Bias

01Definition and Core Concepts

Bias is defined as any systematic error in the design, conduct, or analysis of a study that leads to results or conclusions that differ from the truth. The word "systematic" is critical here because it separates bias from random error. Random error is due to chance variability and can be minimized by increasing sample size, while bias is a structural problem baked into how the study was built or run. No amount of additional participants will correct a biased study.

The reason this distinction matters on the exam is that test writers will describe a study scenario and ask you to identify whether the flaw is "error" or "bias." The key discriminator is directionality: bias pushes results consistently in one direction (overestimation or underestimation of the true effect), whereas random error scatters results unpredictably in both directions.

Bias threatens internal validity, which is the degree to which observed differences between groups can be attributed solely to the variable being studied. When internal validity is compromised, you cannot trust the study's conclusions, regardless of how large the sample was or how statistically significant the p-value appears.

02Classification of Bias

There are three overarching families of bias. Within each family, there are named subtypes that appear frequently in exam vignettes.

03Selection Bias

Selection bias arises when the method used to select or retain study participants results in a sample that is systematically different from the population the study intends to represent. It occurs at the enrollment or retention stage of a study.

Berkson Bias deserves close attention. It is one of the most commonly tested named biases. The classic vignette describes a case-control study conducted entirely within a hospital setting. Hospitalized patients are not representative of the general population because they are sicker and more likely to have multiple comorbidities. This inflates the apparent association between an exposure and a disease. The fix is to use community-based controls rather than hospital-based ones.

Attrition Bias is the other high-yield selection bias. It is relevant to any longitudinal study. The exam tests this when a vignette describes differential dropout between the treatment and control arms. If sicker patients leave the treatment group because of adverse effects, the remaining patients look healthier, and the treatment appears more effective than it truly is. The test writer is checking whether you recognize that the final study population no longer reflects the original randomized groups.

04Information (Measurement) Bias

Information bias occurs during data collection, when the exposure, outcome, or other variables are measured or recorded inaccurately. The key exam concept is that this can be differential (the error differs between groups) or non-differential (the error is equal across groups). Differential misclassification can bias results in either direction. Non-differential misclassification tends to bias results toward the null (making a real association look weaker than it is).

Recall Bias is the single most frequently tested information bias. The classic vignette involves a case-control study where participants are asked to remember past exposures. The exam is testing your understanding that cases (people with the disease) tend to search their memory more thoroughly than controls, producing a differential recall. The distinguishing feature in a vignette is that data is gathered retrospectively via interviews or questionnaires. The fix is to use objective records such as pharmacy databases or medical charts.

Hawthorne Effect is tested differently from other biases. A typical vignette describes a study in which simply enrolling participants or observing them leads to improved outcomes in both the treatment and control groups. The important nuance is that the Hawthorne effect is not about the intervention itself; it is about the act of being watched.

05Confounding

Confounding is often presented alongside bias, but it is mechanistically distinct. A confounder is an extraneous variable that is associated with both the exposure and the outcome but is not on the causal pathway between them. It distorts the apparent relationship between the exposure and the outcome.

The three criteria for a confounder are:

1. It must be associated with the exposure.

2. It must be independently associated with the outcome.

3. It must not be an intermediate step (mediator) on the causal pathway from exposure to outcome.

The most testable concept in confounding is distinguishing it from effect modification (interaction). In confounding, the crude (unadjusted) result differs from the stratum-specific results, but the stratum-specific results are similar to each other. You want to remove confounding by adjusting. In effect modification, the stratum-specific results differ meaningfully from each other, and you want to report the effect modification, not adjust it away.

Exam vignettes test this distinction by showing you a crude odds ratio, then showing stratified odds ratios. If the stratified ORs are roughly equal to each other but differ from the crude OR, the answer is confounding. If the stratified ORs are meaningfully different from each other, the answer is effect modification.

Randomization is the most powerful tool against confounding because it is the only method that controls for both known and unknown confounders. All other strategies (matching, restriction, stratification, regression) require you to identify the confounder in advance.

06Screening-Related Biases

These biases are unique to studies evaluating screening programs. They are tested in vignettes that describe survival statistics after the introduction of a new screening test.

Lead-time bias is one of the most commonly tested concepts in screening-related biostatistics. The classic trap is that the vignette provides survival time from diagnosis and asks you why the screened group appears to live longer. Students who do not recognize lead-time bias will choose "the screening test improves survival," which is the intended distractor. The correct interpretation is that the screening test moved the moment of diagnosis earlier, and the patient's total lifespan was unchanged. To determine whether a screening test truly saves lives, you must compare mortality rates between the screened and unscreened populations, not survival from the point of diagnosis.

Length-time bias is subtler and less frequently tested, but it does appear. The giveaway in the vignette is that screened patients are found to have "less aggressive" or "slower-growing" tumors compared to unscreened patients who present with symptomatic, advanced disease. The screening did not improve outcomes; it simply detected a biologically different subset of disease.

07Strategies to Reduce Bias: Summary

08High-Yield Vignette Patterns

The exam tests bias in one of three predictable formats:

Format 1: "What type of bias is present?" The vignette describes a flawed study and asks you to name the bias. The key is to identify when the error occurred. If it occurred during recruitment, think selection bias. If it occurred during data collection, think information or measurement bias. If it is a third variable distorting results, think confounding.

Format 2: "How could this bias have been prevented?" After identifying the bias, the next question asks what design feature would have eliminated it. Recall bias is prevented by using objective records. Observer bias is prevented by blinding. Selection bias is prevented by randomization and allocation concealment. Confounding is prevented by randomization (design phase) or multivariable analysis (analysis phase).

Format 3: "Is the finding valid?" A vignette presents a screening test that appears to improve survival. The question tests whether you recognize lead-time bias or length-time bias as the explanation for an apparent improvement that is not real.

09Traps and High-Yield Pearls

The most common trap in bias questions is confusing confounding with bias. Students often treat them as interchangeable, but they are not. Bias is a structural flaw in how the study was designed or run, whereas confounding is a statistical phenomenon that can be addressed after the fact using analytical techniques. Randomization is the only strategy that addresses both.

A second frequent trap involves Berkson bias. Students read a case-control vignette set in a hospital and immediately assume the study is valid because it has cases and controls. The correct instinct is to ask: "Were the controls drawn from the same source population as the cases?" If both come from the same hospital, the controls are not representative of the general population.

A third trap appears in screening questions. The vignette provides impressive survival statistics and a new screening test. The intended wrong answer is that the screening test improves outcomes. The correct answer recognizes lead-time bias and asks for mortality data instead. The core competency being tested is your ability to distinguish between "detecting disease earlier" and "actually changing the outcome."

Finally, remember that non-differential misclassification biases toward the null (it dilutes the association), while differential misclassification can bias in either direction. If the vignette describes equal inaccuracy across both groups, the association is being underestimated. If the inaccuracy is different between groups (as in recall bias), the association could be either inflated or deflated.