Study Protocol

Bayesian Autoregressive Mul til evel Model ing of Burden of Diseas- es,
Injuries and Risk Factors in Iran 1990 - 2013 Amir Kasaeian PhD Candidate1,2,
Mohammad Reza Eshraghian PhD1, Abbas Rahimi Foroushani PhD1, Sharareh R. Niakan
Kal hori PhD2,3 #####################1###############################2,4 Abstract
Background: Statistical model ing and devel oping new methods for estimating burden
of diseases, injuries and risk factors is a fundamen- tal concern in studying the
country heal th situation for better heal th management and pol icy making. Bayesian
autoregressive mul til evel model is a strong method for this kind of study though in
compl ex situations it has its own chal l enges. Our study aims to describe the way of
model ing space and time data through an autoregressive mul til evel model and address
chal l enges in compl ex situation.Method: We wil l obtain data from different
publ ished and unpubl ished secondary data sources incl uding popul ation-based heal th
surveys (e.g. NHS, DHS, STEP) at national and provincial l evel s and we al so assess
epidemiol ogical studies via systematic review for each disease, injuries and risk
factor over the period of 1990 - 2013 . These data general l y have a mul til evel
hierarchy and al so time correl ation. However, statistical anal ysis of diseases,
injuries and risk factors data is primaril y facing the probl em of information
scarcity. Data are general l y too scarce to ensure rel iabl e estimates in many
practical probl ems. Al so, there may be nonl inear changes over time, different kind
of uncertain- ties in data and incompatibl e geographical data. We describe Bayesian
autoregressive mul til evel model ing approach that provides a natural sol ution to
these probl ems through its abil ity to sensibl y combine information from several
sources of data and avail abl e prior information. In this hierarchy model l evel s of
each hierarchy borrow information from each other and al so l ower l evel s borrow
information from higher
###################################################################################
#################################################Discussion: Our anal yses wil l
incl ude different existing sources of data in Iran for 24 years through a rational
and reasonabl e model to estimate burden of diseases, injuries and risk factors for
Iran at national , regional and provincial l evel s whil e considering several kinds of
uncertainties. Comprehensive and real istic estimates are al ways an issue of request
that wil l be obtained through a suitabl e statistical model ing considering al l
dimensions and then can be used for making better decision in real situations.
Keywords: Autoregressive time series, burden of diseases, Iran, MCMC, mul til evel
model s, NASBOD############################################???Cite this articl e as:
Kasaeian A, Eshraghian MR, Rahimi Foroushani A, Niakan Kal hori SR, Mohammad K,
Farzadfar F. Bayesian autoregressive mul til evel model - ing of burden of diseases,
injuries and risk factors in Iran 1990 - 2013 . Arch Iran Med. 2014 ; 17(1): 22 - 27.
IntroductionPrecise assessment of gl obal , regional , and country heal th conditions
and trends is crucial for evidence-based decision making for Publ ic Heal th.1 The
Gl obal Burden of DiseaseStudy (GBD) is the l atest and most rel iabl e anal ysis to
reveal the importance of taking different approaches to the chal l enges facing
gl obal heal th.2 The GBD study resul ts provide us a data-rich struc- ture for
comparing the effects and burden of different diseases, injuries, and risk factors
on premature death and disabil ity be- tween popul ations.3 -13 But these resul ts are
not for within popul a-######################1Department of Epidemiol ogy and
Biostatistics, School of Publ ic Heal th, Tehran University of Medical Sciences,
Tehran, Iran, 2Non-Com- municabl e Diseases Research Center, Endocrinol ogy and
Metabol ism Popul ation Sciences Institute, Tehran University of Medical Sciences,
Tehran, Iran, 3 Social Determinants of Heal th Research Centre, School of Heal th,
Ahvaz Jundishapur University of Medical Sciences, Ahvaz, Iran, 4 Endocrinol ogy and
Metabol ism Research Center, Endocrinol ogy and Metabol ism Research Institute, Tehran
Uni- versity of Medical Sciences, Tehran, Iran. ###################################
Kazem Mohammad PhD, Department of Epidemiol ogy and Biostatistics, School of Publ ic
Heal th, Tehran University of Medical Sciences, Tehran, Iran. Address: Poursina
Avenue, P.O. Box 64 4 6, Tehran 14 155, Iran. Tel : 021 889513 96; Email :
mohamadk@tums.ac.ir.Farshad Farzadfar MD DSc, Non-communicabl e Diseases Research
Center, En- docrinol ogy and Metabol ism Research Institute, Tehran University of
Medical
################################################################################
Ave, Tehran, Iran. Postal Code: 1599666615, Tel /Fax: 98-21-88913 54 3 ,E-mail : f-
farzadfar@tums.ac.ir.Accepted for publ ication: 3 December 2013 tion, which means
nothing is known about what's going on within a country, expl icitl y Iran here.
