In low- and middle-income countries, multisite costing studies are increasingly used to estimate healthcare program costs. These studies have employed a variety of estimators to summarize sample data and make inferences about overall program costs.
We conducted a systematic review and simulation study to describe these estimation methods and quantify their performance in terms of expected bias and variance.
We reviewed the published literature through January 2017 to identify multisite costing studies conducted in low- and middle-income countries and extracted data on analytic approaches. To assess estimator performance under realistic conditions, we conducted a simulation study based on 20 empirical cost data sets.
The most commonly used estimators were the volume-weighted mean and the simple mean, despite theoretical reasons to expect bias in the simple mean. When we tested various estimators in realistic study scenarios, the simple mean exhibited an upward bias ranging from 12% to 113% of the true cost across a range of study sample sizes and data sets. The volume-weighted mean exhibited minimal bias and substantially lower root mean squared error. Further gains were possible using estimators that incorporated auxiliary information on delivery volumes.
The choice of summary estimator in multisite costing studies can significantly influence study findings and, therefore, the economic analyses they inform. Use of the simple mean to summarize the results of multisite costing studies should be considered inappropriate. Our study demonstrates that several alternative better-performing methods are available.