DEVELOPING OF EXPECTATION MAXIMIZATION (EM) ALGORITHM ON

ESTIMATING METHOD OF RANDOM AND FIXED EFFECT ON MODEL GLMMs

Adji Achmad Rinaldo Fernandes
1
, I Nyoman Budiantara
2
, Bambang Widjanarko
2
, and Suhartono
2
1
h! Student, !e"artment o# Statistics, Institut $eknologi Se"uluh No"ember Surabaya
%ecturer, !e"artment o# &athematics, Statistics Study rogram, 'ni(ersitas Bra)ijaya
*mail+ #ernandes,ub-ac-id
2
%ecturer, !e"artment o# Statistics, Institut $eknologi Se"uluh No"ember Surabaya
ABSTRACT
Se(eral res"onse (ariables )ere categori.ed only t)o grou"s or classes in e(ery
subject classi#ied obser(ations on being success#ul or unsuccess#ul- Such conditions tend to
the binomial distribution- /enesis binomial distribution is o#ten #ound in the res"onse
(ariables are correlated to longitudinal data- %ongitudinal data )ith binomial distributed
res"onse (ariables can be modeled in /enerali.ed %inear &i0ed &odels 1/%&&s2- /%&&s
used to model the longitudinal data on clinical research and e"idemiological studies such as
cancer and other diseases- $he "ur"ose o# this study )as to test )hether the algorithm
e0"ecation &a0imi.ation 1*&2 to estimate model "arameters /%&&s better than commonly
used algorithms that Rha"son Ne)ton algorithm 1NR2- $his study uses #our simulation o# data
in health research- Based on the research data, it can be concluded that the *& algorithm to
estimate the "arameters better than the models /%&&s NR algorithm-
Keywords: GLMMs, EM algorithm, and the algorithm NR
1. INTRODUCTION
Fre3uently encountered in the real )orld that has t)o "ossible e(ents, eg healthy or
sick, rain or not, and so #orth, )here the data ty"e is called binary data- In general, binary data
is assumed to s"read binomial, )hich is denoted in the #orm o# success 1#igure 12, or #ail 14
"oints2- %ongitudinal data is data obtained #rom re"eated measurements 1re"eated measures2
in some indi(iduals 1cross5sectional units2 )ithin a ro) 1unit o# time2- 6erbekke and
&olenberghs 718 introduced the method o# analysis o# /enerali.ed %inear &i0ed &odels
1/%&&s2 )ere used in longitudinal data )ith binom res"onse-
In the model /%&&s, there are t)o e##ects allegedly estimated, the #irst, the e##ect o#
#i0ed 1#i0ed e##ect2 )hich is the e##ect o# the treatment 1treatment2 and the e##ect o# the
concomittant (ariables 1(ariables broadcaster2, and secondly the random e##ect 1random
e##ect2 )hich is the e##ect o# the di##erences bet)een indi(iduals 1subject s"eci#ic2- Both
methods are e0"ected to a""roach simultaneous &a0imum %ikelihood 1&%2 to estimate the
#i0ed e##ects, and the Restricted &a0imum %ikelihood 1R*&%2 to estimate the random
e##ects, )ith the hel" o# Ne)ton iteration a""roach Rha"son 1NR2- %ately, there are some
disad(antages that occur in the Ne)ton algorithm Rha"son 1NR2 as the results that a""ear in
the algorithm are the negati(e range )hich gi(e rise to inadmissible solution or a solution that
is not acce"table-
&eng and !yk 728 has de(elo"ed an *0"ectation &a0imi.ation algorithm that uses a
t)o5ste" 1estimator, and ma0imi.ation2, the mi0ed e##ect models to the case o# 3uantitati(e
res"onse 1inter(al and ratio scale data2- $here#ore, in this study it raised the de(elo"ment o#
*0"ectation &a0imi.ation algorithm 1*&2 in the #i0ed e##ect estimation method 1#i0ed e##ect2
by using the &%, and the method o# estimation o# random e##ects 1random e##ect2 on the
model using R*&% /%&&s, )hich is the sustainability o# the research conducted by &eng
and !yk 728, )here the di##erence is, in this study in(ol(es binom res"onse (ariables- In this
study )anted at once to com"are )hether the *& algorithm is better than the NR algorithm,
by looking at the (alue o# goodness o# #it is the Akaike In#ormation 9riterion 1AI92 o# the t)o
algorithms-
$he "ur"ose o# the research to be obtained is to estimate the model "arameters in
/%&&s using the *& algorithm, as )ell as testing the goodness o# the *& algorithm )ith
NR, using the smallest AI9 (alue- While the bene#its o# the research is as an alternati(e to the
settlement issue on longitudinal data analysis )ith binom res"onse, and the de(elo"ment o#
the *& algorithm on longitudinal models are e0"ected to be used as the best alternati(e #or
the estimation o# the model "arameters, so it )ill not ha""en again in admissible solution-
2. THEORY REVIEW
2.1. Gee!"#$%e& L$e"! M$'e& M(&e# (GLMMs)
/enerali.ed %inear &i0ed &odels 1/%&&s2 is the de(elo"ment o# the /enerali.ed
%inear &odel 1/%&S2- /%&S models #or res"onse binom kno)n as logistic regression-
Agresti 7:8 states, i# there is a res"onse (ariable ;i and ;i chances o# success #or the (ariables,
then+

