Modelling Individual Patient Hospital Expenditure for General Practice Budgets

The English NHS has introduced a system of budgets for general practices covering hospital expenditure for the patients on their lists. We model individual expenditure using diagnostic information from previous hospital spells, plus a large set of attributed variables measuring population, general practice, and local hospital characteristics. We show that, despite the large proportion of zero expenditures and the heavy right tail of expenditures, estimating models of untransformed expenditure via OLS yields better predictions at practice level than one or two part models using OLS with transformed expenditure or Generalised Linear Models. We describe a procedure for setting budgets for general practices which reduces the problem of the lags in the available data. We examine the distinction between need and non-need variables and the incentive implications of allowing past numbers of hospital encounters to determine practice budgets.