risk-adjusted control charts for non- homogeneous dichotomous events qprc annual conference may 26,...
TRANSCRIPT
Risk-adjusted Control Charts for Non-Risk-adjusted Control Charts for Non-homogeneous Dichotomous Eventshomogeneous Dichotomous Events
QPRC Annual ConferenceQPRC Annual Conference
May 26, 2005May 26, 2005
James Benneyan, Ph.D.Quality and Productivity Lab, Director
Senior Fellow, Institute for Healthcare Improvement334 Snell Engineering Center
Northeastern University, Boston MA
www.coe.neu.edu/Research/[email protected]
www.coe.neu.edu/Research/QPL
Outline
1. Preliminary comments on healthcare SPC
2. Non-homogeneity- Types of applications- Non-homogeneous probability model- Variance inequality proof & other results
3. Risk-adjusted control charts- Shewhart, standardized, EWMA- k-sigma & probability limits- Examples (Infection control, mortality, reliability composite measures,
pain management, hand-washing compliance, others)- SPRT-Cusum
4. Performance & comparisons
www.coe.neu.edu/Research/QPL
Typical healthcare applications of SPC
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep Oct
Nov
Dec Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep Oct
Nov
Dec Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Month
Per
cent of Pat
ients
Rep
lyin
g "
Ver
y Sat
isfied
" or
"Exc
elle
nt"
Multiple changes tested
Subgroup Number
Mor
taliti
es /
1000
Disc
harg
es
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40
Surgical Site Infections
Subgroup (Month) Number
Avera
ge T
ime (
Min
s)
An
tib
ioti
c is
Ad
min
iste
red
Befo
re 1
st
Incis
ion
-200
-100
0
100
200
300
4/9
3
5/9
3
6/9
3
7/9
3
8/9
3
9/9
3
10/9
3
11/9
3
12/9
3
1/9
4
2/9
4
3/9
4
4/9
4
5/9
4
6/9
4
7/9
4
8/9
4
9/9
4
10/9
4
11/9
4
12/9
4
1/9
5
2/9
5
3/9
5
4/9
5
5/9
5
6/9
5
7/9
5
8/9
5
9/9
5
UCL
CL
LCL
Trial X-bar Control ChartPerioperative Antibiotic Timing
X-bar Chart
Patient Satisfaction
Fall Rate
0
0.5
1
1.5
2
2.5
3
3.5
11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9
Fiscal Period
Fal
ls/1
000
pat
ien
t d
ays
Falls and Slips
www.coe.neu.edu/Research/QPL
Public health applications
Scotland Surveillance for Regional MRSA Regional Respiratory Illness (USAF)
Infectious Disease SurveillanceBinary Cusum
Fraction Reporting Used Opioids in Last 30 DaysStandardized p EWMA control chart: Region 5
-1.25
-1.00-.75
-.50
-.25
.00
.25
.50
.75
1.001.25
1.50
24 34 44 54 64 74 84 94
Week
EW
MA
Sta
tist
ic
www.coe.neu.edu/Research/QPL
0
25
50
75
100
125
150
175
5 10 15 20 25 30 35 40 45
Procedure Number
EW
MA
of
Pro
ced
ure
s B
etw
een
C
om
pli
cati
on
s
EWMA: Exponentially weighted moving average
SPC 102: 80/20 rule… (95/5?)
Quarter
Adv
erse
Eve
nt R
ate
UCL = ???
CL = ???
LCL = ???
START-UP SPC METHODSAdverse Event Rate
RARE EVENTSEWMA g chart of Cases Between Complications
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
3
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
z
LCL
CL
UCL
"Incorrect" z
RISK-ADJUSTED SPC
Standardized jb EWMA Chart Incorrect
Correct
www.coe.neu.edu/Research/QPLSample
Tsquare
d
7654321
5
4
3
2
1
0
Median=4.652
UCL=5.143
LCL=0.555
181 - Tsquared Chart of C1, ..., C5
Patient monitoring and control
Observation
Indiv
idual V
alu
e
87654321
10
8
6
4
2
0
_X=4.13
UCL=9.44
LCL=-1.19
Pain
ObservationIn
div
idual V
alu
e
87654321
10
8
6
4
2
0
-2
_X=3.38
UCL=9.07
LCL=-2.32
Fatigue
Observation
Indiv
idual V
alu
e
87654321
7
6
5
4
3
2
_X=4.125
UCL=6.405
LCL=1.845
Nausea
Observation
Indiv
idual V
alu
e
87654321
8
6
4
2
0
-2
-4
_X=2.63
UCL=7.94
LCL=-2.69
Sleep
MCEWMA Chart for Systolic Blood Pressure Data
95
105
115
125
135
145
0 50 100 150 200 250
Pain Fatigue Nausea Sleep Distress
Pain 1.00
Fatigue 0.75 1.00
Nausea 0.60 0.82 1.00
Sleep 0.50 0.56 0.52 1.00
Distress 0.76 0.80 0.77 0.60 1.00
Bounded Adjustment Chart with IMA(1,1) Data
-4
0
4
8
12
0 100 200 300 400 500
z(t)
e(t)
EWMA
+L
-L
Anticoagulation Feedback Adjustment Example
www.coe.neu.edu/Research/QPL
Input Populations
Inputs Risk AdjustmentMethods
(Log-Regression)(others)
X1 ~ Binomial(n1,p1)
X2 ~ Binomial(n2,p2)
Xk~ Binomial(nk,pk)
•
•
•
•
•
(NB: ni = 1 often)
Aggregate Model
T = X1 + X2 + ... + Xk
E(T) nipii1
k
V(T) ni(1 pi)pii1
k
GT (s) (1 pspi)ni
i1
k
j
ij,i nN j
ij,i xT iii /NTF
Fractionx1,i n1,i x2,i n2,i x3,i n3,i x4,i n4,i Ti Ni Fi
1 1 1 1 1 0 3 0 0 2 5 0.400002 0 6 0 14 1 5 0 2 1 27 0.037043 1 11 0 4 0 0 0 1 1 16 0.062504 0 1 0 0 0 4 1 1 1 6 0.166675 1 11 1 9 1 2 0 0 3 22 0.136366 0 1 1 12 1 7 1 5 3 25 0.120007 1 2 0 3 1 1 0 1 2 7 0.28571. . . . . . . . . . . .. . . . . . . . . . . .. . . . . . . . . . . .
