Fig 1.
A child’s family network growing over time.
Panels (a)-(c) show networks constructed using recorded relationships and panels (d)-(f) show networks constructed using whole-life-relationships. The networks are shown at: (a,d) the time of the child’s first notification; (b,e) the time of the child’s second notification; (c,f) the end of the data collection period. Individuals who are part of the child’s family network at each time point are shown in dark colours; individuals who exist but are not part of the network are shown in light colours; individuals who are not yet born are not shown. The values of the network predictor variables (see Table 1 for definitions) are provided in each case. This is an illustration only and does not correspond to a real child.
Table 1.
Summary of predictor variables used in the statistical classification models.
Event variables are the numbers of various types of event that occurred for the focal child prior to the notification. Recorded relationship variables are based on relationships recorded prior to the notification (see Fig 1A–1C); whole-life relationship variables are based on close family relationships, whether they were recorded before or after the notification (see Fig 1D–1F). A target event is defined to be a serious intervention by child protection services or a substantiated finding of maltreatment. Abbreviations are included for reference for Table 2 and Figs 1 and 2.
Table 2.
Models using whole-life relationships are always better than those without whole-life relationships.
Results for selected logistic regression models for estimated concern. Columns show the dataset used, the predictor variables included (see Table 1 for definitions), the number of predictor variables and the ROC score. For each dataset, results are shown for the best overall model (i.e. highest ROC score), the best four-variable model, and the best model that does not have any whole-life network predictor variables. Note that the model using events-based predictor variables only was not applied to the first notifications subsample as these predictor variables are all equal to zero (see Methods).
Fig 2.
Whole-life relationships are consistently better than recorded relationships for identifying high-risk groups.
Each panel shows the incidence rate of high estimated concern when children are split into two groups according to a threshold value (determined by a single-split classification tree) of a single network-based predictor variable at the time of their first notification. Above each graph is shown the predictor variable being used (see Table 1 for definitions), the threshold value used to group the children, and the proportion of children in the above-threshold (high-risk) group. The difference in the incidence rate between the high-risk and low-risk groups is always greater using whole-life relationships (red) than using recorded relationships (blue).