Essays on the use of microsimulation for health and economic policy analysis

PDF

Download

Open document

Transcript (if available)

Content ESSAYS ON THE USE OF MICROSIMULATION FOR HEALTH AND ECONOMIC POLICY ANALYSIS By Barbara Blaylock A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (PHARMACEUTICAL ECONOMICS AND POLICY) August 2015 ii ACKNOWLEDGEMENTS I would like to thank my committee members, Dana Goldman, Julie Zissimopoulos, Neeraj Sood, Geoffrey Joyce, and Maria Aranda, for their guidance and support. I extend a special thanks to my committee chairs, Dana Goldman and Julie Zissimopoulos, for their advice and encouragement during this dissertation process. I greatly appreciate the research assistance from my colleagues at the Schaeffer Center Data Core. I would also like to acknowledge the National Institute on Aging through the Roybal Center for Health Policy Simulation (P30AG024968) for funding for this work. iii TABLE OF CONTENTS ACKNOWLEDGEMENTS ........................................................................................................ ii LIST OF TABLES ...................................................................................................................... v LIST OF FIGURES ..................................................................................................................... vi CHAPTER 1. INTRODUCTION ................................................................................................ 1 CHAPTER 2. POPULATION HEALTH BENEFITS OF COLLABORATIVE CARE TREATMENT OF DEPRESSIVE SYMPTOMS IN A NATIONALLY REPRESENTATIVE SAMPLE OF OLDER ADULTS ............................................................. 5 2.1 Abstract ...................................................................................................................... 5 2.2 Introduction ............................................................................................................... 6 2.3 Conceptual Framework for Modeling Depression .................................................... 9 2.4 Microsimulation Inputs and Scenarios ...................................................................... 11 2.4.1 Data and Measures ..................................................................................... 12 2.4.2 Transition and Estimation Regression Models ........................................... 14 2.4.3 Microsimulation Scenarios ......................................................................... 16 2.4.4 Sensitivity Analysis .................................................................................... 19 2.5 Results ....................................................................................................................... 19 2.6 Discussion and Conclusion ........................................................................................ 30 2.7 References ................................................................................................................. 33 CHAPTER 3. IMPLICATIONS OF SOCIAL SECURITY ENTITLEMENT AGE POLICY CHANGE ON THE PRODUCTIVE ACTIVITIES OF OLDER ADULTS ............... 39 3.1 Abstract ...................................................................................................................... 39 3.2 Introduction ............................................................................................................... 39 3.3 Study Data and Methods ........................................................................................... 43 3.3.1 Health and Disability .................................................................................. 44 3.3.2 Government Expenditures .......................................................................... 45 3.3.3 Taxes ........................................................................................................... 46 3.3.4 Bequests and Inter Vivos Gifts ................................................................... 47 3.3.5 Caregiving and Volunteering ..................................................................... 47 3.3.6 Scenarios ..................................................................................................... 49 3.3.7 Limitations .................................................................................................. 49 3.4 Results ....................................................................................................................... 51 3.5 Discussion and Conclusion ........................................................................................ 59 3.6 References ................................................................................................................. 60 iv CHAPTER 4. CONSTRUCTION OF PERCENTILE CONFIDENCE INTERVALS INCLUDING SAMPLING VARIABILITY FOR POPULATION-BASED MICROSIMULATION RESULTS ............................................................................................. 64 4.1 Abstract ...................................................................................................................... 64 4.2 Introduction ............................................................................................................... 65 4.3 Methods ..................................................................................................................... 67 4.3.1 Inputs to the Future Elderly Model ............................................................ 68 4.3.2 Sampling from Survey Input Data .............................................................. 70 4.3.3 Confidence Intervals ................................................................................... 71 4.3.4 Time Series Assumptions ........................................................................... 72 4.3.5 Computational Resources ........................................................................... 74 4.4 Results ....................................................................................................................... 74 4.5 Methods to Reduce Computational Burden .............................................................. 84 4.6 Discussion and Conclusion ........................................................................................ 86 4.7 References ................................................................................................................. 88 APPENDIX ................................................................................................................................. 91 Appendix A. Chapter 2 Transition and Estimation Models ............................................ 91 A.1 Transitions of depressive symptoms, behavioral risks, and consequences of chronic conditions .................................................................... 91 A.2 Transitions of mortality and chronic conditions ........................................... 94 A.3 Models of medical spending ......................................................................... 96 Appendix B. Chapter 3 Transition and Estimation Models ............................................ 98 B.1 Transitions of health and disability: Chronic conditions .............................. 98 B.2 Transitions of health and disability: Mortality, functional impairment and nursing home residence ................................................................................ 100 B.3 Transitions of working for pay and benefit claiming ................................... 102 B.4 Transitions of wealth, income and other transfers ........................................ 104 B.5 Transitions of caregiving and volunteering .................................................. 107 B.6 Models of medical spending ......................................................................... 109 v LIST OF TABLES Table 2.1 Characteristics of 2004 65+ Community Dwelling Population by Depressive Symptoms Status ......................................................................................................................... 20 Table 2.2 Marginal Effects of Health and Demographic Characteristics on Clinically Significant Depressive Symptoms ............................................................................................... 22 Table 2.3 Marginal Effects of Clinically Significant Depressive Symptoms on Mortality, Consequences of Chronic Conditions, Behavioral Risks, and Medical Spending ...................... 24 Table 2.4 Effect of Collaborative Care Treatment on Population-Level and Lifetime 65+ Outcomes (2010 USD) ................................................................................................................ 26 Table 3.1 Public Expenditures and Revenues, Americans Ages 65 and Over in 2010 (2010 Dollars, Billions) ............................................................................................................... 52 Table 3.2 Private Transfers and Unpaid Services, Americans Ages 65 and Over in 2010 (2010 Dollars, Billions) ............................................................................................................... 53 Table 3.3 Public Expenditures and Revenues 2010 to 2050 (2010 Dollars, Billions) ................ 53 Table 3.4 Private Transfers and Unpaid Services, Americans Ages 65 and Over 2010- 2050 (2010 Dollars, Billions) ...................................................................................................... 54 Table 3.5 Effect of Delaying EEA and FRA on Public Expenditures and Tax Revenues (2010 Dollars, Billions) ............................................................................................................... 56 Table 3.6 Effect of Delaying EEA and FRA on Private Transfers (2010 Dollars, Billions) ...... 57 Table 3.7 High and Low Value of Unpaid Services, Americans Ages 65 and Over 2010- 2050 (2010 Dollars, Billions) ...................................................................................................... 58 Table 4.1 Mortality Reduction and Medical Growth Time Series Scenarios .............................. 73 Table 4.2 First-order uncertainty of 51-52 year old stock population, 2010 ............................... 75 Table 4.3 Second-order uncertainty with predicted transition probabilities and estimates at first transition of 51-52 year old stock population, 2010 ........................................................ 78 Table 4.4 Simulation results for life expectancy at age 50 with percentile confidence intervals, cohort ages 51-52 in 2010 ............................................................................................ 81 Table 4.5 Effect of period assumptions on lifetime outcomes with percentile confidence intervals, cohort ages 51-52 in 2010 ............................................................................................ 82 Table 4.6 Estimates of population level stochastic error for 100, 500, and 1200 repetitions of the simulation over 100 replicates ........................................................................................... 83 vi LIST OF FIGURES Figure 2.1 Sensitivity Analysis for Collaborative Care Program Costs and Effectiveness on Spending (2010 USD) ............................................................................................................ 29 Figure 3.1 Percent Change from Status Quo of Work and Claiming in 2010 ............................. 50 1 CHAPTER 1. INTRODUCTION In the US, population based microsimulation models are used to forecast effects of policy changes by government institutions such as the Congressional Budget Office and Social Security Administration, foundations such as the Robert Wood Johnson Foundation, academic institutions, and public policy think tanks such as the Urban Institute and RAND. Internationally, microsimulations have been developed for similar purposes. Microsimulation is a stochastic modeling technique that transitions individuals over time or events. Compared to macroeconomic modeling or component-based modeling, microsimulation can change assumptions at an individual level and is therefore commonly used for health and economic policy analysis. The primary component for the input to these simulations is often a nationally representative panel survey of health and/or economic characteristics. Survey and other types of health and economic data are being collected more often and in more depth. In the future, microsimulations will become more prevalent and more complex. Ideally, they will also become more transparent with better validation and discussion of confidence in the results such as in a typical empirical paper. Corroboration of results with other forecasting models is an exciting opportunity but requires the cooperation of other researchers at different institutions. The microsimulation used in this work is the Future Elderly Model (FEM), primarily based off the Health and Retirement Survey (HRS). The model has been developed over the last 15 years with funding by the National Institute on Aging and the MacArthur Foundation. The FEM is unique in that it starts with a real population of individuals surveyed in the HRS and transitions an extensive set of health and economic characteristics together and at an individual level. The transition models have been estimated with expert opinion on the causal effects 2 between chronic conditions in particular. Significant effort has been taken to complete internal and external validation. The essays presented here expand and strengthen the microsimulation and apply the model in the areas of health services research, labor and retirement, and uncertainty analysis. The first essay investigates the effects of a collaborative care treatment program for late life depressive symptoms on health and spending outcomes. Collaborative care, a care delivery method in which a caseworker coordinates patient-specific treatment to be administered by a primary care physician with the consultation of a mental health specialist, is the recommended treatment for late-life depressive symptoms. However, the national and long-term effects of making collaborative care the first line treatment for late life depressive symptoms is unknown. In order to examine the population-level and lifetime health and medical spending impacts of an improvement in late life depression treatment, this work expands the FEM microsimulation to include depressive symptoms. A model of late-life depressive symptoms is developed controlling for demographics, socioeconomic status, health, and other factors. Care is taken to incorporate depressive symptoms based on the best information about the causal relationship between depression and chronic conditions in the literature. Approximately 14% of the 65 and older community-dwelling population suffer from late-life depressive symptoms. A nationwide treatment program with similar costs and effectiveness as collaborative care trials could decrease the prevalence of depressive symptoms to 11.5% and would cost Medicare approximately $3 billion in 2020. The reduction in depressive symptoms decreases Medicare spending by $5 billion, so the collaborative care program is likely to pay for itself. In the long term, the collaborative care program could increase quality adjusted life years by almost half a month and decrease lifetime expected Medicare spending. 3 The second essay investigates the effects of increasing retirement ages on both public and private transfers. The full retirement age (FRA) is increasing and the SSA has examined the impact of further increases in combination with increases to the early entitlement age (EEA) on the solvency of the Social Security Trust Fund. In addition to old-age income entitlement changes, working longer should directly affect Medicare spending if individuals remain in jobs with healthcare coverage. Additionally, longer careers likely translate to more income tax liability but also greater wealth, which can be transferred to families in the form of bequests or other gifts. However, the effects of retirement age changes are not expected to be solely financial. Working longer also has the potential to affect the unpaid productive activities of older adults. In order to estimate the size and direction of the effect of changing retirement ages on financial as well as services transfers, this work expands the FEM microsimulation to include unpaid services such as informal caregiving and volunteering. Models for caregiving and volunteering are developed controlling for demographics, socioeconomic status, religious affiliations, and living arrangements. Unpaid services are valued at a replacement rate equivalent to the mean wage from the Bureau of Labor Statistics. In 2010, estimated government expenditures for old age benefits net of taxes paid by the over 65 population are $984 billion. Private transfers are estimated and valued at $722 billion. By 2050, a five-year increase in the FRA could increase private financial transfers by $4.5 billion but might also decrease volunteering and informal caregiving by $4 billion. The final essay is an uncertainty analysis of the FEM. Microsimulations are often referred to as ‘black box’ models because the model specifications and assumptions are not always clear. Uncertainty analysis attempts to clarify the relationship between variable model specification and input uncertainty and the reported results. However, the effect of uncertainty in general and 4 sampling variability of inputs in particular is often overlooked both by researchers and by policy makers. This is due to the fact that uncertainty analysis methods must often be adapted to complex microsimulation and the resulting analysis may be excessively computationally burdensome. This work adapts direct bootstrapping uncertainty analysis methods to the FEM. To overcome the computational burden, resources at the USC High Power Computing center are used. The goal is to assess the impact of multiple sources of uncertainty, but primarily sampling variability, on the microsimulation results. Percentile confidence intervals are constructed incorporating sampling variability from survey-based inputs such as the HRS. Bootstrap samples of the HRS – random samples with replacement taking into account survey design characteristics – are used to produce a large number of sets of implicitly correlated simulation inputs, which are then used to estimated population-level cohort life expectancy and lifetime medical spending outcomes. Life expectancy at age 50 is 81.4 and lifetime Medicare spending after age 65 is $183,924. The confidence intervals constructed with sampling variability have a range six times greater than those for the deterministic run of the model. Next, model specifications related to time series assumptions are tested and the remaining stochastic error is estimated. Stochastic error is found to be relatively close to zero by 1200 repetitions of the simulation. Mortality risk adjustment assumptions substantially increase both life expectancy (2.4 years) and Medicare spending ($27,437). 5 CHAPTER 2. POPULATION HEALTH BENEFITS OF COLLABORATIVE CARE TREATMENT OF DEPRESSIVE SYMPTOMS IN A NATIONALLY REPRESENTATIVE SAMPLE OF OLDER ADULTS 2.1 Abstract Collaborative care programs allow patients to be treated by primary care physicians with the advice of specialists and a patient-specific treatment plan. In the case of depression treatment, there is strong evidence that collaborative care programs decrease depressive symptoms and total medical spending compared to usual care. This study estimates the population prevalence of clinically significant depressive symptoms in the community-dwelling Medicare population and determines the health benefits of collaborative care treatment at a population level. Using the Future Elderly Model microsimulation and data from the Health and Retirement Study, Medicare Current Beneficiary Survey, and Medical Expenditure Panel Survey, a nationally representative sample of middle-aged and older adults is followed until 2030 for population-level results and from age 51 until death for lifetime results. Population prevalence of clinically significant depressive symptoms in the community-dwelling Medicare population is approximately 14% and disproportionately effects women and Hispanics. At the population level, collaborative care treatment for depressive symptoms reduces the prevalence to 11.5%, decreases the mortality rate, and reduces per capita Medicare spending. Over the long term, the increase in quality adjusted life years due to improved treatment efficacy is approximately half a month. From the Medicare perspective, the collaborative care program pays for itself and could result in annual savings of $2 billion compared to usual care if the efficacy and incremental costs from randomized control trials can be matched. 6 2.2 Introduction Medicare spending is projected to increase to 4.6 percent of gross domestic product (GDP) by 2039, up from 2.9 percent in 2015 (2014). There are several reasons for this spending increase. Because of demographic shifts and increases in life expectancy, the US population 65 years or older (65+ population) is projected to increase from 14 percent of the US population in 2014 to 21 percent by 2040 (Administration on Aging 2012; 2014). Additionally, the proportion of the US 65+ population with two or more chronic conditions increased to 45.3 percent in 2009- 2010 up from 37.2 percent a decade earlier (Freid, Bernstein and Bush 2012). The Centers for Medicare and Medicaid Services (CMS) is investigating ways to rein in spending. The Health Care Innovation Awards provide funding for research for alternative delivery methods for care, which will increase quality, decrease spending or both. The 2015 Medicare fee schedule includes care coordination, or non-face-to-face services, for disease management of primary care patients with two or more comorbid conditions, including depression (2013a; 2014; 2012). However, it is unknown how reimbursing primary care physicians for care coordination will impact net medical spending. Depression is a significant problem in the 65+ population that too often is not recognized (Byers et al. 2010). Clinical significance of depressive symptoms occurs at levels sub-syndromal to major depressive disorder (MDD) in older adults (Ayuso-Mateos et al. 2010). Estimates of the prevalence of clinically significant depressive symptoms (CSDS) in the community-dwelling 65+ population range from 9-16 percent (Blazer 2003). Depressive symptoms increase mortality, negatively impact quality of life, increase the likelihood of developing MDD, and increase healthcare spending by complicating the treatment of comorbid conditions (Donohue and Pincus 2007; Karsten et al. 2011; Katon 2003; Lyness et al. 2002; Lyness et al. 2009). Reducing 7 depressive symptoms should be targeted as a way to reduce healthcare costs and improve the length and quality of life (Young et al. 2010). The gold standard of depression diagnosis at all ages is the Diagnostic and Statistical Manual of Mental Disorders (DSM) criteria (American Psychiatric Association 2013b). However, late- life depressive symptoms are more likely to include anhedonia, cognitive impairment, functional impairment, sleep problems, agitation, weight loss, health-related anxiety, and agitation (O’Donnell and Kaszniak 2011). Under-diagnosis is common among older adults whose symptoms may be conflated with comorbid conditions (González et al. 2010; Saint Onge, Krueger and Rogers 2014). Even with a clinical diagnosis, older adults often do not receive adequate treatment of depressive symptoms (Barry et al. 2012; Byers et al. 2010). When treated, older adults prefer to be seen by their primary care physician (Steinman 2007), but primary care physicians often do not have the expertise to identify and effectively treat depression (Solway et al. 2010). Collaborative care programs for depression treatment were developed to aid primary care physicians in treating depressed older adult patients by coordinating care with specialists and developing patient-specific treatment plans (Katon et al. 2010). Programs have adapted over time to include evidence-based improvements. Collaborative programs incorporate patient preferences during the treatment plan creation step and assign care managers to monitor relevant outcomes for stepped protocols (Katon et al. 2010; Unutzer 2002). Randomized clinical trials (RCTs) have shown that collaborative care programs for depression reduce depressive symptoms and are cost- effective (Gilbody et al. 2006; Jacob et al. 2012; Thota et al. 2012). Collaborative care treatment is also cost effective for subthreshold depression (Wells et al. 2007), groups with comorbid 8 disease such as diabetes (Huang et al. 2013), and racial/ethnic minorities (Fuentes and Aranda 2012). Collaborative care programs increase the cost of depression treatment but are considered cost effective in many populations. The incremental costs of collaborative care programs for depression treatment include the specialist and primary care charges in addition to patient education expenses. Per person per year, the incremental and overall costs of collaborative care programs published between 1980-2009 were $204 ($104/$850) and $436 ($104/$2,160), respectively, on average in 2008 USD (Jacob et al. 2012). Of six studies that completed a cost per quality adjusted life year (QALY) analysis, collaborative care programs were cost effective at the $150,000 per QALY level for at least one group in each study (Jacob et al. 2012). The collaborative care literature is limited in several ways. First, RCTs are limited by the small number of outcomes in the trial design. Based on the interactions between depression and chronic conditions, other outcomes of interest include the impact of depression on risky behaviors such as obesity and smoking as well as on the symptoms of chronic conditions, functional impairment, and health-related quality of life. Longer-term follow up would allow for examination of the impact of depression on the incidence of chronic conditions and life expectancy. The cost impact of providing collaborative care treatment of depressive symptoms to the full Medicare eligible population has not been estimated. This study examines the population-level, long-term impacts of collaborative care treatment for depression on Medicare spending and the healthy life expectancy of the 65+ community dwelling population. The primary objective is to estimate the impact of collaborative care programs for the treatment of depression on population level healthcare spending, life expectancy and quality of life in older adults. First, a model of depressive symptoms is 9 developed for older adults taking into account the complex relationship between depression and chronic conditions as well as other health and economic characteristics. Second, a population- based microsimulation is used to age a nationally representative sample of middle aged and older adults to predict the healthcare spending and changes in mortality and other health characteristics associated with collaborative care treatment of depressive symptoms. 