'

i i
i i
i
1 2 4 ; 1 , 4
2 1 ; 1 , 1
;
I# the number o# trials )hich is denoted by n by 1 then ;i #ollo) the Bernoulli distribution,
and i# n < 2 then ;i #ollo) the distribution Binomial 1ni, i2- So that the logistic model is
obtained as #ollo)s 1Fahrmeir, and /erhard, 7=82+
i" " 1 i 1 4
i
i
> --- >
1
ln + + +

,
_

2 0 --- 0 e0"1 1
2 0 --- 0 e0"1
i" " 1 i 1 o
i" " 1 i 1 o
i
+ + + +
+ + +

g1
i

2 ?

,
_

i
i
1
ln

g 1
i

2 is the logit link #unction o# the distribution binom

/enerali.ed %inear &i0ed &odels 1/%&&s2 is an e0tension o# /enerali.ed %inear &odels
1/%&S2 #or data correlated as in longitudinal data )ith random e##ects added to his e3uations-
Res"onse (ariable in /%&&s assumed to be inde"endent )ith the addition o# random e##ects
#or each subject- In general, the model in /%&&s is+
g1
i

2 ? X
T
) @ Z
T
b
i
@ e
i


b
i
N 14,D2, dan e
i
N 14,R
i
2
)here +
S
i
+ subject o# obser(ation, i?1,2,:,A,N
W
j
+ obser(ation o# each subject, j?1,2,:,A,n
g1
i