Knowing and comparing heal th situation within regions and provinces hel ps to
understand the dif- ferences and simil arities better and al so better map the
emerging epidemics of diseases which in turn hel ps heal th pol icy makers to
################################################################- ous effects and
extra burden of those diseases. The onl y study of burden of diseases and injury in
Iran dates back to 2003 which
############################################################## disparity between
these six provinces and al so indicated a transi- ##### ##### ############# ###
################# #### ##### ####### injuries.14 In l ine with the GBD study,
National and Subnational Burden of Diseases study 2013 (NASBOD) is a systematic
effort to ef- #####################################################################
in Iran.15 It al so takes into account care systems, current avail abl e data on
heal th systems and viabl e, systematic and rel evant nation- wide studies carried out
in the previous years.This study is an endeavor to assess and eval uate the burden
of diseases at national and provincial l evel s in Iran by means of the most recent
val id and rel iabl e qual itative and quantitative research methods and experiences
taken from Gl obal Burden of Diseases 2010 (GBD). Moreover, knowl edge, expertise,
and skil l s of heal th
###################################################################
##############################################################-??22
ArchivesofIranianMedicine,Vol ume17,Number1,January2014 tions of burden of diseases.
###########################################################-bil ity of comparison
and contrast of heal th conditions in different provinces and regions, an advantage
or prerogative which is con- ducive to the fostering heal th and hygiene in these
regions, which provides Heal th Pol icy Makers with the necessary documents for
better heal th pol icy making and resource al l ocation.Al so, a comparison of the
situations at provincial l evel s wil l ul ti- matel y boost the heal th condition
throughout the country in a fair and bal anced manner.
##################################################################- tical model ing
and improved methods for estimating time trends of diseases, injuries, and risk
factors. We wil l use the advanced methods and when necessary expand the current
methods to de- vel op Bayesian time series mul til evel model s for 3 1 provinces from
1990 to 2013 . We wil l present the data, methods and the
################################################################# and provincial
1990 - 2013 trends and their uncertainties in popu- l ation's mean (whatever measure
is) of diseases and injuries or risk factors for al l provinces in four regions
incl uded in NASBOD study to al l ow meaningful time and provincial comparison. Iran
is divided into four regions (eastern south, north and eastern north, west, and
center) on the basis of two criteria: epidemiol ogical ho- mogeneity, and
geographical contiguity. The study covers urban and rural areas of the country.The
main purpose of this articl e is to expl ain a Bayesian autore- gressive mul til evel
model and al l its components together with chal l enges in compl ex data which wil l
appear in the study of bur- den of diseases, risk factors and injuries.MethodsStudy
designStatistical anal ysis of diseases, injuries and risk factors data is primaril y
facing the probl em of information scarcity. Data are general l y too scarce to ensure
rel iabl e estimate in many practical probl ems. In the present study there are 24
provinces at the begin- ning year in 1990, however during a period of 24 years,
there are 3 1 provinces at the ending of this study. This means there shoul d be at
l east 576 data points that this is very unl ikel y in a study of diseases, risk
factors and injuries. This probl em is more serious #### #### ##### ##### ### #####
############ ########## ### ### ##### ######## that we encounter geographical
incompatibil ity which is the other issue of concern.We describe a Bayesian
autoregressive mul til evel model ing ap- proach that provides a natural sol ution to
these probl ems through its abil ity to sensibl y combine information from several
sources of data and avail abl e prior information. Such model ing strategies that
capture geographical and time patterns in the data wil l reduce estimation error.We
wil l devel op this model to estimate preval ence of diseases and injuries or mean of
risk factors by age group, sex and province over the time period of 1990 - 2013 . We
do anal yze each gender indepen- dentl y and make estimates for each age group-
province-year unit.In this mul til evel model provinces are nested in subregions,
subregions are nested in regions, and regions nested in country l evel . Accordingl y,
l ower l evel (s) borrow information from higher l evel s and al so l evel s of each
hierarchy borrow information from each other. In fact, there is an concurrence for
borrowing informa- tion depending
on the l evel of avail abil ity and scarcity of data sothat the richer is the data
the l ess borrowing within and across l evel s wil l be needed and vice versa.The other
point is that trends might not be l inear over time; this non-l inearity wil l be
model ed in the form of a l inear trend pl us a smooth non-l inear trend, both
hierarchical l y.Al so, because of heterogeneity between community-based stud- ies
they might have l arger variation than national l y representative studies. The model
is abl e to capture this variation through incl ud- ing a time-varying offset for
non-provincial data. These variation components were estimate empirical l y.Another
probl em that might occur is the non-l inear association between preval ence and/or
mean measurements and age since the association might change in different ages
especial l y in ol der age groups. In such a condition, we wil l use cubic spl ine age
model or ##########################################################We wil l
determine the val ues of different kinds of uncertainties such as sampl ing
uncertainty in the original data, uncertainty as- sociated with inconsistency
between years in national data, uncer- tainty rel evant to data sources that are not
provincial , uncertainty associated with statistical methods for crosswal king
between preval ence (categorical measure) and mean (continuous measure).Final l y, a
Bayesian model with Markov Chain Monte Carl o
##############################################################- rior distribution
of model parameters, which represent uncertainties, wil l be used to achieve
posterior distribution of preval ence or mean ### ##### ######### ### #####
############ ########## ####### ### ######### ####### #### ############ ##### ####
### ##### ############ ########### indeed from the data itsel f in an integrated and
direct way. Uncer- tainty interval s are al so computed for preval ence and mean.Data
sourcesWe wil l obtain data from different publ ished and unpubl ished secondary data
sources incl uding popul ation-based heal th sur- veys (e.g. NHS, DHS, STEP) at
national and provincial l evel s and al so epidemiol ogical studies via systematic
review for each disease, injury and risk factor. Some data are obtainabl e from cen-
suses, househol d expenditure surveys, demographic surveil l ance, and disease and
death registries. Data from systematic review are eval uated via a qual ity
assessment process used in GBD to review the incl uded studies and to excl ude the
poor studies. This process has three parts incl uding general information of the
study, qual - ity of sampl ing, and qual ity of measurement. Data from popul a- tion-
based or community-based surveys, househol d expenditure surveys and censuses al so
wil l be incl uded in the study after data cl eaning for pl ausibl e ranges of variabl es
and outl iers detection.
