Selecting Cases and Controls

Overview

The credibility of a case-control study depends less on whether cases and controls look similar than on whether they represent the same underlying population. Cases should capture eligible outcomes arising from a defined study base. Controls should be sampled independently of exposure from people who would have been eligible and detected as cases had they developed the outcome.[1–3]

Selection therefore requires a linked set of decisions: outcome definition, incident versus prevalent case status, ascertainment system, geographic and temporal boundaries, control source, index dates, eligibility criteria, participation pathways and replacement rules. These decisions should be made before exposure data are examined.

This lecture provides a practical framework for selecting cases and controls and for recognising when apparently convenient groups create selection bias.

Why this matters

A perfectly measured exposure cannot rescue an inappropriate comparison group. For example, women with breast cancer treated at a national referral centre may come from several states, while controls recruited from nearby primary-care clinics may represent only local residents. Differences in socioeconomic status or reproductive history may then reflect referral and recruitment pathways rather than disease occurrence.

Selection can also distort results when:

• severe or rapidly fatal cases are missed;
• prevalent cases must survive long enough to enrol;
• participation depends on exposure-related beliefs;
• hospital controls have conditions caused by the exposure;
• controls are screened more intensively than the source population; or
• case and control eligibility rules are applied at different reference times.

Key concepts

Case definition and ascertainment

A case definition should include clinical, laboratory, imaging or pathological criteria as appropriate. Broad criteria may include false-positive cases; narrow criteria may omit genuine cases and reduce applicability. The protocol should define uncertain diagnoses and whether adjudication is required.

Case ascertainment may be:

• population registry based;
• facility or network based;
• laboratory based;
• disease registry based; or
• identified within a defined cohort.

Completeness matters. If exposure affects where or whether a case is diagnosed, incomplete ascertainment can produce selection bias.

Incident and prevalent cases

Incident cases are newly occurring outcomes during the study period. They usually support clearer time order. Prevalent cases already have the condition when sampled and must have survived with it. A factor associated with survival, recurrence or care access may therefore appear associated with being a prevalent case even if it did not cause disease onset.

Prevalent cases can be useful for rare chronic conditions or exploratory work, but the target inference should be stated honestly.

The control eligibility principle

Control eligibility should mirror case eligibility except for the outcome. If cases must have lived in the district for two years, controls should satisfy the same condition at their reference date. If cases are restricted to people with accessible maternity records, applying that restriction only after control recruitment creates an unfair comparison.

Sources of controls

Population controls

Population registers, household sampling frames, electoral rolls or random-digit methods can approximate the geographic study base. Advantages include fewer disease-related reasons for selection. Challenges include incomplete frames, low response, cost and uncertain healthcare catchment.

Neighbourhood controls

Neighbourhood controls may share environmental and social characteristics, but field selection can be subjective and overmatching may reduce exposure contrast. Interviewers must not choose the most approachable neighbour.

Hospital or clinic controls

Facility controls can be efficient and may have similar recall, catchment and willingness to participate. However, their conditions must not be strongly caused by the exposure under study. Using women with diabetic complications as controls for an exposure related to diabetes would distort exposure prevalence.[2,3]

Friend or relative controls

These controls can improve participation but may share behaviours, genetics and social environments with cases. This resemblance can overmatch the exposure or related determinants and complicate independence.

Registry or database controls

Administrative databases can identify large samples and support identical record extraction. Their validity depends on coverage, eligibility coding, outcome exclusion and the timing of available information.

DECISION GUIDE

Evaluating a proposed control source

• Does the source cover the population that generated the cases?
• Would members have been detected as cases if the outcome occurred?
• Is selection independent of the exposure after accounting for design variables?
• Are eligibility and index-date rules comparable with cases?
• Can exposure be measured using the same data source and look-back period?
• Are non-participation and missing-record mechanisms likely to differ?

Research question and methodological fit

The question is unchanged: whether previous gestational diabetes is associated with first myocardial infarction among eligible women in the source population. Case and control selection should preserve this population contrast.

