Transcript Slide 1

Time – Immortal Bias in the analysis of
“Influenza and COPD Mortality Protection as
Pleiotropic, Dose-dependent effects of
statins” by Floyd J, Frost et. al. Chest 2007:
131 (4) 1006-1012.
Contents:
1.
2.
3.
4.
Basic concept of time-immortal bias
Description of the analysis used in this paper
Proper study design – Matched cohort design
Statistical Analysis
- Cox PH regression with time varying covariate
- Poisson regression
5. Simulation using ICU delirium study
Time-immortal bias
KM graphs for time to deaths
Time-varying (actual data)
Time-invarying approach
Non-user
User
User
0
Time after enrollment in a cohort
Initiation of drug use
0
Time after enrollment in a cohort
Bias can be introduced
If analyzed as ever / never user or
non-user.
Time immortal bias
In the time-invarying (ever/never) drug (statin) user or
non-user approach, patients who survived longer were
followed longer which systematically increases the
likelihood of using the drug. They are systematically
more likely to live longer.
Wrong analysis, but still popular in clinical research.
Examples:
(1) the effect of inhaled corticosteroid use on chronic obstructive pulmonary
disease (COPD)(4), Suissa S. Am J Respir Crit Care Med. 2003
Time – invarying
RR: 0.73 (95%CI: 0.57-0.93)
Time- varying
RR:0.98 (95%CI: 0.77-1.25)
(2) the effect of statin use in reducing mortality among patients with acute
coronary syndromes. Laupacis A, Mamdani M. Observational studies of treatment
effectiveness: some cautions. Ann Intern Med. 2004 Jun 1;140(11):923-4.
Time – invarying
OR: 0.78 (95%CI: 0.70 – 0.86)
Time - varying
NA
(3) Do OSCAR winners live longer than less successful peers? A reanalysis
of the evidence. Sylvestre MP, Huszti E, Hanley JA. Ann Intern Med. 2006
Time – invarying
HR: 23 (95%CI: 8 - 40)
Time-varying
HR: 15 (-5 - 32)
Van Walraven et al. conducted a literature review of nine prominent medical
journals between 1998 and 2002 and found that 41% (95% CI = [32%, 50%]) of
reported observational studies using survival analyses were susceptible to
immortal-time bias.
Two popular methods to deal with time-immortal bias:
(1) Proper study design - i.e., Matched Cohort
Method
In the analysis of the effect of nosocomial infection on ICU length of stay,
controls were matched by number of days spent in the ICU before the
onset of the infection corresponding patients with the infection, and
remaining days in ICU were used as outcome. Pittet D et. al. (1994)
Nosocomial Bloodstream Infection in Critically Ill Patients. JAMA (1994) 271: 1598-1601
Unbiased but loss of statistical power by matching
(2) Through proper regression technique….. (will talk later)
Description of the study design by Floyd et al. –
How did they deal with the time-immortal bias.
1993 Jan
Statin-use
enrolled
Non statin-use
2003 Dec
died
Cohort definition: Patient identified in the above period with COPD with HMO
enrollment > 90 days
Exposure – total 90 + days with statin exposure
Non-exposure - Pick 3 from the rest of the cohort to each of exposed pt matched
on sex, birth year, and HMO enrollment period
Event – died at hospital discharge
Description of the analysis by Floyd et al. –
How did they deal with the time-immortal bias.
Logistic regression was used to compare proportion of
death between exposed and non-exposed groups
adjusting for the number of days enrolled before ( ) and
after statin initiation of the statin-exposed individual ( ).
1993 Jan
Statin-use
Non statin-use
died
enrolled
Questions:
• Are these two time variables,
,
0 for non-exposed patients?
• Are adjustment of time after the initiation properly control different follow-up time?
Are there better models over logistic regression accounting for varying follow-up time?
• Are adjustment of time before the initiation of statin proper to control for timeimmortal bias?
Two popular methods to deal with time-immortal bias:
Through proper regression techinique…..
(2) Analysis with Statin Use as a time-varying covariate
controlling for the onset of delirium.
- Cox PH regression with time-varying covariate
- Poisson regression
Simulation Study using ICU delirium data (1)
Time-invarying Cox:
Median
95% CI
Time-varying Cox:
Median
95% CI
Rate of Delirium
Simulation Study using ICU delirium data (2)
15
Linear regression with log
transformed outcome
5
10
Median
95% CI
Time-varying Cox
Median
0
Exp(beta) for delirium (vs no delirium)
20
ICU Length of Stay
95% CI
0.0
0.2
0.4
0.6
Rate of risk factor
0.8
1.0
ICU Length of Stay
15
Linear regression with log
transformed outcome
Adjusting for time to delirium
5
10
Median
95% CI
Time-varying Cox
Median
0
Exp(beta) for delirium (vs no delirium)
20
Simulation Study using ICU delirium data (3)
95% CI
0.0
0.2
0.4
0.6
Rate of risk factor
0.8
1.0
Simulation Study using ICU delirium data (4)
2.5
Poisson regression:
2.0
Median
95% CI
1.5
Time-varying Cox:
1.0
Median
0.5
95% CI
0.0
Exp(beta) for delirium (vs no delirium)
3.0
ICU Length of Stay
0.0
0.2
0.4
0.6
Rate of risk factor
0.8
1.0
Examples:
(1) the effect of inhaled corticosteroid use on chronic obstructive pulmonary
disease (COPD)(4), Suissa S. Am J Respir Crit Care Med. 2003
Time – invarying
RR: 0.73 (95%CI: 0.57-0.93)
Time- varying (Poisson Regression)
RR:0.98 (95%CI: 0.77-1.25)
(2) the effect of statin use in reducing mortality among patients with acute coronary
syndromes (6, 7). Laupacis A, Mamdani M. Observational studies of treatment effectiveness:
some cautions. Ann Intern Med. 2004 Jun 1;140(11):923-4.
Time – invarying
OR: 0.78 (95%CI: 0.70 – 0.86)
Time - varying
NA
(3) Do OSCAR winners live longer than less successful peers? A reanalysis of the
evidence. Sylvestre MP, Huszti E, Hanley JA. Ann Intern Med. 2006
Time – invarying
HR: 23 (95%CI: 8 - 40)
Time-varying (Cox with time-varying cov)
HR: 15 ((95%CI: -5 - 32)
Van Walraven et al. conducted a literature review of nine prominent medical
journals between 1998 and 2002 and found that 41% (95% CI = [32%, 50%]) of
reported observational studies using survival analyses were susceptible to
immortal-time bias.