Bias, Confounding, and Causation

Confounding Variables

A confounding variable is an extraneous factor that correlates with both the independent variable (exposure) and the dependent variable (outcome), creating a spurious association. Unlike random error, confounding is systematic -- it pushes results consistently in one direction. Example: a study finds that coffee drinkers have higher rates of lung cancer. But coffee drinkers also smoke at higher rates. Smoking -- not coffee -- drives the cancer risk. Smoking confounds the coffee-lung cancer association. Controlling for confounds requires: (1) randomization in experimental studies, which distributes confounds equally; (2) statistical adjustment in observational studies (stratification, multivariable regression); (3) matching in case-control designs. A confound must meet three criteria: it must be associated with the exposure, associated with the outcome, and not be on the causal pathway between them (a mediator is not a confound). The MCAT often asks: 'Which of the following, if true, would most weaken the conclusion?' The correct answer is typically an unmeasured confound.

Selection Bias

Selection bias occurs when the sample included in a study differs systematically from the target population, distorting the exposure-outcome relationship. This is a sampling error, not a measurement error. Common forms: (1) Self-selection bias -- participants choose whether to enroll, and those who do may differ from those who do not (e.g., volunteers for a health study may be healthier than the general population). (2) Berkson's bias (admission-rate bias) -- hospitalized patients have more comorbidities, inflating apparent associations (e.g., a study in a hospital may find a spurious link between two diseases simply because having both increases admission probability). (3) Attrition bias (loss to follow-up) -- participants who drop out differ systematically from those who remain, especially problematic in longitudinal studies. (4) Healthy worker effect -- employed individuals are healthier than the general population, biasing occupational health studies. Selection bias threatens external validity (generalizability) when it distorts who is in the study; it threatens internal validity when the selection mechanism differs between comparison groups. Random sampling prevents selection bias; convenience sampling practically guarantees it.

Recall Bias

Recall bias is a systematic difference in the accuracy or completeness of memories reported by study participants. It is especially problematic in case-control studies where participants with a disease (cases) are more motivated to search their memory for past exposures than healthy controls. Example: in a study of birth defects, mothers of affected children may recall every medication, illness, and environmental exposure during pregnancy far more thoroughly than mothers of healthy children -- not because they were actually more exposed, but because they have searched their memory more intensely. Recall bias inflates the apparent association between exposure and disease. Mitigation strategies: (1) use prospective cohort designs where exposure is measured before outcome occurs; (2) verify self-reports against objective records (medical records, prescription databases); (3) use standardized questionnaires with specific prompts rather than open-ended recall; (4) blind participants to the study hypothesis so they do not differentially search their memory. The MCAT often presents a case-control study and asks whether recall bias could explain the observed association.

Observer Bias

Observer bias (also called experimenter bias, ascertainment bias, or detection bias) occurs when the researcher's expectations, beliefs, or knowledge of group assignment influence how they measure, record, or interpret outcomes. This is not fraud -- it is often unconscious. A researcher who believes Drug A works may subconsciously rate the treatment group's symptoms as slightly milder, measure blood pressure as slightly lower, or probe more thoroughly for side effects in the treatment group. Observer bias can create an effect where none exists or exaggerate a real effect. The primary defense is blinding: in single-blind designs, participants do not know their assignment; in double-blind designs, neither participants nor data collectors know group assignments. Double-blinding is the gold standard. When blinding is impossible (e.g., comparing surgery to physical therapy), researchers use objective outcome measures (mortality, lab values) rather than subjective ratings, and employ independent assessors who were not involved in the intervention. The MCAT asks: given that blinding was impossible in this study, what bias threatens its conclusions?

Placebo Effect

The placebo effect is a genuine physiological or psychological response to an inert treatment driven by the participant's expectation of benefit. It is not 'fake' -- placebo analgesia involves real endogenous opioid release, and placebo-induced bronchodilation has been documented. In drug trials, the placebo effect can produce improvements that rival the active treatment, especially for subjective outcomes (pain, mood, fatigue). This makes the placebo effect a confounding variable: if the treatment group receives both the active ingredient and the expectation of benefit while the control group receives neither, the observed difference overestimates the drug's true pharmacological effect. The standard defense is the placebo-controlled double-blind trial: participants in both the treatment and control groups have the same expectation of benefit (they both believe they might be receiving the real drug), so any difference between groups can be attributed to the active ingredient alone. The MCAT frequently presents a study without a placebo control and asks whether the observed improvement could reflect expectation alone rather than a true treatment effect. Note the nocebo effect -- negative expectations producing real adverse effects -- follows the same logic in reverse.

Correlation versus Causation

Correlation (association) means two variables vary together systematically. Causation means changing one variable produces a change in the other. The distinction is the single most tested concept in MCAT research methods. Three criteria for inferring causation (Hill's criteria, simplified): (1) Temporal precedence -- the cause must precede the effect in time. Cross-sectional studies measure exposure and outcome simultaneously, so temporal order is unknown. (2) Covariation -- changes in the cause must correlate with changes in the effect. If there is no association, there is no causation (but the reverse is not true). (3) Elimination of alternative explanations -- confounds must be ruled out through randomization, control, or statistical adjustment. Experimental designs (RCTs) satisfy all three by design. Observational studies cannot satisfy criterion (3) fully and therefore support association, not causation. Common MCAT traps: assuming a correlational study proves causation, ignoring the directionality problem, and overlooking the possibility of a third variable. Biological plausibility (a fourth Hill criterion) strengthens causal arguments but does not prove them -- even biologically implausible associations can be causal.

