Does transfer to intensive care units reduce mortality for deteriorating ward patients? Estimating Person-centered Treatment (PeT) effects using instrumental variables
1) The study uses instrumental variable methods to examine the effect of transfer to intensive care units (ICU) on mortality for deteriorating ward patients using data from the UK.
2) The instrumental variable is ICU bed availability, which is hypothesized to directly affect the probability of ICU transfer but be independent of patient mortality.
3) Propensity score matching is also used to increase the strength of the instrumental variable by matching patients exposed to many versus few available ICU beds who are similar on observed characteristics.
Statistical Methods for Removing Selection Bias In Observational Studies
Similar to Does transfer to intensive care units reduce mortality for deteriorating ward patients? Estimating Person-centered Treatment (PeT) effects using instrumental variables
Similar to Does transfer to intensive care units reduce mortality for deteriorating ward patients? Estimating Person-centered Treatment (PeT) effects using instrumental variables (20)
Call Girls Jabalpur Just Call 8250077686 Top Class Call Girl Service Available
Does transfer to intensive care units reduce mortality for deteriorating ward patients? Estimating Person-centered Treatment (PeT) effects using instrumental variables
1. Does transfer to intensive care units
reduce mortality for deteriorating ward
patients?
Estimating Person-centered Treatment (PeT) Effects
Using Instrumental Variables
Anirban Basu
University of Washington, Seattle
(Joint work with Stephen O’Neill, National University of Ireland
Galway and Richard Grieve, London School of Hygiene and Tropical
Medicine)
2. Treatment Evaluations
• Literature on treatments effects focused on informing
population level or policy-level decisions.
o Mean effects
o Distributional effects
o Impacts are viewed as informing a social decision maker to
help choose across alternative options
o Discussions will focus on evaluating policies/treatment in use
• Heterogeneity: Optimal Individual choices may vary
from socially optimal choices, based on ex-post
realizations.
1
3. Literature on Distribution of
Treatment Effects
• Imbens and Rubin (1997) and Abadie (2002, 2003)
• Carniero and Lee (2009)
• Heckman and Honoré 1990, Heckman and Smith
1993, Heckman et al. 1997.
• Factor structure models have been used to
establish the joint distribution of potential outcomes
(Aakvik et al. 1999, Carniero et al. 2003).
• Importance of distribution of effects well established
in the literature (Abbring and Heckman, 2007)
2
4. Healthcare Setting
• Dilemma between social versus individual choices,
perhaps, the most stark
• In traditional clinical outcomes research, the focus
has always been on finding average effects either
through large clinical trials or observational datasets.
• Estimating treatment effect heterogeneity has
mostly been relegated to post-hoc analysis, rather
than becoming the central goal of the analysis.
• Growing recognition of the importance of nuanced
and possibly individualized estimates of treatment
effects (Basu 2009, 2011, Basu et al. 2011).
3
5. Observational Data Setting
• Observational data is a valuable resource.
• Selection bias in observational studies is a potential
problem – primary reason for relegation to second-
tier status in evidentiary standards.
• However, growing interests due to large scale
investments in Electronic Health Records.
• Methods for casual inference in such data of high
demand.
4
6. Outline
• Conceptual Identification of treatment effects using
instrumental variable methods
• PeT effects
• PeT effects estimation
• Empirical Example
5
10. Potential Outcomes
• Outcomes to be realized for an individual if that
individual chooses a certain treatment
• If an individual chooses Surgery: YS
• If that same individual chooses Watchful Waiting: YW
• Treatment effect for that individual: YS – YW
• But we do not observed both YS and YW for the
same individual
9
16. IV effect
• With two levels of an IV (e.g. short distance vs long
distance):
o IV effect produces the causal effect for a group of
individuals, whose treatment choices are altered because
of the change of the IV levels
o Local Average Treatment effect (Angrist and Rubin 1996)
o Challenges:
• Who are these people (remember we don’t observed
severity in the data)?
• How generalizable are there effects to other?
o If treatment effects vary over severity, LATE is NOT a very
useful result.
15
17. Where is LATE useful?
• Where the binary instrument represents a policy
change..
