Methods to reduce bias should match the mechanism that could distort a study. Random assignment, allocation concealment, blinding, good measurement and transparent reporting address different problems. None is a general guarantee of unbiased results.
This CPH study guide brings together study-design, measurement, analysis and reporting strategies. Use it with types of bias in epidemiology to explain why a method is appropriate, not merely name it.
Random assignment is not random sampling
Random assignment allocates enrolled participants to interventions by chance. It prevents prognosis from systematically determining assignment and balances measured and unmeasured baseline factors in expectation. A particular trial can still have chance imbalances, especially with a small sample.
Random sampling instead concerns who enters the sample. Randomizing volunteers does not make those volunteers representative of a wider population. It also does not prevent later missing outcomes, unequal measurement or selective reporting.
Allocation concealment comes before assignment
Concealment prevents recruiters from anticipating the next allocation before enrollment is settled. An unpredictable sequence is not enough if staff can see forthcoming assignments and selectively enroll people. Concealment differs from blinding, which concerns knowledge of the assigned intervention afterward. See Cochrane’s trial risk-of-bias guidance.
Blinding and consistent measurement
Blinding can reduce the influence of expectations on behavior, reporting and outcome assessment. Specify who is blinded: participants, clinicians, outcome assessors or analysts. Labels such as “double-blind” can be ambiguous without that detail. See the CONSORT 2025 explanation of trial reporting.
Some interventions cannot conceal participation. That does not automatically make the evidence unusable; investigators can still consider blinded outcome assessment and other safeguards. Blinding reduces particular risks rather than proving that all bias has been removed.
Use clear case definitions, comparable questioning, calibrated instruments and training. Validate tools for the intended use and population. Standardizing data collection is different from statistically standardizing rates.
Recruitment and retention
Choose a comparison group appropriate to the source population and research question. Document eligibility, nonresponse, exclusions and losses to follow-up. Plan practical retention procedures before outcomes are missing.
Weighting and missing-data methods can sometimes address selection problems under assumptions about how data became missing. They do not recover information automatically or repair an unknown selection mechanism by adding more covariates.
Matching, stratification and adjustment
Matching, stratification and regression can help address measured confounding when the design and analysis are appropriate. Matching does not automatically remove selection bias or eliminate confounding; matched designs need an analysis that accounts for the matching.
Choose adjustment variables using the research question, subject knowledge and causal structure. Adding every available variable can be harmful when it includes mediators or variables affected by selection. See the worked examples in confounding and effect modification.
Propensity scores require overlap and assumptions
A propensity score describes the estimated probability of receiving an exposure or intervention given measured baseline covariates. Matching or weighting based on it does not address unmeasured confounding automatically.
Check covariate balance and whether relevant groups have adequate overlap. Restriction or trimming may be appropriate in a particular analysis, but it changes the population or effect being estimated and should be justified and reported. Do not apply a universal rule to trim all extreme scores. See Austin and Stuart’s propensity-score weighting guidance.
Direct and indirect standardization
Direct age standardization compares rates using a common set of age weights. It addresses differences in age composition; it does not remove all other explanations for a difference. Use the same age categories and standard population when comparing standardized rates. See NCHS age-adjustment methods.
Hypothetical calculation: if two age groups have rates of 20 and 100 events per 100,000 and standard weights of 0.60 and 0.40, the directly standardized rate is:
(0.60 × 20) + (0.40 × 100) = 52 events per 100,000.
This is a weighted comparison measure, not the population’s observed crude rate or an individual’s probability.
Indirect standardization applies reference rates to the study population to obtain an expected number of events. For mortality, the standardized mortality ratio is observed deaths divided by expected deaths. In a fictional example, 100 observed deaths and 80 expected deaths give SMR = 100 / 80 = 1.25: 25% more deaths than expected under those reference rates. This is not, by itself, a causal effect of an occupational exposure.
Indirect methods are useful when study-specific age rates are unstable or unavailable, provided the necessary population information and reference rates exist. SMRs from different populations need caution because their effective weighting can differ. See NCHS’s technical discussion of direct and indirect standardization.
Registration and complete reporting
Prespecify hypotheses, outcomes and analysis plans when appropriate, register prospective studies, and report departures from the plan. Report findings irrespective of direction or statistical significance. These steps improve transparency; registration alone neither prevents every reporting problem nor guarantees publication.
Evidence syntheses should search appropriate registries and other sources, investigate missing results, and critically appraise unpublished material. Gray literature is not limited to government datasets and is not automatically validated. See Cochrane’s missing-evidence guidance.
Practice questions
Question 1: intervention assignment
Investigators want baseline prognosis not to systematically determine which intervention enrolled participants receive. Which method most directly addresses assignment?
- Stratification after the study
- Blinding outcome assessors
- Random assignment with allocation concealment
- Publishing only statistically significant results
Answer and explanation
C) Random assignment with allocation concealment. Random assignment balances prognosis in expectation; concealment prevents recruiters from anticipating allocations. Neither makes the recruited sample representative or removes every later source of bias.
Question 2: matching
A case-control study selects controls matched to cases on age. Has the study automatically eliminated age confounding?
Answer and explanation
No. Matching is a design step. The analysis must account for the matching and its assumptions; the design does not eliminate other confounding or selection problems.
Question 3: assessors
Participants know which exercise program they attend, but outcome assessors do not know their assigned program. Which safeguard is present?
Answer and explanation
Blinded outcome assessment. This can reduce assessment influenced by knowledge of assignment. It does not imply that participants or every investigator are blinded.
Question 4: the reference comparison
A study observes 100 deaths where reference rates predict 80. Calculate the SMR and state one limit of its interpretation.
Answer and explanation
SMR = 1.25. Observed deaths are 25% above the reference expectation. The ratio alone does not show why the difference occurred or establish that a particular exposure caused it.
Question 5: prespecification
A trial registers its outcomes before recruitment, then leaves an unfavorable registered outcome out of the report. Has registration prevented selective reporting?
Answer and explanation
No. Registration provides a record against which reporting can be checked. Complete outcome reporting is still required.
Sources and related reading
- Cochrane Handbook: risk of bias in a randomized trial
- Cochrane Handbook: missing evidence
- NCHS: age adjustment
- NCHS: age standardization of death rates
- Age adjustment: comparing rates across populations
Editorial update — October 11, 2026: consolidated the original and expanded bias-reduction guides; corrected claims about randomization, matching, standardization and registration; and revised practice explanations. Numerical examples are hypothetical.
Humanities Moment
The featured image for this article is “‘Adam and Eve’ Tavern, Old Chelsea” by James Abbott McNeill Whistler. James Abbott McNeill Whistler was an American-born painter and printmaker who championed Aestheticism and the credo “art for art’s sake,” favoring refined tonal harmonies and subdued palettes that emphasized composition and mood over narrative or moralizing content. He helped popularize tonalism and modern portraiture in late 19th-century Europe, advanced the revival of etching, and as a theorist and polemicist influenced the move toward modern art in Britain and beyond.
