Title: Estimation of marginal structural survival models in the presence of competing risks
1Estimation of marginal structural survival models
in the presence of competing risks
- Maarten Bekaert and Stijn Vansteelandt
- Department of Applied Mathematics and Computer
Science, Ghent University, - Ghent, Belgium
- Karl Mertens
- Epidemiology Unit, Scientic Institute of Public
Health, Brussels, Belgium
Case study Estimation of attributable
mortality of ventilator associated pneumonia
2Motivation
- Attributable mortality of ventilator associated
pneumonia (VAP) on 30-day ICU-mortality - A nosocomial pneumonia associated with mechanical
ventilation that develops within 48 hours or more
after hospital admission - Controversial results in ICU-literature due to
3Main question To what extent does pneumonia
itself, rather than underlying comorbidity,
contribute to mortality in critically ill
patients.
4Informative censoring
- The decision to discharge patients is closely
related to their health status - Patients are typically discharged alive because
they have a lower risk of death. - These patients are therefore not comparable with
those who stayed within the hospital. - Competing risk analysis
- ICU-death ? event of interest
- Discharge from the ICU ? competing event
- Models based on the hazard associated with the
CIF are used in the ICU setting
5Causal inference
- Confounding
- Infected and non-infected patients are not
comparable because they differ in terms of
factors other than their infection status
Severity of illness
Infection
Mortality
Patients severity of illness increases the risk
of VAP and the poor health conditions leading to
VAP are also indicative of an increased mortality
risk.
6Assumption of no unmeasured confounders
Information that leads to acquiring VAP is
completely contained within the measured
confounders
Severity of illness
VAP
Mortality
No unmeasured confounding
Unmeasured confounders
7Non causal paths between VAP and mortality
In a non-randomized setting at a single time
point, we can adjust for confounding variables by
including them in a regression model
Severity of illness
VAP
Mortality
Causal path
Unmeasured confounders
8Time dependent confounding
- Confounders are time-dependent
- They are also intermediate on the causal path
from infection to mortality because infection
makes an increase in severity of illness more
likely
Severity of illnesst1
Severity of illnesst
VAPt
Mortality
VAPt1
9Time dependent confounding
- Association between infection and mortality is
disturbed by time-dependent confounders - severity of illness at time t1 is a confounder
- ? we need to adjust
Severity of illnesst1
Severity of illnesst
VAPt
Mortality
VAPt1
10Time dependent confounding
- Association between infection and mortality is
disturbed by time-dependent confounders - Severity of illness at time t1 may also be
effected by the patients - infection status at time t (lies on the
causal path) - ? we should not adjust
Severity of illnesst1
Severity of illnesst
VAPt
Mortality
VAPt1
11Importance of modelling evolution in severity of
illness
Severity of illness
ICU admission
12Marginal structural survival model in the
presence of competing risks
- Notation
- Let At and Dt be two counting processes that
respectively indicates 1 for ICU-acquired
infection or ICU discharge at or prior to time t
and 0 otherwise. - Under infection path
( 0,0,0,0,1,1,1,1,1,1, ) we would infect all
ICU-patients 5 days after admission - expresses the counterfactual survival
time, which an ICU patient would, possibly
contrary to fact, have had under a given
infection path - represents the counterfactual event
status at time t (0 still alive in ICU, 1
dead, 2 discharged alive from ICU) - For an event of type k (k 1, 2) we define
-
- which is equal to the time until event k
occurs or infinity when the competing event
occurs
13Marginal structural survival model in the
presence of competing risks
- The counterfactual cumulative incidence function
-
- which is the probability that, under an
infection path , an event of type k occurs at
or before time t. - Discrete time setting ? pooled logistic
regression model for the subdistribution hazard
of death
1
1
Its a marginal model because we do not condition
on time varying confounders because they are
themselves affected by early infections !!
For patients who have not died in the ICU, ß2
describes the effect on the hazard of ICU- death
of acquiring infection on a given day t, versus
not acquiring infection up to that day.
14Estimation principle
- How to fit this model
- Select those patients whose observed data are
compatible with the given infection path - Perform a competing risk analysis on those data,
using inverse probability weighting to account
for the selective nature of that subset
15Selection of patients compatible the infection
path no infection
ICU admission Day 1
Day 30
- No infection
- Patients who died or were discharged without
infection
infection
Discharged alive
Died in ICU
16Discharge without infection
ICU admission Day 1
Day 20
Day 30
At
0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 ? ? ? ? ?