#################################################################### for each year
and province incl uding information for mean or prev- al ence (depending on the
anal ysis), sampl e sizes, standard devia-
##################################################################### embed the
survey weights in age group-sex-province-year groups.Since the mean of measure and
its uncertainty are inputs of the model , in the anal yses of risk factors, for the
studies that reported
#################################################################### ### ##########
######### ###### ### ########### ### ########## ######### width. For studies
reported mean, sampl e size and standard devia- tions (SD), we estimated the
standard error (SE) as SD/(n^0.5).CovariatesA covariate is a variabl e that has a
positive or negative rel ation- ship with a disease, risk factor or injury in the
NASBOD study.###############################################?
ArchivesofIranianMedicine,Vol ume17,Number1,January2014 23 We wil l use covariates to
inform the estimation process in our model s. For conditions with l ots of data,
covariates pl ay a mini- mal rol e in the estimation process however for conditions
with l ittl e data, the rol e of covariates is very important. In fact, time- varying
province-l evel covariates can hel p informing the units ### ## ######## ######
###### #### ####### ##### ### ########### ##### ##### mul tipl e sources.Some of the
frequentl y used covariates associated with the risk fac- tors or diseases under
study are (i) urbanization, measured as propor- tion of province's popul ation that
l ived in urban areas, (ii) province avail abil ity of mul tipl e food types for their
citizens' consumption, (iii) weal th index, estimated from assets, which were asked
in yearl y househol d expenditure surveys, (iv) years of school ing, which is
educational attainment in years obtainabl e from househol d expendi- ture surveys,
(v) popul ation density, proportion of the province with popul ation density over
1000 peopl e per square kil ometer, (vi) mean BMI, mean body mass index (kg/m^2) for
mal es and femal es ol der
#######################################################################- en)
obtainabl e from census, (viii) compl eteness of vital registration (% of deaths
captured) obtainabl e from census and vital registration data, (ix) vehicl es, 2+4
wheel s (per capita) accessibl e from coun- try Road Statistics. However, some
variabl es l ike neonatal mortal - ity rate (per 1000), diabetes preval ence (% of
popul ation), smoking preval ence (% of popul ation), systol ic bl ood pressure (mmHg),
and indoor and outdoor air pol l utions which are indeed estimates from NASBOD study
wil l be used as covariates for estimating of other diseases, risk factors and
injuries.Crosswal king
############################################################### use homogenous
data. Non-homogenous data wil l l ead to wrong estimates. Sometimes depending on the
primary outcome we need to attain continuous measurement from preval ence or vice
versa, for exampl e mean FPG from diabetes preval ence since the rel evant study has
reported just preval ence. Other exampl e is when one measurement can be obtained
from other measurement
############################################################## one measurement is
al so the other issue which is necessary. One
################################################################### intake to point
preval ence of al cohol intake.### ##### ##### ### ######## ## ########### ###
############ ##########
################################################################ metrics) and use
the beta generated as the adjustment factor for a
################################################################## is that the
necessary data shoul d be rel ativel y high and the over- l apping information from the
same source is needed to generate rel ationships. This technic is the so-cal l ed
crosswal k or metadata mapping method.Total l y, crosswal k is a method of data
conversion that enabl es searching data across heterogeneous resources and is a
useful tool for making simil ar data comparabl e.Statistical anal ysisMul til evel
model s that are al so cal l ed hierarchical , mostl y be- cause of the parameters of the
within-l evel regressions at the l ow- est, control l ed by the hyper-parameters of the
upper-l evel model , are the basis of our anal ysis. The mul til evel model ing al l ows
esti- mating heterogeneity within as wel l as across l evel s or units.16
############################################################
##################################################################-ear time sl ope.
