Introduction to medical survival analysis - OtagoEstimated survivor function Months 75+

1

Introduction to medical survival analysis

John Pearson Biostatistics consultant University of Otago Canterbury 7 October 2008

2

Objectives • Describe survival data • Define survival analysis terms • Compare survival of groups • Describe study design

Acknowledgement: Thanks to Colm Fahy for providing the example data.

3

Omissions • Not covered:

– most methodology issues – mathematical justification

• See – Collett: Modelling Survival Data in Medical

Research – Hosmer & Lemeshow: Applied Survival

Analysis – Many other good texts.

4

Example: Metastatic Parotid SCC

• Disease risk factors: – >50 yo – Male – Exposure to sun – Caucasian ancestry

• 61 patients operated on since 1990 • Audit done 1/6/8 • 14 patients died from SCCMP, 20 died

from other causes, 1 couldn’t be found

5

Example: Patient data OpDate Died Status Preserved RadioTx ICOMP

7/05/2002 ALIVE PARTIAL YES N15/11/2007 ALIVE NO YES N12/10/2007 1/03/2008 DOC YES YES N17/04/1992 1/08/1993 DOD YES YES Y7/10/1996 1/04/1997 DOC NO YES N1/05/1991 LOST YES YES N

12/03/2003 1/05/2005 DOC YES YES Y

Only 7 patients shown. Dates have been confidentialized.

6

Example: Patient data

1

2

3

4

5

6

7

1990 1995 2000 2005 6/2008

Parotidectomy patient medical records

Pat

ient

AliveDead OCDead OD

? Lost to follow up audi

t

7

Example: Patient data

1

2

3

4

5

6

7

1990 1995 2000 2005 6/2008

Parotidectomy patient medical records

Pat

ient

AliveDead OCDead OD

? Lost to follow up audi

t

?

?

8

Example: Survival Data

1

2

3

4

5

6

7

0 5 10 15

Parotidectomy patient survival data

Pat

ient

?

AliveDead OCDead OD

Years post operation

9

Example: Survival Data

Date formats and manipulation can cause headaches. Check what happens when your software subtracts dates to get survival time.

1

2

3

4

5

6

7

0 5 10 15

Parotidectomy patient survival data

Pat

ient

?

AliveDead OCDead OD

Years post operation

10

1

2

3

4

5

6

7

0 5 10 15

Parotidectomy patient survival data

Pat

ient

?

AliveDead OCDead OD

Years post operation

Example: Survival Data

censored

censored

Missing data

11

1

2

3

4

5

6

7

0 5 10 15

Parotidectomy patient survival data

Pat

ient

?

AliveDead OCDead OD

Years post operation

Example: Survival Data

censored

censored

Missing data

censored

censored

censored

12

1

2

3

4

5

6

7

0 5 10 15

Parotidectomy patient survival data

Pat

ient

?

AliveDead OCDead OD

Years post operation

Example: Survival Data

censored

censored

censored

Missing data

censored

Censored data is explicitly addressed by survival analysis, using simple linear regression is not recommended. Options: 1. SPSS 2. SAS 3. R 4. Other software

13

1

2

3

4

5

6

7

0 5 10 15

Parotidectomy patient survival data

Pat

ient

?

AliveDead OCDead OD

Years post operation

Example: Survival Data

censored

censored

censored

Missing data

censored

Missing data can have a large effect on results, requires careful management. Options: 1. Omit 2. Impute 3. Model

14

What is survival analysis • Time to event data

– Continuous – Right skewed, ≥0, not normal – Censored – Analyse risk (hazard function)

• Examples – Time to death – Time to onset/relapse of disease – Length of stay in hospital

15

What is survival analysis • Time to event data

– Continuous – Right skewed, ≥0, not normal – Censored – Analyse risk (hazard function)

• Examples – Time to death – Time to onset/relapse of disease – Length of stay in hospital

0

5

10

15

0 2 4 6 8 10

Post operative survival

Pat

ient

s

Years

16

Censoring • Right censoring • Left censoring • Interval censoring

Censoring is also categorised by 1. Fixed study length 2. Fixed number of events 3. Random entry to study

17

Censoring • Right censoring

– observed survival time is less than actual – Study ends before event

• Left censoring • Interval censoring

1

2

3

4

5

6

7

1990 1995 2000 2005 6/2008

Parotidectomy patient medical records

Pat

ient

AliveDead OCDead OD

? Lost to follow up audi

t

?

?