m x1,m n1,m x2,m n2,m x3,m n3,m x4,m n4,m Tm Nm Fm
Totals
Week
Process 1 (j=1) Process 2 (j=2) Process 3 (j=3) Process 4 (j=4)
J Non-Homogeneous Sub-Populations, ni ≥ 1 i
i weekin SSI of number Expected E ii weekin SSI of number of deviation Standard S i
1 0.10732 0.06603 0.0349
. . . . . .
. . . . . .
. . . . . .1 0.10732 0.02043 0.05134 0.09805 0.0955
1
m
Risk-Adjusted SSI LikelihoodWeek Patient
Actual No. SSI Ei Si
0.2082 0.1911
0.3725 0.3392
3
4
Unique likelihoods, ni = 1 i
Two general types of non-homogeneous data
www.coe.neu.edu/Research/QPL
Healthcare examples of this type of data
SSI Stratified by NNIS Category
Time Period
x1,t n1,t … xJ,t nJ,t
1 37 49 18 222 22 32 6 113 37 43 15 194 41 43 16 195 46 76 13 206 55 72 19 227 32 39 19 338 48 53 20 219 35 37 29 3710 56 68 12 2111 57 58 19 2412 18 22 1 213 6 11 5 514 15 19 6 815 16 19 8 816 13 20 3 417 19 22 3 418 19 33 12 1419 20 21 19 1920 29 37 15 1521 12 21 6 622 19 24 19 19l
l
l
T 57 58 19 24
1Hospital / Provider
J
Hand-washing Compliance
Subgroup n1,t x1,t n2,t x2,t n3,t x3,t n4,t x4,t nt xt
1 41 2 38 2 17 5 1 0 97 92 60 5 34 0 17 6 1 1 112 123 50 1 46 3 15 5 2 2 113 114 47 1 32 4 20 4 4 2 103 115 48 0 36 5 10 2 1 1 95 86 36 0 41 4 12 2 2 1 91 77 64 0 25 0 10 4 3 0 102 48 44 1 33 2 20 3 3 1 100 79 45 4 32 2 11 3 4 0 92 9
10 57 2 24 1 17 3 3 2 101 811 52 1 28 4 15 4 4 3 99 1212 54 0 32 1 16 5 0 0 102 613 38 2 32 2 8 1 0 0 78 514 25 3 16 3 9 3 0 0 50 915 20 1 19 6 5 2 0 0 44 916 19 0 18 1 7 3 0 0 44 417 2 0 4 0 4 2 0 0 10 2
Total 702 23 490 40 213 57 28 13 1433 133p1 0.0328 p2 0.0816 p3 0.2676 p4 0.4643 p 0.8463
TotalCategory 1 Category 2 Category 3 Category 4Week
Note: Sample sizes change week-to-weekNote: Sample sizes change week-to-week
www.coe.neu.edu/Research/QPL
Industry examples of this type of data
Time Period
Sample Size
Number Defects
Sample Size
Number Defects
…Sample
SizeNumber Defects
Total sample
Total Defects
1 10 3 15 3 25 6
2 12 2 23 2 35 4
3
:
etc
Process / Product J TotalProcess / Product 1 Process / Product 2
Other applications
• Manufacturing defects of different lines, products, etc.
