POHEM : a framework for understanding and modelling the health of human populations / Michael C. Wolfson

POHEM- a framework for understanding and mode/ling the health of human populations Michael C. Woltsona Introduction The POpulation HEalth Model (POHEM) is both an idea and a computer simulation model. At the level of ideas, it is part of a framework for un der- standing and thinking about human health. In terms ofboth software and data, POHEM is a prac- tical implementation of as many of the concepts and ideas as have proven feasible with the resourc- es invested so far. The genesis of POHEM lay in Statistics Cana- da's concern that its health statistics programme was weak. This diagnosis was not so much due to problems with the data being gathered by Cana- da's national statistical agency, but rather to large gaps in information in important areas. The most important information gap, or imbalance in the health statistics programme, concerned infor- mation on the health status of the Canadian popu- lation. There were far more data on resource in- puts to the health care system, and its throughputs in terms ofvisits to doctors and hospital stays, than on health outcomes. Data on individual health sta- tus were dominated by causes of death and clinical diagnoses for hospital stays. Other major statistical concerns were the lack of coherence among the myriad bits of health- related data being collected and proposed, lack of common concepts and definitions across jurisdic- tions and institutions within Canada, and a narrow- ness of focus reflecting the dominance of clinical medicine relative to other perspectives such as th ose of public health, health promotion, and ap- preciation of the social determinants of health. These statistical concerns reflect deeper sub- stantive concerns with the directions of Canadian health policy. Traditional (for high-income coun- tries) health (actually illness) care is increasingly seen not only as very expensive, but also as having questionable efficacy. There is no denying a core of modern health care interventions that, from a broader historical perspective, are nothing short of a Statistics Canada and Canadian Institute for Advanced Research (ClAR). The views expressed are my own, and not necessarily those of Statistics Canada or the ClAR. The development of PO HEM has been a team effort. I am deeply indebted to my Statistics Canada colleagues for their many contributions, and to the ClAR for providing a unique intellectual milieu. This is a substantially revised version of Ref. (1). I of course rernain responsible for any errors or infelicities. Wld hlth statist. quart., 47 (1994) miraculous. The concern, rather, is with a margin perhaps as large as 25% where otherwise useful interventions are applied inappropriately, and thus have negligible benefits (2). Moreover, even for medical interventions that offer sorne non-negligible benefit, more stringent requirements are being considered and increasing- ly demanded as a prerequisite for public funding- namely that the intervention rank highly according to sorne sort of benefit or cost-benefit criterion (3).b In addition to costs, such benefit analysis re- quires at least 2 major elements. One is a broadly agreed-upon measure of benefit that appropriately reflects impacts of alternative health-affecting in- terventions on population health status. The other is a body of theory of the determinants of popula- tion health such that the implicit "what if' ques- tion of cost-benefit analysis can be credibly an- ~- what would the change in the popula- tion's health status be ifthe specified interven- tion were funded and implemented? Finally, there is a sense in Canada that it is time for the pendulum to swing back somewhat from the current emphasis on medical interventions. Increasingly, evidence on the social determinants ofhealth has entered public debate. In a somewhat unholy alliance, there now appears to be greater common cause among fiscal conservatives who wish to eut back on what is seen as a bloated medi- cal establishment, and social progressives who wish to increase funding for programmes in areas like early childhood development and home care for seniors. Irrespective of the merits of these views, they clearly call for a much broader perspective on the field of health information, and on the range of health-affecting interventions to be considered in both a theoretical and a cost-benefit framework. Our population health model is being devel- oped with these statistical and substantive concerns very much in mind. PO HEM began as a conceptual effort, forming a central part of a proposed system ofhealth statistics (4). These efforts within Statistics Canada have also become integrated with a larger review of health information in Canada (5). b The recent experience in the state of Oregon suggests considerable difference ofviews on the extent to which costs should be considered when ranking health interventions, but clear agreement on the need to measure benefit. See also Ref. (3). 157 PO HEM subsequently moved through a prototype phase, and is now beginning to be used in applied analyses. While much of the value of the modellies in the kinds of analyses it can support, it should be home in mind that the genesis of PO HEM was and continues to be as a core element in a new system ofhealth statistics. POHEM builds on a foundation of computer- intensive techniques - both in its realization as a simulation model, and in its use of inputs drawn from rich and highly multivariate data. While this may make it sound expensive, it need not be. PO- HEM runs on standard persona! computers. Since it is "upwardly compatible" from multi-state life table styles of analysis, it can build on data already collected for such purposes. For example, it can nest, as a special case, estimates of disability-free life expectancy (DFLE) or disability-adjusted life years (DALYs) (6, 7). This article begins with the motivation behind POHEM at the conceptual and theoretical level, then describes microsimulation methods generally and the population health model specifically, and concludes with a few brief examples drawn from analyses using PO HEM. The discussion which follows makes reference to conceptual frameworks, statistical systems, and PO HEM itself. These terms are closely interrelated but different. A conceptual framework is a set of general ideas. Specifie reference will be made to the conceptual framework developed in the Tem- plate for Health Information (8)- both for its ideas and for the pedagogical benefit of sorne of its graphie images. The Template in tum provides a sketch of the beginnings of a system of health statistics. An actual system of health statistics does not exist in Canada, but is currently the objective of renewed efforts, following the impetus of the re- cent National Task Force on Health Information (5). Finally, PO HEM is being developed as one part of the activities involved in creating a system of health statistics. Motivations, concepts and "theories of health" As noted in the introduction, much more statistical effort is spent measuring the inputs to and throughputs of the health care system than in mea- suring how healthy the population is. Thus one central motivation for PO HEM is the development of population health status measures to help reme- dy this imbalance. A desire for coherence is the second motiva- tion. Health certainly rivais the economy in impor- tance, yet in comparison the statistical base is con- fused, fragmentary and incoherent. The System of National Accounts (SNA) provides a coherent sta- tistical framework for economie information. The coherence of the SNA derives from the fact that its data elements ohey a series of arithmetic identities. Furthermore, economie series like unemployment 158 rates and interest rates, while not connected arith- metically to the SNA series, are related in various macroeconometric models developed and housed in nearby organizations. The same kind of mathe- matical structure generally does not exist for health data. Coherence in health statistics is bene- ficiai for two reasons. First, at a conceptuallevel, it aids understanding of the interrelationships among various data elements. Second, at a statisti- cal level, basic arithmetic identities, when com- bined with redundancy in data sources, provide an ongoing check on data quality (a point that is illustrated la ter). In so far as PO HEM is part of a larger process of statistical development, it must at least implicitly have sorne theoretical foundation. Theory and measurement are closely interrelated in ali scientif- ic endeavour. For example, theories in astrophysics explicitly guide the development and construction of specifie kinds of radio-telescopes, particularly for the strengths and frequencies of signais to be detected. Of course, there is no uni-directional causality here-from concepts and theory to a mea- surement system. Rather, the history is one of itera- tion back and forth between theory and concepts on the one hand, and observation and measure- ment on the other. In the case ofhuman health, a large portion of current statistical measurement is based on a dis- ease-oriented theory. However, there is enough accumulated evidence on the broad range of deter- minants ofhealth, and concern about the sequelae of disease processes, to signal the need for major changes in theoretical and conceptual perspective, and therefore changes in approaches to observa- tion and measurement. To give sorne flavour to these empirical results and findings, we might mention that, for example, there is widely accepted evidence that serum cho- lesterol is connected to heart disease - but also to many other important metabolic pathways such as serotonin. Similarly, there is strong evidence that social support and incomes are connected to lon- gevity - including associations with decades-long latencies; and there are specifie "windows" in early childhood when key opportunities for physiologi- cal development (e.g., vision) are present. Yet oth- er evidence suggests major roles for genetic and myriad environmental factors. These kinds of di- verse empirical nuggets ali have sorne superficial connection with one another - for example they are ali pertinent to human health. However, there are no grand theories that knit these diverse empirical findings together; and the piecemeal character of current partial theories tends to generate piecemeal empirical results. There is need at least for a theoretical structure which can hold and index ail the piecemeal results, much as a library holds books on diverse topics. Moreover, it would be even better if the "index- Rapp. trimest. statist. sanit. mond., 47 (1994) ing" in this library of health-related empirical find- ings connected the elements in a logical and co- herent manner. To give a simple example, in addi- tion to indexing tobacco smoking alphabetically under 'T", it could also be connected to lung cancer and heart disease, and it could have con- nections coming from television advertising and peer pressures. A theoretical structure or conceptual frame- work of this sort would not only be broad enough to encompass all the diverse kinds of phenomena associated with human health; it could also play an integrating role, particularly by fostering observa- tions and empirical work that seek to knit together various piecemeal results. Such a conceptual framework was developed as part of the National Task Force on Health Information (5), specifically the Template for Health Information (8) which in turn built on Evans and Stoddart (9). From a theoretical perspective, the basic struc- ture proposed in the Template is in 2 parts - de- scriptions of the variables of interest, and descrip- tions of how they evolve over time. In dynamic systems, these are referred to respectively as the "state space" and the "laws of motion". In human health, the "laws of motion" are generally either unknown or contentious. But the set of variables of interest, the "state space", is more widely accept- ed. Thus, current knowledge can support a theo- retical structure which has a place for all the vari- ables of interest, and provides mechanisms for us- ing or trying out various "laws of motion" as they are proposed