pubhlth 206a- week 6: confounding

confounding

• systematic diff b/w the groups being compared that distorts the true association b/w an exposure & disease

• mixing of effects

• estimate of the effect of exposure on disease is distorted b/c it is mixed w/ the effect of other factors associated w/ the exposure & the disease 

• aka “third variable”

• c —> exposure

• c —> disease

confounding source for experimental & cohort

when the exposed & unexposed groups differ by more than just the exposure-they differ by some other variable

confounding source for case-control

when cases & controls have diff characteristics

confounding vs bias

• bias isn’t an inherent trait of the pop

• confounding is an inherent trait of the pop

counterfactual ideal

• ideal comparison group would be the exact same people who are in the exposed group had they not been exposed

• selection:
  1. w/ respect to other factors that could influence outcome
  2. w/ respect to collection of comparable & accurate info

how is confounding related to the counterfactual ideal

• confounding is a failure to come close to the counterfactual ideal

• confounding occurs when the risk of disease in the unexposed group doesn’t = the risk of disease the exposed group had they been unexposed

3 criteria for when is a variable a confounder

• independent predictor of the outcome (yes)

  • confounder is a risk factor for disease in unexposed group

• associated w/ the exposure (yes)

  • confounder occurs more or less often in the exposed than the unexposed

• can’t be an intermediate on the causal pathway b/w exposure & disease (no)

  • confounder cannot be caused by the exposure.

confounding can be “controlled” or “adjusted”: design phase (before study)

• randomization

• restriction

• matching

confounding can be “controlled” or “adjusted”: analysis phase (after study)

• standardization

• stratified analysis

• multivariate analysis

goal of the design & analysis phase

“break” the association b/w the confounder & the exposure

randomization

• randomly allocate equal change of being assigned to the treatment or comparison group

  • w/ an adequate number of subjects, for
    baseline comparability of groups in
    terms of both known & unknown confounder

  • this only works when:

    • study is large enough

    • treatment assignment is not
      influenced by investigator

randomization strengths

• no limit on the number of confounders that can be controlled for

• don’t need info about unknown confounders

randomization weaknessess

• limited to experimental studies

• less effective w/ smaller sample size

restriction

limit study to people who are within 1 category of the confounder

restriction strengths

• simple conceptually & practically

• effective control of characteristics being restricted

restriction limitations

• only possible for known, measured confounders

• incomplete control for confounding

• can’t evaluate restricted variable

• limits sample size

• limits the generalizability

matching

select study subjects so that confounders are distributed identically among the exposed & unexposed (cohort study) or case and controls (case-control study)

matching strengths

• simple & effective control of characteristics being matched

• useful for vairable that are difficult to capture

• confounding can be addressed w/o matching

• increases the statistical efficiency of confounder
  adjustment & can be useful for strong confounders

matching limitations

• only possible for known measured confounders

• can be difficult, expensive & time-comsuming to find appropriate matches

• can’t evaluate matched variable

stratification

• separate your study population into subgroups where 1 group has the confounder characteristic & 1 group doesn’t. then calculate a measure of association for each subgroup

• ALWAYS USE STRATIFICATION

stratification strengths

• straightforward & easy to perform

• effective control of characteristic being stratified

stratification limitations

• difficult to control for many confounders simutaneously due to sparse data problems

• difficult presentation, esp, if many confounders

• continuous variables not easily stratified

stratified analysis

• calculated crude measure of association

• divide subjects into strata of the confounder

• calculate stratum-specific measures of association

• calculate adjusted measure of association

• determine whether crude measure differs from adjusted measure of association, & by how much (magnitude)

magnitude of confounding (how big confounding is)

(crude rr - adjusted rr) / (adjusted rr) x 100%

ex: crude rr is 20.4% less than the adjusted rr

10% rule

• serves as a guideline

• when confounding occurs, a judgement call has to be made to decide whether it is important

• one common approach is the CHANGE-IN-ESTIMATE GUIDELINE

  • decide that confounding is important when the adjusted RR/OR differes from the crude by X% or more

  • typical cut-point is 10% (ex: -20.4% magnitude of confounding would be consider

note

counfounding in not assessed statistically via p-values

confounding can bias results toward or away from the null?

both

how to know what might be a potential confounder for specific research question

• know subject area

• complete a comprehensive literature review & read

  • previous research will help identify “known” confounders

• “historical” confounders

  • some variables are always considered potential confounders (age, sex, race/ethnicity)

adjusted rr = 5.4

((stratum-specfic w/o confounder) + (stratum specific w/ confounder)) / 2

if crude RR = adjusted RR

no confounding is present

if crude RR ≠ adjusted RR

confounding is present

multivariate regression

• involves construction of statistical model that describes the association b/w exposure, disease, & confounder

• multiple linear regression for continuous outcomes

  • ultivariate logistic regression for dichotomous outcomes

  • cox proportional hazards model for longitudinal data

multivariate regression advantage

simultaneously adjusts for several variables

multivariate regression disadvantage

difficult to conceptualize, data need to fit into an avail statistical model (assumptions needed)

residual confounding

• persists despite efforts to control or adjust for confounding

• should be acknowledged & addressed in discussion section of a published paper

residual confounding sources

• no data or inaccurate data on a confounder

• use of broad categories of a confounder in your analysis
  – wide age groups, ever smokers vs. never smokers

summary

• confounding can lead to incorrect causal conclusions

• confounding is not simply an issue of crude vs. adjusted parameter estimates

  • this information helps diagnose confounding but it is not conclusive

• if we want to make inferences regarding causation, we need to account for the possibility of confounding

• confounding is not the result of an inherent error in the conduct of a study, but rather a causal concept that arises from nature

• no statistical hypothesis tests (or p-values) that explicitly test for confounding

• there can be no substitute for good scientific judgment & expertise in determining confounding

• causal knowledge is a prerequisite for confounding evaluation