Page 43 - Contributed Paper Session (CPS) - Volume 7

CPS2027 Olayan A. et al.
models were used to create a prediction model for incidence of offending in
the development and validation dataset. We first analysed the univariate
associations between the independent variables and criminal convictions.
Backward elimination was then used to reach the final multivariate model, in
which factors with the largest p value were sequentially deleted until only
significant predictors remained (p <0.05). We then created a weighted scoring
system by rounding all regression coefficients up to the nearest integer (i.e.,
the smallest integer greater than the estimate). This method was based on the
β-coefficients [or log of the Hazard Ratios (HRs)] rather than HRs rather than
HRs and is considered to be a more robust estimate.[2, 12, 13] After we
identified the final gender specific models, we created integer weights for each
variable by multiplying the β-coefficients by 10. These integer-weights were
added to create the final scores for each individual. The discriminate power of
the variables was assessed using the standard statistical techniques such as
area under the receiving operating curve (AUC); while the diagnostic
characteristics of various cut-points were evaluated using sensitivity and
specificity in both datasets (i.e. development and derivation). The primary
objective of this analysis was to investigate the accuracy and discriminative
power of the models for the combination of the risk factors that we have
included in the model. In an additional analysis, subject-specific scores were
split into quintiles [1st to 5th]. Crude incidence rates (95% CIs) were calculated
across the quintiles of the risk scores calculated separately for the
development and validation datasets. HRs were also presented across the
increasing quintiles of the scores.

3. Result
A total of 7,743 individuals were included in the study. Table 1 below
summarises the characteristics of the study population. The development and
validation datasets randomly selected included 5162 (67%) and 2,581 (33%),
respectively. There was comparable regarding to group, age group at the first
offence, gender, indigenous status, married status, country of birth, psychosis
type, the status of contact with mental health service after the first offence,
SEFIA and the type of the first offence.

31 | I S I W S C 2 0 1 9