Case-Control Studies

Dr Zo Fox

Institute of Neurology and the Research Department of Primary Care and Population Health, UCL

Epidemiology Lectures in Year 1

Measurement of disease (morbidity & mortality)

Prevalence/incidence/relative risk

Variations in disease

Standardised Mortality Ratios (SMRs)

Causation

Observational/experimental studies

Cohort studies

Objectives of todays lecture

Basic features of case-control studies

Measures of association between exposure and disease

Strengths and limitations

Sources of bias

Different purposes: descriptive (often routine data) versus analytical (often new data collected)

Observational designs:

Cross-sectional

Longitudinal (cohort)

Case-control

Ecological

Experimental/Interventional designs

Randomised trials

Epidemiological studies:

Observational study designs:

Observational studies

Cross-sectional

Cohort

Case-control

Type of analysis developed in the 1950s as an approach to the problem of investigating risk factors

for diseases with long latent periods and rare

diseases

An observational design in which individuals are selected on the basis of whether they do (cases) or do

not (controls) have the disease under investigation

Case-control studies

Case-control studies

Observational

Analytical

Retrospective

Ideal for rare diseases

Quick

Cheap

Can investigate aetiology

Basic steps in a case-control study

Definition of a case (symptoms; duration)

Selection of cases (patients with certain disease condition)

Definition of controls (subjects without the condition)

Selection of controls (hospital, community...)

Measurement of exposure

Comparing frequency of exposure in cases and controls

Cohort Start

Unexposed

Exposed

All healthy Follow-up (wait)

Disease

assessment

Controls

Cases

Start

Look Back

Case-Control

E+ E-

E+ E-

D- D+

D- D+

Disease

(cases)

No disease

(controls)

Exposed to factor

Not

exposed to factor

Exposed to factor

Not

exposed to factor

Compare cases

and controls

Past time Trace individuals

Starting point

A traditional case-control study

Study Population

Ineligible Eligible

Participation

Outcome: Yes (cases)

Exposed Unexposed

Outcome: No (controls)

Exposed Unexposed

No participation

Case-control study

Selection of cases (1)

Set of standardised criteria used to identify cases

Criteria should be precise and unambiguous and might be based on:

Clinical

Histological

Specific category of diagnosis (e.g. CDC AIDS definition, Code on death certificates, etc.)

Selection of cases (2)

Incident or prevalent cases

Incident cases: all new cases identified in the study population over a specific time period

Prevalent cases: cases within the study population who were already diagnosed with the disease at the time of

designing the analysis

Selection of controls (1)

Must not have the outcome in question

Should provide an estimate of the level of exposure in those without the outcome

Controls are selected from the same population group from which the cases are drawn

General population (random digit dialling, registers/voting lists, etc)

People using the same health service (hospital controls)

Friends/relatives/neighbours of cases

Selection of controls (2)

More than one control per case?

Multiple control groups of different types?

Matching?

Cases Controls Representative? Bias

Patients admitted Controls more Spurious

to Hospital with HIV likely to have reduction

Patients low CD4+/CD8+ of OR

with Hodgkin

lymphoma

in Hospital

Patients admitted Controls more Spurious

to Hospital for renal likely to have increase

transplantation high CD4+/CD8+ of OR

Selection bias from a poor choice of

controls: looking at CD4+/CD8+ ratio in NHL

Example: Exposure to armadillos and leprosy Indian J Dermatol Venereol Leprol 2008;74:338-42, Deps PD, Alves BL et al.

