Markov Chain Population Models in Medical Decision Making Gordon B. Hazen Min Huang July 2011 1 Abstract Markov models are widely employed in cost-effectiveness analysis of healthcare interventions. When modeling populations of non-interacting individuals, it is typical to formulate Markov models at the individual level and scale up the results to a fixed cohort or population by weighting by population size. However, we argue that this may not adequately capture population dynamics, as for example when an intervention extends life and thereby increases population size. We examine population-level Markov models of non-interacting individuals using techniques from the literature on stochastic networks. A key construct from this literature is the notion of population equilibrium. We also examine potentially relevant measures of health-related outcome, exploring how they differ from each other and from individual-level measures such as QALYs. As we illustrate, population modeling can provide a more refined picture of outcome. For instance, a beneficial intervention might increase population size but result on average in less healthy individuals. Further interesting consequences are first, that in an equilibrium analysis, the common procedure of weighting QALYs by a representative population size will usually be inappropriate; and second, that in contrast to common practice, there is no need for time discounting in an equilibrium analysis. Key words: Cost-effectiveness, Markov models, Equilibrium, Stochastic networks, Population-level Markov models, Population measures. 1 Introduction Markov models1,2,3 have long been utilized in medical decision making and health economics4, 5. They are designed to describe the course of a disease experienced by a patient in terms of mutually exclusive ‘health states’ and the transitions among them. They are widely employed in disease modeling, prognosis analyses, and in costeffectiveness analysis of healthcare interventions. A May 2011 PubMed search on “Markov” AND “Cost-Effective” for the preceding 10 years resulted in 1134 citations. To date, medical Markov models are usually formulated at the individual level. That is, the model follows progress of a given patient from an initial health state for either a fixed period of time or until death, both with and without the intervention in question. The analyst calculates both incremental cost and incremental individual benefit, the latter often in terms of health-related outcome measures6 such as life expectancy or qualityadjusted life years (QALYs)7,8,9. Often these results are scaled up to the level of a relevant population size by simply multiplying by a population size estimate, or choosing a convenient population level (e.g., 10,000 or 100,000). Equivalently, the Markov model may be applied to a cohort of independent patients of an appropriate representative size10. It is recognized that this naïve treatment of population dynamics is not adequate for infectious disease modeling.10 The position we take in the current paper is that it also may not adequately capture important population dynamics even for populations of noninteracting individuals. In particular, the impact of an intervention on the size of a relevant population may be overlooked, as for example when an intervention extends longevity and thereby increases population count. 2 Mathematical tools for analyzing populations of individuals are available in the operations research literature11,12,13 under the names stochastic networks and queuing networks. In this paper we adapt techniques from this literature to model populations of non-interacting individuals each of whose health progresses as a Markov chain. A key notion we adapt from this literature is the notion of long-run population equilibrium. Once one shifts the focus from an individual to a population, it is natural to ask what the relevant measures of health-related outcome are, how they relate to individual measures such as QALYs, and whether individual measures and population measures may conflict in evaluating an intervention. That is the second focus of this paper. An interesting result from this analysis is that for a population at equilibrium, the naive procedure of weighting QALYs by population size may give inappropriate results, and in fact the most natural measures of ongoing population health benefit and cost involve no time discounting whatsoever. This is in clear conflict with current practice for individual-level models. We begin in the following Section 2 by giving a synopsis of individual-level Markov cohort models. In Section 3 we introduce Markov population models, and in Section 4 we treat measures of health-related outcome. In Section 5 we re-analyze from a population perspective a previously published Markov cohort cost-effectiveness effort. 2 Individual-Level Markov Models Individual-level Markov models, also known as cohort models, have become quite popular approaches to assessments of cost-effectiveness and comparative effectiveness. In Figure 1 we give a simple example of such a model, with arrows denoting possible transitions between health states. Arrows are labeled by transition rates. More sophisticated versions of the model in Figure 1 have been used to evaluate glycoprotein 2b/3a antagonists for acute coronary syndrome.14 Figure 1. An acute coronary individual-level Markov model depicting health risks associated with ischemic heart disease (IHD). Risks include myocardial infarction (MI) occurring at annual rate , and death, occurring at annual rate 0 from the IHD state and at a greater rate 1 from the Post MI state. In a time-homogeneous Markov model such as this one, a transition rate from health state j to health state k can in general be denoted by a parameter λjk . An annual rate jk of transition from j to k induces10 an approximate probability jkdt of transition from state j to state k in a short time interval of length dt.1 Individual-level Markov models are solved by mathematically following a cohort of individuals through the model’s health states for a fixed period of time. One postulates a 1 The exact formula for arbitrary intervals t is pjk = 1 exp(jkt). 