3115 — Length of Stay Calculations: Effect of Algorithm Choice
Berg E (COE Palo Alto), Jackson VW
(COE Palo Alto), Holmes T
(Stanford University), Laungani K
(COE Palo Alto), Frayne SM
(COE Palo Alto)
Methodologic issues arise when investigators use episodes of inpatient Veterans Health Administration (VHA) care to calculate Length of Stay (LOS), of relevance to utilization research. We examined how application of incrementally more sophisticated algorithms refine LOS calculations in a national cohort of VHA diabetes patients.
For all VHA patients with diabetes in fiscal year (FY) 2002 (from DEPIC-Diabetes Epidemiology Cohort, N=972,203), we assembled a record-level database of all instances of FY2003 inpatient care in PTF Bed Section/Census files (Acute Care, Observation, Extended Care) (Algorithm 1); starting cohort was 111,236 patients. We supplemented that file with FY2003 inpatient Non-VA Care and Fee Basis Files and with FY2003 records appearing in FY2004 or FY2005 VHA files (Algorithm 2). Finally, we excluded non-FY2003 care appearing in the combined database (Algorithm 3). We examined effect of these increasingly refined algorithms upon LOS calculations.
Using Algorithm 1, mean total LOS per patient was 55.0 days; 2.8% had LOS > 365 days. Using Algorithm 2, mean LOS was 72.3 days; 4.4% had LOS > 365 days. Totals > 365 came from duplicate records, overlapping records (within- or between-file), and care extending prior to the first day or beyond the last day of FY2003. To address these problems, we deleted duplicate records, consolidated overlapping records into one per contiguous episode of care, and truncated admission and discharge dates at the beginning and end of FY2003, respectively. After these additional steps, Algorithm 3 revealed the 117,914 diabetic patients with at least one inpatient stay in FY2003 had mean LOS 32.8 days; none had LOS > 365 days.
Algorithm choice for identifying episodes of inpatient care alters calculated LOS. Estimates differ depending on extensiveness of record query (e.g., whether fee basis care or records appearing in subsequent-year files are considered) and how record-level problems are addressed.
Failure to include all potential data sources can lead to under-representation of care received. Failure to account for record-level data problems can cause double-counting of some care, leading to over-estimates of LOS and potentially yielding estimated stays exceeding 365 days. VHA investigators should tailor algorithms for identifying episodes of inpatient care to the study question being examined.