session 2: hazard functions, competing risks, cause...
TRANSCRIPT
2-1
Module20:SurvivalAnalysisforObserva9onalData
SummerIns9tuteinSta9s9csforClinicalResearchUniversityofWashington
July,2017
BarbaraMcKnight,Ph.D.
SESSION2:HAZARDFUNCTIONS,COMPETINGRISKS,CAUSE-SPECIFIC
HAZARDS,ANDANDCUMULATIVEINCIDENCE
2-2
OUTLINE
• Nonparametrices9ma9onofhazardfunc9ons• Compe9ngrisks:– Defini9on:whenthereismorethanonecauseofdeath/failure
– Cumula9vefunc9ons:• Event-freesurvival• Cumula9veIncidencees9mator
– Cause-specificandsub-distribu9onhazards– Fine-GrayandCoxregressionmodels– Interpreta9onsubtle9es
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-3
OUTLINE
• Nonparametrices.ma.onofhazardfunc.ons• Compe9ngrisks:– Defini9on:whenthereismorethanonecauseofdeath/failure
– Cumula9vefunc9ons:• Event-freesurvival• Cumula9veIncidencees9mator
– Cause-specificandsub-distribu9onhazards– Fine-GrayandCoxregressionmodels– Interpreta9onsubtle9es
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-4
HAZARDFUNCTION
• Instantaneousrateatwhichdeathoccursattinthosewhoarealiveatt
• Examples:
– Age-specificdeathrate– Age-specificdiseaseincidencerate
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-5
CUMULATIVEHAZARDFUNCTION
⇧(t) =Z t
0⌃(s)ds
= area under the hazard function curve
between 0 and t.
= amount of “hazard" accumulated between 0 and t.
= � log(S(t))
Not usually of interest per se, but estimates useful for diagnostics.
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-6
EQUIVALENTCHARACTERIZATIONS
• Anyoneofthesefourfunc9onsisenoughtodeterminethesurvivaldistribu9on.
• Theyareeachfunc9onsofeachother:
• S(t) =R�t ƒ (s)ds = e�
R t0 �(s)ds
• ƒ (t) = � ddt S(t) = �(t)e�
R t0 �(s)ds
• �(t) = ƒ (t)S(t)
• ⇧(t) =R t0 �(s)ds = � log(S(t))
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-7
EQUIVALENTCHARACTERIZATIONS
0 2 4 6 8
0.0
0.4
0.8
Survival Function
t
S(t)
0 2 4 6 8
0.00
0.10
0.20
0.30
Density Function
t
f(t)
0 2 4 6 8
0.0
0.5
1.0
1.5
Hazard Function
t
λ(t)
0 2 4 6 80
510
15
Cumulative Hazard Function
t
Λ(t)
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-8
EQUIVALENTCHARACTERIZATIONS
0 2 4 6 8
0.0
0.4
0.8
Survival Function
t
S(t)
0 2 4 6 8
0.00
0.10
0.20
0.30
Density Function
t
f(t)
0 2 4 6 8
0.0
0.5
1.0
1.5
Hazard Function
t
λ(t)
0 2 4 6 8
0.0
1.0
2.0
Cumulative Hazard Function
t
Λ(t)
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-9
CUMULATIVEHAZARD
• Nelson - Aalen estimator:
⌥(t) =P
j:t(j)tD(j)N(j)
• Variance:
’V�r(⌥(t)) =P
j:t(j)tD(j)S(j)[N(j)]3
• Standard error:∆’V�r(⌥(t)) can be used to form pointwise CI’s.
• or could use ⌥(t) = � log(S(t))
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-10
CUMULATIVEHAZARDFUNCTION
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-11
CUMULATIVEHAZARDFUNCTION
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d
0.0
0.5
1.0
0.17
0.25
0.33
0.5
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-12
SURVIVALANDIG
• RandomsubsetofthedatafromA.Dispenzieri,J.Katzmann,R.Kyle,D.Larson,T.Therneau,C.Colby,R.Clark,G.Mead,S.Kumar,L.J.MeltonIII,andS.V.Rajkumar.Useofmonclonalserumimmunoglobulinfreelightchainstopredictoverallsurvivalinthegeneralpopula9on.MayoClinicProc,87:512–523,2012.
