Effect of data validation audit on hospital mortality ranking and pay for performance

Quality improvement and epidemiology studies often rely on database codes to measure performance or impact of adjusted risk factors, but how validity issues can bias those estimates is seldom quantified.

To evaluate whether and how much interhospital administrative coding variations influence a typical performance measure (adjusted mortality) and potential incentives based on it.

National cross-sectional study comparing hospital mortality ranking and simulated pay-for-performance incentives before/after recoding discharge abstracts using medical records.

Twenty-four public and private hospitals located in France PARTICIPANTS: All inpatient stays from the 78 deadliest diagnosis-related groups over 1 year.

Elixhauser and Charlson comorbidities were derived, and mortality ratios were computed for each hospital. Thirty random stays per hospital were then recoded by two central reviewers and used in a Bayesian hierarchical model to estimate hospital-specific and comorbidity-specific predictive values. Simulations then estimated shifts in adjusted mortality and proportion of incentives that would be unfairly distributed by a typical pay-for-performance programme in this situation.

Positive and negative predictive values of routine coding of comorbidities in hospital databases, variations in hospitals' mortality league table and proportion of unfair incentives.

A total of 70 402 hospital discharge abstracts were analysed, of which 715 were recoded from full medical records. Hospital comorbidity-level positive predictive values ranged from 64.4% to 96.4% and negative ones from 88.0% to 99.9%. Using Elixhauser comorbidities for adjustment, 70.3% of hospitals changed position in the mortality league table after correction, which added up to a mean 6.5% (SD 3.6) of a total pay-for-performance budget being allocated to the wrong hospitals. Using Charlson, 61.5% of hospitals changed position, with 7.3% (SD 4.0) budget misallocation.

Variations in administrative data coding can bias mortality comparisons and budget allocation across hospitals. Such heterogeneity in data validity may be corrected using a centralised coding strategy from a random sample of observations.