What Exclusion Rules Out
An instrumental variable is useful only if its effect on the outcome operates through the treatment being studied. The exclusion restriction expresses that idea at the level of potential outcomes: if a unit’s treatment were held fixed, changing the instrument would not change that unit’s outcome. It is a strong causal claim, not simply a statement that the instrument is randomized, unrelated to measured covariates, or associated with treatment.
For a binary instrument \(Z_i\), binary treatment \(D_i\), and outcome \(Y_i\), let \(D_i(z)\) denote the treatment unit \(i\) would receive if the instrument were set to \(z\), where \(z\in\{0,1\}\). Let \(Y_i(d,z)\) denote the outcome that unit \(i\) would have if treatment were set to \(d\) and the instrument were set to \(z\). In this notation, the outcome potential outcome allows the instrument to affect the outcome independently of treatment; exclusion rules out that dependence.
The condition is individual-level, not merely an equality of average outcomes. It rules out a direct effect of the instrument for each unit at fixed treatment. A zero average direct effect could arise because positive and negative direct effects cancel; that cancellation does not satisfy the stated restriction.
The observed treatment and outcome are linked to these potential outcomes by consistency: \(D_i=D_i(Z_i)\) and \(Y_i=Y_i(D_i(Z_i),Z_i)\). If exclusion holds, the latter reduces to \(Y_i=Y_i(D_i(Z_i))\). This reduction says that the instrument matters for the observed outcome only insofar as it changes the treatment received. It does not say that the instrument must change treatment for every unit.
Permitted and Forbidden Causal Paths
A useful way to audit exclusion is to draw or describe the causal pathways from \(Z\) to \(Y\). The pathway \(Z\rightarrow D\rightarrow Y\) is permitted: the instrument may change treatment receipt, and treatment may affect the outcome. Exclusion rules out pathways from \(Z\) to \(Y\) that remain open when \(D\) is held fixed. Examples include \(Z\rightarrow Y\), and \(Z\rightarrow M\rightarrow Y\) when \(M\) is a mediator other than the defined treatment.
Whether a pathway is forbidden depends on how the treatment is defined. If the treatment is narrowly defined as attending a counseling session, an instrument that also provides transportation may affect health through travel, even for someone whose attendance is held fixed. If the treatment is defined more broadly to include the transportation package and counseling, that pathway may instead be part of the treatment. A broader treatment definition is not automatically a solution: it must describe a well-defined intervention, and the proposed instrument must still have no route to the outcome outside that intervention.
State exactly what \(D=1\) and \(D=0\) mean, including the relevant treatment components, timing, and versions.
Consider information, money, access, expectations, contact with providers, and other consequences of setting \(Z\).
Ask whether a route can affect \(Y\) while treatment \(D\) is kept fixed. Any such route challenges exclusion.
Mechanism knowledge may support or weaken the assumption, but the restriction is not usually established by a simple observed-data test.
Worked Example: A Reminder and Program Enrollment
Suppose, hypothetically, a research team studies whether enrolling in a job-training program affects later employment. The binary instrument \(Z\) is whether an eligible person receives a reminder letter; \(D=1\) means enrollment, and \(Y\) is the number of weeks employed during the following year. The team proposes that the reminder changes enrollment.
The permitted pathway is \(Z\rightarrow D\rightarrow Y\): the letter prompts enrollment, and enrollment may affect employment. But consider whether a letter could affect employment while enrollment is held fixed. If it contains only a neutral reminder about the program, and conveys no job leads, advice, or other resources, the team may argue that no such route is present. If it includes job-search tips or contact information for employers, there is a plausible route \(Z\rightarrow\) job-search behavior \(\rightarrow Y\) that does not require enrollment.
In the second version, for a person whose enrollment is fixed at \(d\), the letter could still change the potential outcome: \(Y_i(d,1)\ne Y_i(d,0)\). Exclusion would fail. Randomly sending the letter would address a different concern—how \(Z\) is assigned—but would not remove this direct pathway. In the first version, exclusion may be more plausible, but the letter’s actual contents and delivery remain important to the argument.
Worked Checks of the Restriction
Worked Example: A Transportation Voucher and Rehabilitation Attendance
Imagine, hypothetically, that a clinic offers some patients a transit voucher as an instrument for attending rehabilitation. Let \(D=1\) mean attending at least one rehabilitation session, and let \(Y\) be a mobility score measured a month later. Higher scores indicate better mobility. The proposed instrument is \(Z=1\) for receiving the voucher and \(Z=0\) otherwise.
The intended mechanism is \(Z\rightarrow D\rightarrow Y\): the voucher makes attendance easier, and rehabilitation may improve mobility. However, the voucher may also change how much a patient walks or travels, whether the patient can reach other care, or the patient’s financial resources. If any of these changes affect the mobility score while rehabilitation attendance is held fixed, then for some \(i\) and \(d\), \(Y_i(d,1)\ne Y_i(d,0)\).
Thus, with this narrow definition of treatment, exclusion is not well supported merely because the voucher is described as an attendance incentive. The team would need to examine what the voucher can be used for and what other consequences it could have. If it can be used only for travel to rehabilitation, this may reduce some concerns, but does not prove that every direct pathway is absent. The conclusion is a design-based judgment about plausibility, not a test computed from the example’s outcome data.