• Arehighfree-chainIglevelsassociatedwithsurvival?– Popula9on-basedOlmsteadCountyexample– Menandwomen50+yearsofage
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-13
TOPDECILEFLC
0 1000 2000 3000 4000 5000
0.0
0.2
0.4
0.6
0.8
1.0
Days
Prob
abilit
y Su
rviva
l
Bottom 9Top Decile
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-14
FLCEXAMPLE
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 1000 2000 3000 4000 5000
0.0
0.2
0.4
0.6
0.8
1.0
1.2
Days
Cum
ulat
ive H
azar
d
Bottom 9Top Decile
2-15
FLCEXAMPLE
SISCR2017Module20SurvivalObserva9onalB.McKnight
10 20 50 100 200 500 1000 2000 5000
−8−6
−4−2
0
Days
Log
Cum
ulat
ive H
azar
dBottom 9Top Decile
2-16
HAZARDFUNCTION
IDEA:
At each time t, let estimate of �(t) be a weighted average of jumpsin ⇥(t) at nearby times.
Steps:
1. Choose a "bandwidth" ±b outside of which observations arenot averaged.
2. Choose a "kernel" or weight function K(·).
x
K(x)
−1 0 1
00.5
1
x
K(x)
−1 0 1
00.5
1
x
K(x)
−1 0 1
00.5
1
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-17
HAZARDFUNCTION
3. Calculate the estimate:
• �(t) = 1b
PJj=1 K(
t�t(j)b )D(j)N(j)
• se(�(t)) = 1b
⇢PJj=1 K
2( t�t(j)b )D(j)N2(j)
� 12
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-18
CUMULATIVEHAZARDFUNCTION
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-19
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-20
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-21
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-22
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-23
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-24
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-25
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-26
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-27
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-28
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-29
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-30
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-31
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-32
KERNELHAZARDESTIMATE
0 1 2 3 4 5 6
time
Cum
ulat
ive H
azar
d In
crem
ents
0.0
0.5
1.0
● ●● ●● ●
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-33
FLCEXAMPLE
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 1000 2000 3000 4000 5000
0e+0
02e−0
44e−0
46e−0
4
Days since Enrollment
Smoo
thed
Haz
ard
Top DecileOther 9
2-34
OUTLINE
• Nonparametrices9ma9onofhazardfunc9ons• Compe.ngrisks:– Defini.on:whenthereismorethanonecauseofdeath/failure
– Cumula.vefunc.ons:• Event-freesurvival• Cumula.veIncidencees.mator
– Cause-specificandsub-distribu9onhazards– Fine-GrayandCoxregressionmodels– Interpreta9onsubtle9es
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-35
COMPETINGRISKS
• Whenthereismorethanonecauseoffailure:– Recurrenceordeathbeforerecurrence– MI,stroke,PEordeathfromothercauses
• Thedifferenttypesoffailurearecalled“compe9ngrisks”.– They“compete”tobethefirsttomakesubjectsexperienceanevent
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-36
MONOCLONALGAMMOPATHY
• 241MayoClinicPa9ents(MonoclonalGammopathyofUndeterminedSignificance)
• 20-40yearsoffollow-upanerDx• 64developedplasmacellmalignancy(PCM),163diedwithoutit.
• PCManddeathwithoutPCMarecompe9ngrisksRKyle,Benignmonoclonalgammopathy–aner20to35yearsoffollow-up,MayoClinicProc1993;68:26-36
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-37
MONOCLONALGAMMOPATHY
• Insitua9onslikethis,ithasbeencommonprac9cetoapplytheKMmethodtoes9mate“survival”func9ons:– ProbabilityofavoidingPCMover9me– Probabiltyofavoidingdeathw/oPCM
• ForPCMcurve,treatdeathsw/oPCMascensored• Fordeathw/oPCM,treatPCMsascensored• Conceptualdifficul9esindescribingdistribu9onofthe9meofaneventthatmaynotoccur.