2.3 Conceptual Framework for Modeling Depression Depressive symptoms in the community-dwelling 65+ population are positively correlated with being female, chronic conditions such as heart and lung disease, functional impairment, and smoking and are negatively correlated with higher levels of education after controlling for demographic characteristics, socio-economic status, chronic health conditions, symptoms of chronic conditions, and behavioral risks (Zivin et al. 2010). Baseline depressive symptoms predict depression both one and two years later (Karsten et al. 2011; Lyness et al. 2002; Lyness et al. 2009). Although family caregiving may decrease all cause mortality of the caregiver (Roth et al. 2013), it may also be associated with worse mental health of the caregiver (Seegert 2013). However, the causal relationships, particularly between depression and chronic conditions, are less well understood (Blazer et al. 2009). A theoretical model of the pathways between depression and chronic conditions proposes that the causal direction runs both ways but primarily through different pathways. Depression indirectly affects chronic conditions through behavioral risks and chronic conditions indirectly affect depression through the consequences of chronic conditions (Katon 2011). Behavioral risks such as obesity, smoking, and inactivity are positively correlated with both depression and chronic conditions. In the causal diagram developed by Katon (2011), depression 10 increases the level of behavioral risks and in turn increases the probability of onset of chronic conditions. In support of the causal direction primarily leading from depression to chronic conditions through behavioral risks, a prospective study of cardiovascular patients found no significant association between depression and cardiovascular events after controlling for behavioral risks including smoking (hazard ratio (HR)=1.39 (1.17-1.66)), medication non- adherence (HR=1.90 (1.31-2.76)), and self-reported physical activity (HR=1.44 (1.14-1.82)) (Whooley et al. 2008). There may be other pathways in addition to behavioral risks to account for the correlation between depression and chronic conditions. Depression may increase incident diabetes through self-care behavior, complication of symptoms (Egede and Ellis 2010), or biologic mechanisms (Heeramun-Aubeeluck, Lu and Luo 2012). Biological changes associated with depression in older adults are also associated with increased risk of myocardial infarction (Blazer et al. 2009). Leading from chronic conditions to depression, consequences of chronic conditions such as symptom burden, functional limitations, quality of life, and neurological changes are the primary causal pathways (Gayman, Turner and Cui 2008; Hilderink et al. 2012). Chronic condition onset necessarily increases the level of symptom burden and in turn increases the probability of developing depression. Functional limitations as a prodrome for depression and stressful event theory, disruptions of normal life increase the risk for depression, have been investigated as the causal link between functional limitations and depression (Bruce 2001). Suggested paths from symptoms such as pain to depression include feelings like helplessness (Palomino et al. 2007). The relationship between depression and consequences of chronic conditions is complicated by direct and indirect feedback from depression to consequences of chronic conditions. For example, the association between depression and cognitive impairment is likely neurobiological 11 in both directions (Dotson, Beydoun and Zonderman 2010; Saczynski et al. 2010). Depression might directly increase the incidence of functional limitations (Carrière et al. 2011) and lower the reporting thresholds for symptoms of cognitive impairment (Jorm 2001). Depression also affects consequences of chronic illness indirectly through self-care behavior such as diet, exercise, and monitoring individual health. Depression is associated with poor medication adherence (Ciechanowski, Caton and Russo 2000; Zivin et al. 2009), decreased likelihood of quitting smoking (Bock et al. 2010; Katon and Ciechanowski 2002), cancellations of scheduled doctors appointments (Nutting et al. 2002), and reduced likelihood of getting a flu shot or eye exam and decreased monitoring activities in older adults with diabetes (Egede and Ellis 2010; Meeks et al. 2011). These negative self-care behaviors further complicate or delay treatment of consequences of chronic conditions. Based on the evidence for causal relationships, a model of depressive symptoms is developed to forecast population-level late life depressive symptoms. This study adds to the current literature by modeling depressive symptoms and its effect on other health and economic characteristics at the population level and in the long term. To the author’s knowledge, there is no other population-level longitudinal model of depressive symptoms. A further contribution is the estimation of population level health and medical spending changes if collaborative care treatment of depressive symptoms became the standard for the aged Medicare population. 2.4 Microsimulation Inputs and Scenarios Late life depressive symptoms are added to a microsimulation model of health and economic outcomes that individually ages middle age and older adults through 2030 or until death. Parameters of the microsimulation can be altered to simulate policy changes such as the 12 implementation of collaborative care treatment programs. Individual results are aggregated to estimate the impact on population health and economic outcomes. This study builds on previous work using the Future Elderly Model (FEM) in the areas of medical technology advancement, disability, and obesity among other topics (Goldman et al. 2013). Development of the FEM has been supported by grants from the National Institute on Aging, the National Institutes of Aging, and the MacArthur Foundation Network on an Aging Society. This section describes the key measures, model development, and simulation scenarios used in the analysis. 2.4.1 Data and Measures The primary data source of the FEM is the Health and Retirement Study (HRS), a biennial longitudinal survey of individuals 51 and older (Wallace and Herzog 1995). HRS data is supplemented with Medicare Current Beneficiary Survey (MCBS) and Medical Expenditure Panel Survey (MEPS). In the HRS, a binary indicator of clinically significant depressive symptoms was created based on the Center for Epidemiologic Studies Depression Scale (CES-D) questionnaire, which has been validated as an instrument to indicate significant depressive symptoms (Radloff 1977). The HRS CES-D asks eight binary response questions about the respondent’s mood in the last week. The feelings include (1) “Much of the time during the past week, you felt depressed. Would you say yes or no?”; (2) “Much of the time during the past week, you felt that everything you did was an effort.”; (3) “Your sleep was restless.”; (4) “You were happy.”; (5) “You felt lonely.”; (6) “You enjoyed life.”; (7) “You felt sad.”; and (8) “You could not get going.” Clinical significance is defined as a score of four or greater out of a maximum score of eight (Steffick 2000). 13 A binary indicator of depressive symptoms in the MCBS was created using methods adapted from the literature (Bambauer et al. 2007; Zivin et al. 2009). The MCBS question used to address sadness is “In the past 12 months, how much of the time did you feel sad, blue, or depressed?” The responses “all of the time” or “most of the time” are considered positive indicators of depressive symptoms. The question “In the past 12 months, have you had 2 weeks or more when you lost interest or pleasure in things that you usually cared about or enjoyed?” is used to address loss of interest. Individuals are considered to have depressive symptoms if they have reported feeling sad most of the time or have experienced a loss of interest. Additional measures include indicators of behavioral risks such as smoking (never smoked, past smoker, and current smoker) and obesity (body mass index (BMI)>30); chronic conditions such as ever having had cancer, diabetes, heart disease, hypertension, lung disease, or stroke; and consequences of chronic conditions such as difficulty with activities of daily living (ADLs) (one, two, or three or more), difficulty with instrumental activities of daily living (IADLs) (one or two or more), pain level (no pain, mild pain, moderate pain and severe pain), and cognitive impairment (no impairment, cognitive impairment that is not dementia, and dementia). Primary outcome variables include total medical spending — Medicare fee-for-service, Medicare HMOs, Medicaid, employer-based private health insurance, individually purchased private health insurance, private insurance managed care, private insurance with unknown sources, the VA and other public insurance, out-of-pocket payments and uncollected liability — and Medicare spending estimated from the 2007 to 2010, or post Part D implementation, MCBS and quality adjusted life years (QALYs) based on the European Quality of Life 5 Dimension (EQ-5D) instrument estimated from MEPS. Medical spending is adjusted using the medical consumer price index (CPI). 14 2.4.2 Transition and Estimation Regression Models Transition models for use in the microsimulation were developed based on the theoretical relationship between depression and chronic conditions. Every transition model controls for demographic characteristics (age as a spline with break points at 65 and 75; sex; marital status as married, single, or widowed; race/ethnicity as non-Hispanic white, non-Hispanic black, or Hispanic; and education level as less than high school, high school degree, or some college or greater) and health characteristics at age 50 (health conditions diabetes, cancer, heart disease, hypertension, lung disease, and stroke; smoking status; and marital status) in addition to model specific covariates. Specific changes directly related to this project are discussed here. Additional detail on all models can be found in the supplemental technical appendix. The model of clinically significant depressive symptoms additionally includes consequences of chronic conditions including a count of chronic conditions to proxy for symptom burden, functional status, pain level, and cognitive impairment; any helper hours to a spouse or to a parent; and clinically significant depressive symptoms in the previous period. Based on the literature and the theoretical model of depression interactions with chronic conditions, the indicator of clinically significant depressive symptoms is directly included as a covariate in the models of mortality, behavioral risks, and consequences of chronic conditions but not in the models of chronic conditions. Depressive symptoms are expected to increase mortality risk, decrease the probability of quitting smoking, and worsen the consequences of chronic conditions. To avoid the unintended consequences of feedback loops in the Markov model, depressive symptoms are only allowed to increase the incidence of consequence of chronic conditions indirectly. The forced assumption is that the coefficient on non-interacted 15 depressive symptoms will be zero and the term is left out of the models of consequences of chronic conditions. The relationship between depressive symptoms and body mass index (BMI) is less straightforward in an older adult population. Depressive symptoms may lead to weight loss in the elderly, so depressive symptoms are expected to increase BMI for middle-aged adults and decrease BMI for older adults. Models of medical spending and QALYs control for current period variables because they are estimated after the Markov processing in the simulation. Annual medical spending is estimated separately for total medial spending and Medicare Parts A and B controlling for a respondent’s demographic characteristics; chronic conditions; institutionalization in a nursing home; and death. Depressive symptoms are included in the models as well as interactions with chronic conditions to account for the effect of disease interactions on spending. The marginal effect of depressive symptoms on per capita healthcare spending is expected to be positive. The estimation of a QALY model for use in the simulation is a 3-step process. First, the EQ- 5D score is estimated using an OLS regression on the MEPS data controlling for self reported health, functional limitations, chronic conditions, weight, smoking status, and marital status. Second, an EQ-5D score is predicted on the HRS data using the model from the MEPS data. The last step uses the predicted EQ-5D scores to estimate a regression of QALYs controlling for covariates used in the simulation. QALYs are estimated controlling for chronic conditions including cancer, diabetes, heart disease, hypertension, lung disease, and stoke; functional status as one, two, three or more ADLs and one or two or more IADLs; smoking status; weight status; and marital status. 16 2.4.3 Microsimulation Scenarios The FEM microsimulation is used to age a nationally representative population of individuals aged 51 and over through 2030 or until the end of life in two-year time steps. Program effects are estimated by comparing outcomes from a baseline scenario to scenarios using alternative assumptions. Baseline The baseline, or usual care, scenario uses the status quo assumptions of the FEM. Data that starts the simulation is HRS Wave 7 and is weighted to match census estimates for population in 2004 based on age, sex, race/ethnicity. In order to keep a nationally representative sample, new cohorts of 51-52 year olds are added each time step to maintain a full population. New cohorts of 51-52 year olds are simulated for each year 2006 to the end of the simulation based on 1992 and 2004 waves of the HRS then characteristics are altered to match expected trends in health status including smoking rates and obesity. Supplemental data used to trend the health and socio- economic characteristics of the incoming 51-52 year olds include the National Health Interview Study (NHIS), the Census, and sources from the literature. Evidence suggests increasing trends in disability, obesity, and chronic disease and decreasing trends in smoking among younger cohorts (Crimmins and Beltrán-Sánchez 2011; Lakdawalla, Bhattacharya and Goldman 2004). Immigration is included each wave of the simulation based on Census estimates and projections by year, age, and sex. Annual outcomes, such as medical spending and QALYs, are estimated each wave at the individual level. Life expectancy and lifetime medical spending can be estimated for each cohort of 51-52 year olds. The present value of lifetime medical spending is reported in 2013 dollars using a 3% annual discount rate (Gold et al. 1996). 17 Collaborative Care Treatment of Late-Life Depressive Symptoms A scenario of collaborative care treatment of depressive symptoms utilizes the program costs and effectiveness of the Improving Mood–Promoting Access to Collaborative Treatment (IMPACT), a randomized clinical trial (RCT) with usual care control group. Individuals eligible for treatment are the 60+ community dwelling population with late life depressive symptoms but without dementia. The IMPACT RCT has published studies of 12-month outcomes (Unützer et al. 2002), 24-month outcomes (Hunkeler et al. 2006), 24-month cost effectiveness (Katon et al. 2005), and 4-year cost outcomes (Unützer et al. 2008). The IMPACT treatment reduced depressive symptoms by half in 44.67% of the treatment group. This level of clinical effectiveness is similar to other clinical trials of collaborative care interventions on depression (Gilbody et al. 2006; Thota et al. 2012). Individuals in the treatment group had 2 additional months of antidepressant use compared to usual care in 80% of the treatment group, and an average of 6.34 psychotherapy sessions in 30% of the treatment group. Twelve months after treatment, the 50% reduction in depressive symptoms remained in 33.85% of the treatment group. In the collaborative care scenario, individuals eligible for treatment are at least 60 years old, community-dwelling, non-demented, and have clinically significant depressive symptoms. Eligibility for treatment is assessed each wave based on transitioned characteristics. One year of collaborative care treatment is assumed to begin at the beginning of the wave, and treatment response is assigned randomly to the group eligible for treatment. A reduction of depressive symptoms by half is assumed to be equivalent to a move from clinically significant to non- clinically significant depressive symptoms. The scenario uses the two-year treatment response 18 from the IMPACT trial (33.85%). The depressive symptoms cure is temporary and individuals may become depressed and be treated again in a future wave. In addition to the effect on depressive symptoms, medical spending costs are increased by the one-year incremental program and treatment costs. The cost of services provided via the IMPACT treatment was $553 per person per year (Unützer et al. 2002). The average costs of a month of antidepressant medication and a 45- minute psychotherapy appointment are calculated from the 2010 MCBS. The average cost of a 30-day supply of antidepressant medication in the 2010 MCBS is $43.75. Antidepressant NDCs are based on Healthcare Effectiveness Data and Information Set (HEDIS) list of antidepressant medication. The average cost of a 45-minute psychotherapy appointment in the 2010 MCBS is $79.78 (CPT=90806). There was an adjustment in 2013 to the coding of psychotherapy appointments (CPT= 90834), but the effect on average cost is unknown. The total estimated cost of a 12-month collaborative care program is $777 in 2004 USD. Assuming a copay of 20% for all outpatient services, the incremental cost of the program to Medicare Part B is $577 in 2004 USD. Program costs are added to total medical spending and Medicare Part B spending for individuals eligible for treatment in the simulation. The scenario is run with and without incremental program costs in order to separate out cost saving due to decreased levels of depressive symptoms and additional treatment costs due to the collaborative care program. Prevention of Late-Life Depressive Symptoms The prevention of depressive symptoms scenario is representative of additional medical costs due to depressive symptoms. It is meant to be illustrative of an upper bound on population health 19 and spending effects from eliminating depressive symptoms from the older adult population. All individuals eligible for collaborative care are prevented from developing depressive symptoms in the prevention scenario. The results are comparable to the collaborative care scenario without program costs. 2.4.4 Sensitivity Analysis A sensitivity analysis tests the corner states from ranges of collaborative care program costs (incremental treatment costs=($265, $901); incremental costs to Medicare Part B=($167, $676)) and effectiveness (reduction in depressive symptoms=(22.99% or 24-month usual care treatment response, 44.67% or 12-month treatment response)) (Jacob et al. 2012; Unützer et al. 2002). Eight additional scenarios are run to complete a matrix for low, mid, and high estimates of program costs and depressive symptoms reduction. 2.5 Results Table 2.1 shows the prevalence of clinically significant depressive symptoms in the 65+ community-dwelling population and the correlation with demographic and health characteristics. The prevalence of clinically significant depressive symptoms in the 2004 HRS population, the wave used to start the simulation, is 14.4 percent. The depressed population is slightly older (p=0.0002). Sex (p<0.001), race/ethnicity (p<0.001), and education level (p<0.001) are significantly correlated with clinically significant depressive symptoms. Women (67.4% v. 55.6%), Hispanics (9.3% v. 5.3%), and individuals with less than a high school degree (40.8% v. 22.6%) are larger proportions of the clinically significant depressive symptoms population than the population without clinically significant depressive symptoms. Health characteristics such as 20 chronic conditions except cancer (p=0.106), functional status, pain level and cognitive impairment are significantly positively correlated (p<0.001) with depressive symptoms. The 2004 MCBS distribution of patient characteristics by depressive symptoms is similar to that of the 2004 HRS. Table 2.1 Characteristics of 2004 65+ Community Dwelling Population by Depressive Symptoms Status HRS MCBS Characteristic With CSDS Without CSDS p- value a With DS Without DS p- value a Total Total N, in millions 34.6 5 29.6 --- 34 4.3 29.6 --- (%) -100 -14.4 -85.6 -100 -12.8 -87.2 Age, mean 75.2 75.9 75.1 0.0002 75.5 75.8 75.4 0.1828 Sex, % <0.001 <0.001 Female 57.3 67.4 55.6 56.8 65.5 55.5 Male 42.7 32.6 44.4 43.2 34.5 44.5 Race/ethnicity, % <0.001 <0.001 Non-Hispanic White 81.9 76.2 82.9 84.6 80.9 85.1 Non-Hispanic Black 8.2 11.6 7.7 8.2 7.8 8.3 Hispanic 5.9 9.3 5.3 7.2 11.3 6.6 Education level, % <0.001 <0.001 Less than high school 25.2 40.8 22.6 28.0 39.5 26.4 High school degree 37.8 35.7 38.2 37.2 36.1 37.4 Some college or more 36.9 23.5 39.2 34.7 24.4 36.2 Chronic conditions, % Ever had cancer 17.9 18.9 17.8 0.106 18.7 20.1 18.5 0.541 Ever had diabetes 18.7 26.3 17.4 <0.001 20.8 26.3 20.0 <0.001 Ever had heart disease 30.8 41.8 29.0 <0.001 41.6 50.9 40.2 <0.001 Ever had hypertension 59.0 67.9 57.6 <0.001 63.0 69.3 62.0 <0.001 Ever had lung disease 10.7 18.8 9.4 <0.001 14.6 21.3 13.6 <0.001 Ever had stroke 10.0 16.1 9.0 <0.001 11.7 17.8 10.8 <0.001 21 BMI, mean 26.6 27.0 26.5 0.0001 Smoking status, % <0.001 Never smoked 33.6 24.6 35.1 Used to smoke 57.0 61.6 56.2 Currently smoking 9.4 13.8 8.7 ADLs, % <0.001 No difficulty 83 62.2 86.5 One 9.1 17.6 7.7 Two 3.6 8.3 2.8 Three or more 4.2 11.9 3 IADLs, % <0.001 No difficulty 90.4 79.5 92.2 One 6.3 13.0 5.1 Two or more 3.3 7.6 2.6 Pain level, % <0.001 No pain 69.4 47.5 73 Mild 8.2 9.3 8 Moderate 17.1 29.5 15 Severe 5.4 13.8 4 Cognitive impairment, % <0.001 No impairment 72.9 56.1 75.8 CIND 20.3 31.0 18.5 Dementia 6.8 13.0 5.7 a Statistical comparisons of groups with and without clinically significant depressive symptoms uses two-tailed t- tests for continuous variables and chi-squared tests for categorical variables. Abbreviations: HRS Health and Retirement Study, MCBS Medicare Current Beneficiary Survey, CSDS clinically significant depressive symptoms, ADL activities of daily living, IADL instrumental activities of daily living, CIND cognitive impairment without dementia. Table 2.2 shows the marginal effect (ME) of health and economic covariates on clinically significant depressive symptoms evaluated at the average age of the 65+ community-dwelling population in the starting population (mean age=75.2). The model estimates are consistent with the literature. The largest marginal effect is the lag of clinically significant depressive symptoms 22 (ME=33.0 percentage points (pp)). Hispanics (ME=3.0pp) and individuals with a lower level of education (ME=2.6pp) are more likely to report from clinically significant depressive symptoms. Men are less likely than women to report clinically significant depressive symptoms (ME=- 3.3pp). Table 2.2 Marginal Effects of Health and Demographic Characteristics on Clinically Significant Depressive Symptoms a Covariate Marginal Effect b p-value Sex Female reference Male -0.0325 <0.001 Race/ethnicity Non-Hispanic White reference Non-Hispanic Black 0.0076 0.063 Hispanic 0.030 <0.001 Education level Less than high school 0.0258 <0.001 High school degree reference Some college or more -0.022 <0.001 Marital Status Married reference Widowed (lag) 0.0103 0.016 Single (initial) 0.0151 <0.001 Depressive symptoms (lag) None or not clinically significant (CES-D<4) reference Clinically significant depressive symptoms (CES-D≥4) 0.330 <0.001 Number of chronic conditions (lag) 0 reference 1 0.0188 <0.001 2 0.0274 <0.001 3 0.0436 <0.001 4 0.0554 <0.001 23 5 0.0743 <0.001 6 0.0760 0.182 Functional limitations - IADLs No difficulty reference One 0.0309 <0.001 Two or more 0.0234 0.017 Functional limitations - ADLs No difficulty reference One 0.0365 <0.001 Two 0.0591 <0.001 Three or more 0.0643 <0.001 Cognitive impairment Not impaired reference CIND 0.0288 <0.001 Dementia 0.0297 <0.001 Pain status No pain reference Mild pain 0.0354 <0.001 Moderate pain 0.0511 <0.001 Severe pain 0.0639 <0.001 Help provided to parents No care provided reference Any hours of care provided 0.0079 0.054 Help provided to spouse No care provided reference Any hours of care provided 0.0448 <0.001 N 53,860 Pseudo R-squared 0.233 a Clinically significant depressive symptoms is modeled as a non-absorbing condition using a probit model. Age is included in the model as a spline (<65, 65-74, 75+). Characteristics at age 50 are used to control for cohort effects in the simulation model. b Marginal effects are evaluated at the average age of the 65+ community dwelling population in the 2004 HRS data (mean age=75.2). Abbreviations: CES-D Center for Epidemiologic Studies Depression Scale, ADL activities of daily living, IADL instrumental activities of daily living, CIND cognitive impairment without dementia. 