The data received from pilot sites in the CAMHS PbR project will be analysed to:
- Determine which assessment factors drive resource use.
- Determine how resource use relates to outcomes and from that decide how outcome measures can feed into the CAMHS PbR process.
- Develop an algorithm to predict resource needs based on assessment information (it is proposed that clinicians will be able to override the cluster proposed by the algorithm, but the algorithm will indicate/suggest the most likely resource usage given the assessment information supplied).
There is on-going work on retrospective analyses, investigating the predictors of resource need. Negative binomial regression is being used to predict the numbers of sessions attended as a function of a range of variables:
- Presenting problems (using the CORC Snapshot list)
- CGAS (Children’s Global Assessment Scale)
- SDQ (Strengths and Difficulties Questionnaire), including the impact supplement
- A replicated finding is that presenting problem and CGAS provide independent predictors of sessions attended.
- Eating disorder and psychosis predict the most need.
- Comorbidity also predicts more need, e.g. children with mixed conduct and emotional problems attended more sessions than those with only a recorded emotional problem.
- Greater difficulties on SDQ predicts more resource need (this has now been replicated in two datasets). Impact, especially perceived family burden, predicts more sessions.
Where is this leading?
- Quantified estimates of resource need.
- Hypotheses to test in the prospective study.
- Testing various methodological approaches, including model-based clustering, quantile regression, and negative binomial regression (see above), and multiple comparisons for categorical predictors with many levels – the outcomes of which will feed into the prospective modelling.
The analytic team is also continuing to research existing literature and building links with other research groups, for instance in Sweden where there is data on diagnosis, CGAS, and sessions attended. The adult cluster prediction algorithm has also been intensively studied to see what lessons can be learned.