2+ link #unction1logit2
ij

+ mean o# the subject to5i, and obser(ation 5j

X + co(ariate (ector o# #i0ed e##ects #or subject i5th, j5th obser(ation
+ #i0ed e##ects estimator
Z + co(ariate (ector ko(ariat o# random e##ect #or subject i5th, j5th obser(ation
+ random e##ect estimator #or subject i5th
e
i
+ error model
! + 6ariance o# random e##ects
R
i
+ 6ariance o# *rror &Bdel
2.2. C(*+"!",$-e I&$.",(!s (/ NR "& EM A#0(!$,1*
Selection o# the best algorithms by using AI9 1Akaike In#ormation 9riteria2, )hich is
de#ined in the #ollo)ing e3uation+
AI9 ?52" @ 2loglikelihood
" is the number o# "arameters to be estimated- According to Agresti 124422, the best algorithm
is the algorithm that "roduces the smallest AI9 (alue-
2. METHODS
$he data obtained are t)o "rimary data #rom "atients )ith dengue #e(er, and "atients
)ith decubitus )ound, and t)o secondary data o# "atients Age Related &acular !egeneration
1AR&!2, and e0"eriments on "regnant mice-
$he method used in this study is summari.ed in the #ollo)ing ste"s+
1- *stablishment o# a logistic regression model initiali.ation #i0ed e##ects in /%&&s
2- !ata *0"loration+
a- &ean "ro#ile, to determine the e##ect o# #i0ed e##ects
b- 6ariance "ro#ile, to determine the e##ect o# random e##ects
c- 9oncomitant (ariables, to determine the e##ect o# concomitant (ariables-
:- $entati(e model o# #ormation o# the initial determination o# the number o# #i0ed e##ects
and random e##ects )hile- Fi0ed e##ects and random e##ects as obtained #rom the
e0"loration o# the data, i# the mean and (ariance "ro#ile in#luence the need #or the
inclusion o# random e##ects and #i0ed e##ects in the model-
=- *stimation o# #i0ed and random e##ects "arameters using &a0imum %ikelihood 1&%2 and
Restricted &a0imum %ikelihood 1R*&%2 )ith NR and *& algorithm- I# the (alues
obtained are e0hibited signi#icantly then mo(e to the ne0t "hase, but i# this condition is
not met returned to the initial model building "hase 1"hase :2-
C- *0amination o# the signi#icance o# the e##ects o# "arameter estimators remain in the initial
model, i# there is a #i0ed e##ect 1in addition to the time #i0ed e##ects2 are not signi#icant
then go back to ste" : and resha"ing early models )ithout including #i0ed e##ects in
addition to the time it is not signi#icant in the model-
D- 9alculating the (alue o# log likelihood all random e##ects model- %og likelihood (alues
)ere com"ared using a likelihood ratio test, a random e##ects model deri(ed #rom the
likelihood ratio test (alues are signi#icant )ith the smallest (alue 2loglikelihood-
E- Selection o# #i0ed e##ects based on random e##ects s"eci#ied in ste" C- Fi0ed e##ects model
AI9 (alues obtained #rom the model as a )hole to include random e##ects are kno)n-
F- Signi#icance o# the re5e0amination o# the e##ects remained a#ter the inclusion o# random
e##ects into the model-
G- 9om"aring the results o# "arameter estimation a""roach )ith the &% method using NR
and *& algorithm using AI9 (alues-
%ongitudinal modeling o# /enerali.ed %inear &i0ed &odels 1/%&&s2 using SAS G-1-:
so#t)are hel" and S%'S :- : S%'S use macros to #orm NR algorithm, )hile the use o#
SAS G-1-: to run /%&&s )ith *& algorithm-
3. RESULTS AND DISCUSSION
3.1. Ne4,( "#0(!$,1* R1"+s( (NR)
According to Saa(edra 7C8, the assum"tion in /%&&s is+
1- *0"ected (alue o# the res"onse (ariable and the co(ariates associated )ith the random
e##ects such as+
g1
i

2 ?
] [ 2 1
i ij
b Y E g
? X
T
) @ Z
T
b
i
2- Assume to each o# the subjects that are assumed to be inde"endent, and #ollo) /%&S
)here y has the density #unction o# the e0"onential #amily
# 1y H I, !, J22 )here ! is the co(ariance matri0-
:- With inde"endent random e##ects #ollo) a normal distribution, /%&&s can be used to
analy.e discrete longitudinal data, including res"onse to binom 1Kardin and Kilbe, 7D82-
$he relationshi" bet)een the res"onse (ariable and the #i0ed e##ects "arameters as
random e##ects in the #ollo)ing e3uation+
2 1 Bernoulli L b H ;
ij i ij