24 The probability of reporting clinically significant depressive symptoms increases with an increase in the number of chronic conditions (ME=1.9pp, 2.7pp, 4.4pp, 5.5pp, 7.4pp, 7.6pp for 1 to 6 chronic conditions, respectively). Similarly for pain status, more severe pain (ME=3.5pp, 5.1pp, 6.4pp for mild, moderate, and severe pain, respectively) has a greater effect on clinically significant depressive symptoms. Any level of cognitive impairment increases the probability of reporting clinically significant levels of depressive symptoms by approximately 2.9pp-3.0pp. Providing unpaid care to a spouse (ME=4.5pp) also increases the probability of reporting clinically significant depressive symptoms. The indicator of widowhood does not distinguish recent widowhood and so does not capture bereavement. Table 2.3 Marginal Effects of Clinically Significant Depressive Symptoms on Mortality, Consequences of Chronic Conditions, Behavioral Risks, and Medical Spending Characteristic Marginal Effect p-value Mortality a 0.006 0.01 Consequences of Chronic Conditions Functional limitations – IADLs a One 0.015 <0.001 Two or more 0.004 <0.001 Functional limitations - ADLs a One 0.014 <0.001 Two 0.006 <0.001 Three or more 0.004 <0.001 Cognitive impairment a CIND 0.015 <0.001 Dementia 0.004 0.034 Pain status a Mild pain 0.003 <0.001 Moderate pain 0.015 <0.001 Severe pain 0.015 <0.001 25 Behavioral Risks BMI a -0.40% 0.058 Smoking status a Used to smoke 0.006 0.224 Currently smoking 0.002 0.151 Medical Spending (2004 USD) Total medical spending $4,571 <0.001 Medicare Part A $2,015 <0.001 Medicare Part B $1,037 <0.001 a Marginal effects are evaluated at the average age of the 65+ community dwelling population in the 2004 HRS data (mean age=75.2). Abbreviations: ADL activities of daily living, IADL instrumental activities of daily living, CIND cognitive impairment without dementia. Table 2.3 shows the marginal effect of depressive symptoms on mortality, consequences of chronic conditions, behavioral risks, and medical spending. Including clinically significant depressive symptoms in other transition and estimation models gives results consistent with the literature. Depressive symptoms increase the mortality rate (ME=0.6pp); increase level of pain (ME=0.3pp, 1.5pp, 1.5pp for mild, moderate, and severe pain, respectively), functional difficulty (ME=1.5pp, 0.4pp for 1, 2+ IADLs and 1.4pp, 0.6pp, 0.4pp for 1, 2, 3+ ADLs, respectively), and cognitive impairment (ME=1.5pp, 0.4pp for CIND and dementia, respectively); but have no significant marginal effect on BMI (p=0.058) and smoking status (smoke ever p=0.224 and smoke now p=0.151). Depression complicates the treatment of comorbid conditions, so the depressive symptoms measure was interacted with chronic conditions in the models estimating medical spending. The marginal effect of depressive symptoms increases total medical spending (ME=$4,571), Medicare Part A ($2,015) and Part B ($1,037) spending. Table 2.4 shows the population-level and lifetime results of the baseline scenario and changes from the baseline for other simulation scenarios. The population-level results are 26 reported for the 65+ community-dwelling population in 2020, and the lifetime 65+ results are reported for a cohort of individuals ages 51-52 in 2010. In the baseline scenario, the prevalence of clinically significant depressive symptoms (14.0%) has only a small change from the rate calculated in the 2004 HRS sample (14.4%). Per capita Medicare spending is $9,778 in 2010 USD. Life expectancy at age 65 is 83.9 years old, or 18.9 additional years. The remaining QALYs at age 65 are 14.15. The present value of lifetime Medicare spending after age 65 is estimated to be $184,786. Table 2.4 Effect of Collaborative Care Treatment on Population-Level and Lifetime 65+ Outcomes (2010 USD) Difference from Baseline… Scenario Baseline CC including program cost CC without program cost Prevention Population-level outcomes in the 65+ Community Dwelling Population 2020 Prevalence of clinically significant DS 14.0% -2.5% -2.5% -12.9% 2-year incidence of death 7.25% -0.03% -0.03% -0.17% Population Spending (in billions) - Total medical $894 -$1 -$7 -$37 Population Spending (in billions) - Medicare $500 -$2 -$5 -$26 Per capita spending - Total medical $17,469 -$34 -$151 -$809 Per capita spending - Medicare $9,778 -$37 -$103 -$548 Lifetime 65+ Outcomes in a Cohort Age 51-52 in 2010 Life expectancy at 65 83.9 0.05 0.05 0.25 Lifetime episodes of clinically significant DS 2.3 -0.3 -0.3 -1.5 Lifetime QALYs 14.15 0.04 0.04 0.21 Lifetime total medical spending a $341,272 $302 -$1,238 -$6,729 Lifetime Medicare spending a $184,786 -$138 -$1,125 -$5,779 a Present value of medical spending using a 3% discount rate. Abbreviations: DS depressive symptoms, QALY quality adjusted life year, CC collaborative care. 27 The collaborative care scenario reduces the prevalence of depressive symptoms in the 65+ community dwelling population to 11.6 percent. Program expenses per capita are calculated as the per capita spending in the collaborative care scenario including program costs minus the per capita spending in the collaborative care scenario without program costs. The costs benefits of the decrease in depressive symptoms (total medical spending=$151; Medicare=$103) are slightly larger than expenses due to program costs (total medical spending=$117; Medicare=$66). As expected, the collaborative care impact on functional difficulties is small. The QALY estimation captures the effect of changes in several dimensions of health status in one measure. The increase in life expectancy (0.05 years) is largely an increase with good quality of life (0.04 QALYs). The decrease in depressive symptoms due to collaborative care treatment reduces present value of lifetime Medicare spending by $1,125. The cost of the collaborative care treatment to Medicare is $988 over the course of a lifetime after age 65. Preventing clinically significant depressive symptoms in the population eligible for collaborative care would have a larger impact on mortality and medical spending than treating the condition after it occurs. Population prevalence of clinically significant depressive symptoms decreases to 1.2 percent in the 65+ community-dwelling population. The two-year mortality rate drops by 0.17%. The cost savings to Medicare from preventing depressive symptoms ($26 billion) is over five times larger than the collaborative care scenario not including program costs. Compared to the collaborative care scenario, the prevention scenario increases the change in life expectancy five times (0.25), quadruples the increase in lifetime QALYs (0.21), and produces a savings for total medical spending and Medicare spending of $6,729 and $5,779, respectively. Figure 2.1 shows sensitivity analysis results for the collaborative care treatment scenario medical spending outcomes. If the effectiveness of the program was no better than the depressive 28 symptoms response of the usual care group from the IMPACT trial, there was no decrease in population-level Medicare spending for the mid cost scenario, and there was an increase in spending for the high cost scenario. The increase in Medicare spending for the high cost low effectiveness scenario holds for lifetime spending estimates and also occurs for the mid cost scenario. The net impact of the program on lifetime Medicare costs more than triples in the mid cost scenario if the treatment response at one year is maintained. 29 Figure 2.1. Sensitivity Analysis for Collaborative Care Program Costs and Effectiveness on Spending (2010 USD) a) Total medical spending 30 b) Medicare spending 2.6 Discussion and Conclusion This study adds to the late-life depression and health services research literatures. The simulation model forecasts late-life depressive symptoms taking into account the complex relationship between depression and chronic conditions as well as other health and economic characteristics. The model is then used to assess the population-level and long term impacts of 31 collaborative care treatment on a large number of health characteristics, life expectancy and QALYs, and total medical and Medicare spending in the 65+ population. Projected demographic changes that would be expected to increase or decrease the prevalence of depressive symptoms – a larger Hispanic population or higher levels of education – offset each other. The prevalence of clinically significant depressive symptoms in the 65+ community-dwelling population is steady at approximately 14 percent. On average, individuals suffer from two waves of depressive symptoms after the age of 65. Therefore, the number of individuals affected by late-life depressive symptoms over their lifetimes is much higher than the annual prevalence. By incorporating depressive symptoms in the FEM, the impact of collaborative care treatment could be monitored in a large number of health and economic outcomes including QALYs. Collaborative care treatment of depressive symptoms with the efficacy rate of the IMACT RCT (33.85% treatment response at 24-months) would decrease the population prevalence of clinically significant depressive symptoms to approximately 11.5%. Other health benefits such as incidence of chronic conditions, reduced behavioral risks, and reductions in the consequences of chronic conditions would be small. However when considered together in the QALY measure, the improvement is approximately 15 days over the course of the remaining lifetime after age 65. The total gain in life expectancy would be about the same meaning individuals successfully treated with collaborative care would be regaining healthy life. At the population level, collaborative care treatment of depressive symptoms could save Medicare $2 billion annually after $3 billion of increased costs due to the collaborative care program. Over the long term, lifetime total medical spending may increase due in part to increased life expectancy. From the perspective of Medicare, the collaborative care program pays 32 for itself with approximately $1,000 in program costs over an individual’s lifetime after age 65 but with greater cost savings from the reduction in depressive symptoms. A sensitivity analysis shows that cost savings to Medicare could be achieved at any level of effectiveness if the program cost less than the IMPACT estimates. Alternatively, savings would also be possible if the 12-month effectiveness was maintained even at the higher cost estimates. Prevention of clinically significant depressive symptoms in the collaborative care eligible population would decrease the prevalence to 1.2% among 65+ community dwelling individuals with annual savings to Medicare of approximately $26 billion. Although complete prevention of clinically significant depressive symptoms is unrealistic, the scenario offers an estimate of the additional Medicare costs due to under treated depressive symptoms. There are several data limitations to consider. First, the outcomes of the IMPACT study were not reported by age, sex, race/ethnicity, or other characteristics, so interpretation of results for subgroups of the population is difficult. The IMPACT program design has been studied in comorbid (Katon et al. 2006) and minority populations (Gilmer et al. 2008) with similar levels of effectiveness. The range of treatment effects and costs are in line with other collaborative care treatments for depression (Jacob et al. 2012; Thota et al. 2012). Future work should examine the effect of collaborative care programs on subgroups disproportionately affected by depression or depressive symptoms. Second, MDD cannot be evaluated for all individuals in the HRS prior to wave 9. The Composite International Diagnostic Interview Short Form (CIDI-SF), an instrument validated for MDD diagnosis, was only used at entry to the survey prior to 2008. Because the CIDI-SF will be included for all patients in every wave, there will be the potential to use a more complicated definition of depression and treatment effectiveness in the future. 33 Third, the MEPS survey does not include a depression-specific questionnaire, so mental health is not explicitly included in the calculation of quality of life. Health Utilities Index (HUI) scores are being developed for the HRS and could be used in future work. Another limitation from the modeling method is the forced assumption about impact of depressive symptoms on functional status. This assumption was empirically tested by comparing the coefficients on the interaction terms with and without the non-interacted clinically significant depressive symptoms variable. Clinically significant depressive symptoms are significant predictors of all four consequences of chronic conditions variables. However, the marginal effect of clinically significant depressive symptoms on incidence is small (ADL ME=1pp, IADL ME=0.9pp, pain status ME=3.6pp, cognitive impairment ME=1.6pp). The conclusions of the analysis do not change if the assumption is relaxed. Depressive symptoms will continue to affect approximately 14% of the US 65+ population. The recent addition of collaborative care to the Medicare fee schedule for primary care physicians may help to more effectively treat comorbid depression. Modeling depressive symptoms in the population allows us to examine the consequences of more effective but also more costly treatment programs before administrative data is available. Recommended treatment methods such as collaborative care, if implemented with similar effectiveness and cost achieved in clinical trials, could be cost neutral to Medicare while also increasing QALYs by half a month. The ideal scenario would be the prevention of late-life depressive symptoms with potential annual benefits to Medicare of $26 billion. 2.7 References Administration on Aging. 2012. "A Profile of Older Americans: 2012." U.S. Department of Health and Human Services. 34 American Psychiatric Association. 2013a. "American Psychiatric Association Summary Final Rule for the 2014 Medicare Physician Fee Schedule." —. 2013b. Diagnostic and statistical manual of mental disorders (5th ed.). Arlington, VA: American Psychatric Publishing. Ayuso-Mateos, J.L., R. Nuevo, E. Verdes, N. Naidoo, and S. Chatterji. 2010. "From depressive symptoms to depressive disorders: the relevance of thresholds." The British Journal of Psychiatry 196:365-371. Bambauer, K.Z., D.G. Safran, D. Ross-Degnan, F. Zhang, A.S. Adams, J. Gurwitz, M. Pierre- Jacques, and S.B. Soumerai. 2007. "Depression and Cost-Related Medication Nonadherence in Medicare Beneficiaries." Arch Gen Psychiatry 64(602-608). Barry, L.C., J.J. Abou, A.A. Simen, and T.M. Gill. 2012. "Under-treatment of depression in older persons." Journal of Affective Disorders 136:789-796. Blazer, D.G. 2003. "Depression in Late Life: Review and Commentary." Journal of Gerentology 58A(3):249-265. Blazer, D.G., D.C. Steffens, H.G. Koenig, and (eds). 2009. "Chapter 15. Mood Disorders." in The American Psychiatric Publishing Textbook of Geriatric Psychiatry, 4th Edition Bock, B.C., K.S. Hudmon, J. Christian, A.L. Graham, and F.R. Bock. 2010. "A tailored intervention to support pharmacy-based counseling for smoking cessation." Nicotine & Tobacco Research 12(3):217-225. Bruce, M.L. 2001. "Depression and Disability in Late Life Directions for Future Research." Am J Geriatr Psychiatry 9:102-112. Byers, A.L., K. Yaffe, K.E. Covinsky, M.B. Friedman, and M.L. Bruce. 2010. "High Occurrence of Mood and Anxiety Disorders Among Older Adults." Arch Gen Psychiatry 67(5):489-496. Carrière, I., L.A. Gutierrez, K. Pérès, C. Berr, P. Barberger-Gateau, K. Ritchie, and M.L. Ancelin. 2011. "Late life depression and incident activity limitations: Influence of gender and symptom severity." Journal of Affective Disorders 133:42-50. Centers for Medicare and Medicaid Services. 2014. "Medicare Program; Revisions to Payment Policies Under the Physician Fee Schedule, Clinical Laboratory Fee Schedule, Access to Identifiable Data for the Center for Medicare and Medicaid Innovation Models & Other Revisions to Part B for CY 2015." Federal Register. Ciechanowski, P., W.J. Caton, and J.E. Russo. 2000. "Depression and Diabetes Impact of Depressive Symptoms on Adhrence, Function, and Costs." Arch Intern Med 160:3278-3285. Congressional Budget Office. 2014. "The 2014 Long-Term Budget Outlook." 35 Crimmins, E.M.and H. Beltrán-Sánchez. 2011. "Mortality and morbidity trends: is there compression of morbidity?" The Journals of Gerontology Series B: Psychological Sciences and Social Sciences 66(1):75-86. Donohue, J.M.and H.A. Pincus. 2007. "Reducing the Societal Burden of Depression A Review of Economic Costs, Quality of Care and Effects of Treatment." Pharmacoeconomics 25(1):7-24. Dotson, V.M., M.A. Beydoun, and A.B. Zonderman. 2010. "Recurrent depressive symptoms and the incidence of dementia and mild cognitive impairment." Neurology 75:27-34. Egede, L.E.and C. Ellis. 2010. "Diabetes and depression: Global perspectives." Diabetes Research and Clinical Practice 87:302-312. Freid, V.M., A.B. Bernstein, and M.A. Bush. 2012. "Multiple Chronic Conditions Among Adults Aged 45 and Over: Trends Over the Past 10 Years." edited by CDC: NCHS Data Brief. Fuentes, D.and M.P. Aranda. 2012. "Depression Interventions Among Racial and Ethnic Minority Older Adults: A Systematic Review Across 20 Years." Am J Geriatr Psychiatry 20:915–931. Gayman, M.D., R.J. Turner, and M. Cui. 2008. "Physical Limitations and Depressive Symptoms: Exploring the Nature of the Association." Journal of Gerontology: SOCIAL SCIENCES 63B(4):S219-S228. Gilbody, S., P. Bower, J. Fletcher, D. Richards, and A.J. Sutton. 2006. "Collaborative Care for Depression A Cumulative Meta-analysis and Review of Longer-term Outcomes." Arch Intern Med. 166:2314-2321. Gilmer, T.P., C. Walker, E.D. Johnson, A. Philis-Tsimikas, and J. Unützer. 2008. "Improving Treatment of Depression Among Latinos With Diabetes Using Project Dulce and IMPACT." Diabetes Care 31:1324-1326. Gold, M.R., J.E. Siegel, L.B. Russell, and M.C. Weinstein. 1996. Cost-effectiveness in health and medicine: Oxford University Press, USA. Goldman, D.P., D. Cutler, J.W. Rowe, P.-C. Michaud, J. Sullivan, D. Peneva, and S.J. Olshansky. 2013. "Substantial Health And Economic Returns From Delayed Aging May Warrant A New Focus For Medical Research." Health Affairs 32(10):1698-1705. González, H.M., W.A. Vega, D.R. Williams, W. Tarraf, B.T. West, and H.W. Neighbors. 2010. "Depression Care in the United States Too Little for Too Few." Arch Gen Psychiatry 67(1):37- 46. Heeramun-Aubeeluck, A., Z. Lu, and Y. Luo. 2012. "Comorbidity of Depression and Diabetes: In a Nutshell." Psychology 3(Special Issue):787-794. 36 Hilderink, P.H., H. Burger, D.J. Deeg, A.T. Beekman, and R.C.O. Voshaar. 2012. "The Temporal Relation Between Pain and Depression: Results From the Longitudinal Aging Study Amsterdam." Psychosomatic Medicine 74:945-951. Huang, Y., X. Wei, T. Wu, R. Chen, and A. Guo. 2013. "Collaborative care for patients with depression and diabetes mellitus: a systematic review and meta-analysis." BMC Psychiatry 13:260. Hunkeler, E.M., W. Katon, L. Tang, J.W.W. Jr, K. Kroenke, E.H.B. Lin, L.H. Harpole, P. Arean, S. Levine, L.M. Grypma, W.A. Hargreaves, and J.r. Unützer. 2006. "Long term outcomes from the IMPACT randomised trial for depressed elderly patients in primary care." BMJ. Jacob, V., S.K. Chattopadhyay, T.A. Sipe, A.B. Thota, G.J. Byard, D.P. Chapman, and C.P.S.T. Force. 2012. "Economics of Collaborative Care for Management of Depressive Disorders A Community Guide Systematic Review." Am J Prev Med 42(5):539-549. Jorm, A.F. 2001. "History of depression as a risk factor for dementia: an updated review." Australian and New Zealand Journal of Psychiatry 35:776-781. Karsten, J., C.A. Hartman, J.H. Smit, F.G. Zitman, A.T.F. Beekman, P. Cuijpers, A.J.W.v.d. Does, J. Ormel, W.A. Nolen, and B.W.J.H. Penninx. 2011. "Psychiatric history and subthreshold symptoms as predictors of the occurrence of depressive or anxiety disorder within 2 years." The British Journal of Psychiatry 198:206-212. Katon, W. 2003. "Clinical and Health Services Relationships between Major Depression, Depressive Symptoms, and General Medical Illness." Biological Psychiatry 54:216-226. —. 2011. "Epidemiology and treatment of depression in patients with chronic medical illness." Dialogues Clin Neurosci 13:7-23. Katon, W.and P. Ciechanowski. 2002. "Impact of major depression on chronic medical illness." Journal of Psychosomatic Research 53:859-863. Katon, W., J. Unutzer, K. Wells, and L. Jones. 2010. "Collaborative depression care: history, evolution and ways to enhance dissemination and sustainability." General Hospital Psychiatry 32:456-464. Katon, W., J.r. Unützer, F. Ming-Yu, J.W. Williams, Jr, M. Schoenbaum, E.H.B. Lin, and E.M. Hunkeler. 2006. "Cost-Effectiveness and Net Benefit of Enhanced Treatment of Depression for Older Adults with Diabetes and Depression." Diabetes Care 29:265-270. Katon, W.J., M. Schoenbaum, M.-Y. Fan, C.M. Callahan, J. Williams, E. Hunkeler, L. Harpole, X.-H.A. Zhou, C. Langston, and J.r. Unützer. 2005. "Cost-effectiveness of Improving Primary Care Treatment of Late-Life Depression." Arch Gen Psychiatry 62:1313-1320. Lakdawalla, D.N., J. Bhattacharya, and D.P. Goldman. 2004. "Are the young becoming more disabled?" Health Affairs 23(1):168-176. 37 Lyness, J.M., E.D. Caine, D.A. King, Y. Conwell, P.R. Duberstein, and C. Cox. 2002. "Depressive Disorders and Symptoms in Older Primary Care Patients One-Year Outcomes." Am J Geriatr Psychiatry 10:275-282. Lyness, J.M., Q. Yu, W. Tang, X. Tu, and Y. Conwell. 2009. "Risks for Depression Onset in Primary Care Elderly Patients: Potential Targets for Preventative Interventions." The American Journal of Psychiatry 166:1375-1383. Meeks, T.W., I.V. Vahia, H. Lavretsky, G. Kulkarni, and D.V. Jeste. 2011. "A tune in “a minor” can “b major”: A review of epidemiology, illness course, and public health implications of subthreshold depression in older adults." Journal of Affective Disorders 129:126-142. Nutting, P.A., K. Rost, M. Dickinson, J.J. Werner, P. Dickinson, J.L. Smith, and B. Gallovic. 