Expanding the National Health Expenditure Accounts (NHEA) Technical Documentation

Transcription

1 Background The National Health Expenditure Accounts (NHEA) have been produced annually by the Office of the Actuary at the Centers for Medicare and Medicaid Services (CMS) and its precursor organizations, since They track medical expenditure by source of payment (private, public, out-of-pocket) and type of expenditure (hospital, physician, etc.). Information is summarized and data tables are available to the public via download from the Centers for Medicare and Medicaid Services (CMS) website at The NHEAs contribute substantially to our understanding of medical spending. However, because they focus only on spending, the NHEAs provide no information on the value of health care spending, as they do not track the desired output of investment in health care improved health. Further, the data are not necessarily at the right level of aggregation to measure value. To make these productivity calculations, one needs to understand spending at the same level as health outcomes, which are most naturally measured by disease. Thus, a central issue in expanding the NHEA and our purpose is adding the more disaggregated (or micro) data needed to estimate prevalence of disease and costs. Chapter 1. Data sources The data used to create the analytic files for our purposes originated from the Centers for Medicare & Medicaid Services (CMS) and the National Center for Health Statistics (NCHS). Medicare Current Beneficiary Survey (MCBS) We obtained information on medical care (hospital, physician, home health, etc.) related to disease groups or screening and/or preventative services from claim data from the Medicare Current Beneficiary Survey (MCBS). Data was obtained through a data use agreement with CMS. MCBS is a nationally representative survey of Medicare beneficiaries, conducted by CMS for the entire Medicare population including the aged, disabled, and institutionalized. Information on demographic and participant characteristics, income, health insurance and health care expenditures, and health status and functioning are provided in the MCBS survey datasets. Two data files are released annually from the MCBS, the Access to Care and the Cost and Use. For this exercise we used the Cost and Use data files which contain a sample of ever enrolled participants with survey data linked to participant claims. The MCBS provides linked claim data for beneficiaries participating in the survey. The Medicare claims data included information on seven different types of claims: from Part A, inpatient, outpatient hospital care, home health care, skilled nursing home services and hospice care claims, and from Part B, claims for the physician supplier/carrier and durable medical equipment. Claim files have detailed information including the date of the claim, Medicare payments, coinsurance amount, other payments, medical conditions as identified by principal diagnosis (ICD9-CM code) and secondary diagnoses codes, and procedure codes. There are limitations in the MCBS as medical claim data are thought to be imperfect because it cannot be used to identify individuals with disease when care for those conditions was not sought.

2 In addition, care for many chronic diseases are grouped with routine visit care and are not always identified separately in the medical claim and event data files. National Health and Nutrition Examination Survey (NHANES) The National Health and Nutrition Examination Survey (NHANES) is a publicly available national representative survey conducted by NCHS. The NHANES data contain demographic and participant characteristics as well as survey questions to identify individuals who have (ever) been told they have a select group of diseases, and have been screened or received preventative services for a select group of diseases. In addition, laboratory and medical examination data for items such as blood glucose levels and blood pressure measurements are collected by trained medical personnel and available in the NHANES public release datasets. Information collected by NHANES at the medical examination allows for the identification of undiagnosed disease, such as undiagnosed diabetes and high blood pressure. Mortality data are also available in the NHANES. NHANES public release datasets were downloaded from: In the NHANES we had what was considered the gold standard information for the presence of disease. We used the NHANES to validate existing information on disease groups in the claim data source and to impute for missing data where rates were below the expected level, and where data was missing or seemed illogical. Population For this exercise, we defined the population as participants 65 years or older as of January 1 of the year of study, and included both community dwelling people as well as those living in an institution. Derivation of data was performed separately for each population by year and data source. Datasets were then combined for analyses. Years of study included

