Académique Documents
Professionnel Documents
Culture Documents
This appendix has been provided by the authors to give readers additional information about their work.
Supplement to: Amato MBP, Meade MO, Slutsky AS, et al. Driving pressure and survival in the acute respiratory
distress syndrome. N Engl J Med 2015;372:747-55. DOI: 10.1056/NEJMsa1410639
Cardio-Pulmonary Department, Pulmonary Divison, Heart Institute (Incor), University of So Paulo, So Paulo, Brazil;
Departments of Clinical Epidemiology, Biostatistics and Medicine, McMaster University, Hamilton, Ontario, Canada;
Keenan Research Centre for Biomedical Science, St. Michaels Hospital, Toronto, Ontario, Canada;
Interdepartmental Division of Critical Care Medicine, and Department of Medicine, University of Toronto, Ontario, Canada;
Massachusetts General Hospital Biostatistics Center, Harvard Medical School, Boston, MA;
Basel Institute for Clinical Epidemiology and Biostatistics, University Hospital Basel, Switzerland;
Department of Anesthesia, Critical Care, and Pain Medicine, Beth Israel Deaconess Medical Center, Harvard Medical School,
Boston, MA;
Department of Intensive Care and Hyperbaric Medicine, Angers University Hospital, Angers, France;
10
Emergency Department, General Hospital of Annecy, Annecy, France and INSERM UMR 955, Creteil, France;
11
Pulmonary and Critical Care Medicine, Johns Hopkins University School of Medicine, Baltimore, Maryland.
1. Accounting for residual (intrinsic) heterogeneity across the trials (Figure S2)
2. Univariate analysis (Table S3)
3. Length of risk exposure and test of proportional hazards assumption (Tables S4-S6)
4. Sensitivity analysis for different estimates of baseline elastance (Figure S3)
5. Homogeneous P-risks across the trials (Figure S4)
6. Consistency of higher P-risks in the validation cohorts (Tables S7-S8)
7. Tidal volume predicts survival only if normalized to compliance (CRS) (Figure S5)
8. Survival in patients under protective ventilator settings (Figure S6)
9. P (but not VT) predicts Barotrauma after randomization (Figure S7)
10. Mediation analysis: more than grading the severity of lung disease
Please, refer to the tables and figures within the next pages
Tables S1-S2:
Studied cohorts and baseline patient characteristics recalculated from individual
patient data
Figure S1:
Overview of the results of randomization in each of the trials
Table S1(website): Studied cohorts and baseline patient characteristics recalculated from individual patient data.
The trials in the first four rows were pooled and formed our hypothesis generation sample, used to elect a multivariate model for survival (Model-1, Table 1). The ARDSnetVT
study was used as a first validation sample. The studies in the last four rows (testing a higher vs. lower-PEEP strategy) were pooled and used as a second validation sample.
Years of
recruitment
Patients
(N)
Randomization
Cont. / Treat.
Age
mean (SD)
Sepsis at
Entry (%)
Pneumonia/
Aspiration*
MV.Days
at entry
Interventions
(within treatment-arm)
Outcome Treatment-arm
(RR; 95%CI)
Amato et al.1
1991-1995
53
24 / 29
34 (13)
83%
28%
VT 6mL/kg; P 20cmH2O
PPLAT 40cmH2O;
0.38 (0.180.79)
Stewart et al.2
1995-1996
118
59 / 59
59 (18)
40%
58%
VT 8mL/kg;
PPEAK 30cmH2O
0.99 (0.601.70)
Brochard et al.3
1994-1996
113
57 / 56
57 (15)
n.a.
n.a.
VT < 10mL/kg;
PPLAT 25cmH2O
1.28 (0.732.25)
Brower et al.4
1994-1996
52
26 / 26
48 (16)
23%
54%
n.a.
VT 8mL/kg;
PPLAT 30cmH2O
1.11 (0.482.57)
ARDSnetVT5
1996-1999
861
429 / 432
51 (17)
27%
49%
VT 6mL/kg;
PPLAT 30cmH2O;
0.74 (0.580.93)
ARDSnetPEEP6
1999-2002
545
271 / 274
51 (17)
38%
55%
1.11 (0.801.54)
EXPRESS7
2002-2005
0.87 (0.691.09)
LOVS8
2000-2006
983
508 / 475
Talmor et al.9
2004-2007
61
31 / 30
VT-trials :
PEEP-trials :
767
382 / 385
60 (15)
61%
72%
1.5
56 (17)
47%
64%
0.88 (0.711.08)
refract. hypoxemia
53 (20)
48%
20%
n.a.
0.49 (0.201.24)
oxygenation
compliance, rs
ARDSNetTV: First ARDSNet study5 comparing high versus low tidal volume strategies
ARDSNetPEEP: Second ARDSNet study6 comparing high versus low PEEP strategies
*: P < 0.001 - Chi-squared test comparing differences in prevalence of primary ARDS across the trials.
RR: non-adjusted relative-risk (mortality-rate) associated with the treatment arm - calculated by Cox Proportional Hazards regression.
: In the supplement (Figure S1) we present the results of the adjusted relative-risk according to Model-1 (Table 1), where we noticed two important findings:
- the intervention arm in the EXPRESS trial presented a significant reduction in the relative-risk: 0.75 (95%CI: 0.590.96; P=0.02).
- the intervention arm in the ARDSnetPEEP trial presented an inversion of the trend for the relative-risk:0.82 (95%CI: 0.561.12; P=0.29).
(This was related to an important imbalance in covariates at baseline, as reported in the original publication of this trial).
95% C.I. 95% confidence interval;
Table S2 (website): Baseline patient characteristics: recalculated from individual patient data.
The trials in the first four rows were pooled and formed our hypothesis generation sample, used to elect a multivariate model for survival (Model-1, Table 1 - main text).
The ARDSnetVT study (testing a high versus low V T strategy) was used as a first validation sample. The studies in the last four rows (testing a high versus low PEEP
strategy) were pooled and used as a second validation sample.
Risk of Death*
arterial pH at entry
PaO2/FIO2 at entry
Tidal-compliance at entry or
at first day
Amato et al.1
48(35-71)
7.32 (7.24-7.40)
0.40 (0.32-0.53)
Stewart et al.2
40(26-64)
7.39 (7.33-7.43)
0.53 (0.39-0.66)
Brochard et al.3
23(15-36)
7.36 (7.30-7.42)
0.52 (0.38-0.64)
Brower et al.4
44(19-64)
7.42 (7.36-7.45)
0.45 (0.36-0.56)
ARDSnetVT5
39(24-64)
7.41 (7.36-7.45)
0.46 (0.36-0.60)
ARDSnetPEEP6
49(29-64)
7.39 (7.32-7.43)
0.47 (0.36-0.60)
EXPRESS7
42(23-68)
7.37 (7.31-7.42)
0.47 (0.39-0.60)
LOVS8
54(35-74)
7.37 (7.30-7.42)
0.45 (0.36-0.58)
Talmor et al.9
60(46-72)
7.38 (7.33-7.42)
0.47 (0.39-0.59)
Trial:
Values represent the median, with the interquartile range inside parenthesis
*Risk of Death was calculated according to the equations of APACHE II, APACHE III and SAPS II, depending on the individual scores available in each of the trials.
