Item — HN Reader

Part of my job I build predictive models to identify individuals at high risk of 'readmission' (it impacts billing, hence why healthcare systems are interested).

Conditional on an individual's history (e.g. do you have diabetes, other comorbidities, etc.), I haven't found much evidence that leaving against Dr advisement matters all that much for predicting readmission.

2 comments

The number in the GP is conditional on the patient being readmitted, so it's inflated by adverse selection: it doesn't include all the zeros where the patient recovered, but it does include all of the cases where something went wrong and they needed more/different treatment.

Yep, looks like classic “survivorship bias”: https://en.m.wikipedia.org/wiki/Survivorship_bias

> I haven't found much evidence that leaving against Dr advisement matters all that much for predicting readmission

maybe cos they're dead

So although tongue-in-cheek, it is an interesting question when dealing with medical records data (in particular I deal with insurance claims data in the US).

So to be clear, I know the person is alive when discharged (there is a code for death, as well as to hospice for example, in which you don't want to look at readmission either).

So a scenario in which someone is discharged, has a follow up heart attack, goes to ER and dies I would observe. The case the person dies and does not go to the hospital though I would not observe (no follow up insurance claim).

The latter scenario certainly happens -- how often it happens and how much it would bias my estimate in this case I am not sure.