• E.g. Oregon lottery as a instruments for Medicaid
receipt
• LATE = the average effects on outcomes on those
who choose to get Medicaid if Medicaid eligibility
were to be expanded
16
18. Interpretation of IV effect
• If treatment effects vary over unobserved
confounders:
o “essential heterogeneity” (Heckman & Vytlacil, 1999)
o IV effect has generalizability issues
o IV effect varies with specific IVs used
o Alternative methods required to estimate population mean
effect parameters : ATE, TT or TUT
o Heckman & colleagues propose Local instrumental
variable (LIV) methods
• Help to identify and estimate Marginal Treatment
Effects (MTEs)
17
19. Choice Model
• To solve this issue of interpreting IV effect, an explicit
choice model is required (Heckman1997; Heckman
and Vytlacil, 1999)
• Impose a discrete choice model:
D* = g(Distance) + h(Severity) D* ~ (-∞, ∞)
D = 1 if D* > 0
• So, distance and severity kind of balance each
other in terms of the propensity of an individual to
choose surgery.
• If D* = 0, a “marginal individual” who chooses WW.
18
20. Identification strategy for a
marginal treatment effect
• Find marginal individuals
o HOW: those who would change treatment due to small
perturbation in the instrumental variable
o Since IV is independent of all confounders, this small
perturbation does not change the level of confounders
• Estimate treatment effects for marginal individuals –
“Marginal treatment effect (MTE)”
o HOW: compare outcomes between two groups of
individuals only separated by small perturbation of the IV
• Aggregate MTEs across all margins
o HOW: Simple average across all margins will produce
Average Treatment Effect
19
24. LIV Estimator
1st stage
2nd stage
UD = Propensity for treatment selection based on unobserved
confounders
0 1 2( )logit D X Z
0 1 2 3
ˆ ˆ ˆ( | , ( , )) ( ( ; ))E Y X p x z g X x p K p
ˆ (1 )
ˆ ˆ( | , ( , ))
( , )
ˆ
D
D D
P p u
dE Y X P x z
MTE x U u
dP
23
25. Extension to Person‐centered
Treatment (PeT) effect
Marginal patients:
D* = g(Distance) + h(Severity) = 0
Therefore, the severity level of a marginal patients:
Severity = h-1(-g(Distance))
Standard calculus can allow one to characterize the
level of severity as a function of the instrument level,
where the marginal patient is identified.
Why do we care?
Basu A. Journal of Applied Econometrics (2014)
24
26. Extension to Person‐centered
Treatment (PeT) effect
Why do we care?
Go back to data.
Take patient 1
• Observe Z, D
• Calculate the range of Severities for which D satisfies
D* = g(Z) + h(Severity)
• Average the MTE over only those range of severities
for patient 1
Obtain a person-centered treatment effect for patient 1!
Basu A. Journal of Applied Econometrics (2014)
25
27. Extension to Person‐centered
Treatment (PeT) effect
• PeT effect are conditional treatment effect that:
o Not only conditions on observed risk factors
o Averages over the conditional distribution of unobserved risk factors,
conditioned on choices.
Basu A. Journal of Applied Econometrics (2014)
26
28. Extension to Person‐centered
Treatment (PeT) effect
• PeT effect are conditional treatment effect that:
o Not only conditions on observed risk factors
o Averages over the conditional distribution of unobserved risk factors,
conditioned on choices.
• They help to comprehend individual-level
treatment effect heterogeneity better than CATEs.
• They are better indicators for the degree of self-
selection than CATE.
• They can explain a larger fraction of the individual-
level variability in treatment effects than the CATEs.
Basu A. Journal of Applied Econometrics (2014)
27
29. Estimation of PeT effects
• Basu A. Person-Centered Treatment (PeT) effects using
instrumental variables: An application to evaluating prostate
cancer treatments. Journal of Applied Econometrics 2014;
29:671-691.
• Basu A. Person-centered treatment (PeT) effects:
Individualized treatment effects with instrumental variables.
Stata Journal In Press.
• -petiv- command in STATA
28
30. PeT Estimation Algorithm
• Numerical integration: For each individual i:
o Draw 1000 deviates u~Uniform[min(P), max(P)]
o Compute and evaluate it by replacing P with
each value of u.
o Compute D* = Φ-1(P)+Φ-1(1-u) also generating 1000 values
for each individual.
o Compute PeT by averaging over values of u for
which (D* > 0) if D=1, otherwise, by averaging over values
of u for which (D* ≤ 0) if D=0.
• Estimated PeT effects provide us with individualized effects of
treatment effects. Mean treatment effect parameters were
also computed. Averaging PeTs over all observation gave ATE.
Averaging PeTs over over D=1 or D=0 gave us TT and TUT
respectively.