Tidal compliance is represented in milliliters per centimeter of water / ideal body weight. In the hypothesis generation sample (first four rows) as well as the ARDSnet VT
trial, the value was obtained from the first measurement after randomization, taken a few hours after entry;
ARDSnetVT: First ARDSNet study 5 comparing lower versus higher tidal-volume strategies
ARDSnetPEEP: Second ARDSNet study 6 comparing higher versus lower PEEP strategies
Figure-S1:
Hypothesis-generation
Cohort ( N = 336)
Higher vs. Lower VT
study -trials
Adjusted survival
(survival curves were pre-adjusted according to covariates 1-5 of model-1 ; Cox proportional-hazards)
*
Amato et al.
Stewart et al.
First validation
Cohort ( N = 861)
Higher vs. Lower VT
study -trial
Adjusted survival
Brochard et al.
Brower et al.
Days after randomization
Treatment arm
Control arm
ARDSnetVT
(the study of Amato et al. tested a combined strategy of higher PEEP and lower VT)
Second validation
Cohort ( N = 2360)
Higher vs. Lower PEEP
study -trials
Adjusted survival
ARDSnet PEEP
Days after randomization
EXPRESS
Days after randomization
LOVS
Days after randomization
EPVENT
Days after randomization
II)
Additional details:
As shown in Figure S2, in spite of covariate adjustments according to Model-1, there was some
residual, unexplained heterogeneity in the pooled mortality (both arms considered together)
across trials (P < 0.001). Overall the mortality in the derivation cohort (45.3%) was higher than
that observed in the validation cohorts (33.6% and 34.2%, for ARDSnetVT cohort and second
validation cohort, respectively). This higher mortality was a general trend observed in both arms
of each of the studies (from derivation cohort) and could not be fully explained by their baseline
disease (expressed by the covariates Age, APACHE III, arterial pH, PaO2/FIO2ratio) or by P
applied. The source of this heterogeneity is beyond the scope of this study and might relate, for
instance, to improvements in general patient support, independent of ventilation strategies.
We must stress that this residual heterogeneity did not cause any bias in Table 1 or Figures 1-2
presented in the main manuscript, since we pre-adjusted our survival models according to a
categorical variable "Trial". The reported effects of P on mortality were, therefore, necessarily
calculated in proportion to this intrinsic pooled mortality of each trial.
1
2
study-trial:
1
10
11
(N = 336)
VARIABLES:
(N = 861)
P-value
RR (95% C.I.)
(N = 2360)
P-value
RR (95% C.I.)
0.27
Trial *
P-value
RR (95% C.I.)
---
< 0.001
Randomized arm
0.67
0.01
0.13
Days on MV before
0.16
---
---
---
---
Age
0.68
< 0.001
< 0.001
APACHE/SAPS risk
< 0.001
< 0.001
< 0.001
---
---
< 0.001
< 0.001
Arterial pH at entry
< 0.001
< 0.001
< 0.001
PaO2/FIO2 at entry
< 0.001
0.01
< 0.001
---
---
---
---
< 0.001
P at entry
---
---
---
---
< 0.001
0.02
0.29
< 0.001
st
0.22
0.05
< 0.001
< 0.001
< 0.001
< 0.001
0.37
0.16
0.92
Organ Failures
PaCO2 - 1 day
st
FIO2 - 1 day
VT - 1st day
st
0.12
< 0.001
< 0.001
st
< 0.001
< 0.001
< 0.001
st
PEEP - 1 day
0.09
< 0.001
0.003
P - 1 day
< 0.001
0.001
< 0.001
< 0.001
< 0.001
< 0.001
st
12
13
Table S4:
Non-parametric Correlation Between Individual Values Observed During the First Day of
Mechanical Ventilation (Ventilation-Variables), and the Individual Values Observed in
the Following Days
Table S5:
Multivariate Cox Regression Model (60-day Hospital Mortality) comparing the
performance of Ventilation-Variables on Day 1 versus Days 1 to 3.
Table S6:
Comparison of the Original Model 1 (with constant hazards) versus Alternative Models
with Time-Dependent Covariates Included..
14
Table S4 (website):
Non-parametric Correlation Between Individual Values (Ventilation-Variables Observed During the First Day of Mechanical
Ventilation) and the Individual Values Observed in the Following Days
Mean value
st
FIO2 - 1stday
st
VT - 1 day
Respir. rate - 1stday
Plateau Press. - 1st day
PEEP - 1st day
Driving Press. - 1st day
Mean PAW 1stday
1 day
2ndday
3rdday
4thday
1
1
1
1
1
1
1
0.64*
0.87*
0.70*
0.66*
0.73*
0.64*
0.69*
0.51*
0.79*
0.59*
0.56*
0.62*
0.51*
0.63*
0.47*
0.75*
0.54*
0.56*
0.58*
0.52*
0.57*
7th Day
0.29*
0.68*
0.36*
0.51*
0.48*
0.52*
0.50*
* :P < 0.001 ; P-value of the two-tailed test of significance for the Spearmans-Rho correlation-coefficient. The correlation was
calculated between individual data collected at day one (first 24 hs. after randomization) and data at each of the following days,
for the same respective patients.
15
Table S5 (website):
Multivariate Cox Regression Model (60-day Hospital Mortality) comparing the performance of VentilationVariables on Day 1 versus Days 1 to 3.
(data from the ARDSNetPEEP trial)
Considering Ventilation-Variables
to 1st day.
Considering Ventilation-Variables to
3rd day
- Multivariate -
- Multivariate -
RR (95% C.I.)
P-value
RR (95% C.I.)
P-value
Model:
(1) Age
< 0.001
< 0.001
< 0.001
< 0.001
0.39
0.12
< 0.001
< 0.001
0.81
0.72
0.77
0.27
(7) Driving-pressure
0.001
0.001
139.9
(P =2 x10-26)
85.1
st
Model Chi-Square
(change after including all covariates)
(P = 5 x10-15)
The average values in time were used for ventilator-variables collected from days 1 to 3.
Only patients surviving longer than 1or 3 days were respectively included in the Cox survival model. This explains why the overall Chi-Square decreased,
despite a preserved association between individual covariates and survival.
RR: adjusted relative risk associated to 1 standard-deviation increment in the respective variable.