i
ii
ij
1
ln

,
_

? X
T
) @ Z
T
b
i
$he di##erence bet)een the e3uation abo(e )ith the "re(ious e3uation is the inclusion
o# random e##ects bi in the model, the #unction o# the random e##ects in the abo(e e3uation is
to address the correlation bet)een each o# the obser(ations that may arise in longitudinal data-
Random e##ect is a com"onent o# (ariation that is not e0"lained in the "redictor (ariable-
While #i0ed e##ects are caused by the in#luence o# the "redictor (ariables-
According &olenberghs and 6erbekke 718 random e##ects can be "redicted by the
method o# &a0imum %ikelihood 1&%2, )hich is obtained by random e##ects integrate-
According to Miang 7E8, the e3uation #or the likelihood o# each subject is+
( )

N
i
i
D y f D L
1
, 2 , 1

$o com"lete the e3uations in the model is not easy because Ij /%&&s )hich )ill
allegedly be nonlinear, it is necessary #or iterati(e methods- According to Nhuri 7F8, iteration
is necessary i# o"timum (alue cannot be obtained directly- Iterati(e method used is the
Ne)ton Ra"hson algorithm 1NR2- NR a""roach is generally de#ined as #ollo)s+ A "oint 0 in a
#unction # 102 are nonlinear a""ro0imated using Ne)ton Ra"hson method is 16erbekke and
&olenberghs, 7182+
[ ] 2 1 O 2 1 O O
2 1
1
2 1 2 1 2 1 1 t t t t
x f x f x x

+
+
)here +
x
t!"#
+ results "oint iteration t @1
x
t#
+ starting "oint or a "oint o# iteration t5th
f$x# + the #irst deri(ati(e o##unction fx#
f$$x#+ the second deri(ati(e o# #unction fx#
Analogous to the abo(e e3uation, the e3uation a""roaches #or "arameter I is +
2 1 1 2 1 2 1 2 1 1
2 1
t t t t
g %
+

)here

& j
j t
%

2 1
2
2 1

1
]
1

i
i
ij j
i
ij j i i& ij
& j
j
x
x N x x
2
2
2 e0"1 1
2 e0"1
2 1

i
i i i i& ij
N x x 2 1 1
? X
T
V
51
X
V
51
is diagonal matri0 )ith element
2 1 1
i i i
N
H
51
is koragam )ide matri0 )hich is the in(erse o# H

i i
i ij i
j
i
j
j
t
x Nx y x g 2 1
2 1
2 1

i
i i i ij
N y x 2 1
? X
T
1Y
$
5
i

2
so the #i0ed e##ects estimator
)
(!61)

? )
(!)
@ 1X
T
V
51
X2
51
1X
T
1Y
$
5
i

22
in the same )ay, #i0ed e##ects estimator is obtained +
7
(!61)

? 7

(!)
@ DZ
T
V
51
1Y
$
5
i

2
$his e3uation must be sol(ed iterati(ely 1r is the inde0 #or each iteration, )ith r ? 4, 1,
2, ---2, the "rocess is re"eated until the obtained I and b are con(ergent or- I is considered
con(ergent i# the (alue o# P is less than 145D- $he relationshi" bet)een the abo(e t)o
e3uations lead to the "ossibility o# NR algorithm gi(es results that are not con(ergent, and
e(en create a negati(e range because o# the t)o kinds are used #or both e3uations, #or
e0am"le a )ide range o# errors and random e##ects models-
3.2. E'+e.,",$( M"'$*$%",$( "#0(!$,1* (EM)
*& algorithm is essentially as it has been studied by &eng 728 is a re#inement o# the
NR algorithm #or the longitudinal model a""roach to continuous res"onse- In the "re(ious
e3uation, Ri is the (ariance o# the error, using the #ollo)ing a""roach+
R
$
?
2
I
ni
, #or i ? 1, 2, A, n
)here bi and ei Q under the assum"tion o# normality is bi N 14, !2, and ei N 14, Ri2,
then the &a0imum %ikelihood method to estimate 2 and ! become+

i
$
i
1 2
e e N R

$
i i
1
b b n !
R
$he abo(e e3uation is &5ste" 1&a0imi.ation Ste"2 on the *&5Algorithm, because ei and bi
are not kno)n- *& algorithm is a ste" change 2 and ! )ith+