3 Chapter 2. Types of Variables We identified common variables across surveys. Common variables varied from year to year as the content of surveys has changed over time. Response scales were recoded to reflect consistency across surveys and years where necessary. Demographics and Socio-economic status (SES) Self-reported demographics and SES variables reported in the MCBS and NHANES included age, sex, race, education, poverty status and income, employment status, health insurance status, service in the armed services, marital status, and living conditions including type of dwelling, number of rooms in dwelling and number of people living in the dwelling. Participant characteristics Self-rated health and comparative health status, height, weight, smoking status, and items related to hearing and use of a hearing aid were available in each survey. Self-reported activities of daily living (ADLs) and instrumental activities of daily living (iadls) were also included in the data files. Both surveys also included self-report information related to medical care received including if the respondent had a particular location where routine medical care was received, the last time blood pressure and cholesterol were checked, and if the participant received preventative and screening procedures such as a pneumonia vaccination, pap smear, mammogram, and blood test screening for prostate specific antigen. Information on history of hysterectomy was also collected. The number of stays and nights a participant spent in a hospital and in an institution was collected. In the MCBS, this information was also derived from the claim data files. Disease Groups Self-report of select diseases including occurrences of select cancers, heart attack, stroke, diabetes, high blood pressure, coronary heart disease, Parkinson s disease, hyperplasia of the prostate, osteoporosis, rheumatoid and non-rheumatoid type arthritis, broken hip, cataracts and paralysis were provided in the survey components of both of MCBS and NHANES. Diagnosis code variables obtained from the 7 MCBS claim data files included ICD-9-CM diagnosis codes, ICD procedure codes and HCPCS and CPT procedure codes. Diagnosis code variables were used to identify diseases for which care was sought. A description of disease definition and classification is included in Chapter 3 of this document. Medical expenditures Expenditure variables were obtained from the MCBS administrative data. Cost variables included total payment and breakdown by Medicare payment and payment made by other insurance sources, coinsurance/copayment amounts, and patient out-of-pocket payments. The Person Summary file summarizes utilization and expenditure data (1) in total by type of service and (2) in total by payer. To impute for the presence of disease, we matched on variables common to the MCBS and NHANES. See Appendix 2a for formats of common variables where applicable.

4 Chapter 3. Disease group classification We developed a disease group classification schema based primarily upon the Agency for Healthcare Research and Quality s (AHRQ), Healthcare Cost and Utilization Project (HCUP), Clinical Classification Software (CCS) for ICD-9-CM. The CCS collapsed the over 14,000 diagnosis codes and 3,900 procedure codes into a much smaller number of clinically meaningful categories. Information describing the CCS can be found here: Creation of the mutually exclusive, collectively exhaustive disease groups were based primarily on the ICD-9-CM categorization assignment by the CCS and required the clinical expertise of five physicians, and extensive data management and analytic investigation. We further collapsed the CCS categories into 101 disease groups and 4 screening and preventive services categories. While the disease groups were based primarily on the ICD-9-CM categorization assignment by the CCS there were a few diseases identified by ICD-9-CM code included in the larger CCS disease categories that the physician group determined (1) should be stand-alone disease groups because of clinical significance (urinary incontinence because of the costs attributed to the condition (ICD-9- CM=788.3(x) was removed from the CCS category 163 Genitourinary symptoms and ill-defined conditions )), and (2) should be classified in a disease group different than assigned by the CCS (most commonly due to the significant changes to the mental health categories assigned by the CCS with the 2009 data release to more accurately reflect the DSM disease classifications). While CCS codes are not provided on the Medicare claims files provided with the MCBS, we mapped to the CCS categories via the CCS mapping provided as a public use file by HCUP on the web. The mapping was completed using the full 5-digit ICD-9-CM diagnosis codes, full 4-digit ICD procedure codes and full HCPCS, CPT procedure codes. Availability of the full ICD-9-CM diagnosis codes, ICD procedure codes and HCPCS, CPT procedure codes allowed for greater flexibility when the classification in the CCS categories was not sufficient for our research purposes. For example, the mental health disease classifications for depression and bipolar disorder were best defined using the 4 th and 5 th digit of the ICD-9-CM so that major depression would be defined as: (MDD single episode), (MDD recurrent episode), 311 (MDD nos); and bipolar disorder would be defined as: (Bipolar I, single manic episode), 296.1, (Bipolar I, recurrent mania), (Bipolar I, now manic), (Bipolar I, now depressed), (Bipolar I, mixed episode), (Bipolar I, episode unspec), (Bipolar Disorder, unspecified), (Bipolar II or Manic-Depressive psychosis mixed). In order to best make the distinction between these diagnoses, the 4 th and 5 th digit of the ICD-9-CM are required. Procedure codes related to screening were selected based primarily upon covered codes listed in the Medicare Claims Processing Manual (Chapter 18 - Preventive and Screening Services; Rev. 1953, ). If any one of the diagnosis or procedure codes appeared on the physician supplier or outpatient claim file the participant was considered to have undergone screening for the disease of interest.