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-38
MONOCLONALGAMMOPATHY
0 10 20 30 40
0.0
0.2
0.4
0.6
0.8
1.0
Probability no PCM
years
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-39
MONOCLONALGAMMOPATHY
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 10 20 30 40
0.0
0.2
0.4
0.6
0.8
1.0
Probability no Death without PCM
2-40
MONOCLONALGAMMOPATHY
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 10 20 30 40
0.0
0.2
0.4
0.6
0.8
1.0
years
PCMDeath without PCM
Probability of avoiding PCM or death w/o PCM
Whatiswrongwiththispicture?
2-41
CUMULATIVEINCIDENCE
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 10 20 30 40
0.0
0.2
0.4
0.6
0.8
1.0
years
Death without PCMPCM
2-42
CUMULATIVEINCIDENCE
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 2000 4000 6000 8000 10000 12000 14000
0.0
0.2
0.4
0.6
0.8
1.0
Days Since Diagnosis
Cum
ulat
ive In
cide
nce
DeathPCM
Cumulative IncidenceKM
2-43
ESTIMATINGCUMULATIVEINCIDENCE
SISCR2017Module20SurvivalObserva9onalB.McKnight
• We can write
1� S(k)(t) =X
j:t(j)t
D(k)(j)N(j)
S(k)(t(j�1))
• At the second failure time of type k,
1� S(k)(t(2)) = 1� N(1)�D(k)(1)N(1)
· N(2)�D(k)(2)
N(2)=
D(k)(1)N(1)+
D(k)(2)N(2)· N(1)�D
(k)(1)
N(1)
• If any failures of another type have occurred between t(1) and
t(2), theN(1)�D(k)(1)
N(1)term is too big.
• This bias will accumulate and get larger, as we move to largerand larger t(j).
2-44
ESTIMATINGCUMULATIVEINCIDENCE
SISCR2017Module20SurvivalObserva9onalB.McKnight
• Letting D(k)(j) = the number of failures of types other than k att(j), an unbiased estimate of F(k)(t) is given by
X
j:t(j)t
D(k)(j)N(j)
j�1Y
�=1
N(�) �D(k)(�) �D(k)(�)
N(�)=X
j:t(j)t
D(k)(j)N(j)
j�1Y
�=1
N(�) �D(k)(�)N(�)
·N(�) �D(k)(�)
N(�)
"no ties between failures of different types
• Compare to biased upward
1� S(k)(t) =X
j:t(j)t
D(k)(j)N(j)
S(k)(t(j�1)) =X
j:t(j)t=D(k)(j)N(j)
j�1Y
�=1
N(�) �D(k)(�)N(�)
.
2-45
PREFERRED:TOGETHER
0 2000 4000 6000 8000 10000 12000 14000
0.0
0.2
0.4
0.6
0.8
1.0
Days Since Diagnosis
Cum
ulat
ive P
roba
bilit
y Dead
Plasma Cell Malignancy
Alive and Malignancy−Free
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-46
CHOICEOFOUTCOMEEVENT
• Maynotwantcause-specificeventasprimaryoutcome.
• Interpreta9oncloudy,par9cularlyforsurvivalcurves,sincethosewhodieearlymaycountas“survivors”ofacompe9ngevent.
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-47
CUMULATIVEFUNCTIONS
SISCR2017Module20SurvivalObserva9onalB.McKnight
• When there are competing risks, functions that describe thedistribution of the event-specific time Tk do not make sense:
– If the subject fails of another cause before t, Tk is not de-fined.
• Some other cumulative functions do make sense, depending oncontext:
– The probability that a subject is alive and event-of-interest-free at t.
∗ This means re-defining the event of interest to be theoriginal event of interest or death.
