In observational studies, participants are not
randomly assigned to intervention groups. In fact, individuals receiving
a given treatment may be markedly different than those not receiving
treatment. Covariates that are independently associated with both
treatment and outcome variables are called confounders. Illness
severity, for example, would be considered a confounding variable if it
influences whether or not a patient receives a given treatment and is
also associated with the outcome of interest. Important covariates may
not be available in existing datasets. Ignoring group differences in
important covariates, whether available or not, can lead to biased
estimates of treatment effects. It is important to remember that random
error (chance) leads to imprecise results, whereas systematic error
(bias) leads to inaccurate results.
Common approaches to control for group
differences include stratified analyses, matching, or multivariable
modeling using observed covariates, but these strategies are limited in
the number of covariates that can be included, and none address
unobserved covariates. Alternative techniques to deal with confounding
include sensitivity, propensity score, or instrumental variable
Sensitivity analysis identifies what
the strength and prevalence of an unmeasured confounder would have to
be to alter the conclusion of the study. In other words, sensitivity
analysis does not rule out the possibility that confounding exists; it
describes the circumstances necessary for an unmeasured confounder to
negate the observed effect of the treatment (or exposure) on the
Propensity score analysis uses any and
all observed covariates to determine the likelihood (conditional
probability) that a person belongs to the treatment group. The
propensity scores can then be used, through a variety of options, to
balance observed covariates and thus, reduce observed confounding.
Instrumental variable (IV) analysis involves
identifying a variable (instrument) that is associated with treatment,
but not directly associated with the outcome. Since all unmeasured
factors are part of error term, selection bias is (likely) present when
error term is correlated with both the outcome and the treatment
variable. IV analysis involves 1) modeling treatment as a function of
covariates and instrument, and 2) use this information to 'break link'
with unobserved confounder(s). The unique feature of IV analysis is that
it reduces confounding from both observed and unobserved factors.
MA Brookhart et al. (2010) Instrumental variable methods in
comparative safety and effectiveness research. Pharmacoepidemiol Drug
Saf. 19: 537-554
RB D'Agostino (1998) Tutorial in biostatistics propensity
score methods for bias reduction in the comparison of a treatment to a
non-randomized control group. Statist Med. 17: 2265-2281.
JP Leigh & M Schembri (2004) Instrumental variables
technique: cigarette price provided better estimate of effects of
smoking on SF-12. J Clin Epidemiol. 57(3): 284-293. EP Martens et al.
(2006) Instrumental variables application and limitations. Epidemiol.
PR Rosenbaum (2005) Sensitivity analysis in observational
studies. Encyclopedia of Statistics in Behavioral Science. Vol 4:
MG Stineman et al. (2008) The effectiveness of inpatient
rehabilitation in the acute postoperative phase of care after
transtibial or transfemoral amputation: study of an integrated health
care delivery system. Arch Phys Med Rehabil. 89: 1863-1872.
JA Stukel et al. (2007) Analysis of observational studies in
the presence of treatment selection bias: Effects of invasive cardiac
management on AMI survival using propensity score and instrumental
variable methods. JAMA. 297(3): 278-285.
Other Links - Video
The videos below cover analytic procedures for
dealing with confounding and recorded during the Comparative
Effectiveness Research with Population-Based Data conference in the
Baker Institute at Rice University on July 13, 2012.