Why Match Participants?
Suppose a health program is offered to some participants, while others receive usual care. If the two groups differ at the start—for example, in age or initial symptom severity—a difference in their later outcomes could reflect those initial differences as well as the program. Matching is a design strategy: before comparing outcomes, it forms sets of participants who are similar on selected characteristics measured before treatment or exposure.
In one-to-one matching, each participant in one group is paired with one participant in the other group. Often the first group consists of participants who received a treatment, and the second consists of eligible participants who did not. The aim is to make the paired participants alike on specified baseline characteristics, not to make their outcomes alike.
For example, age at enrollment and a symptom score recorded before a program begins might be matching variables. A symptom score recorded twelve weeks later might be an outcome. The measurement’s role depends on when it was taken and how it is used: a baseline score can be used to select matches, while a later score is not a baseline characteristic.
Planning a One-to-One Match
Matching starts with a decision about which baseline characteristics are important for the comparison. The variables should be available for participants in both groups and measured in a comparable way. A plan might require an exact match on a category, such as smoking status, and then choose the closest available participant by age. Alternatively, it might combine several quantitative characteristics into a distance score.
The rule should be set without using the outcome values. If researchers choose pairs because their later outcomes happen to be similar, the resulting comparison is selected using the very information it is meant to study. Matching is also not a substitute for careful study design: it only directly matches on the characteristics measured and included in the rule.
Identify which participants are to be matched and which participants may serve as their comparison partners.
State which pretreatment characteristics will be used, and define how similarity will be judged for each one.
Form one-to-one sets according to the stated rule. If controls cannot be reused, account for that restriction when assigning matches.
After the matched sets are formed, use the outcome measurements for the planned comparison; do not use them to decide who matches whom.
A match can be exact on a categorical characteristic, meaning that both participants have the same category. For a quantitative characteristic, a common simple rule is to choose the eligible participant with the smallest absolute difference. If there are several quantitative characteristics, their units matter: a difference of five years and a difference of five millimeters of mercury are not directly comparable just because both differences equal five.
Worked Example: Matching Exactly on Smoking Status and Closely on Age
A hypothetical clinic compares participants who enroll in a new support program with participants receiving usual care. The clinic plans to match each program participant to one usual-care participant, without reusing anyone. It requires the same baseline smoking status and, among eligible participants, chooses the closest age.
| Participant | Group | Age at enrollment (years) | Baseline smoking status |
|---|---|---|---|
| T1 | Program | 40 | Not currently smoking |
| T2 | Program | 52 | Currently smoking |
| T3 | Program | 45 | Not currently smoking |
| C1 | Usual care | 42 | Not currently smoking |
| C2 | Usual care | 50 | Currently smoking |
| C3 | Usual care | 47 | Not currently smoking |
| C4 | Usual care | 39 | Currently smoking |
For T1, only C1 and C3 have the required smoking status. Their age differences are \( |40-42|=2 \) years and \( |40-47|=7 \) years, so T1 matches to C1. For T2, the eligible controls are C2 and C4, both currently smoking. Their age differences are \( |52-50|=2 \) years and \( |52-39|=13 \) years, so T2 matches to C2. T3 must match to a remaining nonsmoking control: C1 is already used, so T3 matches to C3. C4 remains unmatched.
The resulting sets are T1 with C1, T2 with C2, and T3 with C3. Age and baseline smoking status are the matching variables. If the clinic later records a twelve-week symptom score, that score is an outcome; it did not determine these assignments.
Using More Than One Quantitative Characteristic
When matching on multiple quantitative characteristics, one way to define closeness is to standardize each characteristic first. For a characteristic \(x\), a difference of \(x_T-x_C\) can be expressed in standard-deviation units by dividing it by a chosen scale \(s_x\), such as the standard deviation of that characteristic in the eligible participants. The scale must be positive and defined consistently for the comparison.
A simple distance for two characteristics, \(x\) and \(y\), is the square root of the sum of the squared standardized differences. This is a defined matching score, not a universal rule; other distances and rules are possible. The point of standardizing is that the score combines differences measured in comparable units rather than letting a variable’s numerical scale dominate automatically.
This distance treats a one-standard-deviation difference in either characteristic as the same size of difference in the score. It does not mean the two characteristics are equally important for the study question; that is a substantive design choice. A distance rule should therefore be described, rather than presented as if the data determine a uniquely correct definition of “similar.”
