Risk Adjustment What data are there in administrative files for risk adjustment and how can we code them? Paul L. Hebert, Ph.D. Department of Health Services University of Washington School of Public Health Basic elements of a Quasi-Experimental CE study Basic elements of CER include Rigorous study design Appropriately chosen study sample Treatment Appropriately chosen treatment groups Dfind Defined outcomes Measures of patient demographics, co-morbidity and severity of illness Statistical analysis to deal with potential confounding Demographics β Observed Comorbid Conditions Potential unobserved Confounders Outcomes Measuring Covariates Demographics Socio-demographics Severity (good luck) 1
Demographics Focus on Medicare data because coding of race on other administrative data sources are idiosyncratic Example: HCUP. Some states do not require the collection of race and so race is coded infrequently or not at all for all hospitals in the state The CMS Beneficiary Summary File (BSF) contains the following demographic information Date of birth (from which age can be calculated) Gender Office of Management and Budget (OMB) and RTI defined race groups White Asian/Pacific Islander African American Native American Hispanic Other /Unknown Percentage Distribution of Medicare Enrollees by Race, 2008 BSF race code and RTI Hispanic Code 1.86% Asian 1.84% Other 2.49% Hispanic 0.43% N American Native 2.43% Asian/Pl 1.19% Other 7.80% Hispanic 0.42% American Indian/Alaska Native 10.12% Black 9.77% Black 83.11% White 77.64% Non-Hispanic White 5 Demographics Problems with the coding of race in the beneficiary summary file Hispanic is a mutually exclusive category. There is no such thing as Black-Hispanic, Non-Hispanic Black, or Hispanic Asian in Medicare Based mostly on an historic programming decision, not on policy Asians are a single category 2
Sociodemographics (SES) What measures of socioeconomic status (SES) are available in Medicare files? Two Sources of SES information Medicare buy-in indicator County and ZIP code identifiers to link with Census-based measures of SES Sociodemographics For low income beneficiaries, Medicare will pay for Part B benefits and the copayments and deductibles for Medicare allowed expenses. Note: this does not necessarily mean theh beneficary is receiving full Medicaid benefits (e.g., Nursing home care) Buy in Code in BSF % of beneficiarie s None 124,006 75.2%. Part A only * <0.01 Part B only 3,229 20% 2.0% Part A+B 37,678 22.8% Total 164,918 Sociodemographics (SES) What measures of socioeconomic status (SES) are available in CMS files? County and ZIP code identifiers to link with Census-based measures of SES http://factfinder.census.gov/home/saff/main.html?_lang=en 5-digit ZIP code can be merged to Census based ZIP Code Tabulation Areas (ZCTA) Important SES measures to consider: Median household income Percent of population with income below federal poverty limit Percent of population over 18 with < high school education 3
Sociodemographics (SES) What measures of socioeconomic status (SES) are available in CMS files? Problems with Census-based measures of SES ZIP Code Tabulation Areas (ZCTA) are not perfectly matched to ZIP codes (Krieger 2002) Only available for Census years Between Census years county level data are available from BLS, but counties can be huge and diverse Krieger N, et al. American Journal of Public Health, July 2002, Vol 92, No. 7 Measuring Covariates Demographics Socio-demographics Severity (good luck) Definition: A comorbidity is a disease or condition that coexists with the primary condition of interest and affects health outcomes Measuring comorbidity using CMS data Diagnosis based measures Cost-validated measures How do you incorporate comorbidities into a statistical model in CER? 4
