Instrumental Variable Methods and Two-Stage Least Squares in Clinical Research

Randomized treatment assignment is the cleanest route to causal inference, but many clinical questions arise from settings where randomization is unavailable, unethical, or operationally impossible. Patients and clinicians select treatments, treatment access varies between hospitals, and important factors such as disease severity, adherence, or frailty may be incompletely recorded. Regression adjustment can reduce measured confounding, but it cannot remove bias from variables that were never observed.

Instrumental variable methods address this problem through a different source of variation. Instead of requiring every confounder to be measured, IV analysis identifies an external factor that changes the probability of receiving treatment while having no independent route to the outcome. In clinical research, this might be geographic prescribing preference, physician treatment tendency, policy-driven eligibility, or a logistical feature that shifts access without directly changing patient prognosis. The design is powerful, but its credibility depends on assumptions that must be defended rather than implied.

1. The three core assumptions of an instrumental variable

A valid instrument must satisfy three principal conditions. Relevance requires that the instrument predicts treatment receipt or treatment intensity. A weak relationship produces unstable estimates and magnifies finite-sample bias. Independence requires that the instrument is not associated with unmeasured causes of the outcome. This assumption cannot be proven from observed data alone, so researchers use subject-matter reasoning and balance diagnostics to support it.

The third condition is the exclusion restriction: the instrument affects the outcome only through its effect on treatment. If a regional prescribing preference also changes referral quality, follow-up intensity, or access to supportive care, the instrument may influence the outcome through multiple pathways and the IV estimate loses its causal interpretation. A strong first-stage association is therefore necessary but not sufficient for validity.

2. How two-stage least squares works

2SLS operationalizes IV estimation in two linked regressions. In the first stage, treatment is regressed on the instrument and the pre-specified covariates. This produces each patient's treatment value predicted from the instrument-driven component of exposure. In the second stage, the clinical outcome is regressed on that predicted treatment and the same covariates. The coefficient on predicted treatment is the IV estimate under the model assumptions.

The two stages should not be interpreted as two independent analyses. The predicted exposure is a constructed quantity whose uncertainty must be carried through the second-stage inference. Standard errors from a naive regression of the outcome on fitted treatment can be incorrect; researchers should use an IV estimator or software procedure that accounts for first-stage estimation.

3. Evidence summary table

4. LATE and the meaning of a clinical IV estimate

Unlike an intention-to-treat effect, an IV estimate usually does not describe the effect for every patient. Under the monotonicity assumption, it identifies the local average treatment effect among compliers: patients whose treatment status changes because of the instrument. Always-treated and never-treated patients do not contribute the same identifying variation, and patients who move against the instrument's direction would violate monotonicity.

This population-specific interpretation matters in clinical manuscripts. If the instrument is physician preference, the estimate may describe patients whose treatment choice is sensitive to physician preference, not patients who would receive the therapy regardless of clinician. Authors should state the likely complier population and avoid presenting LATE as a universal treatment effect without justification.

5. Weak instruments and first-stage diagnostics

A weak instrument has limited predictive power for treatment. In this setting, the IV estimator can be biased toward the conventional confounded estimate and can have confidence intervals that do not behave as expected. A first-stage F-statistic is commonly reported as a screening diagnostic, but it should not be treated as a universal pass-fail threshold. The appropriate assessment depends on sample size, model structure, number of instruments, and the inferential method.

Researchers should report the first-stage coefficient, uncertainty, partial R-squared, instrument distribution, and the covariate balance associated with the instrument. When weak-instrument risk is material, weak-IV-robust procedures such as Anderson–Rubin-type inference or conditional likelihood approaches may be more appropriate than relying only on conventional 2SLS intervals.

6. Actionable steps for a defensible IV analysis

7. Common design errors in clinical manuscripts

The most serious error is treating any variable associated with treatment as an instrument. Association establishes neither independence nor exclusion. Another frequent problem is adjusting for post-instrument variables that lie on the pathway from the instrument to treatment, which can remove the variation needed for identification. Researchers also sometimes report a large 2SLS coefficient without describing the complier population, making the clinical meaning of the estimate unclear.

Good reporting places the causal diagram and assumption argument before the regression table. It explains why the instrument changes treatment, why it should not alter prognosis through another route, how instrument strength was evaluated, and whether the estimated effect is clinically plausible. Negative control outcomes, alternative instruments, and sensitivity analyses cannot prove the assumptions, but they can expose contradictions that a single adjusted model would miss.

Conclusion

Instrumental variable methods provide a valuable route to causal inference when treatment selection is confounded and randomization is unavailable. Their strength comes from isolating treatment variation that is plausibly unrelated to unmeasured prognosis, but that strength is conditional on relevance, independence, exclusion restriction, and a transparent interpretation of LATE. A high-quality clinical IV manuscript therefore reports more than a 2SLS coefficient: it documents the causal design, quantifies first-stage strength, addresses weak instruments, defines the target complier population, and tests whether the conclusion survives credible alternatives.