Why the Counterfactual Is Missing
Suppose a study compares a treatment with a control condition. For each unit, the potential-outcomes framework describes an outcome under each treatment state: \(Y_i(1)\) if unit \(i\) receives treatment and \(Y_i(0)\) if that unit receives control. The notation refers to two possible outcomes for the same unit, not to two different people.
But in a study, a unit receives one treatment state at a time. The outcome recorded for that unit therefore corresponds to its received treatment. The potential outcome under the other state is counterfactual: it describes what would have happened to that same unit under a treatment it did not receive. The central difficulty is not usually that the second outcome was measured and accidentally lost. It was not simultaneously realized for that unit in the study.
This can be recorded compactly. When \(A_i=1\), the observed outcome is \(Y_i(1)\); when \(A_i=0\), it is \(Y_i(0)\). Equivalently, a bookkeeping expression for the observed outcome is:
If \(A_i=1\), the expression reduces to \(Y_i(1)\). If \(A_i=0\), it reduces to \(Y_i(0)\). In either case, only one potential outcome contributes to the observed record. This equation describes which outcome is observed; it does not provide the value of the other potential outcome.
A Single Unit Can Reveal Only One State
Consider a patient who receives a medication. The patient’s outcome after treatment can be observed, but the outcome that the same patient would have had over the same period without medication is not also available. A different patient who did not receive the medication can supply an observed control outcome, but that is the other patient’s outcome, not the treated patient’s counterfactual.
This distinction matters because treatment outcomes vary between people for many reasons. Even if two patients have the same measured characteristics, their outcomes need not be identical. Substituting one patient’s observed outcome for another patient’s counterfactual therefore requires a justification; it is not simply a direct observation of what would have happened to the first patient.
Worked Example: Recording a Patient’s Observed Outcome
Imagine a hypothetical study of a short-term medication. The outcome is symptom score after one week, with lower scores indicating fewer symptoms. Patient R receives the medication, so \(A_R=1\), and has an observed score of 4. The two potential outcomes can be written as \(Y_R(1)\) and \(Y_R(0)\).
Because the patient received treatment, the observed score supplies \(Y_R(1)=4\). Substituting into the bookkeeping expression gives:
The value \(Y_R(0)\), the score this same patient would have had without the medication during the comparison period, is not observed. The study record does not let us calculate it as \(4\), as the average of other patients, or as any other particular number without additional information and assumptions. The example specifies only the treatment outcome; it does not specify the counterfactual.
If Patient S instead received control and had a score of 6, then \(Y_S(0)=6\) would be observed. That does not make \(Y_R(0)=6\). The two values refer to different patients, even though they concern the same treatment state.
Counterfactuals Are Not Ordinary Missing Data
A useful diagnostic question is: Was this value defined by the study process but not recorded, or does it refer to an alternative treatment state that the unit did not receive? If a participant received treatment but a device failed to record the outcome, the outcome under treatment may be missing because of a measurement failure. If the participant received treatment and the target is the outcome that would have occurred under control, that is a counterfactual outcome.
These situations should not be conflated. Ordinary missing-data procedures can sometimes address unrecorded outcomes under explicit assumptions about the missingness process and available information. They do not automatically reveal a unit’s outcome under a treatment state that the unit did not receive. In particular, labeling a counterfactual “missing” does not make it an ordinary blank cell that can be filled by observing the same event again.
| Situation | What is unobserved? | Why it is unobserved |
|---|---|---|
| Outcome device fails after treatment | The treated outcome for that unit | The outcome may have occurred, but it was not recorded |
| Unit receives treatment; question concerns its control outcome | The unit’s outcome under control | The unit did not receive control in the comparison |
| Unit receives control; question concerns its treatment outcome | The unit’s outcome under treatment | The unit did not receive treatment in the comparison |
The distinction also sets a limit on what a data table can show. A study may contain many people, each with an observed outcome, while still lacking the unreceived-treatment outcome for every individual. A large sample can provide more observations across units; it does not create a second realized treatment state for each unit.
Worked Data: What Each Row Can and Cannot Show
Worked Example: Marking the Unobserved Potential Outcome
Suppose a hypothetical study records whether each of four participants received a program (\(A_i=1\)) or control (\(A_i=0\)), along with an observed score. The score is higher when the measured outcome is larger.
| Unit | \(A_i\) | Observed score | Observed potential outcome | Counterfactual potential outcome |
|---|---|---|---|---|
| 1 | 1 | 12 | \(Y_1(1)=12\) | \(Y_1(0)\), not observed |
| 2 | 0 | 9 | \(Y_2(0)=9\) | \(Y_2(1)\), not observed |
| 3 | 1 | 10 | \(Y_3(1)=10\) | \(Y_3(0)\), not observed |
| 4 | 0 | 11 | \(Y_4(0)=11\) | \(Y_4(1)\), not observed |
For Unit 1, \(A_1=1\), so \(Y_1^{\mathrm{obs}}=1\cdot Y_1(1)+0\cdot Y_1(0)=12\). The value \(Y_1(0)\) remains unobserved. For Unit 2, \(A_2=0\), so \(Y_2^{\mathrm{obs}}=0\cdot Y_2(1)+1\cdot Y_2(0)=9\); \(Y_2(1)\) remains unobserved. The same reasoning gives \(Y_3(1)=10\) as observed and \(Y_3(0)\) as counterfactual, and \(Y_4(0)=11\) as observed and \(Y_4(1)\) as counterfactual.