Measuring Diagnosis based measures Look for diagnoses for conditions on Medicare claims Group conditions into a small number of conditions Optionally: Give a weight to each of these conditions reflecting its importance to the outcome Result is a series of indicators for each condition, and an overall comorbidity score or index Widely used comorbidity indexes Charlson comorbidity Elixhauser Measuring Diagnosis based measures Charlson comorbidity 18 conditions abstracted from hospital charts used to predict in one-year mortality Adapted for use with administrative data in several iterations. Romano is commonly used (Romano et la., 1993) Software is available here: http://healthservices.cancer.gov/seermedicare/program/charlson.comorbidity.macro.txt MARSHALL: WE SHOULD PUT HEIDI s ON the RESDAC SITE Elixhauser (Elixhauser et al, 1998) 30 conditions developed to predict length of stay, hospital charges, and in-hospital death Software is available here: http://www.hcupus.ahrq.gov/toolssoftware/comorbidity/comorbidity.jsp Romano PS, Roos LL, Jollis JG. Adapting a clinical comorbidity index for use with ICD 9 CM administrative data: differing perspectives. J Clin Epidemiol 1993;46:1075e9; discussion 1081e90. Elixhauser A, Steiner C, Harris DR, Coffey RM. measures for use with administrative data. Med Care 1998;36:8e27. Charlson/Romano conditions Some important conditions aren t included in Romano/Charlson SOURCE: Gagne JJ, Glynn RJ, Avorn J, Levin S, Schneeweiss S, A combined comorbidity score predicted mortality in elderly patients better than existing scores. J Clin Epi, IN PRESS 5
Charlson/Romano conditions SOURCE: Gagne JJ, Glynn RJ, Avorn J, Levin S, Schneeweiss S, A combined comorbidity score predicted mortality in elderly patients better than existing scores. J Clin Epi, 2010 Some important conditions aren t included in Romano/Charlson SOURCE: Gagne JJ, Glynn RJ, Avorn J, Levin S, Schneeweiss S, A combined comorbidity score predicted mortality in elderly patients better than existing scores. J Clin Epi, IN PRESS or in Elixhauser. SOURCE: Gagne JJ, Glynn RJ, Avorn J, Levin S, Schneeweiss S, A combined comorbidity score predicted mortality in elderly patients better than existing scores. J Clin Epi, IN PRESS 6
Measuring Diagnosis based measures Which should you use? Luckily you no longer have to choose because Gange et al (2010), have combined them and the combined measure works better than either individually And, software is available here: http://www.drugepi.org/downloads/ Gagne JJ, Glynn RJ, Avorn J, Levin S, Schneeweiss S, A combined comorbidity score predicted mortality in elderly patients better than existing scores. J Clin Epi, 2010 SOURCE: Gagne JJ, Glynn RJ, Avorn J, Levin S, Schneeweiss S, A combined comorbidity score predicted mortality in elderly patients better than existing scores. J Clin Epi, IN PRESS Distribution of comorbid conditions Percent 10 15 0 5 0 5 10 15 20 Number of Gagne comorbid conditions Note: A good percentage of patients have zero or only 1 or two comorbid conditions recorded in a single year. (37% had <=2 comorbid conditions in our sample) 7
Measuring Get diagnosis codes from administrative data to create comorbidity scores Many administrative databases are developed from the same basic claim form On claims based on HCFA 1500 claims form (e.g., CMS Physician/Supplier) there are 4 diagnosis fields Measuring Diagnosis based measures Many institutional (e.g., hospital) administrative claims databases are developed from the UB-04 claim form CMS MedPAR (hospital), skilled nursing facility, home health, and outpatient hospital claims On MedPar claims, Admitting diagnosis: initial reason given for hospitalization. Don t generally use this to create comorbidity scores because it is superseded by the primary diagnosis Primary diagnosis: the condition most responsible for the hospitalization. Secondary diagnoses: Other important conditions. Use primary and secondary dx s to create comorbidity score Present on admission indicators exist (and are almost always yes) Measuring Diagnosis based measures based on CMS data If you are using CMS data, use indicators from Chronic Conditions Warehouse (CCW) http://ccwdata.org/ Included in the Beneficiary Annual Summary File (request this separately from the Beneficiary Summary File) Indicators for the presence of 21 chronic conditions Yearly indicators going back to 1999. Standard definitions: specific criteria for reference time periods, diagnosis and procedure codes, number/type of qualifying claims (e.g., must have 2 Carrier claims during reference time period), and coverage Ever date: Date when the standard definition was first met. 8