– The probability that an event of type k has (or has not)occurred by time t.
∗ It has not occurred if the subject dies before t.
2-48
CUMULATIVEFUNCTIONS
SISCR2017Module20SurvivalObserva9onalB.McKnight
Event-free Survival:
Estimating the probability a subject is alive and event-of-interest-free at time t is easy:
1. Redefine the event of interest to be either the original event ofinterest or death
�� =⇢1 event of interest or death from any cause0 censored
T� = time to event of interest, death or censoring
2. Compute the KM estimate of S(t) in the usual way with (T�, ��)data.
2-49
CUMULATIVEINCIDENCE
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 2000 4000 6000 8000 10000 12000 14000
0.0
0.2
0.4
0.6
0.8
1.0
Days Since Diagnosis
Cum
ulat
ive In
cide
nce
DeathPCMPCM or DeathCumulative IncidenceKM
2-50
MALIGNANCY-FREESURVIVAL
0 2000 4000 6000 8000 10000 12000 14000
0.0
0.2
0.4
0.6
0.8
1.0
Days Since Diagnosis
Mal
igna
ncy−
Free
Sur
viva
l
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-51
EXAMPLE
SISCR2017Module20SurvivalObserva9onalB.McKnight
MillerK,WangM,GralowJ,DicklerM,CobleighM,PerezEA,ShenkierT,CellaD,DavidsonNE.Paclitaxelplusbevacizumabversuspaclitaxelaloneformetasta9cbreastcancer.NewEnglandJournalofMedicine.2007;357(26):2666–2676.
2-52
SOMESUBTLETIES
• Cumula9veincidence:theprobabilitythataneventoftypekhasoccurredby9met:– Makessensewithoutrequiringthata9metothekthtypeofeventbedefinedforallsubjects
– Dependsonthepor9onofthepopula9ons9llatriskateach9me,soitsvaluewilldependnotonlyontheriskoftheeventofinterest,butalsoontheriskofalltheothercausesoffailure.
– Isapopula9on-specificquan9tythatdependsonwhatotherrisksareopera9nginthepopula9onandhowtheyarerelatedtotheriskoftheeventofinterest.
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-53
OUTLINE
• Nonparametrices9ma9onofhazardfunc9ons• Compe9ngrisks:– Defini9on:whenthereismorethanonecauseofdeath/failure
– Cumula9vefunc9ons:• Event-freesurvival• Cumula9veIncidencees9mator
– Cause-specificandsub-distribu.onhazards– Fine-GrayandCoxregressionmodels– Interpreta9onsubtle9es
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-54
REGRESSIONMODELS
• Therearetwotypesofregressionmodelsforcompe9ngrisksoutcomedatathatcanhaveusefulinterpreta9ons– Fine-Graymodels,thatmodeltheassocia9onbetweenindependentvariablesandthehazardfunc9onassociatedwiththecumula9veincidencesub-distribu9onfunc9on
– Coxregressionmodel,thatmodelstheassocia9onbetweenindependentvariablesandthecause-specifichazardfunc9on
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-55
FINE-GRAYHAZARD
SISCR2017Module20SurvivalObserva9onalB.McKnight
�FG(k)(t) = Pr[T 2 [t, t + �t), c = k|T � t or both T < t and c 6= k]/�t
Interpretations:
• the risk of failure of type k among those event free at t andthose who have experienced all events other than a type kevent. (Note if type k is not death, this would include subjectswho had already died.)
• the hazard function associated with the sub-distribution givenby the cumulative incidence function for a type k failure
2-56
FINE-GRAYMODEL
SISCR2017Module20SurvivalObserva9onalB.McKnight
Fine-Gray Hazard:
�FG(k)(t) = Pr[T 2 [t, t + �t), c = k|T � t or both T < t and c 6= k]/�t
Model:
�FG(k)(t|x) = �FG(k)(t|0)e�x
2-57
INTERPRETATION
SISCR2017Module20SurvivalObserva9onalB.McKnight
Before, we noted that the interpretation of the Fine-Gray hazard isthe risk of failure of type k among those event free at t and thosewho have experienced all events other than a type k event.