The numbers 12 and 9 are both observed, but they belong to different units. The table does not tell us that Unit 1 would have scored 9 under control, or that Unit 2 would have scored 12 under treatment. Those claims would assign one person another person’s outcome. Nothing in the displayed data justifies that assignment.
Why Not Observe Both Outcomes?
A natural proposal is to give each unit treatment and control, then record both outcomes. The difficulty is that the two outcomes are alternatives for the same comparison, not two labels attached to one unchanged event. If treatment is given first and control later, the second outcome occurs at a different time and may be affected by the earlier treatment, changes in the unit, or other events. Reversing the order creates its own version of this problem. The later observation is not automatically the outcome that would have occurred under the alternative state at the original time.
Some designs do expose the same person to different conditions at different times. Such designs can be useful, but a later measurement is not, merely by being from the same person, a direct observation of the earlier counterfactual. A valid interpretation must address the timing and design assumptions that connect those measurements. The basic missing-counterfactual problem remains: for a specified unit and comparison, the study observes the outcome under the treatment state actually received.
Random assignment changes how treatment states are allocated across units, but it does not give any one unit both states. A randomized study can therefore have observed outcomes in both treatment groups while still having only one observed potential outcome per person. Randomization concerns the assignment process; it does not turn the unreceived-treatment outcome into an observed value.
What the Problem Does—and Does Not—Prevent
The missing counterfactual is a fundamental limitation on direct observation, not a claim that causal questions cannot be studied. Researchers can compare groups, use study designs, and rely on assumptions to learn about patterns relevant to treatment effects. Those approaches address what can be inferred from observed data; they do not retroactively reveal the unobserved potential outcome for each particular unit.
This distinction helps keep conclusions appropriately scoped. An observed outcome is a fact recorded for a unit under its received treatment. A counterfactual is a statement about an alternative treatment state for that same unit. Evidence from other units may help answer broader questions, but it must not be described as if it directly measured the counterfactual for a particular person.
Common Mistakes and Diagnostics
- Wrong: “The control-group participant’s outcome is the treated participant’s missing outcome.” Correction: It is an observed outcome for the control participant. Treating it as a stand-in for someone else’s counterfactual requires a separate justification.
- Wrong: “The unreceived outcome is just a blank caused by failed measurement.” Correction: A failed measurement and an unreceived-treatment counterfactual arise for different reasons and should be described separately.
- Wrong: “A large sample lets us observe both outcomes for each person.” Correction: A larger sample adds units and their received-treatment outcomes; it does not supply each unit’s alternative outcome.
- Wrong: “A second measurement on the same person is automatically the counterfactual.” Correction: The treatment state, timing, and intervening circumstances may differ. The second measurement is not automatically the outcome that would have occurred under the alternative state in the original comparison.
- Diagnostic check: For every row, identify the received treatment, the observed outcome, and the potential outcome that remains counterfactual. If the explanation assigns a value to that final quantity without stating a basis for doing so, it claims more than the observed data provide.
Worked Example: Separating a Recording Failure From a Counterfactual
In a hypothetical two-arm study, Participant M receives treatment, but a sensor fails before the end-of-study outcome is recorded. Participant N also receives treatment, and the sensor records a score of 8. A researcher asks for (a) Participant M’s outcome under treatment and (b) Participant N’s outcome under control.
(a) Participant M’s treated outcome: The treatment state was received, so this is an outcome that the study intended to measure under treatment. The sensor failure means the value was not recorded. It is missing observed data; the value cannot be read from the study record.
(b) Participant N’s control outcome: Participant N received treatment, not control. The requested value is \(Y_N(0)\), the outcome for this same participant under the unreceived state. It is counterfactual, not a recorded control outcome that was lost by sensor failure.
The score 8 is \(Y_N(1)\), not \(Y_N(0)\). Nor does Participant M’s sensor failure make M’s unreceived-treatment outcome the same kind of missing quantity as M’s unrecorded treated outcome. The reason each value is unavailable determines how it should be described.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A unit receives control, so \(A_i=0\), and its observed outcome is 5. Which potential outcome is observed, and which is counterfactual?
- Why can’t an observed outcome from a control participant automatically be called the counterfactual outcome for a treated participant?
- A treated participant’s outcome was not recorded because a laboratory sample was lost. Is this the same as that participant’s outcome under control? Explain.
- Does random assignment make both potential outcomes observable for every unit? Why or why not?
- Why is a later measurement after switching treatment conditions not automatically the original counterfactual?
Solutions. (1) Since \(A_i=0\), the observed outcome is \(Y_i(0)=5\); \(Y_i(1)\) is counterfactual. (2) The control participant is a different unit, so the outcome observed for that person is not a direct observation of the treated participant’s outcome under control. (3) No. The treated outcome was intended to be measured but was not recorded; the outcome under control is an alternative-treatment counterfactual. (4) No. Assignment determines which state a unit receives, so its observed outcome corresponds to that state; random assignment does not make the other state occur for that unit. (5) The later measurement occurs at a different time and may follow different prior conditions or events, so it does not automatically represent the outcome under the alternative state at the original comparison time.