Skip to main content
Bureau of AI
Category: Fairness & Bias

Selection Bias

Simply put

Selection bias occurs when the individuals, groups, or data chosen for analysis are not representative of the broader population the analysis is meant to describe. Because the sample is skewed, conclusions drawn from it can be distorted and may not hold true for the intended population. In AI and model contexts, this typically arises when training or evaluation data is gathered in a way that systematically over- or under-represents certain cases.

Formal definition

Selection bias is a distortion in a measure of association (such as a risk ratio or odds ratio) that arises when the selection of individuals, groups, or data into a sample does not accurately reflect the target population, causing the estimated association between an exposure (or input) and an outcome to diverge from the true association in that population. It results from non-random or improper sample selection—including how a cohort is assembled or how records enter the analysis dataset—rather than from measurement error in the variables themselves. Note that selection bias is distinct from other bias categories (for example, measurement or fairness-related bias); the sources here frame it primarily in an epidemiological and study-design context, and its application to AI training and evaluation data is an extension of that concept rather than a separately defined term in the evidence provided.

Why it matters

Selection bias matters because conclusions drawn from a skewed sample can be systematically wrong even when the analysis itself is technically sound. If the data chosen for study does not represent the population the analysis is meant to describe, the resulting estimates of association can diverge from the truth, leading decision-makers to act on distorted evidence. In model contexts, this concern extends to training and evaluation data: when the records that enter a dataset over- or under-represent certain cases, a model may learn relationships that hold for the sample but not for the population it is deployed against.

For AI governance and model risk management, selection bias is a data-quality and validity concern that can undermine both model performance and the reliability of validation results. A validation exercise conducted on a non-representative evaluation set may overstate how well a model will generalize, giving false assurance to those relying on it. Because selection bias arises from how a sample is assembled rather than from measurement error in the variables themselves, it can be difficult to detect after the fact and may persist undetected through model development and testing.

It is worth noting that the evidence available here frames selection bias primarily in an epidemiological and study-design context, and its application to AI training and evaluation data is an extension of that concept rather than a separately codified AI term. Practitioners should also take care not to conflate selection bias with fairness-related bias or measurement bias; while these can interact, they are distinct categories and address different failure modes.

Who it's relevant to

Data Scientists and Model Developers
Those building models rely on training and evaluation data being representative of the deployment population. Understanding selection bias helps developers scrutinize how their datasets were assembled and recognize when systematic over- or under-representation of cases could distort learned associations, rather than assuming that a well-fit model on available data will generalize.
Model Validators and Model Risk Functions
Validation results can give false assurance if the evaluation set is itself non-representative. Independent validators typically assess whether the data used to develop and test a model reflects the intended population, since selection bias can inflate apparent performance and mask how a model will behave in production.
Auditors and Second/Third Line Reviewers
Reviewers assessing data provenance and analytic conclusions need to consider how samples were selected, because selection bias arises from the selection mechanism itself and may not be visible in the variables or measurements. Documenting how data qualified for inclusion is relevant to evaluating the validity of downstream conclusions.
Researchers and Analysts in Study-Design Settings
In epidemiological and study-design work—the context in which the sources primarily frame this concept—selection bias directly threatens the validity of estimated associations. Analysts must design cohort assembly and record inclusion carefully, as improper selection distorts measures of association independent of measurement quality.

Inside Selection Bias

Sampling Bias
A form of selection bias arising when the process used to draw training or evaluation data systematically over- or under-represents certain segments of the target population, so the sample does not reflect the population the model is intended to serve.
Coverage Gaps
Portions of the intended population that are absent or sparsely represented in the data, which can cause a model's behavior on those groups to be poorly estimated during development and validation.
Survivorship and Attrition Effects
Distortions that occur when only records that 'survived' some prior filtering (for example, approved applicants or retained customers) enter the dataset, excluding cases lost earlier in the process and skewing observed relationships.
Label Selection Effects
Bias introduced when outcomes or labels are observable only for a non-random subset of cases, so the labeled data available for training differs systematically from the full population of interest.
Population Drift Interaction
The interaction between how a sample was originally selected and later shifts in the deployment population, which can make an initially representative dataset less representative over time; relevant to ongoing model monitoring.

Common questions

Answers to the questions practitioners most commonly ask about Selection Bias.

