Case-control studies

At a glance

• A case-control study samples on the outcome: people with the disease and a comparison group without it, then looks back at exposure.
• It is the efficient design for rare diseases and long latency, and can examine many exposures at once.
• The measure is the odds ratio (OR). It approximates the relative risk when the disease is rare.
• It cannot give incidence or absolute risk directly, and it is prone to recall bias and selection bias in the controls.
• Classic successes: smoking and lung cancer (Doll and Hill, 1950), diethylstilboestrol and vaginal adenocarcinoma (Herbst, 1971), aspirin and Reye syndrome.

Design

1. Define cases clearly, with a consistent diagnostic standard. Incident (new) cases are preferred over prevalent ones, which over-represent survivors.
2. Choose controls from the same population that produced the cases, the "study base". The test: if a control had developed the disease, would they have become a case in this study?
3. Measure exposure the same way in both groups, ideally with interviewers blind to case status or from records made before diagnosis.

Control sources each have a cost:

Control source	Advantage	Problem
Population (electoral roll, Medicare enrolment, random digit dialling)	Best match to study base	Low response rates, expensive
Hospital	Convenient, similar recall effort	Biased towards 1 if their illness is caused by the exposure; Berkson's bias from differential admission rates
Neighbours, friends, relatives	Match on social factors	May share exposures, causing overmatching

Matching on strong confounders (age, sex) improves efficiency. You then can't study the matched variable, and the analysis must keep the pairs (McNemar test, conditional logistic regression). Overmatching is matching on a factor linked to exposure but not an independent cause of the outcome. If the analysis keeps the matching, the main cost is lost efficiency. Matching on an intermediate on the causal pathway biases the OR towards 1, and no analysis can undo it.

Worked example: unmatched odds ratio

A study of venous thromboembolism in women aged 20 to 45 has 200 cases and 400 controls.

	Cases (VTE)	Controls
Combined oral contraceptive	120 (a)	120 (b)
No COC	80 (c)	280 (d)

• Odds of exposure in cases = 120 / 80 = 1.5. In controls = 120 / 280 = 0.43.
• OR = ad / bc = (120 × 280) / (120 × 80) = 33,600 / 9,600 = 3.5.
• VTE is rare in young women (around 1 to 10 per 10,000 per year), so the OR is a fair estimate of the RR.
• You can't say "120 of 240 COC users developed VTE". The 1:2 case-control ratio was chosen by the investigator. See Venous thromboembolism for the clinical side.

Worked example: matched odds ratio

One hundred cases of a drug-induced liver injury are each matched to one control by age and sex.

	Control exposed	Control unexposed
Case exposed	30	30
Case unexposed	10	30

• Concordant pairs (30 and 30) carry no information about the association.
• Matched OR = 30 / 10 = 3.0.
• Analysing the same data as if unmatched (cases 60 of 100 exposed, controls 40 of 100) gives (60 × 60) / (40 × 40) = 2.25. Ignoring the pairing is wrong, and because matching on a confounder makes exposure within pairs more alike, breaking the matching usually biases the OR towards 1.

Rare disease assumption

OR = (a/c) / (b/d). When the disease is rare, the number with disease is tiny relative to those without, so the odds of disease approximate the risk and the OR approximates the RR. When the outcome is common, the OR exaggerates the RR (further from 1).

With incidence-density sampling (each control chosen from people still at risk at the time a case occurs), the OR estimates the incidence rate ratio regardless of rarity. This is how most nested case-control studies within cohorts and linked databases are run.

Bias

Bias	Example	Usual direction
Recall bias	Mothers of babies with malformations recall more medicine use	Away from the null
Interviewer bias	Interviewer probes cases harder	Away from the null
Control-selection bias	Hospital controls with smoking-related diseases in a smoking study	Towards the null
Berkson's bias	Exposure and disease each change the chance of admission, so hospital patients show a spurious association (a form of collider bias)	Either way
Survival bias	Prevalent cases exclude rapidly fatal disease	Either way
Reverse causation	Early disease changes the exposure (e.g. weight loss before cancer diagnosis)	Either way

Reduce recall bias by using records made before diagnosis, structured questionnaires, or controls with a different disease that would prompt similar recall effort.

Variants

• Nested case-control: cases and controls are drawn from an existing cohort. Exposure data (stored blood, records) were collected before the outcome, which removes recall bias.
• Case-crossover: each case's exposure just before the event is compared with their own exposure in earlier control periods. Good for transient exposures (heavy exertion, anger, a new drug dose) and acute outcomes (MI, crash). Fixed confounders cancel out.
• Test-negative design: used for vaccine effectiveness. People with symptoms are tested; test-positives are cases and test-negatives are controls.

How it's tested

• Recognise the design: groups defined by disease status, looking back at exposure.
• Calculate an OR from a 2×2 table, or a matched OR from discordant pairs.
• Interpret the OR as an RR only if the disease is rare.
• Identify the bias: recall bias in a birth defect or cancer study, selection bias (often labelled Berkson's) in hospital controls.
• Pick the design: a rare disease with a suspected drug cause (case-control), versus a rare exposure with many outcomes (cohort; see Cohort studies).

Pitfalls

• Reporting incidence, prevalence or absolute risk from a case-control study.
• Using concordant pairs in a matched analysis.
• Treating the OR as an RR for a common outcome.
• Picking controls who couldn't have become cases, such as controls from a different region or health system.
• Placing case-control too high: on the NHMRC 2009 aetiology scale it is level III-3, below prospective (II) and retrospective (III-2) cohorts. NHMRC now advises guideline developers to use GRADE; the 2009 levels still appear in older guidelines and exam questions.