2

? *1
2
R H y, ?
R
2

2

? ?
[ ] { }

+
1
i
2
i
2
i
$
i
1
6 1 tr n e e N
! ? *1
!
R
H y, ?
R
2
! ?
[ ] { }

+ ! S 6 !S ! b b n
i
1
i
$
i
$
i i
1
$he right side o# the abo(e e3uation, the (ariety 2 and ! is not kno)n, so it )as
re"laced )ith and the abo(e e3uation as the initial (alue- By iteration, and the (alue )ill be
u"dated using the abo(e e3uation by using the *& algorithm a""roach, so that the *&
algorithm scheme are as #ollo)s+
Set r ? 4, 1, 2, --- is a se3uence o# iterations, and the e0"ected (alue o# the "arameter and is the
#i0ed e##ects and random e##ects at iteration r- arameter estimation ste"s )ith the *&
algorithm is as #ollo)s+
Ste" 4+
set r ? 4, 2 r 1
2
R ? 1 and
2 r 1
!
R
? I
n
ste" 1+
set r ? r @ 1, u"date e3uation 2 r 1
R

and 2 r 1
b
R
using+
( ) ( )
i
1
2 1 r 1 i
$
i
1
i
1
2 1 r 1 i
$
i 2 r 1
y 6
R
> > 6
R
>
R

2
R
> y 1 6
R
S ! b
R
2 r 1 i
1
2 1 r 1 i
$
i 2 1 r 1 2 r 1



)here
n
2 1 r 1
2 $
i 2 1 r 1 i 2 1 r 1 i
I R S !
R
S 6
R


+
ste" 2+
rene) (alue 2 r 1
2
R and
2 r 1
!
R

2 r 1
2
R
?
[ ] { }

+
1
2 1 r 1 i
2 1 r 1
2
i
2 1 r 1
2
2 1 ir
$
2 1 r 1 i
1
6
L
1 tr R n R e e N
2 r 1
!
R
?
[ ] { }

+
2 1 r 1 i
1
2 1 r 1 i
$
i 2 1 r 1 2 1 r 1
$
2 ir 2 r 1 i
1
!
R
S 6
R
S !
R
!
R
b b n
)here
e
i1r2
? y
i
5 >
i
2 r 1
R
5 S
i
2 r 1
b
R
ste" :+
re"eat ste" 1 and 2 u" to con(ergen
D
2 r 1 j 2 1 r 1 j
14