• Automobile accidents across different driver types
• Number on-time shipments for different vendors
• Airline no-shows for different types of passengers
• k of n system reliability
• Basketball team’s successful free throws
www.coe.neu.edu/Research/QPL
Let: T = X1 + X2 + … + XJ and F = T/N, where Xj ~ binomial(nj,pj)
All (Xi, Xj) independent but not identically distributed
E(Xj) = njpj, V(Xj) = njpj(1 - pj), GXj(s) = (1 – pj + spj)nj
Then:
(MGF, skewness, kurtosis: See paper)
Note: GT(s) is not of binomial form (except for trivial cases J = 1 or pi = pj (i, j)
E(T) has binomial form but V(T) does not have binomial form (see below)
Binomial Approximation
Assume T’ binomial(N,P), with
Then:
But:
In fact V(T) < V(T’) (see below)
J-binomial random variable
J
i iiJJ pnpnpnpnTE12211 ... )(
J
i iiiJJJ ppnppnppnppnTV1222111 )1( )1(...)1()1( )(
J
i
nii
nJJ
nnT
iJ sppsppsppsppsG12211 )1( )1(...)1()1( )( 21
211
2
11)1( /)( )(, )/( )(
J
i i
J
i iii
J
i i
J
i ii nppnNTVFVnpnNTEFE
iJiii
Jiii
Jii
Ji npnNnpPnN 1111 and approx
~
)()'(1111
TEpnnpnnNPTEJ
i ii
J
i i
J
i ii
J
i i
)()1(1)1()'(111111
TVppnnpnnpnnPNPTVJ
i iii
J
i i
J
i ii
J
i i
J
i ii
J
i i
www.coe.neu.edu/Research/QPL
Mixed-risk probability modelJ-binomial probability distribution3
nconvolutio recursive , )()(0 1
11
1
1
t
x
J
xii
ii xtXPxXPtTP
t
x
xt
x
xt
xkk
xxt
x
k
i i
k
xXPxXPxXPxXP0 0 00
332211
1
1
2
121
3
)...( ... ...)()()(
.........)()( ... )...( ... 1
10
110
2
1
1
1
Ji iJ
xt
xJJ
xt
xkk xtXPxXPxXP
J
i i
J
k
i i
k
yrecursivel ,
)1(
)1(1
0 , 11where
1
i
i
i
iii x
k i
ini
in
i-xn
ix
ii
iii
kp
knp)-p(
x)-p( )-p(p
x
n)x P(X
P(F = f) = P(T = t), where f = t/N
cases lbut trivia all in eintractabl very although ,)()(or 0
s
Tt
t
sGds
dtTP MLE) (MOM, ˆ
1 ,1 ,
m
j ji
m
j jii nxp
www.coe.neu.edu/Research/QPL
0.00
0.05
0.10
0.15
0.20
0.25
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
t
P(T
=t)
0.00
0.05
0.10
0.15
0.20
0.25
J-BINOMIAL
BINOMIAL
NORMAL-BIN
NORMAL-JB
TAD ValuesJ-Bin vs Bin: 0.1931J-Bin vs J-Bin Norm al: 0.0074
n0=2 p0=0.01n1=2 p1=0.1n2=2 p2=0.2n3=3 p3=0.3n4=3 p4=0.4n5=3 p5=0.5n6=3 p6=0.6n7=3 p7=0.7n8=3 p8=0.8n9=3 p9=0.9
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
t
P(T
=t)
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40J-BINOMIAL
BINOMIAL
NORMAL-BIN
NORMAL-JB
TAD ValuesJ-Bin vs Bin: 1.0485J-Bin vs J-Bin Norm al: 0.2404
n0=10 p0=0.99n1=100 p1=0.01
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
t
P(T
=t)
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
J-BINOMIAL
BINOMIAL
NORMAL-BIN
NORMAL-JB
TAD ValuesJ-Bin vs Bin: 1.3001J-Bin vs J-Bin Norm al: 0.1004
n0=10 p0=0.99n1=10 p1=0.01
0.00
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
0.20
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
t
P(T
=t)
0.00
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
0.20J-BINOMIAL
BINOMIAL
NORMAL-BIN
NORMAL-JB
TAD ValuesJ-Bin vs Bin: 0.0000J-Bin vs J-Bin Norm al: 0.0121
ni=2 pi=0.5(i=0,1,…10)
Why Does/Might This Matter?
Note: JB ≤ B
J
j jjjN
J
j jjN ppnTVpnTE1
2
1)1()(,)( :ionapproximat Normal
www.coe.neu.edu/Research/QPL
Shewhart Control Charts(same examples)
Example4, Shewhart
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
p
p
UCL-bin
UCL
CL
LCL
LCL-bin
Example1, Shewhart
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
p
p
UCL-bin
UCL
CL
LCL
LCL-bin
Error
Error
Example2, Shewhart
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
p
p
UCL-bin
UCL
CL
LCL
LCL-bin
Error
Error
Example3, Shewhart
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
p
p
UCL-bin
UCL
CL
LCL
LCL-bin
Error
Error
3 standard deviation control limits, Lower power (1 – ), lower false alarm rate (),
