Assessment of Strategies for Managing Expansion of Diagnosis Coding Using Risk-Adjustment Methods for Medicare Data

Abstract

This study of 100% Medicare Inpatient Files assesses strategies for managing coding of comorbidities associated with hospitalization before and after Centers for Medicare & Medicaid coding changes.

Since the passage of the Affordable Care Act (ACA) in 2010, many studies have used national Medicare data to examine associations between national hospital pay-for-performance programs and quality and costs of care.^1,2,3,4 In January 2011, as the ACA was being implemented, the Centers for Medicare & Medicaid Services increased the number of available diagnosis billing codes from a maximum of 9 diagnosis codes (the primary diagnosis plus 8 comorbidities; a tenth code was reserved for coding external causes of injury and usually left blank⁵) to 25 diagnosis codes (the primary diagnosis plus 24 comorbidities).

Given that many risk-adjustment models identify comorbidities using these diagnosis codes, this increase may in turn increase the measured severity of illness assigned to each patient. For example, Zuckerman and colleagues⁴ found that Medicare’s Hospital Readmission Reduction Program was associated with lower readmission rates for targeted conditions. However, more recent studies suggest that the increased number of codes available to hospitals for coding allowed by the change in Centers for Medicare & Medicaid Services policy made patient populations appear to be sicker, which may be responsible for a large portion of the observed reduction in readmissions previously attributed to the Hospital Readmission Reduction Program.^2,6 These findings have raised important issues about the analysis of the data when data before and after January 2011 are included, such as the evaluation of national policy changes that occurred around the ACA. Therefore, we sought to identify a suitable strategy to manage a large increase in the number of comorbidity codes allowed.

Methods

Using 100% Medicare Inpatient Files from 2008 through 2012, we examined coding of comorbidities over time, and how to manage the risk adjustment using the coded comorbidities after the Centers for Medicare & Medicaid Services change in policy went into effect in January 2011. First, we evaluated the distribution of the number of comorbidity codes associated with each hospitalization before and after the coding change (2008-2010 vs 2011-2012), limiting evaluation to hospitalized patients with 1 of 3 primary diagnoses that have been the initial focus of policy changes: acute myocardial infarction, congestive heart failure, and pneumonia. Then, we calculated the proportion of admissions that had 1 of 5 comorbidities included in most risk-adjustment models including the Elixhauser Comorbidity Index, the Hierarchical Condition Category, and the Charlson Comorbidity Index: congestive heart failure (excluding patients whose primary diagnosis is congestive heart failure), chronic obstructive pulmonary disease, diabetes, hypertension, and renal failure. We compared 3 methods for identifying comorbidities after January 2011 using (1) all 25 conditions (1 primary plus 24 comorbidities) available in Medicare claims, (2) the primary condition and first 8 comorbidities (from 24 coded comorbidities), and (3) the primary condition and randomly selecting 8 comorbidities from 24 coded comorbidities. We used SAS, version 9.4 (SAS Institute Inc) for the data preparation and Stata, version 14 (StataCorp) for the analyses. To account for multiple comparisons (5 coexisting conditions), a P < .01 was considered to be statistically significant. The study was approved by the institutional review board at Harvard T.H. Chan School of Public Health, which waived informed consent for deidentified data.

Results

Our sample (mean [SD] age, 76.7 [12.2]; 52.5% female) included 1 347 070 acute myocardial infarction, 3 035 748 congestive heart failure, and 2 497 803 pneumonia hospitalizations. We found that the median (interquartile range [IQR]) number of coded comorbidities increased from 8 (8-8) in 2008 to 2010 to 12 (9-17) in 2011 to 2012 (Wilcoxon-Mann-Whitney test, P < .001) (Figure 1). For all 5 conditions we investigated, the trajectories using the first 8 comorbidities after 2011 were consistent with the pre-2011 trends (Figure 2). On the contrary, using all 24 comorbidities led to an overestimation of the presence of coexisting conditions, and using random 8 comorbidities led to an underestimation. For example, assuming a linear trend, no evidence was found that the proportion of patients with congestive heart failure as a coexisting condition changed discontinuously in January 2011, when the first 8 comorbidities were used (−0.6 percentage point; 95% CI, −1.1 to −0.07; P = .03). However, a significant increase or decrease in the proportion was observed when all 24 comorbidities (2.4 percentage points; 95% CI, 1.8-2.9; P < .001) or random 8 comorbidities (−5.9 percentage points; 95% CI, −6.6 to −5.3; P < .001) were used.

Figure 1. Number of Coded Comorbidities Compared Between 2008 and 2010 and 2011 and 2012 in Medicare Patients Hospitalized With Acute Myocardial Infarction or Congestive Heart Failure or Pneumonia.

Light blue bars represent the number of coded comorbidities for patients admitted to a hospital owing to acute myocardial infarction, heart failure, or pneumonia from 2008 to 2010; dark blue bars for patients admitted from 2011 to 2012.

Figure 2. Proportion of Hospitalizations With Comorbid Conditions Among Medicare Patients Hospitalized With Acute Myocardial Infarction or Congestive Heart Failure (CHF) or Pneumonia.

COPD indicates chronic obstructive pulmonary disease.

Discussion

Our findings suggest that when researchers and policy analysts analyze Medicare data across multiple years—including data before and after January 2011—they should consider using the first 9 diagnosis codes; this setup allows for a less biased and more consistent risk-adjustment method. Using all 24 comorbidities or selecting 8 random comorbidities could make it appear as if patients got unrealistically sicker or healthier abruptly in January 2011 and potentially bias the estimated consequences of the policies introduced around the same time period. For future studies that include only the data after 2011, using all 24 comorbidities would be appropriate. Given that our study investigated the Centers for Medicare & Medicaid specific policy change made to the diagnosis billing codes in 2011, our findings may not be generalizable to similar changes in the future.