Objective: To see if exposure to armadillos is associated with leprosy

Setting: Case-control study of 506 individuals with leprosy and 594

individuals without leprosy in four clinics that participated in the national

leprosy control programme in Brazil. Controls were patients with other

chronic diseases from these clinics

Risk factor of interest: Exposure to armadillos

Outcome: Leprosy

Follow-up: Followed back in time from a diagnosis of leprosy

Conclusions: Direct exposure to armadillos was reported by 68% of

leprosy cases and by 48% of controls. Exposure to armadillos was

significantly higher among those with leprosy compared to those without

leprosy

No Leprosy

N=594

Leprosy

N=506

Start

Look Back

Case-Control

285

(48%)

160

(32%)

309

(52%)

346

(68%)

Armadillo exposure in cases and controls (%)

Pe

rce

nta

ge

(%

)

Odds ratio of exposure

An approximation of relative risk (risk ratio)

A good estimate of the risk ratio when the disease is rare

How much more likely are the cases to have been exposed as compared to the controls?

So, in this example we want to know how much more likely

is exposure to armadillos among those with leprosy

compared to those without leprosy

Odds ratio = Odds of exposure among controls

Odds of exposure among cases

Calculating the odds ratio:

Probability of being

exposed among cases

Probability of NOT being

exposed among cases

Odds of exposure among cases =

Odds of exposure

among cases a

= c

a/(a+c)

= c/(a+c)

Disease +

(Cases)

Disease -

(Controls)

Exposure + a b

Exposure - c d

Calculating the odds ratio:

Odds Ratio = a/c ad

b/d bc =

Odds of exposure among cases = a/c

Odds of exposure among controls = b/d

Disease +

(Cases)

Disease -

(Controls)

Exposure + a b

Exposure - c d

Calculating the odds ratio:

Odds ratio for leprosy according to direct

armadillo exposure:

Leprosy Cases

Controls

Direct exposure to armadillos

346

285

No exposure to armadillos

160

309

Odds ratio = 346 x 309 / 285 x 160 = 2.34

Interpretation of OR

People who have direct contact with armadillos have a 2.34 times higher odds of contracting leprosy than those

with indirect or no exposure to armadillos

Those with indirect or no exposure have less than half the odds (i.e. 1/2.34 = 0.43) of contracting leprosy compared

to those with direct exposure

You can convert the odds back to probabilities to work out the probability of getting the outcome depending on

numerous factors

Confounding (1)

Contact with

armadillos Leprosy

BCG vaccination

Confounding (2)

No real relationship between contact with armadillos and risk of leprosy

The relationship may be confounded by factors such as having a BCG vaccination

Having a BCG vaccination is a confounder

BCG vaccination needs to be adjusted for (controlled for) in the analysis or patients need to be matched on BCG

vaccination

Matching

Matching is done to increase the power of the study and to control for confounding factors

It is important to restrict matching to potential confounding factors and not to match patients on the

exposure under investigation

It is not possible to evaluate the association between matching factors and outcome

Matched case-control studies

Match on factors that are likely to be common causes for the exposure and the disease

It may be impractical to match patients on some factors (e.g. if there are many strata, it is difficult to find suitable

controls)

Conditional logistic regression required

More than one control can be selected for each case

Example: matching in the study of

armadillos and leprosy Risk of leprosy depends on education and BCG vaccination;

BCG vaccination less likely to get leprosy

Worse education more likely to get leprosy

Matching cases and controls on education level and BCG

vaccination will eliminate confounding by these factors

For each case [no education; BCG = Y] that is recruited, one or

more controls [no education; BCG = Y] is recruited

For each case [no education; BCG = N] that is recruited, one or

more controls [no education; BCG = N] is recruited

Other ways to control for confounding

Adjustment in analysis:

Stratified analysis (e.g. within different education levels no education, minimal education,

Example: MMR and autism Lancet 2004;364;9438:963-69, Smeeth L, Cook C et al.

Objective: To see whether MMR vaccination is associated with risk of autism or other pervasive developmental disorders (PDD)

Setting: Case-control study of 1294 individuals born >1973 with a

diagnosis of pervasive developmental disorders (PDD) while registered

with a GP between 1987-2001. Controls were 4469 individuals with no

diagnosis of PDD recorded matched on age, sex, and GP to cases.