3 initial proportion pj(0) of the cohort in state j at time t = 0. Calculating forward in time, if one knows the proportions pj(t) for all j for some time epoch t, then one may calculate the proportions pj(t+dt) in the next time epoch t+dt via the rate equations: pj(t+dt) = pj(t) + Rate into j Rate out of j = pj(t) + i pi(t) probability of transition ij k pj(t) probability of transition jk = pj(t) + i pi (t )ij dt k p j (t ) jk dt . all states j (1) In the limit as dt 0, these become a system of differential equations d p j (t ) i pi (t )ij k p jk (t ) jk dt all states j. It is not widely recognized in the medical community that the equations (1) implement the most basic method for solution of ordinary differential equations, the Euler method.15 For our acute coronary model of Figure 1, if one starts 100% of the cohort in state IHD at t = 0, then this solution method gives the proportions pj(t) for j = IHD and j = PostMI as shown in Figure 2(a). If quality coefficients Qj for each health state j are available, then one may calculate quality-adjusted life years (QALYs) per patient. For instance, when QIHD = 0.9 and QPostMI = 0.7, the result is 6.62 QALYs per patient (undiscounted). In general, the relevant formula, including discounting, is QALY = 0 e rt j Q j p j (t )dt . Here QALYs accrued at time t are j Q j p j (t )dt , and these are discounted to the present by the factor ert corresponding to discount rate r. The result is summed by integrating from t = 0 to . Usually this integral would be approximated by a finite sum over a finite duration, but for mathematical simplicity in our development we use the integral form.2 2 In addition to continuously accrued quality as described here, is sometimes desired to include one-time quality tolls. We cover this in Appendix 1. 4 (a) (b) 1 1 0.8 0.8 p IHD( t ) p IHD( t ) 0.6 0.6 p PostM I( t )0.4 p PostM I( t )0.4 0.2 0.2 0 0 0 5 10 15 20 0 5 10 t t (yr) (yr) 15 20 Figure 2. The proportions pj(t) of a starting cohort in the Markov model of Figure 1 who at time t are in state j = IHD or j = PostMI, shown at times t from 0 to 20 yr, obtained via the rate equations (1). Input parameters were 0 = 0.04/yr, = 0.12/yr, 1 = 0.4/yr for (a), and in (b), the Post-MI mortality rate 1 = 0.1 is decreased by 75% and the MI incidence rate is decreased by 50%. Now consider, in our acute coronary model, a hypothetical intervention (such as glycoprotein 2b/3a antagonists) that not only lowers the mortality rate 1 post MI by 75%, but also lowers the incidence of MI by 50%. The results of a cohort analysis are shown in Figure 2(b). Lowering incidence and mortality seems obviously beneficial, and one may quantify this benefit by calculating improvement in QALYs under the intervention. The result, calculated in the same way, is 10.28 QALYs per patient, an improvement of 3.66 QALYs. The usual method for estimating the effect of an intervention on a population is to scale up cohort results by multiplying by population size. For instance, the 3.66 QALY/patient net benefit just discussed would allegedly scale up to a benefit of 366,000 QALYs in a population of 100,000. However, this scaling-up procedure does not capture dynamical population effects. An actual population will renew itself continually as individuals enter and later depart, and an intervention may change these dynamics. For example, in the model of Figure 1, individuals may enter by developing IHD, and depart by death. How can we estimate the effect of an intervention on such dynamic populations? One approach that suggests itself intuitively is to assume that a renewing population is equivalent to a cohort of individuals, each of whom re-enter the population immediately after death. We call this the closed routing process corresponding to the original Markov model. For the model of Figure 1, the corresponding closed routing process is shown in Figure 3. Figure 3. The closed routing process corresponding to the cohort model of Figure 1. Departures by death from either state IHD or Post MI immediately re-enter the IHD state. Population dynamics for Figure 3 can be calculated using the same rate equations (1) as before, but with transition rates suitably modified to include re-entry transitions. The 5 results with and without intervention are shown in Figure 4. In each case, the population reaches equilibrium. The standard method1 for calculating the equilibrium probabilities j = lim p j (t ) is to set Rate into j Rate out of j = 0 in equations (1) and solve, that is, t solve the so called balance equations = 1. i i ij j j k j jk =0 for all states j (2) The results in our acute coronary model are IHD = 1PostMI = 77% without intervention, and IHD = 1PostMI = 62.5% with intervention. (a) (b) 1 1 0.8 p IHD( t ) 0.8 p IHD( t ) 0.6 0.6 p PostM I( t )0.4 p PostM I( t )0.4 0.2 0.2 0 0 5 10 15 20 0 0 5 10 t t (yr) (yr) 15 20 Figure 4. Dynamics for the closed routing process described in Figure 3. Parameter values from Figure 2 are used. The population without intervention is shown in (a) and with intervention in (b). These population results are unsettling: A clearly beneficial intervention appears to have made the population worse off in equilibrium. The equilibrium proportion of the population in the undesirable Post MI state is 23% without intervention and 37.5% with intervention. This is because the intervention extends the lifetime of Post MI patients, increasing their prevalence in the population in spite of the lower incidence of MI. Incorporating the quality coefficients QIHD = 0.9, QPostMI = 0.7, one may calculate the average QALYs accrued per year at equilibrium using AQ/yr = j jQ j . The results are 0.854 per patient per year pre-intervention and 0.825 post-intervention, a decline of 0.029 QALYs per patient per year. Is the re-entry assumption we used to construct the closed routing process the source of this anomaly, that is, does the closed routing process not adequately represent population dynamics? We shall see below that on the contrary, formulating a closed routing process is a valid approach. The problem is rather that AQ/yr is not the “right” measure of benefit for a population. In section 4 below we shall address what we think the right measures should be, and discuss implications. We first turn, however, in the next section to the existing theory on Markov population processes. Along the way we show that this theory justifies the re-entry assumption for constructing the closed routing process. 