+

3.2. D"," e'+#(!",$(
*0"loration data include the mean "ro#ile, and the "ro#ile (ariance structure
concomitant is the initial stage o# model building longitudinal data using /enerali.ed %inear
&i0ed &odels 1/%&&s2 are use#ul to establish the model describing the initial e0"loration
"ro#ile changes mean changes in the "ro"ortion o# success#ul e(ents #or the )hole subject
each unit o# time- *0"loration (ariance "ro#ile aims to in(estigate the in#luence o# random
e##ects- Random e##ects need to be added to the model i# the (ariance "ro#ile gra"hs are
#luctuations in the residual (alue o# each unit o# time- *0"loration concomitant (ariables need
also to be done- Bridesmaids (ariables in#luence cannot be ignored because it can a##ect the
res"onse (ariable accom"animent-
Kere is "resented the #irst data on the e0"loration o# "atients )ith dengue #e(er 1#irst
data2, as #ollo)s+
F$08!e 1. Me" +!(/$#e
F$08!e 2. V"!$".e +!(/$#e
F$08!e 3. C(.(*$,",9 A0e
Based on Figure 1 abo(e, it sho)s the trend o# increase in the "ro"ortion o# "atients
)ho sho)ed reco(ery to the addition o# #i0ed e##ects in /%&&s models to be #ormed- Figure
2 sho)s the di(ersity o# res"onses indicates a change needs to include random e##ects in
addition to the #i0ed e##ects model o# tentati(e- From Figures : and = it can be seen that each
o# the (ariables gender 1se02 ha(e the same relati(e e##ect bet)een men and )omen, )e ha(e
considered the inclusion o# the (ariable se0- While the age (ariable 1age2, )hich #ormed
relati(ely di##erent gra"hs #or each age, so the age (ariables continue to be included in the
model-
3.3. Res8#, (/ P"!"*e,e! Es,$*",$(
From e0"loration results it sho)s that the #ourth data are )orth to do modeling-
Results o# /enerali.ed %inear &i0ed &odel "rediction models 1/%&&s2 #or the #irst data ie
data on "atients )ith dengue #e(er, are "resented as #ollo)s+
T"7#e 1 P"!"*e,e! Es,$*",$( GLMMs *(&e# 4$,1 /$!s, &","
P"!"*e,e!
NR EM
Es,$*",(! SE P5-"#8e es,$*",(! SE P5-"#8e
Interse" 5=-1C2D 4-F22E 4-4441 5=-11=C 4-F121 4-4441
$ime 1-C2:: 4-22=: 4-4441 1-C2:: 4-22=: 4-4441
$reatment 4-41:= 4-42=C 4-CF:C 4-41:= 4-42== 4-DFDC
Se0 54-4:F1 4-=12E 4-G2D= 54-4:F1 4-=12E 4-G2DG
From the table abo(e )e can see that the time (ariables signi#icant at T o# CU #or both
algorithms- 6alue estimate is "ositi(e indicating an increase in healing "atients )ith dengue
#e(er all the time 1 to = days- As #or the concomitant (ariables o# age, there has been a
signi#icant and "ositi(e im"act on both algorithms- $hat is, the younger the "atient, the higher
the cure rate- Seen also in the algorithms #or concomitant (ariables gender seen signi#icant
e##ect on the increase in healing "atients )ith dengue #e(er- /i(en the gender is a dummy
(ariable 11+ male, and 4+ #emale2, indicates a )oman has a #aster cure rate than men- B(erall,
the t)o "arameter estimators, using Ne)tonOs method and *0"ectation &a0imi.ation
Rha"son sho)ed no signi#icant di##erence-
Results o# /enerali.ed %inear &i0ed &odel "rediction models 1/%&&s2 #or the
second data on "atients )ith decubitus )ound as #ollo)s+
T"7#e 2 P"!"*e,e! Es,$*",$( A#0(!$,1* GLMMs *(&e# 4$,1 /$!s, &","
P"!"*e,e!
NR
Es,$*",(! SE P5-"#8e Es,$*",(! SE P5-"#8e