The second component of the model is the nonl in- ear time effect. Covariate effect
is the third one. Age is the other important component which wil l be smoothed via a
cubic spl ine. Since there are different kinds of data sources one component is
########### #### ###### ####### ####### #### ####### ######### #####- nent which is
mul tipl icative with study random effect.The mul til evel hierarchy component of the
time trendsAn important trait of mul til evel model s is that each parameter
########### ### ## ######## ###### ### ##### ######## ############ ##### comparabl e
parameters of other groups or units with simil ar char- acteristic. In other words,
a shrinkage effects towards the popul a- tion mean is present whil e using mul til evel
model s. The vol ume of the shrinkage depends on the variance between the random
################################################################# number of
individual s is observed in some groups. In such cases, there is l arge reduction of
the uncertainty since information from other groups or units with smal l er
variabil ity is incorporated in the posterior estimates.17 This is our main
rational e for using Bayes- ian mul til evel model s.In our project, studies are nested
in provinces, provinces are nested in sub-regions, sub-regions, are in turn nested
in regions and al l nested in the country. This is the structure of the data. The
################################################################## pool ing
estimates from the model . Partial -pool ing is a compro- mise between two extremes;
non-pool ing and compl ete pool ing. Compl ete pool ing is when we combine al l
observations of a given l evel and non-pool ing is the opposite. In this scenario,
mul til evel estimate of a given province is approximated by a weighted aver- age of
the observation in the province (the un-pool ed estimate, yJ) and the mean over al l
provinces (the compl etel y pool ed es- timate, Yal l ). So, depending on the
avail abil ity and sparseness of
################################################################# by means of non-
pool ing, partial pool ing and compl ete pool ing.18 This situation is repeated in each
hierarchy.Nonl inear time effect componentNonl inear changes in time at each province
wil l be captured using a term which is the sum of province, sub-region, region and
the country and each of these four components is assigned a Gaussian autoregressive
prior to al l ow the model to distinguish the extend of nonl inearity exist at each
l evel .19, 20In particul ar, the vectors of each component have a normal
prior with zero mean and precision parameters. The model -estimated precision
parameters wil l determine the degree of smoothing at each l evel . We wil l expect the
provincial precision parameter to be the l owest and the country precision parameter
to be the high- est as the provincial trends of a disease has more extra-l inear
vari- ######## ##### #### ######## ####### ############### ### ########### ### ###
issue of concern here which wil l be achievabl e by constraining
##############################################################- l ing orthogonal ity
between the l inear and nonl inear part of the time trend is that each can be
expl ained independentl y. For prov- inces with no data, we wil l take the Moore-
Penrose pseudoinverse for computation because of some technical matters.21Covariate
effects componentThe covariates which we wil l use in our model are categorized in
two group; province-l evel covariates and study-l evel covari- ates. Province-l evel
covariates incl ude covariates l ike (i) weal th
############################################??24
ArchivesofIranianMedicine,Vol ume17,Number1,January2014 index, (ii) urbanization,
(iii) mul tipl e food types based on a prin- cipal component anal ysis (PCA) on
Househol d Expenditure Data, (iv) years of school ing, (v) body mass index, (vi)
compl eteness of vital registration and etc.The effects of some of these province-
l evel covariates on the risk factors or diseases wil l be al l owed to change l inearl y
over time. Theses covariates wil l be smoothed using moving average
################################################################ of yearl y
variation of covariates.However, the study-l evel covariates incl ude study coverage
and study-l evel urbanization. The study-l evel coverage covariate which expl ains
types of data used has four categories: (i) provin- cial data with sampl ing weight
(ii) provincial data without weight (iii) district data, and (iv) individual
community data.The next study-l evel covariate expl ains whether the study has been
conducted in rural , urban or rural and urban area popul ation. These two covariates
wil l hel p us account for data sources bias- ness. Since non-provincial studies
mostl y are performed in areas of special regard or thought because of a heal th
probl em, their re- sul ts wil l not be representative of the whol e province. They
might al so have l arger variation than provincial representative studies. As
mentioned before the model considers a time-varying offset for district and
community data, and additional variance com- ponents for district and community
data and for provincial data without sampl ing weights. These variance components
were esti- mated empirical l y and l et provincial data with sampl ing weightsto have a
stronger effect on estimates than other sources.The covariates and their
interactions wil l be chosen based on substantive thoughtful ness and their
predictive power through in-
#####################################################22 We arenot seeking causal
effects of these independent variabl es.Age association componentAl most al l risk
factors and diseases have a nonl inear association with age, for exampl e for some
diseases age association might ####### ### ##### ######### ### ###### ###### ####
### ##### #### ## ###### spl ine model to smooth this association.23 We wil l use a
basel ine age and then subtract al l age val ues from that basel ine.Since the age
association between provinces might change fur-
###############################################################-
############################################################## normal distribution
of zero mean and ## variances that each ## has #####################24 We treated
age as a continuous variabl e in this model . This is the reason we extracted age
groups from studies as narrow bands (5 years) to use their mid-point as continuous
measurements.######################################ing for sampl ing variabil ity.