www.coe.neu.edu/Research/QPL
Comparison of VariancesJ-binomial vs. Binomial
Example 1
J = 2 n1 = 10 p1 = .99
n2 = 10 p2 = .01
n = 10, sn = 0, p = .50, sp = .693
V(T) =
V(T’) = NP(1-P) = 5.0
V(T)/V(T’) = .0396
KL =
TAD = 1.3001
198.)1(1
J
i iii ppn
Example 3
J = 10 n1 = 2 p1 = .01 n6 = 3 p6 = .5
n2 = 2 p2 = .1 p7 = 3 p7 = .6
n3 = 2 p3 = .2 p8 = 3 p8 = .7
n4 = 3 p4 = .3 p9 = 3 p9 = .8
n5 = 3 p5 = .4 p10 = 3 p10 = .9
n = 2.7, sn = .483, p = .451, sp = .301
V(T) = 4.720 V(T’) = 6.797
V(T)/V(T’) = .700 KL = TAD = .1931
Example 4 (binomial)
J = 10 n1 = 2 p1 = .5 n6 = 2 p6 = .5
n2 = 2 p2 = .5 p7 = 2 p7 = .5
n3 = 2 p3 = .5 p8 = 2 p8 = .5
n4 = 2 p4 = .5 p9 = 2 p9 = .5
n5 = 2 p5 = .5 p10 = 2 p10 = .5
n = 2, sn = 0, p = .5, sp = 0
V(T) = 5 V(T’) = 5
V(T)/V(T’) = 1.000 KL = TAD = 0.000
Example 2
J = 2 n1 = 10 p1 = .99
n2 = 100 p2 = .01
n = 55, sn = 63.64, p = .50, sp = .693
V(T) = 1.089
V(T’) = 9.820
V(T)/V(T’) = .111
KL =
TAD = 1.0485
www.coe.neu.edu/Research/QPL
Let: T ~ J-binomial(p1, p2, …, pJ, n1, n2, …, nJ) with pi ≠ pk for at least some (i, k)
T’~ binomial(N,P) with
Assume: p1 ≤ p2 ≤ … ≤ pJ without loss of generality such that p1 = p(1) (min), …, pJ = p(J) (max)
pi ≠ pk for at least one (i, k)
Define: 1 ≥ 2 ≥ … ≥ J-1 such that i = pJ - pi
i.e., p1 = pJ – 1, p2 = pJ – 2, …, pJ-1 = pJ – J-1
Then: since i k k i and i > k for at least one (i, k)
After some algebra it can be shown that
and then after additional work that
V(T) < V(T’)
(Also: Skewness of T < skewness of T’)
Under-dispersion: V(T) ≤ V(T’)Outline of Proof3
iJiii
Jii
Ji npnPnN 111 and
0)( 2211 kikiik
Ji nn
211
221 ii
Jii
Jiii
Ji pnnpn
J
i i
J
i ji
J
i i
J
i jiJ
i ijj
J
i in
pn
n
pnnPNPppn
1
1
1
111
1)1()1(
www.coe.neu.edu/Research/QPL
Statistical jb Control Charts2,4 k-sigma Shewhart Limits
LCLt
UCLt
Shewhart np-type jb chart
CLt
Plotted Statistict
Shewhart p-type jb chart
where xj,t ~ binomial(nj,t,pj)
nj,t = sub-subgroup size of sub-population j at time t
xj,t = number of adverse events in sub-population j at time t
J = number of sub-populations, t = time period
J
j jtjt pnTE1 , )(
J
j jjtj
J
j jtj ppnkpn1 ,1 , )1(
J
j tjt xT1 ,
J
j tj
J
j jtjttt npnNTEFE1 ,1 , )/( )(
21 ,1 ,1 ,1 , )1(
J
j tj
J
j jjtj
J
j tj
J
j jtj nppnknpn
21 ,1 ,1 ,1 , )1(
J
j tj
J
j jjtj
J
j tj
J
j jtj nppnknpn
J
j tj
J
j tjt nxF1 ,1 ,
J
j jjtj
J
j itj ppnkpn1 ,1 , )1(
www.coe.neu.edu/Research/QPL
Probability jb Limits
CLi
UCLi
np-type jb chart
LCLi
Plotted Statistict
p-type jb chart
Note: CL (mean and median) of all non-standardized charts (including p-type) vary longitudinally unless all ni’s remain constant
See papers2,4 for standardized cases and examples
Find limits numerically, possibly starting with LCLT’, UCLT’, and CLT’ approximations
LCLi tTPTtLCL
)(argmax,
)(
)()(
5.0,5.0,
1 ,
2
1medianTT
meanpnTE
ii
J
j jjii
J
j jtt xT1 ,
J
j jt
J
j jtt nxF1 ,1 ,
UCLi tTPTtUCL
)(argmin,
J
j jii nTLCL 1 ,,
J
j jii nTUCL 1 ,,
)(
)()(
1 ,5.0,5.0,
1 ,1 ,
2
1mediannTT
meannpnFE
J
j jiii
J
j ji
J
j jjii
www.coe.neu.edu/Research/QPL
J-Binomial UCL:Standard UCL:
J-Binomial CL:Standard CL:
J-Binomial LCL:Standard LCL:
iti
t iti
t iti
t iti
t iti
t iti
t iti
n
n
x
n
x
n
x
,
,
,
,
,
,
,
1
3
t iti
t iti
n
x
,
,
iti
t iti
t iti
t iti
t iti
t iti
t iti
n
n
x
n
x
n
x
,
,
,
,
,
,
,
1
3
iti
it
ti
tti
tti
tti
ti
iti
it
ti
tti
ti
n
n
x
n
xn
n
n
xn
,
,
,
,
,
,
,
,
,
,1
3
iti
it
ti
tti
ti
n
n
xn
,
,
,
,
iti
it
ti
tti
tti
tti
ti
iti
it
ti
tti
ti
n
n
x
n
xn
n
n
xn
,
,
,
,
,
,
,
,
,
,1
3
Parameters Estimated jb Chartsnp-type jb Shewhart case
periods timeofnumber where, ...,2,1, ˆ :binomial
ˆ :Binomial
1 ,1 ,
1 1 ,1 1 ,1 ,1 ,
mJjnxpJ
nxNtp
m
i ji
m
i jij
m
i
J
j ji
m
i
J
j ji
m
i ji
m
i ji
