Purpose: A recurrent challenge in health-related quality of life (HRQoL) research is how to handle comorbidity, as mean EQ-5D tariff-values (mtv) are usually only known for single health states. Several methods have been proposed to derive joint-state values (jsv) from single-state values (ssv), but comparison of the predicted jsvs with the elicited jsvs from groups of patients who actually have the target comorbidities reveals limited success. The aim of this research is investigating a new jsv predictor (jsp).
Method: We analysed data from the Medical Expenditure Survey Panel (MEPS, N≈3·104), revealing a Pearson’s R of .9995368 when correlating the number of comorbid disorders (Clinical Classification-Codes) with the mtvs for respondents with the same number of comorbid disorders. We proceeded to construct a novel jsp P based on the idea that the mtv u for a single state population is best viewed as a measure of that single state’s number m of units of morbidity (um) by a conversion f(u)=m obtained by an initial regression as suggested by the strong correlation above. The predictor aggregates the converted mtvs u and v of two single-states, before predicting the corresponding jsv’s value: P(u,v,z) = f –1(f(u)+f(v)-z) Above, z is a parameter intended to account for some overlap between conditions, which for the current research have been set to 1. The resulting predictor is on the form u+v-c where c is a constant depending on f. Though reminiscent of the additive predictor, conceptually it differs in treating the ssv’s as proxies for ums rather than as utilities: additivity arise from the linear relation empirically observed and not on any a priori assumption.
Result: The jsp outperformed various traditional predictors (additive, multiplicative, minimum and a general predictor proposed in  w.r.t. the MEPS data set. We next tested the concept on data (N≈104) elicited from a Norwegian inpatient population. Again P outperformed the other.
Conclusion: It is striking that the predictor, conceived by analysing a general population, transfers so readily to the inpatients. Although its construction relies on a linear regression for each data set, it does not rely on distinguishing between morbidities. We recommend adopting P as a canonical candidate predictor, as well as further research into how the parameter z can improve accuracy. _Bo_Hu_and_Alex_Z._Fu:_Predicting_Utility_for_Joint_Health_States:_A_General_Framework_and_a_New_Nonparametric_Estimator,_in_MDM_SEP-OCT_2010