95% C.I. 95% confidence interval
16
Table S6 (website):
Original
model
( RR and P-values below refer to the long-term hazard observed during the 60-day period)
RR P-value
RR
P-value
RR
(2) Age
1.71
< 0.001
1.59
(3) APACHE/SAPS-risk
1.38
< 0.001
0.68
P-value
RR
P-value
RR
P-value
RR
P-value
< 0.001
1.58
< 0.001
1.59
< 0.001
1.58
< 0.001
1.24
< 0.001
1.39
< 0.001
1.38
< 0.001
1.38
< 0.001
< 0.001
0.69
< 0.001
0.80
< 0.001
0.68
< 0.001
0.68
< 0.001
0.87
< 0.001
0.87
< 0.001
0.87
< 0.001
0.92
0.33
0.87
< 0.001
1.41
< 0.001
1.40
< 0.001
1.40
< 0.001
1.40
< 0.001
1.35
< 0.001
Age
APACHE III
1.38
1.74
(0.001)
P-value
(<0.001)
arterial pH
at entry
PaO2/FIO2
at entry
0.56
0.77
1.49
(<0.001)
(0.02)
( 0.12 )
17
18
trial because of the lowest number of early deaths and missing cases (until day 7) of
mechanical ventilation. As shown, the effect-size of most ventilation variables was either
maintained or increased after considering longer periods of time exposure, suggesting an
ongoing, cumulative effect. However, the power of the analysis decreased and confidence
intervals widened due to the smaller sample size. After also considering the potential survival
bias introduced by such procedure -the selection of healthier patients, able to survive the first 3
days11 - we preferred to use the simpler and more powerful analysis, only considering risk
exposure to 24 hours of mechanical ventilation.
Finally, we tested the possibility that ventilation covariates exerted only transient effects that
diminished after a few days. This would represent a violation to the proportional hazard
hypothesis, which assumes an ongoing hazard till the end of the 60-day period. Thus, using the
approach suggested by Kasal et al.12, we assigned, for each covariate, a respective timevarying covariate that allows a different hazard (higher or lower) to be applied over days 0 7,
in addition to the fixed hazard (constant during the 60days) already included in the model.
Whenever this new time-dependent covariate brought additional information to the model (at a
significance level of P < 0.01), we considered that there was a non-proportional hazard (higher
or lower) during the first week of mechanical ventilation. Such analysis is presented in
Table S6. The most relevant non-proportional hazards were observed for lower values of
baseline arterial-pH and higher values APACHE-III/SAPS-risk, both conditions associated with
higher mortality during the first week(in addition to their long term hazard). High drivingpressures, however, did not impose additional risks during the first week.
We concluded, therefore, that high driving-pressure exposure during the first days of
mechanical ventilation was strongly associated to a fixed, ongoing hazard during the first 60day after randomization. There was no need of more sophisticated models, with timedependent variables, to describe such relationship.
19
20
( baseline data )
Figure S3: Correlation between the two estimates for baseline ElastanceRS
Using the data from 1656 patients participating in the higher-PEEP trials, for whom we had measurements of
elastance performed at baseline (average VT = 82 mL/kg/ibw and PEEP = 104 cmH2O), as well as after
randomization, we could check the correlation between the two estimates. In order to avoid post-treatment
bias in the measurements performed after randomization, we calculated elastance ranks within each of the
arms, and within each of the studies. The ranks were then scaled within the [-0.5 to 0.5] interval. This
procedure was performed under the reasonable assumption that the systematic changes in ventilation
parameters due to randomization might affect the absolute values of elastance, but could hardly affect the
ranking of individual elastances within the respective study-arm. As shown, the relationship between both
variables was reasonably linear, with similar slopes and determination coefficients for both arms. This
suggest that the stratified ranking avoided post-treatment bias.
21
22
4. Finally, we observed that after pre-adjusting model-1 for elastanceRS-ranks (posttreatment), the test of entry of baseline-elastanceRS (pre-treatment) in the model was no
longer significant. This suggest a consistent overlap of information coming from both
methods of estimation of underlying lung disease: there was no further independent
information (correlated with outcomes) in baseline-elastanceRS that was missed in the
prediction models, or in the mediation models.
This sequence of tests shows that, in the hypothesis that baseline-elastanceRS was causing
some confounding effect in our mediation analysis, this bias was equally removed from P
by both adjustments (i.e. using pre or post randomization data).
23
Figure S4:
Relative risk of death associated to increments in P within each of the trials
Additional details:
Model adjustments for unexplained differences in the pooled mortality (both arms together) for
each trial and for differences in effect-size across the trials did not change the relative risk
associated with general increments of one standard deviation in P. The first adjustment was
performed by assigning one dummy variable for each trial in the Cox model (Figure S2); the
effect-size adjustment was performed by assigning one dummy interaction term for each trial
expressing the freedom for each trial to present a peculiar response to P.
After all such adjustments1, the relative risk associated to increments of P was consistently
1.45 (95% CI, 1.28 to 1.64; P < 0.0001), similar to the numbers presented in Table 1, main text.
This analysis suggests a consistency of effects: whatever the individual severity of disease
(expressed by baseline covariates), or whatever the baseline mortality of a specific population
(expressed by the dummy variables representing each trial), increments in P were always
deleterious, and associated with the same risk magnitude across the 9 trials (Figure S4).
The inclusion of such dummy interaction terms in the Cox model did not contribute to its predictive power
(likelihood-ratio test: P = 0.13) suggesting that the effect-size of increments of P was homogeneous across the
different trial populations.
24
Figure S4: Relative risk of death associated to increments in P within each of the trials
Figure-S2: Relative risk of death associated to increments in P within each of the trials
Benefit
Amato etAmato
al.
n=53
Stewart
Stewart et
al.
n = 120
Brochard Brochar
et al.
n = 116
ARDSnet
ARDSnet
VT
n = 861
ALVEOLI
ARDSnet
PEEP
n = 549
EXPRESS
EXPRESS
n = 767
LOVS
LOVS
n = 983
Talmor EPVENT
et al.
OVERALL
COMBINED EFFECTS
Heterogeneity test:
P = 0.13
n = 52
Y Data
Brower
Brower et
al.
Hazard
n = 61
P < 0.0001
n = 3562
0.33
0.5
0.7
1.4
X Data
Adjusted Relative-risk
of death
( for one-standard-deviation increment in P )
25
Table S7:
Multivariate Cox Regression Model (60-day Hospital Survival)
Original derivation model and posterior test in the ARDSNetVT population (first validation
cohort)
Table S8:
Multivariate Cox Regression Model (60-Day Hospital Survival)
Posterior test of derivation model in the VT trials (derivation cohort combined with the
ARDSNetVT cohort) and in the PEEP trials (second validation cohort) .
This table complements Table 1, main manuscript.
