Bias, Confounding & Validity

Systematic vs random error

Bias is systematic. Bias is a systematic error that pushes results in one direction and threatens internal validity (accuracy). Random error scatters results in both directions and only reduces precision. A larger sample shrinks random error but leaves bias untouched, however many subjects are enrolled.

Internal vs external validity. Internal validity asks whether the result is true within the study, free of bias and confounding. External validity asks whether it generalizes to other populations. RCTs maximize internal validity, but strict eligibility criteria can limit external validity. For example, a tightly controlled trial in young, healthy men may not apply to elderly women.

Selection and information bias

Selection bias distorts who ends up in the study or stays in it, so the sample no longer represents the population and the exposure-outcome link is skewed. Information (measurement) bias distorts how exposures or outcomes are recorded.

Recall bias in practice. Mothers of infants with birth defects report more first-trimester drug use than mothers of healthy infants, even when medical records show similar use. Recall bias is worst in retrospective case-control designs, and objective records reduce it.

Blinding. Single, double or triple blinding keeps subjects, investigators and analysts unaware of group assignment. It reduces observer-expectancy and placebo effects. Because blinded participants still know they are being studied, the Hawthorne effect is handled mainly by comparing against a control group observed in exactly the same way.

Screening biases

Lead-time bias. Screening moves the date of diagnosis earlier, so survival measured from diagnosis looks longer even when death is not delayed. The clock simply starts sooner. When a screening program improves 5-year survival but age at death is unchanged, suspect lead-time bias. Judge screening by mortality rates, not survival from diagnosis.

Length-time bias. Slowly progressive disease has a longer detectable preclinical phase. Periodic screening therefore preferentially catches indolent cases and misses fast, aggressive tumors, which appear between screens. The screen-detected group looks healthier, which inflates the apparent benefit of screening.

Overdiagnosis. Overdiagnosis is the extreme form of length-time bias: detecting disease that would never have caused symptoms in the patient's lifetime. Those patients then bear the harms of treatment with no chance of benefit.

Lead-time bias means the same death date with an earlier diagnosis. Length-time bias means the screen finds the slow, forgiving tumors.

Confounding and effect modification

What a confounder is. A confounder is associated with the exposure and is an independent risk factor for the outcome, and it does not lie on the causal pathway. The classic example is that coffee drinkers appear to get more lung cancer because smoking rides along with coffee drinking. A variable on the causal pathway (a mediator) is not a confounder and should not be adjusted away. A clue to confounding is that the crude estimate changes substantially once it is adjusted for the third variable.

Controlling confounding.

• At the design stage: randomization, restriction (e.g., enroll only nonsmokers), matching and crossover designs, in which each patient serves as their own control after a washout.
• At the analysis stage: stratification, multivariable regression and standardization. These adjust only for confounders that were measured.
• Randomization is the strongest fix because it balances both known and unknown confounders at allocation.

Matching trade-offs. Matching pairs cases and controls on confounders such as age and sex, so those variables cannot distort the comparison. The matched variable can then no longer be studied. Overmatching on a factor too closely tied to the exposure biases the result toward the null and dilutes a true effect.

Effect modification (interaction). Here the exposure's effect genuinely differs across strata of a third variable, giving different stratum-specific risk ratios. A drug that lowers MI risk in men but raises it in women is an example. So is a supplement that prevents fractures only in vitamin D-deficient subjects. These results are reported by stratum, not adjusted away.