So, the term #w enabl es us to expl ain this variabil ity.#w can al so expl ain study
design and qual ity mat- ters. We assume random effects from community studies have
greater variance than random effects from district studies and soforth i.e.: #w
###u # #d # #c. This constrain indicates that studies with l imited coverage are not
onl y have greater or l esser than the province mean or preval ence, but al so have
more variabil ity.Residual age-study variationAge patterns inside communities within
a given province may differ and may not be consistent with its province age
pattern. This kind of within-study variation wil l not be captured by e terms as
they are the same across al l observations in a given study. Thus, an additional
variance component for each study, #2i , wil l be ac- commodated in the model :####
############## ####### ####### ######## #### ####### ##### #### ##### ###### ###
##### ############ ###### ########### #### ###### ####- abl es in the model . We
appoint a normal prior with variance de- ######## ### #### ######### ### ####
###### ## #### ##### ############## random e#ffect, ei :#w #u #d #c
############################################################### or preval ence of
the measurement under study even after account-c2 i##study i is communityAgain
there is l ess variation in weighted provincial studies than unweighted provincial
studies and so on i.e.: #w ###u # #d # #c. This consideration for model comprises
the smooth age in residu- al terms not onl y for each province but al so for each
study to have its own cubic spl ine in age.Computation
########################################################### method. Al l statistical
computation programs wil l be written and done in R l anguage. As we know wel l , to
achieve better estima- tions from the model we shoul d jointl y sampl e random effects
with their hyperparameters since there is a heavy dependency between parameters.25
We wil l not marginal ize over mean param- eters in the model since this may cause
off-diagonal structure into the l ikel ihood covariance and need manipul ating l arge
variance- covariance matrices to cal cul ate this marginal l ikel ihood.A main step in
running MCMC is to ensure the MCMC sampl er wil l converge to the posterior
distribution and that estimating is
################################################################ draws.24 For each
model , we wil l start with 20 chains in paral l el at randoml y-sel ected starting
val ues. Then, after 5000 iterations of burn-in to harmonize the Metropol is proposal
variances, we wil l run each chain 50000 more iterations. Next, we wil l combine
##################################################################
################################################################# MCMC is that
uncertainty generates natural l y from the data via estimation in an integrated and
simpl e manner.Model checking
################################################################- ting and
achieving tradeoff between these two needs a great atten- tion. The perfect model
is el astic enough to capture important com- pl ications whil e stil l keeping its
external val idity and interpretabil ity.We wil l examine our model using posterior
predictive checks to verify that we have not negl ected any key interaction out and
al so ##### #### ################# ### ####### ### ##### #### ######## #### ######
Posterior predictive checks are wel l -designed and smart tool for in-
###################################################26 We wil l com- pare observed
datasets with a given repl icated datasets, e.g. 500,var (ei)= i##study i is weighted
provincial i##study i is unweighted provincial i##study i is districti##study i is
community# #di #2 i##study i is district
################################################w2 i##study i is weighted
provincial #2 = #u2 i##study i is unweighted provincial ?
ArchivesofIranianMedicine,Vol ume17,Number1,January2014 25from model 's posterior
predictive distribution for other risk factors. Whenever the difference between
this prediction and the observed data becomes smal l er, this means our model is
consistent with data.For cross-val idating the model , we wil l divide the provinces
into #### ################ ###### ### ###### ###### ### ##### #### #### #######
become simil ar regarding rich and sparse density. For each group of provinces we
wil l do a 10-fol d cross-val idation so that we drop
################################################################- ###### ###### ###
########## #### ########### ###### ### #### ###### ###### when predicting that 10
percent of data. We wil l do this for every 10 percent and combine the 10 percent
estimates of prediction errors.27At each iteration of the MCMC we wil l draw a
predic- tion from the main model and wil l buil d 95 % prediction interval s from
predictions across al l iterations.Discussion
######################################################### burden of diseases, risk
factors and injuries across provincial and regional l evel s over recent years in
Iran.15 The onl y study of burden of diseases and injuries in Iran dates back to
2003 which ################################################################- cant
disparity between these provinces and al so transition from communicabl e diseases to
non-communicabl e diseases and road
################14 ############################################### burden of
diseases study in Iran and even in the Middl e-East and one of the few subnational
studies al l over the worl d.28-3 1 We wil l obtain l ong-term trends of preval ence of
diseases, risk factors and injuries under NASBOD study for each age group, sex,
province, sub-region, region and the whol e country. Then we wil l estimate heal th
inequal ities respectivel y. Al l the time trends wil l be report- ed together with
their uncertainty interval s. We wil l report esti- mates for al l province-years,
subregion-years, region-years that many of them suffers from poor data.As mentioned
before provincial and subnational studies of bur- den of diseases inside countries
provided heal th pol icymakers with a sol id perspective of heal th situation al l over
