Article Review

1
Article Review

• Cara Carty
• 09-Mar-06

2
Confounding by indication in non-experimental
evaluation of vaccine effectiveness the example
of prevention of influenza complications

• Hak E, Verheij TJM, Grobbee DE, Nichol KL, Hoes
  AW.
• J Epidemiol Comm Health 2002 56951-955.

3
Background

• Health impact of flu
• Outcome of interest post-flu complications
• Few randomized trials
• low incidence of flu-related complications
• virulence is variable and unpredictable
• ethical concerns
• Problems with observational studies
• conflicting results
• confounding by indication
• other confounding

4
Background

• Confounding by indication
• a variable that is a risk factor for disease
  among non-exposed persons and is associated with
  exposure of interest in the population from which
  cases derive, without being an intermediate step
  in the causal pathway between exposure and
  disease

5
Background

• Confounding by indication
• a variable that is a risk factor for disease
  among non-exposed persons and is associated with
  exposure of interest in the population from which
  cases derive, without being an intermediate step
  in the causal pathway between exposure and
  disease
• measured differences in patient groups receiving
  alternative therapies are more attributable to
  differences in patient characteristics than they
  are to differences in effectiveness of therapies

6
Causal diagram
Old age, cardiovascular disease, asthma
Exposure Flu vaccine
Pneumonia, Death
7
Strategy

• Design
• Natural experiments
• difficult to find!
• Ecological study
• communities need to be similar
• Restriction and stratification
• compare groups with similar prognosis
• may limit generalizability, but enhance internal
  validity
• Quasi-experiment
• individual matching within strata of important
  prognostic variables
• costly because it requires sufficient
  participants in each stratum

8
Strategy

• Design
• Analyses
• Control of confounding variables in multivariable
  regression model
• Use of an instrumental variable to enable
  statistical pseudo randomization and to account
  for any residual confounding
• Subclassifying or matching on levels of
  propensity scores

9
Strategy

• Design
• Analyses
• Control of confounding variables in multivariable
  regression model
• Use of an instrumental variable to enable
  statistical pseudo randomization and to account
  for any residual confounding
• Subclassifying or matching on levels of
  propensity scores

10
Strategy

• Design
• Analyses
• Control of confounding variables in multivariable
  regression model
• Use of an instrumental variable to enable
  statistical pseudo- randomization and to account
  for any residual confounding
• Subclassifying or matching on levels of
  propensity scores

11
Propensity Scores Definition

• Replace collection of confounding covariates in
  an observational study with one function of these
  covariatescollapse confounders into a single
  variable
• The score, e(X), is then used as only confounder
• e(X) is estimated using logistic regression or
  discriminant model with binary exposure (Z0 or
  Z1) and observed covariates X so that
  e(X)prob(Z1X)
• Create strata of e(X)
• Compare cases and controls within a stratum to
  calculate stratum-specific risk ratios

12
Propensity Scores Basic Concept

• Purpose
• association between vaccine and outcome
• Problem
• most vaccinees are different than unvaccinated
• few outcomes relative to number of adjustment
  factors
• Approach
• find out what factors predict vaccination by
  calculating propensity scores for every
  participant
• classify participants by quintiles of increasing
  probability of vaccination (propensity score)
• compare outcome in vaccinated and unvaccinated
  with equivalent propensity scores

13
Propensity Scores Properties

• Propensity scores balance observed covariates
• If it suffices to adjust for covariates X, then
  it suffices to adjust for their propensity score
  e(X)
• Estimated propensity scores may remove both
  systematic bias and chance imbalance in
  covariates
• Unlike random assignment, propensity score
  typically doesnt balance unobserved covariates

14
Propensity Scores Comments

• If scores are relatively constant within each
  stratum, then within each stratum, the
  distribution of all covariates should be
  approximately the same in both treatment groups
• Balance can be checked and the score reformulated
  until better balance is achieved

15
Example
Hak et al., 2002
16
Example
Hak et al., 2002
17
Example
Hak et al., 2002
18
Example
Hak et al., 2002
19
Discussion

• Cons
• Design methods are standard practice
• One worked example is not entirely convincing
• Pros
• Nice summary of non-randomization analytic issues
• Gentle introduction to propensity scores and
  their utility

20
Bibliography

• Joffe MM, Rosenbaum PR. Invited commentary
  propensity scores. Am J Epidemiol. 1999 Aug 15
  150(4)327-333.
• Rubin DB. Estimating causal effects from large
  data sets using propensity scores. Ann Int Med.
  1997 Oct 15127(8)757-763.
• Salas M, Hofman A, Stricker BH. Confounding by
  indication an example of variation in the use of
  epidemiologic terminology. Am J Epidemiol. 1999
  Jun 1149(11)981-3.