Measuring comorbidity Example of algorithm from Chronic Conditions Warehouse Definition: A comorbidity is a disease or condition that coexists with the primary condition of interest and affects health outcomes Measuring comorbidity using CMS data Outcomes-validated based measures Cost-validated measures How do you incorporate comorbidities into a statistical model in CER Cost-based measures (very briefly) Software is available to group patients according to the cost category Diagnostic Cost Groups/Hierarchical Condition Codes (Pope et al, 2004). 70 codes that represent groups of diagnosis codes that are similar in etiology and cost implications Ambulatory Care Groups (Weiner et al, 1991) are 51 categories developed from ambulatory claims Validated for use in cost studies and therefore maybe useful for CER, but rarely used Pope GC, Kautter J, Ellis RP, Ash AS, Ayanian JZ, Lezzoni LI, Ingber MJ, Levy JM, Robst J: Risk adjustment of Medicare capitation payments using the CMS HCC model. Health Care Financ Rev 25:119 141, 2004 Weiner JP, Starfield BH, Steinwachs DM, Mumford LM: Development and application of a population oriented measure of ambulatory care case mix. Med Care 29:452 472, 1991 9
Cost-based measures (very briefly) Software is available to group patients according to the cost category Diagnostic Cost Groups/Hierarchical Condition Codes (Pope et al, 2004). 70 codes that represent groups of diagnosis codes that are similar in etiology and cost implications Ambulatory Care Groups (Weiner et al, 1991) are 51 categories developed from ambulatory claims RxRisk (aka Chronic Disease Score) (Fishman et al, 2003): Uses prescription drug claims to groups medications into hierarchical groups similar. Each measure has been validated for use in cost studies and are sometimes used for CER, but usually they perform worse than other measures (e.g., Schneeweiss et al., 2001) Pope GC, Kautter J, Ellis RP, Ash AS, Ayanian JZ, Lezzoni LI, Ingber MJ, Levy JM, Robst J: Risk adjustment of Medicare capitation payments using the CMS HCC model. Health Care Financ Rev 25:119 141, 2004 Weiner JP, Starfield BH, Steinwachs DM, Mumford LM: Development and application of a population oriented measure of ambulatory care case mix. Med Care 29:452 472, 1991 Fishman PA, Goodman MJ, et al., Risk adjustment using automated ambulatory pharmacy data: the RxRisk model, Medical Care 2003 41(1): 84 99 Definition: A comorbidity is a disease or condition that coexists with the primary condition of interest and affects health outcomes Measuring comorbidity using CMS data Outcomes-validated measures Cost-validated measures How do you incorporate comorbidities into a statistical model in CER Incorporating comorbidities measures into a statistical model in CER measures are dummy (0/1) variables that indicate whether they have the condition Charlson/Romano and Gagne can be combined into a summary score based on the sum of the weights provided by the authors. Two options: 1. Use the summary score only in the statistical model, i.e.: logit( Yi ) 0 1Demog 2SES 3Treatment 4Charlson 2. Use each of the dummy variables individually, i.e., logit( Y ) Demog SES Treatment AMI CHF i PVD CVD... HIV 6 0 7 1 18 2 3 4 5 10