6 3 Markov Population Models The theory for treating Markov chain population models has been available in the stochastic networks literature for some time.11,13 We review that theory here and add our own result on closed routing processes. We assume we have an individual-level Markov model with transition rates jk in the model from health state j to health state k. Because individuals may enter a population as well as depart it, we need to include arrival rates j to each health state j, representing individuals who arrive in that health state from outside of the population. For example, in the individual level model of Figure 1, we add entrance arrows with rates 0 and 1 to the states IHD and Post MI respectively, representing individual who arrive with ischemic heart disease, and individuals who arrive already having a myocardial infarction. The result is the diagram in Figure 5, which is referred to as the open routing process. We have omitted the Dead states, which represent exit from the population, and therefore are not explicitly considered, although the individual exit rates j remain. 0 1 IHD 0 (MI) Post MI 1 Figure 5. The model of Figure 1 augmented by arrival rates 0,1 to the two possible health states. The result is called the open routing process for the individual-level Markov model of Figure 1. To summarize, the open routing process has the same transition rates jk from j to k as the individual-level model, has exit rates j from state j equal to the mortality rates in the individual-level model, and in addition has arrival rates j to each state j. Just as we can solve for the equilibrium probabilities in the closed routing process of Figure 3 using balance equations (2), we can also set up and solve balance equations for the open routing process. These equations are based on the same principle Rate into j Rate out of j = 0 as in (2). However, instead of equilibrium probabilities j, the new balance equations involve the mean number j of individuals in health state j. The revised balance equations are as follows. j i i ij j k jk j = 0. for all states j (3) For instance, suppose in the model of Figure 5 that we take 0 = 1,000/yr, that is, one thousand new cases of ischemic heart disease per year; and 1 = 0, that is, no arrivals to the population already having MI. At equilibrium by solving (3), the mean numbers 0,1in states IHD, Post MI respectively are (0,1) = (6,250, 1,875) without intervention, and (0,1) = (10,000, 6,000) with intervention. The proportions in the two health states are the same as calculated from the balance equations (2), but the equilibrium total number of individuals alive has increased on average from 0+1 = 8,125 without intervention to 0+1 = 16,000 with intervention, a net increase of 7,875. Using the same quality coefficients Q0 = 0.9, Q1 = 0.7, we may calculate population total QALYs accrued per year at equilibrium via the equation 7 TQ/yr = j jQ j . The result is 6,937.5/yr accrued without intervention and 13,200/yr accrued with intervention, an increase of 6,262.5/yr. This result confirms our intuition that the intervention has to be beneficial. The result is in this sense is consistent with the cohort analysis presented in Section 2, a fact we shall comment on below. We have yet to formally define what we mean in general by Markov population model. By this we follow the stochastic networks literature and formulate a Markov model whose states are vectors n = (nj), where nj is a count of the number of individuals in the individual-level health state j at a given time. In this model, transitions correspond to the possible transitions of the individual-level model, with additional transitions added to represent individuals entering the population. For example, for the individual-level model of Figure 1, the corresponding Markov population model would have states consisting of the infinite number of possible vectors n = (nIHD,nPostMI), where nIHD is the number of individuals in health state IHD, and nPostMI is the number of individuals in health state Post MI. Deaths in either health state would be represented by transitions that decrease either nIHD or nPostMI by one. A transition from IHD to Post MI would induce a population transition that decreases nIHD by one and increases nPostMI by one. In addition, individuals entering the population would increase either nIHD or nPostMI by one. Such population models are dealt with in the general theory of stochastic networks,11,13 to which the reader is referred for details. In particular, what we have described are known as Jackson networks, which constitute an important branch of queuing theory.12 The models we construct here are specific kinds of Jackson networks that in queuing terminology would be known as networks of M/M/ queues. The following is a basic result from this literature (e.g., see Serfozo11 Example 1.29), and provides a justification for our assertion above that the solutions j to the balance equations (3) are equal to the equilibrium mean number of occupants of state j. Theorem 1: In a Markov population model with health states j, the counts nj of individuals in health states j are, at equilibrium, independent Poisson variables with means j given by the solution to the balance equations (3), that is, at equilibrium P(nj = b) = bj b! e j b = 0,1,2,… We now turn to the validity of the re-entry assumption introduced for the closed routing process discussed in the previous section (Figure 3). Given an open routing process such as Figure 5, we construct a corresponding closed routing process by routing all departing transitions back to state j with probability proportional to the entry ratej to state j. In other words, letting = j j be the total arrival rate, we define the modified