Suppose a type k failure is not death, and high values of � are asso-ciated with higher risk of death. Then if � has no causal associationwith type k failure, it will still be negatively associated with the Fine-Gray hazard of type k failure, since subjects with high values of �will be less likely to live long enough to experience a type k failure.
• Cumulative incidence of type k events will be lower for highvalues of �.
• Appropriate model when concern is about population burden,cost of type k events or estimating probabilities of the differentpossible outcomes for a patient.
• Not appropriate for causal modeling.
2-58
COXCAUSE-SPECIFICHAZARD
SISCR2017Module20SurvivalObserva9onalB.McKnight
�(k)(t) = Pr[T 2 [t, t + �t), c = k|T � t]/�tInterpretation:
• Risk of type k event among subjects who have not experiencedany of the types of events.
• Not influenced by the number of event-free subjects BUT can beinfluenced by who the event-free subjects are (Tsiatis, 1975).
• No real interpretation in terms of cumulative functions. Theanalogue would be the KM-survival curve, which is not appro-priate.
• Appropriate for causal modeling with interpretation restrictedto the population from which data are drawn (Prentice et al.,1978)
2-59
PROPERTIES
SISCR2017Module20SurvivalObserva9onalB.McKnight
T = time to first "failure" of any type
�(k)(t) = lim�t!0
Pr[T 2 [t, t + �t), c = k|T � t]/�t
• The different events defined by c must be mutually exclusive
• The different events defined by c must be exhaustive
• The hazard function for the distribution of T is given by :
�(t) =KX
k=1�(k)(t)
2-60
ESTIMATION
SISCR2017Module20SurvivalObserva9onalB.McKnight
Cause-specific Hazard Functions:
• For estimation, we can treat failures from other causes as cen-sored, and estimate the cause-specific hazard �(k)(t) and cu-mulative cause-specific hazard �(k)(t) in the usual way.
Q: Why does this work?
A:
Q: How are failures from other causes conceptually different fromthe censoring we have talked about earlier in this course?
A:
2-61
ESTIMATION
SISCR2017Module20SurvivalObserva9onalB.McKnight
Cumulative Incidence:
• We can treat failures from other causes as censored, and esti-mate the cause-specific hazard �(k)(t) and cumulative cause-specific hazard �(k)(t) in the usual way.
• If we do the same thing for the cumulative incidence, the Kaplan-Meier 1� S(k)(t) is a biased estimate of F(k)(t).
Q: Why?
A:
2-62
MGUSDATA
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 2000 4000 6000 8000 10000 12000
0.00
000
0.00
005
0.00
010
0.00
015
0.00
020
Days Since Diagnosis
Haz
ard
Rat
e
Death FirstPCM
2-63
OUTLINE
• Nonparametrices9ma9onofhazardfunc9ons• Compe9ngrisks:– Defini9on:whenthereismorethanonecauseofdeath/failure
– Cumula9vefunc9ons:• Event-freesurvival• Cumula9veIncidencees9mator
– Cause-specificandsub-distribu9onhazards– Fine-GrayandCoxregressionmodels– Interpreta.onsubtle.es
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-64
IDENTIFIABILITYANDINTERPRETATION
• Tsia9s(1975)showedthatwecannotiden9fyfrom(T,c=k)datawhethersubjectswhofailfromonecausewouldhavebeenmoreorlesssuscep9blelatertofailurefromanothercause,hadtheysurvived.– Cannottellwhetherthosewhodiefromheartdiseasewouldhavebeenmoreorlesslikelytodevelopcancerlater.