Reverse Causation

Reverse causation (reverse causality) occurs when the presumed direction of causation is incorrect: instead of X causing Y, Y causes X. This is a specific form of the directionality problem. Example: a cross-sectional study finds that people with depression have lower levels of physical activity. The media reports that lack of exercise causes depression. But the reverse may be true: depression causes reduced motivation to exercise. Without temporal data (knowing which came first), you cannot distinguish forward from reverse causation. Reverse causation is especially common in: (1) cross-sectional studies, which measure everything at one time point; (2) studies of biomarkers where the disease process itself alters the biomarker level; (3) studies of behaviors and health where illness changes behavior. Mitigation: prospective cohort studies establish temporal order by measuring exposure before the outcome develops; instrumental variable analysis and Mendelian randomization can help distinguish direction in observational data. The MCAT frequently asks you to identify whether a study design can rule out reverse causation -- look for temporal ordering (was the exposure measured before the outcome?). If not, reverse causation remains a plausible alternative explanation.

How it works

When the MCAT presents a study and asks whether its conclusion is valid, run a bias audit: (1) Could a confound explain the association? Look for unmeasured variables correlated with both exposure and outcome. (2) Was the sample selected in a way that distorts the relationship (selection bias)? (3) In case-control designs, could cases have recalled past exposures differently than controls (recall bias)? (4) Did anyone measuring outcomes know group assignments (observer bias)? (5) Could expectation alone produce the result (placebo effect, nocebo effect)? (6) Does the design establish temporal order (correlation vs. causation)? (7) Could the arrow point the other way (reverse causation)? If the answer to any of these is 'yes' and the study cannot rule it out, the conclusion is at best tentative.

How it works

When the MCAT presents a study and asks whether its conclusion is valid, run a bias audit: (1) Could a confound explain the association? Look for unmeasured variables correlated with both exposure and outcome. (2) Was the sample selected in a way that distorts the relationship (selection bias)? (3) In case-control designs, could cases have recalled past exposures differently than controls (recall bias)? (4) Did anyone measuring outcomes know group assignments (observer bias)? (5) Could expectation alone produce the result (placebo effect, nocebo effect)? (6) Does the design establish temporal order (correlation vs. causation)? (7) Could the arrow point the other way (reverse causation)? If the answer to any of these is 'yes' and the study cannot rule it out, the conclusion is at best tentative.

Comparisons

• B/B (Epidemiology passages): Case-control and cohort studies test your ability to spot recall bias, selection bias, and confounding in disease-exposure relationships.
• C/P (Drug studies): Placebo-controlled trials appear in passage form -- expect questions about whether results reflect pharmacological effect or expectation.
• P/S (Research methods): The entire section tests bias recognition. Passages describe flawed studies and ask you to identify which bias threatens validity.
• RM-002 (Variables and Controls): Confounding is introduced in RM-002 as a threat to internal validity; RM-005 provides the detailed typology of biases and the causation framework.
• RM-003 (Study Design): Each study design (case-control, cohort, cross-sectional) has characteristic biases -- RM-005 explains which biases plague which designs.
• RM-010 (Correlation and Regression): Spurious correlations and statistical vs. causal interpretation build directly on the correlation-vs-causation framework established here.

Common confusions

• Calling a confound a mediator: A confound is a third variable driving both exposure and outcome. A mediator is on the causal pathway (exposure -> mediator -> outcome). Adjusting for a mediator removes the very effect you are trying to measure.
• Assuming randomization fixes everything: Randomization prevents confounding at baseline but does not fix selection bias, recall bias, observer bias, or attrition. A poorly blinded RCT can still have biased outcome assessment.
• Treating correlation as causation because p < 0.05: Statistical significance means the association is unlikely to be due to chance. It says nothing about whether the association is causal. A confounded correlation can be highly significant.
• Ignoring temporal ambiguity: Cross-sectional studies cannot establish temporal order. When a passage reports 'X is associated with Y' from a single-timepoint survey, reverse causation and third-variable explanations remain fully viable.
• Overlooking the direction of recall bias: Recall bias is not random -- cases over-report exposures they believe might be relevant. This inflates (not deflates) the apparent association. If a passage says cases and controls misremember equally, there is no recall bias.
• Assuming placebo effects are 'not real': The MCAT treats placebo effects as genuine physiological phenomena. A placebo response does not mean the patient was faking or the symptom was imaginary -- it means expectation produced measurable biological change.
• Confusing 'no association found' with 'no causation': Observational studies can miss true causal effects due to confounding, measurement error, or insufficient power. 'We found no correlation' is not proof of no causal relationship.

Quick review

• Confound: third variable correlated with both exposure and outcome; creates spurious association
• Selection bias: sample differs systematically from target population (self-selection, Berkson's, attrition, healthy worker)
• Recall bias: cases remember past exposures more thoroughly than controls; especially in case-control studies
• Observer bias: researcher expectations influence measurement; prevented by double-blinding
• Placebo effect: genuine physiological response to expectation; placebo-controlled RCTs isolate pharmacological effect
• Correlation vs. causation: association does not imply causation; requires temporal precedence, covariation, and elimination of alternatives
• Reverse causation: Y causes X instead of X causing Y; cross-sectional studies cannot rule this out
• Temporal precedence: the cause must come before the effect; cross-sectional studies fail this criterion
• Double-blinding: neither participant nor data collector knows group assignment; prevents observer bias and placebo confounding
• Mediator vs. confound: mediator is on causal pathway (do not adjust); confound is outside it (must adjust)