1.47). In this model, however, age at onset no longer significantly predicted risk for MD even at a trend level (χ2l= 2.2; P=.14; HR, 0.93; 95% CI, 0.85-1.02) and was dropped from the model. The final significant improvement in the model came with the addition of degree of impairment during the worst episode (χ2l=3.9; P=.05; HR, 1.13; 95% CI, 1.00-1.28). We attempted to then add to this model our remaining variables, but none were significant: age at onset (χ2l=−2.0; P=.16), level of distress (χ21=3.1; P=.08), number of A criteria (χ21=0.1; P=.81), and help seeking (χ21=0.4; P=.54).