2002. "Barriers to Initiating Depression Treatment in Primary Care Practice." J Gen Intern Med 17:103-111. O’Donnell, R.i.n.M.and A.W. Kaszniak. 2011. "Charting Late-Life Affective Disorders." Generations 35(2):46-57. Palomino, R.A., P.M. Nicassio, M.A. Greenberg, and E.P. Medina, Jr. 2007. "Helplessness and loss as mediators between pain and depressive symptoms in fibromyalgia." Pain 129:185-194. Radloff, L.S. 1977. "The CES-D Scale: A Self-Report Depression Scale for Research in the General Population." Applied Psychological Measurement 1(3):385-401. Roth, D.L., W.E. Haley, M. Hovater, M. Perkins, V.G. Wadley, and S. Judd. 2013. "Family Caregiving and All-Cause Mortality: Findings from a Population-based Propensity-matched Analysis." American Journal of Epidemiology:1-8. Saczynski, J.S., A. Beiser, S. Seshadri, S. Auerbach, P.A. Wolf, and R. Au. 2010. "Depressive symptoms and risk of dementia The Framingham Heart Study." Neurology 75:35-41. Saint Onge, J.M., P.M. Krueger, and R.G. Rogers. 2014. "The Relationship Between Major Depression and Nonsuicide Mortality for U.S. Adults: The Importance of Health Behaviors." Journals of Gerontology, Series B: Psychological Sciences and Social Sciences. Seegert, L. 2013. "Clock is ticking for commission charged with addressing comprehensive long- term care." Association of Health Care Journalists. Services, C.f.M.a.M. 2012. "Chronic Conditions among Medicare Beneficiaries, Chartbook, 2012 Edition." Baltimore, MD. Solway, E., C.L. Estes, S. Goldberg, and J. Berry. 2010. "Access Barriers to Mental Health Services for Older Adults from Diverse Populations: Perspectives of Leaders in Mental Health and Aging." Journal of Aging & Social Policy 22:360-378. 38 Steffick, D.E. 2000. "Documentation of Affective Functioning Measures in the Health and Retirement Study." in HRS/AHEAD Documentation Report. Ann Arbor, MI: Survey Research Center University of Michigan. Steinman, L.E.e.a. 2007. "Recommendations for Treating Depression in Community-Based Older Adults." Am J Prev Med 33(3):175-181. Thota, A.B., T.A. Sipe, G.J. Byard, C.S. Zometa, R.A. Hahn, L.R. McKnight-Eily, D.P. Chapman, A.F. Abraido-Lanza, J.L. Pearson, C.W. Anderson, A.J. Gelenberg, K.D. Hennessy, F.F. Duffy, M.E. Vernon-Smiley, D.E.N. Jr., S.P. Williams, and C.P.S.T. Force. 2012. "Collaborative Care to Improve the Management of Depressive Disorders A Community Guide Systematic Review and Meta-Analysis." Am J Prev Med 42(5):525-538. Unützer, J., W. Katon, C.M. Callahan, J.W. Williams, Jr, E. Hunkeler, L. Harpole, M. Hoffing, R.D.D. Penna, P.H. Noel, E.H.B. Lin, P.A. Arean, M.T. Hagel, L. Tang, T.R. Belin, S. Oishi, and C. Langston. 2002. "Collaborative Care Management of Late-Life Depression in the Primary Care Setting A Randomized Controlled Trial." JAMA 288:2836-2845. Unützer, J.r., W.J. Katon, M.-Y. Fan, M.C. Schoenbaum, E.H.B. Lin, R.D.D. Penna, and D. Powers. 2008. "Long-term Cost Effects of Collaborative Care for Late-life Depression." Am J Manag Care 14(2):95. Wallace, R.B.and A.R. Herzog. 1995. "Overview of the Health Measures in the Health and Retirement Study." The Journal of Human Resources 30(Special Issue of the Health and Retirement Study: Data Quality and Early Results):S84-S107. Wells, K.B., M. Schoenbaum, N. Duan, J. Miranda, L. Tang, and C. Sherbourne. 2007. "Cost- Effectiveness of Quality Improvement Programs for Patients With Subthreshold Depression or Depressive Disorder." Psychiatric Services 58:1269-1278. Whooley, M.A., P.d. Jonge, E. Vittinghoff, C. Otte, R. Moos, R.M. Carney, S. Ali, S. Dowray, B. Na, M.D. Feldman, N.B. Schiller, and W.S. Browner. 2008. "Depressive Symptoms, Health Behaviors, and Risk of Cardiovascular Events in Patients With Coronary Heart Disease." JAMA 300(20):2379-2388. Young, P.L., R.S. Saunders, L. Olsen, and (eds.). 2010. "The Healthcare Imperative: Lowering Costs and Improving Outcomes: Workshop Series Summary." Washington, DC: Institute of Medicine. The National Academies Press. Zivin, K., D.J. Llewellyn, I.A. Lang, S. Vijan, M.U. Kabeto, E.M. Miller, and K.M. Langa. 2010. "Depression among older adults in the United States and England." Am J Geriatr Psychiatry 18(11):1036-1044. Zivin, K., J.M. Madden, A.J. Graves, F. Zhang, and S.B. Soumerai. 2009. "Cost-related medication nonadherence among beneficiaries with depression following Medicare Part D." Am J Geriatr Psychiatry 17(12):1068-1076. 39 CHAPTER 3. IMPLICATIONS OF SOCIAL SECURITY ENTITLMENT AGE POLICY CHANGE ON THE PRODUCTIVE ACTIVITIES OF OLDER ADULTS (Co-authored with Julie Zissimopoulos, PhD, Dana Goldman, PhD, and John Rowe, MD) 3.1 Abstract An aging America presents challenges but also brings social and economic capital. We quantify public revenues from and expenditures on Americans ages 65 and older and the value of their unpaid productive activities and private transfers to family members. We project the value of these activities and government’s revenue and expenditure under different scenarios of policy change to OASI eligibility age through 2050 using dynamic micro simulation. We find productive activities and private gifts of Americans ages 65+ are $722 billion in 2010 while net (of tax revenues) spending on them is $984 billion. A 5-year delay in the full retirement age leads to 10% lower federal spending while a 2-year delay in early entitlement age results in 1.5% more spending. The effect of a 5-year delay on unpaid activities and transfers is small: a $4 billion decrease in services and a $4.5 billion increase in bequests and monetary gifts. 3.2 Introduction Americans are living longer than at any other time in history (Administration on Aging 2013). This is a public health success story. The boom in the population aged 65 years and older has, and will continue to have, major consequences for Americans of all ages and for the social and economic institutions of the US (Congressional Budget Office 2014). While an aging America presents economic and social challenges (Zissimopoulos et al. forthcoming), this boom 40 in older Americans also brings large-scale social and economic capital that has the potential to benefit everyone. Older Americans today are maintaining good health until later in life than earlier birth cohorts. With good health, work and life experiences, older Americans are capable of making economic and social contributions that benefit themselves, their families, and their communities (Johnson 2005). These contributions may include, for example, paid work, volunteering, education, fitness and exercise, leisure and travel, advocacy and political action, and consumerism. The consequences of the longevity revolution and population aging however, are most often framed in terms of rising costs of Medicare and Social Security (The Board of Trustees Federal Old-Age and Survivors Insurance and Federal Disability Insurance Trust Funds 2014). Threats to the solvency of the Social Security Trust fund was brought on by a much larger number than expected number of beneficiaries (the baby boom generation), a larger than anticipated increase in life expectancy, and a political reluctance to increase payroll taxes. The consequences are also negatively framed in terms of the old verses the young. Median wealth of U.S. households headed by a person age 65 or older is about 47 times greater than a household headed by someone under 35 according to data from the Census (Taylor et al. 2011). People typically accumulate assets as they age. However, this wealth gap is more than double what it was in 2005 and nearly five times the 10-to-1 inflation-adjusted disparity a quarter-century ago. One policy tool at the center of the discussion on how to lower Social Security expenditures is a change in the age at which individuals receive benefits thereby reducing the total number of years of benefit receipt and thus lifetime benefit amounts (Social Security Administration 2013). These changes have implications for the financial security of beneficiaries and their families – 41 today two-thirds of beneficiaries rely on Social Security for more than half of their total income, and 25% rely on Social Security for over 90% of their income (Social Security Administration 2014). Policy changes that increase the age at which individuals become eligible for Old Age and Survivors Insurance (OASI) benefits may however increase the amount of time individuals remain in the labor force and thus their private savings and improve their overall financial security. Social Security has evolved extensively since its inception. Most recently, amendments approved in 1983 authorized gradual increases in the age of full eligibility for workers born after 1937 – with provisions fully effective for all workers born after 1959 (Svahn and Ross 1983). This amendment gradually increased the age of eligibility for full Social Security benefits from 65 to 67 (hereinafter full retirement age or FRA) and lowered the benefits for those who choose to receive them at age 62 (hereinafter earliest eligibility age or EEA). The potential of policy change to influence work may differ depending on whether the change is to the EEA or FRA. Approximately 72 percent of new beneficiaries draw benefits before the full retirement age, and 46 percent draw benefits at the earliest eligibility age of 62 (Kingson and Morrissey 2012). The effect of policy change to the FRA on work may be blunted by this fact. Changes to the age of eligibility for benefits may also have other impacts on individuals, families and society that are often neglected in the policy discussions and are less well- understood. Longer work lives induced by changes in benefit age may reduce individuals’ engagement in other productive activities such as volunteering and caring for grandchildren or disabled parents or spouses (Furstenberg et al. forthcoming). According to the Bureau of Labor Statistics, in 2011, 16 percent of the 65+ population provided unpaid care to elderly individuals (Bureau of Labor Statistics 2012). While more work may increase government revenues through 42 taxes, less caregiving by family members may induce increased government spending on care as individuals substitute family care with publicly financed care through Medicaid or Medicare. Longer work lives may lower volunteering which may have negative implication for an individual’s well-being. Volunteering later in life is associated with health benefits, delayed physical disability, enhanced cognition, and lower mortality (Greenfield and Marks 2004; Konrath et al. 2012; Lum and Lightfoot 2005; Van Willigen 2000). Increases in physical activity, cognitive engagement, and social interactions are likely important mechanisms through which these benefits are achieved (Carlson et al. 2008). Organizations, such as Experience Corps, encourage older adults to engage with children at schools, and randomized experiments of older adults into Experience Corps has been shown to increase strength and cognitive ability of volunteers (Freedman et al. 2004; Fried et al. 2004; Rebok et al. 2014). Moreover, this policy focus on the implications of an aging America on public expenditures overlooks other important private contributions by older Americans such as the gift of money from older individuals to their children and grandchildren and the transfer of money to their family members through bequests at death. Indeed, these private transfers by older parents to their children may help finance post-secondary education, reduce liquidity constraints in homeownership and smooth consumption during periods of unemployment (Zissimopoulos and Smith 2011). This study analyzes the public and private transfers to and from older Americans and the implications of policy changes to OASI age of eligibility on these transfers. Using multiple waves of the Health and Retirement study, we estimate models of both paid work and unpaid productive activities such as volunteering and caregiving, and direct financial transfers of the 65 and older population. We also estimate models of take-up of government programs and 43 calculate expenditures on the population 65+ as well as state and federal tax revenue received from older adults. Utilizing health, demographic and economic trends of younger cohorts, we project the value of these activities as well as government’s revenue and expenditure through 2050 using dynamic micro simulation. We then explore the implications of policy change to Social Security’s full and early eligibility ages on private transfers, unpaid productive activities and government revenue and expenditures. 3.3 Study Data and Methods To estimate the potential benefits and costs of Social Security policy changes, we used the Future Elderly Model (FEM), a population-based microsimulation that projects the health and economic outcomes for middle-aged and older adults. The FEM has previously been used to examine the impact of new medical technologies (Goldman et al. 2005), changes in disability (Chernew et al. 2005), improved prevention of diseases (Goldman et al. 2009), the benefits and costs of delayed aging (Goldman et al. 2013) and the value of the delay in the onset of Alzheimer’s disease (Zissimopoulos, Crimmins and St. Clair 2014). Additional detail for the models and methods described briefly here is available in the Technical Appendix. The FEM uses data from a nationally representative sample based on the Health and Retirement Survey (HRS), a biennial survey of Americans age fifty-one or older that began in 1992 (Wallace and Herzog 1995). Multivariate models of two-year changes in health conditions, functional status and economic outcomes such as income, wealth, and work take as inputs health risk factors, race, marital status, education and prior wave health and economic states. Receipt of Social Security and other public benefits as well as taxes paid are also modeled. Furthermore, we estimate multivariate models of monetary gifts made to children and other productive activities such as 44 caregiving and volunteering over two-year intervals. Further details on the empirical models are given below. In the FEM simulation, a new cohort of 50 year olds enters the model each year with health, socio-economic status and demographic characteristics reflecting trends observed in younger populations based on data from the National Health Interview Survey and the Current Population Survey (Crimmins and Beltrán-Sánchez 2011; Lakdawalla, Bhattacharya and Goldman 2004). Increases in life expectancy follow the same assumptions as Social Security Administration projections (The Board of Trustees Federal Old-Age and Survivors Insurance and Federal Disability Insurance Trust Funds 2013). We simulate the effect of changes to the EEA and FRA on public revenues and expenditures, private transfers and unpaid productive activities. Details on the scenarios are given below. 3.3.1 Health and Disability Health conditions, functional status, and mortality are derived directly from HRS questions. The health conditions include ever having been diagnosed with heart disease; any cancer except skin cancer; chronic bronchitis or emphysema; diabetes; hypertension; and stroke or transient ischemic attack. The functional status measures include difficulty performing activities of daily living (ADLs); difficulty performing instrumental activities of daily living (IADLs); and nursing home residency. Both functional status and the likelihood of developing a health condition are modeled based on a person’s demographic characteristics, economic status, other health conditions, functional status, marital status, and health characteristics at age 50. Health characteristics at age 50 control for baseline unobserved factors and allow estimated transition probabilities to account for the decreasing health of future cohorts. 45 3.3.2 Government Expenditures Government expenditures are derived from the costs of major entitlement programs— specifically, federal and state spending for Medicare and Medicaid, and federal income support through Old-Age & Survivors Insurance (OASI), Disability Insurance (DI) and Supplemental Security Income (SSI). We first modeled participation in each program and then used age, marital status, and other economic outcomes to calculate benefits based on the rules for the particular programs. Annual costs are given in constant 2010 dollars. All cumulative costs are discounted using a 3 percent annual discount rate (Gold et al. 1996). Individual medical spending and insurance enrollment is estimated from the Medicare Current Beneficiary Survey (MCBS) for Medicare eligible individuals and from the Medical Expenditure Panel Survey (MEPS) for the non-Medicare eligible individuals. Individuals are automatically enrolled in inpatient benefits when they become eligible for Medicare. Enrollment is modeled separately using MCBS for ambulatory and prescription drug coverage. The Medicare spending models are also estimated separately for inpatient, ambulatory, and prescription costs. Each of the Medicare enrollment and spending models is based on a person’s demographic characteristics, health risk factors, functional status, and marital status. Social Security is available at the full retirement age or at the earliest eligibility age. Individuals claiming disability must switch to Social Security benefits after reaching FRA. The models of Social Security claiming are based on a person’s demographic characteristics, economic status, health risk factors, functional status, and marital status. The Social Security monthly benefit is calculated when an individual starts claiming and is based on their lifetime 46 earnings history available through linked SSA restricted earnings data and following SSA’s benefit calculation algorithm. Other government benefits include veteran’s benefits, welfare, and food stamps. The HRS imputes other government benefits if continuous values are not reported. We model other government benefits based on person’s demographic characteristics, economic status, health risk factors, functional status, and marital status. 3.3.3 Taxes We estimate federal, state, and local tax liabilities paid by older adults. Tax models are developed for federal and state income taxes and property tax. Federal and state income taxes are first estimated for HRS respondents using the online TAXSIM software (NBER, Cambridge, MA, Version 8). TAXSIM was developed to estimate the US federal and state income tax liabilities and marginal tax rates based on individual taxpayer characteristics (Feenberg and Coutts 1993). We report results for Michigan, which is assumed to be representative of average tax levels for all states. Federal and state income taxes are then modeled based on earnings, capital income, wealth, claiming of government benefits, and marital status. HRS respondents report property tax payments for primary and secondary residences either as a continuous variable or within a bracketed range. When a range is reported, the continuous property tax variable is imputed based on respondent characteristics. Property tax is modeled using two waves of the HRS based on a person’s demographic characteristics, economic status, health risk factors, functional status, and marital status. Payroll taxes – specifically, Social Security and Medicare contributions – and city taxes are calculated as a percent of estimated earnings. Social Security payroll taxes are set to 6.2% of earnings up to $97,500, and Medicare payroll taxes are 47 1.45% of all earnings. City tax is 2.55% of net income. Estate and gift taxes paid by recipients of bequests and inter vivos gifts are not considered. 3.3.4 Bequests and Inter Vivos Gifts Bequests are calculated as the financial and housing wealth remaining at death or, for couples, at death of the last living spouse. Wealth is modeled based on a person’s demographic characteristics, economic status, health risk factors, functional status, marital status, and number of children. Additionally, year indicators are included in the estimation model to control for economy-wide effects during the estimation period. Inter vivos transfers are reported for financial transfers from parents to children or to grandchildren. In some cases, inter vivos transfers are reported by HRS respondents using a bracketed range rather than as a continuous variable. In those cases, the transfer value is imputed as a continuous variable from the bracketed ranges. We model inter vivos transfers as a continuous variable based on a person’s demographic characteristics, economic status, health risk factors, functional status, marital status, and number of children. 3.3.5 Caregiving and Volunteering Respondents in the HRS report the annual hours of care they provide to their spouses, grandchildren or parents. To determine the number of childcare hours provided, respondents were asked if they live with their grandchildren or not and how many hours of care they provided to both co-resident and non-coresident grandchildren. We imputed missing number of hours for caregivers based on the mean number of hours of sex-matched individuals with co-resident grandchildren in addition to the basic characteristics of the respondent. We model informal hours 48 of care based on a person’s demographic characteristics, economic status, health risk factors, functional status, marital status (except for care to spouse), location of residence, and proximity to children. Respondents report annual hours spent volunteering for educational, charitable or health- related organizations in HRS waves 4 to 6. Wave 7 forward, a range of annual hours spent volunteering is reported at <50, 50-100, 100-200, and 200 hours. For waves in which only a range is reported conditional on volunteering, the hours spent volunteering are imputed based on wave 6 continuous values matched on demographic health, economic, religious, and location characteristics. We model volunteer hours based on a person’s demographic characteristics, economic status, health risk factors, functional status, marital status, location of residence, proximity to children, religious affiliation, and religiosity. We value hours of caregiving and volunteering in 2010 dollars based on three wage rates. The replacement rate value for highly skilled workers, or what a family member would spend on caregiving services if a paid worker replaced the informal care or an organization would spend to replace a volunteer, is used as a high valuation. The minimum wage is used as the low value for caregiving services and volunteering (Arno, Levine and Memmott 1999; Johnson 2005). The value of caregiving to a spouse or parent is given a moderate value equal to the average wage reported by the BLS for home health aids ($10.46 per hour), and the maximum value is taken from a survey of hourly rates of caregiver agencies ($20.00 per hour). The middle value for childcare is the national average wage for childcare workers ($10.15 per hour), and the maximum is the average wage for skilled preschool workers ($14.04 per hour) (Bureau of Labor Statistics 2010). We set the moderate value for volunteering to the national average wage for clerical office workers ($13.58 per hour), and the high value is the average wage for social 49 service workers ($20.76 per hour). The 2010 federal minimum wage was $7.25 (Bureau of Labor Statistics 2011). We report results based on a moderate value. Results based on high and low values are included as a sensitivity analysis. We assume wages grow by three percent each year. 3.3.6 Scenarios We developed four scenarios, the first one representing the status quo, or baseline and three others representing different changes to the EEA and FRA of eligibility for Social Security benefits. We compared the public expenditures and revenues each would involve and the value of private financial transfers and unpaid productivities each would involve. For each scenario we conducted the simulation one hundred times and averaged the outcomes. The second scenario delays only the FRA and we focus on a delay of 3 or 5 years although results are available upon request for a one-year delay. The third scenario delays only the EEA by 2-years (similarly results from both fewer and more years of EEA delay are available). The EEA delay is stepped so that number of years between EEA and FRA is never less than 2. The forth scenario delays both the EEA and FRA by 3-years. We assumed that 10% to 30% of workers shift to DI benefits with an increase in the EEA. The results shown in the paper use the mid-range assumption of 20% of workers shift to DI with an increase in the EEA. 3.3.7 Limitations There are limitations to our approach. First, various Social Security policy changes are tested using simulation, which includes assumptions about the steadiness of underlying parameters used to model health and economic outcomes. Most important, the policy change scenarios are intended to shed light on the implications of changing the retirement age on both public 50 expenditures and revenues as well as the impact on private transfer and other unpaid productivities activities. More study would be required before the implementation of any policy change — including an official scoring of such a change by the Congressional Budget Office, fuller consideration of distributional and health outcomes beyond the major entitlement programs, consideration of the impact of any financing reforms such as those in the Affordable Care Act and the impact on economic growth of a larger and older labor force. For example, the scenarios modeled do not account for potential increases in Medicaid expenditures if it is a substitute for fewer services provide by family members. Moreover, we do not account for potential future changes in the structure of the family or household arrangements. Our focus on Americans ages 65 and older does not account directly for increased work for individuals younger than 65 brought about by increases in EEA that we show in Figure 3.1. Figure 3.1 Percent Change from Status Quo of Work and Claiming in 2010 -‐70.00% -‐60.00% -‐50.00% -‐40.00% -‐30.00% -‐20.00% -‐10.00% 0.00% 10.00% 20.00% 62 63 64 65 66 67 68 69 70 71 72 73 74 75 Percent change from status quo 5-‐years delay in FRA DI Claim SS Claim Work 51 Figure 3.1 Percent Change from Status Quo of Work and Claiming in 2010 (continued) 3.4 Results Table 3.1 shows estimated public expenditures for Medicare, OASI, Medicaid, Disability Insurance, Supplemental Security Income and other programs (e.g. veteran’s benefits, welfare, and food stamps) benefiting Americans ages 65 and older in 2010. It also shows estimated revenues from federal, state, city, property, Social Security payroll and Medicare payroll taxes paid by Americans ages 65 and over in 2010. Total expenditures were $1,217 billion and total -‐70.00% -‐60.00% -‐50.00% -‐40.00% -‐30.00% -‐20.00% -‐10.00% 0.00% 10.00% 20.00% 62 63 64 65 66 67 68 69 70 71 72 73 74 75 Percent change from status quo 2-‐years delay in EEA DI Claim SS Claim Work -‐70.00% -‐60.00% -‐50.00% -‐40.00% -‐30.00% -‐20.00% -‐10.00% 0.00% 10.00% 20.00% 62 63 64 65 66 67 68 69 70 71 72 73 74 75 Percent change from status quo 3-‐years delay in EEA and FRA DI Claim SS Claim Work 52 revenues were $233 billion. Eighty-nine percent is spending on Medicare and OASI and 63% of revenues are derived from federal taxes. Table 3.1 Public Expenditures and Revenues, Americans Ages 65 and Over in 2010 (2010 Dollars, Billions) Public transfers 2010 USD Public benefits (billions) Medicare 461 OASI 616 Medicaid 97 SSI 9 Disability 1 Other 33 Total public benefits 1,217 Public revenues (billions) Federal tax 147 State tax 11 City tax 18 Property tax 42 SS tax 12 Medicare tax 3 Total public revenues 233 Table 3.2 shows significant private giving and unpaid productive activities by Americans ages 65 and older whose total value is 73 percent of public expenditures less revenues. In 2010, bequests total $475 billion and inter vivos gifts of money to children totaled $151 billion dollars. Older Americans also engaged in productive unpaid activities valued at $96 billion, composed of $27 billion of volunteer activities and $69 billion in caregiving services for a total of $722 billion dollars in 2010. Much of this (90%) directly benefited their children. 