www.coe.neu.edu/Research/QPL
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
3
26 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 44
Week
Zt
Zt
Zt bin
UCL
CL
LCL
-1.5
-1
-0.5
0
0.5
1
1.5
2
Week
SS
I Ra
te
p-fract
UCL-bin
UCL
CL
CL-bin
LCL
LCL-bin
Surgical Site Infection Examples: Moving CL (Non-constant mean)
Russian SSI’s (j: NNIS scores)
0
0.05
0.1
0.15
0.2
0.25
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Subgroup
Fra
cti
on
No
nc
on
form
ing
p
LCL
CL
UCL
Std LCL
Std CL
Std UCL
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
3
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
z
LCL
CL
UCL
"Incorrect" z
Georgia Hospital SSI’s (j: hospitals)
Not standardized Not standardized
Standardized Standardized
Incorrect
Correct
www.coe.neu.edu/Research/QPL
Standardized jbin Charts
LCLt
UCLt
Standardized Shewhart*
CLt
Plotted Statistict
Standardized EWMA *
where E(Z) = 0, V(Z) = 1, f(z) unknown (j-binomial)
= EWMA weight
* Standardized charts are the same for fraction (F) or total (T)
tt kLCL 2)1(1)2(
tt kUCL 2)1(1)2(
0tCL
12
1 ,1 ,
1 ,1 ,)1(
)1(
tJ
j tj
J
j jjtj
J
j tj
J
j jtjt
t ynppn
npnFy
3 kLCL t
3kUCLt
0tCL
2
1 ,1 ,
1 ,1 ,
)1(
J
j tj
J
j jjtj
J
j tj
J
j jtjt
t
nppn
npnFz
2
1 ,1 ,
1 ,1 ,
1 ,
1 ,
)1(
)(
)(
)1(
)(
)(
J
j tj
J
j jjtj
J
j tj
J
j jtjt
t
tt
J
j jjtj
J
j jtjt
t
ttt
nppn
npnF
FV
FEF
ppn
pnT
TV
TETZ
www.coe.neu.edu/Research/QPL
Standardized EWMA Charts(Same Examples)
Example2, Standardized EWMA
-1.5
-1
-0.5
0
0.5
1
1.5
Zt
Zt
Zt bin
UCL
CL
LCL
Example3, Standardized EWMA
-1.5
-1
-0.5
0
0.5
1
1.5
Zt
Zt
Zt bin
UCL
CL
LCL
Example4, Standardized EWMA
-1.5
-1
-0.5
0
0.5
1
1.5
Zt
Zt
Zt bin
UCL
CL
LCL
Example1, Standardized EWMA
-1.5
-1
-0.5
0
0.5
1
1.5
Zt
Zt
Zt bin
UCL
CL
LCL
2
1 ,1 ,
1 ,1 ,
)1(
J
i ti
J
i iiti
J
i ti
J
i itit
t
nppn
npnFz
Note: Difference now is in plotted points (limits are the same)
www.coe.neu.edu/Research/QPL
Interpretation errorsOther applications
Pain Management Bundle
Percent Patients with Pain Assessment and Plan of Care
-5
-4
-3
-2
-1
0
1
2
3
4
Jun-00
Jul-00
Aug-00
Sep-00
Oct-00
Nov-00
Dec-00
Jan-01
z*
z*bin
UCL
CL
LCL
Percent Patients with Pain Assessment and Plan of Care
-1.5
-1
-0.5
0
0.5
1
1.5
Jun
-00
Jul-0
0
Au
g-0
0
Se
p-0
0
Oct-0
0
No
v-00
De
c-00
Jan
-01
Zt
Zt bin
LCL
CL
UCL
Handwashing Compliance forj = 6 MD’s)
Standardized
Standardized-6
-5
-4
-3
-2
-1
0
1
2
3
4
11
/27
/20
00
12
/20
/20
00
12
/26
/20
00
12
/28
/20
00
1/3
/20
01
1/1
0/2
00
1
1/2
0/2
00
1
2/2
0/2
00
1
3/2
1/2
00
1
4/1
9/2
00
1
5/3
1/2
00
1
6/3
0/2
00
1
7/3
0/2
00
1
z*
z*bin
UCL
CL
LCL
-2
-1.5
-1
-0.5
0
0.5
1
1.5
11
/27
/20
00
12
/20
/20
00
12
/26
/20
00
12
/28
/20
00
1/3
/20
01
1/1
0/2
00
1
1/2
0/2
00
1
2/2
0/2
00
1
3/2
1/2
00
1
4/1
9/2
00
1
5/3
1/2
00
1
6/3
0/2
00
1
7/3
0/2
00
1
Zt
Zt bin
LCL
CL
UCL
Standardized
Standardized
www.coe.neu.edu/Research/QPL
60
70
80
90
100
110
Subgroup Number
Incorrect limits Correct limits
CL
105.93
98.18
71.88
64.13
Error
Error
n2 = 103p2 = .06
k = 2n1 = 95p1 = .83
Example
www.coe.neu.edu/Research/QPL
.00
.20
.40
.60
.80
1.00
.45 .50 .55 .60 .65 .70 .75 .80 .85 .90 .95
Shifted Value of p1
Pro
ba
bil
ity
of
Sig
na
l Correct limits Incorrect limits
Example of performance
Example
2 sub-populations (J = 2)
n1 = 95 p1 = .83
n2 = 103 p2 = .06
Correct variance = 77.87 (jb)
Incorrect variance = 48.51 (bin)
• Correct limits roughly 1/3 closer to CL
• Significantly faster change detection
• Average time-to-detect (pop 1):
Shift to Correct Conventional.75 10 points 300 points.70 2-3 points 20 points.65 1-2 points 5 points
www.coe.neu.edu/Research/QPL
Relative performanceAverage run lengths. More detailed results in appendix & paper.