The design should explicitly name the estimand. With risk-set sampling, the comparison estimates an incidence rate ratio. With controls sampled from those outcome-free at the end of follow-up, the odds ratio estimates the exposure odds ratio and may approximate a risk ratio only under appropriate conditions. Selection and analysis must be planned together.[1]

Design, setting and study population

DESIGN PROFILE

Case and control selection within one study base

Core structure: Cases are ascertained among all eligible outcomes, while controls are sampled from the population-time experience capable of producing those cases.

Best suited for: Efficient etiological comparisons when the source population, outcome pathway and control frame can be described.

Important limitations: Referral, diagnosis, survival, participation and record availability can each alter inclusion probabilities.

The case series should document all participating sites, referral patterns, diagnostic capacity and the dates each site contributes. If hospital catchments overlap, duplicate case detection must be prevented.

Controls require an index date. In individually matched or risk-set designs, this may be the matched case’s date. In cumulative designs, it may be the end of the risk period. The date determines age, eligibility, disease-free status and exposure look-back.

Sampling and recruitment

PROTOCOL CHECKLIST

Case-selection protocol

• State the outcome definition and diagnostic evidence required.
• Specify incident, recurrent or prevalent status.
• Define the source population, ascertainment period and site coverage.
• Include fatal, severe and mild cases where consistent with the target outcome.
• Apply eligibility before reviewing exposure history.
• Record eligible, approached, consenting, excluded and analysed cases with reasons.

PROTOCOL CHECKLIST

Control-selection protocol

• Name the control frame and justify its relationship to the study base.
• Apply the same geographic, age, calendar and other eligibility rules.
• Define outcome-free or at-risk status at the reference date.
• Select controls without knowledge of exposure.
• Prespecify sampling ratios, replacement rules and repeated eligibility.
• Record sampling probabilities and participation flow.

Replacement of a non-participating control with an easily available person can change the sampling mechanism. A replacement should be drawn using the same documented procedure, and original non-response must still be reported.

Variables, constructs and measurements

Comparing candidate control sources

illustrative

Table 1. Appraisal of candidate controls

Candidate source	Relationship to study base	Main advantage	Main concern
Electoral-register residents	Covers eligible district residents	Population-based frame	Non-response and outdated addresses
General outpatient attendees	Partial hospital catchment	Efficient recruitment and record access	Attendance may relate to exposure and comorbidity
Hospital employees	Employment-selected subgroup	Easy access	Poor representation of age, occupation and healthcare access

Note: The final choice requires local evidence about hospital catchments and frame coverage.

Variables describing selection should be retained: site, date identified, eligibility reason, approach status, consent, exclusion after enrolment and data completeness. These variables permit flow reporting and limited assessment of participation differences.

Procedures and data collection

Cases and controls should receive equivalent explanations, consent processes and exposure instruments. Recruitment scripts should avoid revealing the preferred hypothesis. Interview timing after the case event should be standardised because acute illness can affect recall and participation.

A pilot can test whether the proposed control frame produces eligible participants, whether cases and controls have comparable historical records, and whether response differs by key characteristics. Pilot findings may reveal that the nominal source population does not match actual referral pathways.

Bias, validity and quality assurance

Selection bias is best described as a mechanism rather than a label. Investigators should draw a pathway from target population to case detection, control-frame inclusion, approach, consent and complete analysis. At each transition, ask whether probability depends on exposure, outcome or common causes.[4]

APPRAISAL CHECKLIST

Appraising case and control selection

• Does the case definition identify the intended outcome consistently?
• Are cases incident and sufficiently complete for the target inference?
• Is the control source connected to the same population and time period?
• Could exposure influence admission, registry inclusion or willingness to participate?
• Are index dates and exposure windows comparable?
• Are participation numbers and exclusions shown separately by group?

Ethical and governance considerations

Interface with statistical analysis

The analysis must reflect the sampling design. Unmatched sampling may use logistic regression with prespecified confounder adjustment. Individual matching requires matched-set identifiers and conditional analysis. Frequency-matched variables should generally enter the model. Sampling fractions may matter in specialised case-cohort or two-phase designs.

