1 / 27

EPI 5344: Survival Analysis in Epidemiology Left Truncation & Other Issues April 1, 2014

EPI 5344: Survival Analysis in Epidemiology Left Truncation & Other Issues April 1, 2014. Dr. N. Birkett, Department of Epidemiology & Community Medicine, University of Ottawa. Objectives. Left truncation, left censoring, unknown ‘ 0 ’ time Imprecise outcome time Interval censoring.

ellis
Download Presentation

EPI 5344: Survival Analysis in Epidemiology Left Truncation & Other Issues April 1, 2014

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. EPI 5344:Survival Analysis in EpidemiologyLeft Truncation & Other IssuesApril 1, 2014 Dr. N. Birkett, Department of Epidemiology & Community Medicine, University of Ottawa

  2. Objectives • Left truncation, left censoring, unknown ‘0’ time • Imprecise outcome time • Interval censoring

  3. Background (1) • Time ‘0’ is the starting point for follow-up • ‘time-on-study’ • Measured from the time of entry into study • Commonly used in RCTs • ‘age-at-follow-up’ • Time ‘0’ is date of birth • useful for • observational studies • when no clear point of initial exposure

  4. Background (2) • Subject observed at time ‘t’ means: • If subject had an event at time ‘t’, then subject would have been recorded with an event in the study. • Point of initial observation may not be time ‘0’ • age-at-follow-up time scale • enter study at age 45 • first follow-up is at age 45, not 0 • subject must have survived to the time of first observation

  5. Left Truncation (1) • Left truncation (delayed entry) • subject enters at time ‘0’ • Initial observation is at later time ‘t’ • Subject must be outcome free at time of initial observation • Distribution of survival times is ‘truncated’ or ‘cut-off’ in the left tail

  6. Left Truncation (2) • Example #1 • 500 subjects with an acute MI enter hospital (time ‘0’) • 461 subjects were discharged alive • Only these subjects are in study • BUT, survival time is measured from the time of entry to hospital Truncation time

  7. Left Truncation (3) • Left truncation (delayed entry) • What do to with the person-time between time ‘0’ and first observation? • include it bias of h(t) towards ‘0’ • exclude it can  bias away from ‘0’ • depends on the hazard during the interval • Analysis depends on assuming that the hazard starting at the first observation time is unaffected by the delayed entry

  8. Left Truncation (4) • Example study (a bit artificial but shows the methods) • Residential drug-abuse treatment programme • 2 RCT’s at 2 sites • Site ‘A’ • 6 months vs. 12 months • Site ‘B’ • 3 months vs. 6 months • sites combined (for this example) • Short vs. long treatment • n=628 • follow-up from date of admission • up to 1,170 days of follow-up

  9. Left Truncation (5) • In this example, we actually know subject history from time ‘0’ • Could analyze using time varying covariate • ‘0’ for time during treatment • ‘1’ for time after treatment. • Includes the entire sample (n=628) • Short vs. long Rx effect (HR): 1.02 (0.85 – 1.23) • On vs. off-Rx effect (HR): 13.1 (9.6 – 17.8) • Length of treatment has no effect but being ‘off-treatment’ has a strong adverse effect

  10. Left Truncation (6) • Suppose that we know nothing about outcomes during treatment. • Study includes only people who complete treatment and remain drug-free at the end of treatment • n = 546 • Why is this an issue? • Using the standard analysis, ‘long-term’ Rx subjects are assured of having more months of drug-free follow-up time than those in ‘short-term’ Rx • We ignore events during these ‘immortal’ periods. • Introduces a serious bias in favor of ‘long term’ treatment.

  11. Left Truncation (7) Truncation time End Prog: recruit Recid Enter Prog Minimum event time

  12. Left Truncation (8) • No analytic method can recover information on events during the truncation period • If we can ignore this, solution is ‘easy’ • Set the ‘immortal person-time’ as outside the period of follow-up. • The assumption is that the hazard starting from the time of initial observation is unaffected by the delayed entry

  13. Left Truncation (9) • Two approaches can be used in SAS. • Define a truncation time variable • Add a command to the Model statement giving entry time to risk set • Use the Counting process input style • Need to specify (for each subject): • Start of their follow-up time • End of their follow-up time • Outcome status (event vs.. censored) • No need to do any data re-structuring in this case

  14. Left Truncation (10) • Continue Example #2 • Define a variable (LOS) which is the number of days from treatment start to entry into follow-up • Method 1: • add to model statement as ‘entrytime’ • Method 2: • No need to rearrange data • only one interval per subject • starts at ‘LOS’ • ends at study time

  15. Left Truncation (11) PROC PHREG DATA=uis1; MODEL time*censor(0)= age becktota hercoc race treat site / TIES=EFRON; run; PROC PHREG DATA=uis1; MODEL time*censor(0)= age becktota hercoc race treat site / TIES=EFRON entrytime=los; run; PROC PHREG DATA=uis1; MODEL (los, time)*censor(0)= age becktota hercoc race treat site / TIES=EFRON; run;

  16. Ignores late entry

  17. Late entry, method #1

  18. Late entry, method #2

  19. Left Censoring (1) Left Censoring • Subject already has the outcome event when they enter the study • Differs from left truncation • In truncation, subject hasn’t had the event yet. • Truncation involves a subject selection process applied to all subjects • Left censoring is an individual subject event

  20. Left Censoring (2) • Example • Outcome is ‘age of smoking uptake’ • Interview 12 year olds (time ‘0’ = age 12) • One child reports starting regular smoking before age 12

  21. Left Censoring (3) • Approach • Inclusion/exclusion criteria • Exclude prevalent cases in cohort recruitment • Specialized regression methods • Parametric regression. • PH models • Too complex to explore further

  22. Unknown ‘0’ time (1) Unknown ‘0’ time • Subject meets entry criteria but it is unclear when they met the criteria • Example • Study is to examine prognosis in people with AIDS • Some recruited subjects (diagnosed with AIDS) lack information on when they converted from HIV+ to AIDS  lack of a clear time ‘0’ • A common issue which is usually just ignored • Use date of recruitment instead • Models can be used but are complex

  23. Imprecise measure of outcome date (1) • Some examples (share common features) • Tied data • Interval censoring • Sero-conversion to HIV+ status • Jan 1: OK • July 1: (+)ve only know conversion was between Jan 1 and July 1 • Discrete time for event • Event can only occur at fixed points in time • Examination failure can only occur when exam is administered

  24. Interval censoring (1) • Issue is frequently ignored and regular Cox models are used. • Use the time when first known to have outcome as the time-to-event • can build more complex models • estimate distribution of event time within interval • Better approach converts to a discrete Cox model • Uses binary regression models to get the estimates, etc. • Logistic model • Complementary log-log model • Allows analysis using existing software

  25. Interval censoring (2) • Divide the follow-up time into intervals • Equal size is easiest but is not required • Now, we assume that there is a continuous hazard underlying the data • Proportional hazards is assumed: • Now, define:

  26. Interval censoring (3) • It can be shown that: • ‘αt’ depends on the baseline hazard • This is a regression model similar to: but with a different ‘link’ function: ln(-ln(1-pit)) rather than logit(pit) 1

More Related