Date post: | 19-Aug-2015 |
Category: |
Health & Medicine |
Upload: | max-moldovan |
View: | 105 times |
Download: | 3 times |
Australian Institute of Health Innovation
Composite quantitative analysis of international healthcare accreditation survey data
Max Moldovan, Charles Shaw and Jeffrey Braithwaite
Healthcare accreditation:
a process that a health care institution, provider, or program undergoes to demonstrate compliance with standards developed by an official agency.[Miller-Keane Encyclopedia and Dictionary of Medicine, Nursing, and Allied Health, Seventh Edition. © 2003 by Saunders, an imprint of Elsevier, Inc. All rights reserved.]
Study and data set description
The first international healthcare accreditation survey was undertaken in 2000, commissioned by the World Health Organisation [WHO Review 2003, WHO/EIP/OSD/2003.1]The current data set has been collected over 2009-201144 accreditation agencies, located in 38 countries10 broad categories of attributes (e.g. “Policy and governance” and “Funding of the organization”)173 questions leading to categorical, date/numeric or free text responses
Older than 20 years Between 10 and 20 y.o. <10 y.o.;
* If more than one program per country, the oldest program is displayed
Geographical and age distribution of accreditation agencies*
Methods of analysis
Stratified categorical analysis
Supervised learning via penalised regression
Unsupervised learning via hierarchical clustering
Stratified categorical analysis
Organisations are stratified into two or more groups based on observed factorsResponses arranged into contingency tables Fisher’s exact conditional testing is appliedThe results are assessed based on the plots of sorted P-values
Stratified categorical analysis
Stratification factors:
1.Purchasing Power Parity adjusted GDP per capita2.Government relatedness3.Regional factor4.Age of organisation5.Country of origin population
Stratified categorical analysis
0.0
0.2
0.4
0.6
0.8
1.0
GDP PPP factor (US$20,000 split value)
Question IDs
P-v
alue
s
pol1
pol3
tra5
rep1
pol7
agh1
2re
p10
phc1
4fu
n1po
l16
phc1
2re
p4 int6
phc1
7fu
n5de
v8re
p13
int3
rep1
1fu
n4 tra1
phc8
agh8
pol1
9po
l10
rep2
pol6
agh1
4po
l13
rep1
2fu
n2re
p16
tra9
fun8
fun6
phc1
6po
l18
rep5 tra
6fu
n10
rep3 int1
pol1
7fu
n9re
p7re
p14
rep1
5ag
h16
agh1
7
0.05
Supervised learning via penalised regression
R package penalized: Penalised regressions via L1and L2 penalties [Goeman (2010): Biometrical Journal 52 (1), 70–84]L1: LASSO; L2: ridge; L1 and L2: elastic netsOnly L1 penalty (LASSO) has been usedSupervisory (response) variable is an ordinal four-level categorical variable defined by field experts (similar to phenotypes defined by clinical experts)Explanatory variables are categorical responses coded as factorsLogistic link function has been used
Supervised learning via penalised regression
LASSO [Tibshirani (1996): J. Royal. Statist. Soc. B 58 (1): 267–288] :
