EPB PHC 6000 EPIDEMIOLOGY FALL, 1997

1
Unit 6 Standardization and Methods to Control
Confounding
2
Unit 6 Learning Objectives 1. Understand the
design and analysis methods used to control
confounding. --- Randomization --- Restriction
--- Matching --- Stratification --- Multivariat
e analysis. 2. Understand pros and cons of the
methods used to control confounding.
3
Lesson 6 Learning Objectives (cont.) 3.
Understand the rationale for rate
adjustment (standardization). 4. Apply and
interpret the technique of direct standardization
. 5. Apply and interpret the technique of
indirect standardization. 6. Recognize
differences between direct and indirect
standardization.
4
Assigned Readings Textbook (Gordis) Chapter
4, pages 60-68 (Age adjustment) Chapter 15,
pages 230-232 (More on confounding) Hennekens
and Buring Evaluating the role of confounding.
In Epidemiology in Medicine, pages 304-323.
5
CONFOUNDING - REVIEW
DEFINITION A third variable (not the exposure
or outcome variable of interest) that distorts
the observed relationship between the exposure
and outcome. Confounding is a confusion of
effects that is a nuisance and should be
controlled for if possible. Age is a very
common source of confounding.
6
CONFOUNDING - REVIEW
E
D
Confounding IS present
CF
Confounding NOT present
E
?CF
D
7
CONFOUNDING
Reason for controlling confounding To obtain
a more precise (accurate) estimate of the true
association between the exposure and disease
under study. As a general rule, age and gender
should always be considered as potential
confounders of an association.
8
CONFOUNDING
POSITIVE CONFOUNDING The confounding factor
produces an estimate that is more extreme
(positive or negative) than the true
association. NEGATIVE CONFOUNDING The
confounding factor results in an under-estimate
of the true association.
9
CONFOUNDING
METHODS TO CONTROL CONFOUNDING DESIGN 1. Rando
mization 2. Restriction 3. Matching (Analysis
also) ANALYSIS 4. Stratification 5. Multivaria
te Analysis
10
CONTROL OF CONFOUNDING

• RANDOMIZATION (Design)
• Definition Subjects or groups of subjects are
  randomly assigned to a hypothesized preventive or
  therapeutic intervention.
• Pro With sufficient sample size, virtually
  assures that both known and unknown confounders
  are controlled.
• Con Sample size may not be large enough to
  control for confounding since many persons are
  unwilling to be randomized.

11
CONTROL OF CONFOUNDING

• RESTRICTION (Design)
• Definition Study participation is restricted to
  individuals who fall within a specified category
  or categories of the confounder.
• Pro Straightforward, convenient, inexpensive
• Con Sufficiently narrow restriction range may
  severely reduce the number of eligible
  participants
• Con If restriction criteria are not sufficiently
  narrow, possibility of residual confounding exists

12
CONTROL OF CONFOUNDING

• RESTRICTION (cont.)
• Con Does not permit evaluation of the
  association between exposure and disease for
  varying levels of the factor.
• Note Although restriction may limit
  generalizability, it does not affect the internal
  validity of any observed association between the
  groups included in the study.

13
CONTROL OF CONFOUNDING

• MATCHING (Design/analysis)
• Definition All levels of the confounding factor
  are allowable for study inclusion, but subjects
  are selected in a way that potential confounders
  are distributed equally among the study groups.
• Pro Great intuitive appeal may provide greater
  analytic efficiency by insuring adequate number
  of cases and controls at each level of the
  confounder.
• Con Can be difficult, time consuming, and
  expensive to find a comparison subjects with
  right set of characteristics on each matching
  variable.

14
CONTROL OF CONFOUNDING

• MATCHING (cont.)
• Con Does not control potential confounding by
  factors other than those matched on
• Con Not needed as much as in the past due to
  alternative techniques (e.g. multivariate
  analysis)

15
CONTROL OF CONFOUNDING
INDICATIONS FOR MATCHING Factors for which
there would otherwise be insufficient overlap
between study groups (e.g. nominal-level
variables such as race). Small case series in
which baseline characteristics are likely to
differ between study groups. Most often
employed in case-control studies.
16
CONTROL OF CONFOUNDING

• MATCHING (ANALYSIS)
• Note Matching on several confounders can make
  the study groups more alike on the exposures of
  interest than would have occurred had independent
  series of cases and controls been selected.
• This requires use of statistical techniques
  that make explicit provision for the matched
  nature of the data (e.g. conditional odds ratio)

17
CONTROL OF CONFOUNDING

• STRATIFICATION (Analysis)
• Definition Evaluation of the exposure/disease
  association within homogeneous categories or
  strata of the confounding variable.
• Pro Intuitively appealing, straightforward, and
  enhances understanding of intricacies of the data
• Con Impractical for simultaneous control of
  multiple confounders, especially those with
  multiple strata