transition rates 8 jk jk j j and let the closed routing process be the process with transition rates jk . Note that if entries to the population occur only at a particular state, say state j = 0, it then follows that 0 = , all other j are zero, and the only modification to the transition rates is that all departures are routed to state 0 at rate j 0 j 0 j . This is precisely how the closed routing process of Figure 3 was constructed. Theorem 2: Consider a Markov population model with a specified open routing process and a corresponding closed routing process as described above. Let (j) be the equilibrium means of the population model calculated by solving the balance equations (3) of the open routing process, and let (j) be the equilibrium probabilities for the closed routing process calculated by solving the balance equations (2) with transition rates jk . Then (a) the equilibrium expected proportion of the population in state j is (b) j j ; j j (c) j j where = j ; j j j j j is the equilibrium average departure rate. We are unaware of any previous version of this result, and provide a proof in the appendix. Taken together, parts (a) and (b) of the theorem indicate that the re-entry assumption used to construct the closed routing process is valid, because the solution (j) to the corresponding balance equations (2) do in fact constitute the equilibrium expected proportion of the population in state j. Parts (b) and (c) of the theorem indicate how one can move between the equilibrium of the closed routing process and the equilibrium of the open routing process. One can solve either balance equation (2) or (3) and then directly recover the solution to the other. 4 Measures of Benefit Suppose as above that we have an individual-level Markov model with quality coefficients Qj for health state j, whose solution by cohort analysis yields quality-adjusted life years QALYj for individuals beginning in state j. Suppose also that we formulate and solve a Markov population model to obtain equilibrium mean numbers j of individuals in health state j, equilibrium mean total population size = jj and equilibrium 9 proportions j = j/ of the population in health state j. In this situation, what would be reasonable measures of benefit, and how do such measures compare? As we have noted, the conventional approach to this question is to scale up the cohort analysis benefit QALY0 beginning in some distinguished initial health state j = 0 by multiplying by a population size estimate ̂ . If we use the equilibrium population size as the estimate, then we obtain the measure Population weighted QALY: PWQ = QALY0. One may object that at equilibrium there may be many individuals in health states j other than j = 0. It would therefore be more reasonable to weight the mean number j of such individuals by their specific QALY beginning in state j, thereby obtaining Total lifetime QALY: TLQ = j j QALY j . The corresponding average across the equilibrium population would be Average Lifetime QALY: ALQ = j j QALY j . One may also examine not lifetime QALY, but only expected QALY accrued in one year at equilibrium, thereby obtaining Total QALY per year: TQ/yr = j jQ j . The corresponding average across the equilibrium population would be Average QALY per year: AQ/yr = j jQ j . Note that because j = j, the average and total measures are proportional: TLQ = ALQ TQ/yr = AQ/yr. All of these measures assume a population at equilibrium, but one may consider nonequilibrium measures as well. In particular, suppose we can compute the expected number Enj(t) of individuals in health state j at every future time t. If we discount future QALYs at rate r, then total infinite-horizon discounted expected QALY would be given as follows. Discounted total population QALY: DTQ e rt j Q j En j (t )dt . 0 10 Please note that we list these measures only as ones that might seem reasonable at first glance. In fact, some may not be so reasonable after all. We know, for instance, from the example in Section 2 that AQ/yr does not serve by itself as a reasonable measure of benefit. Note also that if instead of quality coefficients Qj we had cost rates Cj for each health state j, then we could define measures for population cost analogously. That is, we could define TLC, ALC, TC/yr, AC/yr, DTC just as above. The discussion below would apply to these measures as well. Of course, the listed measures may conflict in ranking interventions. The “average” measures can be particularly misleading, as they do not account for intervention-induced increases or decreases in population size, as we saw in Section 3 for AQ/yr. It is not difficult to construct examples in which ALQ yields inappropriate conclusions as well. We therefore restrict our attention to the remaining “total” measures. Among the total measures, DTQ seems the most complete, as it accounts for impacts on the current and all future generations and not on just one generation, like TLQ, or in one year like TQ/yr. Unfortunately, DTQ may be more bothersome to compute. This difficulty is ameliorated by the following result. Theorem 3: DTQ and TQ/yr are proportional and therefore essentially equivalent when the population is initially in equilibrium. Specifically, if E n j (t ) j for all t ≥ 0, then DTQ = 1 TQ / yr . r In this result the reciprocal 1/r of the discount rate is equal to the discounted value 0 e rt dt of an infinite-horizon lifetime. The result therefore takes the form DTQ = (discounted number of years)(total QALYs per year). This relationship to DTQ gives TQ/yr a conceptual advantage as a measure of benefit over the other total measure TLQ. Another advantage obtains due to the following result. Theorem 4: In a Markov population model with entry rates j into health states j we have TQ/yr = QALY j j j where QALYj is calculated with zero discounting. The import of this result obtains when entries to the population can only happen into a distinguished state j = 0. Then all other j are zero, and we have 11 TQ/yr = 0QALY0. In this case, suppose the intervention in question alters QALY0 and not the entry rate 0 (as would usually be true unless the individual