Interce"t 5E-GF14 1-:E:1 4-4441 522-:1D1 G-F4F= 4-42DE
$ime 4-=DD: 4-4=14 4-4441 1-=2GE 4-1EGC 4-4441
$reatment 4-:EE: 4-22GC 4-4GFC 4-E:E: 1-EC4= 4-DEC2
Age 4-4:DE 4-41F4 4-4=1C 4-4ECE 4-1::C 4-CE:4
Se0 1-E4=F 4-2=:C 4-4441 C-D21G 1-G4G2 4-44=E
From $ables 2, it sho)s that both algorithms estimate "arameters sho)ed considerable
di##erences, both seen #rom the signi#icance (alue estimate, and the (alue o# the notion- In the
estimation o# the "arameters )ith NR algorithm sho)s the e##ect o# time, age, and se0 o# the
"atient to "atient res"onse decubitus )ound- Whereas the "arameter estimation )ith the *&
algorithm sho)s only the e##ect o# time and se0 o# the "atient that in#luence the res"onse o#
"atients )ith decubitus )ound-
/%&&s estimation results #or the third entry o# "atients )ith Age Related &acular
!egeneration 1AR&!2 as #ollo)s+
T"7#e 2 P"!"*e,e! Es,$*",$( A#0(!$,1* NR GLMMs M(&e# 4$,1 ,1$!& &","
P"!"*e,e!
NR EM
Es,$*",(! SE P-"#8e Es,$*",(! SE P-"#8e
Interce"t 5=-GCCD 1-21G2 4-4441 5:-4:G= 2CC-DG44 4-GG4G
$ime 4-=42= 4-4:CE 4-4441 4-2C=4 4-44CG 4-4441
Age 4-41== 4-41CF 4-:FG4 1-22:= 1-2=EC 4-:2EF
From the table belo), it sho)s that both methods o# "arameter estimation in /%&&s
models sho)ed similar results in estimating the (alue o# signi#icance, but sho)ed di##erent
results on the (alue sus"ected- It can be seen #rom the e(olution 1changes each time2 o# the
res"onse o# "atients )ith Age Related &acular !egeneration 1AR&!2- $o estimate the NR
algorithm, sho)s that #rom )eek to )eek, "atients )ith AR&! can be cured by e0" 14-=42=2
or 1,=GC times better than the "re(ious )eek- As #or the estimation using the *& algorithm, it
is seen that #rom )eek to )eek, "atients )ith AR&! can be cured by e0" 14-2C=42 or 1,2FG
times better than the "re(ious )eek-
/%&&s estimation results #or #ourth on the "ro(ision o# data o# betel lea# and
hydrogen "ero0ide in "regnant mice as #ollo)s+
T"7#e 3 P"!"*e,e! Es,$*",$( A#0(!$,1* GLMMs M(&e# 4$,1 /(8!,1 D","
P"!"*e,e!
NR EM
Es,$*",(! SE P-"#8e Es,$*",(! SE P-"#8e
Interce"t 54-F=C1 1-2FFF 4-C124 5=-2C:: 14-FDG4 4-DGGG
$ime 4-FEC1 4-1:22 4-4441 =-1GGF 1-CD1C 4-41=C
$reatment 54-D4F2 4-1E1C 4-444= 52-EEC1 1-DD:4 4-111D
)eight 54-414C 4-44G2 4-2C2G 54-4C:E 4-4F41 4-C112
In the table abo(e, it sho)s that both algorithms estimate "arameters sho)ed
considerable di##erences, )hether it is seen #rom signi#icance o# estimating (alue, and the
(alue o# the notion- In the estimation o# the "arameters )ith NR algorithm, it sho)s the
in#luence o# time and treatment e##ect gi(ing betel lea# and hydrogen "ero0ide )ith an
increase in necrotic tissue in "regnant mice- Whereas the "arameter estimation )ith the *&
algorithm, it sho)s only the e##ect o# time that in#luence the increase in necrotic tissue in
"regnant mice-
C(*+"!$s( (/ NR "& EM A#0(!$,1* ( M(&e# GLMMs
Based on the results o# the model "arameter estimation /%&&s, )hether it is #rom
signi#icance o# estimating (alue, and the magnitude o# "redicti(e (alue tends to make a
di##erence in the t)o algorithms are algorithms NR and *& Algorithm- $o test )hich is the
best algorithm, using the criteria o# Akaike In#ormation 9riterion 1AI92, in )hich the best
