Why the Unit Matters
A regression analysis does more than pair an outcome with one or more predictors. It also uses a set of records, and each record needs to represent something. Is one row a person, a household, a school, a clinic visit, or a measurement site? The answer determines what the analysis describes and what a regression conclusion can sensibly be about.
The unit question is easy to overlook when a data table contains variables measured at different levels. For example, a table might have one row per student and include the spending level of the student's school. The outcome is recorded for students, while one predictor describes schools. That combination can be useful, but it does not turn the rows into schools or make a school-level outcome into a student-level one.
In a straightforward individual-level data set, each row represents one individual, and the regression describes variation in an outcome from individual to individual. In a household-level data set, each row might instead summarize a household. The variables could still be about people, but the analysis would then describe differences between households, not differences between the individuals within them.
Read the Rows Before Interpreting the Regression
A useful way to state a regression target is to complete this sentence: “The outcome is recorded for each ___, and the analysis describes how that outcome varies across ___.” If both blanks name the same kind of entity, the observational unit and unit of analysis align. If they differ, ask whether the table really supports the intended conclusion.
For a simple fitted regression, the notation can make the unit explicit. Let \(Y_i\) be the outcome and \(X_i\) the predictor for analysis unit \(i\). A fitted line is written as:
Here, \(i\) labels the units represented by the rows used in the regression. The fitted value \(\hat{Y}_i\) is a predicted outcome for a unit with predictor value \(X_i\). The equation does not by itself determine what kind of entity that unit is: that comes from the data and the research question. If \(i\) indexes students, the fitted outcome is a student outcome. If it indexes schools, the fitted outcome is a school outcome.
Worked Example: A Regression About Individual Outcomes
Suppose a hypothetical study records sleep duration and a reaction-test score for each of 30 volunteers. One row contains the measurements for one volunteer. A fitted line from these data is \(\hat{Y}_i=18+2.4X_i\), where \(X_i\) is sleep in hours and \(\hat{Y}_i\) is the predicted reaction-test score in points. Identify the units and interpret a fitted value for someone who slept 6.5 hours.
Observational unit. Each row represents one volunteer, so the observational unit is an individual volunteer. The outcome, reaction-test score, is measured for each individual.
Unit of analysis. The regression compares individual volunteers’ scores across their sleep durations. Its unit of analysis is therefore also the individual volunteer.
Substitution. For a sleep duration of 6.5 hours:
The fitted line gives a predicted score of 33.6 points for an individual with 6.5 hours of sleep. Its slope describes a 2.4-point increase in fitted score per additional hour of sleep, on average along this fitted line. This is a description of individual-level outcomes; it is not a claim that sleeping longer causes a higher score.
When Predictors and Outcomes Are Measured at Different Levels
A predictor can describe a different level from the outcome. For example, a person-level outcome might be paired with a household characteristic shared by everyone in that household. In that case, the analysis may still have one row per person and an individual-level outcome. The household predictor gives context to each individual record, but does not change what the outcome represents.
This distinction matters when interpreting a coefficient. A fitted change in an individual outcome associated with a one-unit difference in a household predictor is still a comparison of individual outcomes across values of that shared predictor. It does not describe how an individual household member's outcome changes when the household characteristic is altered, nor is it automatically a conclusion about household-level outcomes.
Worked Example: Household Data Versus Individual Data
Imagine a hypothetical utility analysis with one row per household. The outcome is the household’s monthly electricity use in kilowatt-hours, and the predictor is the number of residents. The fitted line is \(\hat{Y}_j=310+55X_j\), where \(j\) indexes households. What is the unit of analysis, and what does the line predict for a four-person household?
Unit of analysis. Each row summarizes one household, and the outcome is household electricity use. The unit of analysis is the household, not each household member.
Substitution. For \(X_j=4\) residents:
The fitted line predicts 530 kWh of monthly electricity use for a household with four residents. Its slope is 55 kWh per month for each additional resident in the household, as described by this fitted line. It does not say that each person uses 55 kWh more, or predict a particular household member’s electricity use: neither of those is the outcome recorded in this data set.
Aggregation Changes the Unit—and the Question
A common way to change the observational unit is to aggregate several records into a summary. Suppose individual outcomes \(Y_{ij}\) are recorded for person \(i\) in school \(j\), and every student in school \(j\) is assigned the same school-level predictor \(X_j\), such as annual library spending per student. With one row per student, the outcome remains individual-level. If the data are summarized into one row per school, the outcome becomes a school summary, such as the mean student score.
