AI Glossary / Governance, ethics and safety / Intermediate

Selection And Sampling Bias

Definition

Selection And Sampling Bias is the combined label used in the Artificial Intelligence Ontology for distortion introduced by a selection process that fails to represent the relevant population or data. It draws attention to how observations enter a dataset, rather than assuming that the available records are a neutral picture of the intended use.

What it means

The label groups closely related ideas. Sampling concerns how units are drawn from a population; selection is broader and can include eligibility rules, voluntary participation, record availability, and filters applied after collection. Different disciplines draw the boundary differently, so the process causing the distortion should be described explicitly.
For AI projects, selection can occur repeatedly. A person must enter a service, generate a record, meet inclusion rules, receive a usable label, and survive cleaning before their data reaches training. An apparently large dataset may represent only people and situations that passed all of these gates.

How it works

Start with the population and decision the model is intended to support. Trace the collection and filtering pipeline, including exclusions that seem routine. Ask whether selection depends on characteristics related to the outcomes, features, or errors being studied.
Document the differences between the final dataset and the intended population. Where possible, redesign collection, retain appropriate missingness information, or obtain data for excluded conditions. Statistical adjustments require assumptions about the selection process and available information. If those assumptions cannot be supported, narrow the claims or change the intended use rather than hiding the gap behind a single performance score.

Three practical examples

• A model is trained only on completed service requests, omitting requests abandoned because the process was difficult.
• An evaluation removes every document with a missing field, even though missing fields are common in the planned deployment.
• A satisfaction analysis uses people who both opened an invitation and completed a survey, so the final sample reflects several selection stages.

Think of it this way

Think of a series of gates leading into a room. Looking around the room tells you about those who passed every gate, not automatically about everyone outside. The analogy has limits: selection can be probabilistic and invisible, and the effect depends on the question being asked rather than simply on the number of gates.

Worked example

A fictional service receives 1,000 requests. Of these, 800 are straightforward and 200 are complex. Suppose 90% of straightforward requests and 50% of complex requests reach completion. The completed-data set therefore contains 720 straightforward and 100 complex requests, or 820 total.
Complex requests make up 20% of all requests but only 100 divided by 820, approximately 12.2%, of completed records. A model evaluated only on completions faces a different mixture from the incoming workload. Weighting the completed records to 80:20 would correct this mixture only under additional assumptions about how completed and uncompleted cases compare within each type. It would not recover outcomes that were never observed, nor establish that abandoned complex requests resemble the completed ones.

Illustrative case study

At fictional Stonebridge Services, analysts prepare a training dataset by keeping only records with complete outcome fields. A review finds that missing outcomes are concentrated among cases transferred between departments. Those cases are central to the proposed routing assistant's purpose.
The team records the selection mechanism, investigates whether transferred cases can be followed up, and separates claims about completed records from claims about all incoming cases. The case illustrates a collection problem that ordinary train-test splitting would not resolve. It does not imply that every excluded record can or should be restored without checking its validity and permitted use.

Common misconceptions and mistakes

Do not interpret the combined label as a claim that all non-random samples are useless. Carefully designed non-random collection may answer a bounded question, and probability samples can still suffer from nonresponse or frame problems. What matters is whether the inference is justified.
Also distinguish selection bias from random sampling error and from inaccurate measurement within included records. These problems can coexist but require different responses. Equal group counts do not automatically remove selection mechanisms, and collecting more records through the same biased route may simply make a misleading estimate appear more precise.

Business use

The combined concept is useful during data acquisition, pilot review, and model validation. It prompts teams to inspect the entire path into the dataset, including business rules that data scientists may not control. A practical data record should describe eligibility, recruitment, response, label availability, exclusions, and intended inference. This gives reviewers a basis for judging whether the available evidence supports expansion.

People also ask

Why combine selection and sampling in one entry?

The source ontology groups them as related mechanisms affecting representativeness; the separate entries provide more specific perspectives.

Does a random train-test split solve the problem?

No. Both portions can inherit the same selection mechanism from the original dataset.

Can filtering improve data quality and introduce bias at the same time?

Yes. Removing unusable records may be reasonable, but if exclusions systematically remove important cases, the resulting scope must be examined.

Check your understanding

An AI team removes all unfinished applications before evaluating a tool intended to help people finish applications. What concern follows?

Answer: The evaluation excludes the very cases that may need the tool most. Results for completed applications cannot automatically establish performance for unfinished ones.

Related terms

• Sampling bias
• Selection bias
• Coverage bias
• Missing data
• Survivorship Bias
• External validity
• target population

Learn next

Create a simple flow count showing how many records remain after each inclusion rule. For every reduction, identify what is known about excluded cases and whether the rule changes the population represented. This turns a vague concern about biased data into specific questions about collection and inference.

More in Governance, ethics and safety

• Safety - Safety is the condition and practice of preventing unacceptable harm from a system's operation, failures, or interactions with its surroundings.
• Scalability - Scalability is a system's ability to handle growth in workload while meeting defined performance, quality, and resource requirements.
• Security - Security is the protection of systems and information against unauthorized access, use, alteration, disruption, or other harmful interference.
• Selective Adherence Bias - Selective Adherence Bias is the tendency to accept algorithmic advice more readily when it agrees with pre-existing beliefs or stereotypes, while...
• Self-healing system - A self-healing system detects certain faults and automatically takes predefined or policy-governed actions to restore acceptable operation.
• Semantic mapping - Semantic mapping connects items according to their meaning, rather than relying only on matching words or formats.

Sources

Sources: Artificial Intelligence Ontology, Selection And Sampling Bias, including its NIST reference: https://raw.githubusercontent.com/berkeleybop/artificial-intelligence-ontology/main/aio.owl . NIST SP 1270, bias terminology and lifecycle framework: https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=934464 . Google Machine Learning Glossary, selection bias: https://developers.google.com/machine-learning/glossary#selection-bias .

Previous: Selection Next: Selection bias