Lecture 7 (Bias + Confounding) Flashcards
Bias (def)
Systematic errore in design/conduct of study –> mistaken asso
-Major prob in obsv epi studies
Selection bias (2 main reasons, major issue in ____)
- Error resulting f/selection procedures or factors that influence participation
- Major issue in case-control studies
Info Bias (def)
-Error due to collection of incorrect info about subjects (misclassification)
Compensating for selection bias (case-control)
-Equalizing selection bias by selecting cases/control using same selection process
Minimizing selection bias
- Carefully consider selection criteria
- Maximize participation rate
Selection bias and cohort studies (when is likely/unlikely)
- Unlikely w/internal comp grps
- External comp grps more prone (diff pops, might have diff risks aside f/exposure)
- Healthy workers effect: issue when general pop = external comp grp, healthier
- Loss to follow is an issue
Nondiff exposure/disease misclassification (bias towards of away)
Biases asso towards Ho (no asso)
Diff misclassification (related to both exposure and disease)
-Biases towards/away Ho (depends)
Nondiff exposure missclass (cohort)
- Biases RR toward Ho if exposure misscl is unrelated to the future development of disease
- Not likely
- Won’t bias if happens to same degree among those who develop/don’t
Disease missclass (cohort)
- Issue when based on self-report, esp for more subjective things
- Concern if those collecting data are aware of subjects’ exposures
Exposure missclass in case-controls (what’s the effect)
- Important source of this bias
- Bias towards null if missclass unrelated to disease stat (non diff)
- Bias either direction if it missclass depends on disease status (diff)
Types of info bias
- Recall bias
- Reporting bias
- Observer bias (sub: interviewer bias, abstractor bias)
Reducing bias
- Carefully define things
- Choose valid measurement methods
- Train ppl
- Quality control
- Maximize participation
- Apply methods in same way/detail to everyone
- Don’t improve quality of data in one but not other
Detection bias
=surveillance bias
-When close med surveillance –> higher prob of detection of disease
Most common confounder
Age
Difference btwn confounding and bias
confounding is real, bias is not
Checking for confounding
- Compare unadjusted and adjusted ORs (weighted average of stratum-specific ORs)
- If difference is more than 10%: confounding
Confounding (def)
- Distortion of asso due to another extraneous exposure associated w/both disease and exposure
- Cannot be intermediate variable in causal pathway though
Positive confounding
-Overestimation of asso
Negative confounding
-Underestimation of asso
Qualitative confounding
-Inversion of direction of asso
Prevention of confounding (design) (3)
- Randomization
- Restriction (single category of confounding exposure)
- Matching
Controlling for confounding in analyses
Stratification methods:
- Direct adj (cohort)
- Indirect (occupational retro cohort)
- Mantel haenszel method (most common, case control/cohort)
Multivariate models:
- Logistic reg (case-control)
- Cox proportional hazards model (cohort)
- Poisson regression (cohort)
Limitations of controlling for confounding using stratification
- Can only adjust for 1-2 w/small # of categories each
- Or else data would be sparse
- Only categorical variables
Residual confounding (def and sources)
- Confounding that remains after adjustment
- Sources: categories in adj too broad; variable was imperfect surrogate