Table 5.
Development of a risk prediction model for each dataset using model selection1
| Variable Selected | Australian Model, Odds Ratio (95% CI)2 | Leeds Model, Odds Ratio (95% CI)2 |
|---|---|---|
| Traditional risk factors | ||
| Family history of melanoma | ||
| None | 1.00 | 1.00 |
| 1 or more relative | 1.61 (1.05–2.48) | 3.38 (1.33–8.59) |
| Hair color | ||
| Dark brown/black | 1.00 | 1.00 |
| Light brown | 1.01 (0.71–1.44) | 1.15 (0.79–1.68) |
| Fair or blonde | 1.82 (1.16–2.86) | 2.13 (1.32–3.42) |
| Red | 2.76 (1.36–5.60) | 1.86 (0.96–3.58) |
| Nevus density | ||
| None | 1.00 | 1.00 |
| Few | 1.19 (0.57–2.48) | 1.84 (1.09–3.11) |
| Some | 3.13 (1.50–6.52) | 3.93 (2.27–6.79) |
| Many | 5.36 (2.43–11.83) | 4.64 (2.36–9.15) |
| Nonmelanoma skin cancer | ||
| No | 1.00 | 1.00 |
| Yes | 2.28 (1.19–4.37) | 3.86 (0.77–19.40) |
| Blistering sunburn as a child | ||
| None | 1.00 | — |
| 1 or more episodes | 0.80 (0.58–1.11) | — |
| Sunbed use | ||
| None | 1.00 | — |
| 1–10 sessions | 0.96 (0.61–1.51) | — |
| >10 sessions | 1.79 (1.01–3.20) | — |
| Freckling as an adult | ||
| None/very few | — | 1.00 |
| Few/some/many | — | 0.73 (0.52–1.02) |
| Eye color | ||
| Brown or black | — | 1.00 |
| Green or hazel | — | 1.05 (0.68–1.63) |
| Blue or grey | — | 1.39 (0.89–2.16) |
| Sun exposure hours on weekends and vacations | ||
| Quartile 1 (lower exposure) | 1.00 | |
| Quartile 2 | — | 0.52 (0.34–0.81) |
| Quartile 3 | — | 0.61 (0.39–0.96) |
| Quartile 4 (higher exposure) | — | 0.44 (0.27–0.72) |
| Genomic variants | ||
| rs7412746 (ARNT) | 0.85 (0.67–1.06) | 0.81 (0.65–1.02) |
| rs62211989 (ASIP) | 1.90 (1.33–2.71) | 1.91 (1.28–2.84) |
| R151C (MC1R) | 2.59 (1.77–3.79) | 2.75 (1.83–4.13) |
| R160W (MC1R) | 1.47 (1.00–2.16) | 1.70 (1.15–2.51) |
| rs2487999 (OBFC1) | 1.40 (0.95–2.06) | 1.37 (0.93–2.01) |
| rs132985 (PLA2G6) | 1.19 (0.95–1.48) | 0.85 (0.68–1.06) |
| rs1393350 (TYR) | 1.32 (1.03–1.70) | 1.20 (0.95–1.52) |
| rs6949072 (AGR3) | 1.28 (0.94–1.74) | — |
| rs7274597 (ASIP) | 0.50 (0.31–0.81) | — |
| rs76699054 (CCND1) | 1.40 (0.93–2.12) | — |
| rs12527588 (CDKAL1) | 1.56 (0.95–2.55) | — |
| rs3731217 (CDKN2A) | 0.79 (0.57–1.09) | — |
| D84E (MC1R) | 2.18 (1.02–4.67) | — |
| I155T (MC1R) | 2.60 (1.09–6.18) | — |
| V60L (MC1R) | 1.74 (1.21–2.50) | — |
| V92M (MC1R) | 1.70 (1.17–2.49) | — |
| rs45430 (MX2) | 0.72 (0.57–0.90) | — |
| rs3219090 (PARP1) | 0.73 (0.58–0.93) | — |
| rs2736100 (TERT) | 0.74 (0.59–0.94) | — |
| rs34585474 (AGR3) | — | 1.27 (0.89–1.83) |
| rs7781130 (AGR3) | — | 1.59 (0.85–2.96) |
| rs1801516 (ATM) | — | 0.77 (0.55–1.07) |
| rs700635 (CASP8) | — | 1.27 (0.99–1.63) |
| rs7776158 (CDKAL1) | — | 1.36 (1.06–1.75) |
| rs16953002 (FTO) | — | 1.27 (0.93–1.73) |
| R163Q (MC1R) | — | 0.62 (0.36–1.09) |
| D294H (MC1R) | — | 4.11 (1.62–10.46) |
| rs6517661 (MX2) | — | 0.75 (0.53–1.07) |
| rs113908778 (RAD23B) | — | 0.55 (0.24–1.24) |
| rs4436178 (RAD23B) | — | 1.87 (0.82–4.26) |
Abbreviation: CI, confidence interval.
A risk prediction model was developed separately for each dataset using a backward selection process in which traditional and genomic risk factors with P < 0.20 were retained in the multivariable model in addition to forced variables age, sex, city of recruitment, and European ancestry. The same genetic variants and traditional risk factors were assessed for inclusion in both models.
Odds ratios derived from the respective dataset, adjusted for all other variables in the model. For genomic variants, the per-allele odds ratio is presented. Values left blank indicate that the factor was not included in the final model for that dataset (Australia/Leeds). The areas under the curve for the Australian model were 0.80 (95% CI = 0.77–0.83) from the development model, 0.77 (95% CI = 0.74–0.80) from internal validation (10-fold cross-validation), and 0.72 (95% CI = 0.69–0.75) from external validation using the Leeds dataset. The areas under the curve for the Leeds model were 0.77 (95% CI = 0.73–0.80) from the development model, 0.72 (95% CI = 0.69–0.75) from internal validation, and 0.77 (95% CI = 0.74–0.80) from external validation using the Australian dataset. Both models were well calibrated in the external datasets (Hosmer-Lemeshow P = 0.57 for both).