and refined through careful re- search. Wld hlth statist. quart., 47 (1994) Fig. 1 provides an image from the Template (actually a monochrome "still" drawn from the animated graphics in the Template software) that can serve as the basis for describing such a struc- ture - a comprehensive state space for population health. The image in Fig. 1 divides the health field into 3 broad domains. At the centre of the image (delib- erately so) is a representation of an individual's life cycle. The array of cubbyholes is a visual meta- phor for a hypothetical individual's biography of events and states. The shading illustrates an anima- tion sequence for a hypothetical individual's bi- ography. It starts with a blank matrix of cubbyholes or cells. The cells are then filled in from left to right, representing the evolving biography of the individual as he or she passes through the life cycle. For example, there are events like entering school, getting married, being exposed to risk factors, hav- ing a heart attack, partial recovery, but then declin- ing health and eventually dying. Surrounding but not completely enveloping the individual in Fig. 1 (visually and metaphorical- ly) is the "External Milieu"- a major source of influences on individuals' health.c The external milieu has in turn has been classified into 4 kinds of environments. The popular media tend to identify "the envi- ronment" only with the first of these, the physical- c While the time dimension is explicit for the individual in the centre of Fig. 1, it is only implicit for the external milieu and for health-affecting interventions. Fig. 1 159 chemical environment. However, there is strong evidence that socio-cultural and economie expo- sures, the second and third kinds of environment, are at least as important to human health as risks derived from physical-chemical exposures. Fourth, sorne aspects of the health system are best consid- ered as an environment (e.g., the existing stock of hospital buildings and medical equipment). However, the bulk of society's health-related activity is shawn in the vertical bar to the right of the basic Template image. The visual nuance is that while the individual in the centre ofthe image is surrounded by the "External Milieu", an indi- vidual's health is not completely out of his or her control at the whim of various environmental ex- posures. Health is also influenced by individuals' and society's consciously intended actions - a large collection of "Health-Mfecting Interven- tions". This phrase was deliberately chosen to be broader than "health care", since there is much more that influences the health of individuals in society than opportunities to visit health care pro- viders. The Template sub-divides this broad domain of interventions into 2 groups. The first, "individual" health-affecting interventions, act on us as individ- uals typically in one-on-one settings - for example encounters with providers of various health care services like a dental visit or hospitalization for an appendectomy. The second is "collective" health- affecting interventions. These interventions take the form of government programmes and regula- tions that act on us collectively, although indirectly (and often inadvertently) through the external mi- lieu. Examples include regulations of food quality, tobacco advertising, pharmaceuticals, water quality and particulate emissions, or expenditures on traf- fic safety and headstart-type programmes. d Returning to the central portion of the image, the individual biography is represented much like the record layout in a computerized longitudinal microdata set. Each row corresponds to a group of variables or fields in the record layout, and thus to a set of state variables in a dynamic system. Each column refers to a rime interval with age along the horizontal axis. The main groups of variables de- scribing this hypothetical individual's stylized bi- ography are: • socio-economic status (SES) attributes - vari- ables known to have strong statistical and likely causal associations with health; • various risk factors including genetic predisposi- tions, physical factors (e.g., serum cholesterol, obesity, hypertension), lifestyle factors (e.g., d To extend the analogy with dynamic systems to the ideas of mathematical control theory, health-affecting interventions constitute the "control variables" for influencing the trajectories of individuals bath directly, and indirectly via the trajectories of the environments of the extemal milieu. 160 smoking), and risks deriving from environmen- tal exposures (including noxious physical-chem- ical agents, and socio-cultural factors such as the availability of social support); • clinically-defined diseases like ischaemic heart disease, Jung cancer, osteoarthritis and Alzhei- mer' s ( classified by the International Classifi- cation ofDiseases-ICD); • vernacular health problems- such as not having sufficient lower limb or cognitive capacity to get around without a wheelchair, and chronic pain (classified by the impairment and disability por- tions of the International Classification of Im- pairments, Disabilities and Handicaps-ICIDH); • direct costs of health services used;e and • finally the bottom row, as weil as the "bottom line"- a summary health status value- a number between zero = dead (shawn as black in the figure) and one= fully healthy (shawn as white) summarizing each individual's overall health status for the year. Note that environmental exposures (i.e., con- tacts with the external milieu) are largely implicit in risk factor exposures, the evolution of SES at- tributes like incarne from paid work, and in the case of handicap the interaction between vernacu- lar health problems (i.e., impairments and disabili- ties) like being wheelchair-bound, and the physical and social environment. Also, handicap in the sense of the ICIDH is not considered an intrinsic individual attribute. Rather it is captured or "ex- pressed" in the individual's interactions with the external milieu. The dynamics or laws of motion for these state variables in an individual's biography are not explicit in this image. Ail that is shawn is the top level of a classification structure. However, descrip- tions of the dynamics of various health-related pro- cesses are an essential part of the "theory" and health information that should be encompassed by this framework. For example, at school age, the socio-cultural environment of school, family and television may influence attitudes toward healthy behaviours. Later, these behaviours may influence diet, which then over many years may influence susceptibility to heart disease. One well-known marker could be serum cholesterollevels, but ath- er factors like chronic stress may also be significant to the individual's health history. The Template image is agnostic about which specifie causal sto- ries are correct. The key point is that it provides a framework within which such stories, or better still scientific evidence ifit is available, can be codified, quantified, and accumulated. e Indirect economie costs like foregone earnings can also be considered, but they must be measured by a simulated comparison of eamings to a hypothetical scenario where a given disease or health problem is "deleted". Rapp. trimest. stat1st. sanit. mond., 47 (1994) For example, the shading in Fig. 1 indicates (illustratively) the levels of scalar values for a set of individual attributes over age intervals in the indi- vidual's life course. The dynamics representing various theories or knowledge relating to health then consist of those rules, mapping systems or formulae that allow any column of an individu- al's biographical array (assuming discrete time) to be "filled in" (i.e., estimated, calculated or im- puted) as a function of all the information on the individual's history to the left, plus any relevant exposures or influences from the external milieu and health-affecting interventions. In reality, these dynamic processes involve a heterogeneous population. Th us, even moderately realistic statistical descriptions will likely be com- plex, multi-level and multivariate. They will inevita- bly have substantial stochastic components, not least to reflect myriad factors that have not been explicitly included in the analysis. The (various and partial) theories comprehended by the concep- tual framework are embodied in these explicit de- scriptions of the dynamics of the various individual attributes. For example, the fertility "theory" that under- lies many demographie projections is naturally in- corporated by having the event of giving birth rep- resented by one of the rows in each individual's biographical record layout, and taking as the law of motion or dynamics for this type of event a stochas- tic process described by a set of age-specifie fertility rates. Of course, this is a very simple theory of the determinants of fertility. Other variables like edu- cational attainment, marital status, prior birth spac- ing, ethnicity and labour market experience are omitted even though analysis of appropriately rich data would likely show them to be significant. Knowledge in a case like this is limited by the availability of multivariate longitudinal microdata. The framework in Fig. 1 is open in this regard. As long as independent or determining variables like educational attainment are part of the state space, they can be incorporated as inputs to a (re- cursive) algorithm describing the dynamic process. The same point applies to health processes. There is general agreement that hypertension and elevat- ed cholesterol increase the risk of heart disease; there is continuing debate in the epidemiological literature about the magnitudes of these effects, the form of interactions if any, and the range of other variables that are also significant determi- nants of coronary events. The framework in Fig. 1 in principle can absorb or include any such theory or observed empirical regularity, and can in fact encompass alternative or contendingversions. The only proviso is that they are well-defined, i.e., they can be expressed algorithmically. More broadly, there is clear evidence that the socio-cultural milieu and economie circumstances have a profound effect on health (10.13). Increas- Wld hlth statist. quart., 47 (1994) ingly, public po licy is turning from the question of how many people have high cholesterol to what is it about our communities that predisposes mem- bers to certain dietary patterns (14). Again, the framework in Fig. 1 is agnostic. By design, it can incorporate significant results of this form, namely dependencies of individuals' dynamics on the attributes of their communities. The state space can be extended to include the relevant variables, and the dynamic processes can be expressed as functions of these community level variables. The only proviso is that they be measurable- i.e., they can either be derived as a well-defined function of a set of individuallevel variables (e.g., the neigh- bourhood's poverty rate) or are separately col- lected (e.g., ambient air pollution levels or crime rates). General role of microsimulation The discussion so far has set out our appreciation of the basic situation with regard to statistical infor- mation in the health field, and has sketched a broad conceptual framework based on the Tem- plate. In turn, this conceptual framework can guide the construction of a system of health statis- tics which both reflects current theory and under- standing, and supports further development of new theory and understanding. This conceptual framework and proposed system ofhealth statistics also serve as the basis for a theoretical structure expressed as a computer simulation model. This idea is developed in 3 stages. First, we describe the utility of computer simulation models as one cen- tral aspect of a statistical system like that proposed for health information. Second, the importance of microsimulation modelling methods is developed. Then, in a later section, the current version of PO HEM as one specifie instance of a microsimula- tion model is described, along with the analyses which it supports. One basic reason for reliance on simulation in the health area is to "observe" life expectancy (LE). LE is a synthe tic indicator that requires ob- servation (populations at risk and mortality rates) plus the simulated answer to the specifie "what if' question, "How long would a birth cohort live on average if everyone were exposed to a given series of age-specifie mortality rates?" In turn, "statistics" that are generalizations of LE also require numeri- cal simulation to be "observed". More concretely, we are interested in a family of generalizations of LE, including "cause-deleted" LE, as well as indi- viduals' (in a given population) expected dura- tians in various health states. Our premise is that such indicators are fundamental to creating good summary population health status statistics, and thereby remedying the imbalance of current health statistics. As a result, sorne sort of simulation modelling structure is an essential component of a system of health statistics. 