Bias and confounding

Bias systematically distorts a comparison. It can arise from who enters or remains in a study, how exposures and outcomes are measured, how time is counted, or factors that influence both the exposure and the outcome.

Bias, confounding and validity

Bias is systematic error, distinct from random error or lack of precision. ^{1, 2} Internal validity concerns whether the exposure–outcome association is assessed correctly; a study’s internal validity does not guarantee that its findings generalise to other populations. ¹ Oxford’s Centre for Evidence-Based Medicine maintains a Catalogue of Bias with definitions of biases affecting health research. ³

A confounder is associated with both an exposure and its outcome but does not lie on the proposed causal pathway between them. ⁴ For example, a hypothetical association between drinking expensive kopi luwak coffee and lower mortality could reflect the drinkers’ greater financial resources and access to high-quality healthcare. ⁴ Adjustment for income may leave residual confounding from education, health beliefs, literacy and health-promoting behaviour. ⁴

Matching, stratification, covariate adjustment, regression and sensitivity analyses are approaches to controlling or accounting for bias and confounding. ^{5, 6} Neither regression nor propensity-score matching controls confounding by unmeasured variables. ⁷

Indication, severity and treatment selection

Confounding by indication arises when disease risk, symptoms or prognosis influence the treatment selected and are also associated with the outcome. ⁸ Channeling occurs when treatments with similar indications are preferentially prescribed to patients with different baseline prognoses, potentially making a treatment appear harmful because its recipients have poorer prognoses. ⁹ Conversely, patients with more severe disease may receive more intensive or advanced care and have better outcomes than less severely ill patients receiving less intensive care. ^{10, 11}

Propensity-score matching seeks to balance groups on observed baseline characteristics associated with treatment receipt. ⁷ Matching can be incomplete or inexact and loses data from patients without comparable controls. ⁷ Its ability to balance measured characteristics does not remove its limitation with unmeasured confounders. ⁷

Entry, retention and survival

Sampling and referral

Hospital patients may differ from community cases because of healthcare access, providers’ interest in particular diseases, institutional reputation, referral of complex cases to tertiary centres, and socioeconomic, geographic or cultural access to treatment. ¹² The healthy-worker effect arises because employed people tend to be healthier than the general population, including unemployed people. ¹³ Healthy volunteers may likewise have better health than the general population. ¹⁴

People reached through expensive phones or applications may be more affluent than those excluded by the technology. ^{15, 16} Patient-advocacy-group survey respondents may be especially health-conscious, motivated and able to afford treatment, inflating apparent adherence relative to a broader population. ^{16, 11} Simple random and stratified sampling are generally less susceptible to bias than convenience sampling, which uses readily available data, or snowball sampling, which recruits acquaintances of participants. ¹⁶

Hospital controls and eligibility

Berkson’s bias arises in hospital-based studies because exposure and disease can both influence admission, producing an exposure distribution unlike that in the community. ^{13, 17} People with several diseases are more likely to be admitted than people with one, and the resulting association may be biased towards or away from the null. ¹⁷ Sampling incident rather than prevalent cases from outpatient clinics and registries as well as hospitals, and selecting community controls, can reduce this bias. ^{13, 17}

Inclusion bias can occur when hospital controls have conditions associated with the exposure, raising exposure frequency in the reference group and potentially biasing the association towards the null. ¹⁵ Exclusion bias can arise when cases have concomitant conditions that are excluded among controls. ¹³ Excluding patients who experience serious adverse events during a treatment-free run-in can also create a spurious association between a new treatment and survival. ¹³ Matching on nonconfounders associated with exposure but not disease can introduce overmatching and bias an association towards the null. ¹⁵

Nonresponse and withdrawal

Nonresponse bias occurs when respondents differ meaningfully from nonrespondents. ^{14, 18} Imbalanced losses to follow-up can affect the validity of experimental and cohort-study findings. ^{14, 2} Retrospective analyses requiring complete records may select patients with longer hospital stays, more serious illness or both. ¹⁴ Requiring continuous health-plan enrolment for a minimum period may overestimate adherence because continuity of care is associated with adherence. ¹¹

Cohort selection can occur on entry or departure when inclusion or exclusion is influenced by both exposure and outcome risk; inverse probability weighting is one possible adjustment. ¹⁶

Selection of survivors

Neyman’s bias, also called incidence–prevalence or selective-survival bias, occurs when an exposure affects prognosis and the surviving prevalent cases no longer represent the exposure distribution among incident cases. ^{12, 19} For example, heavy drinkers’ higher mortality after hepatocellular carcinoma can weaken the apparent association between alcohol abuse and prevalent carcinoma. ¹⁹ Cross-sectional snapshots can underrepresent diseases with high mortality because their prevalence may be low despite high incidence. ²⁰ Survivor treatment-selection bias can make treatment appear associated with longer survival when people who survive longer have more opportunity to receive it. ¹³

