.Estimating population attributablefractionsimpact fraction - PDF document

.Estimating population attributablefractionsimpact fractionÓ (PIF) in Equation 1a (Drescher and Becher 1997; Eideand Heuch 2001; Walter 1980).(x):relative risk at exposure level x):population distribution of exposure):counterfactual distribution of exposure, and :maximum exposure levelThe Þrst and second terms in the numerator of Equation 1a representthe total exposure-weighted risk of disease or mortality in the popula-tion under current and counterfactual exposure distributions. The cor- RRxPxdxRRxPxdxRRxPxdx()()()() Estimating attributable burden of disease from exposure andhazard dataStephen Vander Hoorn,Majid Ezzati,Anthony Rodgers,Alan D.Lopez and Christopher J.L.Murray The PIF equation can be used to estimate the population attributablefraction (PAF), deÞned as the proportional reduction in disease or deaththat would occur if exposure to the risk factor were reduced to the coun-terfactual exposure distribution. The remainder of this chapter outlineshow data on exposure, hazard and disease burden were combined toderive estimates of attributable disease burden, with estimation of thepopulation attributable fraction as the intermediate step. The applica-tion of Equations 1a and 1b in the context of the comparative risk assessment (CRA) project is discussed and several issues regarding its.Estimating attributable mortality andburden of diseaseFor each risk factorÐdisease outcome pair, PAFs for each of the 224 age,sex, subregiongroups were calculated using the relationships in Equa-tion 1, separately for mortality (PAF) and incidence (PAFrelative risks for mortality and incidence were different. For each of these 224 groups, the estimates of mortality (attributable to risk factor below. Burden of disease, reported annually in the annexes of the World Health Report(DALYs), with methods and assumptions described elsewhere (MurrayPAFAYLLPAFAYLDPAFAYLLAYLDWhere ÒAÓ indicates ÒattributableÓ and years of life lost to premature mortalityFor those risk factors with insufÞcient data to estimate a relative riskmodel (e.g. occupational or alcohol-caused injuries or the effects of leadexposure on blood pressure), disease burden or mortality was estimatedusing existing registers or corresponding hazard relationships. Estimateswere then aggregated across age groups to obtain subregional estimates, PRRPRRPRR Comparative Quantification of Health Risks and across subregions to obtain global estimates. The details of this ag-gregation are described later in this chapter..Counterfactual exposure distributionThe estimates of burden of disease and injuries due to risk factors in theCRA project are based on a counterfactual of theoretical-minimum-riskexposure distribution, deÞned in chapter 1 and described in individualrisk factor chapters. By using the theoretical-minimum-risk exposure distribution, which by deÞnition has a relative risk of 1.0, as the reduced to:Theoretical-minimum-risk exposure distribution forcontinuous exposure variablesThe theoretical-minimum-risk exposure distribution for continuous riskfactors is itself often a distribution of exposure levels, vs a constant base-line. Figure 25.1, for example, illustrates a scenario for systolic bloodpressure (SBP) with typical exposure levels in an older population (mean:150mmHg; SD: 9mmHg) compared with the theoretical-minimum-riskexposure distribution (mean: 115mmHg; SD: 6mmHg). The non-zerostandard deviation of the theoretical-minimum-risk distribution reßectsthe reality that there always is some inter-person variability within anyin Figure 25.1.The optimal exposure distribution for a population would overlap precisely with the theoretical-minimum-risk exposure distribution. BydeÞnition of theoretical-minimum risk, such a population would be col-burden due to the risk factor of interest. Any population containing indi- PRRPRR RRxPxdxRRxPxdx()()()() Stephen Vander Hoorn et al.2131 theoretical minimum will have attributable burden tending towards zero.of exposure) in the population would be determined by the differencebetween her/his current exposure (SBP level) and the SBP level that s/heical-minimum-risk exposure distribution.Estimating the total hazard at the population level can be achievedrandomly from current and theoretical-minimum-risk exposure distribu-tions. For