Table 5.
Area under ROC Curve for Different Outcomes/Ages/Groups Using Either the Scores Derived from the Primary Risk Model or Using Scores from Specific Risk Models.
| Prediction Model Primary Outcome | Area under the ROC Curve for Different Outcomes/Ages/Groups When Using the Primary Risk Model Scores a | Area under the ROC Curve for Different Outcomes/Ages/Groups When Using Risk Model Specific Scores to Each Outcome/Age/Group b |
|---|---|---|
| Predict d3mfs >0 (cavitated caries experience) at age 4 using age 1 responses, all participants (primary outcome risk score) | 0.68 | |
| Predict d1mfs >0 (cavitated and noncavitated caries experience) at age 4 using age 1 responses, all participants | 0.63 | 0.63 |
| Predict d5mfs >0 (extensive cavitated caries experience only) at age 4 using age 1 responses, all subjects | 0.68 | 0.73 |
| Predict d3mfs >0 (cavitated caries experience) at age 2.5 using age 1 responses, all subjects | 0.70 | 0.75 |
| Predict d3mfs >0 (cavitated caries experience) at age 4 using age 2.5 questionnaire responses | 0.66 | 0.71 |
| Predict d3mfs >0 (cavitated lesions) at age 4 using age 1 responses, including only participants enrolled in Medicaid | 0.59 | 0.66 |
| Predict d3mfs >0 (cavitated caries experience) at age 4 using age 1 responses, including only participants not enrolled in Medicaid | 0.60 | 0.68 |
d1mfs, decayed, missing and filled surfaces; d = ICDAS ≥ 1; d3mfs, decayed, missing, and filled surfaces; d = ICDAS ≥ 3; d5mfs, decayed, missing and filled surfaces; d = ICDAS ≥ 5; ROC, receiver operating characteristic.
The “primary” risk score used baseline data (age 1 questionnaire responses) from all subjects to predict d3mfs >0 (cavitated caries experience) at age 4. The data presented are after internal validation (bootstrapping). That same scoring algorithm was then used to predict other caries outcomes/ages (e.g., applied to the questionnaire responses at the 2.5-y-old visit, etc.) and was evaluated for subgroups of subjects (e.g., Medicaid enrolled and non-Medicaid enrolled). The intent was to understand how implementation of a single scoring risk algorithm would perform versus requiring different scoring algorithms for different outcomes/ages/groups.
In addition, separate risk models with distinct scoring algorithms were also created for each of the above situations to use as comparisons.