18

Censoring • Right censoring • Left censoring

– Time to relapse

– Time to event is less than observed t < 3 • Interval censoring

Surgery

0

Recurrence

3 month exam t

19

Censoring • Right censoring • Left censoring • Interval censoring

– Time to relapse

– 3 < t < 6

Surgery

0

Free of disease

3 month exam t

Recurrence

6 month exam

20

Censoring

Independent censoring

Survival time is independent of censoring process. A censored patient is representative of those at risk at censoring time. The methods described here assume independent censoring

21

Censoring

Independent censoring

Survival time is independent of censoring process. Informative censoring

Patients removed from study if condition deteriorates.

22

Censoring example How are the SCCMP patients censored?

23

Censoring example How are the SCCMP patients censored? • Enter study on surgery date • Last known status is at audit Random right censoring.

24

Survival function The survival function S(t) is the probability of surviving longer than time t.

S(t) = P(T>t)

Where T is the survival time.

patients of number total than longer surviving patients of Number t

S(t)

25

Hazard function The hazard function λ(t) is the probability of dying “at” time t.

Also called the instantaneous failure rate and force of mortality.

S(t)

f(t)(t)

)(log tS(t)Usually plotted is the cumulative hazard function, that is the accumulated hazard until time t.

26

Survival function For censored data the survival function can only be estimated.

1

2

3

4

5

6

7

0.0 0.5 1.0 1.5 2.0 2.5 3.0

Parotidectomy patient survival data

Pat

ient

AliveDead OCDead OD

Years post operation

27

Survival function Life table estimates WHO, StatsNZ

All causes mortality

0

20

40

60

80

100

0 10 20 30 40 50 60 70 80 90 100

Age

Perc

en

t su

rviv

ing

NZ

Australia

Chad

28

Survival function Kaplan Meier estimates

Months n d (n-d)/n S(t)

1 2.2 57 1 0.982 0.9822 6.12 51 1 0.980 0.9633 10.32 46 1 0.978 0.9424 10.78 45 1 0.978 0.9215 10.88 44 1 0.977 0.96 13.08 41 1 0.976 0.8787 13.35 39 1 0.974 0.8568 16.11 37 1 0.973 0.8339 26.2 34 1 0.971 0.808

10 29.42 31 1 0.968 0.78211 37.48 26 1 0.962 0.75212 45.86 23 1 0.957 0.71913 59.08 19 1 0.947 0.68214 65.33 14 1 0.929 0.633

29

Survival function Kaplan Meier estimates

Months n d (n-d)/n S(t)

1 2.2 57 1 0.982 0.9822 6.12 51 1 0.980 0.9633 10.32 46 1 0.978 0.9424 10.78 45 1 0.978 0.9215 10.88 44 1 0.977 0.96 13.08 41 1 0.976 0.8787 13.35 39 1 0.974 0.8568 16.11 37 1 0.973 0.8339 26.2 34 1 0.971 0.808

10 29.42 31 1 0.968 0.78211 37.48 26 1 0.962 0.75212 45.86 23 1 0.957 0.71913 59.08 19 1 0.947 0.68214 65.33 14 1 0.929 0.633

1. Order data by time to event (death) 2. Number at risk of

event is number surviving less number censored.

3. Estimate of probability of surviving to next event

4. Multiply probabilities to estimate survival

30

Kaplan Meier plot

0.0

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100 120

Kaplan Meier estimate

Est

imat

ed s

urvi

vor f

unct

ion

Months

31

Kaplan Meier plot SCCMP

0.0

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100 120

Kaplan Meier estimate

Est

imat

ed s

urvi

vor f

unct

ion

Standard errors and 95% CI’s calculated by most software (SPSS, R, SAS)

Usually use Greenwood’s or Tsiatis’ formula, software dependent.

32

Cumulative Hazard SCCMP

0.0

0.1

0.2

0.3

0.4

0 20 40 60 80 100 120

Cumulative Hazard Function

Cum

ulat

ive

haza

rd

Months

33

Summary statistics 1. Median survival: time when S(t) = 0.5

• Must have enough data 2. Mean survival: area under the survival

curve 3. 5 year survival is survival rate at 5 years

34

Kaplan Meier estimate KM and lifetables are non-parametric methods: no assumptions are made about the distribution on the survival times. Typical distributions are exponential and Weibull. More powerful but can be sensitive to getting the distribution right.