18
CONTROL OF CONFOUNDING
Hypothesis Sedentary lifestyle is associated
with risk of myocardial infarction (cohort study)
RR (40 / 120) / (100 / 850)
RR 2.83
It appears that persons with a sedentary
lifestyle are 2.83 times more likely to
experience myocardial infarction compared to
persons without a sedentary lifestyle. BUT WHAT
ABOUT SMOKING?
19
CONTROL OF CONFOUNDING
NON-SMOKERS
SMOKERS
RR (5 / 30) / (50 / 575)
RR (35 / 90) / (50 / 275)
RR 1.92
RR 2.14
Is there evidence that smoking confounds the
relationship between sedentary lifestyle and
myocardial infarction?
20
CONTROL OF CONFOUNDING
CRUDE RRMI 2.83
STRATA 1 RRNS 1.92
STRATA 2 RRSM 2.14
In general If Strata 1 RR lt Crude RR gt
Strata 2 RR OR If Strata 1 RR gt Crude RR
lt Strata 2 RR then confounding is present.
21
CONTROL OF CONFOUNDING
CRUDE RRMI 2.83
STRATA 1 RRNS 1.92
STRATA 2 RRSM 2.14
Now, the question is Should the
stratum-specific estimates be combined to obtain
an unconfounded (adjusted) estimate of the
relationship between sedentary lifestyle and
risk of myocardial infarction?
22
CONTROL OF CONFOUNDING
CRUDE RRMI 2.83
STRATA 1 RRNS 1.92
STRATA 2 RRSM 2.14
Axiom If the stratum-specific estimates are
similar (homogeneous), the estimates can be
combined to obtain an unconfounded (adjusted)
estimate. However, if the stratum-specific
estimates are sufficiently different, they should
not be combined, as this would obscure useful
information (lecture 7).
23
CONTROL OF CONFOUNDING
CRUDE RRMI 2.83
STRATA 1 RRNS 1.92
STRATA 2 RRSM 2.14
Note Statistical tests of homogeneity exist to
test the similarity of the stratum-specific
estimates, however, these tests are heavily
affected by sample size, and often
under-powered. Thus, the stratum-specific
estimates should be eyeballed.
24
CONTROL OF CONFOUNDING
CRUDE RRMI 2.83
STRATA 1 RRNS 1.92
STRATA 2 RRSM 2.14
Mantel-Haenszel pooled RR estimate (uniform
strata) S a(c d) / T RRMH
----------------- S c(a b) / T Where T
total sample in each stratum
25
CONTROL OF CONFOUNDING
CRUDE RRMI 2.83
STRATA 1 RRNS 1.92
STRATA 2 RRSM 2.14
5(50 525) / 605 35(50 225) / 365 RRMH
--------------------------------------------
------ 50(5 25) / 605 50(35 55) /
365 4.75 26.4 31.1
-------------- ------ 2.10 2.48
12.3 14.8
26
CONTROL OF CONFOUNDING
CRUDE RRMI 2.83
ADJUSTED RRMH 2.10
Axiom The magnitude of confounding is evaluated
by observing the degree of discrepancy observed
between the crude and adjusted estimates. Presen
ce of confounding should not be assessed using a
test of statistical significance. Generally,
when the crude estimate changes by at least 10,
meaningful confounding exists.
27
CONTROL OF CONFOUNDING

• MULTIVARIATE ANALYSIS (Analysis)
• Definition A technique that takes into account a
• number of variables simultaneously.
• Involves construction of a mathematical model
• that efficiently describes the association
  between exposure and disease, as well as other
  variables that may confound or modify the effect
  of exposure.
• Examples Multiple linear regression model
• Logistic regression model

28
CONTROL OF CONFOUNDING
MULTIVARIATE ANALYSIS (Analysis) Multiple
linear regression model Y a b1X1 b2X2
bnXn Where n the number of independent
variables (IVs) (e.g. Exposure(s) and
confounders) X1 Xn individuals set of
values for the IVs b1 bn respective
coefficients for the IVs
29
CONTROL OF CONFOUNDING

• MULTIVARIATE ANALYSIS (Analysis)
• Logistic regression model
• ln Y / (1-Y) a b1X1 b2X2 bnXn
• Where
• Y probability of disease
• n the number of independent variables (IVs)
• (e.g. exposure(s) and confounders)
• X1 Xn individuals set of values for the IVs
• b1 bn respective coefficients for the IVs

30
CONTROL OF CONFOUNDING

• MULTIVARIATE ANALYSIS (Analysis)
• Pro Can simultaneously control for multiple
• confounders when stratified analysis is
  impractical
• Pro With the logistic regression model, beta
• coefficients can be directly converted to odds
  ratios
• Con Process of efficient mathematical modeling
  can
• occur at the expense of clear understanding of
  the data