Risk factor of interest: Exposure to MMR vaccination

Outcome: Autism diagnosis

Conclusions: MMR vaccination was not associated with an increased

risk of PDD. The odds ratio for association between MMR and pervasive

developmental disorder was 086 (95% CI: 068109)

Selection bias:

It may be difficult to find appropriate controls

Measurement bias (recall bias/observer bias):

The retrospective measurement of exposure may be

problematic:

Inaccurate - will underestimate any relation

Reporting bias - overestimation

Reverse causation - spurious association

Case-control study: Bias

Reverse causality example:

Low physical activity is related to higher risk of hip fracture in

a case control study

But people with hip fracture (or history of hip fracture) have

lower physical activity

This can lead to a spurious association - temporality must be

clarified

Case-control study: Bias

Cases may be more likely to remember events that

occurred around the time in which they were diagnosed

with the disease (recall bias)

Example: cigarette smoking and lung cancer; people with lung cancer may, consciously or not, tend to exaggerate their

level of exposure to smoking if they believe that cancer was

caused by smoking

Assessment of exposure blind to case/control status is

not always possible (observer bias)

Case-control study: Bias

Multiple choice questions (1):

A) Case-control

B) RCT

C) Cohort study

D) None of the above

What types of study can you perform to identify predictors of

struggling (i.e. academic difficulties, course termination or

failure of 3 preclinical exams) during medical training?

Multiple choice questions (2):

Overall, 123 students were identified as strugglers and 492 as controls. There were 56/123 (45.5%) strugglers and 130/492 (26.4%) controls who had negative comments on

their UCAS form. What is the estimated odds ratio of a

negative UCAS comment for strugglers vs. controls?

A) -1.3

B) 1.7

C) 2.3

D) 6.0

E) None of the above, because odds ratios cannot be

estimated from a case-control study.

Multiple choice questions (3):

A Chi-squared test gave a p-value of

Adjusted odds ratio

In this study, there was an OR (95% CI) of 2.33 (1.55 to 3.50)

After adjustment for male gender, the odds ratio marginally decreases to 2.25 (1.44 to 3.50). Why?

Case-control study nested in cohort/RCT

Using an existing cohort study/RCT

Cases: patients who developed the disease

Controls: a random sample of patients without the disease from the cohort easier to find relevant matches

Controls are chosen among those at risk of becoming cases

Cost saving by collecting data on a sample of controls instead of the entire cohort at risk

Rationale: to reduce cost with laboratory measurements

Advantage: no reporting / measurement bias because data have been recorded as part of the cohort study

Nested case-control study

Quick (cases already exist, no need to wait)

Cheap (not necessary to examine large number of people)

Can examine many exposures

Suitable to study rare diseases or diseases where there is a long latency between exposure and manifestation

Suitable to study stable exposures (e.g. genetic markers)

Sufficient power achievable with relatively small sample size

Case-control study: strengths

Selection of cases/controls can be troublesome

Not suitable for rare exposure

Cannot calculate incidence risk or death rates

Prone to bias (selection bias, recall bias, observer bias)

In controls, exposure to risk factors and confounders should be

representative of that in the population "at risk" of becoming cases

Exposures of controls should be measurable with similar accuracy

to those of cases

Limited to a single outcome variable

It is not possible to validate responses

Correct interpretation of results is difficult

Case-control study: limitations

It is not usually possible to establish a sequence of events so you are not likely to be able to estimate incidence rates

It is not possible to estimate the risk or risk ratios because we have artificially selected the number of cases relative to

the number of controls.

You can calculate:

Conditional Odds (matched analysis)

Odds of having the exposure and Odds ratios

Statistical Considerations

Conclusions

Case-control studies are valuable for studying disease aetiology

Like other designs, a case-control study has unique strengths and limitations that must be considered when selecting this

particular strategy

More susceptible to bias than other designs; potential sources of bias should ideally be recognised in the design phase in order

to minimise their effects

Could be used as a quick/cheap first step to identify potential risk factors for a disease