algorithm is the algorithm that "roduces the smallest AI9 (alue- $he #ollo)ing table
summari.es the AI9 (alue and the number o# iterations #or each algorithm, and the "ercent
e##ecti(eness-
T"7#e : C(*+"!$s( (/ !es8#,s (/ NR "& EM A#0(!$,1*
N(
N8*7e! (/
I,e!",$(
AIC
NR EM NR EM
!ata 1 2: 22 1DG-1 1D:-4
!ata 2 =1 2E =14-2 :12-F
!ata : :G 11 12CC-F G1C-=
!ata = =D 21 11:-2 F1-G
From the table abo(e it can be seen that the lo)est AI9 (alue in the #ourth research
data generated by the algorithm *0"ectation &a0imi.ation 1*&2- Also e(ident is the number
o# iterations re3uired to reach the *& algorithm to estimate the "arameters con(erging "oint
less than the number o# iterations )ith the NR algorithm on the same data- It can be
concluded that the estimation o# the model "arameters /enerali.ed %inear &i0ed &odels
1/%&&s2 )ith *0"ectation &a0imi.ation algorithm 1*&2 gi(es better results than the
Ne)ton algorithm Rha"son 1NR2-
CONCLUSIONS AND RECOMMENDATIONS
C(.#8s$(
From the results o# research conducted, it can be concluded as #ollo)s+
1- *0"ectation &a0imi.ation algorithm 1*&2 can be used to estimate the model "arameters
in /enerali.ed %inear &i0ed &odels 1/%&&s2- $his )as sho)n by the #ormation o# the
#our models in the research data using the *& algorithm-
2- *0"ectation &a0imi.ation algorithm 1*&2 algorithm is better than Ne)ton Rha"son
1NR2 )hich is currently o#ten used in model "arameters /%&&s sus"ect- $his is e(ident
#rom the #our research data, the *& algorithm AI9 (alue is smaller than the (alue o# AI9
NR algorithm-
S800es,$(
From the results o# this study, it can be suggested some o# the #ollo)ing+
1- *& algorithm can be used as the settlement o# the "roblem in longitudinal data analysis
)ith res"onse binom, and tend to im"ro(e model accuracy better than the NR algorithm
is no) more #re3uently used- So it can be suggested that *& de(elo"ment is the best
alternati(e models "arameters estimating /%&&s-
2- Bn #urther research it is recommended to use the "5(alue criterion )hich hel"s to de#ine
the selection o# the best algorithm- It is recommended to study the de(elo"ment o# the
*& algorithm to the analysis o# longitudinal data )ith ordinal res"onse and res"onse
"oison-
REFFERENCES
718 &olenbergs-/-, dan 6erbekke, /-, 244C- M(&e# /(! D$s.!e,e L(0$,8&$"# D","-
S"ringer Series in statistics- Ne) V;ork+S"ringer V6erlag-
728 &eng, >, dan !yk, !- 244F- F"s, EM T;+e I*+#e*e,",$(s /(! M$'e& E//e., M(&e#s-
Mournal o# the Royal Statistical Society- Series B- 6ol- D4, No- :- " CCG5CEF-
7:8 Agresti, A- 2442- A I,!(&8.,$( ,( C",e0(!$."# D"," A"#;s$s. Mohn Wiley W Sons-
Ne) ;ork-
7=8 Fahrmeir, %- dan $- /erhard- 1GG= - M8#,$-"!$",e S,",$s,$."# M(&e##$0 B"se& (
Gee!"#$%e& L$$e! M(&e#s- Mohn Willey dan Sons , Ne) ;ork-
7C8 Saa(edra, -A-$- 244D- Pe!.e,$#e C8!-es I B$"!; L(0$,8&$"# D",".
htt"+XXgrad-u"rm-eduXtesisXtorressaa(edra-"d#- $anggal akses + 1F Bktober 2414-
7D8 Kardin, M- W- dan M- Kilbe- 244E- Gee!"#$%e& L$e"! M(&e#s "& E',es$(s- Stata
ress- $e0as
7E8 Miang, M- 244E- L$e"! "& Gee!"#$%e& L$e"! M$'e& M(&e#s "& T1e$! A++#$.",$(.
S"ringer Series in Statistics- Ne) ;ork-
7F8 Nhuri, A- 244:- A&-".e& C"#.8#8s 4$,1 A++#$.",$( I S,",$s,$.s. Mohn Wiley and Son-
Ne) Mersey-