161 Simulation models are also of central strategie importance because they give coherence. Without sorne sort of integrating analytical framework like the network of arithmetic identities in the System of National Accounts, or a simulation model, data series risk being a hodgepodge, as is the case cur- renùy in the health area. Turning this point around, it is relatively easy to consult with various data-using constituencies and then compile a wish list of needed data, as was do ne in the course of the Task Force consultations (5). But if the wish list is at least parùy framed by, or can be mapped into a coherent structure like a simulation model, the whole may well be greater than the sum of its parts. The various constituent data series will be of inter- est in and of themselves. They can also be "in- dexed" by the overall conceptual framework (e.g., the classification structure embodied in the Tem- plate), by having each kind of data associated with a specifie element or part of the framework. But the real power of coherent data will only be fully realized when combined in an explicit quantitative simulation model. There are 2 main reasons. First, simulations allow the posing and answering of rigorously constructed "what if' questions. Exploration of these kinds of hypothetical scenarios is fundamental to decision making and basic research. Second, with proper de- signed-in redundancy in the data feeder systems, sim- ulation can enhance the generation of "data con- frontations". These are situations where the same concept or number can be estimated in 2 different ways. In principle, the results should be identical. However, they are typically different (as in the "re- sidual error of estimate" in the System of National Accounts). Such "confrontations" serve as an in valu- able check on statistical error, particularly non-sam- pling error which is far more difficult to assess (15). This kind of data confrontation is illustrated in the case of lung cancer data in the penultimate section on PO HEM applications below. The utility of microsimulation follows essentially from the heterogeneity of individuals and their behaviours. The conventional, partially-aggregated or cell-based approaches of macroeconomies and much of demography are simply inadequate to capture the richness and texture of the ph en orne- na of interest in the health area- and in economies and demography for that matter.f A microanalytic approach is not only computationally feasible, but also needed to begin reflecting realistic population heterogeneity. It is also needed to provide the corn- mon foundation for the generalizations of LE to be described below. f Wolfson, M.C. Implications of evolutionary economies for measurement in the SNA, towards a system of social and economie statistics, Twenty-third General Conference of the International Association for Research in Income and Wealth, St. Andrews, New Brunswick, August 21-27, 1994. 162 The data on individuals envisioned in Fig. 1 above comprise an ideal microdata set. Unfortu- nately, these data cannot be observed direcùy. The implied longitudinal household survey is impracti- cal - not least for reasons of respondent burden, privacy and confidentiality concerns, and the cen- tury or so we would have to wait until it was com- plete. Moreover, by the time the century oflongitu- dinal follow-up was complete, the information would likely be useless because so much had changed. We need to be able to observe recent trends and regularities in behaviour, and then make extrapolations and predictions about their implications. In this context, microsimulation modelling is a methodology for synthesizing a cohort of the req- uisite biographies using more practical, albeit frag- mentary, data sources. Microsimulation modelling can be thought of as a form of super imputation where a variety of partial pieces of data and partial descriptions of dynamic processes are woven to- gether in to a set of realistic- but synthe tic - individ- uallife cycle histories. We sketch the details of this synthesizing process below. It must be supported by a large effort not only of meta-analysis, as in- creasingly undertaken in the epidemiologie litera- ture, but more generally a process of meta-synthesis. In meta-analysis, a number of cohort studies con- sidering the same phenomenon, say the relation- ship between serum cholesterol and heart disease events, are examined and efforts are made to pool the results so as to increase the effective sample size. Microsimulation modelling (and large scale modelling generally) goes beyond this by taking quantitative results from a variety of domains ( e.g., labour force participation transitions, risk factor dynamics, disease incidence hazard functions) and seeking to draw out their joint implications. The result is not just one "baseline" instance of the longitudinal microdata set implicit in Fig. 1. It is also a simulation model capable of constructing hypothetical alternative versions of this data set where one or more factors have been changed. A much more familiar example of this kind of hypo- thetical simulation experiment is cause-deleted LE - how long could we expect to live if there were no mortality at all from a particular kind of cancer, for example.g Microsimulation modelling is able to construct generalizations of this kind of indicator, and to base the computations on explicit models of g Note that these kinds of cause-deleted !ife expectancies are qui te simplistic. They are based implicitly on a mode! that treats each cause of death as strictly independent. One example where this is clearly false is for heavy cigarette smokers. If they are hypotheticallypreventedfrom ever dying oflung cancer, as in a Jung cancer-deleted !ife expectancy calculation, theywould still be at elevated risk of dying from heart disease, chronic obstructive pulmonary disease, etc. Such underlying factors causing elevated riskare rompletelyignored in the usual cause- deleted !ife table estimates. Rapp. trimest. stat1st. sanit. mond., 47 (1994) the complex interactions among risk factors and co-morbidities. In addition, by using microsimulation madel- ling methods, the underlying (synthetic) longitudi- nal population data are always available. Thus, oth- er kinds of related statistics can be readily comput- ed. Moreover, these synthetic individual biogra- phies are always available for inspection, in turn allowing the plausibility or "face validity" of the simulation results to be assessed in much greater detail. This is not possible with conventional (par- tially aggregated or cell-based) multi-state life table methods, which invisibly yield life paths which are highly implausible.h The hypothetical alternative versions of the ba- sic longitudinal microdata set of Fig. 1 generated by microsimulation modelling can provide the basis for policy analysis and decision making. For exam- ple, cause-deleted LEs have often been used to form a "league table" of the most "important" diseases - with heart disease in first place, and can- cers second, since they typically account for the largest expectations of years of life lost in high- income countries. This ranking in turn influences the allocation of resources for health care and health research. However, this is not the best indi- cator for setting priorities for resource allocation. First, the ranking measure should account for more than mortality effects, which are ali that af- fect increments in LE due to the hypothesized elimination of a given disease. lt should also take account of morbidity and disability while alive. The core idea is to ad just measures based on LE for how ill or unhealthy people are year by year. A wide variety of health-status-adjusted life expectancy measures have been proposed. For convenience and simplicity, we shall refer to such adjusted LE measures as healthy life expectancy (HLE).i One consequence of adopting such adjustments is that h For example, in an increment-decrementlife tablewith atleast two states, one for healthy and one for disabled, if the transition probabilityis non-zero at ali ages, therewill be implicitlife paths that oscillate back and forth between healthy and disabled every year, though the proportion of the cohort with such !ife paths will be very small. i Other major acronymic contenders include QALYs (quality· adjusted life years), ALE (active life expectancy), QWB ( quality of well-being), and DFLE (disability-free life expectancy). QALYs and QWB are inappropriate because "quality" has a much broader connotation than the health basis should imply. ALE and DFLE are unsatisfactory because they treat disability too simplistically as ayes/no condition with no recognition of gradations of disability. DALE (disability adjusted life expectancy) isa possibility thatwould build on the major efforts inestimating DALYs bythe WorldBank (See Ref.(7)) .However, in our judgement, HLE seems the most straightforward and accessible acronym for the desired concept. A useful review of this terminology is found in Mathers, C.D. et al. (Health expectancy indicators: recommendations for terminology, Seventh Meeting of the International Network on Health Expectancy (Réseau Espérance de Vie et Santé), Canberra, 23-25 February 1994), where HALE (health-adjusted !ife expectancy) corresponds to our HLE. Wld hlth statist. quart., 47 (1994) in league tables based on cause-deleted HLEs rath- er than LEs, chronic disabling diseases which are not typically fatallike arthritis and dementias will rank as much more important. Second, our thinking should be broadened so that diseases are not the only category of health- related phenomenon that can affect HLE. For ex- ample, tobacco smoking behaviour has just as much daim to be considered a "cause" in cause- deleted HLE as conventional diseases classified by the ICD. Generally, the capacity to treat as wide a range as possible of risk factors and health-affect- ing interventions as "causes" is desirable in gener- ating estimates of cause-deleted HLEs. Then, not onlywould arthritis and dementia rise in the result- ing league table ranking, so too would smoking, and perhaps chronic stress. Moreover, health care interventions like coronary artery bypass graft sur- gery could also be placed in the same ranking. Indeed, items currently outside the usual health care discourse like "neighbourhood cohesive- ness" might enter the league table. Of course, this is all contingent on having plausible causal stories represented algorithmically, for example by for- mally-estimated stochastic processes. In effect, HLE thus becomes the common stan- dard for measuring population health status. It is also a simple summary index of population health whose regular publication would meet the basic concern expressed at the outset regarding imbal- ance in health statistics. Microsimulation modelling can play a role be- yond policy analysis and the construction of sum- mary indices. It is also significan t for basic research and the development of statistical priorities. Hav- ing microsimulation model-based indicators at the core of a statistical system will induce a tighter coupling and hence more fruitful basic research. This effect can be illustrated as follows. A micro- simulation model is built to estimate HLE. In a number of areas, data are very limited or even non- existent. In these areas, "guesstimates" are made, and then a sensitivity analysis is undertaken to see