2/20/2012 1 Brook I. Martin PhD MPH Dartmouth College Introduction to Survival Analysis About me • Health services PhD from University of h Washington. • Health services faculty at Dartmouth College. • Affiliated with Dartmouth‐Hitchcock Medical Center Department of Orthopaedics. • Primary research interest is in the quality of care for musculoskeletal & spinal problems. 2 Overall goals Review framework for choosing an analysis 1) Introduction to survival analysis. 2) Descriptive analysis of survival data. 3) Applied introduction to Cox‐Proportional Hazard regression models. 1) Modeling an outcome 2) Diagnostic tests of PH assumptions 3
19
Embed
10CER Module10 SurvivalAnalysis Martin€¦ · Hazard regression models. 1) Modeling an outcome 2) ... B. Goals of survival analysis C. Survival time and censored data D. Calculating
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
2/20/2012
1
Brook I. Martin PhD MPHDartmouth Collegeg
Introduction to Survival Analysis
About me
• Health services PhD from University of hWashington.
• Health services faculty at Dartmouth College.
• Affiliated with Dartmouth‐Hitchcock Medical Center Department of Orthopaedics.
• Primary research interest is in the quality of care for musculoskeletal & spinal problems.
2
Overall goals
Review framework for choosing an analysis
1) Introduction to survival analysis.
2) Descriptive analysis of survival data.
3) Applied introduction to Cox‐Proportional Hazard regression models.
1) Modeling an outcome
2) Diagnostic tests of PH assumptions
3
2/20/2012
2
Data types
Type Example
Continuous Age; annual salary; WBC
Dichotomous Infection (yes/no); Death (yes/no)
Count Number (rates) of surgical procedures.
Survival # days until death or other event.
Categorical (Ordinal)
Health rating:1 = Excellent; 2 = Very good; 3 = Good; 4 = Fair; 5 = Poor
Count Incidence & Rate‐difference or ratio Poisson regressionprevalence
g
Survival Kaplan‐Meier survival Log‐rank; Wilcoxon(other) for survival data
Cox‐proportionalhazard regression
Categorical (Ordinal)
Proportion;Wilcoxon signed rank
Spearman’s test;Mann‐Whitney test;
Ordered logistic regression
Categorical (Nominal)
Proportion Chi‐square; Mantel‐Haenszel test
Multinomial logistics
7
Examples
Martin BI, Mirza SK, Comstock BA, Gray DT, Kreuter W, Deyo RA. Reoperation rates following lumbar spine surgery and the influence of spinal fusion procedures. Spine (Phila Pa 1976). 2007 Feb 1;32(3):382‐7.
Bederman SS, Kreder HJ, Weller I, Finkelstein JA, Ford MH, Yee AJ. The who, what and when of surgery for the degenerative lumbar spine: a population‐based study of surgeon factors, surgical procedures, recent trends and reoperation rates. Can J Surg. 2009 Aug;52(4):283‐290.
8
Part I: Survival Analysis
1) Introduction to survival analysis
A. What is survival analysis
B. Goals of survival analysis
C Survival time and censored dataC. Survival time and censored data
D. Calculating survival
E. Why use survival analysis
9
2/20/2012
4
What is survival analysis?
• Statistical procedures used when the outcome of interest is time until an event occurs.
• Example: Time until death, Time until repeat surgerysurgery.
• Definitions:
Time = survival time (days/years)
Event = indication of failure
Censoring = when survival data is not fully known.
10
Goals of survival analysis?
• Estimate and interpret survivor and/or hazard functions.
• To compare survivor and/or hazard functions between groups of interestbetween groups of interest.
• To assess the relationship of explanatory variables to survival time ( Cox‐Proportional Hazard regression)
11
Survival data:
A
B
C
subjects
Time Event Censor
No 12 0 1
12
Time (weeks)
0 2 4 6 8 10 12
D
EStudy s
2/20/2012
5
Survival data:
A
B
C
subjects
Time Event Censor
YES
No 12 0 1
8 1 0
13
Time (weeks)
0 2 4 6 8 10 12
D
EStudy s
Survival data:
A
B
C
subjects
Time Event Censor
YES
No
No
12 0 1
8 1 0
8 0 1
14
Time (weeks)
0 2 4 6 8 10 12
D
EStudy s
Survival data:
A
B
C
subjects
Time Event Censor
YES
No
No
12 0 1
8 1 0
8 0 1
15
Time (weeks)
0 2 4 6 8 10 12
D
EStudy s No 4 0 1
2/20/2012
6
Survival data:
A
B
C
subjects
Time Event Censor
YES
No
No
12 0 1
8 1 0
8 0 1
16
Time (weeks)
0 2 4 6 8 10 12
D
EStudy s No
Yes
4 0 1
6 1 0
Survival data:
Time Event Censor
A 12 0 1
B 8 1 0
C 8 0 1
ID δ = (0,1) Where 1 if failure0 if censored
T = random variable for survival time (>=0)t = specific value for T
17
D 4 0 1
E 6 1 0
Survival > 8 weeks?H0: T > t = 8
Survival data:
Time Event Censor
A 12 0 1
B 8 1 0
C 8 0 1
ID δ = (0,1) Where 1 if failure0 if censored
T = random variable for survival time (>=0)t = specific value for T
18
D 4 0 1
E 6 1 0
Survival > 8 weeks?H0: T > t = 8
n tn δn
2/20/2012
7
Survivor function:
Survivor function S(t): The probability thata subject survives longer than a specific t
• Always decrease as t increases
S(t)
S(t) = 11
19
Always decrease as t increases• At t= 0; S(t) = 1• At t = ∞; S(t) = 0• In practice represented by a step curve.
