Find the Third Variable That Connects to Both
In “Observational Studies and Confounding,” you learned that a potential confounding variable is related to both the explanatory variable and the response. This tutorial focuses on applying that idea when a regression scenario describes several possible influences but does not label one as a confounder.
Start by identifying the two variables in the regression. The explanatory variable is the one used to predict or explain the response; the response is the outcome being predicted. Then examine the other variables in the scenario. A strong candidate is not simply one that sounds important. The scenario should give a plausible reason that it is connected to each of the two regression variables.
The word “plausible” matters. A short scenario may not provide enough evidence to establish exactly how a third variable affected the observed data. Your job is to identify a reasonable candidate and make both links explicit. Avoid presenting a guess as a demonstrated fact.
A Two-Link Check for Each Candidate
For each possible third variable, ask two separate questions. First, why might it vary with the explanatory variable? Second, why might it be related to the response? If you cannot explain both links in context, the candidate is not yet a convincing confounder choice.
State what the explanatory variable \(x\) measures and what the response variable \(y\) measures, including the individuals or cases when the scenario says.
Look for a characteristic, condition, or group difference in the scenario that could be related to the values of \(x\).
Describe why cases with different \(x\)-values might also differ in the candidate variable.
Describe why the candidate variable could also be associated with \(y\), the response.
Call it a potential or plausible confounder unless the scenario provides stronger evidence. Do not claim that it explains the entire association or proves causation.
The two links do not have to be described with a calculated correlation or a numerical model. A reasonable contextual mechanism is often enough for this identification task. For instance, a variable may be related to \(x\) because people select different activities, or because locations differ in resources. It may be related to \(y\) because it influences access, opportunity, condition, or behavior relevant to the outcome.
A variable can be named in a scenario without being a useful confounder candidate. If it is not plausibly related to one of the regression variables, explain why the connection is missing rather than assuming that every measured characteristic matters. Also be alert to variables that are actually part of how \(x\) or \(y\) is defined. A restatement of the response is not a distinct third variable.
Worked Examples: Make Both Connections Explicit
Worked Example: Park Access and Weekly Physical Activity
A fictional city planning group records data for 120 neighborhoods. The explanatory variable \(x\) is the number of public parks per square kilometer, and the response \(y\) is the mean number of minutes of recreational physical activity per resident per week. The overall regression shows a positive association: neighborhoods with more parks tend to have higher mean activity. The scenario also notes that neighborhood income differs across the city.
Identify the variables. The cases are neighborhoods, \(x\) measures park density, and \(y\) measures mean weekly recreational activity in minutes per resident. Neighborhood income is a possible third variable.
Explain the link to park density. If neighborhoods with greater income tend to have more resources for purchasing, maintaining, or advocating for public parks, income could be related to the number of parks per square kilometer. This is a plausible connection, not a fact established for every city or every neighborhood.
Explain the link to activity. Income could also be related to residents’ recreational activity through differences in available time, transportation, recreation programs, or access to other exercise facilities. Those are possible pathways by which income and mean activity might vary together.
Give a careful conclusion. Neighborhood income is a plausible confounding variable because it could be related to both park density and mean recreational activity. The positive regression association alone does not show that adding parks causes residents to exercise more; differences in income or other neighborhood characteristics could contribute to the pattern. The scenario does not establish that income actually accounts for the association or that it is the only potential confounder.
Worked Example: Daily Temperature and Ice-Cream Sales
A fictional shop records data for 90 days. Its regression uses daily high temperature \(x\), in degrees Celsius, to predict ice-cream sales \(y\), in dollars per day. The shop reports that sales are generally higher on warmer days. Thirty of the recorded days are in the warm season and 60 are in the cool season; the shop also has more customers outdoors during the warm season.
Identify the variables. The cases are days. Temperature is the explanatory variable, and daily sales in dollars are the response. Season is a distinct third variable mentioned in the scenario.
Explain the link to temperature. Season is related to the daily temperature distribution: warm-season days tend to have higher temperatures than cool-season days. Thus, days in different seasons can have systematically different values of \(x\).
Explain the link to sales. Season may also be related to ice-cream sales for reasons beyond that day’s temperature. The scenario says more customers are outdoors during the warm season, which could increase opportunities to buy ice cream. That gives a contextual reason for season to be associated with \(y\) as well.
Give a careful conclusion. Season is a plausible confounder of the temperature–sales association because it is related to temperature and could also be related to sales through outdoor customer traffic. This does not show that temperature has no association with sales, nor that the regression association is entirely due to season. It means the pooled pattern may combine a temperature relationship with differences between seasons.