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1
X0
X1
X0
Avg n = 200, SD n = 14.1, Avg p = .5, SD p = .566, J = 2SR = 0.6005, T AD = 0.4850, KL = 0.0828
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
B X1B X0
JB X1
JB X0
Avg n = 200, SD n = 141.4, Avg p = .5, SD p = .566, J = 2SR = 0.6547, T AD = 0.4054, KL = 0.0601
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X0X0
X1
X1
Avg n = 200, SD n = 141.4, Avg p = .0505, SD p = .07, J = 2SR = 0.9263, T AD = 0.03704, KL = 6.16E-4
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1X0
Avg n = 200, SD n = 141.4, Avg p = .5, SD p = .028, J = 2SR = 0.9994, T AD = 5.816E-4, KL = 1.57E-7
Risk-Adjusted Not Risk-Adjusted
www.coe.neu.edu/Research/QPL
Factors affecting difference(Preliminary – more than just p)
25 factorial design: 5 main effects (J, n, sn, p, sp) and all interactions
Responses:
1. Kullback-Leibler information:
2. Total Absolute Deviation:
3. Standard Deviation Ratio:
Y = 0 + 1J + 2 n + 3ssn + 4 p + 5sp + 12J n + 13J sn + … + 123J snsp + …
J
i in
tBJB tTPtTPTAD
1
0
)()(
N
t B
JBJB tTP
tTPtTPKL
0 )(
)(log)('
J
i i
J
i ii
J
i i
J
i ii
J
i i
J
i iii
B
JB
npnnpnn
ppnSR
11111
12
22
1
)1(
Parameters Interactions Response
p J
Ave p (pbar)
Ave n (nbar)
n
p x J p x pbar
p x pbar x nbar
p x nbar
x J
p x pbar
n x nbar All others
SR 0 * * * * 0 * .046 * .001 *
TAD * * * * * 0 * 0 .008 * *
KL * * * * * 0 0 * .004 * *
* Not statistically significant (.05)
www.coe.neu.edu/Research/QPL
Other applications: Dr. Harold Shipman (GP)
• Britain’s most prolific serial killer
• Murdered 236 (215-289) elderly patients between 1971(75)-1998
by lethal injections of heroin (diamorphine)
• > 189 females, > 55 males
• Postulated related to mother’s death from cancer when Shipman was young
• National inquiry, media, analysis
• 15 life sentences, 1/31/00 Hangs himself in jail, 1/13/04
www.coe.neu.edu/Research/QPL
Risk-adjusted mortality rates
Male Standardized Risk-Adjusted Mortality Rate
-6
-3
0
3
6
9
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
Year
Mo
rtal
ity
Rat
e (S
tan
dar
diz
ed)
LCL
UCL
CL
Female Standardized Risk-Adjusted Mortality Rate
-6
0
6
12
18
24
30
36
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
Year
Mo
rta
lity
Ra
te (
Sta
nd
ard
ize
d)
LCL
UCL
CL
• Adjusted for patient acuity using comparison groups
• Standardized p charts
• Higher than expected mortality, increasing trends for both males and females
• Noticeable difference between male and female populations
• ARL = 1.7 years to detect this magnitude of rate difference (pre-trend, binary sprt)5
www.coe.neu.edu/Research/QPL
100K Lives Campaign^
• Save 100,000 lives by 6/14/06 via ‘bundles’ of proven interventions in six areas:
• Adverse drug events (ADE)• Surgical site infection (SSI)• Myocardial infarction (AMI)• Ventilator pneumonia (VAP)• Central line infection (CLI)• Rapid response teams
• More than 2,000 US hospitals participating
• Accomplish by consistent ‘bundle’ implementation using a reliability model
^ www.ihi.org/ihi/programs
Institute for Healthcare Improvement (IHI)
Background
• National leader of healthcare system improvement
• Safety, medical error, waits & delays, appointment access, specific clinical diagnoses, …
• Large-scale multi-hospital (15-70) problem area projects
• National network of leading hospitals, faculty, researchers
• Ties to Institute of Medicine, JCAHO, CMS, NAE, VA, BMJ…
• Modern Healthcare: 3rd most powerful person in healthcare
www.coe.neu.edu/Research/QPL
IHI growth & impact: 1986 - present
• Active in over 50 countries, internationally, and in developing countries
• Roughly 14,000 organizations involved
• Annual forum conference > 5,000 attendees worldwide
• > 200 nationally-recognized faculty, researchers, experts
1986
15,000
50
www.coe.neu.edu/Research/QPL
Examples of composite bundles^
AMI CHF CAP CABG TJ VAP
ASA (aspirin) within 24 hrs
BETA blocker within 24 hrs
Thrombolytics within 30 min
PCI within 120 minutes
ACE Inhibitor for LVSD
Smoking counseling
ASA at d/c
BETA blocker at discharge
LVEF assessed
ACE Inhibitor for LVSD
Detailed discharge instructions
Smoking cessation counseling
O2 assessed within 24 hours
Blood culture prior to abx
First dose of antibiotic within 4 hours
Influenza screening / vaccination
Pneumococ. vaccination
Smoking cessation counseling
First dose antibioitc 1 hour prior to first incision
Appropriate antibiotic selection
Discontinued antibiotics within 24 hours
Use internal mammary artery (IMA)
ASA at discharge
First dose antibioitc 1 hour prior to first incision
Appropriate antibiotic selection
Discontinued antibiotics within 24 hours
Head of bed elevation
DVT prophylaxis
Sedation vacation
PUD prophylaxis
Others for SSI, other disease diagnoses, rapid response teams, pain mgmt (see JCAHO, CMS, & IHI websites)
^ Center for Medicare & Medicaid Services (CMS), “Clinical conditions and measures for reporting and incentives” 2003; AHA; Leapfrog Group; National Quality Forum; AHQR; JCAHO Core Measures
www.coe.neu.edu/Research/QPL
Composite measures
Approach Explanation Example
Individual measures
% patients meeting that measure
- Fraction TJ patients receiving prophylactic antibiotic received one hour prior to surgical incision
CompositeTotal measures not met
Total # opportunities
(across all patients)
Total not met
3 measures x Number patients
“All-or nothing”
(Aggregate bundle)
Fraction patients meeting all measures (total bundle met,
‘perfect encounter’)
Fraction TJ patients with ALL 3:
- Prophylactic antibiotic 1 hour prior
- Prophylactic antibiotic selection for surgical patients
- Antibiotics discontinued within 24 hours of surgery end time
Front-line improvement and analysis
Ultimate goal
Used to measure
improvement
www.coe.neu.edu/Research/QPL
Composite example
Patient Measure 1 Measure 2 Measure 3 Measure 4 # Met All measures?
