Why Plan for Confounding?
When an observational study compares people who received different exposures or treatments, the groups may differ in other ways that matter to the outcome. For example, age could be related both to who receives a treatment and to the outcome being studied. A difference in outcomes could then reflect age as well as the treatment. Earlier material on confounding distinguishes such a variable from a mediator, which lies on a pathway from exposure to outcome. The methods here aim to address measured confounders, not mediators.
A study team can take steps before collecting outcomes to make the comparison more informative. It can restrict who is eligible, match people with similar characteristics, or plan to measure potential confounders so the analysis can account for them. These methods differ in how they use the data, but share a key limitation: they can address only characteristics that have been recognized and measured well enough to use.
Restriction: Make the Study Group More Comparable
Restriction is the most direct of these methods. If age is a potential confounder, a study might enroll only adults aged 40 to 49. Within that age range, age differences may be smaller than they would be in a study including people from adolescence through old age. Restriction does not make people identical, and a range can still contain meaningful age variation; it narrows the source of differences.
Restriction is useful when a clear eligibility rule is practical and the research question is about the restricted population. Its cost is that results may not apply to people who were excluded. It may also make recruitment harder, reduce the number of eligible participants, or prevent the study from assessing how outcomes differ across the restricted characteristic. Restriction on one variable also does not automatically balance other confounders.
Worked Example: Restricting Eligibility by Age
Suppose investigators want to study the association between a dietary supplement and a fatigue score. They are concerned that age is related to supplement use and fatigue. For this hypothetical study, they decide before recruitment to include only adults aged 40 to 49.
What restriction accomplishes. The study will not compare supplement users aged 40 to 49 with nonusers in a group that also includes many much younger or older adults. This reduces the range of age differences that could contribute to the comparison.
What it does not accomplish. The two groups could still differ in exact age within the decade, or in other factors such as underlying illness. Restriction therefore does not establish that age or other confounding has been eliminated.
Scope of the result. A result from this design directly describes the included age range. It cannot automatically be generalized to adults outside that range, because those adults were not studied.
Matching: Build Similar Comparisons
With matching, investigators identify comparison units that resemble one another on selected measured characteristics. In an observational study of an exposure, a team might select an unexposed person with the same age category and smoking status as each exposed person. The aim is to avoid a comparison in which one group has, for example, many more older participants or smokers than the other.
Matching can be done person by person or by making the overall distributions of characteristics similar across groups. It requires measuring the characteristics used for matching and finding suitable comparison units. If suitable matches cannot be found for some exposed people, those people may not contribute to the matched comparison. Matching also does not make treatment assignment random: people may still differ in characteristics that were not matched or measured. In addition, the analysis should respect how units were matched rather than treating a matched data set as if its selection process did not matter.
Worked Example: Matching on Age and Smoking
In a hypothetical observational study of an exposure and a later respiratory outcome, investigators consider age category and smoking status as possible confounders. They identify 30 exposed participants who are current smokers aged 50 to 59. Before outcomes are assessed, they seek 30 unexposed participants who are also current smokers aged 50 to 59.
Check the matching variables. Each selected exposed participant and each selected unexposed participant share the specified age category and smoking status. Thus, the groups are comparable on those characteristics by design, within the limits of the categories used.
Interpret the comparison carefully. If the exposed group later has a different outcome rate, that difference is not explained simply by one group containing more participants in a different age category or smoking-status category, because the groups were matched on those categories.
Recognize what remains uncertain. The groups could still differ in health history, intensity of smoking, or another unmeasured factor. Further, matching on broad age categories does not ensure equal exact ages. The procedure improves comparability on the specified measured characteristics; it does not guarantee that all confounding is gone.
Statistical Control: Measure and Account for Confounders
Statistical control is an analysis strategy that depends on data collection being planned well. Investigators must record the potential confounder accurately, along with exposure and outcome, so that comparisons can account for it. One introductory approach is to divide the data into strata, such as age categories, compare exposed and unexposed units within each stratum, and then combine those comparisons using the same weights for both groups.
Using common weights matters because the exposed and unexposed groups may have different distributions of the confounder. A crude overall rate averages within-category rates using each group’s own category proportions. If those proportions differ, the crude comparison mixes differences in outcomes within categories with differences in the groups’ category composition.