the country and therefore hel ped them in better heal th management and fu- ture
pl anning to control the progressive epidemics of al l domi- nant diseases. The other
advantage of the present study compared with the onl y previous one in Iran and
other subnational studies in the worl d is that its methodol ogical and anal ytical
approach is very cl ose to GBD study 2010 guidel ines together with their main
investigators invol vement. What mentioned above are just the epidemiol ogical
achievements of such a study which wil l be a hel pful l andmark for pol icy makers in
heal th systems.The NASBOD project achievements are not onl y very important from
epidemiol ogical perspective but al so from statistical point of view because of
handl ing the compl exities existing in the na- ture of this study. These
compl exities wil l be model ed with new advanced statistical model s especial l y
Bayesian autoregressive mul til evel model s as expl ained in this paper.Though the
detail ed main discussion on the resul ts wil l be pro- vided after running the model
and rel easing the resul ts, we can
##################################################################Mul til evel model s
are of the rare approaches for model ing ag- gregated data l ike what we encountered
in NASBOD study.One of the main advantages of mul til evel model s is assessing
different l evel s effects. Considering higher l evel units as a random
###################################################################-el variation in
the total popul ation and therefore l eads to unbiased standard error estimates and
independent residual s of the model .16 Another advantage is that missing data which
are frequentl y oc- curred in l arge surveys are handl ed very simpl y via these
model s.16Though handl ing missing data is one of the advantages of mul - til evel
model s and we wil l just use these model s together with informative priors to impute
missing information, our model suf- fers from data scarcity especial l y in ol der age
groups and the ear- l ier time of the study. Our model al so suffers from l ow qual ity
and non-representativeness of data at the earl ier time of the study mainl y before
2000. Thus, rel ativel y l arge amount of data wil l make our inferences more robust.It
is cl ear from the l iterature that many approaches have been devel oped for missing
data imputation but al most al l of them use simpl e methodol ogy l ike bootstrapping
just l ike Amel ia3 2 and ### #### ##### ## ####### ############ ### ##############
### ### ##### ### compl ex situations resembl ing NASBOD study. The onl y disad-
vantage is that model ing process and interpretation of the resul ts may be compl ex
which both can be passed off through advanced ########### ######### ############
############ #### ############# Data gaps may be the main l imitation of our study
just l ike what occurred in model ing GBD study 2010.10-13 The other l imitation is
the geographical incompatibil ity that occurs at the provincial l evel s which is not
a serious probl em in mul til evel model ing since we have onl y sl ight changes during
the study period and it can be handl ed with tricky techniques. But it may be a
serious probl em at district l evel s and more advanced model s shoul d be devel oped at
this phase of study in near future.The other sensibl e model s which can be engaged
in NASBOD study is the Spatio-temporal model s.3 3 which wil l be devel oped ####
###### ### ######### #### ##### ##### ### ###### ##### #### ######## Bayesian
autoregressive mul til evel model s to devel op ensembl e model s which wil l produce
independent model and more rel iabl e and accurate estimations. Ensembl e model s are
weighted com- binations of the posterior distributions of individual model s and
provide l ower error for point estimates and more accurate uncer- tainty
interval s.3 4 -3 6 Moreover, ensembl e model s catch uncertainty due to both the
parameters in any singl e model and the uncertainty
###################################################################General l y
speaking, the main advantages of the mentioned mod- el is estimating l ong-term
trends using a Bayesian autoregressive mul til evel model to predict mean and
preval ence of risk factors and diseases by incl uding non-l inear age associations
and time trends, incorporating study coverage as wel l as variance compo-
################################################################# our model to use
al l the data and track provincial representative
################################################################# variance
components to be greater and have l arger uncertainty for data sources with l ess
representativeness and ul timatel y uncer- tainl y interval s achieved from the
Bayesian model that represent the true avail abil ity of information.Though we are to
devel op a sophisticated model based on real needs and existing compl exities in real
situations to estimate missing information this does not obviate the need for
gathering ############################Al l mentioned about model ing and its
chal l enges in compl ex conditions itsel f creates careers for young researchers to
l earn and train more and more and this capacity buil ding ul timatel y wil l l ead to
knowl edge production in the country.As a bottom l ine, achieving estimations of time
trends after mod-############################################?26
ArchivesofIranianMedicine,Vol ume17,Number1,January2014 el ing al l diseases, risk
factors and injuries under NASBOD study can hel p anybody who works in heal th
systems, special l y Heal th Pol icy Makers and al so pol iticians to trace, understand
and monitor epidemiol ogical transition of non-communicabl e diseases in al l over the
country and then l aunch prevention pl an to reduce the burden of non-communicabl e
diseases, risk factors and injuries and conse- quentl y achieve the new heal th goal
of the Worl d Heal th Assembl y in 2012,2 which is reducing avoidabl e mortal ity from
non-communi- cabl e disease (NCDs) by 25 % by 2025 (the 25 by 25 goal ).Authors'
ContributionsGeneral design prepared by Farshad Farzadfar and Amir Ka- saeian.
