Hester van Eeren

| Chapter 2 2 | 28 measures, like the number of police contacts or youth self-report of committed crimes. Since not all committed crime, irrespective of the seriousness of the crime, is reported to the police, the difference in definition could give different effectiveness and cost- effectiveness results. Preference weighted measures (like the QALY) would be preferred in this context. Such measures could add a weight to different types of criminal activities and be more comprehensive in terms of the benefits they include (which could even entail a mix of health and crime-related outcomes). Reducing delinquent behavior is an important outcome of systemic interventions, but multiple other outcomes may be relevant as well, among which for example the ability to live at home after treatment, school attendance or family functioning (Henggeler, 1999; Sindelar et al., 2004). As these multiple outcomes are not considered in the current model, it could be valuable to extend the model or broaden the outcome measure. Before further use, the model would require improvement, since our analysis had a number of limitations. First, the model was limited to three states. Although a model is always a simplification of reality, and the current model even was an illustration, it should be investigated whether three states are sufficient to provide reasonable estimations of reality. Secondly, the states used now were dichotomous (criminal or non-criminal behavior). The severity of criminal offenses is likely to be important as well, also as a predictor of future criminal activity (Farrington, 2003). The frequency or the types of crime could be an important differentiating factor to discriminate more detailed states (Farrington, 2003). Using more differentiated states would therefore add validity to the model. However, a necessary condition for the formulation of amore complexmodel is the availability of more and detailed trial data. Third, an individual’s history of offenses could be used to predict future behavior and, thus, it may be useful to relax the ‘memoryless’ feature of the Markov model (Briggs et al., 2006). This feature encompasses that once a subject has moved from one state to another, the Markov model will have ‘no memory’ regarding which state the subject has come from or the timing of that transition. Using the history of earlier offences in the model could also improve the resulting estimates. The incorporation of long-term effects in the model was based on the coarse assumption individuals reach a stable state of criminal behavior after an age of 30 (Moffitt, 1993). However, the impact of using this theory in the current model was minor. In future research one could consider incorporating other relevant theories like the one used here (Moffitt, 1993) to improve long-term effect modelling. Various other theories and studies about the development of offending and antisocial behavior exist (Farrington, 2003), that could be used to incorporate long-term effects into the model. For example, Sampson and Laub (1993) suggest that offending depends on the strength of bonding to society, like bonding to family, peers, school and social institutions (Moffitt, 1993). In addition, an early age of onset predicts a relatively long criminal career (Farrington, 2003; Loeber & Farrington, 2000) and several risk factors for the early onset of offending are acknowledged (Farrington, 2003). Besides using studies like those mentioned, a stabilizing effect could be modelled more smoothly over time or could be based on empirical, long-term follow-up data to add more detail to modelling long-term effects. Furthermore, Value of Information (VoI) analyses should explore the additional value of