1 / 33

INTRODUCTION TO CLINICAL RESEARCH Survival Analysis – Getting Started Karen Bandeen-Roche, Ph.D.

INTRODUCTION TO CLINICAL RESEARCH Survival Analysis – Getting Started Karen Bandeen-Roche, Ph.D. July 20, 2010. Acknowledgements. Scott Zeger Marie Diener-West ICTR Leadership / Team. Introduction to Survival Analysis. Thinking about times to events; contending with “censoring”

chads
Download Presentation

INTRODUCTION TO CLINICAL RESEARCH Survival Analysis – Getting Started Karen Bandeen-Roche, Ph.D.

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. INTRODUCTION TO CLINICAL RESEARCH Survival Analysis – Getting Started Karen Bandeen-Roche, Ph.D. July 20, 2010

  2. Acknowledgements • Scott Zeger • Marie Diener-West • ICTR Leadership / Team JHU Intro to Clinical Research

  3. Introduction to Survival Analysis • Thinking about times to events; contending with “censoring” • Counting process view of times to events • Hazard and survival functions • Kaplan-Meier estimate of the survival function • Future topics: log-rank test; Cox proportional hazards model

  4. “Survival Analysis” • Approach and methods for analyzing times to events • Events not necessarily deaths (“survival” is historical term) • Need special methods to deal with “censoring”

  5. Typical Clinical Study with Time to Event Outcome Loss to Follow-up Event Start End Enrollment End Study 0 2 4 6 8 10 Calendar time

  6. Switching from Calendar to Follow-up Time >3 5 >8 1 >6 Follow-up time 0 2 4 6 8 10

  7. The Problem with Standard Analyses of Times to Events • Mean: (1 + 3 + 5 + 6 + 8)/5 = 4.6 - right? • Median: 5 – right? • Histogram

  8. Censoring > 3 is not 3, it may be 33 Mean is not 4.6, it may be (1 + 33 + 5 + 6 + 8)/5 = 10.6 Or any value greater than 4.6 > 3 is a right “censored value” – we only know the value exceeds 3 > x is often written “x+”

  9. Censoring • Uncensored data: The event has occurred • Event occurrence is observed • Censored data: The event has yet to occur • Event-free at the current follow-up time • A competing event that is not an endpoint stops follow-up • Death (if not part of the endpoint) • Clinical event that requires treatment, etc. • Our ability to observe ends before event happens

  10. Contending with Censored Data Standard statistical methods do not work for censored data We need to think of times to events as a natural history in time, not just a single number • Issue: If no events are reported in the interval from • last follow-up to “now”, need to choose between: • No news is good news? • No news is no news?

  11. One Option: Overall Event Rate • Example: 2 events in 23 person months = 1 event per 11.5 months = 1.04 events per year = 104 events per 100 person-years • Gives an average event rate over the follow-up period; actual event rate may vary over time • For a finer time resolution, do the above for small intervals

  12. Switching from Calendar to Follow-up Time >3 5 >8 1 >6 Follow-up time 0 2 4 6 8 10 3+5+8+1+6 person months of observation; 2 actual events

  13. Second Option: Natural history “One day at a time” 0 0 0 >3 0 0 0 0 1 5 0 0 0 0 0 0 0 0 >8 1 1 0 0 0 0 0 0 >6 Follow-up time 0 2 4 6 8 10

  14. Thinking about Times to Events

  15. Survival Function “Survival function”, S(t), is defined to be the probability a person survives beyond time t S(0) = 1.0 S(t+1) S(t)

  16. Hazard Function • Hazard at time t, h(t), is the probability per unit time of having the event in a small interval around time t • Force of mortality • ~ Pr{event in (t,t+dt)}/dt • Need not be between 0 and 1 because it is per unit time • h(t) ~ {S(t)-S(t+dt)}/{S(t) dt}

  17. Hazard Function • Basic idea: Live your life one interval (day, month, or year) at a time • Example: S(3) = Pr(survive for 3 months) = Pr(survive 1st month) × Pr(survive 2nd month | survive 1st month) × Pr(survive 3rd month | survive 2nd month) • Thus, = Pr(survive for 1st month & 2nd & 3rd)

  18. Estimating the Survival Function: Kaplan-Meier Method Pr(survive past 5) = Pr(survive past 5|survive past 4) *Pr(survive past 4) [ = Pr(survive past 5 and survive past 4) ]

  19. Displaying the Survival Function

  20. Notes on Estimating Survival Function • Estimate only changes in intervals where an event occurs • Censored observations contribute to denominators, but never to numerators • Intervals are arbitrary; want narrow ones • Kaplan-Meier estimate results from using infinitesimal interval widths

  21. Acute Myelogenous Leukemia Example Data: 5,5,8,8,12,16+, 23, 27, 30+, 33, 43,45 5 5 8 8 12 16+ 23 27 30+ 33 43 45

  22. Kaplan-Meier Estimate of S(t) – AML Data

  23. Graph of K-M Estimate of Survival Curve for AML Data

  24. K-M Estimate for Risp/Halo Trial

  25. Comparing Survival Functions • Suppose we want to test the hypothesis that two survival curves, S1(t) and S2(t) are the same • Common approach is the “log-rank” test • It is effective when we can assume the hazard rates in the two groups are roughly proportional over time

  26. Logrank test: “Drug trial” data Logrank: 1.72 p-value: .19 Conclusion: We lack strong support for a drug effect on survival

  27. Comparing Survival Functions • Suppose we want to test the hypothesis that two survival curves, S1(t) and S2(t) are the same • Common approach is the “log-rank” test • It is effective when we can assume the hazard rates in the two groups are roughly proportional over time • Regression analysis—“Cox” model: more to come

  28. Regression Analysis for Times to Events • Cox proportional hazards model • Hazard of an event is the product of two terms • Baseline hazard, h(t), that depends on time, t • Relative risk, rr(x) that depends on predictor variables, x, but not time • Each person’s hazard varies over time in the same way, but can be higher or lower depending on their predictor variables x

  29. Cox Proportional Hazards Model • (t,x) = hazard for people at risk with predictor values x = (x1,x2, …..xp) • (t,x) = • ln[(t,x)] =

  30. Cox Proportional Hazards Model • Relative Hazard (hazard ratio) interpretation of the ’s = relative risk for one unit difference in x1 with same values for x2, …. xp (at any fixed time t)

  31. Cox Proportional Hazards Model • Proportional hazards over time: x1 =1 (t;x) x1 =0 0 t

  32. Main Points Once Again • Time to event data can be censored because every person does not necessarily have the event during the study • Think of time to event as a natural history, that is 0 before the event and then switches to 1 when the event occurs; analysis counts the events • Survival function, S(t), is the probability a person’s event occurs after each time t

  33. Main Points Once Again • Kaplan-Meier estimator of the survival function is a product of interval-specific survival probabilities • Hazard function, h(t), is the risk per unit time of having the event for a person who is at risk (not previously had event) • Logrank tests evaluate differences among survival in population subgroups • Cox model used for regression for survival data

More Related