? ? ? ? ?
- Patients who are discharged by time t stay in the
risk set - Survival time of infinity (30 days)
- We need to expand the data set
- Several possible infection paths after discharge
infection
Discharged alive
Died in ICU
17Discharge without infection
ICU admission Day 1
Day 20
Day 30
At
0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 ? ? ? ? ?
? ? ? ? ?
Yt
0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 ? ? ? ? ?
? ? ? ? ?
t
1 20 ? ? ?
? ? ? ? ? ? ?
wt
w1 w20 ? ? ?
? ? ? ? ? ? ?
Observed information
18Discharge without infection
ICU admission Day 1
Day 20
Day 30
At
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 1
0 0 0 0 0 0 0 0 1 1
0 0 0 0 0 0 0 1 1 1
0 0 0 0 0 0 1 1 1 1
(30 - time of discharge) 1 possible infection
paths
0 0 0 0 0 1 1 1 1 1
0 0 0 0 1 1 1 1 1 1
0 0 0 1 1 1 1 1 1 1
0 0 1 1 1 1 1 1 1 1
0 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1
0 0 0 0 0 0 0 0 0 0
Yt
0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0
21 30
t
1 20
wt
w20w20
w1 w20
Observed information
19Selection of patients compatible with getting
infection on day 5
ICU admission Day 1
Day 5
Day 30
0 0 0 0 1 1
0 0 0 0 1 1
- Infection on day 5
- Patients who died before day 5
- Patients who acquired infection on day 5 and died
in the ICU within 30 days - Patients who were discharged after day 5 with an
infection acquired on day 5 - Patients who were discharged before day 5
20Estimating equation
21Estimating equation
Weights
- Calculation of the patient specific time
dependent weights - Estimate
using a logistic regression - For patients who are discharged
1 - Calculate the weights as
- where K discharge time
22Data analysis
- Data set
- Data from the National Surveillance Study of
Nosocomial Infections in ICU's (Belgium). - A total of 16868 ICU patients were analyzed.
- Of the 939 (5,6) patients who acquired VAP in
ICU and stayed more than 3 days, 186 (19,8) died
in the ICU, as compared to 1353(8,4) deaths
among the 15929 patients who remained VAP-free in
ICU
23Confounders included in the analysis
- Baseline confounders
- age, gender, reason for ICU admission, acute
coronary care, multiple trauma, presence and type
of infections upon ICU admission, prior surgery,
baseline antibiotic use and the SAPS score - Time dependent confounders
- Invasive therapeutic treatment indicators
collected on day t - indicators of exposure to mechanical ventilation,
central vascular catheter, parenteral feeding,
presence and/or feeding through naso- or
oro-intestinal tube, tracheotomy intubation,
nasal intubation, oral intubation, stoma feeding
and surgery
24Preliminary result
- Crude analysis
- Ignoring informative censoring pooled logistic
regression - When not take into account time dependent
confounding, the OR associated with infection is
equal to 0,67 with 95 CI (0,57 0,79) - Including time dependent confounders as
covariates in the model the OR equals 0,75 with
95 CI (0,63 0,89) - ? infected patients have a significant decreased
mortality
25(No Transcript)
26Competing risk analysis ignoring time dependent
confounding
271. Separated analysis per potential infection path
- We selected patients compatible with a given
infection path - Analyse the data with a weighted pooled logistic
regression model with a flexible time trend. - Plot the cumulative incidence function
28(No Transcript)
292. Results after solving the weighted estimating
equation
- We defined a simple model for the effect of
infection and a quadratic time trend without
taking into acount the baseline confounders - OR equals 1,15 (no estimation of SE available
yet) - Still working on models with a more complex
impact of infection
30(No Transcript)
31Discussion and future work
- When ignoring the informative censoring we get
biased results - In order to get insight into the problem of time
dependent confounding we will do a competing risk
analysis by including the confounders as time
dependent covariates in the model - Work in progress
- Calculation of sandwich estimators of the
standard error - We will develop semi-parametric estimators for
the time-evolution in severity of illness - Using the COSARA data set we will be able to
account for a lot more time dependent confounders - Check results with simulation studies