Report crude and adjusted odds ratios with 95% confidence intervals, reference categories and the control-sampling strategy. Do not estimate population incidence from the sampled case and control counts.

Multiple control groupsMMed extension

Two control groups from different sources can reveal whether findings are sensitive to one selection mechanism, but they do not automatically eliminate bias. Each group must independently represent a defensible study base, and discordant results require investigation of selection, measurement and confounding rather than choosing the preferred estimate.[1,3]

Worked example

WORKED EXAMPLE

Choosing controls for the myocardial-infarction study

The hospital network reviews referral data and confirms that most eligible myocardial infarctions among district residents are treated at the four sites. The electoral register covers the same districts and permits age-stratified random sampling. Hospital employees are rejected because employment, occupational class and healthcare access differ substantially from the study base.

For each case, four population controls are randomly sampled from women alive, resident in the district and without prior myocardial infarction on the case’s index date. The same age range, residency duration and maternity-record requirements apply. Non-responders remain in the recruitment denominator, and replacements are drawn by the original random procedure.

The team also compares basic register characteristics of participants and non-participants where legally permitted. This does not prove absence of selection bias, but it makes the recruitment pathway more transparent.

Suggested protocol or reporting wording

Common mistakes

Knowledge check

Which candidate control source is automatically best: population, neighbourhood or hospital controls?

Answer and reasoning

None is automatically best. The appropriate source is the one that represents the population-time experience that generated the cases and permits comparable exposure assessment. Local referral, coverage and participation mechanisms determine suitability.

Why are prevalent cases vulnerable to survivor bias?

Answer and reasoning

They must survive from disease onset until recruitment. If exposure affects survival, recurrence or care access, exposure prevalence among prevalent cases can differ from that among all incident cases.

Should controls be screened to exclude every medical condition?

Answer and reasoning

No. Controls need to be eligible members of the study base without the index outcome at the relevant time. Excluding unrelated conditions can create an unusually healthy group and alter exposure prevalence.

A non-participating population control is replaced by the next clinic attendee. What is wrong?

Answer and reasoning

The replacement comes from a different sampling mechanism. It may have exposure-related healthcare attendance and no longer represents the original population frame. Any replacement should follow the prespecified random procedure.

Key take-home messages

SUMMARY

Key take-home messages

• Case definition, ascertainment and control selection must be designed as one system.
• Incident cases usually provide clearer etiological interpretation than prevalent survivors.
• Controls should be eligible to become cases under the same detection pathway.
• Comparable index dates, eligibility rules and exposure windows are essential.
• Population and hospital controls each have strengths and distinct selection mechanisms.
• Recruitment and non-participation should be documented separately for cases and controls.

Transparent reporting of participant selection should follow the case-control recommendations in STROBE.[5,6]

References

1. Keogh RH, Cox DR. Case-control studies. Cambridge: Cambridge University Press; 2014. https://doi.org/10.1017/CBO9781139094757
2. Wacholder S, McLaughlin JK, Silverman DT, Mandel JS. Selection of controls in case-control studies. I. Principles. Am J Epidemiol. 1992;135(9):1019–1028. https://doi.org/10.1093/oxfordjournals.aje.a116396
3. Wacholder S, Silverman DT, McLaughlin JK, Mandel JS. Selection of controls in case-control studies. II. Types of controls. Am J Epidemiol. 1992;135(9):1029–1041. https://doi.org/10.1093/oxfordjournals.aje.a116397
4. Hernán MA, Hernández-Díaz S, Robins JM. A structural approach to selection bias. Epidemiology. 2004;15(5):615–625. https://doi.org/10.1097/01.ede.0000135174.63482.43
5. von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP; STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. PLoS Med. 2007;4(10):e296. https://doi.org/10.1371/journal.pmed.0040296
6. Vandenbroucke JP, von Elm E, Altman DG, Gøtzsche PC, Mulrow CD, Pocock SJ, et al. Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): explanation and elaboration. PLoS Med. 2007;4(10):e297. https://doi.org/10.1371/journal.pmed.0040297