The following identity shows why group composition matters. Let \(r_{gj}\) be the outcome rate in group \(g\) and stratum \(j\), and let \(w_{gj}\) be the proportion of group \(g\) in stratum \(j\). The overall rate in group \(g\) is the sum of its stratum-specific rates, each weighted by that group’s own stratum proportion:
This follows by counting the outcome cases in each stratum, dividing by the group’s total size, and writing each stratum’s size as a proportion of that total. If the weights differ between groups, the overall rates need not compare like with like. Instead, choose common weights \(q_j\), with \(q_j \geq 0\) and \(\sum_j q_j=1\), and calculate standardized rates:
The standardized difference compares the groups using the same distribution across strata. This can clarify how the observed within-stratum differences combine for a chosen reference mix. It does not, by itself, turn an observational comparison into a causal effect.
Worked Example: Standardizing Outcome Rates by Age
In a hypothetical observational study, investigators record whether patients achieved blood-pressure control and whether they received a particular medication. They suspect age is a confounder, so they collect age category and plan to compare rates within younger and older patients. Suppose the data are:
| Age category | Medication group | Controlled | Total | Rate |
|---|---|---|---|---|
| Younger | Received medication | 50 | 100 | 50% |
| Younger | Did not receive medication | 20 | 50 | 40% |
| Older | Received medication | 30 | 100 | 30% |
| Older | Did not receive medication | 30 | 150 | 20% |
Check the concern about age. The medication group is 50% younger and 50% older. The group without medication is 25% younger and 75% older. Age category is therefore distributed differently across groups. The within-category control rates are also different: 50% versus 40% among younger patients, and 30% versus 20% among older patients.
Calculate the crude rates. Among patients who received medication, the overall rate is \((50+30)/(100+100)=80/200=40\%\). Among those who did not, it is \((20+30)/(50+150)=50/200=25\%\). The crude difference is \(40\%-25\%=15\) percentage points.
Choose common weights and standardize. Use an equal younger–older mix as the reference, so each age category receives weight \(0.5\). The standardized rate among patients who received medication is
The standardized rate among patients who did not receive medication is
The standardized difference is \(40\%-30\%=10\) percentage points. In this hypothetical data set, the age-standardized rate of blood-pressure control is 10 percentage points higher in the medication group under the selected equal age weights. This is an adjusted association, not proof that the medication caused the difference. The choice of equal weights defines the reference age mix; different weights could yield a different standardized comparison.
Assumptions and Limits of These Methods
Each method relies on careful choices made before or during data collection. Investigators must identify plausible confounders using subject-matter knowledge, define how they will be measured, and avoid adjusting for variables that are mediators when the goal is to estimate the total effect of an exposure. The methods also require enough overlap to make a meaningful comparison. For example, if no unexposed older participants exist, data cannot support a direct within-older comparison of exposed and unexposed participants without additional assumptions.
Restriction, matching, and statistical control address measured characteristics in different ways. Restriction narrows the study population. Matching selects comparable units. Statistical control retains variation in the data and adjusts the comparison using recorded covariates. Their common aim is to reduce differences in measured confounders between the groups being compared. None can guarantee that all relevant confounders have been identified or measured correctly. An unmeasured factor may still differ between groups and contribute to the observed outcome difference.
| Method | How it addresses measured confounding | Important limitation |
|---|---|---|
| Restriction | Studies only one level or a narrow range of a characteristic | Narrows the population to which results directly apply |
| Matching | Selects comparison units similar on chosen characteristics | Cannot balance characteristics that were not measured or matched |
| Statistical control | Accounts for recorded characteristics in the analysis | Depends on measurement, overlap, and appropriate analysis |
Common Mistake: Treating Adjustment as Proof
A common error is to say, “We adjusted for age, so age cannot confound the result and the remaining difference must be causal.” That reasoning overstates what statistical control establishes. An adjustment can address age as recorded and represented in the analysis, but it may leave residual age differences if categories are broad or measurements are inaccurate. It also does nothing by itself about an unmeasured confounder.
A more accurate conclusion is that the groups were compared after accounting for the measured age information under the chosen method. The result remains subject to possible unmeasured confounding and other limitations of the study. A careful plan improves the comparison; it does not remove the need to qualify conclusions.
Check Your Understanding
Try these questions before moving on.
- A study includes only adults aged 60 to 69 to study an exposure and an outcome. Which method is being used, and what limitation does it create for generalizing results?
- In an observational study, what is the purpose of matching exposed and unexposed people on smoking status?
- Why should investigators decide in advance which potential confounders to measure, even when statistical control occurs during analysis?
- Two groups have different age distributions. Explain why comparing age-specific outcome rates with common weights can answer a different question from comparing crude rates.
- Does matching on several measured confounders guarantee that no confounding remains? Explain why or why not.