Exploring two cost-adjustment methods for selection bias in a small sample: using a fetal cardiology dataset.

Abstract

In economic evaluations of healthcare technologies, situations arise where data are not randomized and numbers are small. For this reason, obtaining reliable cost estimates of such interventions may be difficult. This study explores two approaches in obtaining cost estimates for pregnant women screened for a fetal cardiac anomaly. Two methods to reduce selection bias in health care: regression analyses and propensity scoring methods were applied to the total mean costs of pregnancy for women who received specialist cardiac advice by means of two referral modes: telemedicine and direct referral. The observed total mean costs of pregnancy were higher for the telemedicine group than the direct referral group (4,918 versus 4,311 GBP). The regression model found that referral mode was not a significant predictor of costs and the cost difference between the two groups was reduced from 607 to 94 GBP. After applying the various propensity score methods, the groups were balanced in terms of sizes and compositions; and again the cost differences between the two groups were smaller ranging from -62 (matching "by hand") to 333 GBP (kernel matching). Regression analyses and propensity scoring methods applied to the dataset may have increased the homogeneity and reduced the variance in the adjusted costs; that is, these methods have allowed the observed selection bias to be reduced. I believe that propensity scoring methods worked better for this dataset, because after matching the two groups were similar in terms of background characteristics and the adjusted cost differences were smaller.