53 Table 3.2 Private Transfers and Unpaid Services, Americans Ages 65 and Over in 2010 (2010 Dollars, Billions) Private transfers 2010 USD Bequests 475 Inter vivos 151 Help to spouse 27 Care to parents 41 Care to grandkids 6 Total private transfers and unpaid services 722 Note: Value of unpaid services is $13.58 for volunteering, $10.46 for care to parents or spouse and $10.15 for care to grandchildren. Public expenditures almost quadruple between 2010 and 2050 rising from $1,217 to $4,525 while revenues and the population age 65 and older about double over this same time period (Table 3.3). Fifty-three percent of the dollar increase in expenditures is driven by increased Medicare expenditure. Private gifts of money and the value of unpaid productive activities double between 2010 and 2050, from $722 billion to $1,447 billion (Table 3.4). Sixty-four percent of the increase is attributable to rising wealth and gifts in the form of bequests. Table 3.3 Public Expenditures and Revenues 2010 to 2050 (2010 Dollars, Billions) Public transfers 2010 2020 2030 2040 2050 Population 65+ (millions) 43.76 57.44 71.91 77.04 79.98 Public benefits (billions) Medicare 461 651 1,125 1,684 2,223 OASI 616 871 1,178 1,380 1,541 Medicaid 97 133 233 419 661 SSI 9 11 13 15 19 Disability 1 1 4 4 5 Other 33 48 67 73 75 Total public benefits 1,217 1,715 2,620 3,576 4,525 54 Public revenues (billions) Federal tax 147 234 313 314 313 State tax 11 20 26 23 24 City tax 18 27 35 39 43 Property tax 42 61 80 83 82 SS tax 12 20 23 22 23 Medicare tax 3 5 5 5 5 Total public revenues 233 365 482 485 489 Table 3.4 Private Transfers and Unpaid Services, Americans Ages 65 and Over 2010-2050 (2010 Dollars, Billions) Private transfers 2010 2020 2030 2040 2050 Bequests 475.39 610.54 839.47 961.19 937.59 Inter vivos 150.85 212.14 275.88 304.64 310.33 Volunteer hours 26.81 36.57 45.69 45.38 43.71 Help to spouse 41.39 56.61 78.32 95.42 108.59 Care to parents 6.34 9.12 11.10 10.51 10.57 Care to grandkids 21.09 28.14 35.53 35.56 35.99 Total private transfers 722 953 1,286 1,453 1,447 Note: Value of unpaid services is $13.58 for volunteering, $10.46 for care to parents or spouse and $10.15 for care to grandchildren. Figure 3.1 shows the impact of policy changes to the EEA and FRA on the percent of individuals claiming DI, OASI benefits and doing paid work. Increasing the EEA by 2-years increases the percent working at ages 62 and 63 compared to the status quo by just 5%, increases DI claiming by about 4% and reduces OASI claiming by a much larger 34% at age 62 and 43% at age 63. There is a very small impact on OASI claiming at ages 64 and 65 compared to the status quo and no effect at older ages. In contrast, delaying the FRA by 5-years increases work at ages 62 to 65 by about 10% compared to the status quo, at ages 66 by 12% and there is a 55 sustained increase in working through age 75. Social Security claiming is reduced significantly compared to the status quo, particularly at ages 64 (43%) through 66 (65%) and even to age 68 (33%). Disability also increases through age 70 and the most compared to the status quo at age 67. Increasing both the EEA and FRA by 3-years has a smaller effect on the percentage working compared to the status quo than does a FRA increase of 5-years. For example work at age 66 or 67 increases by about 6%. OASI claiming is reduced at rates slightly smaller but somewhat comparable to changes in the FRA of 5-years and DI claiming increases at rates similar to those under the 5-year FRA delay scenario. The results in Tables 3.5 and 3.6 show how these changes in working for pay and claiming due to changes in the early and full eligibility ages translate into changes in public expenditures and revenues as well as monetary gifts to family members and unpaid productive activities. The scenarios reveal that increasing the FRA by 5-years produces the largest decrease in expenditures and increase in revenues in 2050 compared to the status quo (Table 3.5). Expenditures decrease by 10% ($4,525 to $4,059) and revenues increase by 19% ($489 to $581). Increasing the EEA without changing the FRA results in slightly higher expenditures as more individuals move to DI and remain on DI until the FRA without actuarially fair reductions to benefit amount. Increasing both the EEA and FRA by 3-years leads to similar overall changes as increasing FRA by 5-years, however the distribution of spending is different. Increasing FRA by 5-years results in OASI spending declines of 24% compared to status quo while expenditures on DI increase by 413% compared to status quo. In contrast, increasing the EEA and FRA by 3 years results in OASI spending declines of 14% compared to status quo while expenditures on DI increase by 295% compared to status quo. 56 Table 3.5 Effect of Delaying EEA and FRA on Public Expenditures and Tax Revenues (2010 Dollars, Billions) Scenario Year - Public Transfers Status Quo FRA 3-years FRA 5-years EEA 2-years EEA & FRA 3-years 2010 Total public revenues 233 246 254 234 246 Total public benefits 1,217 1,087 1,006 1,222 1,092 Medicare 461 454 449 462 457 OASI 616 489 410 620 490 Medicaid 97 97 97 97 97 SSI 9 9 9 9 9 Disability 1 6 10 1 7 Other 33 32 31 33 32 2030 Total public revenues 482 536 577 485 537 Total public benefits 2,620 2,382 2,232 2,679 2,422 Medicare 1,125 1,084 1,051 1,127 1,090 OASI 1,178 974 852 1,235 1,005 Medicaid 233 233 234 233 233 SSI 13 13 13 13 13 Disability 4 15 20 5 17 Other 67 63 61 67 63 2050 Total public revenues 489 541 581 492 542 Total public benefits 4,525 4,226 4,059 4,592 4,271 Medicare 2,223 2,160 2,110 2,226 2,171 OASI 1,541 1,294 1,172 1,605 1,327 Medicaid 661 662 663 661 663 SSI 19 19 19 19 19 Disability 5 19 28 6 21 Other 75 71 69 75 71 57 Table 3.6 Effect of Delaying EEA and FRA on Private Transfers (2010 Dollars, Billions) Scenario Year - Private Transfers Status Quo FRA 5-years EEA 2-years EEA & FRA 3-years 2010 Bequests 475 476 475 476 Inter vivos 151 151 151 151 Volunteer hours 27 27 27 27 Care 69 68 69 68 2030 Bequests 839 841 839 840 Inter vivos 276 279 276 278 Volunteer hours 46 45 46 45 Care 125 122 125 123 2050 Bequests 938 939 938 939 Inter vivos 310 313 310 312 Volunteer hours 44 43 44 43 Care 155 152 155 153 Changes in the FRA and EEA will increase the time spent in the labor force, increases income and wealth and thus may increase financial gifts to family. It will however, reduce the time available to volunteer and provide caregiving services. Table 3.5 shows the results of from our scenarios on private transfers of money and unpaid productive activities. A 5-year delay in FRA increases bequests and inter vivos transfers compared to the status quo by $4.5 billion. It also reduces the value of volunteer hours by $1 billion and caregiving services by $3 billion. Increasing both the EEA and FRA by 3-years has a smaller effect on private transfers and productive activities. It increases bequests and inter vivos transfers compared to the status quo by $2.6 billion. It also reduces the value of volunteer hours by $0.6 billion and caregiving services by $1.8 billion. 58 The value of unpaid productive activities varies based on the hourly wage attributable to the services. Table 3.7 presents the value of services for a low valuation at minimum wage and for a high valuation at the replacement wage for highly skilled workers. In 2010, the value of unpaid services contributes $62 billion to private transfers at minimum wage and $161 billion at highly skilled wages. The value of unpaid services accounts for 9% to 20% of private transfers depending on the replacement wage used for valuation. A 5-year delay in FRA reduces the value of unpaid services in 2050 by between $2.7 billion and $6.9 billion. Increasing both the EEA and FRA by 3-years reduces the value of unpaid services by between $1.6 billion and $4.1 billion. Table 3.7 High and Low Value of Unpaid Services, Americans Ages 65 and Over 2010-2050 (2010 Dollars, Billions) Scenario Year - Unpaid Services Status Quo FRA 5-years EEA 2-years EEA & FRA 3-years Low Valuation 2010 Volunteer hours 14.3 14.2 14.3 14.3 Care 48.1 47.7 48.1 47.9 2030 Volunteer hours 24.4 23.8 24.4 24.0 Care 87.4 85.3 87.3 86.1 2050 Volunteer hours 23.3 22.8 23.3 23.0 Care 108.3 106.2 108.3 107.0 High Valuation 2010 Volunteer hours 40.6 40.3 40.6 40.4 Care 120.4 119.2 120.4 119.7 2030 Volunteer hours 69.2 67.6 69.1 68.2 Care 220.1 215.0 220.0 217.1 59 2050 Volunteer hours 66.2 64.6 66.1 65.2 Care 277.6 272.3 277.6 274.5 Note: Low value of unpaid services is minimum wage ($7.25/hour) for volunteering, care to parents or spouse and care to grandchildren. High value of unpaid services is $20.56 for volunteering, $20.00 for care to parents or spouse and $14.04 for care to grandchildren. 3.5 Discussion and Conclusion Research studies and policy proposals addressing the implications of an aging society should include the economic and social contributions of Older Americans and the benefit of these contributions to themselves, their families, and their communities. A discussion of these contribution alongside that of the economic challenges to public entitlement programs that support older Americans will better inform policy. In this study we report our findings on these productive activities and direct monetary gifts to family of Americans ages 65 and older as well as public expenditures to this age group and revenues from them. We found in 2010 Americans ages 65 and older provided: $626 billion in inter vivos transfers and bequests, 2 billion hours of volunteering valued at $27 billion and 6.6 billion hours of caregiving to spouses, parents and grandchildren valued at $68 billion. The sum total of private financial transfers and productive activities was $722 billion. Much of this (90%) directly benefited their children. They also provided society $233 billion in tax revenues. At the same time, public expenditures on this age group in 2010 was $1.2 trillion. At a societal level, increases in the early entitlement age increased public expenditures. Both spending on DI and OASI increased. A 5-year increase in the full retirement age eligibility age over time lowered public expenditures: increases in DI spending were offset by lower OASI spending as individuals delayed claiming as a response to the delay. It also increased public 60 revenues as workers responded by remaining in the labor force longer. A five-year delay in FRA had an impact on the private transfers and the productivities activities of Americans ages 65 and older. The size, relative to the size of the impact on public revenues and expenditures, is relatively small. There was a $4.5 billion increase in bequests and monetary gifts and a $4 billion reduction in caregiving and volunteer services. The near parity of the switch from services to financial transfers is disrupted if the true value of unpaid services is closer to minimum wage ($2.7 billion) or the replacement rate for highly skilled labor ($6.9 billion). At the individual level, the increase in monetary gifts due to a delay in FRA may positively impact the welfare of their adult children through for example, financing post-secondary education of young adults, reducing liquidity constraints in homeownership and smoothing consumption during periods of unemployment or other shocks to income. It may also negatively impact their family through a lower provision of caregiving services. The impact is likely to be heterogeneous by socioeconomic status. That is, high income and wealth families are more likely to make financial gifts to children while low income families are more likely to provide services. The result of policy change that increases eligibility age for OASI benefits may mean fewer caregiving needs being met among low SES families or a shift to paid caregiving financed by families or government (e.g. through Medicaid). Initiatives or policy change should not only focus on mitigating financial challenges to public entitlement programs but should also take into account the value of unpaid activities as well as the direct monetary gifts to family and the impact of policy change on these services and gifts. 3.6 References Administration on Aging. 2013. "A Profile of Older Americans: 2013." U.S. Department of Health and Human Services. 61 Arno, P.S., C. Levine, and M.M. Memmott. 1999. "The economic value of informal caregiving." Health Affairs 18(2):182-188. Bureau of Labor Statistics. 2010. "Occupational Employment Statistics (OES) Survey." U.S. Department of Labor. —. 2011. Characteristics of Minimum Wage Workers: 2010: U.S. Department of Labor. —. 2012. "American Time Use Survey -- 2011 Results." U.S. Department of Labor. Carlson, M.C., J.S. Saczynski, G.W. Rebok, T. Seeman, T.A. Glass, S. McGill, J. Tielsch, K.D. Frick, J. Hill, and L.P. Fried. 2008. "Exploring the Effects of an ‘‘Everyday’’ Activity Program on Executive Function and Memory in Older Adults: Experience Corps." The Gerontologist 48(6):793-801. Chernew, M.E., D.P. Goldman, F. Pan, and B. Shang. 2005. "Disability And Health Care Spending Among Medicare Beneficiaries." Health Affairs 24(Suppl Health and Costs of the Future Elderly):W5R42-52. Congressional Budget Office. 2014. "The Budget and Economic Outlook: Fiscal Years 2014 to 2024." Crimmins, E.M.and H. Beltrán-Sánchez. 2011. "Mortality and morbidity trends: is there compression of morbidity?" The Journals of Gerontology Series B: Psychological Sciences and Social Sciences 66(1):75-86. Feenberg, D.and E. Coutts. 1993. "An Introduction to the TAXSIM Model." Journal of Policy Analysis and Management 12(1). Freedman, M.M., M.C. Carlson, M.J. Hill, K.D. Frick, N. Ialongo, M.S. McGill, G.W. Rebok, T. Seeman, J.M. Tielsch, B.A. Wasik, and others. 2004. "Experience Corps: design of an intergenerational program to boost social capital and promote the health of an aging society." Journal of Urban Health 81(1):94-105. Fried, L.P., M.C. Carlson, M.M. Freedman, K.D. Frick, T.A. Glass, M.J. Hill, S. McGill, G.W. Rebok, T. Seeman, J. Tielsch, and others. 2004. "A social model for health promotion for an aging population: initial evidence on the Experience Corps model." Journal of Urban Health 81(1):64-78. Furstenberg, F.F., C.S. Hartnett, M. Kohli, and J. Zissimopoulos. forthcoming. "The Future of Intergenerational Relations in Aging Societies." Daedalus. Gold, M.R., J.E. Siegel, L.B. Russell, and M.C. Weinstein. 1996. Cost-effectiveness in health and medicine: Oxford University Press, USA. Goldman, D.P., D. Cutler, J.W. Rowe, P.-C. Michaud, J. Sullivan, D. Peneva, and S.J. Olshansky. 2013. "Substantial Health And Economic Returns From Delayed Aging May Warrant A New Focus For Medical Research." Health Affairs 32(10):1698-1705. 62 Goldman, D.P., B. Shang, J. Bhattacharya, A.M. Garber, M. Hurd, G.F. Joyce, D.N. Lakdawalla, C. Panis, and P.G. Shekelle. 2005. "Consequences Of Health Trends And Medical Innovation For The Future Elderly." Health Affairs:W5R5-17. Goldman, D.P., Y. Zheng, F. Girosi, P.-C. Michaud, S.J. Olshansky, D. Cutler, and J.W. Rowe. 2009. "The benefits of risk factor prevention in Americans aged 51 years and older." Journal Information 99(11). Greenfield, E.A.and N.F. Marks. 2004. "Formal volunteering as a protective factor for older adults' psychological well-being." The Journals of Gerontology Series B: Psychological Sciences and Social Sciences 59(5):S258-S264. Johnson, R.W. 2005. "Value of unpaid activities by older Americans tops $160 billion per year." Kingson, E.and M. Morrissey. 2012. "Can Workers Offset Social Security Cuts by Working Longer?" in EPI Briefing Paper: Economic Policy Institute. Konrath, S., A. Fuhrel-Forbis, A. Lou, and S. Brown. 2012. "Motives for volunteering are associated with mortality risk in older adults." Health Psychology 31(1):87. Lakdawalla, D.N., J. Bhattacharya, and D.P. Goldman. 2004. "Are the young becoming more disabled?" Health Affairs 23(1):168-176. Lum, T.Y.and E. Lightfoot. 2005. "The effects of volunteering on the physical and mental health of older people." Research on Aging 27(1):31-55. Rebok, G.W., K. Ball, L.T. Guey, R.N. Jones, H.-Y. Kim, J.W. King, M. Marsiske, J.N. Morris, S.L. Tennstedt, F.W. Unverzagt, and S.L. Willis. 2014. "Ten-year effects of the ACTIVE cognitive training trial on cognition and everyday functioning in older adults." Journal of the American Geriatrics Society 62:16-24. Social Security Administration. 2013. "Actuarial Publications - Provisions Affecting Retirement Age." Long Range Solvency Provisions. —. 2014. "Social Security Basic Facts." SSA Press Office. Svahn, J.A.and M. Ross. 1983. "Social Security Amendments of 1983: Legislative History and Summary of Provisions." Social Security Bulletin 46(7). Taylor, P., R. Fry, D.V. Cohn, G. Livingston, R. Kochhar, S. Motel, and E. Patten. 2011. "The Old Prosper Relative to the Young: The Rising Age Gap in Economic Well-being." in Social & Demographic Trends: Pew Research Center. The Board of Trustees Federal Old-Age and Survivors Insurance and Federal Disability Insurance Trust Funds. 2013. "The 2013 Annual Report." Washington, DC. —. 2014. "The 2014 Annual Report." Washington, DC. 63 Van Willigen, M. 2000. "Differential benefits of volunteering across the life course." The Journals of Gerontology Series B: Psychological Sciences and Social Sciences 55(5):S308-S318. Wallace, R.B.and A.R. Herzog. 1995. "Overview of the health measures in the Health and Retirement Study." Journal of Human Resources:S84-S107. Zissimopoulos, J., E.M. Crimmins, and P. St. Clair. 2014. "The Value of Delaying Alzheimer’s Disease Onset." Forum for Health Economics and Policy. Zissimopoulos, J.and J. Smith. 2011. "Unequal Giving: Monetary Gifts to Children Across Countries and Over Time." Pp. 289-328 in Persistence, Privilege, and Parenting: The Comparative Study of Intergenerational Mobility: Russell Sage Foundation. Zissimopoulos, J., D.P. Goldman, J. Olshansky, R. Rother & R.W. Rowe (forthcoming). Individual and Social Strategies to Mitigate the Risks and Expand Opportunities of an Aging America. Daedalus. 64 CHAPTER 4. CONSTRUCTION OF PERCENTILE CONFIDENCE INTERVALS INCLUDING SAMPLING VARIABILITY FOR POPULATION-BASED MICROSIMULATION RESULTS 4.1 Abstract Population-based microsimulations are used for policy analysis and decision making worldwide. When the two primary components of microsimulation – the starting population and the rules for transitioning individual characteristics between time steps –are created from survey data, sampling variability is a large source of uncertainty. Uncertainty analysis identifies sources of uncertainty, such as sampling variability, and assesses the impact on microsimulation results. This study applies a direct bootstrapping uncertainty analysis technique to the Future Elderly Model (FEM), a first order Markov process Monte Carlo stochastic model that forecasts health and economic outcomes for middle age and older adults, in order to propagate sampling variability from microsimulation inputs through to the results. Percentile confidence intervals (PCI) are constructed and used to interpret substantive change across simulation scenarios. Compared to the deterministic run of the model, the PCI range with sampling variability is six times larger for life expectancy and almost five times larger for remaining lifetime QALYs. Assumptions about mortality risk reduction in the FEM substantially increase life expectancy (2.4 years), remaining QALYs (1.5 QALYs), and Medicare spending ($27,437). Medical cost growth assumptions substantially increase expected lifetime total medical spending and Medicare spending. After 1200 repetitions, the contribution of stochastic error is relatively close to zero. I conclude that the net contribution of uncertainty analysis to improve model clarity may be worth the additional computational burden and research time commitment. 65 4.2 Introduction Simulation has been widely used to aid health policy decisions worldwide (Astolfi, Lorenzoni and Oderkirk 2012a; Astolfi, Lorenzoni and Oderkirk 2012b), and the continued development of simulations that take into account health and social welfare variables is still a priority (Oderkirk et al. 2012). Microsimulation, as opposed to macro simulation, is particularly useful for forecasting and policy analysis of health and economic issues because interventions are often expected to affect outcomes at an individual level. Health (economic) microsimulations may focus on a specific disease progression (life event). However, the most complex microsimulations – also the most useful to address wide ranging health policy scenarios – model the health and economic characteristics of entire populations and rely on a variety of inputs and data sources to do so. The increasing use of complex health and economic microsimulation models has raised technical challenges, particularly when addressing uncertainty, which should be addressed in order to meet the needs of researchers and policy makers alike. Common contributors to microsimulation uncertainty include sampling variability from input sources, model specification, and stochastic error (National Research Council 1991; Wolf 2001). Stochastic error results from the random process, such as Monte Carlo draws, but can be reduced to nearly zero by running many replicates. Model specification contributes to uncertainty because models, by definition, simplify the real world. When conducting policy analysis, using a consistent specification over all scenarios minimizes the contribution of model misspecification to uncertainty of results. Testing model specifications, such as assumptions about future reductions in mortality risk, can identify areas in which greater care should be taken to correctly specify the model. 