Table 3.
Applications of the l1-penalized least squares with hierarchical group LASSO variable selection (l1-PLS-HGL), l1-penalized least squares with group LASSO variable selection (l1-PLS-GL), adaptive contrast weighted learning (ACWL), and direct learning (D-learning) to data from male donors in the INTERVAL trial assuming the target is to maximize the utility. The trade-off parameter b in the utility function varies from 1 to 5 at an increment of 1.
|
Assignment Percentages |
ITR Effects |
||||||
|---|---|---|---|---|---|---|---|
| Trade-off Parameter | Method | 12 weeks | 10 weeks | 8 weeks | Donation | Deferral | Utility |
| b = 1 | l1-PLS-HGL | 0.9 (0.1) | 1.2 (0.1) | 97.9 (0.1) | 1.309 (0.008) | 0.024 (0.001) | 1.064 (0.009) |
| l1-PLS-GL | 0.3 (0.1) | 2.4 (0.8) | 97.2 (1.0) | 1.289 (0.014) | 0.025 (0.001) | 1.040 (0.014) | |
| ACWL | 0.0 (0.0) | 0.0 (0.1) | 100.0 (0.1) | 1.314 (0.003) | 0.027 (0.000) | 1.055 (0.002) | |
| D-learning | 0.7 (0.2) | 1.2 (0.5) | 98.1 (0.5) | 1.309 (0.012) | 0.025 (0.001) | 1.058 (0.013) | |
| b = 2 | l1-PLS-HGL | 3.4 (0.1) | 4.2 (0.3) | 92.4 (0.4) | 1.242 (0.016) | 0.021 (0.001) | 0.809 (0.020) |
| l1-PLS-GL | 1.7 (0.4) | 7.2 (1.6) | 91.1 (2.0) | 1.217 (0.027) | 0.022 (0.002) | 0.774 (0.024) | |
| ACWL | 2.7 (0.8) | 3.4 (1.7) | 93.9 (1.3) | 1.266 (0.019) | 0.022 (0.001) | 0.814 (0.016) | |
| D-learning | 1.5 (0.4) | 5.8 (1.1) | 92.7 (1.0) | 1.260 (0.022) | 0.022 (0.001) | 0.816 (0.023) | |
| b = 3 | l1-PLS-HGL | 8.6 (0.3) | 11.9 (0.5) | 79.5 (0.6) | 1.091 (0.022) | 0.011 (0.001) | 0.689 (0.028) |
| l1-PLS-GL | 4.8 (1.1) | 15.0 (3.3) | 80.2 (4.4) | 1.069 (0.056) | 0.017 (0.003) | 0.569 (0.041) | |
| ACWL | 9.3 (1.4) | 8.2 (2.7) | 82.5 (2.2) | 1.100 (0.034) | 0.014 (0.001) | 0.627 (0.032) | |
| D-learning | 3.8 (0.8) | 17.5 (1.6) | 78.6 (1.3) | 1.067 (0.027) | 0.016 (0.001) | 0.607 (0.027) | |
| b = 4 | l1-PLS-HGL | 17.0 (0.4) | 23.3 (0.5) | 59.7 (0.5) | 0.745 (0.023) | 0.001 (0.001) | 0.623 (0.030) |
| l1-PLS-GL | 10.5 (2.3) | 27.9 (3.2) | 61.6 (5.2) | 0.782 (0.070) | 0.008 (0.004) | 0.468 (0.081) | |
| ACWL | 16.8 (1.9) | 16.2 (3.5) | 67.0 (3.1) | 0.793 (0.055) | 0.007 (0.002) | 0.475 (0.033) | |
| D-learning | 9.9 (1.6) | 30.0 (1.7) | 60.2 (0.7) | 0.783 (0.027) | 0.006 (0.001) | 0.543 (0.039) | |
| b = 5 | l1-PLS-HGL | 26.4 (0.4) | 33.4 (0.5) | 40.3 (0.3) | 0.410 (0.022) | −0.007 (0.001) | 0.648 (0.031) |
| l1-PLS-GL | 18.2 (3.2) | 48.4 (6.2) | 33.4 (3.4) | 0.324 (0.059) | −0.004 (0.002) | 0.485 (0.084) | |
| ACWL | 30.1 (2.5) | 22.6 (4.2) | 47.3 (3.2) | 0.422 (0.053) | −0.005 (0.002) | 0.541 (0.046) | |
| D-learning | 19.3 (1.6) | 37.4 (1.3) | 43.3 (0.7) | 0.505 (0.031) | −0.004 (0.001) | 0.622 (0.045) | |
Note: Means and standard deviations (in parenthesis) of assignment proportions in % and empirical ITR effects on donation, deferral, and utility across 100 repetitions of 5-fold cross-validation are reported. ITR effects measure the difference in the average outcome between donors whose assigned inter-donation intervals in the trial are optimal (with respect to the method used to estimate the ITR) and those whose assigned inter-donation intervals are non-optimal. A larger ITR effect on donation/utility and a smaller ITR effect on deferral are more desirable.