Worked Example: Comparing Candidates on Two Baseline Measurements
In a hypothetical study, a program participant is 50 years old and has a baseline systolic blood pressure of 140 mmHg. Two eligible usual-care participants are available. For illustration, the matching plan uses age and systolic pressure, with scales \(s_{\text{age}}=10\) years and \(s_{\text{BP}}=20\) mmHg.
| Candidate | Age (years) | Baseline systolic pressure (mmHg) |
|---|---|---|
| C1 | 48 | 145 |
| C2 | 52 | 138 |
For C1, the standardized differences are \((50-48)/10=0.20\) for age and \((140-145)/20=-0.25\) for pressure. Thus,
For C2, they are \((50-52)/10=-0.20\) and \((140-138)/20=0.10\). Thus,
Under this particular rule, C2 is the closer match because \(0.2236<0.3202\). The scores do not describe a difference in blood pressure or an outcome effect; they summarize pretreatment similarity on the two characteristics included in this matching rule.
When Matches Cannot Be Chosen Independently
If each comparison participant may be used only once, selecting the closest available control for each treated participant in isolation can lead to a poor overall assignment. A sensible assignment considers the possible matches across the whole group. The following example uses distances already calculated from a stated baseline-characteristic rule.
Worked Example: Assigning Two Controls Without Reuse
Two hypothetical program participants, T1 and T2, are to be matched to controls C1 and C2. The matching plan prohibits reusing a control. Distances are smaller for more similar participants:
| C1 | C2 | |
|---|---|---|
| T1 | 1 | 4 |
| T2 | 2 | 100 |
If T1 is assigned its individually closest control, C1, T2 must then be assigned C2. The total distance is \(1+100=101\). The alternative assignment pairs T1 with C2 and T2 with C1, giving a total distance of \(4+2=6\). Since \(6<101\), the alternative has the smaller total distance and is preferable under a rule that minimizes the sum of distances across the two sets.
This illustrates why “choose the nearest control” needs a precise assignment rule when controls cannot be reused. The rule might minimize total distance or use another prespecified criterion. Different rules can produce different matches, so the method should be reported clearly.
What Matching Can and Cannot Establish
Matching is a way to construct a comparison group on observed baseline characteristics. It does not assign treatment at random, and it does not guarantee that matched participants are alike in every relevant respect. For example, participants may differ on an important characteristic that was not measured or was not included in the matching rule. Therefore, matching alone does not prove that a difference in outcomes was caused by treatment.
A causal interpretation requires more than finding similar-looking pairs. In particular, the measured baseline variables used for matching must be relevant to comparability, and the comparison group must contain eligible participants sufficiently similar to the treated participants under the chosen rule. If suitable comparison participants are unavailable for some treated participants, those participants may remain unmatched; the comparison then pertains to the matched participants rather than automatically to everyone originally considered.
The timing of measurement is essential. A matching variable must precede treatment or exposure. A variable measured afterward may have been affected by treatment, so using it to select matches can make the groups comparable on something that is itself part of the treatment’s consequences. Keep the matching rule grounded in pretreatment information and reserve the post-treatment measurements for outcomes.
Common Mistake
Wrong reasoning: “The program group and control group should be matched on their twelve-week symptom scores, because those scores are important to the study.” This uses the outcome to determine which participants are compared. It can obscure or distort the outcome difference the study is intended to assess.
Correction: Match using relevant characteristics measured before the program begins, such as age or baseline symptom score. Then compare the twelve-week symptom scores as outcomes, without letting those scores change the matching assignments.
Check Your Understanding
Answer each question using the definitions and examples above.
- A study records age and symptom severity before treatment, then records symptom severity eight weeks later. Which measurements could be matching variables, and which measurement is the outcome?
- For a treated participant aged 60, two eligible controls are aged 58 and 65. If age is the only matching variable and the rule chooses the smallest absolute age difference, which control is selected? Show the differences.
- Why can measurements in years and millimeters of mercury not simply be combined as raw differences in a distance score without specifying how their scales are handled?
- Two treated participants are to be matched without reusing controls. Why might assigning each treated participant their individually closest control fail to minimize the total distance?
- Does matching on several measured baseline characteristics by itself establish that treatment caused a later outcome difference? Explain one reason it may not.