5 Codes designated for use as billing for diagnostic testing were not included in the screening definitions. While screening rates based upon screening codes from Medicare claims data are considered to underestimate use, including diagnostic codes would artificially inflate screening rates (Freeman, et al 2002). Therefore, to correct for the imperfection in the claim data the screening variables are imputed following the methodology used for other non-sr diseases. We followed the algorithm of any 1 claim to be defined as presence of disease. The algorithm was assigned to all diagnosis code in each of the 7 available claim data files. Appendix 3a contains definitions of disease group classifications. Appendix 3b contains definitions of screening variables. Reference Freeman, et al. (2002). Measuring breast, colorectal and prostate cancer screening with Medicare claims data. Medical Care 40(8),

6 Chapter 4. Within Imputation processes First, missing data on demographic, socio-economic and self-reported disease in MCBS and NHANES are imputed using a sequential regression multiple imputation procedure. Covariates listed in Table 4a and self-reported (SR) disease in Table 4b were missing for some subjects in both NHANES and MCBS data files. A sequential regression multiple imputation procedure (Kennickel (1992), Van Buuren and Oudshoom (2000) and Raghunathan, et al (2001)) as implemented in the software package IVEWARE (Raghunathan, et al (2002)) was used multiply impute the missing values. These are iterative procedures in which the missing values in each variable are imputed conditional on all other variables using appropriate regression models. Random draws from an approximate predictive distribution of the missing values under these models are then used as imputations. Surveys for community dwellers and institutionalized populations differ, hence the sets of common NHANES-MCBS variables were different for institutionalized and community populations. We separately imputed missing values of common covariates in NHANES, MCBS-community, and MCBS-institutionalized populations. The rates of missing values varied by survey ranging between [0%-81%] and were generally highest in the MCBS institutionalized population. Tables 4a and b show the percentage of missing values in each survey for the year Some variables with the highest rate of missingness in MCBS were imputed by combining the community and institutionalized MCBS populations to increase stability. Table 4a. Rates of missingness of covariates across the surveys (2009) Label NHANES MCBS MCBS Inst Age (continuous) [0.0%] [0.0%] [0.0%] Asthma/Emphysema/COPD (combined in MCBS) [0.0%] Last blood cholesterol check [2.0%] [3.0%] Last blood pressure check [0.0%] [0.7%] Health compared to 1 year ago [0.0%] [0.7%] [4.3%] Expenditure label [0.0%] [0.0%] Served in Armed Forces [0.0%] [0.2%] [7.0%] Difficulty lifting/carrying 10 pounds [0.7%] [2.0%] Difficulty reaching up over head [0.1%] Difficulty stooping/crouching/kneeling [0.0%] [0.7%] [1.6%] Difficulty walking 1/4 mi [0.5%] [0.8%] [1.7%] Education [0.3%] [0.6%] [12.3%] Ever smoked [0.0%] [21.8%] Flu shot last year [1.1%] [10.1%]

7 Label NHANES MCBS MCBS Inst Has job [0.0%] [0.3%] Have particular place for medical care [0.0%] [0.7%] General health status [0.0%] [0.8%] [1.6%] Quality of hearing [0.2%] [0.5%] Wear hearing aid [32.1%] [0.8%] [3.6%] Height (cm) [2.5%] [0.2%] [0.0%] Had Hysterectomy [6.8%] [8.8%] [17.5%] Number of Days Patient in Institution [0.0%] [0.0%] [0.0%] Inpatient nights [0.0%] [0.0%] [0.0%] Inpatient stays [0.0%] [0.0%] [0.0%] Gender: 1=male [0.0%] [0.0%] [0.0%] Mammogram/breast X-ray in the last yr [0.8%] [12.9%] Marital status [0.0%] [0.2%] [1.8%] Total number of people in household [0.0%] [0.2%] # of rooms in home [1.0%] [0.6%] Pap smear in the last year [0.9%] [13.5%] Pneumonia vaccination [0.7%] [13.1%] Poverty Level Category [8.3%] [0.3%] [80.8%] Any difficulty dressing [0.0%] [0.7%] [0.9%] Any difficulty eating [0.0%] [0.7%] [0.9%] Private Health Insurance [0.0%] [0.0%] [0.0%] Private Health Insurance [0.0%] [0.0%] [0.0%] PSA test in last year [2.6%] [5.7%] Race [0.0%] [0.3%] [0.1%] Current smoker [0.0%] [0.4%] [21.9%] Weight (kg) [1.4%] [0.7%] [2.9%] Table 4b. Prevalence(SE) [Rates of missingness] of SR disease groups available across the surveys (2009) Label NHANES MCBS MCBS Inst Cervical Cancer 0.60 (0.23) [0.1%] 0.46 (0.10) [0.0%] Prostate Cancer 4.08 (0.61) [0.3%] 4.93 (0.30) [0.0%] Hematological Cancer 0.92 (0.37) [0.4%] Diabetes (1.39) [0.1%] (0.60) [0.0%]

