Prediction model
The backward stepwise selection process selected three predictors (p<0.10) for calcium intake: cheese; milk; yogurt and curd. The EPV was satisfactory at 29 (232 cases, 7 degrees of freedom). Table 3 presents the full logistic prediction model. The regression coefficient corresponds with the odds of having an adequate calcium intake for a person who’s consumption matches any other category, compared to the reference category. To calculate the predicted probability of having an adequate calcium intake, the beta coefficients are entered in the formula stated in the footnote of table 3.
The AUC was 0.858 (Figure 2), indicating excellent discrimination. The Hosmer and Lemeshow test (p=0.499) and Brier score (0.136) both indicated a good fit of the model. Figure 3 displays the calibration plot, with a null line representing predictions that perfectly match the observed incidence. The size of the dots represents the number of participants within the corresponding intake group. The model was very well calibrated, with most predictions bordering the null line. Predictions that did not match the observed incidence were generally found in midrange intake groups with small numbers of participants.