which of the various guesstimates is most important to estimates of HLE. The result should be in- creased priority to the collection of data relevant to the most important guesstimate. Even if such sensitivity analysis to various guess- timates does not feed back immediately to data- collection priorities, the process has value to basic social science research. Much of this research con- sists of co~ectures about possible causal pathways and their magnitudes. For example, how impor- tant is unobserved heterogeneity with respect to sorne kind of innate "frailty" ( 16); what would be the impact of relaxing the assumption of indepen- dence of competing risks? These basic questions cannot be addressed by new data collection, either because we simply do not currently know how to collect the data ( e.g., innate frailty), or because it is 163 logically impossible (dependent competing risks (17)). The alternative is to use a numerical simula- tion madel. How to microsimulate Such is the conceptual background for a proposed system of health statistics, in which numerical mi- crosimulation models can play a key role. A micro- simulation madel can serve as a repository for sorne of the requisite information, and provide a method for constructing of a family of HLE-based indicators. In so doing, the microsimulation madel "solve s" or draws out the implications of the la test observations of the population's status, using empirically-based inferences for the various "laws of motion". In this section we illustrate the microsimulation modelling method using a simple life table exam- ple. The illustration appeals to the structure shawn in Fig. 1. The rows in the individual's biography represent, in effect, the record layout for the hy- pothesized longitudinal microdata set, and as dis- cussed earlier, the "state space" for the madel. The columns represent the individual at various ages. It is also implicit that there is a series of "slices" or planes going back into the page repre- senting a population of individuals. Microsimulation (at least in a discrete time ver- sion) synthesizes the se data set in a series of 3 nest- ed loops. At the innermost level, the simulation process creates one column vector in an individu- al's biography by synthesizing each element in the vector working from top to bottom. This pro- cess starts with the "birth" of an individual at age zero. This is followed by the repeated application of the appropriate dynamic algorithms to fill in column vectors for successive ages until the indi- vidual dies. This is the second loop in the simula- tion. Finally, the third and outermost loop synthe- sizes a large sample of individualsj Life table analysis can be nested as a special case of microsimulation modelling. It is not as efficient computationally to use microsimulation madel in- stead of life table methods where the latter are feasible, and the only desired results are summary statistics like LE. However, microsimulation madel- ling is much more readily generalized than life table analysis. A pertinent example is the estima- ~on of disability-free life expectancy (DFLE) by mcrement-decrement multi-state life tables ( 18). In microsimulation modelling terms, this life table j The order of the 2 outermost loops can be reversed. In other w?r~s, the simula~on could proceed individual by individual Wlthm ~ year unul a full (pre-specified) sample of column vecto~s 1s co~pletely synthesized. Th en as the outermost Ioop, ~e. ~1mula~on could proceed year by year un til the Iast 1~d1V1dual d1es (assuming the simulation applies to a synthetic b1rth cohort as is typical oflife table analyses). 164 analysis corresponds to a very much simplified ver- sion of the biographical record layout in Fig. 1. At most, 3 rows are needed from the image - one for alive or dead, one for healthy or disabled, and one for the health status value of each life-year. In other words, an individual's state space in each year of life consists of a three-tuple or a three element col- umn vector. These three-tuples are simulated using micro- simulation one individual at a time, starting at birth and moving from left to right in the sense of F'_ig. 1 using "laws of motion" or dynamic algo- nthms defined by simple functions as follows: Be- ing in the alive/ dead and healthy / disabled states at age ais assumed to depend stochastically only on age and prior disability. If the individual is alive and healthy at age a-1, the possible transitions to age a are: no change, become disabled, or death. Similarly, if the individual is alive and disabled at age a-1, the possible transitions are: no change, become healthy, or death. The transitions are based on observed probabilities. In other words, this is a first-order Markov process. A Monte Carlo process is applied where ran- dom numbers are drawn, and depending on the draws and the individual's state at age a-1, the individual is simulated to die, or advance one year in age, possibly becoming disabled or getting bet- ter. To represent this (discrete state, discrete time) dynamic process, assume that an individual's bi- ography has been synthesized up to age a-1. Empir- ical observations provide the basis for 2 sets of transition probabilities. These are looked up from the model's input data - tables of m(i,a) and d(i,a), the probability of dying, and the probability of changing disability level respectively, at age a given disability status i at age a-1. The simulation process proceeds by drawing a random number from a uniform distribution over the range 0 to 1. If the number is in the [0, m(i,a)] interval, the individual is simulated to die; if in the [m(i,a), m(i,a) + d(i,a)] interval, the individual survives but changes disability status; if the number is in the [m(i,a) + d(i,a), 1] interval the individual survives and remains in the same disability state. This process of drawing random numbers and testing them against the exogenously estimated transition probabilities is repeated over ages to c?mplete the individual's synthetic biography. Fmally, many such biographies are generated to synthesize a large sample.k We th us have complete- ly synthesized a population of complete life-cycle biographies with 2 of the 3 elements in the state space - alive/dead and healthy/disabled. Lastly, k The sam pie size is chosen to bring the Monte Carlo sampling error down to the level where the resulting estimates have the desired level of precision. Monte Carlo error can be determined by sample reuse methods. Rapp. tnmest. statist. sanit. mond., 47 (1994) the third health status value is computed from the contemporaneous alive/ dead and healthy 1 dis- abled values according to a (simplistic) function that assigns a value of one if alive and not disabled, zero otherwise. In this example, using a microsimulation madel to "solve" for the life table resulting from these simple dynamic algorithms, the conventional sum- mary statistics can be computed as follows: sum- ming across ages in the alive/dead row (assuming alive has a value of one and dead a value of zero), and then averaging over all the individuals in the sample results in an estimate of LE. The same summing and averaging applied to the third health status row gives DFLE. As shawn in Wolfson & Manton,l a very wide range of models of disability, risk factor, and disease processes can be expressed in terms of the state space of Fig. 1 combined with monte carlo microsimulation. Microsimulation modelling clearly nests conventional multi-state in- crement-decrement life tables as a special case. Similarly, microsimulation modelling should be able to nest the recent World Bank DALY analysis (19,20). The multi-state increment-decrement life table madel just reviewed, and cell-based models more generally, tend to embody strong simplifying as- sumptions. These include independence among processes, and processes represented by transition probability functions with only a few simple inde- pendent variables. The practical reason is that more involved process representations entail a combinatorial explosion in the size of the state space, and hence in the number of cells. However, it is more realistic to consider many processes as interdependent and simultaneous. For example, getting married, buying a house, fin- ishing school, and entering the labour market are decisions or socio-economic transitions occurring in early adulthood that are often jointly deter- mined. Such interaction can be included in micro- simulation modelling. The major problem is not any constraint imposed by the microsimulation modelling methodology. Rather, it is the set of difficult empirical questions raised in gathering the appropriate data and estimating any interac- tions among hazard or transition probability func- tions. The POHEM microsimulation structure PO HEM is one particular instance of a microsimu- lation madel, and is very much a work in progress, with a number of areas currently under active de- l Wolfson, M.C. & Manton, KG. A review of models incorporating notions of population health expectancy, presented to the fifth international meeting of REVES, Ottawa, 1992. Wld hlth statist. quart., 47 (1994) velopment. To begin, PO HEM creates not just in- dividuals, but male-female pairs. This is done in anticipation of a marriage or a common-law union. As well, children and prospective remarriage part- ners are explicitly included in this family structure or "case". Thus, individuals are simulated in close to a nuclear family context (i.e., with other individ- uals who will be part of a given individual's nu- clear family at sorne point in his or her lifetime). The fulllife cycle of each case is simulated, not just one single individual at a time. A case is completed with the death of the last adult (and the last child leaving home) before another is commenced.m Unlike the DFLE life table modeljust described in arder to illustrate the microsimulation madel- ling method, POHEM includes a large number of sometimes complex processes or dynamic algo- rithms. For sorne processes, algorithms for several variants are included in the software. Each state variable is listed below, along with an indication of the method used to set the variable, or the variables drawn upon as inputs to the associated transition probability function.n Socio-Economic Status educational attainment- endowed at birth by draw- ing from univariate distributions and husband- wife correlations based on Canadian census data. first union - either legal marriage or common law union (CLU); probability at each age represent- ed by a multivariate hazard function of age, sex, education, fertility (for females), labour force history, CLU history, and pre-ordained mar- riageability (for unobserved heterogeneity, see Rowe and Wolfson°). first spousal age difference - based on age at mar- riage, and observed joint distribution of brides' and grooms' ages. fertility- probability based on age, parity and mar- ital status. rn This particular !ife cycle structure, as weil as the set of socio- economic status variables described later, reflects POHEM's origins as a mode! for analysing public pension policy and demography (See also Ref. 22). n Most efforts have been devoted to enriching the statistical descriptions of various health-related processes in Canada. However, in order for PO HEM to be more broadly accessible, including use in other co un tries, a version un der development will allow users to specifY both simpler and more complex dynamics, depending on the available data. Sorne variables, such as marriage and mortality rates, will have to be unique to each country. Other processes, however, are hopefully more "international", i.e., constant across national populations. One such international process might be Jung cancer survival. 0 Rowe, G. & Wolfson, M.C. Biased divorce: validation of marital status !ife tables and microsimulation models, paper presented to the United Nations Economie Commission for Europe, Seminar on Demographie and Economie Consequences and Implications of Changing Population Age Structure, Ottawa, Canada (1990). 