26
Table S7 (website): Multivariate Cox Regression Model (60-day Hospital Survival) Original derivation model and posterior test in the ARDSNetVT population (first validation cohort)
- Multivariate -
- Multivariate -
- Multivariate -
P-value
RR (95% C.I.)
Refined model
ARDSNETVT trial
Model:
RR (95% C.I.)
P-value
RR (95%
P-value
C.I.)
RR(95
< 0.001
0.001
< 0.001
0.013
< 0.001
(3) P - 1 day
< 0.001
0.001
< 0.001
0.014
0.001
0.81 (0.71 0.92)
0.001
< 0.001
st
st
N.S.
(6) Age
N.S.
N.T.
N.T.
Model Chi-Square
(step change after inclusion
of block of covariates)
78.7
(P =3 x10-16)
81.6
(P =1 x10-16)
145.7
(P =1 x10-29)
ARDSnetVT: First ARDSNet study 5 comparing lower versus higher tidal-volume strategies
RR: adjusted relative risk associated to one standard-deviation increment in the respective variable. Values above 1.00 indicate increased mortality-rate. The values used for standarddeviation were: age (17), death-risk (26), arterial pH (0.09), PaO2/FIO2 (60), P (7), FIO2 (0.19).
27
(N = 2060)
P-value
RR(95% C.I.)
(N = 3080)
P-value
RR(95% C.I.)
Combined analysis
- Multivariate -
P-value
RR(95% C.I.)
Model 1:
--1.51(1.36 1.69)
1.34(1.20 1.49)
0.69(0.63 0.77)
0.85(0.77 0.95)
1.35(1.24 1.48)
< 0.001
< 0.001
< 0.001
< 0.001
0.004
< 0.001
1.64(1.50 1.79)
1.41(1.29 1.54)
0.68(0.63 0.74)
0.88(0.80 0.96)
1.50(1.35 1.68)
0.83
< 0.001
< 0.001
< 0.001
0.005
< 0.001
1.59(1.48 1.70)
1.38(1.29 1.48)
0.68(0.64 0.72)
0.87(0.81 0.93)
1.41(1.31 1.51)
< 0.001
< 0.001
< 0.001
< 0.001
< 0.001
< 0.001
< 0.001
0.12*
1.48(1.28 1.71)
0.98(0.88 1.10)
< 0.001
0.75*
1.41(1.30 1.53)
1.01(0.92 1.10)
< 0.001
0.90*
< 0.001
0.42*
1.51(1.35 1.68)
1.05(0.90 1.23)
< 0.001
0.52*
1.40(1.30 1.51)
1.02(0.95 1.10)
< 0.001
0.58*
0.008
0.65*
1.51(1.31 1.75)
0.99(0.87 1.13)
< 0.001
0.90*
1.37(1.22 1.53)
1.04(0.93 1.15)
< 0.001
0.53*
< 0.001
0.78*
1.50(1.34 1.68)
0.99(0.91 1.09)
< 0.001
0.90*
1.41(1.32 1.52)
1.03(0.95 1.11)
< 0.001
0.51*
(1) TRIAL
(2) Age
(3) Risk of Death
P - 1st day
---
---
28
RR: adjusted relative-risk associated to 1 standard-deviation increment in the respective variable. Values above 1 indicate increased mortality-rate. The values used for
standard-deviation were: age (17), death-risk (26), arterial pH (0.09), PaO2/FIO2 (60), P (7), PEEP (5), Plateau pressure (7), Tidal volume (2), Compliance, RS (0.3). By
normalizing RR in this way, the strength of the association of different variables with survival can be grossly compared as the RR per se (using 1/RR when RR < 1). For
instance, in the combined analysis, P showed stronger association with survival (1.4) than the PaO2/FIO2 (1/0.87 = 1.15)
st
95% C.I. 95% confidence interval; P - 1 day: average driving-pressure during the first 24 Hs after randomization.
* Test of variable inclusion in the model (net contribution to predictive power likelihood-ratio test) where variables 1-6 plus driving-pressure were previously included.
Test of variable inclusion in the model where variables 1-5 plus the extra covariate in the line below were previously included.
Risk of Death was calculated according to the equations of APACHE II, APACHE III and SAPS II, depending on the trial.
Although not shown in the table, the variable mean-airway-pressure was tested before and after inclusion of P in model 1, showing no significant association with survival.
29
30
Figure S5: Survival effects of tidal volume, before and after lung-sizing
ResamplingD
D:
Resampling
ResamplingEE:
Resampling
- matched PPLAT,
- decreasing ranks of VT / PBW
25
10
10
Scaling
VT
20
15
88
to
CRS
Instead10 of PBW
66
DRIVING - PRESSURE
25
( cmH2O )
VT / CRS
(P, cmH2O )
44
611
620
611
623
607
1.2
1.2
1.0
1.0
P = 0.92
0.6
0.6
0
30
30
25
25
25
20
20
20
15
15
15
10
10
10
600
1.6
1.4
1.4
0.8
0.8
30
1.6
Relative risk
Relative risk
( Sub-sample1.6
N ):
1.6
35
S1
S: MATCHED
SPPLAT
S4
SOF
RESAMPLING
, QUINTILES
2
3
5 VT
( mL / kg )
30
VT / PBW
30
12
PPLAT
PLATEAU-PRESSURE
TIDAL - VOLUME
( cmH2O )
35
35
35
P ( cmH2O )
14
35
PPLAT
- matched PPLAT,
- decreasing ranks of VT / CRS (=P)
624
644
598
PLATEAU-PRESSURE
Figure 1B:
614
1.4
1.4
1.2
1.2
1.0
1.0
0.8
0.8
P < 0.001
0.0001
0.6
0.6
S1 : MATCHED
S2
S3 , QUINTILES
S4
S5 P
RESAMPLING
PPLAT
OF
* : mortality rate adjusted for age, APACHE/SAPS risk, arterial-pH, P/F ratio, and Trial (Cox Proportional Hazard Regression)
*: mortality-rate
for age, APACHE/SAPS
risk,after
arterial-pH,
P/F ratio, and study-trial
Figure S5: Survival
impact ofadjusted
tidal volume,
before and
lung-sizing
(Cox Proportional Hazards Regression)
Using double stratification procedures (like in Figure 1, main manuscript), we partitioned our dataset into five distinct sub-samples (each one
with approximately 600 patients with ARDS), and calculated the relative risk for each sub-sample in comparison with the average risk of the
combined population. Patients are the same as those included in the combined analysis of Table 1.
In the upper scatter/error-bar diagrams (open triangles) we show the average values for plateau pressures across quintiles of tidal volumes
(left) or P (right). In the middle scatter/error-bar diagrams (black squares) we show the average values for tidal volume (left panel, VT scaled
to predicted body-weight, PBW) and for driving-pressure (right panel, VT scaled in proportion to respiratory-system compliance, so
representing P), found in each quintile. The error bars represent one standard-deviation. Note that each resampling (D and E) produced subsamples with comparable mean values for plateau pressures, but very distinct values for tidal volume or P.