Check the tempting alternative. The shop’s daily sales target might be mentioned in a longer scenario, but it would be a weaker candidate unless there were a reason for the target to be related to both temperature and actual sales. A variable’s presence in the story is not enough; the two-link check still applies.
Worked Example: Vehicle Mileage and Annual Repair Cost
A fictional service center examines 80 used cars. The explanatory variable \(x\) is a car’s odometer reading in thousands of kilometers, and the response \(y\) is its repair cost in dollars over the next year. Cars with higher mileage tend to have higher annual repair costs. The records also include each car’s age in years.
Identify the variables. The cases are used cars. Mileage is \(x\), annual repair cost is \(y\), and vehicle age is a possible third variable.
Explain the link to mileage. Older cars have generally had more time to accumulate kilometers, so age could be associated with odometer reading. A car’s mileage also depends on how it was used, so age and mileage need not match perfectly; the connection is plausible rather than automatic.
Explain the link to repair cost. Age could be associated with annual repair cost because components may wear or deteriorate over time, even among cars with similar odometer readings. That gives a reason for age to be related to the response as well as to mileage.
Give a careful conclusion. Vehicle age is a plausible confounder because it could be associated with both mileage and annual repair cost. The regression line describes the observed relationship among these cars, but it does not isolate the relationship between mileage and cost from possible age differences. This observational record does not establish that mileage itself causes higher repair costs.
Consider a detail that would help. Comparing cars of similar ages could help show whether a mileage–cost association is also present within age groups. Such comparisons can clarify the pattern, but they would not automatically account for every other difference among cars.
How to Choose Between Several Plausible Candidates
Some scenarios include more than one possible confounder. It is acceptable to name a candidate supported by the information given; you do not need to claim it is the only possible one. Choose the variable for which you can clearly explain both connections, rather than listing many possibilities without explanation.
A useful way to organize the reasoning is to say what the candidate variable does in the scenario. Does it differ across people, days, places, or groups with different \(x\)-values? Could it also be related to the measured outcome \(y\)? A sentence about each relationship makes the logic visible to the reader.
The phrase “may contribute” is appropriately cautious. It recognizes that a confounder could help produce or alter the overall association without asserting that the variable has been measured well, that its effect is known, or that it accounts for the whole pattern.
As discussed in “Association Versus Causation in Regression,” a regression association does not by itself show that changing the explanatory variable would change the response. Identifying a plausible confounder strengthens the reason for caution, but it does not prove that the explanatory variable has no relationship with the response. Nor does identifying one potential confounder rule out other explanations.
Common Mistakes and AP Exam Tip
- Naming a variable but explaining only one link. State why it could vary with the explanatory variable and why it could also be related to the response. Full credit requires both connections.
- Calling any extra variable a confounder. A variable that is merely recorded or mentioned is not automatically a confounder. Tie it to both regression variables using details or a reasonable mechanism from the scenario.
- Claiming the confounder has been proven. If the scenario only makes the relationship plausible, say “potential” or “plausible confounder,” not “the cause of the association.”
- Saying the confounder explains everything. Unless the scenario provides strong evidence for that claim, say it could contribute to the pattern. Other variables or a relationship between \(x\) and \(y\) may also be involved.
- Switching the roles of the regression variables. Clearly name which variable predicts and which is the response, then explain the candidate’s links to those specific variables.
- Making a causal claim from an observational regression. An observed association, even with a plausible explanation, does not show that changing \(x\) causes a change in \(y\).
A complete AP-style response is concise but specific: name the candidate, connect it separately to both variables in context, and use cautious language about what the scenario establishes. “Season may be a confounder” is too brief by itself. A stronger answer explains that season is associated with daily temperature and may also affect sales through changes in outdoor customer traffic.
Check Your Understanding
For each situation, name a plausible confounder if the scenario supports one, and explain how it could be linked to both regression variables.
- A researcher records the number of hours employees work remotely and their job-satisfaction scores. Employees’ job roles differ. What two links would you need to explain before calling job role a plausible confounder?
- A garden center uses weekly rainfall to predict plant sales. The records cover spring and autumn, when the types of plants stocked also differ. Explain how season could be related to both variables.
- A coach studies the association between athletes’ weekly training hours and their race times. Athletes have different levels of prior experience. Describe two plausible links involving experience.
- Why is “the study also recorded participants’ favorite color” not, by itself, a sufficient explanation of a confounder?
- Complete the response frame for one of the worked examples, using “could be related” and “may contribute” appropriately.