level were not adequate for modeling the intervention). Then the equilibrium population measure TQ/yr and the individual-level measure QALY0 (undiscounted) would evaluate the intervention identically (either both would indicate it is desirable or neither would). Note, however, that population weighted QALY might conflict with this evaluation, since it is given by PWQ = QALY0 and an intervention might change both QALY0 and the equilibrium mean population size . Similarly, total lifetime QALY, given by TLQ = j j QALY j , might conflict with TQ/yr. However, an intervention that alters only the quality coefficients Qj would leave the equilibrium population mean unaltered, so PWQ, QALY0 and TQ/yr would evaluate this intervention identically. We summarize as follows. Corollary: Consider a Markov population model with only one entry state j = 0. Then: (a) The desirability of an intervention that does not alter the entry rate 0 is evaluated identically by TQ/yr and by undiscounted QALY0. (b) The desirability of an intervention that alters only the quality coefficients Qj is evaluated identically by TQ/yr, by undiscounted PWQ, and by undiscounted QALY0. On the other hand, it is quite easy to devise a simple example where TQ/yr conflicts with TLQ and PWQ, even with the latter two undiscounted. Example: Consider a Markov population model with only one state 0 having quality coefficient Q0, entry rate and departure rate . It is easy to check that with discount rate r, QALY0 = Q0 r = / TQ/yr = Q0 = Q0/ TLQ = PWQ = QALY0 = Q0 . r Consider an intervention that extends life () at reduced quality (Qo), leaving Q0/ constant. Then TQ/yr remains unchanged, but TLQ and PWQ both increase. By altering the changes in Q0 and slightly, we could make TQ/yr decrease slightly while TLQ and PWQ still both increase. To summarize the results of this section, suppose it is desired to assess the benefit of an intervention on an existing population, not considering costs for the time being. For 12 simplicity, suppose entry to this population occurs at rate 0 (not influenced by the intervention) only at a distinguished state 0. In our view, the “right” approach would be to adopt an equilibrium population perspective as captured by DTQ or its more easily computed relative TQ/yr. However, we can imagine analysts unaware of our perspective proceeding in alternate ways. Suppose in particular that: 1. Analyst A computes increase in discounted QALY0, perhaps scaling it up to a representative population size. 2. Analyst B understands that the intervention may change equilibrium population size . She estimates population entry rate 0, and computes increase in PWQ = QALY0. 3. Analyst C forgets to think about population issues, and doesn’t believe in discounting health benefits. He computes increase in QALY0 (undiscounted). Which analyst is guaranteed to always produce the “right” recommendation concerning the desirability of the intervention? If “right” means consistent with TQ/yr, then ironically only the naïve Analyst C will do so (Corollary (a) above). The take-home point is that seemingly reasonable ways to extend individual-level analyses to the population level may in fact yield worse results than doing nothing at all. Of course, should there be more than one entry point to the population, then none of the analysts above are guaranteed to get the “right” answer, even analyst B who might be smart enough to replace PWQ by TLQ. Cost-Effectiveness As already noted, all of the mentioned measures have cost analogs, and it is interesting to consider how cost-effectiveness analysis might be accomplished in this setting. Suppose we want to compare incremental discounted total QALYs, which we denote DTQ with its analog incremental discounted total cost, or DTC. In a population initially at equilibrium, we would have by Theorem 3, DTQ = 1 TQ / yr rQ DTC = 1 TC / yr , rC where rC and rQ are the discount rates for costs and life years. The cost-effectiveness ratio would be given by DTC 1 rC TC / yr . DTQ 1 rQ TQ / yr Therefore, if costs and QALYs are discounted at the same rate, as is often recommended,16 then the cost-effectiveness ratio reduces to TC/yr / TQ/yr, a figure that is independent of the common discount rate. Therefore, for populations initially at equilibrium, the common discount rate for costs and life-years is irrelevant – all one needs to do is compare TC/yr with TQ/yr. 13 For the other commonly used assumption that life-years are not discounted at all (rQ = 0), then DTQ is infinite and DTQ is not even defined, so equilibrium cost-effectiveness analysis is not possible except in a limiting sense if one adopts the implausible assumption that both costs and life years are discounted at a common vanishing rate. 5 An Illustrative Example In this section we apply the population methods of this paper to a decision analysis by Col and colleagues17 of tamoxifen use to prevent breast cancer. This section also illustrates how one may calculate the short-term effect of an intervention on an existing equilibrium population, a point we have not discussed above. Tamoxifen, an estrogen agonist/antagonist, is an effective therapy against established breast cancer. There is also evidence that it can reduce breast cancer incidence, but unfortunately, its use can produce life-threatening side effects including endometrial cancer and vascular events. The question is whether the benefit of its prophylactic use in healthy women is worth the associated risks. For simplicity here, we examine only benefits and not costs, and the only side effect we include is endometrial cancer. We modify the Col model by using cure rate models18,19 for breast cancer survival (parameters pb, b) and endometrial cancer survival (parameters pe, e). These are shown in Figure 6, formulated as stochastic trees.20,21,22 These two factors are assumed to be independent. For background mortality, we use a constant mortality rate μ0 equal to the reciprocal of life expectancy of a 50-year-old woman. (Compared to a more accurate Gompertz mortality model, the bias for total population counts in equilibrium is 2% and the bias for the