most risk factors, however, individual exposures ÒtrackÓ overrelatively long periods of time (Lauer and Clarke 1988; Voors et al. 1979;Wilsgaard et al. 2001). In other words, those with higher/lower expo-sure levels of a particular risk factor are expected to have higher/lowerexposure levels within the theoretical-minimum-risk exposure distribu-tion (see the hypothetical individuals in Figure 25.1). Random (uncor-related) draws of individuals from current and theoretical-minimum-riskexposure distributions would be inconsistent with the empirical evidenceon tracking. Consistent with this evidence, we assumed that the order-Comparative Quantification of Health RisksFigure 25.1Theoretical-minimum-risk exposure distribution forcontinuous risk factors using systolic blood pressure (SBP) as an example Usual SBP (mmHg) 100 2.04.08.0 5 Proportion of population (%) Relative riskheart disease Note:Each point represents a hypothetical individual or small group of individuals in the population.Thesolid straight lines represent the increasing relative risk,on a log scale,for ischaemic heart disease with increasing SBP. the rank-order correlation of individual exposures equals 1) in the tran-sition to the theoretical-minimum-risk distribution in estimating the PAF.With correlated rank-ordering of individuals in current and theoretical-minimum-risk exposure distributions, if hazards were a linear function of exposure, then for those risks with symmetric distributions,shifting the population to the theoretical minimum distribution wouldof the theoretical-minimum-risk exposure distribution (i.e. the standarddeviation of the theoretical-minimum-risk exposure distribution wouldmean, as a result of changing the standard deviation of the theoretical-minimum-risk exposure distribution, would fully compensate each other.Risk is, however, an exponential function of exposure in most epi-demiological models. With an exponential hazard function, when thebaseline is the mean of theoretical-minimum-risk exposure distribution,the mean (Figure 25.2), compared to the case of treating theoretical-minimum-risk exposure as a distribution with non-zero standard devia-curve (i.e. increased risk per unit increase in exposure) and the standarddeviations of the current and theoretical-minimum-risk exposure For computational reasons, we estimated PAFs for continuous riskfactors relative to the mean of the theoretical-minimum-risk exposuredistribution. In these calculations, the relative risk for any individual inthe population with an exposure below the mean of the theoretical-minimum-risk exposure distribution was set to 1.0 (e.g. in Figures 25.1and 25.2, the lower tail of current blood pressure distribution is insidethe theoretical minimum distribution with some individuals already at alevel below 115mmHg. These individuals were assigned a relative riskCRA project, global PAFs estimated by integrating risk relative to themean of the theoretical-minimum-risk exposure distribution were up to2% larger than those estimated by integrating risk relative to the full dis-.Aggregation of attributable burdenacross age,sex and subregionWithin each of the 14 subregions, all-age-sex population attributablefractions (PAFestimates across the 16 age-sex-speciÞc estimates within the subregionStephen Vander Hoorn et al.2133 Similarly, for each age-sex group, world attributable fractionsPAFacross all the 14 subregion-speciÞc estimates and then dividing by the subregionagesexagesexsubregionagesexagesex Comparative Quantification of Health RisksFigure 25.2Effect of a non-linear hazard function and choice of baseline on total population risk Note:With an exponential hazard function,when theoretical-minimum-risk exposure is a distributionwith a non-zero standard deviation,those falling above the mean of the current distribution (e.g.155mmHg for SBP) contribute more to total population hazard than those below it,relative to thecase when the baseline is a constant level (115mmHg for SBP).In the Þgure,the solid linesrepresent the hazard when the theoretical-minimum-risk exposure is a distribution,and the dottedline when a constant baseline is considered.The difference between the two relative risks on the) is larger than those on the left (RR).As a result of this imbalanced contribution tohazard,using the mean of theoretical-minimum-risk exposure distribution