Comparing Least‐squares and Quantile Regression Approaches to Analyzing Median Hospital Charges
ACADEMIC EMERGENCY MEDICINE 2012; 19:866–875 © 2012 by the Society for Academic Emergency Medicine Abstract
department (ED) and hospital charges obtained from administrative data sets are useful descriptors of injury severity and the burden to EDs and the health care system. However, charges are typically positively skewed due to costly procedures, long hospital stays, and complicated or prolonged
treatment for few patients. The median is not affected by extreme observations and is useful in describing and comparing distributions of hospital charges. A least‐squares analysis employing a log transformation is one approach for estimating median hospital charges, corresponding confidence
intervals (CIs), and differences between groups; however, this method requires certain distributional properties. An alternate method is quantile regression, which allows estimation and inference related to the median without making distributional assumptions.
objective was to compare the log‐transformation least‐squares method to the quantile regression approach for estimating median hospital charges, differences in median charges between groups, and associated CIs.
Methods: The authors performed simulations using
repeated sampling of observed statewide ED and hospital charges and charges randomly generated from a hypothetical lognormal distribution. The median and 95% CI and the multiplicative difference between the median charges of two groups were estimated using both least‐squares and quantile
regression methods. Performance of the two methods was evaluated.
Results: In contrast to least squares, quantile regression produced estimates that were unbiased and had smaller mean square errors in simulations of observed ED and hospital charges. Both methods performed well
in simulations of hypothetical charges that met least‐squares method assumptions. When the data did not follow the assumed distribution, least‐squares estimates were often biased, and the associated CIs had lower than expected coverage as sample size increased.
Quantile regression analyses of hospital charges provide unbiased estimates even when lognormal and equal variance assumptions are violated. These methods may be particularly useful in describing and analyzing hospital charges from administrative data sets.
Open access content
Free trial content