outcome measure TQ/yr is 3% see Huang's Ph.D. thesis for details23). No Disease No Disease b pb Cured Breast Ca (a) 1-pb e b pe Cured Endometrial Ca Death Not Cured e Death Not Cured (b) 1pe Figure 6: Cure models depicting incidence and survival for breast cancer (a) and endometrial cancer (b). Unlike all previous diagrams, this diagram has instantaneous states, namely breast cancer, and endometrial cancer. Straight arrows emanating from these state represent chance events and parameters associated with them are probabilities. Continuous transitions in this diagram are shown by wavy arrows, and the parameters adjacent to the wavy arrows are transition rates. See Hazen.20, 21,22 In each of the factors of Figure 6, there are only three nonfatal states that persist in time, namely ‘No disease’, ‘Cured’ and ‘Not Cured’ (the ‘Breast Cancer’ and ‘Endometrial Cancer’ states are mathematically instantaneous and do not affect the analysis). This produces 3× 3 = 9 nonfatal state combinations. Using Cykert24 and Jones25 we added quality coefficients Qjk for each of these 9 states to the Col analysis. These are presented in Table 1. 14 Table 1. The nine nonfatal state combinations in our replication of the Col et al analysis, with quality coefficients Qjk specified for each state combination j,k. Endo Ca Qjk No Ca Cured Not Cured No Ca 1 0.81 0.39 Cured 0.81 0.656 0.316 Not Cured 0.39 0.316 0.152 Br Ca We conduct our analysis for women in only one of Col’s categories, namely, those who are at fourfold higher and twofold higher risk for developing breast cancer and endometrial cancer, respectively, as might be expected for women with family histories of breast cancer, who constitute an estimated 4.8% of women aged 35-64.26 We assume this population of women begins tamoxifen treatment in the vicinity of age 50. The values of parameters are listed in Table 2. Following Col, we set the incidence rate of breast cancer without tamoxifen to be 0.0086/year, corresponding to a five-year risk of approximately 4.19%, four times higher than average. For simplicity here we depart from Col and presume the effects of tamoxifen on breast and endometrial cancer persist through the patient’s lifetime. Women enter this population at rate 0 in distinguished state 00 = (No Diseaase, No Disease). To give an estimate of 0, we observe that in the year 2010 in the U.S., there were 2.3 million women of age 49 who would enter the 50+ age group in 201127, of whom approximately 4.8% or 110,000 have a family history of breast cancer, and would hypothetically begin tamoxifen. For expositional simplicity, and recognizing the roughness of this estimate, we round to0 = 100,000/yr. Table 2: Risk rate, risk ratio, probability and mortality rate parameters in the Col et al. model17. The risk rates and risk ratios are estimated from the studies by Fisher et al. 28 that Col et al have cited. The probabilities and mortality rates are estimated based on SEER data. The parameter μ0 is estimated such that the average survival time of 50-year-old women is 1/ μ0. Variable λb0 RRb λe0 RRe pb μb pe μe μ0 Description incidence rate of breast cancer without tamoxifen risk ratio of breast cancer with tamoxifen incidence rate of endometrial cancer without tamoxifen risk ratio of endometrial cancer with tamoxifen probability of cure for breast cancer mortality rate of breast cancer if not cured probability of cure for endometrial cancer mortality rate of endometrial cancer if not cured Background mortality Value 0.0086/yr 0.4936 0.00152/yr 4.0132 0.5631 0.0996/yr 0.9019 0.3159/yr 0.03118/yr Results We conduct our analyses without discounting future life years. Table 3 displays the individual-level QALY results with and without tamoxifen treatment. Here QALYjk is expected QALYs with or without tamoxifen for an individual beginning in state combination j,k. The key figures that would be used in a cohort analysis are QALY00 with and without tamoxifen (29.19 yr and 28.79 yr) and the incremental QALY00 = 0.405 yr = 4.9 mo. (This compares to the 3.8 mo. gain in pure life expectancy reported by Col. 15 Table 4 displays equilibrium population counts jk in state combination j,k as well as equilibrium total population count = jkjk, both with and without tamoxifen. As expected, the equilibrium incremental effect of tamoxifen is to decrease breast cancer prevalence and increase endometrial cancer prevalence. The overall increase in equilibrium population size ( = 46.4 thousand) is achieved not because the number never acquiring cancer increases (it actually decreases by 14.4 thousand), but because cases of cured endometrial cancer increase more than cases of cured breast cancer decrease (293.4 thousand versus 239 thousand). It is true, however, that the equilibrium number of cancer-free women (no cancer or cured) does increase by 72.2 thousand. These ambivalent statistics are reflected in the corresponding equilibrium population measures in Table 5, where we see that TQ/yr increases by 40.5 thousand, whereas AQ/yr actually goes down slightly. Much as in the example of Section 2, the intervention keeps more individuals alive but they tend to be in less healthy states. Recall that according to the corollary to Theorem 4, incremental TQ/yr must agree in sign with incremental (undiscounted) QALY00, as it does here. We also include TLQ and ALQ in Table 5. These incremental values are in this case consistent with incremental TQ/yr, although in general they need not be. Table 3. Individual-level QALY benefits of tamoxifen. QALYjk is the expected QALYs accrued by an individual beginning in state combination j,k. No Tamoxifen Endo Ca QALYjk No Ca Br Ca Tamoxifen Endo Ca QALYjk No Ca Cured Not Cured 23.58 1.00 No Ca 29.19 24.65 1.00 No Ca 0.405 1.07 0.004 Cured 25.66 21.04 0.81 Cured 24.87 21.04 0.81 Cured -0.791 0 0 2.73 0.32 2.73 0.32 Not Cured -0.032 0 0 3.36 Not Cured 3.33 Cured Not Cured QALYjk No Ca No Ca 28.79 Not Cured Cured Not Cured Incremental Endo Ca Table 4. Equilibrium population levels with and without tamoxifen. ji is the equilibrium mean population count (in 1000s) in state combination j,k. jk No Tamoxifen Endo Ca No Ca 