66 Sampling variability from input sources is a key contributor to uncertainty when microsimulation inputs are constructed using data from national surveys. Sampling variability can affect both of the primary components in microsimulations, the stock population and transition parameters. The stock population is the population of individuals that initialize the simulation. When the stock population consists of survey data respondents, the starting population is dependent on the sampling design of the survey (first-order uncertainty). Transition parameters are the rules by which individuals age over time and acquire health conditions or experience life events. When the transition parameters are estimated from survey data, the parameters are subject to sampling variability (second-order uncertainty). Varying these inputs in a systematic way is necessary to account for sampling variability in the range of results. However, the challenge of incorporating sampling variability is often overlooked. When every input is treated as completely deterministic, the analysis ignores sampling variability and leads to an overly optimistic confidence interval. Uncertainty analysis can be used to construct confidence intervals that include sampling variability of inputs. Uncertainty analysis first identifies sources of uncertainty in the microsimulation model specification and then assesses the impact of uncertainty on microsimulation results. This identification and assessment of uncertainty makes uncertainty analysis an important part of microsimulation validation (Briggs et al. 2012; Kopec et al. 2010; National Research Council 1991; O'Hagan, Stevenson and Madan 2007). The application of uncertainty analysis to health microsimulation has been limited for several reasons. First, the technical solutions for uncertainty analysis often must be adapted to each specific microsimulation. If a large number of the inputs are estimated from the same sample, it is important to account for the correlation (O'Hagan et al. 2005). The joint parametric 67 distribution of the inputs may be unknown or may not be estimable. Direct bootstrapping implicitly allows for the correlation between multiple input parameters and can be used as conservative technique to increase coverage of confidence intervals without assuming a known distribution of input parameters (Barton 2012). Second, direct resampling methods are computationally burdensome. The direct bootstrapping method requires every individual to be run through the model for every set of parameters (repetitions) before replicating the model to estimate stochastic error. The increase in repetitions requires a large amount of computational time, which is often a limiting factor for researchers. This study contributes to the microsimulation validation literature by adapting uncertainty analysis methods to the Future Elderly Model (FEM) and also expands on the sources of uncertainty assessed in the applied literature. Additionally, the application is to the full model rather than a simplified version of the simulation. First, the sources of first- and second-order uncertainty are identified in the FEM. Second, percentile confidence intervals are constructed by propagating uncertainty from sampling variability through to the outputs using the direct bootstrapping method. Time series assumptions about mortality risk reduction and medical spending growth are assessed for significant impacts on reported results. Next, a large number of replications of the model are completed to estimate stochastic error and confidence in the distribution of the confidence intervals. Finally, methods to reduce the computational burden of the uncertainty analysis are discussed. 4.3 Methods The FEM is a well-established individual-level Markov process Monte Carlo stochastic model of the US population of older adults. It has been used for cohort- and population-level 68 policy analysis purposes (Goldman et al. 2013). The FEM models, in two-year time steps, the incidence of six chronic conditions (cancer, diabetes, heart disease, hypertension, lung disease, and stroke), episodic health states (functional impairment, pain level, cognitive decline) and economic status (work, income, wealth). Policy outcomes (total medical spending, Medicare spending, quality adjusted life years) are also projected based on the modeled characteristics. This section describes the inputs to the FEM that contribute to the uncertainty of reported results, the sampling and simulation strategies used to take the sources of uncertainty into account, and the generation of confidence intervals around the reported results. 4.3.1 Inputs to the Future Elderly Model Large health and economic surveys serve as the foundation of the stock population as well as the source from which the majority of inputs to the FEM are estimated. The Health and Retirement Study (HRS), a biennial nationally representative panel survey of the US population ages 51 and older, is the primary data source of the FEM. The stock population and future cohorts are based on the 2004 wave respondents. Health and economic transitions are estimated from the 1998-2008 waves. Policy outcome inputs are estimated from the Medicare Current Beneficiary Survey (MCBS) and Medical Expenditures Panel Survey (MEPS). Sources of uncertainty including the stock population (first-order) and models of transition parameters and policy outcome estimates (second-order) that are subject to survey data sampling variability are described below. The stock population is a simulated cohort of 51-52 year olds based on HRS characteristics. First-order uncertainty sources include the trends imposed between 1992-2004 using the HRS, trends imposed between 2004-2080 taken from the literature and estimated from NHIS, and a 69 variety of variables that are estimated on the simulated and trended characteristics of the new 51- 52 year old cohorts using the HRS. The basic demographic characteristics of the 1992 wave of the HRS are used with health and economic characteristics trended to the 2004 HRS. A subset of variables is jointly estimated using the 1992 demographic characteristics (hypertension, heart disease, diabetes, weight state, smoking status, functional limitations, employment, health insurance status, pension and wealth). These characteristics are adjusted to 2004 levels based on HRS data. The trends imposed after 2004 are applied to make cohorts look like they will in the future. Variables trended after 2004 include ever having diabetes, ever having heart disease, ever having high blood pressure, weight state, smoking status, pension status, and demographic characteristics from the Census (proportion of the population male, non-Hispanic black, and of Hispanic ethnicity). Additional variables estimated for the cohort based on simulated and trended characteristics include the number of ADLs (IADLs) if predicted to have any ADLs (IADLs), depressive symptoms, as well as economic variables such as wealth, income, and retirement spending. Second-order uncertainty sources include the models used to estimate transition parameters and the models used to predict policy outcomes. The FEM transitions over forty health and economic variables based on previous period characteristics in a first-order Markov process. Transition probabilities are estimated in the simulation from regression equations estimated from the HRS dataset. Binary variables are estimated with probit models, ordered variables are estimated with ordered probit models, and continuous variables are estimated using OLS. For binary and ordered variables, predicted probabilities of transitioning to each state are estimated each wave of the simulation. For continuous variables, the next period value is predicted then transformed when necessary. Each model includes demographic characteristics, socioeconomic 70 characteristics, and initial characteristics at age 50 that control for cohort effects. Health conditions and economic status are included depending on the model and as assessed by experts or based on the literature. Policy outcomes are estimated from regression equations based on current period characteristics. Medicare, Medicaid, out-of-pocket, and total medical spending are modeled by the FEM using data from the MCBS and MEPS. Each model includes demographic characteristics, socioeconomic characteristics, health status, and mortality. When lifetime outcomes are calculated, medical spending is discounted at 2.9% and reported in 2010 USD. Quality adjusted life years (QALYs) are predicted using a three-step process. First, QALYs are modeled using the MEPS data controlling for self reported health and other characteristics. QALYs are then predicted on the HRS data controlling for variables available in both the MEPS and HRS datasets. The QALY model used in the simulation is estimated using the HRS predicted values. 4.3.2 Sampling from Survey Input Data Sampling variability from the survey sources – HRS, MCBS, and MEPS – is considered under the conditions of a resampled survey. The data is randomly sampled with replacement taking into account survey design (Yeo, H and Liu 1999). The HRS is sampled at the household level with 56 strata in recent waves. The MCBS is sampled at the individual level with 100 strata. The MEPS is sampled at the dwelling unit level with 165 strata in recent surveys. Observations are selected for their full panel of interviews. The stock cohort is created based on each HRS sample. The bootstrap samples are also used to estimate the transition and policy outcome models. In the case of the HRS, each input is produced for each sample of the survey. 71 The MEPS and MCBS weights are post-stratified based on the sample and then the weights are used in the estimation of the medical spending regression models. 4.3.3 Confidence Intervals The FEM produces output parameters for each repetition of the simulation. The primary outcomes reported in this study are life expectancy at age 50, remaining healthy life expectancy, lifetime total medical spending after 50, and lifetime Medicare spending after 65. A 95% confidence interval can be calculated for the output parameters as the 2.5 and 97.5 percentiles of the potentially non-parametric distribution of the outcomes. The number of bootstrap samples is started at 1000 based on what is found in the literature, but the total number used depends on the outcome and models being assessed. Watching the variance of the output parameter to see where it settles down offers an estimate of the number of repetitions needed. Running the simulations separately for each source of uncertainty assesses the contributions of sampling variability from the stock population, first-order uncertainty, and the transition and estimation parameters, second-order uncertainty. After including the non-stochastic sources of uncertainty and running the model many times, the stochastic error can be estimated by re-centering the histogram of outputs by replicate at zero (Barton 2012). Repeating the deterministic estimation for 50 to 100 replicates has been considered adequate to reduce the Monte Carlo error of the FEM results to nearly zero. This assumption is test in the final analysis over 100 replicates of the full uncertainty analysis. The mean is calculated as the mean over all the repetitions and replicates. The final confidence intervals are calculated as the mean of the upper and lower percentiles across replicates. 72 4.3.4 Time Series Assumptions Percentile confidence intervals including sampling variability are used to test the impact of other inputs not created from survey data on microsimulation results. Time series assumptions adjust transition probabilities or policy outcome estimations for future expectations. Assumptions about the growth in medical spending over gross domestic product (GDP) growth and mortality reductions directly affect the lifetime outcomes examined in this study. The size and presence of the medical spending growth and mortality reduction time series assumptions is tested using five different scenarios (no mortality adjustment, low cost mortality adjustments, high cost mortality adjustments, no medical spending growth above GDP, no cap on medical spending growth) in addition to the baseline (Table 4.1). In the baseline scenario, the mortality predicted probability is decreased in the future based on estimates from the Trustee’s Report (Social Security and Medicare Boards of Trustees 2012). The mortality model is estimated from waves 5-9 of the HRS, so 2004 (wave 7) is used as the base year for the relative mortality adjustment. The medical growth above GDP growth and base GDP growth assumptions come from Congressional Budget Office (CBO), Centers for Medicare and Medicaid Services (CMS), and Social Security Administration (SSA) reports. After passage of the Affordable Care Act (ACA), medical spending growth is capped at 3%. The scenario without a mortality reduction assumption removes that time series from the analysis. Without mortality reduction, life expectancy and spending are expected to decrease. The level of the mortality reduction assumption is tested using the low- and high- cost estimates of mortality risk taken directly from the Trustee’s Report. The ranges vary from the mid-cost level by approximately +/-5% in 2020 to greater than +/-21% for individuals alive in 2070. 73 The scenario without medical spending growth above GDP growth tests the importance of assuming that medical spending is increasing faster than general spending. I also test the pre- ACA assumption that there is no cap on medical spending growth. The main difference between the two scenarios is higher and lower growth rates. Although, the ‘No Cap’ scenario dips below the ‘With Cap’ scenario in the latest years of the simulation due to a negative growth rate for medical spending growth above GDP growth. Table 4.1 Mortality Reduction and Medical Growth Time Series Scenarios Year Mortality Reduction a Medical Spending Growth Baseline Low Cost High Cost Baseline With Cap No Cap (Med+GDP) (GDP only) (Med+GDP) 2010 0.949 0.949 0.949 1.40% 1.40% 3.79% 2015 0.909 0.929 0.889 0.70% 0.70% 4.31% 2020 0.865 0.911 0.815 3.00% 2.00% 3.22% 2025 0.768 0.832 0.704 3.00% 2.00% 3.14% 2030 0.738 0.815 0.662 3.00% 2.00% 3.05% 2035 0.711 0.799 0.624 2.94% 2.00% 2.94% 2040 0.685 0.783 0.591 2.79% 2.00% 2.79% 2045 0.661 0.768 0.56 2.64% 2.00% 2.64% 2050 0.639 0.753 0.533 2.49% 2.00% 2.49% 2055 0.618 0.739 0.507 2.36% 2.00% 2.36% 2060 0.598 0.725 0.484 2.26% 2.00% 2.26% 2065 0.579 0.712 0.463 2.16% 2.00% 2.16% 2070 0.561 0.699 0.443 2.06% 2.00% 2.06% 2075 0.545 0.687 0.425 1.96% 2.00% 1.96% 2080 0.529 0.675 0.408 1.76% 2.00% 1.76% a Mortality adjustments are based on the Trustee’s Report. 2004 is used as base year. High and Low cost estimates are also taken from Trustee’s Report. b Medical spending growth above GDP growth and GDP growth are come from Congressional Budget Office, Centers for Medicare and Medicaid Services, and Social Security Administration. Abbreviations: Med medical spending growth above GDP growth, GDP gross domestic product. 74 4.3.5 Computational Resources The FEM microsimulation is written in C++ programming language. A single run of the cohort simulation takes approximately 2.5 minutes to complete on a single node and processor. The run time when doing 1000 repetitions of the simulation, each corresponding to a bootstrap sample, can be estimated as a linear function on a single node and processor (approximately 42 hours). Using the resources at the University of Southern California High Power Computing center, the simulation can be run in parallel on multiple processors and nodes. Run time is approximately 2 (1) hours for 1000 repetitions using 25 (50) processors split between 5 (10) nodes. 4.4 Results In a deterministic run of the FEM microsimulation, the stock cohort, transition parameter models, and policy outcome models are the same in every repetition of the simulation. In other words, the inputs to the simulation are treated as the true population cohort and true population models of transitions and policy outcomes. The following results relax those assumptions and allow for the propagation of sampling variability from the inputs to the simulation outputs. First, I relax assumptions for the stock population. Instead of starting the simulation with a cohort of individuals with the same characteristics in every replication, the stock cohort varies based on the sampling variability of the HRS. This adds a source of first-order uncertainty from sampling variability to the FEM results. In this study, the stock population is a cohort of individuals ages 51-52 in 2010. Table 4.2 shows the percentile distribution of mean population characteristics calculated from 1000 samples of the stock cohort compared to the mean from the deterministic run of the 75 FEM. The proportions male and by race/ethnicity do not vary because the sample is reweighted to the 2010 Census estimates. As expected, the means from the deterministic cohort are similar to the means of the sampled cohorts. The largest differences are for variables that are predicted for the cohort based on the trended and demographic characteristics (depressive symptoms 14.5% v. 14.1%; no pain 70.9% v. 69.7%; mild pain 8.8% v. 9.1%; moderate pain 16.2% v. 16.8%; severe pain 4.1% v. 4.3%). The percentile ranges vary in size based by parameter. There is very little deviation in the average cohort age (52.0 years). However, the education level proportions vary greatly over the sampled cohorts (7.1% to 12.8% with no high school degree; 22.2% to 32.3% with a high school degree; 56.8% to 67.7% with at least some college). Rare conditions in the 51-52 year old population have the greatest variation in mean prevalence. Prevalence of stroke (0.8% to 4.1%), difficulty with three or more ADLs (1.0% to 4.0%), difficulty with two or more IADLs (0.3% to 4.0%), and the highest obesity category (0.0% to 11.8%) have prevalence ranges greater than +/-80% of the mean prevalence. Cohort characteristics are controlled for in all transition models, so this variation at the cohort level can be expected to impact the simulation past the initial wave of transitions. Table 4.2 First-order uncertainty of 51-52 year old stock population, 2010 Determi nistic Mean Sampled Characteristic Mean Min 0.05 Median 0.95 Max Observations d , in thousands 18.6 18.5 16.7 17.5 18.5 19.5 20.4 Age d , mean 52 52 51.9 51.9 52 52 52 Sex a , % Female 50.9 50.9 50.9 50.9 50.9 50.9 50.9 Male 49.1 49.1 49.1 49.1 49.1 49.1 49.1 Race/ethnicity a , % Non-Hispanic White/other 77 77 77 77 77 77 77 Non-Hispanic Black 11.9 11.9 11.9 11.9 11.9 11.9 11.9 76 Hispanic 11.1 11.1 11.1 11.1 11.1 11.1 11.1 Education level, % Less than high school 10.1 10.2 7.1 8.6 10.1 11.8 12.8 High school degree d 27.7 27.7 22.2 25.1 27.7 30.4 32.3 Some college or more d 62.1 62.1 56.8 59.3 62.1 64.9 67.7 Work b , % 78.2 78 72.7 75.1 78.1 80.7 82.7 Any health insurance b , % 87 87 83.9 85.5 87.1 88.4 89.5 Chronic conditions, % Ever had cancer 4.7 4.7 2.5 3.6 4.7 6 7.6 Ever had diabetes a,b,d 12.2 12.3 9.1 10.5 12.3 14.1 16.5 Ever had heart disease a,b 10.6 10.5 7.7 8.9 10.5 12.2 14.3 Ever had hypertension a,b,d 37.3 37.4 32.6 34.6 37.4 40.1 42.4 Ever had lung disease d 4.1 4.1 1.8 2.8 4.1 5.4 7 Ever had stroke d 2.2 2.2 0.8 1.5 2.2 3 4.1 Smoking status a,b , % Never smoked 48.3 48.3 44.5 46.2 48.3 50.5 52.8 Used to smoke 51.7 51.7 47.2 49.5 51.7 53.8 55.5 Currently smoking d 21 21.3 17.4 19.5 21.3 23.2 24.7 ADLs, % No difficulty b 91 90.9 87.5 89.2 91.9 92.4 93.2 One c 5.3 5.3 3.6 4.3 5.2 6.4 7.2 Two c 1.6 1.7 1 1.2 1.7 2.3 3 Three or more c 2 2.1 1 1.5 2 2.7 4 IADLs, % No difficulty b,d 95.1 95.1 93 94.2 95.1 95.9 96.9 One c 4.1 4.1 1 3.2 4.1 4.9 6.2 Two or more c 0.8 0.9 0.3 0.5 0.9 1.4 4 Pain level c , % No pain 70.9 69.7 65.5 67.5 69.7 71.8 73.2 Mild d 8.8 9.1 7.3 8.1 9.1 10.2 11.4 Moderate d 16.2 16.8 13.6 15.4 16.8 18.3 20 Severe d 4.1 4.3 3 3.5 4.3 5.2 6.3 Depressive symptoms c,d , % 14.5 14.1 10.9 12.6 14 15.8 17.1 Weight state a,b , % 77 Normal/Under d (BMI≤25) 28 28.2 25.4 26.6 28.2 29.9 31.4 Overweight (25<BMI<30) 36.5 36.5 33.1 34.5 36.4 38.9 48.8 Obese 1 (30≤BMI<35) 21.3 21.6 16.8 19.6 21.5 23.6 28.8 Obese 2 (35≤BMI<40) 8 8.2 3.5 7.1 8.2 9.4 10.5 Obese 3 (BMI≥40) 6.2 5.5 0 3 5.6 7.7 11.8 a Trends are used to adjust prevalence of hypertension, heart disease, and diabetes; weight state and smoking status; male, Hispanic ethnicity, and non-Hispanic black; and pension variables (any DC or DB). b Joint estimation is used to adjust the values of hypertension, heart disease, and diabetes; weight state and smoking status; any functional limitations; employment, health insurance status, pension and wealth in the 1992 cohort of HRS 51-52 year olds based on demographic characteristics. The joint estimation is no re-estimated for each bootstrap sample, but the cohorts are trended from 1992 to 2004 levels of the characteristics based on the sampled HRS. c Characteristics estimated on the trended characteristics of the cohort are modeled controlling for demographic, socioeconomic, and health status. d No rejection of normally distributed means based on Shapiro-Wilk test of univariate normality (p>0.05). Abbreviations: ADL activities of daily living, IADL instrumental activities of daily living, BMI body mass index. Second, I relax the deterministic assumption for the transition parameters and policy outcomes. Instead of the same regression models being used in every replication, the regression models are re-estimated from bootstrap samples of the HRS, MCBS, or MEPS. This adds a source of second-order uncertainty from sampling variability to the FEM results. In this study, there are transition models for binary, ordered, and continuous health and economic outcomes. Policy outcome models are included for QALYs and medical costs. Table 4.3 shows the percentile distribution of mean predicted probabilities for binary and ordered outcomes and mean expected values for continuous outcomes calculated for the first time step (N=1000). As expected, the mean transition probabilities and expected values are similar to the means from the deterministic run of the model. At ages 51-52, the incidence of mortality is pretty low (0.9% two-year mortality rate) with a range between 0.5% and 1.5% depending on the sample. Large percentile ranges as a percent of the mean are seen for estimates with low prevalence in the 51-52 population including incidence of lung disease, incidence of stroke, and living in a nursing home. The largest ranges for continuous models are for helper 78 hours (greater than +- 38%) and other government transfers (greater than +/- 25%). The range of per capita total and out of pocket (OOP) spending is greater than +/- 10% for the Medicare population and greater than +/- 8% for the non-Medicare eligible population when the MCBS and MEPS sampled cost models are randomly assigned to HRS sampled cohorts. The QALY distribution is narrow (+/- 1%). Table 4.3 Second-order uncertainty with predicted transition probabilities and estimates at first transition of 51-52 year old stock population, 2010 Determi nistic Mean Sampled Models a Mean Min 0.05 Median 0.95 Max Transition models Binary - health Incidence of cancer 0.016 0.016 0.011 0.013 0.016 0.019 0.022 Incidence of diabetes 0.036 0.035 0.027 0.031 0.035 0.040 0.074 Incidence of heart disease 0.028 0.028 0.020 0.024 0.028 0.032 0.037 Incidence of hypertension 0.082 0.082 0.068 0.075 0.082 0.09 0.097 Incidence of lung disease 0.014 0.014 0.010 0.012 0.014 0.017 0.021 Incidence of stroke 0.011 0.011 0.007 0.009 0.011 0.014 0.018 Mortality rate 0.009 0.009 0.005 0.007 0.009 0.011 0.015 Depressive symptoms 0.145 0.144 0.124 0.133 0.144 0.156 0.167 Living in nursing home 0 0 0 0 0 0.001 0.002 Binary - economic Any health insurance 0.920 0.919 0.900 0.909 0.919 0.929 0.936 DB claiming 0.061 0.058 0.033 0.043 0.058 0.076 0.100 DI claiming 0.043 0.044 0.030 0.036 0.043 0.052 0.060 Working for pay 0.757 0.755 0.699 0.727 0.755 0.780 0.807 SSI claiming 0.025 0.025 0.016 0.020 0.025 0.031 0.040 Ordered - health ADL limitations One 0.059 0.059 0.05 0.054 0.059 0.064 0.068 Two 0.020 0.020 0.016 0.018 0.020 0.023 0.027 Three or more 0.021 0.022 0.015 0.018 0.022 0.027 0.036 79 IADL limitations One 0.035 0.035 0.028 0.031 0.035 0.040 