Index-event bias is an inverted association between a risk factor and recurrence among people who have already experienced the disease. ⁹ Risk factors or protective factors can cluster differently in these survivors, creating apparent protection, including the obesity paradox in recurrent coronary events. ⁹ A propensity-score-matched analysis eliminated an apparent obesity paradox. ⁹ In cryptogenic-stroke survivors with patent foramen ovale, an apparent reduction in recurrent stroke can reflect younger age and lower frequencies of smoking, high cholesterol, hypertension and diabetes, rather than protection from the defect itself. ⁹

Information and measurement

Misclassification

Misclassification assigns exposure or disease status incorrectly because of imperfect sensitivity or specificity, missing data, recording errors, rounding or other measurement errors. ^{21, 22} Differential misclassification differs between comparison groups and can bias findings towards or away from the null. ²¹ Nondifferential misclassification affects groups equally; it may attenuate associations for binary variables, while misclassification of variables with several categories may bias away from the null. ^{21, 22}

Prospective exposure recording, objective biomarkers and self-administered scales can limit misclassification. ²² Biomarkers have limitations too: different detection limits or persistence in plasma can bias exposure comparisons. ²²

Expectations, recall and reporting

Knowledge of disease, exposure or treatment can influence how an interviewer helps respondents or an observer records information. ^{21, 22} Patients with disease or awareness of treatment assignment may search their memories more intensely than controls, combining recall, rumination and exposure-suspicion biases. ²² Records and objective exposure markers can help confirm recollections. ²²

Participant expectations can influence answers, while obsequiousness bias arises when investigators and participants produce the response the investigator desires. ²³ People may avoid embarrassing questions, or report their most frequent behaviour rather than their average behaviour, potentially understating average consumption. ²³ Detection of a disease in one person may prompt testing and further diagnoses among relatives, producing family-aggregation bias. ²³

Blinding observers or participants can reduce expectation and reporting biases, but conspicuous effects such as niacin flushing or a visibly wasted state may compromise masking. ^{21, 23, 22} Merely knowing that observation is taking place can change behaviour: the Hawthorne effect. ^{24, 22} The measurement setting can also change a result, as in white-coat hypertension. ^{21, 22}

Placebo by proxy occurs when a prescriber’s belief in treatment effectiveness is shared by a parent or animal owner, so both rate it effective even when the patient has not improved and cannot adequately communicate continuing symptoms. ²⁵

Surveillance and diagnostic suspicion

More frequent visits and assessments can increase detection without a corresponding increase in underlying disease occurrence. ^{12, 4} For example, more frequent visits and pelvic examinations among women prescribed oestrogen can raise detection of endometrial cancer. ⁴ Aspirin-associated melena can similarly increase suspicion and investigation of gastric cancer, raising the frequency of diagnosis. ⁴ An exposure may also produce manifestations that prompt investigation or mimic the disease being studied. ²⁶

Ways to minimise medical-surveillance bias include:

1. Stratifying the analysis by disease severity or the type or frequency of healthcare intervention.
2. Masking exposure status before ascertaining outcomes.
3. Selecting cases and controls from populations receiving similar treatment or diagnostic protocols.

⁴

Time and apparent change

Timing of exposure and follow-up

Lead-time bias adds apparent illness duration when diagnosis occurs during the latent phase of disease. ^{24, 22} Better diagnostic tests can move patients into more advanced disease stages, the Will Rogers phenomenon observed when comparing survival across times or clinics. ⁸

Immortal-time bias can arise when cohort definitions require an interval during which eligible participants cannot die or experience the outcome, or exclude patients who experience events before treatment. ²² Counting the interval from diagnosis to first treatment for treated patients can inflate their apparent survival relative to untreated patients. ²²

In protopathic bias, an early symptom of undiagnosed disease prompts treatment, which then appears associated with the disease it was prescribed to address. ^{24, 22} Excluding exposures within a specified interval before diagnosis through a lag-time analysis can help address this bias. ²⁴ Sick-quitter bias can weaken an exposure–disease association when illness prompts people to stop a risky behaviour and they are classified as unexposed. ^{24, 22} Cross-sectional associations may not establish whether the exposure or the outcome came first. ^{20, 22}

Regression to the mean and maturation

With imperfectly correlated repeated measurements, an extreme initial value may be followed by a value nearer the mean. ^{27, 23} Approaches to accounting for the effects of regression to the mean include random allocation to comparison groups, averaging several baseline measurements during a run-in, and an untreated natural-history arm. ²⁷

Regression dilution occurs when a single baseline measurement of a fluctuating risk factor understates its association with a later outcome. ^{27, 24} More observations, particularly during baseline screening, can reduce it. ²⁴ Maturation and practice effects can also account for before-and-after improvements, as in a hypothetical improvement in educational-test performance after three years of ADHD treatment when no comparable control has undergone the same testing. ²⁷ Imbalanced secular trends between groups can create differential maturation in group-randomised trials. ²⁸

Aggregate and individual associations

The ecological fallacy is drawing conclusions about individuals from analyses of groups or populations. ^{22, 23} Group-level associations may reflect misclassification, confounding, selection bias or effect modification within groups. ²³