2423.7 84.4 Not Cured 0.833 2508.9 Cured 419.4 33.9 0.16 453.5 Not Cured 60.1 2.8 0.022 62.9 2903.2 121.1 1.02 3025 (1000s) No Ca Br Ca Cured Tamoxifen Endo Ca jk No Ca 2409.3 377.8 Not Cured 3.4 2790.5 Cured 180.5 66.0 0.279 246.8 Not Cured 28.4 6.0 2618.2 449.8 (1000s) No Ca Cured Incremental Endo Ca jk (1000s) No Ca Cured Not Cured 2.567 281.6 No Ca -14.4 293.4 Cured -238.9 32.1 0.119 -206.7 0.043 34.4 Not Cured -31.7 3.2 0.0206 -28.5 3.72 3072 -285 328.7 2.71 46.4 Table 5. Comparison of equilibrium population measures for tamoxifen intervention. No Tamoxifen Tamoxifen TQ/yr (1000s) AQ/yr TLQ (1000s) ALQ 2878.7 0.9515 83448.6 27.48 2919.2 0.9504 85637.3 27.88 Increase 40.5 -0.0012 2188.7 0.40 Short-term results The results in Table 5 are long-term equilibrium effects, but it could also be useful to ask what the immediate effects of this intervention would be, that is, the effect on the current population. It is reasonable to expect that the current population is likely to be in equilibrium, so that the no-tamoxifen equilibrium population counts in Table 4 match the 16 current population of high-risk women not using tamoxifen prophylactically. One can therefore estimate what the short-term effect would be of putting these women on tamoxifen by combining the tamoxifen and no-tamoxifen QALY estimates from Table 3 with these counts to form TLQ = jkjkQALYjk and its counterpart ALQ = TLQ/. These results are shown in Table 6, and indicate a beneficial effect. Notice that the incremental values of the measures TQ/yr = jkjkQjk and AQ/yr = (TQ/yr)/ are zero because we are fixing the mean counts jk at the current population level, and the tamoxifen intervention does not change the quality coefficients Qjk. Table 6. Short-term effects of prophylactic tamoxifen intervention on the no-tamoxifen equilibrium population in Table 4. No Tamoxifen Tamoxifen TLQ (1000s) ALQ 83448.6 27.58 84186.9 27.83 Increase 7.4 0.24 4 Conclusion In this paper we have presented population-level Markov models for medical applications, and introduced population measures for cost-effectiveness analyses of health interventions. Compared with the individual-based analyses and cohort analyses currently used, we contend that population methods can provide useful adjuncts for populationbased medical decision making. One interesting consequence of our analysis is that the common procedure of weighting QALYs by population size may be inappropriate for equilibrium analysis. In this paper, we focus on time-homogeneous models, which assume constant transition rates, or exponential transition times, between all states. The assumption can be unrealistic since some transitions such as human mortalities are usually time/age dependent. However, the exponential function with constant mortality rate as the reciprocal of life expectancy has been commonly employed in disease models to approximate human mortality since the introduction of the convenient approximation to life expectancy29 by Beck and colleagues in the 1980's. Although constant survival rates can be over simplistic and lead to inaccuracies in estimate of life expectancy30,31,32 especially when the time horizon is long and disease-specific mortality rates are low, they offer reasonable approximations in some decision scenarios. Finally, by expanding the set of Markov states, constant transition rate models can approximate human mortality reasonably well.20,33 For example, in our reconstruction of the Col et al analysis, we employed a cure-rate model18, which has constant transition rates and provides a reasonable approximation to cancer mortality. Finally, we note that the results we present here apply only to populations of noninteracting individuals, which is nevertheless the most common assumption (albeit made implicitly) used in these kinds of analyses. Our results would not apply to populations of interacting individuals such as infectious disease models. 17 Appendix Appendix 1: Individual QALYs With continuous discounting at rate r > 0, the total expected quality-adjusted life year accrued beginning in state i is expressed as QALYi E[ e rt QX (t ) dt ] = e rt j Q j p j (t )dt 0 0 where it is assumed Qj = 0 for mortality states j. In some cases, transition from state j to state k produces a quality toll Q(j,k) ≥ 0. The usual practice in discrete-time models is to incorporate such tolls into the quality coefficient Qj for state j for one cycle only. In the continuous-time case we consider, it is more convenient to attach the toll to the transition j k. The total expected QALY due to transitions, beginning in state i is QALY 'i E[ e rt Q( j, k )dX j ,k (t )] , j ,k 0 where dXjk(t) denotes the number of transitions from state j to k at the infinitesimal time interval (t, t+dt). For time-homogeneous Markov models, these values are convenient to represent in terms of matrix algebra. Suppose we augment the quality coefficients Qj and the overall QALYs by terms accounting for transition QALYs: Qj Qj + jk Q( j , k ) k QALYj QALYj + QALY j Then Qj accounts for both quality accrual and expected transition QALY while in state j, and similarly QALYj now accounts for both types of QALY accrual. It has been derived by Kulkarni 1 that QALY (rI M ) 1 Q, (4) where M is the submatrix of the rate matrix corresponding to all nonfatal states, and QALY (QALY0 , QALY1 ,..., QALYJ )T , Q (Q0 , Q1 ,..., QJ )T . Appendix 2: Proof of Theorem 2 Proof. (a) From Theorem 1 and properties of independent Poisson variables, the equilibrium distribution of the population model n conditional on total population size |n| is a multinomial distribution with parameters (| n |, 0 , J 1 j 0 j ,..., J j 0 j J ). J j 0 j 18 Therefore the expected proportion of the population in each health state j is j , j 0,..., J . J j 0 j (b). Equilibrium population means j , j 0,..., J are given by the balance equations (3), which are J J k 0 k j k 0 k j j ( j jk ) j k kj , j 0,..., J By summing up all the above equations, we get J J j 0 j 0 j j j (5) On the other hand, j , j 0,..., J are given by balance equations (2), which are J J J k 0 k 0 j 0 j ' jk k 'kj , j 0,..., J , and j 1 Or, J J k 0 k j k 0 k j j ( jk j k / ) k (kj k j / ) , j 0,..., J . Which is equivalent to J J J k 0 k j k 0 k j k 0 j ( j jk ) k kj ( k k ) j / , j 0,..., J . By using equation (5), it is easy to see that j , j 0,..., J satisfies the above system of balance equations, since J J k 0 k j k 0 k j j ( j jk ) k kj j J J k 0 k j k 0 k kj ( k k ) j / , J J J k 0 k j k 0 k j k 0 j 0,..., J However, j ( j jk ) k kj ( k k ) j / , j 0,..., J is a system of homogenous linear equations, its solutions only differ by a multiplier. 