as baseline in estimatingtotal population hazard would result in slightly larger PAFs than using the complete theoretical minimum distribution. Relative risk of ischaemicheart disease 100115150180 1.02.04.08.0 10Proportion of population (%) Current distributionTheoretical-minimum-riskdistribution Stephen Vander Hoorn et al.2135This is shown in Tables 25.1Ð25.3 for the case of SBP and ischaemicheart disease (IHD). The non-italic numbers in Table 25.1 are the sub-region-age-sex speciÞc PAFs estimated using Equation 2. Next, thesefractions were applied to the Global Burden of Disease (GBD) 2000 esti-mates of disease burden for IHD, shown in Table 25.2, producing theestimates of IHD disease burden attributable to SBP in Table 25.3(similar estimates could be made for mortality or YLL). Dividing thetotal attributable burden in any subregion (e.g. 1.548 million DALYs forAMR-A in the highlighted cell) by the total IHD burden for the subre-gion in the GBD database (3.506 million DALYs for AMR-A in the high-lighted cell) gives the all-age-sex subregional PAFs (44% for AMR-A inThe all-age-sex PAF estimates for the remaining 13 subregions are alsoshown in Table 25.1 in italics. Similarly, world PAFs were calculatedare shown in the bottom row of Table 25.1. For example, the world PAFable burden in that age-sex group (4.71 million DALYs) by the totalworld IHD burden in the GBD database (9.015 million DALYs), givinga PAF of 52%.Computationally, aggregate PAFs (whether aggregated across age-sexgion-age-sex speciÞc estimates, with weights being the same as the totalnumber of events (i.e. deaths, YLL, or DALYs) for each subregion-age-sex group. In the above example, total subregion-age-sex speciÞc DALYsare the weighting factor. As a result, subregion and world PAFs areweighted more towards the ages and/or subregions that have higherDALYs (rather than those with larger populations). For each risk factor,these separate aggregate PAFs were estimated for deaths, YLL, andDALYs with the corresponding GBD estimates used as the denominator(or weighting factor). As a result, even when the age-sex-subregion-speciÞc PAFs are the same for the three measures, the aggregate onesmay differ.For example, although the subregion-age-sex-speciÞc PAFs are thesame for deaths and DALYs in the case of elevated SBP and IHD, divid-ing the IHD deaths attributable to this risk factor in AMR-A (203000)by the total IHD deaths in this region (618000), gives a regional PAFfor mortality of 33% in AMR-A. In fact, the smaller subregional PAFfor mortality compared to that of DALYs highlights the higher weight-ing towards the older age PAFs (which are smaller) in the case of agesexsubregionagesexsubregionagesex Comparative Quantification of Health Risks Table 25.1PAFs for IHD attributable to increased SBP (%),by age,sex and subregion 0Ð4 (years)5Ð14 (years)15Ð29 (years)30Ð44 (years)45Ð59 (years)60Ð69 (years)70Ð79 (years)80 (years) MalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesTotal SubregionAFR-DÑÑÑÑÑÑ63567073566350581619AFR-EÑÑÑÑÑÑ51395961465340481215AMR-AÑÑÑÑÑÑ46205753495146481518AMR-BÑÑÑÑÑÑ46236160515747541417AMR-DÑÑÑÑÑÑ50316058485243491315EMR-BÑÑÑÑÑÑ55576471536148561519EMR-DÑÑÑÑÑÑ49446168516146571419EUR-AÑÑÑÑÑÑ66477270596253571719EUR-BÑÑÑÑÑÑ63497778647059662023EUR-CÑÑÑÑÑÑ66557480627356681823SEAR-BÑÑÑÑÑÑ43396060515246481415SEAR-DÑÑÑÑÑÑ33305349464242381311WPR-AÑÑÑÑÑÑ60407164595854531817 WPR-BÑÑÑÑÑÑ28235155445141481214WorldÑÑÑÑÑÑ4735615952544853151849 ÑNo data. Stephen Vander Hoorn et al.2137 Table 25.2GBD 2000 estimates of total disease burden for IHD (000s of DALYs),by age,sex and subregion 0Ð4 (years)5Ð14 (years)15Ð29 (years)30Ð44 (years)45Ð59 (years)60Ð69 (years)70Ð79 (years)80 (years) MalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesTotal SubregionAFR-D346531481168824228119018912216637481576AFR-E457749661498827327820018811014829501653AMR-A1110157214747092605142894703912193403506AMR-B2232462121810759430141227925522369962631AMR-D12111072511543342303124911294EMR-B35643217179594111432191281149926291474EMR-D24242112908432217077147552246929232869733746EUR-A10171515181406571467172846715182393893882EUR-B1011351526678734258647440431471931763647EUR-C001178156931271708477149692084712231715638319SEAR-B7343984922212339427931828618818843542259SEAR-D70529958283622110590036882119265723611412148427829117480WPR-A003310640101514013559112854666765 