At the bottom, we show the respective relative-risk calculated for each sub-sample after multivariate adjustment (at patient-level) for covariates
1-5 specified in Model 1 (Table 1). Error bars represent 95% confidence intervals. A relative-risk of 1 represents the average risk of the pooled
population, which presented an adjusted survival at 60-day of 68%.
Note that reductions in tidal volume per se had no impact on mortality risks, whereas reductions of a re-scaled tidal volume (so representing
P) were associated with a marked risk reduction. Note the mathematical equivalence: P = ( VT / CRS) = VT normalized to CRS
31
settings
32
Figure S6:
Figure
2a :
Figure-S4:
95
( adjusted* )
100
90
P < 0.001
85
80
75
stratification:
(N)
P MEDIAN
14 cmH2O
( 989 )
P >> MEDIAN
14 cmH2O
( 756 )
PPLAT MEDIAN
25 cmH2O
( 955 )
( 790 )
100
95
( adjusted* )
70
90
P = 0.98
85
80
75
70
100
95
( adjusted* )
90
85
P = 0.30
80
VT > MEDIAN
6 mL / kg
( 867 )
VT < MEDIAN
6 mL / kg
( 878 )
75
70
10
20
30
40
50
60
Figure
S6: Survival
patients P/F
under
"protective"
ventilator settings
*: adjusted for age,
APACHE/SAPS
risk, in
arterial-pH,
ratio,
and study-trial
Survival curves were obtained after multivariate adjustment at patient level (Cox Proportional Hazards model)
for covariates 1-5 specified in Model-1 (Table 1, main manuscript, and Table S8). For each survival plot, the
selected sub-sample (N=1745) was stratified according to the median values of P, plateau pressure and VT,
respectively (median values = 13 cmH2O, 26 cmH2O and 6 mL/kg PBW, respectively, from top to bottom
plots), producing two strata with similar number of patients. Treating ventilator variables as continuous variables
did not improve the association of tidal volume or plateau pressure with survival, but did improve the
association of P with survival (P<0.001).
Note that lower tidal volumes were associated with a paradoxical trend of higher mortality.
33
34
Figure S7:
Figure 2d:
Figure-S5:
1.4
1.0
2.2
( adjusted * )
P <<P0.0001
0.001
< 0.0001
1.8
P < 0.0001
2.2
( adjusted * )
*)
( adjusted
for Barotrauma
Odds-ratio
PP== 0.87
0.87
1.8
1.4
1.0
0.6
0.6
44
88
12
12
16
16
Driving-pressure
20
24
20
24
(P, cmH2O)
28
28
44
6
6
88
Tidal-Volume (V , mL/kg.PBW)
*: pre-adjusted for age, APACHE/SAPS risk, arterial-pH, P/F ratio and study-trial
(VT and P co-participating in model-1)
**::pre-adjusted
adjusted for
risk,
arterial-pH,
P/F and
ratio,
and Trial
forage,
age, APACHE/SAPS
APACHE/SAPS risk,
arterial-pH,
P/F ratio
study-trial
(multivariate
logistic
regression
wherewhere
both P
and V
in Model-1) in Model-1)
(multivariate
logistic
regression
both
TPco-participate
and VT co-participate
Figure S7: Odds-ratio for Barotrauma during the first 28 days after randomization
Barotrauma was strictly defined as pneumothorax requiring chest tube drainage (with 313 events, or 9% of
the sample). The odds-ratio for each quintile was calculated in relation to the average risk of the combined
population (assumed to be 1.00). The mean odds and 95% confidence intervals (error bars, enclosing the
gray zone) for each quintile were calculated after multivariate adjustment at patient level (Logistic regression
model) for covariates 1-5 specified in Model-1 (Table 1, main manuscript). After including tidal volume and
P in the multivariate model, the adjusted odds-ratio for progressive quintiles of both variables were
calculated (each quintile had approximately 600 patients). The number of percentiles was chosen in order to
have at least 40 events per percentile, guarantying reliable confidence intervals. P-values indicate the overall
differences in risk across quintiles (as categorical variable).
.
10
10
12
12
35
Figure S8:
Mediation in the Lower vs. Higher VT-trials:
Tested mediator: P-changes driven by randomization
Figure S9:
Mediation in the Higher vs. Lower PEEP-trials:
Tested mediator: P-changes driven by randomization
36
P-changes
driven by randomization
Solid arrows in the path diagram represent significant association between variables, with left to right direction representing an independent
model: High vs. Low V trials
Figure 3a
: Mediational
to dependent
relationship.
Red dashed arrows representT non-significant effects.
Tested mediator:
Step 1:
P-changes
driven by randomization
Hazard = 0.60
Randomization
Survival effect
(0.48 0.75; P < 0.001 )
Baseline disease
covariates*
Step 2:
Hazard = 0.62
lower P
Survival effect
( 0.52 0.74 ; P < 0.001 )
ElastanceRS
adjustment
Steps 3 & 4:
-8.4 cmH2O
Randomization
Mediatedproportion:
74%
lower P
P < 0.001
Baseline disease
covariates*
P = 0.004
Survival effect
ElastanceRS
adjustment
Baseline disease
covariates*
Subsequently, we jointly calculated the influence of the mediator on survival, after accounting for baseline risk factors, baselineelastanceRS, and the direct effects of randomization. This last step shows that reductions in P mediated most (75%, P = 0.004) of the
original effect of randomization and, consequently, randomization is no longer associated with survival in an independent manner
(characterizing complete mediation).
Implicitly, this last step with significant ACME (average causal mediation effect) also suggests that variations in P had an independent
impact on survival: i.e. patients exhibiting larger reductions in P obtained a survival benefit that exceeded the average benefits found in
the lower-VT arm.
37
P-changes
driven by randomization
Solid arrows in the path diagram represent significant association between variables, with left to right direction representing an independent
model: arrows
High vs.
Low PEEP
trials effects.
Figure 3c
: Mediational
to dependent
relationship.
Red dashed
represent
non-significant
Tested mediator:
Step 1:
P-changes
driven by randomization
Hazard = 0.83
Randomization
Survival effect
(0.72 0.97; P = 0.02 )
Baseline disease
covariates*
Step 2:
Hazard = 0.57
lower P
Survival effect
( 0.42 0.72; P < 0.001 )
ElastanceRS
adjustment
Steps 3 & 4:
-1.2 cmH2O
Randomization
Mediatedproportion:
45%
lower P
P < 0.001
Baseline disease
covariates*
P = 0.001
Survival effect
ElastanceRS
adjustment
Baseline disease
covariates*
This last step, with significant ACME (average causal mediation effect) also demonstrates that P had a survival effect that exceeds the
main effect of randomization: i.e. patients exhibiting accentuated reductions in P obtained a survival benefit that exceeded the average
benefit found in the respective higher-PEEP arm.
