In their milestone 1 paper, the team "market makers" say the following on page 13:

For each Primary Care Physician (PCP), Vendor and Provider, a value was calculated that was the 
probability that a patient associated with the entity would visit hospital.  Each patient was then 
allocated the highest probability of all the PCPs (Vendors or Providers) that they were associated 
with, generating 3 fields in total.

I'm trying to replicate their methodology.  Is this probability only calculated using the claims data, or is the actual days in hospital for the next year merged in as well?  Here's my first shot at this.  I've already read in the claims data, converted LengthOfStay to numeric, and replaced missing values with zero:

library(plyr)
Claims$visit <- as.numeric(Claims$LengthOfStay>0)
providerProbs <- ddply(Claims,c('ProviderID','Year'),function(x) c('prob'=mean(x$visit)), .progress='text')

The result is the percent of visits to a given provider that resulted in hospitalization:

> head(providerProbs[providerProbs$prob>0,],10)
    ProviderID Year      prob
29       12890   Y1 1.0000000
56       23379   Y1 0.5294118
57       23379   Y2 0.7222222
58       23379   Y3 0.4166667
97       40154   Y1 0.4285714
98       40154   Y2 0.5714286
99       40154   Y3 0.5000000
466     173881   Y1 1.0000000
467     173881   Y2 0.9375000
468     173881   Y3 0.5600000

Am I on the right track?  Or should I be using the "DaysInHospital" table, rather than "LengthOfStay" in the claims table?

Hi Zach,

I've just been reading what I wrote and it is not 100% clear is it!

What is meant when I say 

 'probability that a patient associated with the entity would visit hospital'

should probably be

'probability patients associated with the entity would have at least one day in hospital in the following year'

So basically get all the patients who have visited a particular entity - get the DaysInHospital for those patients, rounding anything above 1 to 1. If you just then take an average of these DaysInHospital then that is the probability for that entity.

It is nothing to do with the length of stay field in the claims data. 

Great, thank you. What function did you use to compute probabilities (if it wasn't the mean)?