Hester van Eeren

General discussion | 6 147 | Secondly, non-randomized data need sophisticated modelling or statistical techniques to control for limitations like uncertainty in the parameter estimates and differences in observed baseline variables. Although our statistical methods were carefully chosen, they are not the only methods available to adjust for data limitations. For example, using information about the parameters available only in the dataset when modelling the cost-effectiveness and estimating the value of conducting further research using a Markov model is one way to fill in the model. An alternative would have been to first systematically search the literature and fill in the model parameters based on this information, then modelling the uncertainty, and then updating the model with the information from the dataset available. An advantage would have been that all available evidence would have been submitted in the model. A disadvantage would have been that the evidence used came from different datasets with different contexts. Then, the question would arise how the conclusions would relate to the specific situation in which the data were gathered. In addition, there are available alternatives to the PS method to adjust for observed baseline differences between intervention groups. This method only controls for measured baseline differences and it is not the only method that can control for such differences. Alternatives such as instrumental variables and multivariate matching methods are also plausible options (Borah et al, 2014; Kreif, Grieve, Radice, Sadique, Ramsahai, & Sekhon, 2012). In addition, among other methods to test the robustness of the findings, like using sensitivity analyses or repeating the analyses in different subsets, one could think of using two or three different statistical techniques to find out whether the results found were robust (Borah et al., 2014; Duncan, Engel, Claessens, & Dowsett, 2014). It could overcome misinterpretations and false conclusions from analyses, especially when only the p-value is used to draw inferences from the analyses (Greenland, Senn, Rothman, Carlin, Poole, Goodman et al., 2016). Results could be presented as a probability that the findings are likely given the selected dataset, and the robustness of the findings can be represented in the method, in the interpretations, and in the selected datasets (Nuzzo, 2014). Conclusion Although the methods used were not new in all aspects (e.g., in applying a cost- effectiveness framework or using statistical methods to control for allocation bias), the studies in this thesis showed the practical applicability of these methods and of the use of clinical practice data to answer relevant research questions. More importantly, this thesis showed that different research fields can learn from each other. Health economic evaluations are not yet widely applied in youth care. Therefore, youth care can learn from and adopt these modelling techniques. The other way around, health economics can learn from youth care practice and adopt strategies from that field when deciding on guidelines. Issues that are initially thought to be specific for a certain research area are probably not that specific and can be used in related fields as well. Thus, this thesis showed that cost-effectiveness analysesprovidevaluable information that can be used to allocate public budgets on available youth care interventions wisely. Furthermore, using clinical practice data in youth care that are routinely gathered is