165 union dissolution - either divorce or separation; probability at each age represented by a multiva- riate hazard function of age, duration of mar- riage, presence of children, labour force experi- ence, age at marriage.0 child leaving home - probability based on age, sex and birth order of child. remarriage - probability based on age, sex and whether divorced or widowed. second spousal age difference - based on marrying person's age at marriage, sex, and prior mari- tal status, then drawn from the observed joint distribution of brides' and grooms' ages. labour force participation - probability of entry or exit at each age represented by a set of multivari- ate hazard functions of age, sex, marital status, presence of children by age group, educational attainment, and duration in state (22). labour market earnings - dollar level each year based on an auto-regressive stochastic process with parameters based on age, sex, and strength oflabour force attachment.P Risksq smoking, cholesterol, blood pressure, and obesity- quadrivariate joint density at age a derived as first order Markov function of quadrivariate joint density at age a-1, age, and sex based on analysis of the 1978 Canada Health Survey (23). Tomiak and Berthelot have updated the analysis to make use of the recent round of provincial Heart Health Surveys.r radon- endowed at birth by drawing from a lognor- mal fit to the observed distribution oflevels with- in residential dwellings. ages of females at menarche and menopause - based on a sample of 90 000 women from the National Breast Screening Study (24). Diseases first coronary event incidence - a multivariate risk function from the Framingham study and Merck model (25,26). heart disease progression, types of events (cardiac arrest, myocardial infarction, angina pectoris, or sorne combination) and case fatality- based on Weinstein et al. (Ref. (27) and persona} commu- P Kennedy, B. The LIPPS Earning Module. mimeo, Social and Economie Studies Division, Statistics Canada, Ottawa (1986) q The particular risk factors modelled represent the historical development of PO HEM, where heart disease was fust. The next area of development was Jung cancer, where we thought it would be useful to include one environmental exposure, hence the inclusion of radon. Currentlyworkis underwayon a breast cancer module, hence ages at menarche and menopause have been added as risk factors. r Tomiak, M. & Berthelot, J-M. Modeling lifetime risk factor histories", mimeo, Social and Economie Studies Division, Statistics Canada, Ottawa (1994). 166 nication). While this model is cell-based, it can be approximated to any desired degree of accu- racy using Monte Carlo microsimulation, and has been transformed into an microsimulation model variant for PO HEM. lung cancer- incidence by age at diagnosis, sex and cell type based on Canadian cancer registry data; stage at diagnosis based on special chart review studies (28); relative risk conditional on cumula- tive radon and tobacco exposure up to age a-10 according to a risk function estimated by Whitte- more and McMillan (29); progression and case fatality conditional on cell type and stage, based on meta-analysis of clinicalliterature (30). breast cancer - incidence by age, nulliparity, and age at birth of first child; progression and case fatality based on higher order Markov transi- tions among disease-free, localized recurrence, and metastatic states; a new version is under development (31). dementia- incidence based on age and sex; pro- gression based on duration since onset (32). osteoarthritis- incidence and progression based on combination of the 1986 post-censal Health and Activity Limitations Survey and expert consen- sus. s,t other cause mortality- based on age, sex and mari- tal status, in turn derived from vital statistics and census data. Functional status This area is incomplete. One approach draws di- rectly on the 1986 post-censal Health and Activity Limitations Survey to define 3 disability states (mild, moderate, and severe) with first-order Mark- ov transitions based on previous disability status (i.e., ignoring risk factors and diseases). The main approach still under development will use data from the Statistics Canada's new National Popu- lation Health Survey to generate a mapping of diseases into 8 dimensions of functional status: gross motor, dexterity, hearing, vision, speech, cog- nition, emotion and pain. These in turn form the basis for the McMaster multi-attribute value and utility scales (33).U For now, data from the 1990 Ontario Health Survey have been used to map diseases directly into health status values using a preliminary weighting function. Costs Only a lung cancer care module has been devel- oped, with treatment algorithms based on expert ' Tugwell, P. et. al. The population health impact of arthritis, a workshop report, mimeo, Analytical Studies Branch, Statistics Canada, Ottawa (1992). ' Chambers, L.W. et al. Physical disability among Canadians reporting musculoskeletal disease. mimeo, Statistics Canada, Ottawa (1991). Rapp. trimest. statist. sanit. mond., 47 (1994) medical oncology advice and 1988 unit costs (34,35) (Fig. 7). The heart disease treatments in Weinstein et al (27) have not been used. A detailed breast cancer costing module is under develop- ment along similar lines to that for lung cancer. Health The plan is to use multi-attribute value and utili_ty scales (33). Currently, preliminary weights for dis- ease based on Torrance et al. have been imple- mented.u It should be evident from these very brief sketches of the modules in PO HEM that the effort is seriously constrained by available data. The gen- eral philosophy has been to push on anyway. Where critical data are needed, we have sought the best possible approximation via expert opinion, consensus panel, or by approaching researchers with relevant data which can be re-analysed to gen- erate the required statistics. The cooperation of many researchers in this endeavour has been in- valuable. These processes or "laws of motion" are ap- plied year by year and individual by individual in POHEM. The simulation is exactly analogous to (but considerably more complex than) the process sketched for DFLE at the beginning of this section. In this way, complete synthetic biographies are built up for a representative sample of the popula- tion (more precisely, a steady-state birth cohort in the sense of a period life table). In effect, a com- plete longitudinal microdata set has bee_n imput~d or woven toge th er from diverse and partial descnp- tions of the dynamics of health and health-related processes. . One general caveat is in order. The processes m PO HEM as just sketched are quite disease-orient- ed. This is in large part by necessity. While we certainly wish for a much broader perspective on health and health-related processes, the vast major- ity of existing data and epidemiological studies are disease-based. Extensions, for example, in the directions of the roles of social determinants of health and disability outcomes, must await new data. Toward a coherent system of health statistics One method for achieving coherence among a set of statistical indicators is to define all the indicators as projections or transforms (in the mathematical sense) of the same underlying data. POHEM achieves this by creating a common underlying synthetic longitudinal data set representing the full u Torrance, G.W. et al. Provisional health index for the Ontario health survey. Final report to Statistics Canada of Project No. 44400900187, McMaster University Centre for Health Economies and Policy Analysis, Hamilton, Ontario (1992). Wld hlth statist. quart., 47 (1994) life cycles of a birth cohort. Each of these individu- al biographies, graphically represented in Fig. 1 by a "rectangle" (state space along the vertical axis, age along the horizontal) can be stacked (~th individuals going back into the page along a th1rd dimension). This results, with sorne poetic license, in a "data cube" like that shown in Fig. 2 (an image also drawn from the Template). This data cube provides a common (synthetic) microdata foundation and hence, a coherent basis for a variety of derived statistics and graphs, includ- ing the HLE family of health outcome measures. For example, the health status information along the bottom row of each individual's biography can be readily transformed into a conventional survival curve by the following algorithm: extract the bottom plane of the data cube; convert all non- zero entries to ones; sum across the columns to derive a vector oflife lengths (LLs); sort these life lengths in ascending order; graph the resulting distribution of LLs starting in the upper right of the survival curve and proceeding toward the lower left. Survival curves are, of course, very convention- al, and of decreasing relevance with the historie increase in chronic disease. It is more valuable to have indicators of how healthy people are while they are alive, as well as the expected distribution of their life lengths. To capture this notion, a some- what different algorithm can be applied. Specifical- ly, the health status information in the bottom plane of the data cube -the values between zero and one in the years when individuals are alive but in less than full health - is not thrown away. The algorithm again extracts the bottom plane of the data cube. But this time, the series of shades of grey is used to represent varying degrees of severity of illness. The conventional survival curve is constructed as before. But we now begin shading the area under the curve. This starts in the lower- left with a pure white area representing life-years spent in full health. Then for each of a series of threshold levels x ofless than full health (e.g., 0.9, 0.8, 0.7, ... ), the durations of intervals (there may be more than one) in each individual's biogra- phy where he or she was alive and had a health status index value above this threshold x are cumu- lated. Call these individuals' life lengths spent in health states valued better than x, or LL(x) (so that LL(O) = LL): sort the sample ofindividuals by these LL(x) durations; plot the contour (or equivalently the survival curve in health with a value score better than x); and shade the space between this and the previous contour with a slightly darker grey. Fig. 3 illustrates this process using another im- age from the Template. The top portion shows the bottom plane of the data cube turned up on its edge. A small set of individual health status vectors are shown, with the varying grey shade levels corre- sponding to varying states ofhealth (in the [ 0, 1 ] 167 Fig. 2 Fig. 3 Alive and in "perfect health" interval). The bottom portion shows the resulting survival curve plus contour plot assuming a num- ber of levels of disability. The area under the conventional survival curve is simply LE. The heavy "disability-free survival" line illustrates an alternative survival curve based on an arbitrary threshold dividing health states 168 between "healthy" and "disabled". The area un- der this curve is precisely DFLE. More generally, the ''weighted" area under the survival curve, with weights corresponding to the health status values represented by shades of grey, is Healthy Life Ex- pectancy (HLE). Of course, HLE could be comput- ed directly from the dat<l cube by summing over Rapp. trimest. statist. sanit. mond., 47 (1994) the entire bottom plane and then dividing by the sample size. HLE can be readily disaggregated, given the full data cube. For example, it could be computed for various subsets of the population - e.g., by gen- der, and whether the individual ever had a given disease like heart disease or arthritis. It could also be broken down by age interval, for example to derive the portion of the discrepancy between HLE and LE attributable to women suffering from ar- thritis after age 65 (analogous to a disaggregation of the "ali items" consumer priee index into com- ponents for food, transportation, etc.). Of course, PO HEM and this synthetic birth co- hort data cube are only part of the proposed system of health statistics. This system should also include a variety of longitudinal microdata sets on real individuals. These real data sets provide the basis for estimates of the various transition probability functions used as inputs to POHEM. For reasons already noted, however, these data sets will inevita- bly be partial- both in their range of variables and length of follow-up. Still, these