Supervised learning via penalised regression
Vitality categorisation (“phenotype”):
0 – on “life support” and unlikely to survive1 – hardly functional and deteriorating2 – functional, with signs of improvement3 – alive and well
Supervised learning via penalised regression
The set of shifted (“phenotype”) response variables:
yi,1 – refined “deprived”: 0 vs. 1-2-3
yi,2 – mid split: 0-1 vs. 2-3
yi,3 – refined “winners”: 0-1-2 vs. 3
Supervised learning via penalised regression
4.0 3.5 3.0 2.5 2.0 1.5
-0.5
0.0
0.5
1.0
1.5
LASSO path
lambda1
coef
ficie
nt
Supervised learning via penalised regression
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
LASSO path: Entrance distances
lam
bda1
tra5a
rep1
dre
p11b
pol3
apo
l3c
pol1
6apo
l10b
pol1
cre
p14a
rep1
3gpo
l1a
rep5
cre
p13h
tra9e
rep1
0bpo
l18e
pol1
3bfu
n4c
fun4
bpo
l19c
fun1
fpo
l19b
pol3
apo
l16c
rep1
6apo
l19d
tra6a
Supervised learning via penalised regression
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
LASSO path: Entrance distances
lam
bda1
tra5a
rep1
dre
p11b
pol3
apo
l3c
pol1
6apo
l10b
pol1
cre
p14a
rep1
3gpo
l1a
rep5
cre
p13h
tra9e
rep1
0bpo
l18e
pol1
3bfu
n4c
fun4
bpo
l19c
fun1
fpo
l19b
pol3
apo
l16c
rep1
6apo
l19d
tra6a
Supervised learning via penalised regression
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
LASSO path: Entrance distances
lam
bda1
tra5a
rep1
dre
p11b
pol3
apo
l3c
pol1
6apo
l10b
pol1
cre
p14a
rep1
3gpo
l1a
rep5
cre
p13h
tra9e
rep1
0bpo
l18e
pol1
3bfu
n4c
fun4
bpo
l19c
fun1
fpo
l19b
pol3
apo
l16c
rep1
6apo
l19d
tra6a
Supervised learning via penalised regression
“Winners” refined 50‐50 “Losers” refined
tra5a pol3a pol3a
rep1d pol13a pol10b
rep11b fun5b pol13a
pol3a dev8a rep11b
pol3c tra9b pol7a
pol16a rep11b fun1a
pol10b pol1a int6b
pol1c rep14b tra1c
rep14a tra5a fun10a
rep13g pol16a pol7d
pol1a fun4d pol19a
rep5c pol16d pol16d
rep13h fun1a rep3a
tra9e pol1c rep15a
rep10b fun8g tra1b
pol18e pol10b fun8g
pol13b pol6k fun4d
Supervised learning via penalised regression
“Winners” refined 50‐50 “Losers” refined
tra5a pol3a pol3a
rep1d pol13a pol10b
rep11b fun5b pol13a
pol3a dev8a rep11b
pol3c tra9b pol7a
pol16a rep11b fun1a
pol10b pol1a int6b
pol1c rep14b tra1c
rep14a tra5a fun10a
rep13g pol16a pol7d
pol1a fun4d pol19a
rep5c pol16d pol16d
rep13h fun1a rep3a
tra9e pol1c rep15a
rep10b fun8g tra1b
pol18e pol10b fun8g
pol13b pol6k fun4d
Supervised learning via penalised regression
“Winners” refined 50‐50 “Losers” refined
tra5a pol3a pol3a
rep1d pol13a pol10b
rep11b fun5b pol13a
pol3a dev8a rep11b
pol3c tra9b pol7a
pol16a rep11b fun1a
pol10b pol1a int6b
pol1c rep14b tra1c
rep14a tra5a fun10a
rep13g pol16a pol7d
pol1a fun4d pol19a
rep5c pol16d pol16d
rep13h fun1a rep3a
tra9e pol1c rep15a
rep10b fun8g tra1b
pol18e pol10b fun8g
pol13b pol6k fun4d
Unsupervised learning via hierarchical clustering
