JA Gardner, CHQOER, Bedford MA, VA Hospital; A Rosen, CHQOER, Bedford MA, VA Hospital; S Loveland, CHQOER, Bedford MA, VA Hospital

Objectives: To determine whether two case-mix measures, ACGs and DCGs contribute essentially the same information to models explaining inpatient length of stay (LOS) as a function of case-mix, age, community resources, nature of the local medical care market, local practice patterns, and VA network.

Methods: A physician expert panel identified conditions for which there was substantial agreement about appropriate treatment practices (AMI, angioplasty, CABG, hip fracture, and the acute portion of knee replacement). All VA patients with FY1998 discharges for these conditions (over 27,000) were identified using ICD-9 CM codes (primary diagnosis) or ICD-9 CM procedures from the VA Patient Treatment and Extended Care files maintained at the Austin Automation Center (AAC). Patients who died were omitted from analysis of LOS. Diagnosis information from VA inpatient and outpatient data at AAC determined a patient’s ADG clinical grouping. The patient’s DCG relative risk score was obtained from files created by the VA Management Science Group. Linear regressions were estimated for LOS (with and without log transformations) with fixed effects estimated for VISNs. The contributions for ACG groupings and the DCG relative risk score were judged by examining R-square statistics for the equation and the coefficients and t-values for other variables in the model.

Results: For each condition, R-square statistics in models with DCG scores differed little (less than 0.02) from models with ADG groupings. R-square statistics ranged from .0961 for AMI to .2185 for CABG. For AMI, angioplasty and CABG, R-square was slightly higher in the model that included the single DCG continuous variable than in the model with 31 ADG grouping binary variables. Among all ten models (5 conditions, LOS and log LOS models for each), each with eight to ten other explanatory variables, only three coefficients changed more than 10% in their point estimates and none differed in statistical significance threshold tests (.10 level) when ADG groupings were replaced by DCG score. For angioplasty, the point estimates on the community resource variable and a measure of medical specialization in the local market changed. For hip fracture, the coefficient on the indicator of age >= 65 changed.

Conclusions: Both measures of case-mix provide information that is not captured in the other explanatory variables listed above. Because the R-square statistics only differ slightly and because coefficients on other variables usually do not materially differ between alternative models, it is likely that both ADG groupings and the DCG relative risk score bring much the same information to explaining LOS. For at least some medical conditions or procedures, the DCG case-mix measure appears to explain LOS more accurately than ADG groupings.

Impact: In models explaining LOS, although the choice of case-mix measure may sometimes affect estimates of contributions made by other explanatory factors in the general categories listed above, for the five conditions we studied, the contribution of these other factors and the percent of variation in LOS explained by the model usually are not substantially altered by choice of case-mix measure.