2012 HSR&D/QUERI National Conference Abstract
3045 — Assessing Bias in Current Methods for Estimating Disease-Attributable Costs across 31 Common Conditions in VA Users
Zeliadt SB, Yee L, Batten AJ, and Chapko MK, VA Puget Sound; Wagner TH, HERC; Zhao XH, VA Puget Sound;
Quantifying how much individual conditions add to healthcare costs is of considerable interest to decision makers. We compared two commonly used methods – matching and regression – to estimate the disease-attributable (DA) costs of 31 conditions in the VA.
A 20% random sample of VA users from FY2008 was identified. Encounter-level costs for FY2008 were extracted using inpatient TRT files and outpatient OPAT files from Decision Support Systems, and were linked to files recording all encounter-level ICD9 diagnosis and procedure codes. The matching approach compared costs for patients with each condition of interest and a demographically similar group of patients without the condition. The regression approach included indicators for each condition and potential confounding factors.
Among 996,869 VA users in FY2008, 83.9% of VA users had at least one of our 31 priority conditions and 67.9% had two or more conditions. The four most prevalent conditions were hypertension (55%), diabetes (26%), dyspepsia (22%) and ischemic heart disease (19%). Among this sample of VA users with at least one chronic condition, the actual FY2008 expenditure was $6.7 billion with a mean cost of $7,039. All approaches yielded unrealistically high estimates. For example, matching estimated the DA costs of hypertension to be $3,507 per person, which would mean hypertension accounts for over $1.9 billion (28%) of the actual FY2008 budget. Regression approaches yielded inconsistent results with negative costs for some conditions. The combined cost estimates for only the 31 conditions significantly exceeded the total FY2008 expenditure; matching by 252%, demographic-adjusted regression by 231%, and comorbidity-adjusted regression by 110%. This is notable given the DA cost estimates for the 31 conditions should not include costs for care beyond this limited group of conditions.
Commonly used and widely cited methods for estimating disease-attributable costs are significantly biased when applied to VA users who have multiple conditions. Adjusting for comorbidity in regression models helps avoid double-counting some components of care for complex patients, although these models appear to also result in over-estimates. Alternative methods, including attributable-fraction, should be explored.
Decision makers should use caution when interpreting cost studies for individual conditions among VA users.