From a Table Pattern to a Causal Claim
A two-way table can show that an outcome occurs at different rates in groups defined by another variable. That is evidence of an association: the observed distribution of one variable differs across categories of the other. But a difference in observed rates does not, by itself, show that changing one variable would change the outcome.
This distinction matters whenever people choose their own exposure, treatment, or behavior. Students who attend tutoring, patients who exercise regularly, and people who answer a survey in a particular way may already differ from those who do not. A table describes the groups as observed; it does not automatically make them comparable in all other respects.
For clarity, call the variable whose groups are being compared the exposure and call the measured result the outcome. These labels do not imply that the exposure caused the outcome. They simply help organize the question. For example, “tutoring” can be the exposure and “passed the course” the outcome, even when students were not randomly assigned to tutoring.
What the Conditional Proportions Tell Us
For a binary outcome, a useful summary is the outcome proportion within each exposure group. The denominator for each proportion is the total number in that exposure group. The difference between these proportions describes the observed gap between the groups; it is not automatically the effect of the exposure.
Worked Example: Tutoring and Course Pass Rates
Suppose a hypothetical campus registrar summarizes course outcomes for two groups of students who chose whether to attend tutoring. The table gives the number who passed or did not pass.
| Passed | Did not pass | Total | |
|---|---|---|---|
| Attended tutoring | 72 | 28 | 100 |
| Did not attend tutoring | 60 | 40 | 100 |
The pass proportion among students who attended tutoring is \(72/100=0.72\), or 72%. Among students who did not attend, it is \(60/100=0.60\), or 60%. The observed difference is
In this hypothetical data set, the pass rate is 12 percentage points higher among students who attended tutoring. This is a correct description of the table. It does not show that tutoring raised the pass rate by 12 percentage points: students chose whether to attend, and the groups may have differed before tutoring began.
The distinction can be stated precisely. Let \(E\) denote the exposure group and \(Y\) the outcome. The table gives quantities such as \(P(Y=1\mid E=1)\): the proportion with the outcome among those observed in exposure group 1. A causal question instead concerns what would happen to outcomes if exposure were changed. Those are different questions, and the table directly answers only the first.
Why a Third Variable Can Change the Story
A confounding variable is a variable that is associated with the exposure and is also related to the outcome, so that differences in the confounder can help explain the observed exposure-outcome association. The name does not mean that every third variable related to either variable is a confounder. Its timing and role matter: a variable caused by the exposure, for example, is not automatically a confounder.
One way to see why the group comparison can be misleading is to divide the population into categories of a third variable \(C\). The overall outcome proportion in exposure group \(e\) is a weighted average of the proportions within those categories:
This follows by splitting the event \(Y=1\), within group \(E=e\), into the mutually exclusive cases \(C=c\), then adding their probabilities. The weights \(P(C=c\mid E=e)\) describe the mix of categories in each exposure group. If those mixes differ, the overall proportions can differ partly because the groups contain different kinds of people. Thus an overall association does not, on its own, isolate the relationship between exposure and outcome within comparable groups.
This reasoning does not prove that a particular third variable explains an association. It shows why a plausible alternative explanation must be considered. Even after comparing groups within categories of measured variables, other differences may remain, and a variable chosen for adjustment may not be a genuine confounder.
Worked Example: Exercise and High Blood Pressure
Imagine a hypothetical health survey of 400 adults. Of 200 who report regular exercise, 30 report high blood pressure. Of 200 who do not report regular exercise, 60 report high blood pressure.
The observed proportions are \(30/200=0.15\), or 15%, for regular exercisers and \(60/200=0.30\), or 30%, for those not reporting regular exercise. The difference is \(0.15-0.30=-0.15\), or 15 percentage points lower in the regular-exercise group.
Age is a plausible confounder to investigate: age may be related to exercise habits and to the prevalence of high blood pressure. If the two exercise groups have different age distributions, part of the overall difference could reflect that imbalance. The table alone does not tell us how much, if any, of the difference is due to age, nor does it establish that exercise caused a reduction in blood pressure. Other explanations could also matter, including how exercise and blood pressure were measured and whether people with health concerns changed their activity.
Other Explanations to Consider
Confounding is important, but it is not the only reason an observational association may fail to represent a causal effect. The order of events may be unclear: an outcome can influence the exposure, a possibility called reverse causation. People may enter or remain in the data differently depending on exposure or outcome, creating selection bias. Misclassification or inaccurate self-reports can also distort the table. In addition, an association observed in a sample can be affected by sampling variation.
These possibilities are reasons to be cautious, not automatic proof that the association is false. A well-designed observational study can provide valuable evidence, and causal conclusions may be supported by a broader body of evidence and appropriate assumptions. The central point is narrower: the observed table pattern alone does not distinguish a causal explanation from these alternatives.
Worked Example: Campus Residence and Club Participation
A hypothetical end-of-term survey asks students whether they live on campus and whether they participated in at least one student club that term. Among 120 students living on campus, 80 report club participation; among 100 students living off campus, 35 report participation.
The participation proportion for on-campus students is \(80/120=0.667\), or about 66.7%. For off-campus students it is \(35/100=0.350\), or 35.0%. The observed difference is \(0.667-0.350=0.317\), about 31.7 percentage points.
A careful conclusion is: “In this hypothetical survey, reported club participation was more common among students living on campus.” The table does not establish that moving onto campus would increase a student’s participation. Year of study is a plausible confounder to consider if, for example, first-year students are more likely to live on campus and to take part in orientation-related activities. The survey also records residence and participation at the same time, so it may not establish which came first.
Use Language That Matches the Evidence
The safest reporting separates the numerical result from its interpretation. First state the groups, the outcome, and the observed proportions or difference. Then describe the result as an association, and avoid turning it into a claim about what would happen if exposure changed.
| Overstated wording | More accurate wording |
|---|---|
| “Tutoring increased the pass rate by 12 percentage points.” | “The pass rate was 12 percentage points higher among students who attended tutoring.” |
| “Exercise prevented high blood pressure.” | “High blood pressure was less common among respondents reporting regular exercise.” |
| “Living on campus made students join clubs.” | “Reported club participation was more common among on-campus students in this survey.” |
The more accurate statements do not claim that no causal effect exists. They simply report what the observational table supports without adding an unsupported causal explanation. When a causal interpretation is important, ask what design and evidence would justify it, and which alternative explanations remain plausible.
Common Mistake
Wrong reasoning: “The outcome proportion is higher in the exposed group, so the exposure caused the outcome.” This treats an observed group difference as if exposure were the only difference between the groups.
Correction: First identify the association and its size using the correct group denominators. Then ask whether the groups may differ in confounders, whether the outcome could have influenced exposure, whether selection or measurement could affect the results, and whether the study design supports a causal interpretation. If the data are observational and those questions are unresolved, report an association rather than a causal effect.
Check Your Understanding
For each question, distinguish what the table can describe from what it can establish causally.
- A hypothetical table shows that 40% of students who use a study app pass an exam, compared with 30% who do not. Write one accurate association statement and one causal statement the table alone cannot justify.
- In a survey, people who report more weekly exercise also report lower rates of a health condition. Name one plausible confounding variable and explain how it could relate to both variables.
- Why is it important to use the total number in each exposure group as the denominator when comparing outcome proportions?
- A survey records campus residence and club participation at the same time. Explain why the timing makes a causal interpretation difficult.
- Give one example of reverse causation or selection bias that could contribute to an observed association.