Units, Treatments, and Outcomes in Causal Studies

Why Define the Study Before Comparing Groups?

A causal question is not fully specified by asking whether a treatment “works.” We need to know for whom the question is being asked, which treatment and comparator are being contrasted, and which outcome at what time is meant by “works.” Different choices can produce different, equally legitimate questions. A study of a medication’s effect on symptom scores after four weeks does not answer exactly the same question as a study of its effect on hospitalizations after one year.

These choices are not merely labels added after data collection. They define what the study is about and how its results can be interpreted. A numerical comparison is meaningful only in relation to the units, treatment assignment, outcome, and follow-up time that produced it. Even a carefully calculated difference cannot answer a causal question that was never clearly defined.

Units: Who or What Is the Question About?

The study unit is often an individual person, but it can also be a household, clinic, school, or community. The unit should be named in a way that matches the research question. “Patients” is too vague if the intended units are adults with a particular diagnosis who begin treatment during a defined period. Naming the units precisely helps clarify both the population to which the question pertains and what one row of an analysis represents.

The study unit need not be the same as the unit that receives an assigned intervention or the unit on which a measurement is first recorded. For example, a program might be assigned to clinics, while patient outcomes are measured and the question concerns patients. These are different roles: the clinic is the assignment unit, and the patient is the outcome-recording and study unit. If observations are summarized at the clinic level, the analysis unit may differ too. Such differences matter because a comparison of clinics is not automatically the same as a comparison of patients.

State the units concretely: “adults newly diagnosed at participating clinics during January through June” is more informative than “people.” Also state whether a unit is eligible before assignment, selected after assignment, or defined by some other rule. These details determine which units the question refers to; they do not, by themselves, guarantee that a comparison between groups has a causal interpretation.

Treatment Assignment: Name Both Sides of the Contrast

A treatment is an intervention or strategy whose effect is of interest. Treatment assignment describes which strategy a unit is assigned to receive. A useful specification names the treatment and the comparator in parallel, such as “assigned to receive the new counseling program” versus “assigned to receive usual services.” Saying only that a study compares “treated” with “untreated” can conceal what either group actually receives.

The distinction between assignment and receipt is important even at the design-description stage. A participant assigned to a program may not attend every session; someone in the comparator group may obtain similar services elsewhere. Those facts do not change the original assignment, though they may matter to other questions about treatment received. Here, the basic causal question should make clear whether it concerns assignment to strategies or actual receipt of an intervention.

A treatment label can also be too broad to tell us what is being compared. “Medication” might mean different drugs, doses, schedules, or durations. For the question to be interpretable, specify enough about the strategy and its comparator to distinguish them in the study. More detailed issues about treatment versions are addressed separately; the immediate task is to avoid treating an unspecified category as if it described one clear intervention.

Outcomes Need a Measure and a Time

An outcome is not simply “health” or “success.” It is a variable recorded for each unit, such as a symptom score, whether an infection occurred, or the number of days in hospital. A good description states the measurement scale and, where relevant, how the variable is constructed. For a binary outcome, identify the event counted as “yes.” For a score, identify what the scale represents and its possible range when that information is available.

The time of measurement is part of the outcome definition. A symptom score at four weeks and the same score at twelve weeks are distinct endpoints. For events, specify the period during which the event is counted, such as hospitalization within 90 days after assignment. Without a time origin and follow-up window, “hospitalized” might refer to very different periods for different units. Time should be anchored to a clearly stated event, commonly assignment or treatment initiation.

A study can measure several outcomes at several times. Each is a separate question, not a interchangeable substitute for the others. Selecting one outcome and time after seeing which comparison looks most favorable can make the reported result difficult to interpret. A clear question states the outcome and timing in advance, and reports other measured outcomes as additional questions rather than silently replacing the original endpoint.

Worked Comparisons: The Choices Change the Question

A Short Justification: Why the Three Choices Must Match

A group comparison can be represented in words as a contrast between an outcome summary for units assigned to one strategy and an outcome summary for units assigned to another. For that contrast to answer a particular causal question, both summaries must refer to the same kind of study unit and the same outcome definition and time. This is a matter of matching the quantities being compared.

If the first summary is a 30-day hospitalization proportion and the second is a 90-day proportion, their difference mixes treatment contrast with outcome-window differences. If one summary is per clinic and the other per patient, they are not summaries on the same unit basis. If the groups differ in more than the assigned strategy, the numerical contrast may reflect those differences as well. Thus the comparison is interpretable only after the unit, assigned strategies, outcome, and timing have been aligned. Alignment makes the question precise; it does not by itself establish causal identification.

Diagnostics, Alternatives, and Limitations

Before calculating or interpreting a comparison, check the study description for mismatches. The following questions often expose an underspecified causal question:

• Is it clear what one study unit is, and does that match the population named in the question?
• Is the assignment unit the same as the study unit? If not, is the distinction stated?
• Are both treatment strategies described, including the comparator?
• Is the outcome a measurable variable with a stated scale or event definition?
• Is the follow-up time or event window anchored to a clear starting point?
• Are the compared summaries calculated for the same type of unit and the same endpoint?

There are alternatives when the first formulation is not adequate, but they answer different questions. A researcher might compare assignment to a program rather than treatment actually received; report a 30-day endpoint rather than a one-year endpoint; or define clinics rather than patients as the units of interest. These choices can all be reasonable if made deliberately and described clearly. They are not interchangeable analysis options that preserve one unchanged causal question.

Clear definitions also have limits. They do not resolve whether assignment was randomized, whether groups differ systematically, whether outcomes were measured accurately, or whether the intervention was sufficiently well specified. Those questions require additional design and causal reasoning. In particular, an observed difference between assigned groups should not be called a treatment effect merely because the units, strategies, and endpoint have been named precisely.

Common Pitfall