1 Yes Yes Yes Yes 4 Yes
2 Yes No Yes Yes 3 No
3 Yes Yes No No 2 No
4 Yes Yes Yes Yes 4 Yes
5 Yes Yes Yes No 3 No
6 Yes No Yes No 2 No
7 No Yes Yes No 2 No
8 Yes Yes Yes Yes 4 Yes
9 Yes No No Yes 2 No
10 Yes Yes Yes Yes 4 Yes
Reliability 9/10 (.9) 7/10 (.70) 8/10 (.8) 6/10 (.6) 30/40 4/10
Composite measure = (# met) / (# opportunities) = (9+7+8+6) / (10x4) = .75
www.coe.neu.edu/Research/QPL
jb Composite Control Chart
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 3 5 7 9 11 13 15 17 19 21 23 25
Week
Co
mp
osi
te S
core
Correct limits
Incorrect limits
Application to clinical evidence ‘bundles’
Individual measure Historical rate
1. Head of bed elevation 61%
2. DVT prophylaxis 75%
3. Sedation protocol 43%
4. PUD prophylaxis 88%
Ventilator Associated Pneumonia Bundle
Suppose tests 10% improvement in each measure
Week X1 X2 X3 X4
1 10 6 8 2 6 40 22 55%
2 10 4 9 5 10 40 28 70%
3 10 7 5 4 8 40 24 60%
4 10 5 7 4 9 40 25 62.5%
5 10 6 8 3 7 40 24 60%
etc etc etc etc etc etc etc etc etc
Number Patients
# Satisfying Measure Xi TotalOps
TotalMet
CompositeScore
Data Format
Undetectedimprovement
Undetected negative spike
www.coe.neu.edu/Research/QPL
Other composite reliability control charts
Columbia ADE Reliability(High hazard and regular risk populations (4) combined)
0
0.1
0.2
0.3
0.4
0.5
1 3 5 7 9 11 13 15 17 19 21 23
Correct
Incorrect
CAMC CAP jb Composite Reliability
0.70
0.80
0.90
1.00
12/2
7/04
1/3/
05
1/10
/05
1/17
/05
1/24
/05
1/31
/05
2/7/
05
2/14
/05
2/21
/05
2/28
/05
3/7/
05
3/14
/05
3/21
/05
3/28
/05
Correct LCL
Correct UCL
Baptist HF Composite Reliability
0.40
0.50
0.60
0.70
0.80
0.90
1.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Correct
Columbia ADE Reliability(High hazard and regular risk populations (4) combined)
-1.5
-1
-0.5
0
0.5
1
1.5
1 3 5 7 9 11 13 15 17 19 21 23
Correct
Incorrect
www.coe.neu.edu/Research/QPL
• If at least k of the n measures are met for a patient, the likelihood of the adverse outcome significantly decreases
• Non-linear “bundle effect” (observation)
• Safety culture and human factors phenomenon
• “Tipping point” of patient-centered care
• All-or-nothing: k = n
• n measures are not i.i.d.
• Usual k of n calculation is not binomial
“Bundle effect” and k of n reliability
Pro
babili
ty N
o A
E
.
0 k n
0
1.0
"tipping point"
www.coe.neu.edu/Research/QPL
Bundles of bundles (aggregate bundle)
Bundle # Measures
VAP 4
AMI 7
CHF 6
SSI 4
Total 21
Week VAP CHF AMI SSI
1 10 6, 8, 2, 6 210
2 10 4, 9, 5, 10 210
3 10 7, 5, 4, 8 210
4 10 5, 7, 4, 9 210
5 10 6, 8, 3, 7 210
etc etc etc etc etc etc etc etc etc
# Pat-ients
TotalOps
TotalMet
CompositeScore
Data Format
• Two aggregation trends
• Meet all individual measures across all campaign bundles
• Meet at least k of the 6 all-or-nothing campaign bundles
# Satisfying Measure Xi
www.coe.neu.edu/Research/QPL
SPRT Cusums
Standard homogeneous Bernoulli SPRT-Cusum. Given xi = m of ni = n failures where Xi=1 w.p. p0 or p1 i.i.d. under H0 or H1, respectively,
^ Sequential probability ratio tests for non-homogeneous Bernoulli and binomial events”, working paper. So=0, bo=B, ao=A
App
pp
pXXP
pXXPB
mnm
mnm
n
n
1log
)1(
)1(log
)0|,...(
)|,...(log
1log
00
11
1
11
)1(
)1(log
1
1log
)1(
)1(log
1log
)1(
)1(log
1
1log
)1(
)1(log
1log
10
01
0
1
10
01
10
01
0
1
10
01
pp
pp
p
pn
pp
ppx
pp
pp
p
pn
pp
ppi
Scenario Example SPRT-Cusum
One item at a time
(J=1, n=1)
Operate on one patient at a time. Interpret after each individual case.
J Bernoulli items at a time
(J>1, nj=1, j=1,…J)
Operate on J patients per day. Interpret at end of each day.a. Know which T of J items failedb. Don’t know (true j-Bernoulli)
J binomial samples at a time
(J>1, nj>1, j=1,…J)
Operate on nj patients in each of J risk categories.a. Know Xi’s in sub-samplesb. Don’t know (true j-binomial)
0
1
110
01
10
1
1
00
00
11
1
1log
)1(
)1(log
1
1log
)1)(1(
)1)(1(log
i
ii
ii
iiii
i
ii
ii
iiii
p
pa
pp
ppXS
p
pb
Apxpx
pxpxB
1log
),|(
),|(log
1log
),|()...,|(),|(
),|()...,|(),|(log
00
11
1
0002
0222
01
0111
111
12
1222
11
1111
iiiiJB
iiiiJBii
iiiiJBJBJB
iiiiJBJBJB
nptTP
nptTPXS
AnptTPnptTPnptTP
nptTPnptTPnptTPB
iii J
j ji
jijii
J
j jiji
jijijii
J
j ji
jijii
jijiJiiJJ
jijiJiiJJ
p
pna
pp
ppXS
p
pnb
AXXXXXXP
XXXXXXPB
10,
1,
,11
1,
0,
0,
1,
,11
0,
1,
,1
,0,,1,,21,2,11,1
,1,,1,,21,2,11,1
1
1log
)1(
)1(log
1
1log
,|),...),...(,...)(,...(
,|),...),...(,...)(,...(log
np
np
Following similar logic:
a.