data sets cease be- ing a hodgepodge and become coherent to the extent they are built with common concepts and definitions, and become systematically related by feeding into the construction ofPOHEM. Th us, a modellike PO HEM aids statistical coher- ence in two ways. First, it clarifies the interrelation- ships among diverse data sources by an explicit set of processes whereby empirical patterns derived from these data sources are woven together. Sec- ond, it creates a common synthetic core of longitu- dinal microdata from which a variety of health indi- cators, particularly the HLE family, are ali derived. Table 1 Summary results from a baseline POHEM simulation Tableau 1 Illustrations of POHEM-based analyses We tum in this penultimate section to a few brief illustrations of POHEM simulations. These results highlight population health status measures, chron- ic disease burdens, statistical coherence, health in- terventions, and health research applications. Burdens of chronic disease To begin, Table 1 shows summary results from a "baseline" POHEM simulation. The morbidity processes associated with 3 major diseases have been modelled: ischaemic heart disease, lung can- cer and osteoarthritis, though only the first 2 can be fatal. Each time an individual in the simulation suffers from one of these diseases, there is an age of onset. This may be followed by death from the disease. fndividuals may die of other causes, where only mortality has been explicitly modelled, based on rates from vital statistics. Table 1 shows, for example, that 4.2% offemales can expect to have an incident case oflung cancer, at an average age of 68.2 years. Most will not sur- vive, with 3.4% dying. Over half of ali individuals can expect to have sorne form of heart disease during their lifetimes, where this is fatal in about half the cases. Heart disease onset is about 1 0 years earlier for men than for women. Over three-quar- ters of ali women can expect to suffer from sorne form of osteoarthritis, beginning on average in their early 50s. The overall result is life expectan- cies of 80.0 and 73.9 years; but when account is taken of years lived but spent in less than full health due to the 3 major diseases explicitly mod- elled, healthy life expectancy is from 3 to 5 years less. Récapitulation des résultats d'une simulation POHEM de référence Females- Femmes Males- Hommes Average age- Population affected {%) - Average age- Population affected {%) - Age moyen Population atteinte{%) Age moyen Population attemte {%) Lung cancer onset - Cancer du poumon à son début 68.2 4.2 68.8 9.0 Lung cancer death - Décès par cancer du poumon 68.5 3.4 69.3 7.0 Heart disease onset- Cardiopathie à son début 74.8 55.4 64.4 53.2 Heart disease death - Décès par cardiopathie 80.5 25.2 73.5 29.3 Arthritis onset -Arthrose à son début 52.7 76.6 64.3 51.7 Other causes of death - Autres causes de décès 80.5 71.4 74.7 63.6 Lite expectancy (LE) - Espérance de vie (EV) 80.0 73.9 Healthy lite expectancy (HLE) - Espérance de vie en bonne santé (EVBS) 75.0 70.5 Wld hlth statist. quart., 47 (1994) 169 Fig. 4 Joint distribution of years with arthritis and years with heart disease (females) Distribution conjointe des années «avec arthrite•• et des années «avec cardiopathie•• (femmes) en 60 Q) "(!) c: c: ~ Q) :E "t;; c. 0 '6 ... ni (.) 1 -~ ni ~ Q) !Q Q) Cl) '6 t:: ni Q) ::I: • • , .. • • •• ........ •... · ... 1,~,. • •1• • ... • ••• c. • Arthritis (years)- Arthrite (années) 90 Figs. 4 & 5 show quite different ''views" or sets of results from the same POHEM simulation, fo- cusing on heart disease and arthritis.v Fig. 4shows a scatter plot of durations of time individuals spend burdened by each disease. Since the underlying simulation actually generated 100 000 cases, the plot only shows a random sub-sample of 2 500 cas- es. This scatter plot indicates a considerable amount of co-morbidity (the dots that are in the positive orthant and not on either of the axes), and suggests that more years are likely to be spent with arthritis than with heart disease. The very ability to construct the plot illustrates the microanalytic as- pect of the modei.w Fig. 5 summarizes the data represented by the scatter plot into a few average expected durations for the hypothetical birth cohort generated by this POHEM simulation. Given survival to age 15, as was assumed in this simulation, the figure shows an overall life expectancy of 65.0 years (i.e., to age 80.0). Of this total expected life length, 21.1 years or about one quarter can be expected to be spent with sorne form of arthritis, of which 4.0 years will v Berthelot,J-M. et al. Modeling the lifetime burden of arthritis using QALY-style measures, mimeo, Social and Economie Studies Division, Statistics Canada, Ottawa (1994). w The females with 80 years of !ife burdened by arthritis, or over 50 years with heart disease shown in Fig. 4 may strain credulity. They are, however, possible outcomes of the specification of the stochastic incidence, progression, and mortality processes. To the extent such individual !ife paths are considered implausible, they can be inspected in greater detail, possibly leading to changes in the way specifie processes are mode lied. Note that in cell-based or !ife table models, such !ife paths remain invisible. 170 be with more severe disabling arthritis, and 4.1 will be co-morbid with heart disease. 1.1 years can be expected with both disabling arthritis and heart disease. Looking at these disease burdens from another perspective, 5.8 years can be expected with heart disease, of which 4.1 will be co-morbid with sorne form of arthritis. Methodologically, these estimates of co-morbid health state expectancies (sojoum times in multi- variate health states, in stochastic process par- lance) are simply tabulated results from the "data cube" underlying a POHEM simulation. At the same time, they reflect quite detailed descriptions of the underlying disease processes, and by virtue of the microanalytic method they capture a consid- erable amount of the heterogeneity of the actual population. (These estimates would be practically infeasible using life table methods, not least be- cause of the complexity of the state space.) Substantively, considerable co-morbidity is evi- dent. However, by their nature, heart disease and arthritis are independent disease processes. More precisely, they are conditionally independent, since they are both functions of age. It is this corn- mon dependence on age that results in co-morbid- ity rates higher than might otherwise be expected. In tum, these results open the possibility of explor- ing whether or not there is any interaction between heart disease and arthritis. The hypothesis repre- sented by this POHEM simulation is that they are conditionally independent (on age). Ifwe were to observe significantly more or less co-morbidity in a cross-sectional survey for a given older age group than is generated by the simulation, this would be suggestive that the conditional independence as- sumption is wrong. Toward coherence in health statistics A central objective of POHEM has been to bring coherence to health statistics. One kind of coher- Fig. 5 Arthritis and heart disease: durations and co-morbidities (females) Arthrite et cardiopathie: durée et co-morbidité (femmes) 15 0 ~~~~~~"~~:! • Dlsabllng lltllrttls Arlhrtte lnVIlldante • Overall arthrllls- Totallltllrlle Lite expeclallcy 11 age 15 (65) Esp6rance de vie t 15 ans (65) 25 35 45 (wllh CHD) - (MC cardiopathie) (~) (17.1) (4.0) -(21.1) CHD- ClrdloPij';.~&5j~ (wlth lltllrllls)- (MC lllllr:';!J.:JJ 55 65 75 ao Durillon ln years - Ourdi en aniiMs ! i Rapp. trimest. statist. sanit. mond., 47 (1994) ence has just been illustrated- Table 1 andFigs 4 & 5 are coherent in so far as they are ali "projections" or views or aspects of the identical underlying data cube. A second kind of coherence arises when statis- tics are assembled from diverse sources into a sys- tematic framework which can then be used to gen- erate "data confrontations". Two different sources or approaches are used to estimate a given magni- tude, where in principle the results should be iden- tical. This provides a method for revealing and assessing the extent of overall error in the many steps of the statistical process. The example presented here focuses on the lung cancer module. Lung cancer incidence, pro- gression and case fatality have been simulated 3 different ways, essentially using 3 alternative ver- sions for the lung cancer module. The first and simplest uses only age/sex-specific mortality rates by cause, with lung cancer as one cause of death (the "mortality only" scenario). The data come from vital statistics and the census. They take no account of incidence and morbidity, so lung can- cer is effectively modelled as having an infinitesi- mally short morbid phase always followed by death, i.e., 100% case fatality. The next variant ("incidence and survival") explicitly distinguishes incidence and case fatality. Incidence is based on Canadian cancer registry data by age and sex. Incident cases are then disag- gregated by cell type and stage based on a special chart review study. Finally, disease progression and case fatality are explicitly modelled based on a detailed literature review and expert clinical judge- ment (35). The most detailed variant ("relative risk") builds on the second. It is the same except that incidence rates are adjusted for risk factor expo- sures. Data from a survey of residential dwellings are used to assign radon exposures, and data from the 1978 Canada Health Survey are used to gener- ate age profiles of cigarette smoking. Then a risk function from the epidemiologicalliterature (29) is used to scale each individual's risk of contracting lung cancer as a function of his or her persona! risk factor history. Given an incident case, lung cancer progression and case fatality are modelled in the same way as the second variant. Table 2shows the results ofPOHEM simulations with each of these 3 variants.x For each simulation, life expectancy should, in principle, be identical, as should average age at death from lung cancer, and this is, in fact, virtually so. However, there does seem to be a problem with incidence and case fatality. The product of these 2 rates, which is the X'fhese simulations are based on somewhat different mortality rate and cancer data that in the base case scenario in Table 1, so the results are not directly comparable. Wld hlth statist. quart., 47 (1994) Table 2 Lung cancer incidence and progression under three simulation scenariosa Tableau 2 Incidence et progression du cancer du poumon selon trois scénarios de simulationa Females- Femmes Lifetime incidence(%)- Incidence pendant la durée de la vie(%) Average age at diagnosis - Mortality only- Mortalité seulement 3.5 Age moyen lors du diagnostic 68.4 Average age at death - Age moyen au décès 68.4 Case fatality (%) - Taux de létalité(%) 100.0 Incidence x fatality (%)- Incidence x létalité(%) 3.5 Overalllife expectancy - Espérance de vie globale 79.6 Males - Hommes Lifetime incidence(%)- Incidence pendant la durée de la vie(%) 8.6 Average age at diagnosis - Age moyen lors du diagnostic 69.1 Average age at death - Age moyen au décès 69.1 Case fatality (%)- Taux de létalité(%) 100.0 Incidence x fatality (%)- Incidence x létalité(%) 8.6 Overall lite expectancy - Espérance de vie globale 73.2 Incidence and SUIVIVal- lnCidence et SUIVIe 3.3 67.9 68.1 83.2 2.8 79.6 9.4 68.8 69.2 81.1 7.6 73.3 Relative nsi<- Risque relatif 3.3 67.8 68.1 84.9 2.8 79.6 9.3 68.3 68.7 81.3 7.5 73.3 a Ali digits shown are significant with respect to Monte Carlo variab1lity -Tous les chiffres md1qués sont significatifs au regard de la variabilité de la méthode de Monte Carlo. proportion of all deaths attributable to lung can- cer, should also be identical across ail 3 scenarios. But the simulations suggest an inconsistency, with registry-based cancer incidence rates yielding lung cancerdeath rates thatare about0.7 to 1.1 percent- age points lower than those coming from mortality rates based directly on death certificates and the population census. Two factors that might account for these discrepancies are under-coverage oflung cancer cases by the cancer incidence registry on the one hand, and classification of sorne uncom- mon lung tumours with sites in the lung as "lung cancer" on death certificates. Such tumours should be excluded, since they are managed differently than "typical" lung cancers. One or other of these factors is suggested by the 3.5% mortality rate for females in the first scenario, which is greater than the 3.3% incidence rate from the incidence regis- try in the other two scenarios. 