• Pren9ceetal(1978)arguedthatthecause-specifichazardfunc9on(Coxmodel)wasthebestbasisforcausalinferenceinthepopula9onasitiscons9tuted,butcannotextendinterpreta9ontoanotherpopula9onwherecompe9ngrisksarenotopera9ng.– Cannotsayhowxmightberelatedtocancerriskinapopula9onwheretherearenodeathsfromMI
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-65
INTERPRETATIONSUBTLETY
SISCR2017Module20SurvivalObserva9onalB.McKnight
Cannot estimate what the hazard �(k) (or survival function S(k)(t))would be if competing causes of failure were removed.
Reason: �(k)(t) = risk of type k failure at t among thosestill at risk at t.
6= risk of type k failure at t among thosestill at risk at t if other causes were removed.
For these to be equal, population at risk at T would need to havethe same risk of event k whether or not other causes of failure wereremoved. This is a strong and unverifiable assumption (Tsiatis,1975).
2-66
INTERPRETATIONSUBTLETY
• Cause-specifichazardisnonetheless,lessdependentontheriskofothertypesoffailureinthefollowingsense– Ifsomeinterven9onincreasestheriskofanothertypeoffailure,butdoesnotchangehowsuscep9blesurvivorsofthatfailuretypearetothetypeofinterest,theinterven9onwillnotinfluencethecause-specifichazardfunc9on
– Ifsomeinterven9onincreasestheriskofanothertypeoffailure,therewillbefewersubjectsatriskinthepopula9onatlater9mes,andthecumula9veincidenceoftheeventofinterestwillbelower.
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-67
COXMODEL
Cause-specific Hazard:
�(k)(t) = Pr[T 2 [t, t + �t), c = k|T � t]/�t
Cox Model:
�(k)(t|x) = �(k)(t|0)e�x
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-68
CHOOSINGACOMPETINGRISKSMODEL
• Coxmodelwithcause-specifichazardgoodwheninterestedinwhatcausestheeventofinterestinthepopula9onasitiscons9tuted.
• Fine-Graygoodwheninterestedinpredic9ngpa9entprognosisorpopula9ondisease/costburden.
• Botharepropor9onalhazardsmodels,butfordifferenthazardfunc9ons
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-69
EXAMPLE
• CoxandFine-Graymodelsfortheassocia9onofsexwithPCMandDeathbeforePCMintheMonoclonalGammopathydata.
• Willshow– Cause-specifichazardfunc9onsbysexandcause– Cumula9veincidencefunc9onsbysexandcause– Es9matedHazardra9os(maletofemale)bycauseunderbothmodels
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-70
CAUSE-SPECIFICHAZARDESTIMATES
0 5 10 15 20 25 30
0.00
0.02
0.04
0.06
0.08
0.10
Years
Haz
ard
Rat
e
malesfemales
Plasma Cell Malignancy
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-71
CAUSE-SPECIFICHAZARDESTIMATES
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 5 10 15 20 25 30
0.00
0.02
0.04
0.06
0.08
0.10
Years
Haz
ard
Rat
emalesfemales
Deaths from Other Causes
2-72
COXMODELS
OutcomeType M/FHazardRa9o 95%CI P-valuePlasmaCellMalignancy 0.95 (0.58,1.56) 0.8441DeathfromOtherCauses 1.55 (1.13,2.14) 0.0064
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-73
CUMULATIVEINCIDENCE
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 10 20 30 40
0.0
0.2
0.4
0.6
Female Death no PCMMale Death no PCMFemale PCMMale PCM
2-74
CUMULATIVEINCIDENCE
SISCR2017Module20SurvivalObserva9onalB.McKnight
0 10 20 30 40
0.0
0.2
0.4
0.6
0.8
1.0
Years
Prob
abilit
y
MalesFemales
Alive PCM−free
PCM
Death no PCM
2-75
FINE-GRAYMODELS
OutcomeType M/FHazardRa9o 95%CI P-valuePlasmaCellMalignancy 0.71 (0.44,1.16) 0.17
DeathfromOtherCauses 1.45 (1.06,1.97) 0.02
SISCR2017Module20SurvivalObserva9onalB.McKnight
PCMhazardra9ofartherfromoneherebecausemenaremorelikelytodiefromothercausesandnotsurvivetodevelopPCM.