Is selection bias the same thing as sampling bias?
Not exactly, and practitioners frequently conflate the two. Sampling bias is commonly treated as one form of selection bias, arising when the process used to draw a sample systematically over- or under-represents certain units. Selection bias is the broader category: it refers to any systematic distortion introduced by how observations enter (or are excluded from) the data or analysis, which can occur through sampling design, self-selection, attrition, survivorship effects, or filtering steps applied during data preparation. Treating the two as interchangeable can lead teams to check only their sampling procedure while overlooking other entry and exclusion mechanisms that produce comparable distortions.
Does selection bias mean the same thing as unfairness or discriminatory bias in an AI model?
No, and the distinction matters for how you address each. Selection bias is a data-provenance and representativeness issue concerning how observations were included or excluded, and it can affect model performance and generalization regardless of any protected attribute. Fairness-related or discriminatory bias concerns whether outcomes differ inappropriately across groups relative to a defined fairness objective. Selection bias can contribute to fairness problems—for example, if certain populations are systematically underrepresented—but it is a separate concept with distinct diagnostics and remedies. Fixing a fairness concern does not necessarily resolve selection bias, and vice versa.
How can a team identify whether selection bias is present in a dataset used for model development?
Common approaches include comparing the distribution of the development data against the target population or deployment population where a reference is available, documenting each step where records were included, filtered, or dropped, and examining attrition or missingness patterns to see whether they appear systematic rather than random. Teams also review the data collection or acquisition process for self-selection or survivorship effects. These methods can surface indicators of selection bias but typically cannot fully confirm its absence, since the units that were never observed are by definition not in the data; conclusions therefore often rely on assumptions that should be stated explicitly.
What mitigation options are available once selection bias is suspected?
Frequently discussed measures include revising the data collection or sampling design to better reflect the intended population, applying weighting or reweighting techniques to adjust for known over- or under-representation, collecting additional data to fill identified gaps, and constraining the model's intended use to the population the data actually represents. Each option carries assumptions and limitations—weighting depends on knowing the relevant distributions, and additional collection may not be feasible—so mitigations reduce or manage the effect rather than guarantee its elimination. The chosen approach should be documented alongside its residual limitations.
How should selection bias be documented for model validation and governance purposes?
Documentation typically covers the data's source and collection process, the inclusion and exclusion criteria applied, any known gaps between the development population and the intended deployment population, and the assumptions underlying any mitigation. This information commonly supports independent validation activities and the review carried out under a second line of defense. Framing it within the organization's model risk and governance processes helps reviewers assess how selection bias may affect the model's reliability in use; the depth of documentation expected can vary by sector and by the risk profile of the model.
How does selection bias relate to monitoring a model after deployment?
Selection bias can arise not only during development but also in the data a model encounters or generates in production, for example when the model's own decisions influence which cases are subsequently observed. Ongoing monitoring may compare production input distributions against the development data to detect divergence, though such monitoring is generally better suited to flagging distribution shifts and potential representativeness problems than to definitively proving a specific bias mechanism. Distinguishing selection bias from general performance degradation during monitoring is important, since they can have different causes and call for different responses.

Common misconceptions

Selection bias is the same thing as unfairness or discrimination in a model.
Selection bias is a data-generation and sampling problem that describes how a dataset departs from the target population. Fairness concerns how outcomes are distributed across groups and involves normative and, in some contexts, legal considerations. Selection bias can contribute to unfair outcomes, but the two concepts are distinct and should not be collapsed; a representative sample can still produce contested fairness results, and a non-representative sample does not automatically constitute unlawful discrimination.
A large dataset eliminates selection bias.
Sample size does not, on its own, correct for systematic selection effects. A large sample drawn through a biased selection process typically reproduces the same distortion at scale. Representativeness relative to the target population, not volume, is the relevant property, so more data can increase confidence in a biased estimate rather than reduce the bias.
Selection bias is only a concern at initial model development.
As commonly framed in model risk practice, selection effects can also emerge or worsen after deployment, for example when the population encountered in production diverges from the sampled data or when feedback loops shape which future observations are collected. This is why ongoing monitoring, rather than a one-time check, is typically appropriate.

Best practices

Document the data-generating and sampling process explicitly, including how records entered the dataset and which parts of the target population may be under- or unrepresented, so reviewers can assess coverage gaps.
Compare the distribution of the sample against the intended deployment population across relevant segments, rather than relying on sample size as evidence of representativeness.
Identify and record potential survivorship, attrition, and label-availability effects where outcomes are observable only for a filtered subset of cases.
Have model validation (as a second-line or independent function) test for selection-related weaknesses separately from overall performance metrics, since strong aggregate performance can mask poor coverage of specific groups.
Establish ongoing monitoring to detect divergence between the deployment population and the original sample over time, and to flag feedback loops that may reshape future data collection.
Treat controls that address selection bias as measures that reduce and manage the risk, not as steps that eliminate it, and record residual limitations and out-of-scope populations.