When Is an Instrumental-Variable Design Appropriate for Causal Inference?

Can You Estimate a Causal Effect When Treatment Is Not Randomly Assigned?

Suppose you want to determine whether participation in an optional educational program improves student achievement. Students who participate may already be more motivated, receive greater family support, or attend better-resourced schools. Even after adjusting for measured characteristics, some differences between participants and nonparticipants may remain.

A conventional regression comparing their outcomes could therefore mistake preexisting differences for an intervention effect. An instrumental-variable design offers a possible alternative: identify a source of variation that influences participation but is otherwise unrelated to the potential outcomes, except through participation itself.

The difficulty is that finding a variable associated with participation is relatively easy. Establishing that it is a valid instrument is much harder. When does the approach genuinely strengthen causal inference, and when does it simply replace one set of questionable assumptions with another?

Use Instrumental Variables When a Credible External Source of Treatment Variation Exists

How Instrumental Variables Identify Causal Effects

Why Ordinary Regression May Not Identify the Treatment Effect

In causal inference, a treatment is endogenous when its relationship with the outcome is entangled with factors that prevent a regression coefficient from having the intended causal interpretation.

Common sources include unmeasured confounding, reverse causation, and certain forms of measurement error. For example, a researcher examining whether additional tutoring improves examination performance might adjust for prior achievement, socioeconomic status, and school characteristics. Yet unmeasured academic motivation could still influence both tutoring participation and subsequent performance.

Adding more covariates does not necessarily eliminate the problem. Some relevant factors may be unavailable, measured poorly, or conceptually difficult to observe.

An IV strategy seeks variation in treatment that does not arise from these problematic sources. Rather than relying on all observed differences between treated and untreated participants, it uses the component of treatment variation associated with the proposed instrument.

What Is an Instrumental Variable?

An instrumental variable, usually represented by Z, is a variable used to identify the causal effect of a treatment or exposure X on an outcome Y under a specified set of assumptions.

Consider a hypothetical education program in which some schools are randomly offered additional tutoring resources. The offer influences whether students receive tutoring, but not every eligible student participates. The randomized offer could potentially serve as an instrument for actual tutoring participation.

The distinction is important. The instrument is the offer, the treatment is actual participation, and the outcome might be a standardized achievement score.

Randomized encouragement is particularly useful for explaining IV logic because random assignment can support the independence assumption. It does not, however, establish the exclusion restriction automatically.

What Assumptions Must an Instrument Satisfy?

Instrumental-variable identification depends on assumptions that should be articulated before interpreting an estimate. Their precise formulation varies with the causal estimand, treatment type, and statistical model.

1. Relevance: Does the Instrument Affect Treatment?

The instrument must influence treatment assignment, treatment uptake, or treatment intensity. If an encouragement does not change participation, it cannot identify the participation effect through the intended mechanism.

Researchers commonly investigate relevance through the first-stage relationship between the instrument and treatment. In linear IV analysis, a weak first stage can produce unstable estimates and unreliable conventional confidence intervals.

A statistically significant relationship is not enough by itself. Instrument strength, sample size, the number of instruments, and the chosen inference procedure all matter.

2. Independence: Is the Instrument Unrelated to Potential Outcomes?

The independence assumption requires that the instrument be appropriately independent of potential outcomes and relevant unmeasured determinants of those outcomes, potentially conditional on prespecified covariates.

Random assignment of an encouragement can make this assumption more credible, provided randomization is implemented correctly. Naturally occurring instruments require a substantive argument explaining why their assignment is plausibly exogenous.

For instance, geographic distance to a training center might predict participation. But distance may also reflect neighborhood wealth, infrastructure, transportation access, or school quality. These relationships could undermine independence.

Researchers should therefore examine whether the proposed source of exposure variation is plausibly as if random rather than relying on the label of a natural experiment.

3. Exclusion Restriction: Can the Instrument Affect the Outcome Only Through Treatment?

The exclusion restriction rules out causal pathways from the instrument to the outcome that bypass the treatment of interest.

Suppose schools offered tutoring resources also receive improved internet access. If internet access independently improves achievement, the offer may influence outcomes even among students who do not receive tutoring. The offer would then violate the exclusion restriction for an analysis treating tutoring participation as the sole causal pathway.

Exclusion is usually not directly testable from observed data. Researchers must justify it through institutional knowledge, program design, causal reasoning, and sensitivity analysis.

4. Monotonicity: Does the Instrument Move Treatment in a Consistent Direction?

For the conventional local average treatment effect interpretation with a binary instrument and binary treatment, monotonicity assumes that no individual would take the treatment when not encouraged but refuse it because of encouragement.

People may respond differently to an instrument. Some always participate, some never participate, and others participate only when encouraged. Monotonicity excludes individuals whose treatment response moves in the opposite direction.