This change is not merely a shorter way to display the same data. To see why, suppose the individual outcomes are represented by the equation \(Y_{ij}=b_0+b_1X_j+e_{ij}\), where \(e_{ij}\) is the difference between person \(i\)'s outcome and the value given by the equation. The school mean is found by summing the individual outcomes in school \(j\) and dividing by the number of students \(n_j\):
The last equality follows because \(b_0\) and \(b_1X_j\) are the same for every student in school \(j\), so their averages remain \(b_0\) and \(b_1X_j\); the average of the individual differences is \(\bar{e}_j\). After aggregation, each outcome is a school mean. A regression using one row per school therefore describes variation among school means, not the variation among individual students. The algebra shows how a school summary is formed; it does not make school-level conclusions interchangeable with individual-level conclusions.
Worked Example: From Student Rows to School Rows
Consider a hypothetical set of student scores from two schools. School A has scores of 70, 80, and 90 points; School B has scores of 60 and 70 points. A second data set is made by replacing the student rows with one row per school, recording each school’s mean score. Find the summaries and identify the units in both data sets.
Student-level data. The original five rows each represent one student, so the observational units are students. For School A, the mean is \((70+80+90)/3=240/3=80\) points. For School B, the mean is \((60+70)/2=130/2=65\) points.
School-level data. The aggregated table has two rows: one for School A with a mean score of 80 points, and one for School B with a mean score of 65 points. Each row now represents a school, and the outcome is a school mean. A regression on this table would describe how school means vary with school-level predictors.
The numerical summaries come from student scores, but the rows in the aggregated table are schools. A conclusion about schools with higher mean scores is not, by itself, a conclusion that every student in those schools has a higher score than every student in the other schools.
Site-Level Records and Repeated Measurements
The same reasoning applies to measurement sites. If a monitoring device records one value at each of 12 locations and each row gives a location’s seasonal average, the observational unit is the site. If the device records a value at each location every day, the rows represent site-days: each row is one site observed on one day. Treating those rows as though there were one row per site would misstate what was recorded.
Worked Example: Measurements at Monitoring Sites
Suppose a hypothetical air-monitoring table has three rows for each site, one for each sampled day. At Site North, the daily readings are 12, 15, and 18 micrograms per cubic meter. A second table contains one row per site, with the three readings averaged. Identify the observational unit in each table and calculate the summary for Site North.
Daily table. Each row represents one site on one sampled day, so the observational unit is a site-day. There are three records for Site North, not three different sites.
Site summary. The average for Site North is:
The summary table has one row for Site North, so its observational unit is the site, and its outcome is the site’s average reading across the three sampled days. A regression using daily readings describes site-day outcomes; a regression using these averages describes site-level averages. Those are different outcomes at different units of analysis.
Risks of a Mismatch
Mismatching the unit in the question with the unit in the data can produce a conclusion that sounds more specific than the analysis supports. It can also distort how much information the data contain. For example, if 40 people from 5 households are recorded individually, there are 40 person rows, but only 5 distinct households contributing household-level predictor values. Counting the same household characteristic repeatedly does not create 40 different household measurements.
The key risk is not that a regression with variables from different levels is always invalid. Rather, its interpretation must match the outcome and records, and the data structure must be described honestly. A regression with student outcomes and school predictors can address student-level outcomes across school contexts. A regression of school-average outcomes addresses differences among schools. Neither analysis automatically answers the other question.
A related error is drawing an individual-level conclusion from a group-level pattern. If schools with higher library spending have higher average scores, that alone does not establish that, within each school, students who use the library more have higher scores. The school and student are different units, and a relationship between school summaries need not describe the relationship between individuals.
Common Mistakes
- Wrong: “The predictor is a school characteristic, so this must be a school-level regression.” Correction: Check the outcome and the rows. If each row is a student and the outcome is a student score, the outcome remains individual-level even if a predictor is shared by students in the same school.
- Wrong: “There are 40 rows, so the data contain 40 separate household measurements.” Correction: If those rows represent people in 5 households, there are 40 person records but only 5 distinct households represented. Repeated household values do not become separate household observations.
- Wrong: “A relationship between school averages proves the same relationship among individual students.” Correction: School averages and individual outcomes are different outcomes at different levels. A group-level pattern alone does not establish an individual-level pattern.
- Wrong: “The daily reading table has one row per site because all readings come from sites.” Correction: If each row is a site on a particular day, the observational unit is a site-day. The entity measured and the complete unit represented by a row must be distinguished.
Check Your Understanding
For each question, identify the unit represented by the records and keep your interpretation at that level.
- A data set has one row per patient and records each patient’s outcome and the number of beds at that patient’s hospital. What is the observational unit, and what level is the outcome?
- A table has one row per household, with total monthly food spending as the outcome. What unit would a regression of that outcome describe?
- A school data set records one row per school, with the percentage of students absent at least once. Is the outcome an individual student’s absence status or a school summary? Explain.
- A sensor records five daily measurements at each of eight sites. What does one row represent in the unaggregated table?
- Why does a relationship between school-average test scores and school spending not, by itself, establish the same relationship among individual students?