0.043 Two or more 0.013 0.014 0.009 0.011 0.014 0.017 0.021 Pain level Mild 0.087 0.088 0.083 0.085 0.088 0.091 0.094 Moderate 0.162 0.165 0.149 0.157 0.164 0.174 0.183 Severe 0.052 0.053 0.042 0.047 0.053 0.053 0.070 Smoking Ever smoked 0.340 0.340 0.314 0.327 0.340 0.353 0.370 Current smoker 0.176 0.179 0.144 0.163 0.179 0.196 0.209 Continuous - health BMI 29.29 29.09 27.75 28.65 29.1 29.44 29.73 Continuous - economic Care hours received 39.83 39.82 14.52 22.07 38.68 62.69 111.05 Caregiving to grandkids 113.25 113.32 72.01 87.23 112.08 142.28 177.16 Caregiving to parents 194.96 195.75 132.23 158.66 194.52 236.5 278.11 Volunteering 116.52 116.3 76.45 90.07 115.23 145.7 177.33 Wealth (in thousands) $341 $340 $307 $323 $340 $358 $380 Earnings (in thousands) $44 $44 $39 $41 $43 $46 $47 Government transfers $839 $830 $476 $625 $820 $1,064 $1,316 Inter vivos transfers $3,659 $3,658 $2,594 $3,168 $3,656 $4,171 $4,745 Estimation models Continuous - health QALY 0.829 0.829 0.817 0.823 0.829 0.834 0.838 Continuous - economic Medicare eligible Total medical $16,780 $16,792 $13,798 $15,135 $16,770 $18,490 $20,961 Medicare $10,885 $10,945 $8,239 $9,608 $10,933 $12,427 $14,389 Part A $4,003 $4,013 $2,445 $3,183 $4,010 $4,949 $5,973 Part B $5,032 $5,032 $3,833 $4,396 $5,037 $5,697 $6,480 Part D $3,072 $3,070 $2,722 $2,898 $3,072 $3,257 $3,429 Medicaid $2,193 $2,186 $193 $1,168 $2,183 $3,273 $4,186 Out of pocket $1,294 $1,299 $893 $1,084 $1,302 $1,513 $1,687 Non-Medicare eligible 80 Total medical $6,803 $6,802 $5,766 $6,188 $6,797 $7,434 $7,960 Medicaid $9,965 $9,960 $6,790 $8,034 $9,966 $12,000 $15,035 Out of pocket $963 $962 $805 $885 $961 $1,038 $1,132 a Binary variables are modeled as probit models. Ordered variables are modeled as ordered probit models. Continuous variables are modeled with ordinary least squares except wealth and income, which are modeled as an inverse hyperbolic sine transformation. Abbreviations: DB defined benefits pension, DI Social Security disability, SS Social Security, SSI Social Security supplemental security income, ADL activities of daily living, IADL instrumental activities of daily living, BMI body mass index, QALY quality adjusted life year. These inputs are used in the FEM microsimulation with and without first- and second-order uncertainty from sampling variability. A test of the simulation was completed with 1000 repetitions, each corresponding to a different bootstrap sample of the HRS inputs and MCBS and MEPS models. The variance of the means was monitored for four outcomes of interest (not shown), including life expectancy at age 50, remaining healthy life expectancy, lifetime medical spending after 50, and lifetime Medicare after 65. The variance on life expectancy, remaining QALYs, lifetime medical spending, and Medicare spending seemed to continue downward at 1000 repetitions, so the exercise was repeated for an additional 1000 samples. The variance settles down at 1200 repetitions, which is the number of repetitions used in the remaining analysis. Table 4.4 shows lifetime outcomes for four different scenarios (deterministic, first-order uncertainty, second-order uncertainty, both first- and second-order uncertainty) with 2.5- and 97.5-percentile confidence intervals (PCI). The life expectancy at age 50 is 81.4 years and the remaining QALYs are 23.6. Total medical spending is estimated as approximately $423-$425 thousand and lifetime Medicare spending after 65 is approximately $184 thousand. The deterministic run of the model using the non-sampled stock cohort and non-sampled parameter models gives very tight confidence intervals around the predicted life expectancy (PCI (81.2, 81 81.5)). In the first-order uncertainty scenario, perturbing the characteristics of the stock cohort increases the PCI range over three times for life expectancy (PCI (80.9, 81.9)). The PCI is wider in the second-order uncertainty scenario in which the non-sampled stock cohort is used but the parameters are allowed to vary. When both first- and second-order uncertainty due to sampling variability are accounted for, the PCI around life expectancy is +/- 1 year, or a range five times greater than in the deterministic scenario. The findings are similar for remaining QALYs. The PCI range is approximately +/-1% of lifetime medical spending and lifetime 65+ Medicare spending when only first-order uncertainty added to the model but expands to +/-6% when first- and second-order uncertainty from sampling variability is accounted for. Table 4.4 Simulation results for life expectancy at age 50 with percentile confidence intervals, cohort ages 51-52 in 2010 Deterministic a First-Order Uncertainty Second-Order Uncertainty First- and Second- Order Uncertainty Life expectancy at 50 81.4 (81.2, 81.5) 81.4 (80.9, 81.9) 81.4 (80.7, 82.3) 81.4 (80.5, 82.4) Remaining QALYs at 50 23.6 (23.4, 23.7) 23.6 (23.1, 24.0) 23.6 (23.0, 24.2) 23.6 (22.9, 24.3) Lifetime 50+ total medical spending b $423,464 (419664, 426995) $423,289 (417970, 428468) $424,876 (404480, 447955) $424,656 (403747, 447875) Lifetime 65+ Medicare spending b $183,765 (182155, 185360) $183,709 (181859, 185650) $184,003 (173840, 193903) $183,924 (174207, 193489) a Average over 1200 replicates of model. b Present value of medical spending using a 2.9% discount rate. Abbreviations: QALY quality adjusted life year. Model specifications of time series assumptions are tested by removing the assumptions from the baseline model. Table 4.5 shows the results for five additional scenarios (see assumptions in Table 4.1) run to test time series assumptions about mortality reduction and medical spending growth. The removal of the mortality reduction time series does have a substantial effect on life expectancy (mid-cost=81.4 PCI(80.5, 82.4), no adjustment=79.0 PCI(78.1, 79.9)), remaining 82 QALYs (mid-cost=23.6 PCI(22.9, 24.3), no adjustment=22.1 PCI(21.4, 22.8)), lifetime total medical spending (mid-cost=$424,656 PCI($403,747, $447,875), no adjustment=$360,707 PCI($342,591, $379,780)), and lifetime Medicare spending (mid-cost=$183,924 PCI($174,207, $193,489), no adjustment=$156,487 PCI($147,884, $165,242)). However, the high- and low-cost mortality reduction scenarios do not have PCIs that overlap the mid-cost scenario for life expectancy (low-cost=80.5 PCI(79.6, 81.5), high-cost=82.5 PCI(81.5, 83.5)) and QALYs (low- cost=23.1 PCI(22.3, 23.8), high-cost=24.2 PCI(23.5, 25.0)). The model specification that excludes the medical spending growth above GDP growth assumption substantially increases both lifetime total medical spending ($339,008 PCI($322,616, $357,259) and lifetime Medicare spending ($140,320 PCI($132,938, $147,817)). The cap on medical spending growth substantial increases both lifetime total medical spending ($556,551 PCI($529,120, $587,440)) and lifetime Medicare spending ($245,757 PCI($232,601, $258,864)). Table 4.5 Effect of period assumptions on lifetime outcomes with percentile confidence intervals, cohort ages 51-52 in 2010 Mortality Adjustment Mid Cost Low Cost High Cost No Adjustment Life expectancy at 50 81.4 (80.5, 82.4) 80.5 (79.6, 81.5) 82.5 (81.5, 83.5) 79.0 (78.1, 79.9) Remaining QALYs at 50 23.6 (22.9, 24.3) 23.1 (22.3, 23.8) 24.2 (23.5, 25.0) 22.1 (21.4, 22.8) Lifetime 50+ total medical spending a $424,656 (403747, 447875) $399,190 (379091, 420968) $458,285 (434379, 482661) $360,707 (342591, 379780) With Cap (GDP only) a $339,008 (322619, 357259) No Cap a $556,551 (529120 587440) Lifetime 65+ Medicare spending a $183,924 (174207, 193489) $173,324 (163855, 182708) $197,318 (186873, 207903) $156,487 (147884, 165242) With Cap (GDP only) a $140,320 (132938, 147817) No Cap a $245,757 83 (232601 258864) a Present value of medical spending using a 2.9% discount rate. Abbreviations: QALY quality adjusted life year, GDP gross domestic product. The PCIs reported include stochastic error in addition to the uncertainty from sampling variability. The assumption that the large number of repetitions (N=1200) reduces the stochastic error of the mean estimates to a negligible amount can be tested. Table 4.6 shows the results of replicating the full analysis typical for FEM results (replicates=100) and calculating the standard deviation of the resulting means. The standard deviations are estimates of the stochastic error for 100, 500 and 1200 model repetitions. The standard deviation of the life expectancy mean at 1200 repetitions (SD=0.006) is close to zero and is less than half the standard deviation for 100 repetitions (SD=0.013). The estimate of the stochastic error contribution to remaining QALYs is similarly close to zero (SD=0.005). Although the medical spending means have a wider distribution, the standard deviations are small relative to the size of the mean. In the case of upper bound for lifetime 65+ medical spending, the 95% confidence interval for 100 repetitions covers an area +/-0.7% of the mean and for 1200 repetitions covers an area +/-0.2%. The tight distributions of the upper and lower bounds are unlikely to disrupt the conclusions from Table 4.5 that the inclusion of mortality reduction assumptions and medical cost growth substantially impacts the FEM results. Table 4.6 Estimates of population level stochastic error for 100, 500, and 1200 repetitions of the simulation over 100 replicates Mean (SD) 2.5-Percentile (SD) 97.5-Percentile (SD) Life expectancy at 50 100 repetitions 81.4 (0.013) 80.5 (0.047) 82.4 (0.047) 500 repetitions 81.4 (0.008) 80.3 (0.030) 82.3 (0.022) 1200 repetitions 81.4 (0.006) 80.5 (0.020) 82.4 (0.018) 84 Remaining QALYs at 50 100 repetitions 23.5 (0.010) 22.8 (0.032) 24.4 (0.048) 500 repetitions 23.6 (0.006) 22.8 (0.024) 24.3 (0.022) 1200 repetitions 23.6 (0.005) 22.9 (0.013) 24.3 (0.015) Lifetime 50+ total medical spending a 100 repetitions $424,840 (195.18) $403,501 (971.37) $450,568 (1590.41) 500 repetitions $424,570 (107.33) $402,522 (561.54) $448,183 (705.61) 1200 repetitions $424,605 (67.55) $403,128 (412.88) $447,638 (500.87) Lifetime 65+Medicare spending a 100 repetitions $184,886 (76.39) $174,528 (420.95) $197,151 (778.27) 500 repetitions $184,788 (47.12) $174,982 (258.59) $194,998 (311.80) 1200 repetitions $184,674 (27.76) $174,849 (167.93) $194,702 (181.39) a Present value of medical spending using a 2.9% discount rate. Abbreviations: QALY quality adjusted life year, SD standard deviation. 4.5 Methods to Reduce Computational Burden When high power computing resources are available, probabilistic uncertainty analysis is possible but computational burden remains an issue. Each scenario discussed in the results Table 4.4 and Table 4.5, nine scenarios in total, requires close to 60 hours of computational time. The estimate stochastic error presented in Table 4.6 requires approximately 6,640 hours of computational time. Running in parallel, the amount of time a researcher must wait is reduced drastically. However if a researcher must consider computational resource costs in addition to time, the computational burden may become cost prohibitive. Two methods of reducing computational burden of the uncertainty analysis of individual- level stochastic modeling have been discussed in the literature. Both methods have been applied to smaller or simplified simulations, but it is unknown whether the analyses could be completed on a model as complex as the FEM. 85 ANOVA methods have been suggested (O'Hagan et al. 2007). Reduced numbers of simulation runs can be combined in a factorial way to capture uncertainty for a desired level of certainty and within computational limits. In the case of the FEM, the number of factorial combinations would be to great to complete the analysis in this way. Additionally, O’Hagan et al. (2007) suggest that the analysis of variance method may not be feasible when there is a significant amount of heterogeneity in the stock population of the simulation. Another suggested method is to reduce the computational burden of uncertainty analysis is the use of post-simulation metamodeling (Friedman and Pressman 1988). Linear regression metamodeling has been used to test sensitivity of simulations (Kleijnen 1975) and has been tested for simplified health models (Jalal et al. 2013). Metamodels must be fitted and validated before they can be used for uncertainty analysis purposes (Friedman and Friedman 1985). Kleijnen and Sargent (2000) present ten steps for linear regression metamodeling in stochastic simulations (Kleijnen and Sargent 2000). The application of the exercise to the FEM is straightforward, but fails to meet the high bar necessary for use in uncertainty analysis. The desired metamodel would aid uncertainty analysis by acting as predictive replacement model and therefore reducing computational burden. Inputs and their characteristics included in this study include the list of first- and second-order uncertainty presented in Table 4.2 and Table 4.3. The analysis was applied to the lifetime outcomes, lifetime total medical/Medicare spending and life expectancy/health life expectancy, of a cohort simulation. The metamodel would need a high level of accuracy measured by R-squared, absolute error, and F-tests because the model would be used to predict outcomes for input combinations not estimated in the initial runs of the model. A high level of prediction was not achieved with a linear model of the sampled FEM inputs and lifetime outcomes. 86 4.6 Discussion and Conclusion This study contributes to the microsimulation validation literature by adapting uncertainty analysis techniques to the FEM. The primary objective is to examine the contributions of the main sources of uncertainty – sampling variability, changes in model specification such as the inclusion of time series assumptions, and stochastic error – to the microsimulation results. The uncertainty analysis also requires the explicit identification of sources of uncertainty, which is itself beneficial to researchers and policy makers that want more clarity about these otherwise ‘black box’ models. The first step is to adapt an uncertainty analysis technique to the FEM. Direct bootstrapping is chosen to sample from nonparametric distributions of the input parameters while implicitly including correlation between the great number of inputs. A direct resampling technique is necessary because estimating a multivariate joint distribution of all the FEM inputs that rely on the HRS is unfeasible. The tradeoff is increased computational burden and likely conservative confidence intervals (Barton 2012). The direct bootstrapping method produces very unlikely scenarios such as no individuals having a BMI≥40 in the cohort of 51-52 year olds in 2010 (Table 4.2) or average Medicaid spending in the dual edible population of $198 per capita (Table 4.3). However, the majority of sampled inputs are realistic and produce mean and median estimates close to the deterministic input means.. Percentile confidence intervals, set at the 2.5- and 97.5-percentiles to mirror 95% confidence intervals in many statistical analyses, are estimated including sampling variability from first- and second-order uncertainty sources. Compared to the deterministic run of the model, the PCI range 87 on life expectancy is six times greater, and the range on QALYs is almost five times greater. When changing simulation assumptions and comparing the outcomes to the baseline, a change in life expectancy would need to be at least 0.9 years in order to fall outside the confidence interval including sampling variability. The equivalent change in remaining QALYs would be at least 0.7. The estimates of lifetime medical spending include sampling variability from the three main FEM survey inputs HRS, MCBS and MEPS. The PCI for spending is relatively large at approximately +/-6% of the mean. Currently, the bootstrap samples of the datasets are independent which gives the most conservative estimate of the confidence interval range. Future work could be target a reduction in the variance of medical cost outcomes by more closely aligning the MCBS, MEPS, and HRS sample characteristics. The model specifications tests include the removal of time series assumptions about mortality risk reduction and medical spending growth above GDP growth. The life expectancy and lifetime spending results are sensitive to the assumptions about mortality reduction and about medical spending growth above GDP. The introduction of the mid-cost mortality risk reductions increase life expectancy by 2.4 years with associated increases in lifetime total medical spending of $63,949 and lifetime Medicare spending of $27,437. Additionally, the specifications of medical cost growth greater than GDP growth substantially impacted the medical spending outcomes. These time series are important the model specification and should be updated frequently based with the best available data. The estimate of the contribution of stochastic error to measured uncertainty is relatively close to zero at 1200 repetitions of the FEM. The standard deviations for all measured outcomes are at least half as small as the estimation of stochastic error at 100 repetitions. A greater number of 88 receptions may further decrease stochastic error, but the tradeoff with computational time may need to be considered. Overall, the uncertainty analysis is beneficial to researchers who use the FEM in two ways. First, uncertainty analysis has explicitly identified sources of uncertainty within the model. The contribution of uncertainty from transition parameters is much greater than uncertainty in the stock cohort. Second, the analysis brings to attention the substantial effects of time series assumptions on life expectancy and lifetime spending results, although the exact level of those effects may not be as significant. This analysis is beneficial to the wider research community because it shows that uncertainty analysis of complex models using the full model specification is possible within the confines of current computational capacity. 4.7 References Astolfi, R., L. Lorenzoni, and J. Oderkirk. 2012a. "A Comparative Analysis of Health Forecasting Methods." OECD Health Working Papers(59). Astolfi, R., L. Lorenzoni, and J. Oderkirk. 2012b. "Informing policy makers about future health spending: A comparative analysis of forecasting methods in OECD countries." Health Policy 107:1-10. Barton, R.R. 2012. "Tutorial: Input Uncertainty in Output Analysis." Presented at Winter Simulation Conference. Briggs, A.H., M.C. Weinstein, E.A.L. Fenwick, J. Karnon, M.J. Sculpher, and A.D. Paltiel. 2012. "Model Parameter Estimation and Uncertainty: A Report of the ISPOR-SMDM Modeling Good Research Practices Task Force-6." Value in Health 15:835-842. Friedman, L.W.and H.H. Friedman. 1985. "Validating the simulation metamodel: Some practical approaches." Simulation September:144-146. Friedman, L.W.and I. Pressman. 1988. "The Metamodel in Simulation Analysis: Can It Be Trusted?" The Journal of the Operational Research Society 39(10):939-948. 89 Goldman, D.P., D. Cutler, J.W. Rowe, P.-C. Michaud, J. Sullivan, D. Peneva, and S.J. Olshansky. 2013. "Substantial Health And Economic Returns From Delayed Aging May Warrant A New Focus For Medical Research." Health Affairs 32(10):1698-1705. Griffin, S., K. Claxton, N. Hawkins, and M. Sculpher. 2006. "Probabilistic analysis and computationally expensive models: Necessary and required?" Value in Health 9(4):244-252. Jackson, C.H., S.G. Thompson, and L.D. Sharples. 2009. "Accounting for uncertainty in health economic decision models by using model averaging." Journal of the Royal Statistical Society 172(2):383-404. Jalal, H., B. Dowd, F. Sainfort, and K.M. Kuntz. 2013. "Linear Regression Metamodeling as a Tool to Summarize and Present Simulation Model Results." Medical Decision Making. Kleijnen, J.P.C. 1975. "A Comment on Blanning's 'Metamodel for Sensitivity Analysis: The Regression Metamodel in Simulation'." Interfaces 5(3):21-23. Kleijnen, J.P.C.and R.G. Sargent. 2000. "A methodology for fitting and validating metamodels in simulation." European Journal of Operational Research 120:14-29. Kopec, J.A., P. Finès, D.G. Manuel, D.L. Buckeridge, W.M. Flanagan, J. Oderkirk, M. Abrahamowicz, S. Harper, B. Sharif, A. Okhmatovskaia, and others. 2010. "Validation of population-based disease simulation models: a review of concepts and methods." BMC public health 10(1):710. National Research Council. 1991. Improving Information for Social Policy Decisions -- The Uses of Microsimulation Modeling: Volume I, Review and Recommendations: Council on National Statistics, National Academies Press. O'Hagan, A., C. McCabe, R. Akehurst, A. Brennen, A. Briggs, K. Claxton, E. Fenwick, D. Fryback, M. Sculpher, D. Spiegelhalter, and A. Willan. 2005. "Incoporation of Uncertainty in Health Econoimc Modelling Studies." PharmacoEconomics 23(6):529-536. O'Hagan, A., M. Stevenson, and J. Madan. 2007. "Monte Carlo Probabilistic Sensitivity Analysis for Patient Level Simulation Models: Efficient Estimation of Mean and Variance Using ANOVA." Health Economics 16:1009-1023. Oderkirk, J., F. Sassi, M. Cecchini, and R. Astolfi. 2012. "Toward a New Comprehensive International Health and Health Care Policy Decision Support Tool." OECD Health Division: OECD. 90 Sharif, B., J.A. Kopec, H. Wong, P. Finès, E.C. Sayre, R.R. Liu, and M.C. Wolfson. 2012. "Uncertainty Analysis in Population-Based Disease Microsimulation Models." Epidemiology Research International 2012. Social Security and Medicare Boards of Trustees. 2012. Status of the Social Security and Medicare Programs A SUMMARY OF THE 2012 ANNUAL REPORTS. Wolf, D.A. 2001. "The role of microsimulation in longitudinal data analysis." Canadian Studies in Population 28(2):313--339. Yeo, D., M. H, and T.-P. Liu. 1999. "Bootstrap variance estimation for the National Population Health Survey." In American Statistical Association: Proceedings of the Survey Research Methods Section:49-57. 91 APPENDIX Appendix A. Chapter 2 Transition and Estimation Models A.1 Transitions of depressive symptoms, behavioral risks, and consequences of chronic conditions Clinically Significant DS Smoking status Log transform of BMI Difficulty with ADLs Difficulty with IADLs Pain level Cognitive impairment Demographics Male -‐0.186*** 0.533*** 0.002* -‐0.058** 0.124*** -‐0.156*** -‐0.081*** Non-‐Hispanic black 0.029 -‐0.091*** 0.002 0.104*** 0.097*** -‐0.112*** -‐0.357*** Hispanic 0.128*** -‐0.264*** -‐0.001 0.125*** 0.126*** 0.007 -‐0.215*** Less than HS degree 0.164*** 0.021 0.001 0.122*** 0.170*** 0.037* -‐0.329*** Some college -‐0.141*** 0.125*** 0.000 -‐0.050** -‐0.022 -‐0.057*** 0.132*** Male * Less than HS degree -‐0.082** 0.022 -‐0.001 -‐0.040 -‐0.008 0.044 0.015 Male * College 0.030 -‐0.193*** -‐0.002 -‐0.020 -‐0.100** -‐0.006 0.103** Male * Non-‐Hispanic black 0.039 -‐0.107*** -‐0.007*** 0.044 -‐0.030 0.004 0.025 Male * Hispanic 0.091* 0.241*** 0.001 0.017 -‐0.077 -‐0.001 0.111* Single at 50 0.084*** 0.042** -‐0.001 0.067*** 0.119*** 0.036** -‐0.046* Lag of Widowed 0.058** 0.021 0.000 0.054** 0.054** -‐0.024 -‐0.023 Age splines <65 -‐0.019*** 0.007*** 0.000*** 0.001 -‐0.011*** -‐0.003* 65-‐74 0.004 -‐0.005** -‐0.000 0.014*** 0.032*** -‐0.002 -‐0.029*** 75+ 0.001 -‐0.008*** -‐0.001*** 0.042*** 0.049*** -‐0.002 -‐0.040*** Status at t-‐1 Lag of Depressive symptoms 1.213*** 0.031 -‐0.016 Lag of Heart disease 0.136*** 0.093*** 0.155*** -‐0.030 Lag of Stroke 0.299*** 0.337*** 0.029 -‐0.190*** Lag of Cancer 0.025 0.011 0.025 0.106*** Lag of Hypertension 0.028* -‐0.052** 0.073*** 0.055*** Lag of Diabetes 0.164*** 0.088*** 0.048** -‐0.057** Lag of Lung disease 0.207*** 0.074** 0.169*** -‐0.021 Lag Chronic conditions 1 0.112*** 0.032** 0.002** Lag Chronic conditions 2 0.159*** 0.128*** 0.002 Lag Chronic conditions 3 0.243*** 0.178*** 0.003** Lag Chronic conditions 4 0.301*** 0.212*** -‐0.003 Lag Chronic conditions 5 0.389*** 0.221** -‐0.001 Lag Chronic conditions 6 0.396 0.596** -‐0.012 Lag of Currently smoking 2.491*** -‐0.009*** BMI <30 -‐0.159** 0.820*** BMI ≥30 0.164 0.826*** 92 Lag of ADL 1 0.194*** 0.022 0.001 1.080*** 0.295*** 0.276*** -‐0.154*** Lag of ADL 2 0.301*** 0.035 -‐0.003 1.551*** 0.498*** 0.380*** -‐0.245*** Lag of ADL 3+ 0.324*** -‐0.023 0.000 2.189*** 0.667*** 0.448*** -‐0.358*** Lag of IADL 1 0.165*** -‐0.016 -‐0.002 0.257*** 1.075*** 0.061** -‐0.346*** Lag of IADL 2+ 0.127* -‐0.022 0.002 0.596*** 1.920*** -‐0.033 -‐0.740*** Lag of Dementia 0.162*** -‐0.007 -‐0.007*** -‐1.559*** Lag of CIND 0.157*** -‐0.029 -‐0.004*** -‐0.849*** Lag Pain mild 0.196*** 0.040* 0.003** 0.847*** Lag Pain moderate 0.273*** 0.017 0.000 1.220*** Lag Pain severe 0.332*** 0.030 -‐0.000 1.645*** Lag of Help to parents > 0 0.045** Lag of Help to spouse > 0 0.234*** Interactions with depressive symptoms * Lag of Currently Smoking 0.011 * Lag of IADL 1 0.054 -‐0.184*** * Lag of IADL 2+ -‐0.176** -‐0.375*** * Lag of ADL 1 -‐0.139*** 0.003 * Lag of ADL 2 -‐0.176*** -‐0.114 * Lag of ADL 3+ -‐0.286*** -‐0.227*** * Lag of Diabetes 0.115*** 0.061 0.059 0.047 * Lag of Cancer 0.060 0.027 0.076* -‐0.111* * Lag of Heart disease -‐0.009 0.001 -‐0.052 -‐0.072 * Lag of Hypertension 0.258*** 0.343*** 0.187*** -‐0.131*** * Lag of Lung disease 0.095** 0.034 0.021 -‐0.023 * Lag of Stroke -‐0.111** -‐0.107* -‐0.060 0.053 * Lag of Pain mild 0.008 * Lag of Pain moderate 0.057* * Lag of Pain severe 0.096** * Lag of Dementia 0.189*** * Lag of CIND 0.026 * Age <65 0.000 * Age 65-‐74 -‐0.001** * Age 75+ 0.001* Status at age 50 Heart disease at 50 0.094* 0.059 0.002 0.032 0.060 0.071 -‐0.022 Stroke at 50 -‐0.223 0.393*** 0.005 -‐0.107 0.014 0.061 3.994 Cancer at 50 -‐0.050 0.095** 0.002 0.086* -‐0.017 0.083** 0.023 Hypertension at 50 -‐0.048 0.015 0.006*** 0.021 0.064 0.036 0.040 Diabetes at 50 0.080** 0.021 -‐0.002 0.134*** 0.198*** 0.040 -‐0.081 Lung disease at 50 0.033 -‐0.069 -‐0.002 0.066 -‐0.117 0.110 4.671 93 Ever smoked 0.025 0.001 -‐0.001 0.013 0.033** -‐0.008 Currently smoking at 50 0.123*** 2.752*** 0.002* 0.090*** 0.024 0.025 -‐0.049** Widowed at 50 -‐0.002 -‐0.017 0.000 0.090** 0.078 0.018 -‐0.110** BMI <30 at 50 0.056 0.132*** BMI ≥30 at 50 -‐0.277** 0.105*** Year of birth 0.001*** Time between interviews 0.195*** -‐0.020 -‐0.011*** 0.167** 0.251*** -‐0.053 -‐0.233*** Constant -‐0.638*** -‐1.239*** Cut 1 0.537** 1.872*** 1.722*** 0.773*** -‐2.847*** Cut 2 4.571*** 2.501*** 2.415*** 1.120*** -‐1.761*** Cut 3 2.953*** 2.229*** N 53,874 53,874 53,874 53,874 53,874 53,874 32,038 Pseudo R-‐squared 0.233 0.438 0.848 (adjusted) 0.243 0.257 0.168 0.246 Model type Probit Ordered Probit OLS Ordered Probit Ordered Probit Ordered Probit Ordered Probit Significance indicated by * p<0.05, ** p<0.01, *** p<0.001. Abbreviations: DS depressive symptoms, BMI body mass index, HS high school, ADL activities of daily living, IADL instrumental activities of daily living, CIND cognitive impairment without dementia. 