38
Additional details:
The reliability of a randomized clinical trial resides in the lack of correlation between treatment
and baseline condition of patients. This is an essential, well-accepted pre-requisite for
accepting causality implication. The lack of correlation arises from an exogenous, random
process for treatment selection. In case we find a significant correlation between treatment and
outcomes, we can suggest that the benefits are directly caused by treatment, and not by the
circumstantial fact that we applied the treatment in patients with better baseline condition.
Similarly, to suggest that P was more than a risk predictor, working as a mediator of survival,
(i.e. to the extra-survival related to randomized treatment) we had to be sure that the variability
of P observed in our samples were not correlated with baseline disease (i.e. independent
from baseline elastance). Ideally, P variations should be mostly related to ventilator strategy,
implemented according to a randomized process. However, since P is mathematically
correlated with baseline elastance of respiratory system (baseline-elastanceRS), this
independency of P was hardly true and the solution was to first remove (filter out) the
component of P correlated with baseline mechanics, later applying the mediation analysis on
the residual P component.
In multivariate regression analysis, this removal or subtraction of the confounding component
(P-component related to baseline disease) is equivalent to the pre-adjustment of a regression
model by baseline-elastanceRS, if adjusting the same model in which P is simultaneously
tested as an independent explanatory variable. A significant correlation between P and
outcomes would indicate the presence of a residual and independent (orthogonal) source of
variability in P, uncorrelated with baseline disease, which is also affecting outcomes.
Accordingly, in the first logical steps of our mediation analysis (step-2: the check if mediator
changes, theoretically assumed as beneficial, correlated with better survival), we explicitly
tested this hypothesis (Figure S8 and S9). After pre-adjustment of survival Model-1 for
39
40
that variations in the mediator explain the whole effect of the randomized treatment (so called
complete mediation), but this is not a pre-requisite to demonstrate mediation.
The mediation analysis must be performed according to a sequence of statistical tests and
logical steps. We here used the sequence described by Kraemer and Shrout, in accordance
with the MacArthur approach 20-22.
The initial step was the analysis of the main effect of randomization, traditionally called as the
direct effect. This was calculated through a Cox survival analysis, using trial as a random effect.
In the original reports, only two1,5 out of the nine trials in our combined sample reported a
statistically significant impact of protective strategies. In a recent meta-analysis24 using a
population that had large overlap with the population of our derivation cohort, investigators
have found a marginal benefit of lower VT strategies. And another patient-level meta-analysis25,
considering only the higher versus lower PEEP trials, showed a benefit of the higher PEEP
strategy, although restricted to a subgroup of patients with more severe hypoxemia.
Differently, our combined analysis showed a more consistent efficacy of protective strategies,
either considering the lower-VT trials separately, or the higher-PEEP trials separately. This
difference was essentially caused by the multivariate adjustment at the patient level, according
to survival Model-1. For instance, the crude relative-risk of lower-VT strategies was 0.82 (95%CI
= 0.671.00) before adjustment, and 0.60 (95%CI = 0.480.75) after adjustment. Similarly,
the crude relative-risk of higher-PEEP strategies was 0.92 (95%CI = 0.791.06) before
adjustment, but 0.83 (95%CI = 0.720.97) after adjustment.
When using the R software for mediation analysis, there is an output called total effect,
estimated by simulation according to the principles described by Imai et al15-18. This output must
provide an equivalent result (for the main effect of randomizarion), although slightly different
41
because of the simulations and different estimation methods. After running 5000 simulations,
the results fitted very well with the estimates provided above:
Relative-risk (Hazard)
P-Value
0.60 (0.480.74)
P < 0.001
0.82 (0.710.96)
P = 0.01
The second logical step was the check described above (page 32, when discussing the
required adjustment for baseline-elastanceRS), where we tested if mediator changes,
theoretically assumed as beneficial, correlated with better survival. At this step, we added the
pre-adjustment of baseline-elastanceRS in the model, making sure that our analysis was not
biased by confounding effects of the severity of underlying lung disease (potentially inflating the
association between P and survival). Among our tested mediator-candidates, PEEP and
plateau pressures failed at this second step within the higher-PEEP trials. After accounting for
baseline elastance, both variables were not significantly associated with survival.
The next two steps involve the analysis of inclusion of the effects of superimposed-P (i.e.
residual variations in P after subtracting the P variations correlated with baseline lung
disease) in a model where the direct effects of randomization are already taken into account:
42
By using the notation indicated in the diagram above, this analysis involves the demonstration
of 2 separate steps (steps 3 and 4 , Figures S8-S9, above)13,20-22 :
Step-3: to show that superimposed-P was significantly correlated to treatment group
assignment. Accordingly, the coefficient "A" produced by the multivariate regression of the
superimposed-P variable over randomization must be significantly different from zero. The
signal of coefficient A is an important issue in this mediation analysis, since we want to prove
that reductions in P cause improved survival. Thus, it has to be negative. As a
counterexample, plateau-pressures did not pass this logical step for the PEEP-trials, since
randomization caused higher plateau-pressure (,i.e. a positive coefficient A, P < 0.001), but an
improved survival.
Step-4: to show that superimposed-P carried significant survival impact (with a significant
coefficient "B") when included in a multivariate model where treatment group was also included
(as well as baseline-elastanceRS). This procedure is equivalent to the Sobel test, or, when using
the R software for mediation, equivalent to the significance test for the ACME (average causal
mediation effect). The ACME considers the combined effects of steps 3 and 4: both steps have
to present some minimum effect-size to guarantee a significant path through the mediator2.
In the R package for Mediation Analysis, two models are fit, one modeling the effects of randomization
on the mediator, and a second one jointly modeling the effects of randomization (directly) and mediator
(indirectly) on outcomes. Using Monte-Carlo simulations, a mediation proportion is estimated, indicating
43
Whenever the ACME has an important effect, the coefficient "C" (representing the direct effect
of randomization, after multivariate adjustment) becomes weaker after accounting for the
mediating effect (indirect path "A*B"). Ideally, in order to demonstrate complete mediation, "C"
should be virtually zero (non-significant), with a mediation proportion ~100%.
For testing all the steps above, we pre-adjusted our mediational model with the covariates
elected in Model-1 (Table 1, main text), also including the baseline-elastanceRS variable. We
observed that all those covariates were non-specific predictors of survival, uncorrelated with
treatment group. We repeated the stepwise checks for all mediation-candidates, using only one
mediator each time.
how much of the whole risk-reduction in the treatment arm could be explained by the indirect path in
which randomization drives a change in the mediator, which then affects the outcome.