Several weights have negative values (e.g., obesity is associated with better survival) SOURCE: Gagne JJ, Glynn RJ, Avorn J, Levin S, Schneeweiss S, A combined comorbidity score predicted mortality in elderly patients better than existing scores. J Clin Epi, IN PRESS Incorporating comorbidities measures into a statistical model in CER Which should you do? If your sample size is large enough, use each individual measure as an explanatory variables (Schneeweiss et al, 2003) If you are doing a logistic regression, you may have to drop or combine some categories of rare diseases (mild liver disease is often a culprit). If every person who has the condition has the same outcome, the model will spit those observations out. If there are other conditions important to your outcome but not coded or coded imprecisely with existing measure then include them as a separate variables as well. Schneeweiss S, Wang PS, Avorn J, Glynn J, Improved Adjustment for Predicting Mortality in Medicare Populations, Health Services Research 2003, 38(4) p. 1103 1120 Measuring Covariates Demographics Socio-demographics Severity (good luck) 11
Severity Measurement Good luck Not as straightforward because severity not directly coded Can be a major source of unobserved confounding in claims based analysis Two people may have generated claims for heart failure, but one is NYHA Class I and spends his winters skiing, while the other is Class IV and is short of breath at rest. CMS claims cannot distinguish between them Can approximate via complications Examples: Diabetes complications severity index (Young 2008 AJMC) Complicated versus uncomplicated pneumonia (DRG 89 vs DRG 90) Summary: Measuring Covariates Demographics in administrative data : Institutional and state-specific differences in race information For CMS data race is getting better but still limited. Socio-demographics: For CMS data, Medicaid buy-in is a good proxy for low income. Otherwise you must use Census-based proxies Several good diagnosis- and cost-based options Use the Chronic Conditions Warehouse if you are using CMS data Severity (good luck) References & Resources 1. Elixhauser, A., Steiner, C., Harris, D. R., & Coffey, R. M. (1998). measures for use with administrative data. Med Care, 36(1), 8-27. 2. Fishman, P. A., Goodman, M. J., Hornbrook, M. C., Meenan, R. T., Bachman, D. J., & O'Keeffe Rosetti, M. C. (2003). Risk adjustment using automated ambulatory pharmacy data: the RxRisk model. [Comparative Study Research Support, Non-U.S. Gov't Research Support, U.S. Gov't, P.H.S.]. Med Care, 41(1), 84-99. doi: 10.1097/01.MLR.0000039830.19812.29 3. Gagne, J. J., Glynn, R. J., Avorn, J., Levin, R., & Schneeweiss, S. (2011). A combined comorbidity score predicted mortality in elderly patients better than existing scores. [Research Support, N.I.H., Extramural Validation Studies]. J Clin Epidemiol, 64(7), 749-759. doi: 10.1016/j.jclinepi.2010.10.004 4. Krieger, N. (2002). Is breast cancer a disease of affluence, poverty, or both? The case of African American women. [Comparative Study]. Am J Public Health, 92(4), 611-613. 5. Pope, G. C., Kautter, J., Ellis, R. P., Ash, A. S., Ayanian, J. Z., Lezzoni, L. I.,... Robst, J. (2004). Risk adjustment of Medicare capitation payments using the CMS-HCC model. [Research Support, U.S. Gov't, Non- P.H.S.]. Health Care Financ Rev, 25(4), 119-141. 6. Romano, P. S., Roos, L. L., & Jollis, J. G. (1993). Adapting a clinical comorbidity index for use with ICD-9-CM administrative data: differing perspectives. [Research Support, Non-U.S. Gov't Research Support, U.S. Gov't, P.H.S.]. J Clin Epidemiol, 46(10), 1075-1079; discussion 1081-1090. 7. Schneeweiss, S., Wang, P. S., Avorn, J., & Glynn, R. J. (2003). Improved comorbidity adjustment for predicting mortality in Medicare populations. [Research Support, U.S. Gov't, P.H.S.]. Health Serv Res, 38(4), 1103-1120. 12
References & Resources 8. Weiner, J. P., Starfield, B. H., Steinwachs, D. M., & Mumford, L. M. (1991). Development and application of a population-oriented measure of ambulatory care case-mix. [Research Support, U.S. Gov't, P.H.S.]. Med Care, 29(5), 452-472. 9. Young, B. A., Lin, E., Von Korff, M., Simon, G., Ciechanowski, P., Ludman, E. J.,... Katon, W. J. (2008). Diabetes complications severity index and risk of mortality, hospitalization, and healthcare utilization. [Research Support, N.I.H., Extramural Research Support, Non-U.S. Gov't]. Am J Manag Care, 14(1), 15-23. 13