2-76
COMMONMISINTERPRETATION
• TheFine-Graymodelisnotthemodelthat“accountsforcompe9ngrisks”
• BoththeCoxmodelwithcause-specifichazardfunc9onsandtheFine-Graymodelaccountforcompe9ngrisks,butaswehaveseentheirtargetsofinferencearedifferent.– Hazardamongthoses9llatriskoftheevent(Cox)– Hazardamongthosewhohavenotyetexperiencedtheeventofinterest,butcouldhaveexperiencedothers,includingdeath,already(F-G)
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-77
EXAMPLE
• Ashburneretal(2017)studiedacohortof13,559subjectsdiagnosedwithatrialfibrilla9on(AF)atKaiserNorthernCalifornia– 1092thromboembolismevents(1017ischemicstrokes)– 4414experienceddeathwithoutthromboembolismevent– Thromboemolism-freeDeathratewas5.5/100PYamongwarfarintakersand8.1/100PYamongnon-takers
– Non-takerswereolderhadhigherstroke-riskscores• TheycomparedCoxandF-Gregressionwith9me-dependent
currentwarfarinuseastheexposure
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-78
EXAMPLE
• TheyconcludedthattheFine-Graymodelthat“accountedfor”compe9ngrisksgaveabeter“real-world”assessmentofthebenefitofwarfarin.
• Whatareyourthoughts?
AshburnerJM,GoAS,ChangY,FangMC,FredmanL,ApplebaumKM,SingerDE.JAmGeriatrSoc.2017Jan1;65(1):35–41.
SISCR2017Module20SurvivalObserva9onalB.McKnight
Event
Model
AdjustedHazardRa9o
95%CI
Thromboembolism Cox 0.57 (0.50,0.65)
Fine-Gray 0.87 (0.77,0.99)
2-79
SOMECOMPETINGRISKSREFERENCES• Aus9nPC,LeeDS,FineJP.Introduc9ontotheAnalysisofSurvivalDatainthe
PresenceofCompe9ngRisks.Circula9on.2016Feb9;133(6):601–609.• FineJP,GrayRJ.APropor9onalHazardsModelfortheSubdistribu9onofa
Compe9ngRisk.JournaloftheAmericanSta9s9calAssocia9on.1999Jun1;94(446):496–509.
• LauB,ColeSR,GangeSJ.Compe9ngriskregressionmodelsforepidemiologicdata.AmericanjournalofEpidemiology.2009;170(2):244–256.
• NoordzijM,LeffondréK,StralenV,JK,ZoccaliC,DekkerFW,JagerKJ.Whendoweneedcompe9ngrisksmethodsforsurvivalanalysisinnephrology?NephrolDialTransplant.2013Nov1;28(11):2670–2677.
• Pren9ceRL,KalbfleischJD,PetersonJrAV,FlournoyN,FarewellVT,BreslowNE.Theanalysisoffailure9mesinthepresenceofcompe9ngrisks.Biometrics.1978;541–554.
• Tsia9sA.Anoniden9fiabilityaspectoftheproblemofcompe9ngrisks.ProceedingsoftheNa9onalAcademyofSciences.1975;72(1):20–22.
SISCR2017Module20SurvivalObserva9onalB.McKnight
2-80
TOWATCHOUTFOR
• Interpreta9oninthepresenceofcompe9ngriskscanbesubtleandrequirescare.
– S(t)definedintermsoftheprobabilitydistribu9onofTkdoesnotmakesense
– Cannotinterpretfunc9onsofthecause-specifichazardasapplyinginapopula9onwithoutcompe9ngriskspresent.
– 1–KMes9matorcangiveupwardbiasedes9mateofcumula9veincidence.
– Cannotinterpretcumula9veincidenceasapplyinginapopula9onwithoutcompe9ngriskspresent
– Fine-GraymodelisnotTHEwaytoaccountforcompe9ngrisks.
SISCR2017Module20SurvivalObserva9onalB.McKnight