Building An Exclusion Restriction Argument

What an Exclusion Argument Must Establish

An instrument can predict an endogenous regressor and still be unsuitable for identifying its effect on an outcome. The exclusion restriction asks a specific causal question: holding the endogenous regressor fixed, can changing the instrument still change the outcome? A credible answer requires more than noting that the instrument was assigned by a rule or is strongly associated with treatment. It requires tracing the routes by which the instrument could influence the outcome and explaining why routes outside the endogenous regressor are absent or adequately ruled out.

Let $Y_i$ be an outcome, $D_i$ an endogenous regressor, $Z_i$ a proposed instrument, and $X_i$ a vector of pre-instrument controls. A linear structural equation used to describe the exclusion restriction is

Y_i=\beta D_i+X_i'\gamma+u_i.

The equation contains no direct $Z_i$ term. In this model, exclusion means that, conditional on $X_i$, the instrument has no effect on $Y_i$ except through $D_i$. More generally, with potential outcomes, exclusion is expressed as $Y_i(d,z)=Y_i(d)$: at a fixed value $d$, changing $z$ does not change the outcome. When controls are needed, this claim is understood within levels of the pre-instrument controls.

Definition: The exclusion restriction requires that the instrument affect the outcome only through the endogenous regressor, conditional on the specified pre-instrument controls. It rules out direct effects of the instrument and effects through other channels that bypass the endogenous regressor.

Exclusion is not the same as relevance or instrument exogeneity. Relevance requires that $Z_i$ predict $D_i$ after accounting for $X_i$. Exogeneity requires an appropriate relationship between the instrument and unobserved determinants of the outcome; in the linear model, a common conditional moment is $E[Z_i u_i\mid X_i]=0$. Exclusion concerns causal pathways. A complete IV argument needs the relevant conditions, but this tutorial focuses on how to reason about exclusion.

The treatment or endogenous regressor must also be defined precisely. If $D_i$ means receipt of a particular medication, a reminder’s effect on taking that medication is a pathway through $D_i$. Its effect on a separate treatment, follow-up care, or symptom management is not. Whether a channel counts as part of $D_i$ depends on the stated estimand and treatment definition, not on whether the channel seems related to treatment.

A Pathway-Based Argument

A useful starting point is a pathway diagram or a written list of causal routes. The intended route is $Z_i\rightarrow D_i\rightarrow Y_i$. Then ask whether any plausible route remains from $Z_i$ to $Y_i$ that does not pass through $D_i$. Such a route might be direct, such as an instrument changing income, or indirect through another variable, such as an instrument changing access to a second service that affects the outcome.

Institutional details matter because they tell us what the instrument actually changes. Who is eligible? What does assignment provide? Can the benefit be used for another purpose? Does the policy change access, information, prices, peers, or expectations as well as the endogenous regressor? These questions help assess exclusion, but a favorable institutional story is not a mathematical proof. The argument should name important alternatives and explain why the institutional design makes them unlikely, impossible, or part of the defined treatment.

Define the variables.
State what $Z_i$, $D_i$, and $Y_i$ mean, including units and timing. A vague treatment definition makes pathways impossible to classify.

State the intended channel.
Describe how the instrument changes the endogenous regressor and how that regressor can change the outcome.

Trace bypass channels.
Consider direct effects and effects through other variables, including information, resources, access, behavior, and institutional responses.

Use design details to evaluate them.
Explain which features block or support each proposed route, and acknowledge what remains uncertain. Do not treat association or a strong first stage as evidence that exclusion holds.

Why a Bypass Pathway Matters

A short algebraic argument shows why an unblocked pathway is consequential. Suppose the actual structural equation contains a direct instrument effect:

Y_i=\beta D_i+X_i'\gamma+\delta Z_i+u_i.