8 Label NHANES MCBS MCBS Inst Undiagnosed Diabetes 2.42 (0.58) [0.1%] Hyperlipidemia (2.37) [2.9%] Undiagnosed hyperlipidemia 5.35 (0.90) [2.9%] Parkinson's Disease 1.59 (0.14) [0.0%] (1.78) [19.4%] Paralysis 2.69 (0.20) [0.0%] Cataract (0.67) [0.2%] Hypertension (1.75) [0.1%] (0.71) [0.1%] Undiagnosed Hypertension 9.16 (1.35) [0.1%] MI (heart attack) 8.57 (0.73) [0.4%] (0.48) [0.0%] Coronary Atherosclerosis and other heart disease (0.45) [0.0%] Stroke 8.17 (1.03) [0.3%] (0.45) [0.0%] Colon Cancer 2.71 (0.48) [0.4%] 2.49 (0.17) [0.0%] COPD 9.85 (1.17) [0.3%] Asthma (1.46) [0.1%] Lung Cancer 0.19 (0.09) [0.4%] 1.29 (0.14) [0.0%] Skin Cancer (0.94) [0.4%] (0.66) [0.0%] Prostate Hyperplasia 9.48 (0.41) [0.8%] Rheumatoid Arthritis (0.48) [0.0%] Arthritis: non-rheumatoid (0.74) [0.0%] Breast Cancer 5.68 (0.74) [0.1%] 5.09 (0.26) [0.0%] Osteoporosis/brittle bones (1.27) [0.6%] (0.58) [0.0%] Broken or fractured a hip 3.29 (0.57) [0.1%] 3.68 (0.18) [0.0%] References Kennickell AB. (1991). Imputation of the 1989 Survey of Consumer Finances: Stochastic relaxation and multiple imputation. Proceedings of the Survey Research Methods Section of the of the American Statistical Association, 1990 Joint Statistical Meetings, Atlanta, GA Raghunathan T.E., Lepkowski J.M., VanHoewyk J., Solenberger P., (2001). A multivariate technique for multiply imputing missing values using a sequence of regression models. Survey Methodology, 27, van Buuren S, Oudshoorn CGM (2000). Multivariate Imputation by Chained Equations: MICE V1.0 User's manual. TNO Report PG/VGZ/ Leiden: TNO Preventie en Gezondheid