35

Disease specific survival

0.0

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100 120

SCCMP survival

Est

imat

ed s

urvi

vor f

unct

ion

Months

Disease specificAll causes

36

Comparing 2 groups Log rank test • Computed in SPSS, SAS, R • Most popular

– (Bland Altman BMJ 2004;328:1073 (1 May) • Limitations

– No estimate of size – Unlikely to detect a difference when risk is not

consistent

37

Immuno compromised

0.0

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100 120 140

SCCMP survival: Immuno Compromised

Est

imat

ed s

urvi

vor f

unct

ion

Months

No

Yes

38

Immuno compromised

0.0

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100 120 140

SCCMP survival: Immuno Compromised

Est

imat

ed s

urvi

vor f

unct

ion

Months

No

Yes

Case Processing Summary

53 9 44 83.0%7 5 2 28.6%

60 14 46 76.7%

ICOMPNYOverall

Total N N of Events N PercentCensored

39

Immuno compromised

0.0

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100 120 140

SCCMP survival: Immuno Compromised

Est

imat

ed s

urvi

vor f

unct

ion

Months

No

Yes

Means and Medians for Survival Time

101.048 7.616 . .22.978 7.653 16.110 3.29391.761 7.842 . .

ICOMPNYOverall

Estimate Std. Error Estimate Std. Error

Meana Median

Estimation is limited to the largest survival time if itis censored.

a.

40

Immuno compromised

0.0

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100 120 140

SCCMP survival: Immuno Compromised

Est

imat

ed s

urvi

vor f

unct

ion

Months

No

Yes

Overall Comparisons

19.579 1 .000Log Rank (Mantel-Cox)Chi-Square df Sig.

Test of equality of survival distributions for the different levels ofICOMP.

41

Age group

0.0

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100 120 140

SCCMP survival: Age group

Est

imat

ed s

urvi

vor f

unct

ion

Months

75+<75

Call:

survdiff(formula = Surv(mths,Status == "DOD") ~ ICOMP)

N Observed Expected (O-E)^2/E (O-E)^2/V

Age75=<75 24 7 5.63 0.332 0.557

Age75=75+ 36 7 8.37 0.224 0.557

Chisq= 0.6 on 1 degrees of freedom, p= 0.455

42

Facial Nerve

0.0

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100 120 140

SCCMP survival: Facial Nerve Preserved

Est

imat

ed s

urvi

vor f

unct

ion

Months

NO

PARTIAL

YES

Log rank p value: 0.09

43

Multiple independent variables Cox proportional hazards model • Most common model • Linear model for the log of the hazard ratio

• Baseline hazard unspecified

2211

)(

)(

0

1 ZBZBe

th

th

44

SCCMP example CPH model: Survival ~ Preserved + Age + ICOMP Preserved and ICOMP categorical Age continuous Plot survival for patients with each of /Y/N/partial nerve preservation adjusted for age and immuno compromised status

45

SCCMP example - SPSS

Analyze > Survival > Cox Regression COXREG Months /STATUS=Status('DEAD') /PATTERN BY Preserved /CONTRAST (Preserved)=Indicator /CONTRAST (ICOMP)=Indicator(1) /METHOD=ENTER Preserved Age ICOMP /PLOT SURVIVAL /SAVE=PRESID XBETA /PRINT=CI(95) CORR SUMMARY BASELINE /CRITERIA=PIN(.05) POUT(.10) ITERATE(20) .

46

SCCMP example - SPSS

Patients with their facial nerve preserved have 12.6 times less hazard ratio, (95% CI 2-70) . Preserving the facial nerve significantly reduces patients risk, (p value <0.001 CPH model).

Variables in the Equation

8.493 2 .0142.535 .871 8.470 1 .004 12.617 2.288 69.5642.091 1.110 3.549 1 .060 8.093 .919 71.2793.588 .918 15.274 1 .000 36.166 5.981 218.676-.011 .028 .149 1 .700 .989 .936 1.046

Preserved No PartialICOMPAge

B SE Wald df Sig. Exp(B) Lower Upper95.0% CI for Exp(B)

47

SCCMP CPH model

0.0

0.2

0.4

0.6

0.8

1.0

0 10 20 30 40 50 60 70

SCCMP survival: Facial nerve preserved

Est

imat

ed s

urvi

vor f

unct

ion

Months

NO

PARTIAL

YES

Adjusted for age and immuno compromised patients

48

Next Steps:

• Check proportional hazards assumption – Residual plots for groups

• Time dependent covariates • More complex models

• we also didn’t do power calculations

49

Summary

• Survival analysis accounts for censoring in time to event data

• Log rank test: difference in survival between 2 groups

• Cox proportional hazard model • More complex/powerful models available • SPSS, R, SAS, Stata