b.
www.coe.neu.edu/Research/QPL
PerformanceAverage Run Lengths (ARL’s)
HHHHL ARLProbability Limits
1.00E+00
1.00E+01
1.00E+02
1.00E+03
1.00E+04
1.00E+05
1.00E+06
1.00E+07
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
B X1B X0
JB X1
JB X0
Avg n = 200, SD n = 141.4, Avg p = .5, SD p = .566, J = 2SR = 0.6547, T AD = 0.4054, KL = 0.0601
Comments
• Excessively low in-control false alarm rate (standard binomial chart)
• Detection power is significantly better (J-bin risk-adjusted chart)
Example: J = (≤) .001 probability limits
www.coe.neu.edu/Research/QPL
Effects and Interactions(p, p)
Low p High p
Lowpbar
Highpbar
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X0
X1
Avg n = 200, SD n = 141.4, Avg p = .01, SD p = .001, J = 2SR = 0.9999, T AD = 3.963E-5, KL = 6.91E-10
Risk-Adjus ted Not Risk-Adjus ted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X0X0
X1
X1
Avg n = 200, SD n = 141.4, Avg p = .0505, SD p = .07, J = 2SR = 0.9263, T AD = 0.03704, KL = 6.16E-4
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1X0
Avg n = 200, SD n = 141.4, Avg p = .5, SD p = .028, J = 2SR = 0.9994, T AD = 5.816E-4, KL = 1.57E-7
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1X0
X1
X0
Avg n = 200, SD n = 141.4, Avg p = .5, SD p = .566, J = 2SR = 0.6547, T AD = 0.4054, KL = 0.0601
Risk-Adjusted Not Risk-Adjusted
www.coe.neu.edu/Research/QPL
Variation in p’s (p, n )
Low p High p
Lownbar
Highnbar
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1
X0
Avg n = 10, SD n = 11.3, Avg p = .5, SD p = .028, J = 2SR = 0.9997, T AD = 0.0282, KL = 3.80E08
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
100000000
1000000000
10000000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1
X1
X0
X0
Avg n = 10, SD n = 11.3, Avg p = .5, SD p = .566, J = 2SR = 0.7809, T AD = 2.915E-4, KL = 0.0281
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1X0
Avg n = 200, SD n = 141.4, Avg p = .5, SD p = .028, J = 2SR = 0.9994, T AD = 5.816E-4, KL = 1.57E-7
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1X0
X1
X0
Avg n = 200, SD n = 141.4, Avg p = .5, SD p = .566, J = 2SR = 0.6547, T AD = 0.4054, KL = 0.0601
Risk-Adjusted Not Risk-Adjusted
www.coe.neu.edu/Research/QPL
Subsample Size (n, n)
Low n High n
Lownbar
Highnbar
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1
X0
X1X0
Avg n = 10, SD n = 1.41, Avg p = .5, SD p = .566, J = 2SR = 0.6019, T AD = 0.0263, KL = 8.58E-2
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
100000000
1000000000
10000000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1X1
X0
X0
Avg n = 10, SD n = 11.3, Avg p = .5, SD p = .566, J = 2SR = 0.7809, T AD = 2.915E-4, KL = 0.0281
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1
X0
X1
X0
Avg n = 200, SD n = 14.1, Avg p = .5, SD p = .566, J = 2SR = 0.6005, T AD = 0.4850, KL = 0.0828
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
B X1B X0
JB X1
JB X0
Avg n = 200, SD n = 141.4, Avg p = .5, SD p = .566, J = 2SR = 0.6547, T AD = 0.4054, KL = 0.0601
Risk-Adjusted Not Risk-Adjusted
www.coe.neu.edu/Research/QPL
Low p High p
LowJ
HighJ
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1X0
Avg n = 200, SD n = 141.4, Avg p = .5, SD p = .028, J = 2SR = 0.9994, T AD = 5.816E-4, KL = 1.57E-7
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1X0
X1
X0
Avg n = 200, SD n = 141.4, Avg p = .5, SD p = .566, J = 2SR = 0.6547, T AD = 0.4054, KL = 0.0601
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X0
X1
Avg n = 200, SD n = 141, Avg p = .5005, SD p = .006, J = 20SR = 0.9999, T AD = 0.0252, KL = 2.17E-4
Risk-Adjusted Not Risk-Adjusted
1
10
100
1000
10000
100000
1000000
10000000
-0.9
p
-0.8
p
-0.7
p
-0.6
p
-0.5
p
-0.4
p
-0.3
p
-0.2
p
-0.1
p p
0.1p
0.2p
0.3p
0.4p
0.5p
0.6p
0.7p
0.8p
0.9p
Percent Change in p(i)
Ave
rage
Run
Len
gth
X1
X0
X0
X1
Avg n = 200, SD n = 141, Avg p = .494, SD p = .351, J = 20SR = 0.8102, T AD = 0.2045, KL = 1.70E-2
Risk-Adjusted Not Risk-Adjusted
Number of Strata (J)