Worked Example: Checking Potential Outcomes at Fixed Treatment
Consider a fully hypothetical illustration with two units, a binary treatment, and a continuous outcome measured in points. Suppose the proposed mechanism is a randomly assigned offer of a tutoring session, with \(D=1\) meaning receiving the session. The following table lists hypothetical potential outcomes under each fixed treatment value.
| Unit | \(Y_i(0,0)\) | \(Y_i(0,1)\) | \(Y_i(1,0)\) | \(Y_i(1,1)\) |
|---|---|---|---|---|
| A | 62 | 62 | 71 | 71 |
| B | 55 | 57 | 66 | 68 |
For unit A, the fixed-treatment comparisons are \(Y_A(0,1)-Y_A(0,0)=62-62=0\) points and \(Y_A(1,1)-Y_A(1,0)=71-71=0\) points. This unit satisfies exclusion in the illustration. For unit B, the corresponding differences are \(57-55=2\) points and \(68-66=2\) points. The offer increases B’s score by 2 points at either fixed treatment value, so exclusion fails for B.
The overall average direct effect in this two-unit illustration is 1 point at either treatment value: \((0+2)/2=1\). More importantly, even if the units instead had direct effects of \(+2\) and \(-2\), the average would be zero while both units violated the individual-level equality. Therefore, an average comparison alone is not equivalent to the exclusion restriction.
Why Exclusion Is Distinct From Other IV Assumptions
Independence of the instrument and potential outcomes concerns how the instrument is assigned or distributed relative to potential outcomes. Exclusion concerns what setting the instrument can causally change at a fixed treatment value. These claims are logically different. A randomized instrument can still have a direct effect on the outcome; conversely, a plausible no-direct-effect mechanism does not establish that the instrument is independent of potential outcomes.
Relevance is also distinct: relevance concerns whether the instrument changes treatment receipt. Exclusion can hold even if the instrument never changes anyone’s treatment, and a strongly treatment-shifting instrument can violate exclusion. Each assumption addresses a separate feature of the design, so evidence for one should not be presented as evidence for another.
In potential-outcome terms, the distinction is visible in what is compared. Exclusion requires the within-unit equalities \(Y_i(d,1)=Y_i(d,0)\) for fixed \(d\). It does not impose \(D_i(1)\ne D_i(0)\), nor does it describe how \(Z_i\) was assigned. Keeping these statements separate prevents the common error of treating an instrument’s random assignment as proof that it affects the outcome only through treatment.
Diagnostics, Limitations, and Alternatives
Exclusion is generally not directly testable from observed data alone. For each unit, only the outcome under the actual instrument and treatment values is observed; the other potential outcomes are missing. A comparison of observed outcomes across instrument groups combines any effect of the instrument through treatment with any forbidden direct effect. It therefore cannot, by itself, reveal whether the fixed-treatment equalities hold.
Diagnostics can still expose weaknesses in the argument. Investigators can map instrument mechanisms before analysis, inspect implementation records, ask whether the instrument changes services or resources other than treatment, and assess whether outcomes that treatment cannot plausibly affect nonetheless vary with the instrument. Such negative-control checks can reveal problematic pathways, but a null result does not prove exclusion: the check may be underpowered, or the direct effect may be specific to the outcome of interest. Sensitivity analyses can show how conclusions change under specified departures from exclusion, but they replace the exact restriction with an explicit alternative model rather than verify it.
When exclusion seems implausible, possible responses include redesigning the instrument to remove the extra pathway, redefining treatment only when a broader intervention is substantively coherent, or choosing another identification strategy. Merely adjusting for observed covariates does not generally eliminate a causal path from the instrument to the outcome. Each alternative has its own assumptions and should be justified rather than treated as a mechanical repair.
Exclusion is consequential because any forbidden pathway means the instrument can influence the outcome without operating through the treatment under study. In that case, attributing outcome differences to treatment alone is not justified by the instrument design. The assumption should therefore be stated explicitly, tied to a precise treatment definition, and defended with substantive knowledge of how the instrument operates.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Write the exclusion restriction using \(Y_i(d,z)\). What does it hold fixed?
- A reminder instrument changes treatment uptake and also gives recipients a phone number for unrelated counseling. What pathway may violate exclusion if the treatment is narrowly defined?
- Why does random assignment of \(Z\) not establish exclusion?
- In the hypothetical table above, does unit B satisfy exclusion at \(d=1\)? Show the relevant difference in points.
- Can an average direct effect of zero establish the unit-level exclusion restriction? Explain.
Solutions. (1) Exclusion requires \(Y_i(d,1)=Y_i(d,0)\) for every unit \(i\) and each fixed treatment value \(d\); it holds treatment fixed while changing the instrument. (2) The instrument may affect the outcome through counseling use, \(Z\rightarrow\) counseling \(\rightarrow Y\), without changing the narrowly defined treatment. (3) Random assignment concerns how \(Z\) is assigned, whereas exclusion concerns whether changing \(Z\) can affect \(Y\) at fixed \(D\); the two are separate claims. (4) No: \(Y_B(1,1)-Y_B(1,0)=68-66=2\) points. (5) No. Unit-specific direct effects can cancel in an average while one or more units have \(Y_i(d,1)\ne Y_i(d,0)\).