9 Chapter 5. HMO Weight adjustment Claim data may be incomplete or missing for Medicare beneficiaries enrolled in a Medicare Advantage plan or Health Maintenance Organization (HMO) because payment of encounters with the health care system are paid by the private insurer, not CMS. Across the years of study, , between 20-30% of MCBS participants were enrolled in an HMO each year. We performed weight adjustments to account for incomplete claim information for each of the community and institutionalized MCBS participants enrolled in an HMO each year. Weights were calculated to adjust for participants not considered as "pure" Medicare enrollees. Pure" was defined as follows: 1. no participation in an HMO for the year of study, and 2. enrollment in Medicare parts A & B for the full 12-month study period unless the participant died during the year. To calculate the adjusted weights, logistic regression modeling was conducted to model "pure" Medicare participation using select covariates including any combination of demographic, SES, health status, and ADL/IADls variables and interaction terms to achieve decent model fit. Covariates included in the models vary across years and may vary across calibrated sets. Appendix 5a contains covariates included in each model. Model fit was assessed by Hosmer-Lemeshow test. Using the predicted probability (p) of "pure", the pure MC weight was calculated as 1/p. To assess balance in the community dwelling population, propensity of pure was estimated using GLM; F-ratios were reviewed for significance. In the institutionalized population, regression models were performed for each covariate to assess the association with propensity for pure Medicare participation. Using the residuals from each model, we calculated effect size to assess balance. For both the community and institutionalized populations, we calculated the final weight as the product of the existing MCBS survey weight and the pure MC weight. To assess the accuracy of the weight adjustment, we compared the population totals from the 2000 Census, aged 65+ to those represented in the year 2000 analytic files for those 65+ using the weight variable provided in the MCBS data files for those with any Medicare participation and the final weight for those with pure participation. The population size differed from the Census by about 5.5% for the 65+ Medicare population and by about 6.5% for those 65+ with pure enrollment. The final analytic file used for the related analyses was restricted to participants considered to have "pure" Medicare enrollment. To account for participants with incomplete claim data due to participation in an HMO and the population restriction we followed, the final weight was used for all analyses.

10 Chapter 6. Reconciliation of expenditures to NHEA: Survey spending adjustment Accurately and comprehensively tracking health care spending by Americans is a primary purpose of U.S. government agencies, economists, and health service researchers. The National Health Expenditure Accounts (NHEA) provides the most comprehensive estimate of heath care spending on personal health care, public health activities, government administration, and public health investment in research and construction. While NHEA provides invaluable aggregate data on health spending trends, it lacks the individual level data for detailed policy analysis, and also understanding the trends or services that are driving the health care spending. A significant effort, in health policy, is to distinguish between efficient and inefficient expenditures. How much of what we spend on cardio vascular disease is appropriate and how much is not? If we spent more on screening for disease at earlier stages, what would be the impact on cost per year of quality-adjusted life? To better address these questions, we need more information on the detailed components of spending than is available in the aggregated National Health Expenditure Accounts alone. Researchers often use personal or household level surveys of medical utilization and expenditure for policy analyses. The most commonly used is the Medical Expenditure Panel Survey (MEPS). The MEPS offer detailed self-reported information on medical spending for each medical service used by survey respondents. The previous reconciliation studies by the Agency for Healthcare Research and Quality (AHRQ), the Centers for Medicare and Medicaid Services (CMS), and other researchers used the MEPS and its precursor surveys (Meara et al., 2004; Selden et al., 2001; Sing et al., 2006). The standard approach has been to align NHEA and the MEPS in terms of their covered population, covered services, and grouping of services, and then compare total medical spending between the sources. Sing et al. (2006) found that when MEPS and the NHEA adjusted on a consistent basis, their expenditure estimates differ by about 14%. We make an adjustment for the variance between MCBS survey spending and NHEA national spending estimates for each year between 1999 and Total health spending reported in national health surveys is lower than the totals reported in the National Health Expenditure Accounts. 1 To account for this, we make three types of adjustments. First, we remove expenditures from the NHEA for goods and services which are out of scope of the surveys: other non-durable medical equipment (2.8%), other personal healthcare (2.6%), graduate medical education and disproportionate share medical payments to hospitals, hospital non patient revenue such as in the gift shop and for parking, and spending by foreign visitors. In total, this accounts for about 11% percent of NHEA spending. Second, we redefine some categories of medical services in the NHEA and MCBS, shifting expenditures as appropriate, to create consistent categories between the two sources. Our work here follows along the lines of Selden et al. (2001) and Sing et al. (2006). Third, we then 1 Selden et al. (2001) and Sing et al. (2006) attempted reconciliations between NHEA and MEPS for 1996 and 2002, respectively. In their work, MEPS-reported expenditures were reconciled with the comparable components of NHEA expenditures, omitting the institutionalized population and their spending.