Designing of model s prepared by Farshad Farzadfar, Ka- zem Mohammad, Amir Kasaeian,
Mohammad Reza Eshraghian and Abbas Rahimi Foroushani. The primary draft was
prepared by Amir Kasaeian and revised by al l co-authors. Al l authors have
#####################################################Acknowl edgmentsThe study is
granted by Ministry of Heal th and Medical Education of Isl amic Republ ic of Iran and
Setad-e-Ejraie Farmane Imam. The authors woul d l ike to express thanks to Dr.Masoud
Moradi for his precise editing of the text and Ms Rosa Hagh Shenas for her efforts
on managing coordinative and administrative processes.References1. Chan M. From new
estimates to better data. The Lancet. 2012; 3 80(9859): 2054 .2. Horton R. Non-
communicabl e diseases: 2015 to 2025. The Lancet. 2013 ; 3 81(9866): 509 - 510.3 . Lim
SS, Vos T, Fl axman AD, Danaei G, Shibuya K, Adair-Rohani H, et al . A comparative
risk assessment of burden of disease and injury attributabl e to 67 risk factors and
risk factor cl usters in 21 regions, 1990-2010: a systematic anal ysis for the Gl obal
Burden of Disease Study 2010. The Lancet. 2013 ; 3 80(9859): 2224 - 2260.4 . Murray
CJ, Vos T, Lozano R, Naghavi M, Fl axman AD, Michaud C, et al . Disabil ity-adjusted
l ife years (DALYs) for 291 diseases and injuries in 21 regions, 1990-2010: a
systematic anal ysis for the Gl obal Burden of Disease Study 2010. The Lancet. 2013 ;
3 80(9859): 2197 - 2223 .5. Vos T, Fl axman AD, Naghavi M, Lozano R, Michaud C, Ezzati
M, et al . Years l ived with disabil ity (YLDs) for 1160 sequel ae of 289 diseases and
injuries 1990-2010: a systematic anal ysis for the Gl obal Burden of Disease Study
2010. The Lancet. 2013 ; 3 80(9859): 2163 - 2196.6. Sal omon JA, Wang H, Freeman MK,
Vos T, Fl axman AD, Lopez AD, et al . Heal thy l ife expectancy for 187 countries,
1990-2010: a system- atic anal ysis for the Gl obal Burden Disease Study 2010. The
Lancet. 2013 ; 3 80(9859): 214 4 - 2162.7. Sal omon JA, Vos T, Hogan DR, Gagnon M,
Naghavi M, Mokdad A, et al . Common val ues in assessing heal th outcomes from disease
and injury: disabil ity weights measurement study for the Gl obal Burden of Disease
Study 2010. The Lancet. 2013 ; 3 80(9859): 2129 - 214 3 .8. Lozano R, Naghavi M,
Foreman K, Lim S, Shibuya K, Aboyans V, et al . Gl obal and regional mortal ity from
23 5 causes of death for 20 age groups in 1990 and 2010: a systematic anal ysis for
the Gl obal Burden of Disease Study 2010. The Lancet. 2013 ; 3 80(9859): 2095 - 2128.
9. Wang H, Dwyer-Lindgren L, Lofgren KT, Rajaratnam JK, Marcus JR,
#################################################################### countries,
1970-2010: a systematic anal ysis for the Gl obal Burden of Disease Study 2010. The
Lancet. 2013 ; 3 80(9859): 2071 - 2094 .10. Danaei G, Finucane MM, Lu Y, Singh GM,
Cowan MJ, Paciorek CJ, et al . National , regional , and gl obal trends in fasting
pl asma gl ucose and diabetes preval ence since 1980: systematic anal ysis of heal th
examination surveys and epidemiol ogical studies with 3 70 country-years and 2? 7
mil l ion participants. The Lancet. 2011; 3 78(9785): 3 1 - 4 0.11. Danaei G, Finucane
MM, Lin JK, Singh GM, Paciorek CJ, Cowan MJ, et al . National , regional , and gl obal
trends in systol ic bl ood pres- sure since 1980: systematic anal ysis of heal th
examination surveys and epidemiol ogical studies with 786 country-years and 5? 4
mil l ion participants. The Lancet. 2011; 3 77(9765): 568 - 577.12. Finucane MM,
Stevens GA, Cowan MJ, Danaei G, Lin JK, Paciorek CJ, et al . National , regional , and
gl obal trends in body-mass index since 1980: systematic anal ysis of heal th
examination surveys and epidemiol ogical studies with 960 country-years and 9? 1
mil l ion par- ticipants. The Lancet. 2011; 3 77(9765): 557 - 567.13 . Farzadfar F,
Finucane MM, Danaei G, Pel izzari PM, Cowan MJ, Pa- ciorek CJ, et al . National ,
regional , and gl obal trends in serum
total chol esterol since 1980: systematic anal ysis of heal th examination sur- veys
and epidemiol ogical studies with 3 21 country-years and 3 ? 0 mil - l ion participants.