8-10
The rationale used by this package is the same described by Imai et al : suppose that randomized
treatment is denoted as a z=0 (control) or z=1 (treatment), and the mediating variable for a patient is
denoted as M(0) and M(1) if they got treatment 0 or 1. Further suppose that the outcome for a patient
with mediating variable M who got treatment 0 is T(M,0) and if they got treatment 1, T(M,1). Then the
average mediated effect is the average of (T(M(1),1) -T(M(0),1) and T(M(1),0) -T(M(0),0). In other words
it is the effect we would see if the mediating variable changed but treatment did not. This effects needs to
be estimated in a model including the treatment group, the mediating variable and all confounding
variables the affect both the mediating variable and the outcome.
44
Figure S10:
Mediation in the Lower vs. Higher VT-trials:
Tested mediator: VT-changes driven by randomization
Figure S11:
Mediation in the Higher vs. Lower PEEP-trials:
Tested mediator: PEEP-changes driven by randomization
Additional details:
Except for P, no other mediation-candidate consistently passed through the stepwise
mediation tests. Of note, tidal volume was not a significant mediator in the lower-VT trials
(P=0.67, ACME, Figure S10), and PEEP was not a significant mediator in the higher-PEEP
trials (P=0.50, ACME; Figure S11).
45
VT-changes
driven by randomization
Solid arrows in the path diagram represent significant association between variables, with left to right direction representing an independent
model: arrows
High vs.
Low VT non-significant
trials
Figure 3b
: Mediational
to dependent
relationship.
Red dashed
represent
effects.
Tested mediator:
Step 1:
VT -changes
driven by randomization
Hazard = 0.60
Randomization
Survival effect
(0.48 0.75; P < 0.001 )
Baseline disease
covariates*
Step 2:
Hazard = 0.78
Survival effect
lower VT
(-1 STD = - 2.8 mL/kg)
ElastanceRS
adjustment
Steps 3 & 4:
-5.1 mL/kg
Mediatedproportion:
0%
lower VT
P < 0.001
Hazard = 0.59
Randomization
Survival effect
Baseline disease
covariates*
Note that randomization keeps an independent predictive effect at this last step, whereas VT does not. This suggests that the predictive
information provided by treatment group assignment (representing a bundle of intention to treat interventions) exceeds the information
provided by individual levels of VT.
Thus, tidal-volume cannot be considered as an independent mediator of the effects of randomization.
46
SolidFigure
arrows in
diagram model:
represent
significant
variables, with left to right direction representing an independent
Mediational
High
vs. Low association
PEEP trialsbetween
3cthe: path
to dependent relationship.
dashed
arrows
represent
TestedRed
mediator:
PEEP
-changes
driven non-significant
by randomizationeffects.
Step 1:
Hazard = 0.83
Randomization
Survival effect
(0.72 0.97; P = 0.02 )
Baseline disease
covariates*
Step 2:
Hazard = 0.95
Survival effect
higher PEEP
ElastanceRS
adjustment
Steps 3 & 4:
+6.1 cmH2O
Baseline disease
covariates*
Mediatedproportion:
0%
higher PEEP
P < 0.001
Hazard = 0.82
Randomization
Survival effect
Baseline disease
covariates*
Note that randomization keeps an independent predictive effect at this last step, whereas PEEP does not. This suggests that the
predictive information provided by treatment group assignment (representing a bundle of intention to treat interventions) exceeds the
information provided by individual levels of PEEP.
Thus, PEEP per se cannot be considered as an independent mediator of the effects of randomization. Other strategies included in
the bundle, or other intermediate variables affected by high-PEEP strategy are likely more important.
47
Table S9:
Mediation effects of P within study-trials (intra-trial mediation effects)
Additional details:
The variable Trial was entered as a random-effect in both mediation models: the mixed-effects
linear regression modeling the effects of randomization on the mediator, and the mixed-effects
Cox regression modeling the effects of both, mediator and randomization on survival. The R
package for mediation was designed to perform multi-level mediation analysis, where
individuals may be correlated within groups (trials), and the different trials may have different
randomization processes. Using a conservative approach, we also considered a moderated
mediation analysis, in which the randomization process could be moderated by the trial
variable, i.e. each study could have a different effect of randomization (direct effect), a different
effect of P (mediating effect), and a different mediation proportion. In fact, this analysis
allowed us to test the intra-trial mediation effect of P, shown in Table S9.
As observed, the ACME had the same consistent trend in all of the nine studies, presenting a
significant value in many of the individual studies. These findings support our pooled analysis
presented in Figure S8 and S9. They also demonstrate that P mediates intra-trials and intertrials effects, i.e. differences between the two arms in each study, and differences in the net
effects of randomization across the studies, with some studies presenting larger benefits of
randomization than others.
48
Table S9 (website):
Total effect
(Hazard)
ACME
(Hazard)
Amato et al.1
0.12
0.26
0.07
Stewart et al.2
0.63
0.72
0.14
Brochard et al.3
0.69
0.60
0.03
Brower et al.4
1.16
0.86
0.28
ARDSnetTV5
0.62
0.78
0.10
VT-trials combined
0.59
0.73
0.01
ARDSnetPEEP6
0.90
0.83
0.001
EXPRESS7
0.76
0.91
0.004
LOVS8
0.89
0.96
0.06
Talmor et al.9
0.52
0.97
0.46
PEEP-trials combined
0.83
0.92
< 0.001
Trial:
P-value
(for ACME)
Note that the ACME had the same consistent trend in all of the nine studies, presenting a significant value in many of the individual studies. Thus,
P was always responsible for part of the observed effects of randomization (shown as the total-effect). These findings support the pooled
analysis presented in Figures S8-S9. They also suggest that P mediates intra-trials and inter-trials effects, i.e. differences between the two arms
in each study, and differences in the net effects of randomization across the studies, with some studies presenting larger benefits of randomization
than others.
49
50
deviations from the population-sample mean) within baseline or ventilation variables were
assessed. If they were consistent with values on the previous or following days (i.e. within the
range of the individual mean along the time 3 STD), they were retained; otherwise they were
either within-patient-interpolated (when possible) or discarded. If there was more than one
value per day, we assumed that the non-outlier value was representative for that day.
In Tables S1-S2 and Figure S1 we present some descriptive statistics for these trials, after the
incompatibility analysis. In Table S3, we tested the univariate association with survival for each
of those screened covariates.
Finally, we tested the impact of excluding additional 87 subjects whose plateau-pressures were
below 10 cmH2O or Ps were below 5 cmH2O or PEEP values were below 5 cmH2O or tidalcompliance was above 1.25 mL/kg/cmH2O, during the first day after randomization. We
reasoned that such values might indicate errors during data collection, or patients with mild
disease or low risk exposure who might decrease the sensitivity of our analysis. The mortality
rate of these excluded patients was lower (25.3%) than the average mortality of the remaining
population (35.3%; P = 0.07 for the comparison of mortality-rates between excluded vs.
remaining patients).