11 Ratio proportionately increase spending in the MCBS by the factors necessary to have total survey spending equal the remaining portion of the NHEA total in each service-by-payer category. This paper focuses on the NHEA-reconciled estimates from MCBS. Figure 6a gives the adjustment factors by each service category. Overall, the NHEA adjusted spending is 11 percent higher than the total spending reported in MCBS. Figure 6a. Ratio of total spending: Adjusted NHEA and MCBS by Service Categories: Hosptial Care Physician and Clinical Services & DME Nursing Home Prescription Drugs Type of Service Dental Home Health Overall Note: This figure gives the adjustment factor for different types of services. Overall, the adjusted National Health Expenditure Accounts (NHEA) spending is 11 percent higher than the total spending reported in 2009 MCBS. We use these adjustment factors by service categories to adjust MCBS reported services to the national level. Table 6a (next page) gives the NHEA adjusted total and average medical spending in 2009 adjusted to $2010 US dollars, using the GDP deflator. We present these results separately for three age groups 65-74, 75-84, and and for six different services including hospital, physician and clinical services and durable medical equipment (DME), nursing home, prescription drugs, dental, and home health services. In 2009, total personal health care spending in United States for

12 the elderly was estimated to be $644 billion (in 2010 US $), with per capita spending of $17,480. On average, Medicare beneficiaries aged 65 to 74 spend $13,500 annually on personal health care. For this group, the average spending on hospital related services is approximately $5,000 and for physician and clinical services including durable medical equipment (DME) the average spending is nearly $4,000. The average spending on nursing homes for this group is much lower as compared to older beneficiaries, about $1,000. Table 6a. NHEA Adjusted Spending by Service Types: 2009 Type of Services Age Group (65-74) Total Spending (billions) Average Spending Hospital $89.55 $4,963 Physician and Clinical Services & Durable Medical Equipment ,952 Nursing Home ,071 Prescription drugs ,636 Dental Home Health Overall ,531 Age Group (75-84) Hospital ,675 Physician and Clinical Services & Durable Medical Equipment ,986 Nursing Home ,737 Prescription drugs ,567 Dental Home Health ,358 Overall ,645 Age Group (85+) Hospital ,625 Physician and Clinical Services & Durable Medical Equipment ,395 Nursing Home ,299 Prescription Drugs ,248 Dental Home Health ,913 Overall ,704 Note: Total and averages are weighted using final sample weights in 2009, MCBS. MCBS is matched to adjusted NHEA for all the above service categories. Here, N=6,200 and weighted N=36,824,486

13 The average spending for beneficiaries aged is 38% higher than for beneficiaries 65-74, about $18,500 annually. Spending is substantially higher for hospital care, nursing home, and home health care. As expected, beneficiaries 85 years and older spend even more. A typical person aged 85+ spends on average $26,700 annually with major spending on nursing home care ($9,000), hospital care ($7,600) and home health ($3,000).

14 Chapter 7. Calibration processes: community dwelling population The goal of the calibration process was to create a set of disease group indicators that reflects the prevalence of disease groups both treated (resulted in a claim), as well as latent (not medically addressed within a given year). This task was accomplished in two steps: first calibrating 25 disease groups with self-report (SR) available in NHANES; then calibrating the remaining 80 disease groups. This chapter provides details of the methods used in each of the 2 steps. Calibration of community dwelling population for disease groups available in NHANES For each year the data from MCBS and NHANES were appended. Without loss of any generality, assume that in the combined data set, the first n observations are from MCBS and the last m observations are from NHANES. Let X i,i = 1,2,,n,n +1,,n + mbe the common covariates for N = n + msubjects in the combined data set. Let C ij = 1denote the presence of claim for disease group j = 1,2,, por set to 0 otherwise for subject i = 1,2,,n where p is the number of disease groups for which we have self-report data in NHANES. Note that this variable is defined only for subjects in MCBS. Let S ij = 1,i = n +1, n + m indicate if the subject self-reported having disease group jor set to 0 otherwise. Note that this variable is available only for subjects in NHANES. Define a new indicator variable for the disease groups in a combined data set as follows: D ij = 1 if S ij = 1 or C ij = 1, D ij = 0 if S ij = 0 and D ij =. if C ij = 0 That is, the subject is assumed to have presence of the disease group, if the self-report (in NHANES portion of the appended data set) or the claim (in the MCBS portion of the appended data set) indicates the presence of disease; and not having the disease, if the self-report indicates the subject does not have disease; and, if there was no claim for the disease the actual disease group status is missing. We now have a standard missing data problem by setting aside variables (S,C) and only working with (X,D) in the combined data set. When imputing the missing values in D(only in MCBS) one needs to make sure that prevalence rates after imputation match with prevalence rates in NHANES. That is, the multiply imputed prevalence rates are calibrated to the observed prevalence rates in NHANES after adjusting for any differences in the covariate distribution. The sequential regression methodology was modified to result in calibrated prevalence indicator variables. Specifically, we defined a variable R = 1for the subjects in MCBS and R = 0for the subjects NHANES. The following steps describe an iteration in the sequential regression approach for calibrated multiple imputation using the variables (X,D,R):