ˆ ˆ(.) /dg dp
ˆ ˆ(.) /dg dp
29
32. Background
• Adult Intensive Care Units (ICU) costly and scarce
resource
o Supply usually lags demand
• No RCT evidence
• Observational study evidence
o Do not deal with the endogeneity of transfer
o Do not recognizing heterogeneity in returns from transfer
• Transfers to ICU typically relies on clinical judgement
o Not perfect proxy for reliable and causal evidence
31
33. Our Study
• Exploit natural variation in ICU transfer according to
ICU bed availability for deteriorating ward patients in
the UK
• The (SPOT)light Study (N = 15,158)
o Prospective cohort study of the deteriorating ward patients
referred for assessment for ICU transfer
o Hospitals were eligible for inclusion if they participated in the
ICNARC Case Mix Programme
o Patients recruited between Nov 1, 2010 - Dec 31, 2011 from 49
UK NHS hospitals
o A variety of exclusion conditions were applied to identify
deteriorating ward patients who are equipoised to be
transferred to ICU
32
34. Data
• Database locked on Sep 2012
• Detailed demographic and physiological data
(collected from the time of ward assessment)
• Date of death (NHS Information Service)
• Critical care provision, ICU bed availability, and
hospital characteristics (CMP and Hospital Episode
Statistics (HES))
• Number of available ICU beds =
(Maximum number of beds reported to ICNARC) ―
(Number of actively treated patients occupying those
beds at the time of ward assessment)
33
35. Data
• Primary Outcome: Death 7 days post assessment
• Secondary Outcomes: Death within 28 and 90 days
• Exposure: ICU transfer vs care on general wards
• Baseline covariates: Age, diagnosis of sepsis, peri-
arrest, dependency at ward assessment and
recommended level of care post assessment (4 levels)
and three physiology measures
o National Early Warning Score (NEWS) : whether respiratory rate,
oxygen saturations, temperature, systolic blood pressure, pulse
rate, a level of consciousness vary from the norm,
o the Sequential Organ Failure Assessment (SOFA), and
o the ICNARC physiology score
34
36. IV
• IV = NBA: Vary across hospital and over time
• Key Assumptions:
o NBA at ward patient’s assessment directly affects one’s
probability of transfer to ICU
o NBA unconditionally independent of mortality of patients
35
38. IV and Near‐Far Matching
• IV = NBA: Vary across hospital and over time
• Key Assumptions:
o NBA at ward patient’s assessment directly affects one’s
probability of transfer to ICU
o NBA unconditionally independent of mortality of patients
• Increase strength of instrument – Near –far matching
o Match similar patients with ‘many’ versus ‘few’ NBA
o Similarity assessed on: age, gender, NEWS SOFA and ICNARC
physiology scores, CCMDS level at assessment, and timing (out
of hours, winter, and weekend or not)
o Keep matched pairs with at least difference of 3 of more NBA
Baiocchi et al. JASA 2010; Small and Rosenbaum JASA 2008; Keele and Morgan AAS 2016
37
42. PeT effects
• First Stage
Probit(ICU) = H(X, Z)
• Second Stage Control Function: compute PeT Effects using –petiv-
Probit(Death) = G(X, X*P(X,Z), K(P(X,Z)))
Basu JAE 2014; Basu and Gore Med Care 2015
41
44. References
Basu A, Meltzer D. Value of information on preference heterogeneity and individualized
care. Medical Decision Making 2007; 27(2):112-127.
Basu A, Heckman J, Navarro-Lozano S, et al. Use of instrumental variables in the presence
of heterogeneity and self-selection: An application to treatments of breast cancer
patients. Health Econ 2007; 16(11): 1133 -1157.
Basu A. 2009. Individualization at the heart of comparative effectiveness research: The
time for i-CER has come. Medical Decision Making, 29(6): N9-N11.
Basu A, Jena AB, Philipson TJ. The impact of comparative effectiveness research on
health and health care spending. J Health Econ. 2011;30(4):695-706.
Basu A. Economics of individualization in comparative effectiveness research and a basis
for a patient-centered health care. J Health Econ. 2011; 30(3):549-559.
Basu A. Person-Centered Treatment (PeT) effects using instrumental variables: An
application to evaluating prostate cancer treatments. Journal of Applied Econometrics
(In Press).
Heckman JJ, Vytlacil EJ. Local instrumental variables and latent variable models for
identifying and bounding treatment effects. Proc Nat Acad Sci 1999; 96(8): 4730-34
Heckman JJ, Urzua S, Vytlacil E. Understanding instrumental variables in models with
essential heterogeneity. Rev Econ Stat 2006; 88(3): 389-432.
Kaplan S, Billimek J, Sorkin D, Ngo-Metzger Q, Greenfield S. Who Can Respond to
Treatment?: Identifying Patient Characteristics Related to Heterogeneity of Treatment
Effects. Medical Care 2010; 48(6): S9-S16
51