The performance of Models 1-5 was rechecked within this more restricted sample of 3466
patients. The relative-risk associated with one standard-deviation increment in driving-pressure
or other covariates minimally changed (maximum change of 1% in the relative-risks for all
covariates). After observing this lack of benefit of a more restricted sample, we ultimately
decided to present our results without this filter (as shown in Table 1). The slightly lower
number of patients presented in Table 1 was caused by missing data (see next session).
Whereas patients in the validation sample had their tidal-volumes adjusted according to the
ideal body weight (IBW), most patients in the hypothesis-generation sample had their tidal-
51
volumes adjusted according to actual body weight. The physiological bias generated by this
procedure, in case of prevalent obesity, would be that of a systematically lower tidalcompliance (affecting both arms) in those studies using actual body weight. Since we found
similar mean values for normalized compliance in both samples (P = 0.23), without significant
differences in other markers of baseline lung disease, we concluded that such bias was unlikely
and that the use of actual body weight in these particular patients had the same physiological
consequences as the use of ideal body weight. In sum, 92% of patients in our combined
sample had their tidal-volumes adjusted according to ideal body weight, but we kept the term
IBW for all.
52
53
54
References
1. Amato MBP, Barbas CSV, Medeiros DM, et al. Effect of a Protective-Ventilation Strategy on
Mortality in the Acute Respiratory Distress Syndrome. N Engl J Med 1998;338:347-54.
2. Stewart TE, Meade MO, Cook DJ, et al. Evaluation of a ventilation strategy to prevent barotrauma
in patients at high risk for acute respiratory distress syndrome. Pressure- and Volume-Limited
Ventilation Strategy Group. N Engl J Med 1998;338:355-61.
3. Brochard L, Roudot-Thoraval F, Roupie E, et al. Tidal volume reduction for prevention of
ventilator-induced lung injury in acute respiratory distress syndrome. The Multicenter Trail Group
on Tidal Volume reduction in ARDS. Am J Respir Crit Care Med 1998;158:1831-8.
4. Brower RG, Shanholtz CB, Fessler HE, et al. Prospective, randomized, controlled clinical trial
comparing traditional versus reduced tidal volume ventilation in acute respiratory distress syndrome
patients. Crit Care Med 1999;27:1492-8.
5. ARDSNet. Ventilation with lower tidal volumes as compared with traditional tidal volumes for
acute lung injury and the acute respiratory distress syndrome. N Engl J Med 2000;342:1301-8.
6. Brower RG, Lanken PN, MacIntyre N, et al. Higher versus lower positive end-expiratory pressures
in patients with the acute respiratory distress syndrome. N Engl J Med 2004;351:327-36.
7. Mercat A, Richard J-CM, Vielle B, et al. Positive End-Expiratory Pressure Setting in Adults With
Acute Lung Injury and Acute Respiratory Distress Syndrome: A Randomized Controlled Trial.
JAMA 2008;299:646-55.
8. Meade MO, Cook DJ, Guyatt GH, et al. Ventilation Strategy Using Low Tidal Volumes,
Recruitment Maneuvers, and High Positive End-Expiratory Pressure for Acute Lung Injury and
Acute Respiratory Distress Syndrome: A Randomized Controlled Trial. JAMA 2008;299:637-45.
9. Talmor D, Sarge T, Malhotra A, et al. Mechanical ventilation guided by esophageal pressure in
acute lung injury. N Engl J Med 2008;359:2095-104.
55
10. Schoenfeld D. Survival methods, including those using competing risk analysis, are not appropriate
for intensive care unit outcome studies. Crit Care 2006;10:103.
11. Zhou Z, Rahme E, Abrahamowicz M, Pilote L. Survival bias associated with time-to-treatment
initiation in drug effectiveness evaluation: a comparison of methods. Am J Epidemiol
2005;162:1016-23.
12. Kasal J, Jovanovic Z, Clermont G, et al. Comparison of Cox and Gray's survival models in severe
sepsis. Crit Care Med 2004;32:700-7.
13. Ditlevsen S, Christensen U, Lynch J, Damsgaard MT, Keiding N. The mediation proportion: a
structural equation approach for estimating the proportion of exposure effect on outcome explained
by an intermediate variable. Epidemiology 2005;16:114-20.
14. Ditlevsen S, Keiding N, Christensen U, Damsgaard MT, Lynch J. Mediation proportion.
Epidemiology 2005;16:592.
15. Imai K, Jo B, Stuart EA. Using Potential Outcomes to Understand Causal Mediation Analysis:
Comment on. Multivariate Behav Res 2011;46:861-73.
16. Imai K, Keele L, Tingley D. A general approach to causal mediation analysis. Psychol Methods
2010;15:309-34.
17. Imai K, Keele L, Tingley D, Yamamoto T. Causal Mediation Analysis using R. In: Vinod HD, ed.
Advances in Social Science Research using R: Springer New York; 2010:129-54.
18. Imai K, Keele L, Tingley D, Yamamoto T. Unpacking the Black Box of Causality: Learning about
Causal Mechanisms from Experimental and Observational Studies. American Political Science
Review 2011;105:765-89.
19. Preacher KJ, Zyphur MJ, Zhang Z. A general multilevel SEM framework for assessing multilevel
mediation. Psychological Methods 2010;15:209-33.
20. Shrout PE, Bolger N. Mediation in experimental and nonexperimental studies: new procedures and
recommendations. Psychol Methods 2002;7:422-45.
21. Kraemer HC, Kiernan M, Essex M, Kupfer DJ. How and why criteria defining moderators and
mediators differ between the Baron & Kenny and MacArthur approaches. Health Psychol
2008;27:S101-8.
56
22. Kraemer HC, Stice E, Kazdin A, Offord D, Kupfer D. How do risk factors work together?
Mediators, moderators, and independent, overlapping, and proxy risk factors. Am J Psychiatry
2001;158:848-56.
23. Kraemer HC, Wilson GT, Fairburn CG, Agras WS. Mediators and moderators of treatment effects
in randomized clinical trials. Arch Gen Psychiatry 2002;59:877-83.
24. Burns KE, Adhikari NK, Slutsky AS, et al. Pressure and volume limited ventilation for the
ventilatory management of patients with acute lung injury: a systematic review and meta-analysis.
PLoS One 2011;6:e14623.
25. Briel M, Meade M, Mercat A, et al. Higher vs lower positive end-expiratory pressure in patients
with acute lung injury and acute respiratory distress syndrome: systematic review and meta-analysis.
JAMA 2010;303:865-73.