This assumption is not required in exactly the same form for every IV estimand or model. Nevertheless, it is central to the familiar interpretation of IV estimates as effects among compliers.

What Causal Effect Does an Instrumental-Variable Design Estimate?

An important limitation is that IV estimation does not necessarily recover the average treatment effect for the entire population.

Under the standard binary-instrument framework with relevance, independence, exclusion, and monotonicity, the Wald estimator identifies the local average treatment effect (LATE): the average effect among individuals whose treatment participation changes because of the instrument.

These individuals are commonly called compliers. Their characteristics and treatment responses may differ from those of always-takers or never-takers.

Consequently, an IV estimate may answer a narrower question than the researcher initially intended. The causal effect for students induced to participate by an offer may not equal the effect for students who would participate regardless of the offer.

With continuous instruments, multiple instruments, heterogeneous effects, or different structural assumptions, the interpretation can become more complicated. Researchers should define the estimand rather than automatically calling every IV coefficient an average treatment effect.

How Is an Instrumental-Variable Effect Estimated?

For a binary instrument and treatment, a simple Wald estimator divides the instrument's effect on the outcome by its effect on treatment uptake.

Under the required assumptions, this estimate can be interpreted as a complier average causal effect. It does not establish that every participant gains 10 points or that the entire student population would gain the same amount.

For more complex settings, researchers frequently use two-stage least squares (2SLS). The first stage models treatment using the instrument and appropriate covariates. The second stage estimates the outcome relationship using the instrument-induced component of treatment variation.

Valid standard errors require an appropriate IV estimation procedure. Simply substituting fitted treatment values into an ordinary regression and treating them as observed without accounting for the estimation structure is not a generally reliable approach to inference.

How Do You Assess Weak Instruments?

Weak instruments explain too little variation in treatment to support reliable conventional IV inference. The resulting estimates may be imprecise, sensitive to specification, and affected by substantial finite-sample bias.

Researchers often report first-stage statistics. The familiar first-stage F-statistic threshold of 10 is a historical screening heuristic, not a universal guarantee of instrument strength. Its adequacy depends on the model, number of endogenous regressors, error structure, and inference method.

Where appropriate, researchers should consider weak-instrument-robust confidence intervals or tests, including procedures based on the Anderson–Rubin approach. With clustered observations or heteroskedasticity, diagnostics and inference should account for those features.

A weak instrument cannot be repaired merely by describing the resulting coefficient as causal.

Can Multiple Instruments Make an IV Design More Credible?

Multiple instruments may improve identification or precision when they contribute valid, relevant variation. However, adding instruments also creates additional assumptions and can worsen finite-sample performance.

Overidentification tests may detect certain inconsistencies among instruments, but failing to reject such a test does not prove that all instruments satisfy the exclusion restriction. Invalid instruments can share similar violations.

Substantive justification remains necessary for every proposed instrument.

When Is an IV Design Preferable to Other Quasi-Experimental Strategies?

An IV design is particularly attractive when treatment selection is problematic but a credible source of exogenous treatment variation exists.

Other strategies rely on different identification opportunities. Difference-in-differences uses changes over time between groups under assumptions about counterfactual trends. Regression discontinuity exploits treatment assignment around a threshold. An interrupted time-series design examines changes associated with an intervention at a defined time point.

These approaches are not interchangeable. The most defensible choice depends on the treatment assignment process, available observations, assumptions, and causal question.

An IV design may also be combined with other strategies, but such combinations require additional identification arguments rather than automatically providing stronger evidence.

Estimating the Effect of Tutoring When Participation Is Voluntary

Suppose the invitation also encourages students to study independently. That additional pathway would threaten the exclusion restriction. The calculation would remain mathematically correct, but its interpretation as the causal effect of tutoring would no longer be justified.

This illustrates a defining feature of instrumental-variable research: identification depends on how the instrument operates, not merely on whether the statistical calculation can be completed.

Common Misinterpretations of Instrumental-Variable Designs

How to Decide Whether an Instrumental-Variable Strategy Fits Your Study

Begin with the causal problem, not the statistical technique. Identify the treatment, outcome, target population, and effect you want to estimate. Then explain why ordinary adjustment may not identify that effect.

Next, investigate how the proposed instrument is assigned. A credible institutional mechanism, randomized encouragement, or plausibly exogenous policy variation may offer stronger justification than a variable selected because it happens to predict treatment well.

For educational research, the most promising instruments often arise from carefully documented allocation procedures or institutional rules. Even then, implementation details matter. Researchers should examine noncompliance, spillovers, missing outcomes, differential attrition, and clustering where relevant.

Predefining the identification assumptions and planned sensitivity analyses can reduce the risk of selecting an instrument merely because it produces an attractive estimate.

Before Using an Instrumental-Variable Design, Check These Conditions

Frequently Asked Questions About Instrumental Variables

Use Instrumental Variables Only When the Instrument's Causal Role Is Defensible

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.