R function daisy (package cluster) has been used207 categorical attributes have been coded as asymmetric binomialsDissimilarity matrix has been computed based on Gower’s dissimilarity coefficient [Gower (1971) Biometrics 27,
857–874 ]Dendrograms are based on Ward’s hierarchical clustering methodWeights are imposed as an inverse of attribute counts within each category
Unsupervised learning via hierarchical clustering
Bos
nia_
Her
zego
vina
_AA
QI
Kaz
akhs
tan
Portu
gal
Fran
ceS
outh
_Kor
eaB
ulga
riaLe
bano
nPh
ilippi
nes
Rom
ania
Sau
di_A
rabi
aC
olom
bia
Cro
atia
Spa
inC
zech
_Rep
ublic
Net
herla
nds Br
azil
Indi
aU
SA_
JCI
Mon
golia
Kyr
gyzs
tan
Ser
bia
Bos
nia_
Her
zego
vina
_AKA
ZPo
land
Sou
th_A
frica
Ger
man
yJo
rdan
Arg
entin
aJa
pan
Sw
itzer
land
US
A_D
NV
HC
Taiw
anU
SA
_Joi
n_C
omis
sion Den
mar
kTh
aila
ndA
ustra
lia_A
CH
SM
alay
sia
Aus
tralia
_QIC
Eng
land
New
_Zea
land
_Tel
arc
Alb
ania
Can
ada
Lith
uani
aA
ustra
lia_A
GP
AL
New
_Zea
land
_HD
AN
Z
All categorical attributes
Unsupervised learning via hierarchical clustering
Bos
nia_
Her
zego
vina
_AA
QI
Kaz
akhs
tan
Portu
gal
Fran
ceS
outh
_Kor
eaB
ulga
riaLe
bano
nPh
ilippi
nes
Rom
ania
Sau
di_A
rabi
aC
olom
bia
Cro
atia
Spa
inC
zech
_Rep
ublic
Net
herla
nds Br
azil
Indi
aU
SA_
JCI
Mon
golia
Kyr
gyzs
tan
Ser
bia
Bos
nia_
Her
zego
vina
_AKA
ZPo
land
Sou
th_A
frica
Ger
man
yJo
rdan
Arg
entin
aJa
pan
Sw
itzer
land
US
A_D
NV
HC
Taiw
anU
SA
_Joi
n_C
omis
sion Den
mar
kTh
aila
ndA
ustra
lia_A
CH
SM
alay
sia
Aus
tralia
_QIC
Eng
land
New
_Zea
land
_Tel
arc
Alb
ania
Can
ada
Lith
uani
aA
ustra
lia_A
GP
AL
New
_Zea
land
_HD
AN
Z
All categorical attributes
Unsupervised learning via hierarchical clustering
Fran
ceSo
uth_
Kor
eaBo
snia
_Her
zego
vina
_AKA
ZB
ulga
riaLe
bano
nAl
bani
aPh
ilipp
ines
Pola
ndLi
thua
nia
Rom
ania
Kaza
khst
anPo
rtuga
lS
erbi
aBo
snia
_Her
zego
vina
_AAQ
IKy
rgyz
stan
Sau
di_A
rabi
aS
pain
Net
herla
nds Mon
golia
Arge
ntin
aSw
itzer
land Ja
pan
New
_Zea
land
_Tel
arc
Sou
th_A
frica
Jord
anC
zech
_Rep
ublic
Ger
man
yB
razi
lN
ew_Z
eala
nd_H
DAN
ZEn
glan
dC
anad
aU
SA_J
CI
Aust
ralia
_QIC
Mal
aysi
aTa
iwan
US
A_Jo
in_C
omis
sion
Aust
ralia
_AG
PAL
US
A_D
NV
HC
Indi
aC
olom
bia
Cro
atia Th
aila
ndAu
stra
lia_A
CH
SD
enm
ark
Policy and governance
Example of results triangulationIt has been identified that the legal status of organization is associated with the GDP PPP per capita factor (stratified analysis) => This association can be attributed to economic/social factors beyond the direct controlRelatively wealthy Portugal fell into a “deprived” cluster in the Policy and Governance motivated dendrogram (cluster analysis)POL3a (What is the legal status of the accreditation organization?: government agency) has been identified as highly important attribute contributing to failure (LASSO)It can be recommended to change the legal status
Challenges and extensions
Missing variablesAttributes interactionsNonlinear associations
Risk factors analysisPredictive modelingRegulatory impact analysis