19 J Since j 0 j 1 , we have j j , j 0,..., J . J j j 0 J j J (c). From (5), we have j . j 0 j 0 j 0 Since j j j 0 J J J j 0 j 0 J j 0 j / .Therefore j j j j j j 0 j j . j j , or , we have J j . J . Appendix 3: Proof of Theorem 3 Proof. By definition DTQ e rt j Q j En j (t )dt . If the population is in equilibrium from 0 the current time t=0, then En j (t ) j for all t. Thus DTQ e rt dt j Q j j 0 1 Q j j . r j Therefore 1 DTQ TQ / yr r as claimed. Appendix 4: Proof of Theorem 4 Proof: 1) From Equation (4) we have QALY M 1 Q . We also have M 1 if writing equation (3) in matrix form. Here M is a submatrix of the rate matrix of the routing process, and also a submatrix of the rate matrix of the underlying individual model corresponding to all nonfatal states, and ( 0 , 1 ,..., J ), ( 0 , 1 ,..., J ) .Therefore, TQ/yr= Q M 1 Q QALY j j QALY j 20 References 1 Kulkarni VG. Modeling and analysis of stochastic systems. Champman& Hall 1996. 2 Beck JR and Pauker SG. The Markov process in medical prognosis. Med Decis Making 3 (1983) 419-58. Sonnenberg JB and Beck JR. Markov models in medical decision making: a practical guide. Med Decis Making 13 (1993) 322-38. 3 4 Ludbrook A. A cost-effectiveness analysis of the treatment of chronic renal failure. Appl Econ 1981; 13: 337-50. Simon DG . A cost-effectiveness analysis of cyclosporine in cadaveric kidney transpalanation. Med Decis Making 1986; 6; 199-207. 5 Deverill M, Brazier J, Green C, Booth A. The use of QALY and Non-QALY measures of health-related quality of life. Pharmacoeconomics 1998; 13: 411-420. 6 7 Pliskin JS, Shepard DS and Weinstein MC, Utility functions for life years and health status, Operations Research 28 (1980), 206-224. Mehrez A, Gafni A. “Quality adjusted life years, utility theory, and healthy-years equivalents”, Med Decis Making 1989; 9: 142-149. 8 9 Hunink M, Glasziou P, Siegel J, Weeks J, Pliskin J, Elstein A, Weinstein M, Decision Making in Health and Medicine: Integrating Evidence and Values, Cambridge University Press, 2001. Briggs A, Sculpher M, Claxton K. Decision Modeling for Health Economic Evaluation. Oxford University Press 2006. 10 11 Serfozo R. Introduction to Stochastic Networks. Springer 1999. Gross D, Shortle JF, Thompson JM, Harris CM. Fundamentals of Queuing Theory. John Wiley and Sons 2008. 12 Chen H, Yao DD. Fundamentals of Queuing Networks: Performance, Asymptotics, and Optimization. Springer-Verlag 2001. 13 Palmer S, Sculpher M, Phillips Z, Robinson M, Ginnelly L, Bakhai A et al. Management of non-ST elevation acute coronary syndrome: how cost-effective are glycoprotein IIb/IIIa antagonists in the U.K. National Health Service?. International J Cardiology 100 (2005) 229-40. 14 Ascher UM, Petzold LR. Computer methods for ordinary differential equations and differential-algebraic equations. 1998. SIAM. ISBN 0898714125. 15 Gold MR, Siegel JE, Russell LB, Weinstein MC, Cost-Effectiveness in Health and Medicine, Oxford University Press, 1996. 16 Col NF, Orr RK, Fortin JM. Survival impact of tamoxifen use for breast cancer risk reduction: projections from a patient-specific Markov model. Med Decis Making 2002; 22: 386-393. 17 21 18 Ibrahim JG, Chen M-H, Sinha D. Bayesian Survival Analysis. Springer, 2001. 19 Hazen GB, Huang M. Parametric Sensitivity Analysis Using Large-Sample Approximate Bayesian Posterior Distributions Analysis. Decision Analysis, 2006:3,208219. 20 Hazen GB. Stochastic trees: A new technique for temporal medical decision modeling. Med Decis Making 1992; 12:163-178. 21 Hazen GB. Factored stochastic trees: a tool for solving complex temporal medical decision models. Med Decis Making 1993; 13:227-236. 22 Hazen GB. Stochastic trees and the StoTree modeling environment: models and software for medical decision analysis”, Journal of Medical Systems 2002; 26:399-413. Huang M, Hazen GB., Markov Chain Population Models in Medical Decision Making, IEMS, Northwestern Univ. 2007 23 24 Cykert S, Phifer N, Hansen C. Tamoxifen for Breast Cancer Prevention: A Framework for Clinical Decisions. The American College of Obstetricians and Gynecologists. 2004: 104: 433-442. 25 Jones L, Hawkins N, Westwood M, Wright K, Richardson G, Riemsma R. Systematic review of the clinical effectiveness and cost effectiveness of capecitabine (Xeloda (R)) for locally advanced and/or metastatic breast cancer. Health Technology Assessment 2004;8: 1-176. Johnson N, Lancaster T, Fuller A, Hodgson SV. 287-9.The prevalence of a family history of cancer in general practice. Fam Pract. 1995 Sep;12(3):287-9. 26 27 U.S. Census Bureau, 2010 Census Summary File 1. 28 Fisher B, Costantino JP,Wickerham DL, et al. Tamoxifen for prevention of breast cancer: report of the National Surgical Adjuvant Breast and Bowel Project P-1 Study. J Natl Cancer Inst. 1998; 90:1371–88. Beck JR, Kassirer JP, Pauker SG. A convenient approximation of life expectancy (the DEALE"). I. Validation of the model. Am J Med. 1982;73:883-8. 29 Robert R. Holland, Charles A. Ellis, Berta M. Geller, Dennis A. Plante and Roger H. Secker-Walker. Life Expectancy Estimation with Breast Cancer: Bias of the Declining Exponential Function and an Alternative to Its Use. Med Decis Making 1999; 19; 385. 30 31 Jochanan Benbassat, Gershom Zajicek, Gerrit J. Van Oortmarssen, Isachar Ben-Dov and Mark H. Eckman Inaccuracies in Estimates of Life Expectancies of Patients with Bronchial Cancer in Clinical Decision Making Med Decis Making 1993; 13; 237. 32 J. Robert Beck and Stephen G. Pauker Does DEALE-ing Stack the Deck? Med Decis Making 1999; 19; 503. van den Hout WB. The GAME estimate of reduced life expectancy. Med Dec Making 2004; 24: 80-88. 33 22