7 Aug 2002 – DRAFT – some issues that need to be resolved are highlighted; please provide comments/suggestions to Kevin Sullivan (cdckms@sph.emory.edu); Copyright © 2001 Data Description Inc.




Issue: how do deal with several risk factors that may confound an E→D relationship


Assumption: no risk factors are effect modifiers


Two Fundamental Principles for Confounding with Two or More Risk Factors


1.        Joint control of 2+ variables may give different results from controlling each variable separately.  The adjusted estimate that controls for all risk factors is the “standard” for which conclusions of confounding are based.


2.        Not all variables may need to be controlled.


Data-based joint confounding – meaningful difference between crude and adjusted effect when controlling for all potential confounders


aRR(age, smoking) =           2.4

cRR =                                    1.5


Data-based marginal confounding – meaningful difference between crude and adjusted effect controlling for only one  confounder


cRR = 1.5                              aRR(age) =                 1.5

aRR(smoking) =          2.4



Considering principle 2, an adjusted estimate for a subset of risk factors may adequately control for confounding.


Example: cRR = 2;      aRR(K, L) = 1.0;      aRR(K) = 1.0;      aRR(L) = 1.0


Why not always control for all potential confounders?  Validity vs. Precision


Confounding: Validity versus Precision



Identifying a subset of confounders giving a precise estimate (yet still “valid”) is important enough to make such examination worthwhile.


Study Questions (Q11.7)


A clinical trial to determine effectiveness of a treatment.


1.     The cRR = 6.28 and the aRR(AGE, SERH, TSZ, INSG) = 8.24.  Confounding?


Four subsets shown below.


2.     What criteria may we use to reduce the number of candidate subsets?  Note the answer there is +/- 10% around adjusted estimate.


Below are the results from the gold standard and the 4 candidate subsets whose aRR is within 10% of the gold standard:



3.     The most valid estimate results from controlling which covariates? 

4.     The most precise estimate results from controlling which covariates?

5.     Which covariates do you think are most appropriate to control?


A valid estimate of effect is most important.

Consider the trade-off between controlling for enough risk factors to maintain validity and loss in precision from control of too many variables.



Study Designs by Bias


Comparison of study designs for assessing exposure-disease relationships


     Strongest                                                            Weakest


Clinical Trial     Cohort      Case-Control  Cross-Sectional



Systematic Error


Selection Bias

                                √√                    √√√                   √√√

                   Loss to Follow-up                Berkson’s             Survival


Information Bias

                                                                   √√√                    √√√

               E→D                  E→D                   E←D                 E↔D

                                                           Recall               Self-report



              Unlikely                   √√                     √√                           √√      





Note: Sampling Error

            Affects all designs; can reduce by increasing sample size