StatsMiniBlog: Recursive partitioning

20140205-091454.jpg If you want to know who does, and who does not, need a bone marrow biopsy to detect malignant infiltration if the patient has rhabdomyosarcoma, you might want to start by taking a very large cohort of patients who had RMS and had a load of tests, including marrows. Then construct a decision tree that settles on identifying the group without marrow disease.

One (very good) stats process to do this is called ‘recursive partitioning’. It does what it says – splits the group up (partitions) it and then splits those up (recursively) until you’ve got a ‘good enough’ answer. Where the split is placed is by calculating and recalculating a threshold value, seeing how well that discriminates, and then moving on. For simple dichotomous measures this ‘threshold finding’ is incredibly easy as there are only two categories…

Now, how you decide what ‘good enough’ is is a matter of clinical judgement (e.g. what % of BM +ve patients would you miss – 2%?), and that’s worth a few arguments.

– Archi

(Visited 114 times, 1 visits today)