15 1. Let D (- j) denote the collection of disease group indicators for all disease groups except disease group j. Construct a propensity score based on fitting a logistic regression model predicting Rwith (X, D (- j) )as covariates and create strata based on the propensity scores. This step groups the subjects in the two surveys based on similarity of the covariates and other disease groups. 2. Within each propensity score class, estimate the prevalence rate based on the self-report, S j and the claims C j. If the prevalence rate based on the claims is greater than or equal to that based on the self-report then set all missing D j to 0. That is, no additional imputation is necessary and all those without a claim for that disease group are considered as not having that disease. 3. If the prevalence rate based on the claims is smaller than the self-report prevalence rate then randomly some missing D j were set to 1 so that the prevalence rates after the imputation will match the self-report prevalence rates. We used several Bernoulli draws within each propensity score class to achieve this calibration. 4. Note that medical expenditure and disease groups without self-report are missing in the NHANES portion of the appended data. To be fully conditional, we imputed these missing values in the NHANES. 5. These steps were iterated across all diseases several times until the multiply imputed prevalence rates stabilized. Calibration of community dwelling population for disease groups not available in NHANES For calibrating disease groups that are not available in NHANES, relationships between the multiply imputed D j and claims based C j, j = 1,2,, p were developed. This can be viewed as an measurement error model and this relationship is then used to calibrate the disease groups, k = p +1, p + 2,,105. The following steps describe the imputation procedure for these disease groups: 1. Fit a measurement error regression model, 2. Fit two propensity models for the claims based disease groups, Pr(C j = 1 X) and Pr(C k = 1 X). 3. Match the disease groups j and kbased on the propensity score to find the closest match of the measurement error model developed in step Use the closest match measurement error model to impute D k,k = p +1, p + 2,,105. This process was repeated for each year. Resulting claim-based and calibrated disease prevalence estimates from MCBS and NHANES SR estimates for each year are provided in Appendices 7_8ak.

16 Chapter 8. Calibration processes: institutionalized population Indicators of calibrated disease groups in the community population were used to calibrate disease groups in the institutionalized sample. After the calibration/imputation of the disease groups D j, j = 1,2,,105 in MCBS as described in Chapter 7 for the community dwelling population, the same 105 disease groups were calibrated/imputed for institutionalized subjects in the MCBS. At this step, the imputed data on community dwelling subjects in MCBS were appended to the unimputed data on institutionalized subjects. Define I = 1, for institutionalized subjects and 0 for community dwelling subjects. The following steps describe the iterative process for calibrating claim based disease groups for institutionalized subjects: 1. For disease groups j, define the covariates as (X, D (- j) )and fit a propensity score model for institutionalization through a logistic regression model with I as the dependent variable. Let p j = Pr(I =1 X,D (- j) ). 2. Define as the probability of disease j without the presence of a claim for that disease group given the similarity of the institutionalized subjects to the community dwelling subjects. 3. Estimate the probability on the right hand side of the equation in step two by assuming and evaluating this density as the logit of the propensity score. 4. Draw a uniform random variable and impute new claims for the institutionalized subject with C j = 0,if. This process was repeated for each year. Resulting claim-based and calibrated disease prevalence for community and institutionalized populations are provided in Appendices 7_8a-k.

17 Chapter 9. Estimation of Calibrated Disease Prevalence ( ) To estimate rates based on calibrated disease groups we combined the community and institutionalized populations. We estimated prevalence and standard errors of calibrated disease groups from each multiple imputed (MI) dataset, and combined estimates from the 5 MI datasets using standard MI rules. The final weight was calculated as the product of the existing MCBS survey weight and the calculated selection weight. The results across the years, , are shown in Appendix 9a.