171 While the inconsistency between the vital statis- tics and cancer incidence registry data are clear for females, and this complex POHEM simulation is not necessary to reveal it, sorne kind of modelling would be required. Moreover, the inconsistency for males would not be revealed by such a simple comparison. In sum, this is an example of data confrontation from otherwise disparate data sourc- es that has been rendered feasible by the coherent structure ofPOHEM. Lung cancer treatment The lung cancer module is currently the most de- tailed in PO HEM, especially in so far as it includes a detailed description of care processes. This, in tum, has enabled the simulation of benefits and costs for specifie kinds of therapeutic interven- tions. As an example, Fig. 6 shows the standard "treatment schema" for Stage III and IV non-small celllung cancer. Based on detailed 1988 costs, this treatment schema and corresponding schema for other kinds of lung cancer suggest that the total direct costs of lung cancer treatment in Canada in 1988 were $325 million, with an estimated average five year cost per case of $21 000. This latter figure is an average, and ranged from $16 500 for a case of Stage IV non-small cell lung cancer (NSCLC) to $29 900 for limited small celllung cancer (SCLC). It is naturally difficult to obtain detailed data on the course of lung cancer in the absence of treat- Fig. 6 Treatment schema for stage Ill and IV non-small celllung cancer Schéma de traitement de cancer non microcellulaire du poumon, stades Ill et IV • 85% of Stage 111 A and 80% of Stage Ill B recelve radiotherapy. Staging tests Tests déterrn. du stade Terminal care Soins terminaux • 85 'k des malades au stade Ill A et 80% au Stade Ill B reçoivent une radloth6rapie. 172 ment. Nevertheless, given estimates of such "un- treated" survival curves, POHEM simulations sug- gest that compared to a base li ne of no lung cancer treatment at all, the estimated cost per life-year gained was about $11 000 for NSCLC and $19 600 for SCLC (35).Y POHEM has also been used to estimate the costs per life-year gained of several kinds of neo- adjuvant chemotherapy. For example, Evans et al. have estimated that the use of this chemotherapy as part of combined modality therapy involves costs between $2 900 and $5 000 per life-year gained.z Impacts of cholesterol-lowe ring interventions We tum next to estimated impacts of a qui te differ- ent kind of health intervention, cholesterollower- ing strategies (36). The starting point is Frank et al' s re-analysis of the Honolulu Heart Study data (37). They replicate the usual elevated mortality risk from heart disease in the case of high serum cholesterollevels at baseline. However, their analy- sis goes beyond the common heart disease end point to consider mortality from other causes. As shown in Table 3, their results indicate elevated mortality risk both for high and for low levels of serum cholesterol. One important corollary of these data is that it is not clear a priori that a cholesterol-lowering programme will be beneficiai for an entire population when ali causes of death are considered. Y Houle, C. et al. Évaluation des impacts et des coûts associés au cancer du poumon au Canada. ACFAS 1994: Colloque - Méthodes et applications de la statistique, Montréal, 1994. z Evans, W.K. et al. An estimate of the cost-effectiveness of combined modality therapy for Stage Illa/IIIb NSCLC (non- small celllung cancer) in Canada, presented at the American Society of Clinical Oncology Annual Meeting, Dallas, May, 1994. Table 3 Relative mortality risk by disease types other than heart disease and by level of cholesterol (males only) Tableau 3 Risque relatif de mortalité par type de maladies autres que les cardiopathies et par taux de cholestérolémie (hommes seulement) Cholesterol range (mg/dl)- Fourchette de la cholestérolémie (mg/dl) > 260 200-260 180-200 160-180 140-160 < 140 Cancers (including lung)- Cancers (y compris cancer du poumon) 0.9 1.0 1.0 1.0 1.3 2.0 Other circulatory diseases- Autres maladies cardio- vasculaires 1.5 1.2 1.0 1.0 1.5 1.5 Other diseases- Oivers 1.0 1.0 1.1 1.1 1.2 3.5 Rapp. trimest. statist. sanit. mond., 47 (1994) Algorithms to reflect these specifie relative risk patterns were added to the baseline disease inci- dence modules already in PO HEM. This is equiva- lent to the assumption that the observed relation- ships in Table 3 are causal, and not just associations. Then 3 PO HEM simulations were run -a base case and 2 health-promoting intervention scenarios. The first intervention was a dietary policy which had the effect of lowering everyone's total cho- lesterol, whatever it was, by 5%. The second added to this dietary policy selective prescription of cho- lesterol-lowering drugs for those with higher risks. This latter scenario is actually a bit complex be- cause it assumed drug prescription and use would depend on individual risk factor profiles as well as partial compliance by patients. The assumed im- pacts of both of these intervention scenarios on cholesterollevels follow the published literature in this area. Table 4 shows results for males. Everyone "de- parts" by age 75, either by one of 3 causes of death, or by surviving to age 75. Lowering cholesterol does lower death from heart disease. But the in- creased mortality from other causes, assuming the Honolulu Heart Study relative risks are causal, leaves the proportion surviving to age 75 under the first dietary intervention essentially unchanged. The second, more aggressive intervention results in almost a 10% reduction in the proportion of deaths from heart disease before age 75 (-1.28% against 14.82% in the base case). But this interven- tion also raises deaths from other causes, so that Table 4 Simulation results for a base case and two cholesterol- lowering scenarios (what males can expect at age 75) Tableau 4 Résultats de la simulation pour un cas de référence et deux scénarios d'abaissement de la cholestérolémie; à quoi peuvent s'attendre les sujets masculins à 75 ans Heart disease deaths - Décès par cardiopathie Lung cancer deaths - Décès par cancer du poumon Other-cause deaths- Autres causes de décès Survival ta age 75 - Survie jusqu'à 75 ans Total Base case- Cas de référence 14.82 5.14 32.33 47.71 100.00 Dietary intervention - 1 ntervenllon portant sur le régime alimentaire 14.17 5.19a 32.76 47.78a 100.00 D1et + Drug- Rég1me alimentaire+ médicament 13.54 5.41 32.76 48.29 100.00 • Ali numbers except these are stat1st1cally s1gnlf1cantly different from the base case with respect to monte carlo variability at a 95% confidence level.- Tous les autres chiffres accusent une différence statistiquement significative par rapport au cas de référence, compte tenu de la vanab1hté de la méthode de Monte Carlo au niveau de confiance de 95% Wld hlth statist. quart., 47 (1994) half of the reduction in heart disease deaths is matched by increases in other causes of mortality. It should be emphasized that these are not neces- sarily rigorous results (see Ref. (41) for alternative analyses). However, they highlight the importance for health policy of a comprehensive, population- based perspective and a concern with all end- points, not just those ofinterest to a specifie disease or research group. Explaining socio-economic status gradients in morta/ity In our final example, POHEM has been used to explore the sources of the clearly observed positive association in Canada of a variety of socio-econom- ic status (SES) measures and health. One possibili- ty is the observed inverse correlation of smoking with educational attainment. If smoking is more prevalent in lower SES groups, this factor alone might account for a very large portion of the ob- served gradient in mortality by SES. The base case scenario in PO HEM has smoking independent of any SES variables. (Smoking does depend on age and sex, and is correlated as ob- served in the 1978 Canada Health Survey with pri- or smoking, and other risk factors like obesity and hypertension.) To explore this question, an alternative scenar- io was constructed where smoking prevalences were adjusted to reflect observed patterns by edu- cational attainment (39). The question was then how much of a mortality gradient, using education- al attainment as a marker for SES, could be gener- ated sim ply by this observed correlation of SES and smoking. As points of comparison, 3 other scenari- os were also constructed: one where no one ever smoked, one where everyone was a heavy smoker from age 15 onward, and one where smoking was as in the base case but there was no death at all from lung cancer. The results, in terms ofHLE, are shown in Fig. 7. The bold horizontal lines in the two graphs represen t overall HLE (not LE) , given survival to age 25 and assuming the distribution of smoking patterns observed in Canada independent of SES- 70.6 years for men and 74.9 years for women. Next, the top two horizontallines represent hypothetical scenarios where either no one smoked at all, or smoking was as usual but no one ever contracted lung cancer. These two scenarios have very similar impacts, and raise HLE by 0.5 and 0.6 years respec- tively for women, and 0.9 and O. 7 years for men. They show that a fully successful "war against smoking" would have at least the same payoff in terms ofHLE gains as instantly eliminating alllung cancer mortality, taking account only of the effects of smoking on lung cancer and ischaemic heart disease (IHD), about half the tobacco-sensitive causes of death. At the other extreme, the bottom line repre- sents a hypothetical scenario where everyone was a 173 Fig. 7 Simulated healthy lite expectancies (years) for baseline, lung cancer (LC) deleted, and three modified smoking scenarios Espérances de vie en bonne santé (en années) simulées comme ligne de base, suppression des cancers du poumon (CP) et trois scénarios de tabagisme modifiés F1m111s- Femmes 76- ., 1==;;-;;;....-::-==="'=1 75.5 (no LC- pas de CP) -ti 7M 75.4 (no smoking- non fumeuses) ~ 75"""1!!!!!"'!!~p.;;---...;;;;;~ 74.9 1 74.7 i 74-_ - 73.8 (ali heavy- toutes grosses fumeuses) > 73- 72 :Il 71 .., c: ~ 1 70 ~ "' ~ 69 68 SES- SSE _____. Mlle - Hommes 71.5 (no smoking- non tumeurs) 1--------:::7-t 71.3 (no LC- pas de CP) 1--.... ---'1....--1 70.6 68.6 (ali heavy- tous gros tumeurs) SES- SSE _____. heavy smoker ail their lives. This heavy smoking scenario has a greater impact on men, lowering their HLE by 2.0 years, and that ofwomen by 1.1 years. Overall, the two smoking extremes account for a range of about 3 years in HLE for men, and almost 2 years for women. This smaller range for women is associated with their lower (inferred) average age-specifie incidences for lung cancer and IHD, holding smoking rates constant. Finally, the stepped line in the middle gives the scenario where each of three educational attain- ment groups smokes at the distribution of rates actually observed. For women, the larger effect is for the lowest educational group, lowering HLE by about 0.2 years, while for the middle and upper educational status groups, HLE is about 0.1 years greater than average. For men, we see a more pro- nounced gradient, with a range in HLE of O. 7 years from the lowest (70.2 years) to the highest (70.9 years) educational attainment group. The rough conclusion was that about one-fifth of the observed gradient in mortality by SES in Canada might be "explained" by the smoking-SES gradient working through lung cancer and IHD. Conclusion This article has described the origins and use of the POHEM microsimulation model in the context of a number of broad statistical and health science issues. 