The Lancet. 2011; 3 77(9765): 578 - 586.14 . Naghavi M, Abol hassani F, Pourmal ek F,
Lakeh MM, Jafari N, Vaseghi S, et al . The burden of disease and injury in Iran
2003 . Popu- l ation heal th metrics. 2009; 7(1): 9.15. Farzadfar F, Del avari A,
Mal ekzadeh R, Mesdaghinia A, Jamshidi #### ########### ### #### ####### ######
######## ############ #### ###- rics. Arch Iran Med. 2014 ; 17(1): 7 - 15.16.
Gol dstein H. Mul til evel statistical model s: Wil ey. com; 2011.17. Ntzoufras I.
Bayesian model ing using WinBUGS: Wil ey. com; 2011. 18. Gel man A. Data anal ysis
using regression and mul til evel /hierarchical model s: Cambridge University Press;
2007.19. Bresl ow NE, Cl ayton DG. Approximate inference in general ized l in-ear mixed
model s. Journal of the American Statistical Association.1993 ; 88(4 21): 9 - 25.20.
###############################################################-tions: CRC Press;
2005.21. Harvil l e DA. Matrix al gebra from a statistician's perspective: Spring-er;
2008.22. Spiegel hal ter DJ, Best NG, Carl in BP, Van Der Linde A. Bayesian
######################################################################Society:
Series B (Statistical Methodol ogy). 2002; 64 (4 ): 583 - 63 9. 23 . Durrl eman S, Simon
R. Fl exibl e regression model s with cubic spl ines.Statistics in Medicine. 1989;
8(5): 551 - 561.24 . Marin J-M, Robert CP. Bayesian core: a practical approach to
compu-tational Bayesian statistics: Springer; 2007.25. Chib S, Carl in BP. On MCMC
sampl ing in hierarchical l ongitudinal model s. Statistics and Computing. 1999; 9(1):
17 - 26.26. Gel man A, Meng X-L, Stern H. Posterior predictive assessment of model
###################################Statistica Sinica. 1996; 6(4 ): 73 3 - 760. 27.
Hastie T, Tibshirani R, Friedman J, Frankl in J. The el ements of statisti- cal
l earning: data mining, inference and prediction. The Mathematical Intel l igencer.
2005; 27(2): 83 - 85.28. Bradshaw D, Nannan N, Groenewal d P, Joubert J, Laubscher
R, Ni-jil ana B, et al . Provincial mortal ity in South Africa, 2000-priority- setting
for now and benchmark for the future. South African Medical Journal . 2008; 95(7):
4 96.29. Stevens G, Dias RH, Thomas KJ, Rivera JA, Carval ho N, Barquera S, et al .
Characterizing the epidemiol ogical transition in Mexico: na- tional and subnational
burden of diseases, injuries, and risk factors. PLoS Medicine. 2008; 5(6): e125.3 0.
Begg SJ, Vos T, Barker B, Stanl ey L, Lopez AD. Burden of disease and in- jury in
Austral ia in the new mil l ennium: measuring heal th l oss from diseas- es, injuries
and risk factors. Medical journal of Austral ia. 2008; 188(1): 3 6.3 1. Asaria P,
Fortunato L, Fecht D, Tzoul aki I, Abel l an JJ, Hambl y P, et al . Trends and
inequal ities in cardiovascul ar disease mortal ity across 793 2 Engl ish el ectoral
wards, 1982-2006: Bayesian spatial anal ysis. International journal of Epidemiol ogy.
2012; 4 1(6): 173 7 - 174 9.3 2. Honaker J, King G, Bl ackwel l M. Amel ia II: A program
for missing data. R Package version 1.5-5. 2011.3 3 . Parsaeian M, Farzadfar F,
Zeraati H, Mahmoudi M, Rahimighazika- l ayeh G, Navidi I, et al . Appl ication of
spatio-temporal model to esti- mate burden of diseases, injuries and risk factors
in Iran 1990 - 2013 . Arch Iran Med. 2014 ; 17(1): 28 - 3 2.3 4 . Vrugt JA, Robinson BA.
Treatment of uncertainty using ensembl e methods: Comparison of sequential data
assimil ation and Bayesian model averaging. Water Resources Research. 2007; 4 3 (1):
W014 11.3 5. Gneiting T, Raftery AE. Strictl y proper scoring rul es, prediction, and
estimation. Journal of the American Statistical Association. 2007; 102(4 77): 3 59 -
3 78.3 6. Raftery AE, Gneiting T, Bal abdaoui F, Pol akowski M. Using Bayes- ian model
averaging to cal ibrate forecast ensembl es. Monthl y Weather Review. 2005; 13 3 (5):
1155 - 1174 .###############################################?
ArchivesofIranianMedicine,Vol ume17,Number1,January2014 27Copyright of Archives of
Iranian Medicine (AIM) is the property of Academy of Medical Sciences of I.R. Iran
and its content may not be copied or email ed to mul tipl e sites or posted to a
l istserv without the copyright hol der's express written permission. However, users
may print, downl oad, or email articl es for individual use.