94 A.2 Transitions of mortality and chronic conditions Mortality Cancer Diabetes Heart Disease Hypertension Lung Disease Stroke Demographics Male 0.236*** 0.168*** 0.072* 0.220*** -‐0.102*** 0.030 0.105** Non-‐Hispanic black -‐0.048 -‐0.124** 0.087* -‐0.060 0.078 -‐0.149*** 0.021 Hispanic -‐0.160** -‐0.107 0.321*** -‐0.143** -‐0.020 -‐0.195** -‐0.093 Less than HS degree 0.070** -‐0.027 0.069* 0.110*** 0.038 0.109** 0.051 Some college 0.004 0.063 -‐0.086** 0.020 -‐0.061*** 0.046 -‐0.008 Male * Less than HS degree -‐0.062 0.081 0.007 -‐0.104* 0.034 -‐0.084 -‐0.104 Male * College -‐0.059 -‐0.033 0.081 -‐0.021 0.036 -‐0.244*** -‐0.010 Male * Non-‐Hispanic black 0.123** 0.172** 0.092 -‐0.140** 0.027 -‐0.061 0.010 Male * Hispanic 0.180** -‐0.096 -‐0.160* -‐0.029 0.081 0.005 0.006 Single at 50 0.186*** 0.064* -‐0.000 0.024 0.023 0.073* 0.111*** Lag of Widowed 0.158*** 0.094** 0.002 0.056* 0.074** 0.072* 0.027 Age splines <65 0.038*** 0.020*** 0.014*** 0.022*** 0.012*** 0.007 0.013** 65-‐74 0.020*** 0.020*** 0.010** 0.021*** 0.012*** 0.028*** 0.025*** 75+ 0.051*** -‐0.007* -‐0.011** 0.020*** -‐0.003 -‐0.003 0.025*** Status at t-‐1 Lag of Depressive symptoms 0.066*** Lag of Heart disease 0.228*** 0.150*** Lag of Stroke 0.084*** Lag of Cancer 0.370*** -‐0.001 Lag of Hypertension 0.113*** 0.168*** 0.162*** Lag of Diabetes 0.225*** 0.166*** 0.147*** 0.168*** Lag of Lung disease 0.334*** Lag of Currently smoking 0.133*** 0.048 -‐0.013 0.098** 0.052 0.263*** 0.091* BMI <30 -‐0.175 1.182*** -‐0.061 0.841*** -‐0.004 -‐0.197 BMI ≥30 0.189 1.112*** 0.278 0.388* 0.994*** -‐0.124 Lag of ADL 1 0.202*** Lag of ADL 2 0.384*** Lag of ADL 3+ 0.630*** Lag of IADL 1 0.175*** Lag of IADL 2+ 0.600*** Status at age 50 Heart disease at 50 0.040 0.004 0.213** -‐0.019 0.185* 0.327*** Stroke at 50 -‐0.245 -‐0.147 -‐0.201 0.397* 0.147 0.515** Cancer at 50 -‐0.167** -‐0.009 0.119* 0.033 0.267*** 0.031 Hypertension at 50 0.075 0.019 0.173*** 0.087 -‐0.035 -‐0.065 Diabetes at 50 0.058 -‐0.052 0.105* 0.116* 0.251*** 0.148** 95 Lung disease at 50 -‐0.10 0.239 0.197 0.417** 0.067 0.361* Ever smoked 0.063** 0.046 0.016 0.047* -‐0.006 0.265*** -‐0.010 Currently smoking at 50 0.172*** 0.051 0.064* 0.109*** 0.032 0.231*** 0.081* Widowed at 50 0.040 -‐0.048 0.121* -‐0.120* -‐0.043 -‐0.138 0.126 BMI <30 at 50 0.154 0.718*** 0.223 -‐0.021 -‐0.073 0.181 BMI ≥30 at 50 -‐0.004 0.209 0.232 0.057 0.085 0.383 Time between interviews 0.405*** 0.401*** 0.223** 0.264*** 0.163 0.476*** Constant -‐5.002*** -‐3.612*** -‐9.415*** -‐4.145*** -‐4.906*** -‐2.898*** -‐3.661*** N 57,420 47,390 45,437 42,173 27,390 49,440 50,122 Pseudo R-‐squared 0.242 0.021 0.068 0.044 0.017 0.062 0.058 Model type Probit Probit Probit Probit Probit Probit Probit Significance indicated by * p<0.05, ** p<0.01, *** p<0.001. Abbreviations: BMI body mass index, HS high school, ADL activities of daily living, IADL instrumental activities of daily living. 96 A.3 Models of medical spending Total Medical Spending Medicare Part A Medicare Part B Demographics Male -‐774** 99 -‐416*** Non-‐Hispanic black 1975*** 1550*** 967*** Hispanic -‐912 -‐48 178 Less than HS degree -‐690** -‐179 -‐367*** Some college -‐76 -‐201 -‐74 Male * Less than HS degree 708 723 514** Male * Some college 632 99 518*** Male * Non-‐Hispanic black -‐171 -‐235 65 Male * Hispanic 1508** -‐567 272 Widowed 265 419** 80 Single 344 578** 141 Living in nursing home 42,693*** 4030*** 1495*** Died 22,678*** 16,108*** 7017*** Age group dummies <64 768* -‐44 23 70-‐74 973*** 443*** 159 75-‐79 2988*** 1317*** 760*** 80-‐84 3231*** 1579*** 506*** 85+ 4280*** 2141*** 486*** Status at t Depressive symptoms 4744*** 1672*** 811*** Cancer 2742*** 296* 1653*** Diabetes 3063*** 831** 844*** Heart disease 3607*** 1325*** 1277*** Hypertension 857*** 56 222** Lung disease 3390*** 928*** 938*** Stroke 3668*** 1591*** 904*** Interactions with depressive symptoms * Cancer 1310 1787** 907** * Diabetes 205 178 539* * Hypertension -‐304 417 160 * Heart disease 5 -‐313 54 * Lung disease -‐921 -‐999* -‐525* * Stroke -‐838 -‐13 -‐949*** 97 Demographic interactions Died * Living in nursing home -‐31,755*** -‐9198*** -‐3312*** Died * Age <64 2766 5530 678 Died * Age 70-‐74 7545 11,591** -‐1405 Died * Age 75-‐79 -‐5222 1938 -‐3156* Died * Age 80-‐84 -‐7125* -‐1314 -‐3992*** Died * Age 85+ -‐12,993*** -‐4574 -‐5577*** Disease interactions Died * Cancer 1123 -‐758 1264 Died * Diabetes 2178 1718 1289 Died * Hypertension 2799 2746* 301 Died * Heart disease -‐270 1174 -‐845 Died * Lunge disease -‐6525*** -‐4772** -‐2215*** Died * Stroke -‐3627 -‐2244 -‐1130* Diabetes * Heart disease 307 427 41 Diabetes * Hypertension 682 53 190 Hypertension * Heart disease 701 515* 160 Hypertension * Stroke 950 194 424 Constant 4150*** -‐80 1240*** N 39,174 29,255 28,067 R-‐squared 0.206 0.126 0.088 Model type OLS OLS OLS Significance indicated by * p<0.05, ** p<0.01, *** p<0.001. Abbreviations: HS high school. 98 Appendix B. Chapter 3 Transition and Estimation Models B.1 Transitions of health and disability: Chronic conditions Cancer Diabetes Heart Disease Hypertension Lung Disease Stroke Demographics Male 0.152*** 0.127*** 0.209*** -‐0.078*** -‐0.092*** 0.105*** Non-‐Hispanic black -‐0.124** 0.094** -‐0.046 0.071 -‐0.155*** 0.026 Hispanic -‐0.114 0.341*** -‐0.131** -‐0.015 -‐0.160** -‐0.076 Less than HS degree -‐0.027 0.091** 0.099*** 0.048 0.058 0.055 Some college 0.050* -‐0.045 0.005 -‐0.051** -‐0.046 -‐0.008 Male * Less than HS degree 0.082 -‐0.041 -‐0.077 0.002 0.053 -‐0.083 Male * Non-‐Hispanic black 0.158** 0.064 -‐0.153* 0.036 -‐0.060 0.004 Male * Hispanic -‐0.098 -‐0.203** -‐0.051 0.088 -‐0.035 0.007 Single at 50 0.063* -‐0.011 0.027 0.027 0.074* 0.114*** Lag of Widowed 0.089** -‐0.005 0.053 0.072** 0.053 0.040 Age splines <65 0.020*** 0.017*** 0.023*** 0.013*** 0.007 0.017*** 65-‐74 0.020*** 0.009** 0.021*** 0.012*** 0.028*** 0.024*** 75+ -‐0.007* -‐0.009** 0.019*** -‐0.004 -‐0.001 0.026*** Status at t-‐1 Lag of Heart disease 0.137*** Lag of Stroke Lag of Cancer -‐0.008 Lag of Hypertension 0.164*** 0.151*** Lag of Diabetes 0.158*** 0.146*** 0.144*** Lag of Lung disease Lag of Currently smoking 0.063 -‐0.033 0.108*** 0.073** 0.252*** 0.106** BMI <30 -‐0.107 1.093*** 0.000 0.831*** -‐0.039 -‐0.248 BMI ≥30 0.204 1.153*** 0.315* 0.377* 0.968*** -‐0.106 Status at age 50 Heart disease at 50 0.038 0.208** 0.074 0.170 0.284*** Stroke at 50 -‐0.181 0.158 0.430** 0.010 0.513** Cancer at 50 0.010 0.107 0.027 0.275*** 0.028 Hypertension at 50 -‐0.001 0.234*** 0.080 -‐0.109 0.077 Diabetes at 50 -‐0.069 0.117** 0.114* 0.243*** 0.185*** Lung disease at 50 0.284 0.003 0.469*** 0.094 0.251 Ever smoked 0.045* -‐0.016 0.059** 0.002 0.259*** -‐0.017 Currently smoking at 50 0.038 0.115*** 0.077** -‐0.000 0.266*** 0.088** Widowed at 50 -‐0.052 0.124* -‐0.099 -‐0.043 -‐0.122 0.140* 99 BMI <30 at 50 0.088 0.777*** 0.160 -‐0.008 -‐0.040 0.202 BMI ≥30 at 50 -‐0.023 0.172 0.216 0.081 0.133 0.392* Time between interviews 0.402*** 0.389*** 0.242*** 0.246*** 0.168 0.454*** Constant -‐3.572*** -‐9.475*** -‐4.212*** -‐4.941*** -‐2.841*** -‐3.789*** N 48,691 46,643 43,208 28,089 50,778 51,367 Pseudo R-‐squared 0.021 0.068 0.043 0.017 0.062 0.057 Model type Probit Probit Probit Probit Probit Probit Significance indicated by * p<0.05, ** p<0.01, *** p<0.001. Abbreviations: BMI body mass index, HS high school. 100 B.2 Transitions of health and disability: Mortality, functional impairment, and nursing home residence Mortality ADLs IADLs Nursing home residence Demographics Male 0.218*** -‐0.063*** 0.095*** 0.027 Non-‐Hispanic black -‐0.030 0.077*** 0.123*** -‐0.345*** Hispanic -‐0.155*** 0.117*** 0.114*** -‐0.581*** Less than HS degree 0.066** 0.121*** 0.159*** -‐0.068 Some college -‐0.300 -‐0.052*** -‐0.079*** 0.064 Male * Less than HS degree -‐0.052 -‐0.025 0.056 -‐0.050 Male * Non-‐Hispanic black 0.110** 0.072* -‐0.065 0.251** Male * Hispanic 0.182** 0.027 -‐0.064* 0.184 Single at 50 0.173*** 0.054*** 0.106*** 0.266*** Lag of Widowed 0.154*** 0.047** 0.044* 0.234*** Age splines <65 0.037*** 0.005* -‐0.007** 0.064*** 65-‐74 0.021*** 0.019*** 0.034*** 0.044*** 75+ 0.049*** 0.043*** 0.044*** 0.049*** Status at t-‐1 Lag of Heart disease 0.220*** 0.134*** 0.096*** -‐0.024 Lag of Stroke 0.091** 0.285*** 0.289*** 0.118** Lag of Cancer 0.373*** 0.027 0.009 -‐0.056 Lag of Hypertension 0.110*** 0.051*** 0.026 0.013 Lag of Diabetes 0.206*** 0.132*** 0.121*** 0.187*** Lag of Lung disease 0.344*** 0.228*** 0.080*** -‐0.050 Lag of Currently smoking 0.157*** 0.067*** 0.095*** BMI <30 -‐0.394*** -‐0.773*** BMI ≥30 0.863*** -‐0.393*** Lag of ADL 1 0.212*** 1.022*** 0.323*** 0.238*** Lag of ADL 2 0.401*** 1.458*** 0.502*** 0.470*** Lag of ADL 3+ 0.638*** 2.059*** 0.625*** 0.501*** Lag of IADL 1 0.188*** 0.323*** 1.048*** 0.451*** Lag of IADL 2+ 0.568*** 0.646*** 1.891*** 0.872*** Lag of living in nursing home 2.082*** Lag of household wealth non-‐zero -‐0.098 Lag of household wealth (IHS transform) -‐5.286*** Status at age 50 Heart disease at 50 0.096 0.074 0.087 -‐0.091 101 Stroke at 50 -‐0.287 -‐0.141 0.067 0.201 Cancer at 50 -‐0.175*** 0.091** 0.003 -‐0.035 Hypertension at 50 0.056 0.049 0.162*** 0.221* Diabetes at 50 0.086* 0.128*** 0.193*** 0.190** Lung disease at 50 0.074 0.024 -‐0.198 Ever smoked 0.075*** -‐0.003* 0.009 -‐0.009 Currently smoking at 50 0.150*** 0.067*** -‐0.009* 0.046 Widowed at 50 0.023 0.092** 0.101** 0.098 BMI <30 at 50 0.607*** 0.445*** BMI ≥30 at 50 0.262** 0.279* AIME at 50 -‐0.000 Quarters worked at 50 -‐0.000 Time between interviews 0.179*** 0.251*** 1.194*** Constant -‐4.894*** -‐7.541*** Cut 1 2.885*** 0.838*** Cut 2 3.512*** 1.526*** Cut 3 3.963*** N 59,206 55,345 55,345 55,120 Pseudo R-‐squared 0.246 0.257 0.284 0.488 Model type Probit Ordered Probit Ordered Probit Probit Significance indicated by * p<0.05, ** p<0.01, *** p<0.001. Abbreviations: BMI body mass index, HS high school, ADL activities of daily living, IADL instrumental activities of daily living, IHS inverse hyperbolic sine, AIME average indexed monthly earnings. 102 B.3 Transitions of working for pay and benefit claiming Working for pay Claiming OASI Claiming DI Claiming SSI Demographics Male 0.176*** -‐0.300*** 0.052 0.263*** Non-‐Hispanic black 0.056 -‐0.216*** 0.124* 0.322*** Hispanic -‐0.114** -‐0.140 0.004 0.185 Less than HS degree -‐0.087** 0.054 -‐0.109 0.059 Some college 0.100*** -‐0.200*** -‐0.111** -‐0.094 Male * Less than HS degree 0.076 -‐0.091 0.036 0.051 Male * Non-‐Hispanic black -‐0.113* 0.259** 0.013 -‐0.493*** Male * Hispanic 0.050 0.161 -‐0.239 -‐0.079 Single at 50 0.044 -‐0.065 -‐0.011 0.172*** Lag of Widowed 0.057 -‐0.023 0.204* 0.040 Age variables At EEA -‐0.224*** 0.435*** At FRA -‐0.091** 1.423*** Years to FRA 0.042*** -‐0.235*** Years past FRA -‐0.038*** -‐0.099*** 1 year to FRA 0.413*** 2 years to FRA 0.551*** 3 years to FRA 0.649*** 4 years to FRA 0.580*** 5 years to FRA 0.418*** 6 years to FRA 0.680*** 7 years to FRA 0.493*** 8 years to FRA 0.390*** 9 years to FRA 0.585*** 10+ years to FRA 0.407*** <65 -‐0.000 65-‐74 -‐0.020 75+ -‐0.010 Status at t-‐1 Lag of Heart disease -‐0.029 0.033 0.133** 0.043 Lag of Stroke -‐0.135** -‐0.159* -‐0.048 -‐0.021 Lag of Cancer -‐0.076** -‐0.049 0.20** 0.060 Lag of Hypertension -‐0.071*** 0.041 0.107** 0.016 Lag of Diabetes -‐0.091*** 0.073 0.148** -‐0.022 Lag of Lung disease -‐0.150*** -‐0.090 0.058 -‐0.181** Lag of ADL 1 -‐0.108*** -‐0.151* 0.415*** 0.090 103 Lag of ADL 2 -‐0.276*** -‐0.031 0.381*** -‐0.071 Lag of ADL 3+ -‐0.543*** -‐0.170 0.582*** 0.103 Lag of IADL 1 -‐0.254*** 0.030 0.216*** 0.127 Lag of IADL 2+ -‐0.224*** -‐0.040 -‐0.047 0.041 Lag of working for pay 1.801*** 0.043 -‐0.355*** 0.234* Lag of claiming OASI -‐0.040 -‐0.006 Lag of claiming DI -‐0.447*** -‐1.016*** 2.850*** 0.435*** Lag of claiming from DB 0.035 -‐0.082 -‐0.088 Lag of claiming SSI 2.991*** Lag of individual earnings (IHS transform) 7.525*** -‐18.163*** -‐7.491*** -‐13.320*** Lag of household wealth non-‐zero 0.463*** -‐0.135 0.011 -‐0.010 Lag of household wealth (IHS transform) -‐1.912*** 1.624* -‐2.391*** -‐4.856*** Status at age 50 Heart disease at 50 -‐0.137 -‐0.063 0.167 -‐0.113 Stroke at 50 -‐0.231 0.026 -‐0.095 -‐0.019 Cancer at 50 0.041 -‐0.016 -‐0.021 0.165 Hypertension at 50 -‐0.027 0.015 0.002 0.016 Diabetes at 50 -‐0.034 -‐0.016 0.047 -‐0.098 Lung disease at 50 -‐0.024 0.140 0.234 -‐0.122 Ever smoked -‐0.059** 0.079 0.063 -‐0.022 Currently smoking at 50 -‐0.024 -‐0.023 0.131*** 0.074 Widowed at 50 -‐0.092 -‐0.039 -‐0.299* -‐0.001 AIME at 50 -‐0.000** -‐0.000*** -‐0.000 Quarters worked at 50 0.009*** 0.006*** -‐0.004*** Time between interviews -‐0.238*** 0.503*** -‐0.023 0.064 Unemployment rate 0.007 Constant -‐1.498*** -‐0.502*** -‐2.829*** -‐2.356*** N 30,465 5,376 20,114 36,741 Pseudo R-‐squared 0.493 0.212 0.659 0.727 Model type Probit Probit Probit Probit Significance indicated by * p<0.05, ** p<0.01, *** p<0.001. Abbreviations: OASI Old Age and Survivors Insurance, DI Disability Insurance, SSI Supplemental Security Income, BMI body mass index, HS high school, EEA early entitlement age, FRA full retirement age, ADL activities of daily living, IADL instrumental activities of daily living, DB private defined benefits pension, IHS inverse hyperbolic sine, AIME average indexed monthly earnings. 104 B.4 Transitions of wealth, income and other transfers IHS transform of Household wealth IHS transform of Individual earnings Property tax Inter vivos transfers Other government benefits Demographics Male -‐0.776*** 0.408*** -‐266.997*** 177.546 6510.446*** Non-‐Hispanic black -‐4.417*** 0.059 -‐213.030 -‐68.044 -‐186.792 Hispanic -‐2.269*** -‐0.413 -‐49.857 -‐107.021 -‐1249.245* Less than HS degree -‐1.921*** 0.104 90.868 19.914 54.579 Some college 3.533*** 0.624*** 286.062*** 1172.509*** 4273.303*** Male * Less than HS degree 0.240 -‐0.164 -‐28.655 -‐4359.938*** Male * Non-‐Hispanic black 0.541 0.067 248.690 -‐1817.527** Male * Hispanic -‐1.534** 0.369 100.293 -‐3149.525*** Single at 50 -‐3.620*** 0.399*** 266.543** -‐480.521 Lag of Widowed -‐3.767*** 0.777*** 473.945*** 334.786 Lag of Married -‐842.906*** Age variables <60 0.208*** 20.374 235.014** 60-‐74 0.141*** -‐14.126 -‐214.599*** <65 -‐0.037*** 65-‐74 -‐0.004 75+ -‐0.190*** -‐0.008 -‐1.934 42.683 Age at July 1 -‐277.150*** Age-‐squared 2.067*** Status at t-‐1 Lag of Heart disease -‐0.611*** -‐0.488*** -‐114.161 33.787 Lag of Stroke -‐1.454*** 0.197 -‐29.093 -‐207.240 Lag of Cancer 0.447 -‐0.198 220.037** -‐630.996 Lag of Hypertension -‐1.068*** -‐0.091 -‐60.110 -‐232.594 Lag of Diabetes -‐1.510*** -‐0.220 -‐85.240 156.871 Lag of Lung disease -‐1.888*** -‐0.216 -‐48.842 -‐523.078 Lag 1 chronic disease 137.490 Lag 2 chronic diseases -‐338.102** Lag 3 chronic diseases 308.388 Lag 4 chronic diseases -‐478.557 Lag 5 chronic diseases 70.144 Lag 6 chronic diseases 739.335 Lag of ADL 1 -‐1.400*** -‐0.270 -‐21.104 -‐71.037 Lag of ADL 2 -‐1.307*** -‐0.549 -‐77.581 876.720 Lag of ADL 3+ -‐1.203*** -‐0.031 22.263 343.438 105 Lag of IADL 1 -‐0.412 -‐0.539 123.298 740.049 Lag of IADL 2+ -‐1.338** -‐2.630*** -‐88.272 -‐235.981 Lag of any ADL 78.733 Lag of any IADL 234.559 Lag of currently smoking 276.386 Lag of working for pay 0.211 -‐0.711*** -‐202.926 -‐586.187** -‐1726.392** Lag of claiming OASI -‐1.436*** 91.808 774.024 Lag of claiming DI -‐1.563*** 19.365 -‐868.979 Lag of claiming from DB -‐1.542*** 276.598** -‐602.159 Lag of claiming SSI 1730.804*** Lag of living in a nursing home -‐131.022 767.845 Lag of individual earnings (IHS transform) 1.567 118.478*** 8023.667** 11,610.502* 70,519.559*** Lag of household wealth non-‐zero -‐0.380 -‐822.350 -‐240.879 Lag of household wealth (IHS transform) 748.652*** 1.286 18,153.832*** 363.954*** 49,786.694*** Lag of property tax 0.113*** Lag of household capital income (IHS transform) 134.008*** Lag of inter vivos transfers 0.246*** Status at age 50 Heart disease at 50 -‐0.400 0.346 57.437 879.839 Stroke at 50 -‐2.421 -‐0.196 132.860 1295.050 Cancer at 50 0.906 0.093 -‐116.438 1376.865 Hypertension at 50 -‐1.417*** 0.114 -‐107.723 -‐1566.401* Diabetes at 50 -‐0.847 -‐0.293 -‐64.883 -‐759.567 Lung disease at 50 1.966 -‐0.086 216.766 -‐1217.873 Ever smoked 0.193 -‐0.170 124.849 1994.640*** Currently smoking at 50 -‐1.957*** -‐0.049 -‐183.371** -‐1988.036 Widowed at 50 0.342 -‐0.502 68.402 856.699 AIME at 50 0.002*** 0.001*** 0.261*** 2.666*** Quarters worked at 50 -‐0.046*** -‐0.018*** -‐4.699*** -‐51.464*** 1 child at 50 -‐1.726*** 1851.764*** 2 children at 50 -‐0.668 2584.337*** 3 children at 50 -‐1.298*** 2492.090*** 4+ children at 50 -‐2.225*** 2396.643*** Individual earnings at 50 (IHS transform) 31.332*** Time between interviews -‐1.041 0.216 694.049 Wave 5 -‐0.381 -‐631.339 Wave 6 -‐0.628** -‐132.898 Wave 7 -‐0.005 106 Wave 8 0.555* Constant -‐2.770 6.548*** -‐539.844 6280.092*** -‐9941.007 N 25,348 6,360 8,614 55,345 3,207 Adjusted R-‐squared 0.050 0.080 0.364 Model type GHREG GHREG OLS OLS OLS Significance indicated by * p<0.05, ** p<0.01, *** p<0.001. Abbreviations: OASI Old Age and Survivors Insurance, DI Disability Insurance, SSI Supplemental Security Income, BMI body mass index, HS high school, ADL activities of daily living, IADL instrumental activities of daily living, DB private defined benefits pension, IHS inverse hyperbolic sine, AIME average indexed monthly earnings. 107 B.5 Transitions of caregiving and volunteering Hours of caregiving received from spouse Hours of caregiving to parents Hours of caregiving to grandchildren Hours volunteering Demographics Male -‐5.708 -‐0.370 -‐61.932*** 2.549 Non-‐Hispanic black 7.863 17.013** 34.810*** -‐3.144 Hispanic 5.676 0.115 -‐0.271 -‐16.268*** Less than HS degree 17.929 -‐20.379*** 3.310 -‐16.318*** Some college -‐2.744 3.649 -‐5.382 44.178*** Single at 50 7.682 -‐6.548 -‐8.293** Lag of Widowed 7.481 -‐22.228*** 7.039* Age variables Age at July 1 -‐34.378*** -‐15.641*** 1.640 8.634*** Age-‐squared 0.293*** 0.079*** -‐0.049** -‐0.067*** Status at t-‐1 Lag of Heart disease 19.876* Lag of Stroke 166.284*** Lag of Cancer 13.119 Lag of Hypertension -‐15.448* Lag of Diabetes 75.393*** Lag of Lung disease 72.943*** Lag 1 chronic disease -‐1.074 4.187 -‐3.862 Lag 2 chronic diseases -‐8.666 10.498 -‐14.507*** Lag 3 chronic diseases -‐11.893 3.686 -‐19.974*** Lag 4 chronic diseases -‐8.401 -‐13.095 -‐18.796** Lag 5 chronic diseases -‐12.985 -‐7.621 -‐42.137** Lag 6 chronic diseases -‐43.794 -‐23.238 -‐52.283 Lag of ADL 1 127.077*** Lag of ADL 2 360.817*** Lag of ADL 3+ 609.545*** Lag of IADL 1 230.838*** Lag of IADL 2+ 791.570*** Lag of any ADL -‐5.131 -‐8.823 -‐19.479*** Lag of any IADL -‐7.685 -‐13.923 -‐9.356* Lag of working for pay -‐32.061*** -‐5.347 18.476* 17.522*** Lag of individual earnings (IHS transform) -‐11.430 61.013 -‐1068.909*** -‐820.971*** Lag of household wealth (IHS transform) -‐8.135*** 276.463*** 171.387 394.234*** Lag of household capital income -‐0.829 (IHS transform) -‐0.000 -‐0.000 0.000 Lag of hours of caregiving received from spouse 0.244*** 108 Lag of hours of caregiving for parents 0.271*** Lag of hours of caregiving to grandchildren 0.225*** Lag of hours of volunteering 0.358*** Status at age 50 1 child at 50 5.667 2 children at 50 28.516 3 children at 50 23.921 4+ children at 50 17.929 Number of children living within 10 miles 1.690 2.280 23.611*** -‐2.478** Lives in suburbs 11.188** -‐7.393 1.102 Lives in exurbs 11.879** -‐7.426 -‐4.900 Religion not important -‐44.594*** Religion somewhat important -‐41.034*** Catholic -‐15.949*** Jewish -‐12.483 Other religion -‐9.578 No religion -‐12.871** Constant 1045.333*** 739.318*** 237.538** -‐201.061*** N 36,697 55,345 51,783 55,345 Adjusted R-‐squared 0.164 0.092 0.082 0.186 Model type OLS OLS OLS OLS Significance indicated by * p<0.05, ** p<0.01, *** p<0.001. Abbreviations: HS high school, ADL activities of daily living, IADL instrumental activities of daily living, IHS inverse hyperbolic sine. 109 B.6 Models of medical spending Medicare Part A Medicare Part B Medicare Part D Dual eligible Medicaid Non-‐Medicare eligible Medicaid Demographics Male 17 -‐131 -‐223*** -‐1248** -‐2875** Non-‐Hispanic black 668 351 483*** 1011** 618 Hispanic 649 184 158*** 376 -‐212 Less than HS degree -‐478 -‐371*** 495*** -‐249 -‐1632 Some college -‐102 146 -‐87*** 1231** -‐96 Male * Less than HS degree 546 242 -‐245 1758 Male * Non-‐Hispanic black 1170 586 24 867 Male * Hispanic -‐1375 405 -‐456 2278 Widowed 269 51 164*** -‐1978*** -‐2759** Single 146 20 -‐1424*** -‐932 Living in nursing home 4241*** 1870*** 30,210*** Died -‐9779 -‐2327 3915 Age group dummies 55-‐59 -‐553 60-‐64 1335 65-‐69 365 -‐230 3237* 70-‐74 619 -‐200 4075** 75-‐79 1594*** 521 4261** 80-‐84 1768*** 93 3874** 85+ 2072*** 6 5409*** Age splines <75 -‐11*** 75+ -‐11*** Status at t Cancer diagnosis 6943*** 6362*** 3470*** Diabetes diagnosis 1668** 659* 4846** Heart disease diagnosis 4291*** 2490*** 1918* Hypertension diagnosis 1246*** 595*** 3468*** Lung disease diagnosis 2813*** 1094*** 2055** Stroke diagnosis 6371*** 509 4458* Cancer maintenance 431** 1443*** -‐345 Diabetes maintenance -‐239 775*** 1652 Heart disease maintenance 1043*** 1431*** 13 Hypertension maintenance -‐18 407*** -‐809* 110 Lung disease maintenance 406* 767*** 40 Stroke maintenance -‐203 397 1049 Cancer 3625* Diabetes 205*** 929 Heart disease 160*** 3810*** Hypertension 163*** -‐992 Lung disease 218*** 2270* Stroke 114*** 4142*** Currently smoking 143*** IADL 1 158*** IADL 2+ 291*** ADL 1 65** ADL 2 93** ADL 3+ 3837*** 2434*** 228*** 450 Claiming disability 792 229 631*** 3331* Earnings (log transform) -‐12,048 Demographic interactions Died * Living in nursing home -‐5091** -‐2499*** -‐18,506*** Died * Age 65-‐69 26,800*** 9394*** 243 Died * Age 70-‐74 34,851*** 7537*** -‐3559 Died * Age 75-‐79 26,068*** 5493*** -‐2540 Died * Age 80-‐84 23,723*** 4676*** -‐1243 Died * Age 85+ 19,887*** 3109** -‐2969 Died * Claiming disability 29,226*** 10,847*** -‐2377 Living in nursing home * Claiming disability -‐1269 -‐904 -‐6218** Disease interactions Died * Cancer -‐85 2455*** -‐787 Died * Diabetes 1589 2280*** 2600 Died * Hypertension 1596 1082* 21 Died * Heart disease 2789 486 427 Died * Lunge disease -‐4021* -‐1916*** 590 Died * Stroke -‐1121 -‐353 987 Diabetes * Heart disease 357 -‐107 -‐860 Diabetes * Hypertension 1125** 438* -‐1245 Hypertension * Heart disease 266 -‐221 332 Hypertension * Stroke 1010 338 -‐1075 Constant -‐58 1394*** 1,551*** -‐913 5906*** N 20,469 19,668 19,344 4,418 1,740 111 Adjusted R-‐squared 0.139 0.114 0.239 0.484 0.048 Model type OLS OLS OLS OLS OLS Significance indicated by * p<0.05, ** p<0.01, *** p<0.001. Abbreviations: HS high school, ADL activities of daily living, IADL instrumental activities of daily living.

Abstract (if available)

Abstract The essays expand and strengthen microsimulation research and apply a microsimulation model in the areas of health services research, labor and retirement, and uncertainty analysis.

Linked assets