Nonsampling measurement error in administrative data: implications for economic evaluations

Administrative databases are increasingly being used to measure resource use in economic evaluations. Traditionally, it is assumed that any measurement error within the resource data source is stochastic and uncorrelated with group assignment. If the error is correlated with characteristics of the service delivery system and/or correlated by group, the estimated differences in consumption between groups may be reflecting errors in measurement rather than treatment effects. This paper is concerned with the effect of nonsampling measurement error on the internal and external validity of cost estimates based on data drawn from administrative record systems. Two service delivery characteristics, ownership form and financial incentives, are likely to influence systemically an agency's data collection effort. Using data from three community mental health centres located in three different states, evidence of systematic differences in data quality was found; private agencies with reimbursement property rights had higher quality data than public agencies without property rights. Simple tests for detecting variation in service use and costs data and cost-effective solutions for managing these problems are proposed.