174 Numerical microsimulation models like PO- HEM can play severa! key roles in health statistics. One is to produce summary indicators of popula- tion health, by drawing on such generalizations of life expectancy as healthy life expectancy (HLE). In this role, POHEM can help remedy the imbal- ance in much of current health statistics where far more data are available on inputs and throughputs of the health care system than on population health outcomes. A second role for a simulation model like PO- HEM is to provide coherence to health informa- tion, in a context where many data series are little more than a hodgepodge. Such a coherent struc- ture, combined with sensitivity analyses, can also help guide development of new statistical sources. A third role for PO HEM is to support decision- making, particularlywith regard to resource alloca- tion for health-affecting interventions. POHEM can be used to model the impact of possible inter- ventions, and then generate quantitative estimates of costs and benefits, where benefits are measured in terms of health expectancy. A fourth role is as a tool for basic health science research. As shown in the co-morbidity and SES gradient examples above, POHEM can be used to obtain indirect, quantitative estimates regarding the importance or character ofvarious health-relat- ed phenomena. PO HEM is a work in progress. Among its weak- nesses are that it is still primarily disease-centred; it makes extraordinary demands on data; and it pre- sumes far more knowledge than currently exists regarding the causal pathways influencing human health. At the same time, this ambitious character may help spur the data development and basic health science research POHEM requires, thereby facilitating advances in new areas of health science and health policy analysis. Summary A variety of developments have come together to serve as both an impetus to and foundation for the develop- ment of a new POpulation HEalth Model (POHEM) at Statistics Canada. Part of the impetus iS statistical and derives from weaknesses in Canada's health statistics programme- particularly the lack of balance between information on health outcomes and health care re- source consumption, and the absence of a coherent statistical structure. The other major impetus is the need for rational processes for managing and allocating re- sources to improve the health of Canadians. The foun- dation for the development of this model has come from the revolution in computing. Dramatic improvements have opened up new methodological opportunities, particularly sophisticated simulation modelling and de- tailed analyses of large volumes of microdata. POHEM is designed to build on these increasingly powerful methods in order to meet health statistical and policy needs. At this time, POHEM is like a partially-completed Rapp. trimest. statist. sanit. mond., 47 (1994) building. This article reviews its motivation, the overall architectural plan, and the portion of the structure al- ready completed. A major portion of POHEM is devoted to the explicit modelling of chronic disease processes, using monte carlo microsimulation methods. The article concludes with illustrations of a few recent applications, focusing on the joint patterns of smoking, cholesterol and heart disease, osteoarthrit1s and lung cancer mor- bidity. Wh ile PO HEM has been developed in a Canadian context, work is under way to create a version that can be used in other countnes. Résumé Plusieurs faits nouveaux ont contribué à stimuler l'élabo- ration par Statistique Canada d'un modèle de santé des populations (POpulation Health Model: POHEM). L'ins- piration en est en partie d'origine statistique et trouve plus précisément sa source dans les lacunes du pro- gramme de statistiques sanitaires du Canada, en parti- culier le déséquilibre entre les données sur les résultats sanitaires d'une part, et les données sur la consomma- tion des ressources destinées aux soins de santé d'autre part et, en plus, l'absence d'une structure statis- tique cohérence. L'autre motivation principale est la nécessité de disposer d'un processus rationnel de gestion et de répartition des ressources en vue d'amé- liorer la santé des Canadiens C'est la révolution dans le domaine de l'informatique qui a permis de mettre au point ce modèle. Des améliorations spectaculaires ont ouvert la voie de nouvelles méthodologies, en particulier une modélisation perfectionnée par simulation et l'ana- lyse détaillée d'un grand volume de microdonnées. Le POHEM est conçu pour tirer parti de ces méthodes de plus en plus performantes afin de répondre aux besoins en matière de politiques et de statistiques sanitaires. A l'heure actuelle le POHEM évoque un bâtiment encore inachevé. Le présent article décrit sa motivation, le plan architectural global et la partie de la structure déjà achevée. Une part importante du POHEM est consa- crée à la modélisation explicite de l'évolution des mala- dies chroniques, avec utilisation des méthodes de microsimulation de Monte Carlo. L'article conclut en illustrant quelques applications récentes. Elles sont axées sur le tableau commun de la morbidité due au tabagisme, au cholestérol et aux cardiopathies, à l'arth- rose et au cancer du poumon. Le POHEM a été conçu dans un contexte canadien, mais des travaux sont en cours pour en créer une version susceptible d'être utilisée dans d'autres pays. References/Références 1. Wolfson, M.C. POHEM - A framework for understa.nding and modeling the health ofhuman populations, U.S. Bureau of the Census annual research conference, proceedings, Washington, DC, U.S. Bureau of the Census, 1992 pp. 261- 282. 2. Lomas, J. Report of the international conference on quality assurance and effectiveness in health care, centre for health economies and policy analysis, Hamilton (Ontario), McMaster University, 1989. 3. Radom, D.C. Setting health care priorities in Oregon. JAMA, 265 (17): 2218-2225 (1991). Wld hlth statist. quart., 47 (1994) 4. Wolfson, M.C. A system of health statistics - toward a new conceptual framework,Reoiew of income and wealth, 37 ( 1): 63- 80 (1991). 5. Wllk, M. Report of the national task force on health information. Ottawa, Statistics Canada, 1991. 6. Robine,J.M. Disabilityfree lifeexpectancy, generalindicators of the health of populations, scientific report. Québec, Conseil des affaires sociales et de la famille, 1986. 7. Wor1d Bank, World development report 1993: investing in health, Washington, D.C., World Bank & Oxford University Press, 1993. 8. Wolfson, M.C. A template for health information. World health statistics quarterly, 45(1): 109-113 (1992). (includes software diskette, also available from the author). Wolfson, M.C. Un modèle structurel pour l'information sur la san té. [résumé] Rapport trimestriel de statistiques sanitaires mondiales, 45(1): 112 (1992) (y compris progiciel micro- informatique sur disquette, également disponible auprès de l'auteur). 9. Evans, R.G. & Stoddart, G. Producing health, consuming health care. Social science and medicine, 31 (12): 1347-1363 (1990). 10. Marmot, M.G. Social inequalities in mortality: the social environment. In: Wilkinson, RG.(ed.). Class and health: research and longitudinal data. London, Tavistock Press, 1986, pp 21-33. 11. Power, C. et al. Class and health: the early yeaTS. London, Chapman and Hall, 1991. 12. Wolfson, M.C. et al. Career eamings and death - a longitudinal analysis of older Canadian men. Journal of gerontolog;y, 48(4): S167-Sl79 (1993). 13. Evans, R.G. etal. Wh y are somepeople healthy and others not? The determinants of the health of populations. New York, Aldine de Gruyter, 1994. 14. Ontario Premier's Council on Healtb Strategy Nurturing health: a Jramewurlt. on the determinants of health. Toronto, Healthy public policy committee, 1991. 15. Wllk, M.B. The concept of error in statistical and scientific work. In: Proceedings of the Census Bureau annual research conference, Washington, D.C., 1987. 16. Vaupel, J.W. & Yashin, A.I. Heterogeneity's ruses: sorne surprising effects of selection of population dynamics. American statistician, 39: 176-185 (1985). 17. Elandt, R.C. & Johnson, N.L. Survival models and data analysis. New York, John Wiley & Sons Inc., (1980). 18. Rogers, A. et al. Amultistate analysis of active !ife expecta.ncy. Public health reports, 104: 222-225 (1989). 19. Murray, C.J.L Quantirying the burden of disease: the technical basis for Disability-Adjusted Life Years. Bulletin of the World Health Organization, 72(3): 429-445 (1994). 20. Murray, C.J.L. & Lopez A.D. Quantifying disability: data, methods, and results.Bulletin of the WorldHealth Organization, 72(3): 481-494 (1994). 21. Wolfson, M.C. Divorce, homemaker pensions, and !ife cycle analysis. Population research and policy review, 8: 25-54 (1989). 22. Picot, G. Modeling the lifetime employment patterns of Canadians. In: Legare,J. et al. (eds), Thefamily in crisis: a populauon crisis 1 Proceedings of a colloquium organized by the Federation of Canadian Demographers and Sponsored by the R.uyal Society of Canada, Ottawa, The Royal Society of Canada, 1989. 23. Gentleman, J.F. et al. Smoothing procedures for simulated longitudinal microdata, Analytical Studies Branch Research Paper Series No. 32 Ottawa, Statistics Canada, 1990. 24. Miller, A.B. et al. The national study of breast cancer screening: protocol for a Canadian randomized controlled trial of screening for breast cancer in women. Clinicat and investigative medicine, 4: 227-258 (1981). 25. Abbott, R.D. & McGee, D. The Framingham study. An epidemiological invesUgation of cardiovascular disease: Section 37: The probability of developing certain cardiovascular diseases in eight yeaTS at specified values ofsome characteristics. U.S. Department of Health and Human Services, National Institutes of Health, Bethesda, MD. Publication No. 87-2284. 175 26. Glick, H. et al. A model for evaluating the cost- effectiveness of cholesterol-lowering treatment. Interna- tional journal of technology assessment in health care, 8 ( 4): 719- 734 (1992). University of Pennsylvania Leonard Davis Institute, 1990. 27. Weinstein, M.C.et al. Forecasting coronary heart disease incidence, mortality and cost: the coronary heart disease policy mode!. American journal of public health 77: 1417-1426 (1987) (and persona! communication). 28. Gentleman, J.F. et al. The development of staging data for use in the microsimulation of Jung cancer. Health reports 4(3): 251-268, (1992). 29. Whittemore, A.S. & McMillan, A. Lung cancer mortality among U.S. uranium miners: a reappraisal. Journal of the National Cancer lnstitute, 71:489-499 (1983). 30. Gentleman, J.F. et al. Microsimulation of health data: simulation of incidence, progression and mortality for Jung cancer. Proceedings of the statistical computing section of the A merican Statistical Association, 1991 Joint Statistical Meetings, Atlanta, Georgia, American Statistical Association, pp 60-69 (1991). 31. Will, B.P. et al. A mode! for estimating the costs and burdens ofbreast cancer diagnosis and treatment in Canada. Health reports 5(4): 399-408 (1993). 32. Forbes, W.F. &Barham,J.Aninvestigationoftheprroalenceand incidence of dementia in populations aged 65 years and over. 176 Research Report 2. Waterloo (Ontario), Prograrn in Gerontology, University of Waterloo, 1989. 33. Tôrrance, G.W. Utility approach to measuring health- related quality of!ife.Journal of chronic diseases40 (6): 593-600 (1987). 34. Evans, W.K. et al. Diagnostic and therapeutic approaches to Jung cancer in Canada and their costs (in review, 1994). 35. Evans, W.K. et al. Estimating the cost of Jung cancer diagnosis and treatment in Canada: the PO HEM mode! (in review, 1994). 36. Wolfson, M.C. & Berthelot, J-M. The burden of CHD and cholesterollevels: exploratory simulations using population health mode! (POHEM). Canadian journal of cardiology, Supp. D, May, 1993. 37. Frank, J. et al. Will lowering population levels of serum cholesterol affect total mortality? - expectations from the Honolulu heart program. journal of clinicat epidemiology 45(4): 203-204 (1992). 38. Jacobs, D. Models of effects of low blood cholesterol on the public health, implications for practice and policy. Circulation, 87(3): 1033-1036 (1993). 39. Wolfson, M.C. & Berthelot, J.-M. Accounting for socio- economic status gradients in health- an application of the PO HEM microsimulation mode!. In: Morgenstern, W. ( ed.), Modeling chronic disease morbidity, Berlin, Heidelberg, New York, Springer-Verlag, (forthcoming 1995). Rapp. trimest. statist. sanit. mond., 47 (1994)