2005 Volume 17 Pages 43-58
Health insurance claims commonly contain multiple diagnoses. However, doctors had not been required to specify the primary diagnoses until April 2002. Before then, coders of insurers arbitrarily chose primary diagnoses with no explicit criteria, subject to criticism that such classification may be different from classification by doctors. As part of the fee schedule revision in April 2002, doctors were required to specify a primary diagnosis in claims with more than one diagnosis. The author took advantage of this opportunity and analyzed if the change of classifiers affected the classification on a disease-specific manner using national survey results of claims. Since the number of days per claim and per diem cost were found to follow a log-normal distribution, we adopted the multiplicative model, instead of the regular additive model, in which geometric mean (GM) and geometric standard deviation (GSM) were used instead of arithmetic means and standard deviation, a method commonly used for working environment measurement GSD was calculated using three indices, number of claims classified into a category, number of days per claim and cost per claim for each of the 119 diagnostic categories. Intra-classifier (yearly variance among insurers' coders) GSDs were calculated between the geometric mean (GM) over seven years (1995-2001) and each year. Inter-classifier (variance between coders and doctors) GSDs were calculated between the GM over seven years (1995-2001) and GM over two years (2002-3) for each of 119 diagnostic categories. Inter-classifier GSDs were larger than intra-classifier GSDs in the power of 2.01, signifying that inter-classifier variances were larger than yearly fluctuation among insurers' coders. These findings led one to conclude the change of classifiers did affect the classification. "Pancreatic disease" and "mood disorders" showed the largest discrepancy between doctors and insurers' coders while renal failure and dental diseases showed the least discrepancy, indicating that both doctors and insurers' coders were in agreement in terms of choosing renal failure and dental diseases as primary diagnoses but in disagreement in choosing pancreatic disease or mood disorders. Lack of proper coding skills as well as proliferation of rule-out or diagnoses-for-convenience might have contributed to such disagreement. The proposed method will have a wide applicability including detection of potential misclassification of DPCs.