For simplicity, consider the just-identified case after removing the linear effects of $X_i$. Write the residualized variables as $\tilde Y_i,\tilde D_i,\tilde Z_i$. The IV ratio is $\operatorname{Cov}(\tilde Z,\tilde Y)/\operatorname{Cov}(\tilde Z,\tilde D)$. Substitution of the structural equation gives

\frac{\operatorname{Cov}(\tilde Z,\tilde Y)} {\operatorname{Cov}(\tilde Z,\tilde D)} = \beta+ \delta\frac{\operatorname{Var}(\tilde Z)} {\operatorname{Cov}(\tilde Z,\tilde D)} + \frac{\operatorname{Cov}(\tilde Z,\tilde u)} {\operatorname{Cov}(\tilde Z,\tilde D)}.

To obtain this expression, take the covariance of the structural equation with $\tilde Z_i$. The term involving $X_i$ vanishes after residualization; the direct-effect term contributes $\delta\operatorname{Var}(\tilde Z)$. Dividing by $\operatorname{Cov}(\tilde Z,\tilde D)$ yields the displayed ratio. Thus a direct effect can shift the IV ratio away from $\beta$, even if the instrument predicts $D_i$. The final term also shows why exclusion alone does not establish the required orthogonality with unobserved determinants.

Worked Example: A Reminder as an Instrument for Medication Use

Hypothetical setting. A clinic randomly sends some patients a text reminder. Let $Z_i=1$ indicate receiving the reminder, $D_i=1$ indicate taking a prescribed medication during the next month, and $Y_i$ be a symptom score, measured in points, at the end of the month. The proposed instrument is the reminder, and the endogenous regressor is medication use.

The intended route is clear: the reminder increases medication use, and medication use can affect symptoms. Suppose, for illustration, the reminder raises the medication-use rate by $0.20$, while the average symptom score falls by $3.0$ points. The simple Wald ratio is

\frac{\Delta Y}{\Delta D} = \frac{-3.0\text{ points}}{0.20} = -15.0\text{ points per unit increase in medication use}.

These figures are hypothetical and do not establish that the ratio is a causal medication effect. Now suppose the text also provides symptom-management advice that lowers the score by $1.0$ point even for patients whose medication use is unchanged. That is a bypass pathway from $Z_i$ to $Y_i$, so exclusion fails. Under the illustrative additive decomposition, the portion of the outcome change attributable to the medication channel is $-3.0-(-1.0)=-2.0$ points. Dividing by the first-stage change gives $-2.0/0.20=-10.0$ points per unit increase in medication use. The difference between $-15.0$ and $-10.0$ illustrates the distortion from the direct channel; it is not a proposed correction method for general IV analyses.

The institutional argument would therefore need to establish what the messages contain and what other care they trigger. A neutral reminder that only states the medication name and dose gives a more favorable exclusion story than a message containing advice, appointment links, or prompts to contact clinicians. Even the neutral version is not automatically valid: it might change adherence to other care or patients’ expectations. The relevant question is whether those changes affect symptoms outside medication use.

Institutional Features Can Reveal Hidden Channels

Worked Example: Transit Subsidy as an Instrument for Training

Hypothetical setting. A city gives eligible residents a transit pass that can be used for trips to an approved job-training center. Let $Z_i$ denote eligibility for the pass, $D_i$ attendance in the training program, and $Y_i$ annual earnings in dollars. Suppose eligibility increases training attendance by $0.10$, and the average earnings difference associated with eligibility is $\$120$ per year.

If the pass only enables attendance at the training center, the intended pathway is $Z_i\rightarrow D_i\rightarrow Y_i$. But suppose the pass also helps recipients reach job interviews. That route, $Z_i\rightarrow$ interview access $\rightarrow Y_i$, bypasses training attendance. To see the scale of the issue under an illustrative additive model, suppose this interview-access route raises annual earnings by $\$40$ per eligible person, independently of training attendance. The outcome change attributed to the training channel would then be $\$120-\$40=\$80$. The corresponding ratio would be

\frac{\$80\text{ per eligible person}}{0.10\text{ additional training attendance}} = \$800\text{ per unit increase in attendance}.

