Time-dependent variables Stat 532 Presentation JM Gamble, BSc, BSc(Pharm), MSc
Time-dependent variables
Stat 532 Presentation
JM Gamble, BSc, BSc(Pharm), MSc
Outline
1. A Historical example
2. Definitions and theory
3. A scientific question: bisphosphonates and pneumonia-related adverse events
A Historical example
1. Do Oscar winner’s live longer?
A Historical example
• 28% (95% CI: 10% - 42%) relative reduction in death rates between Oscar winners and less successful performers
• 4 year survival advantage
• 20% (95% CI: 0% - 35%) relative reduction using time-dependent variable
• Data has been re-analyzed extensively
• Debate on the best analytical approach
A Historical example
Oscar Winner
No Award
death or censored
Birth
death or censored
Date of Award
A Historical example
Subjects must remain event free until the start of exposure to be classified as exposed
Received transplant
Did not receive transplant
death
Accepted for transplant
Cohort exit; death or censored
transplant
A Historical example
exposed
unexposed death
Cohort exit exposure.( i.e. first drug use)
Cohort entry (i.e. diagnosis)
Cohort entry (i.e. diagnosis)
Adapted from figure 1. Suissa S. Pharmacoepidemiology & Drug Safety 2007;16:241-249.
Definitions & Theory
Time-dependent:
• Value of variable differs over time
Time-independent:
• Value of variable is constant over time
Definitions & Theory
Cox PH model
Extended Cox PH model
Hazard Ratio
Definitions & Theory
A. Defined time-dependent variable
• Commonly a product of a time-independent variable and time (or some function of time)
• Values are completely defined over study period
• Sex x time
• Baseline drug level x exp(-0.35 x time)
• Exposure x g(t) where g(t) is a heavy-side function
Types of Time-dependent variables:
Definitions & Theory
B. Internal
• Values may change over time due to subject specific characteristics
• Subject must be under periodic observation
• Physiological status
• Drug treatment
Types of Time-dependent variables:
Definitions & Theory
C. External
• May or may not be subject specific
• Subject not required to be under observation
• Environmental factor (i.e. air pollution)
• Age
• time
Types of Time-dependent variables:
Definitions & Theory
• The hazard at time t depends on the value of Xj(t) at the same time … not an earlier or later time – May account for lag-time effect if
appropriate
• Reason for change in value of time-dependent variable (i.e. treatment switching) must be unrelated to risk of the outcome event
Some key assumptions of the Cox PH model using a time-dependent variable:
Research question
In patients admitted to the hospital with community-acquired pneumonia, does exposure to a bisphosphonate decrease the risk of pneumonia-related mortality or hospitalization?
Background
• Bisphosphonates – Class of medications used to prevent
fractures • Signal from RCT data
– Colón-Emeric et al. Potential Mediators of the Mortality Reduction with Zoledronic Acid After Hip Fracture. J Bone Miner Res (2009) online early.
– Decreased the relative risk of death by 25% (95% CI: 0.58-0.97)
– Found a statistically significant decrease in pneumonia-related death (p=0.04).
Study Design
Time 0
Cohort design: time-varying drug exposure
Statistical analysis
• Cox Proportional hazards model • Primary predictor/explanatory/exposure
variable – bisphosphonate use
• Dichotomous variable (coded as 1 for exposed) • Time-dependent • Heavy-sided form
• Primary outcome variable – Composite of pneumonia-related mortality and
hospitalizations
• Other covariates included in the model based on scientific/clinical rationale
Cox PH model
pneum_comp: composite outcome of CAP-related death and CAP-related hospitalization
Ebisp: time-dependent bisphosphonate exposure (0=no exposure; 1= exposed)
age: age in years at date of presentation sex: male=0; female=1 Functional_Status: 0=walking with/without
assistance; 1=wheelchair/prosthesis; 2=bedridden
Risk_class_combined: Pneumonia severity index; categories 1 through 4 indicated more severe pneumonia.
Variables in the model:
Results
Cohort Selection:
• 3415 subjects admitted for community-acquired pneumonia
– 27 subjects excluded due to prior bisphosphonate exposure
– 131 subjects dropped due to unable to link to long-term outcomes
• N=3257 subjects available for analysis
Results Cohort Characteristics Variable Bisphosphonate
N= 200 No Bisphosphonate
N=3057
Age, years (mean, sd) 75 (9) 68 (18)
Age, n(%) < 45 yrs 45 – 64 yrs ≥ 65 yrs
0 (0) 27 (14)
173 (87)
426 (14) 648 (21) 648 (21)
Female, n(%) 145 (73) 1391 (46)
Functional Status, n(%) Walking Wheelchair/prosthesis Bedridden
193 (97) 3 (2) 4 (2)
2711 (89) 223 (7) 123 (4)
Risk Class, n(%) Class I/II Class III Class IV Class V
24 (12) 58 (29) 94 (47) 24 (12)
591 (19) 546 (18)
1200 (39) 720 (24)
Results Descriptive results
Results Descriptive results
Results Descriptive results
Setting up data for survival analysis:
Data structure
Setting up exposure variable as time-dependent:
. replace bisph_date1=mdy(4,1,2006)+1 if bisph_date1==.
. stsplit Ebisp, after(time=bisph_date1) at(0)
. replace Ebisp=Ebisp+1
Setting up data for survival analysis:
Data structure
Results: crude HR
Results - aHR
Results: time-independent exposure
Results
Conclusion
• Bisphosphonates do not appear to decrease
the risk of pneumonia-related death or
hospitalizations.
References
• Kleinbaum & Klein 2005. Survival Analysis: A Self-Learning Text. 2nd Edition. Chapter IV: The Extended Cox Model for Time-Dependent Variables. Springer: New York, USA
• Hosmer & Lemeshow 1999. Applied Survival Analysis: Regression modeling of time to event data. Chapter 7: Extensions of the Proportional Hazards Model. John Wiley & Sons: New York, USA.
• Kalbfleisch & Prentice 2002. The Statistical Analysis of Failure Time Data. Second Edition. Chapter 6: Likelihood Construction and Further Results. John Wiley & Sons: New Jersey, USA.
• Cleves, Gould, & Gutierrez. An Introduction to Survival Analysis Using Stata. Revised Edition. Stata Press: Texas, USA.