Here is a thorough explanation of matching in pairs in a case-control study, based directly on Park's Textbook of Preventive and Social Medicine:
Matching in a Case-Control Study
What is Matching?
Matching is the process of selecting controls so that they are comparable to the cases on one or more confounding variables (e.g., age, sex, occupation, smoking status). Its purpose is to neutralize the effect of these confounders so that the true association between the exposure and the disease can be measured without distortion.
For example:
- If investigating alcohol and oesophageal cancer, smoking is a confounder (it is associated with both alcohol use and cancer risk). Matching on smoking removes this distorting effect.
- If investigating steroid contraceptives and breast cancer, age is a confounder. Matching ensures both groups have the same age distribution.
Important rule: You must NOT match on the suspected aetiological factor you are investigating - doing so eliminates that factor's role from the study entirely.
Types of Matching
There are two main procedures:
| Type | Description |
|---|
| Group (Frequency) Matching | Cases are divided into subgroups (strata) by characteristics (age, occupation, social class), and controls are selected so their frequency distribution mirrors that of the cases |
| Individual (Pair) Matching | For each individual case, one control is chosen who matches that case as closely as possible on specific characteristics |
Matching in Pairs - Explained
Pair matching means selecting one control for every case, matched closely on key variables. The result is a series of case-control pairs.
Example from Park's textbook:
"If we have a 50-year-old mason with a particular disease, we will search for a 50-year-old mason without the disease as a control."
Each pair shares the same:
- Sex
- Age
- Occupation
- Duration and severity of illness (where applicable)
- Any other potential confounders
Practical Challenges of Pair Matching
- Difficulty finding matches - The more variables you match on, the harder it is to find a suitable control.
- Time-consuming - You may have to wait a considerable time to accumulate enough matched pairs.
- Over-matching risk - Excessive matching can make it impossible to find controls, and may artificially reduce the odds ratio (bias toward the null).
"If matching is overdone, it may be difficult to find controls. Further, with excess zeal in matching, there may be a tendency to reduce the odds ratio." - Park's Textbook of Preventive and Social Medicine
Analysis of Matched-Pair Data
When pairs are formed, the statistical analysis must account for the pairing. The standard methods are:
- McNemar's test - for 1:1 matched pairs (discordant pairs are analyzed, concordant pairs contribute no information)
- Mantel-Haenszel odds ratio estimator - for stratified matched sets
- Conditional logistic regression - for multivariable analysis with matched sets
The key concept in paired analysis is that only discordant pairs (where the case was exposed but control was not, or vice versa) provide information about the exposure-disease association.
Summary Table
| Feature | Group Matching | Pair Matching |
|---|
| Unit of selection | Group/strata | Individual pairs |
| Controls per case | Variable | Usually 1 (can be 1:n) |
| Analysis | Stratified analysis | McNemar's test / conditional logistic regression |
| Flexibility | More flexible | Stricter, harder to implement |
| Risk of over-matching | Lower | Higher |