By contrast, using the entire earnings difference gives $\$120/0.10=\$1{,}200$ per unit increase in attendance. These calculations rely on the stipulated hypothetical effects and a simple additive decomposition; they demonstrate why the bypass route matters, not how to estimate its size in an actual application.

The institutional case turns on the pass’s permitted uses and the transportation network. A pass valid only for a dedicated shuttle to the center, with no useful job-search destinations along its route, would make the interview channel less plausible. A general transit pass usable across the city would make it more plausible. A researcher should also ask whether eligibility changes access to other services or job placement assistance. Merely controlling for observed interview attendance would not automatically solve the problem: that variable may be caused by the instrument, and conditioning on it changes the causal comparison rather than proving that the original instrument has no bypass effect.

Worked Example: A Scholarship Offer as an Instrument for Years of Schooling

Hypothetical setting. A college uses a lottery to offer a tuition scholarship. Let $Z_i=1$ denote receiving an offer, $D_i$ be years of completed schooling, and $Y_i$ be annual earnings five years later. Assume the offer raises average schooling by $0.25$ years and the earnings difference by $\$500$ per year. The raw ratio is $\$500/0.25=\$2{,}000$ in annual earnings per unit increase in schooling. The lottery may support a credible argument about how the offer is assigned, but that fact alone does not establish exclusion.

Consider how the scholarship works. If the award is paid directly to the college, cannot be converted to cash, and is available only while enrolled, that design limits a simple direct-income route from the offer to later earnings. It does not rule out all alternatives. The award could change which college a student attends, reduce borrowing, or allow a student to work fewer hours while enrolled. If those changes affect later earnings without being captured by completed years of schooling, they are bypass channels for the stated $D_i$.

The conclusion depends on the outcome model and treatment definition. If $D_i$ is years of schooling alone, changing college quality while holding years fixed is a possible route outside $D_i$. If the target treatment is instead a broader educational package, the treatment and estimand must be redefined explicitly; it is not enough to quietly treat every scholarship effect as “schooling.” Institutional details that restrict award use strengthen some parts of the argument, while changes to school choice, debt, or work experience leave concerns to address. The hypothetical ratio does not settle those concerns.

Common Mistakes and a Disciplined Conclusion

A common mistake is to argue, “The instrument was randomly assigned, so exclusion holds.” Random assignment may support independence from baseline causes of the outcome, but it does not prevent the assignment itself from changing the outcome through a route other than $D_i$. Another mistake is, “The instrument affects the outcome only because it predicts treatment.” Prediction establishes neither the absence of direct effects nor the absence of omitted channels. A first-stage relationship is relevant to relevance, not proof of exclusion.

A more careful argument states the claim and its limits: identify the intended pathway; name plausible direct and indirect alternatives; use concrete institutional features to assess them; and distinguish what those features rule out from what remains uncertain. If the argument relies on controls, specify that they are pre-instrument variables and explain why conditioning on them addresses the particular concern. Do not assume that controlling for a post-instrument variable erases a causal pathway without changing the estimand or introducing other problems.

Key takeaway: An exclusion-restriction argument is a causal pathway argument, not a statistical association argument. It is strongest when the instrument’s institutional operation is described precisely, the treatment is defined narrowly enough to classify channels, and plausible routes from the instrument to the outcome outside that treatment are addressed directly.

Check Your Understanding

For each question, focus on the pathway and the institutional claim rather than on whether the instrument predicts treatment.

A randomly assigned reminder increases vaccination, but also sends recipients a link to book a preventive-health appointment. If the outcome is later health-care use, identify the intended route and a possible bypass route.
In the transit-pass example, why does random eligibility not by itself establish the exclusion restriction?
Suppose a scholarship offer changes completed years of schooling and college attended. If the endogenous regressor is years of schooling, explain why college attended could be an exclusion concern.
In the structural equation with a direct instrument effect, what term in the IV ratio is added by $\delta Z_i$?
Give one institutional feature that could make a proposed bypass channel less plausible, and explain why it supports but does not prove exclusion.