WPR-B871471557749527810996659457836827761823406513 World12510716811194610494225215411484575590156704573761271510252657743 Comparative Quantification of Health Risks Table 25.3Burden of IHD attributable to increased SBP (000s of DALYs),by age,sex and subregion 0Ð4 (years)5Ð14 (years)15Ð29 (years)30Ð44 (years)45Ð59 (years)60Ð69 (years)70Ð79 (years)80 (years) MalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesMalesFemalesTotal SubregionÑÑÑÑÑÑ7349168205107119619669893ÑÑÑÑÑÑ76351601719299457238760AMR-AÑÑÑÑÑÑ971540313725114621518934611548AMR-BÑÑÑÑÑÑ1012536418021115811912010161303AMR-DÑÑÑÑÑÑ12333192016131212132EMR-BÑÑÑÑÑÑ983426410111678555645811EMR-DÑÑÑÑÑÑ1567547332426528513418610141922EUR-AÑÑÑÑÑÑ1201947610242217535429641742079EUR-BÑÑÑÑÑÑ1673856520141730825630918402320EUR-CÑÑÑÑÑÑ456711270381927670475826311325239SEAR-BÑÑÑÑÑÑ95482371681611498790681050SEAR-DÑÑÑÑÑÑ36727019721030122298859556835337080ÑÑÑÑÑÑ2441072680346046812400 ÑÑÑÑÑÑ1386555736642039927837022492664WorldÑÑÑÑÑÑ198375070493412471036252746323422946328201 ÑNo data. mortality, because greater numbers of IHD deaths occur in these ageto larger loss of life (YLL), when DALYs are considered, the contribu-tion of PAFs at younger ages to the all-age-sex PAF becomes greater..Exceptions to the general estimation procedureThe following list brießy describes the major departures from the stan-dard analysis framework which were required so that all risk factorscould be adequately assessed within the project. Further details are pro-¥Theoretical minima varied by age, sex and subregion for iron deÞ-ciency, since this was the haemoglobin distribution that would beobserved if iron deÞciency were eliminated from each population.¥Theoretical minima varied by age, sex and subregion for lack of con-¥Fruit and vegetable intakes in any population were truncated at zero.In other words, all individuals falling below zero in the distribution¥The burden of cardiovascular diseases attributable to lead exposuresure due to lead and then estimating the total mediated effect through.Other methodological issuesrelationship in Equation 1 lead to biased estimates when the relative Robins 1988). This bias is in fact a result of the correlation among mul-tiple risks (the risk factor of interest and other risk factors that act asconfounders), as well as the diseases affected by them (Ezzati et al. 2003).Accounting for this correlation in the estimation of attributable burden,however, would require the availability of exposure and disease dataare not available and therefore reliance on the formula with direct useof the adjusted relative risk estimates was necessary. In the case of pos-itive correlation among risk factors, this would generally result in anunderestimation of population attributable fraction.Stephen Vander Hoorn et al.2139 1See preface for an explanation of this term.Drescher K, Becher B (1997) Estimating the generalized impact fraction from.BiometricsEide GE, Heuch I (2001) Attributable fractions: fundamental concepts and theirStatistical Methods in Medical ResearchEzzati M, Vander Hoorn S, Rodgers A, Lopez AD, Mathers CD, Murray CJ,Comparative Risk Assessment Collaborating Group (2003) Estimates ofglobal and regional potential health gains from reducing multiple major risk.The LancetGreenland S (1984) Bias in methods for deriving standardized morbidity ratio.Statistics in MedicineGreenland S, Robins JM (1988) Conceptual problems in the deÞnition and inter-pretation of attributable fractions.American Journal of EpidemiologyLauer R, Clarke W (1988) A longitudinal view of blood pressure during child-hood: the Muscatine Study.Statistics in MedicineMurray CJL, Lopez AD, eds. (1996) The global burden of disease: a compre-hensive assessment of mortality and disability from diseases, injuries and riskfactors in 1990 and projected to 2020. Global Burden of Disease and Injury,Vol 1. Harvard School of Public Health on behalf of WHO, Cambridge, MA.Voors A, Webber L, Berenson G (1979) Time course studies of blood pressurein childrenÑthe Bogalusa Heart Study.American Journal of EpidemiologyWalter SD (1980) Prevention of multifactorial disease.American Journal of Wilsgaard T, Jacobsen BK, Schirmer H et al. (2001) Tracking of cardiovascularrisk factors: the Tromso Study, 1979Ð1995. American Journal of Epidemiol-Comparative Quantification of Health Risks