S(∞) = 00
Survivor function:
Survivor function S(t) – The probability thata subject survives longer than a specific t
• Always decrease as t increases
S(t)
S(t) = 11
20
Always decrease as t increases• At t= 0; S(t) = 1• At t = ∞; S(t) = 0• In practice represented by a step curve.
S(∞) = 00
S(t)
S(t) = 1
0
1
t
Hazard function:
The limit, as Δt approaches 0, of a probability statement about survival, divided by Δt
h(t) = lim P( t ≤ T < Δt | T ≥ t)/ Δt
h(t) ≥ 0; has no upper limit.h(t)
∞
21
View the hazard as an “instantaneous” rateof the event per a unit of time, given that the subject has survived to time t.
S(t) : not failingH(t) : failing
h(t)
0
t
2/20/2012
8
Calculating survival:
ID T δ cen
A 12 0 1
B 8 1 0
C 8 0 1
D 4 0 1
ID T δ cen
‐‐ 0 0 0
D 4 0 1
J 4 1 0
K 5 0 1
t(j) #fail
# cen # at risk
t(0) 0 0 11
Table of ordered failure times
22
D 4 0 1
E 6 1 0
F 10 1 0
G 10 0 1
H 9 0 1
I 8 1 0
J 4 1 0
K 5 0 1
E 6 1 0
B 8 1 0
C 8 0 1
I 8 1 0
H 9 0 1
F 10 1 0
G 10 0 1
A 12 0 1
t(4) 1 2 11
t(6) 1 0 8
t(8) 2 2 7
t(10) 1 2 3
Calculating survival:
ID T δ Cen
‐‐ 0 0 0
D 4 0 1
J 4 1 0
K 5 0 1
t(j) #fail
# cen # at risk
t(0) 0 0 11
t(4) 1 2 11
Table of ordered failure timesAverage survival time (ignoring censoring) = 84/ 11 = 7.6 weeks
Average Hazard time = # fail / ∑ ti = 5/ 84 = 0.06
Count Incidence & Rate‐difference or ratio Poisson regressionprevalence
g
Survival Kaplan‐Meier survival Log‐rank; Wilcoxon(other) for survival data
Cox‐proportionalhazard regression
Categorical (Ordinal)
Proportion;Wilcoxon signed rank
Spearman’s test;Mann‐Whitney test;
Ordered logistic regression
Categorical (Nominal)
Proportion Chi‐square; Mantel‐Haenszel test
Multinomial logistics
37
Cox Proportional Hazard Model:
General form
h(t, X) = h0(t) exp[β1X1 + … + βnXn]
38
Hazard Ratio:The measure of effect used in survival analysis that describes the association between an exposure and an outcome.
Calculated as the hazard for one set of characteristic (exposure) compared to an alternative set of characteristics (unexposed).
HR = ĥ(t, X*)/ ĥ(t, X) = exp[∑ βi(X*I – Xi)]
39
( , )/ ( , ) p[∑ βi( I i)]
Typically expressed as an exponential from a regression coefficient:HR = 1 no associationHR > 1 exposure associated with greater risk of outcomeHR <1 exposure associated with lower risk of outcome.
2/20/2012
14
Cox‐Proportional Hazard Regression
Stata command for Cox‐PH Regression:
stcox i.PAYGRP female i.AGE4 positive_quan
40
PH assumptionThe requirement for the HR to constant over time…
…or, that the hazard for one exposure is proportional (not dependent on time) to the hazard for an alternative exposure.
41
ĥ(t, X*) = constant * ĥ(t, X)
Can be evaluated •Graphically•Through GOF statistics
PH assumptionThe requirement for the HR to constant over time…
…or, that the hazard for one exposure is proportional (not dependent on time) to the hazard for an alternative exposure.
h(t, X)
∞
X = 1
X = 0
42
ĥ(t, X*) = constant * ĥ(t, X)
Can be evaluated •Graphically•Through GOF statistics
0
0 4 8 12weeks
2/20/2012
15
PH assumptionThe requirement for the HR to constant over time…
…or, that the hazard for one exposure is proportional (not dependent on time) to the hazard for an alternative exposure.
h(t, X)
∞
X = 1
X = 0
43
ĥ(t, X*) = constant * ĥ(t, X)
Can be evaluated •Graphically•Through GOF statistics
0
0 4 8 12weeks
HR 2 weeks: ĥ(t=2, X = 1)/ ĥ(t=2, X = 0) < 1
HR 10 weeks: ĥ(t=10, X = 1)/ ĥ(t=10, X = 0) > 1
Thus, HR is not constant over time, and the PH assumption is not met! & Cox‐PH regression is not appropriate.
Coefficients are for variables that meet PH assumption, adjusted for the variables that they are stratified on; not possible to obtain HR for the effect on PAYGRP
2/20/2012
18
Use time‐varying coefficientsOriginal model assumes no interaction with time:h(t, X) = h0(t) exp[β1X1 + … + βnXn]
Time‐varying model allows for hazard to vary with time:h*(t, X) = h0(t) exp[β1X1+ …+ βnXn + δjXj(t)]