Instrumental variable analysis
Instrumental variable analysis uses a cause of treatment to estimate an effect when ordinary adjustment cannot remove confounding.
Fit
When to use it, and when not to
Use it when
- You have a credible source of treatment variation, such as randomized encouragement, that affects the outcome only through treatment.
Not when
- Your proposed instrument can affect the outcome through another path. Use a design with more credible causal assumptions. Read: Directed acyclic graphs
- Measured confounders suffice for your question. Consider standardization instead of relying on an instrument. Read: Marginal standardization and g-computation
Before you start
What you need
- A clearly defined instrumentExplain what changes treatment and why that source of variation is credible.
- Treatment, outcome and target grouptargetoutcomeTypeName the effect sought and whose treatment the instrument could change.
- The first-stage modelRecord how strongly the instrument predicts treatment after the planned covariate adjustment.
- The instrument changes treatment.How to check: Estimate the first-stage association and report its uncertainty and an appropriate strength statistic.
- The instrument has no shared causes with the outcome.How to check: Justify this from the design, conditional on any required baseline covariates. Measured balance cannot prove it.
- The instrument affects the outcome only through treatment.How to check: Describe and challenge other plausible pathways. This exclusion restriction cannot be established by a strong first stage.
- A complier effect needs no people who move against the instrument.How to check: For binary treatment and instrument, monotonicity rules out defiers. Also require well-defined treatment and no interference.
- Choose a missing-data method because of a stated assumption.How to check: Missing completely at random (MCAR) means gaps depend on neither observed nor unseen data. Missing at random (MAR) means recorded data explain gaps. Under MAR, suitable multiple imputation may help; it is not inherently better. Missing not at random (MNAR) means unseen values still predict gaps. Under plausible MNAR, add a delta or tipping-point sensitivity analysis. Dropping people with missing exposure can cause selection bias.
Plain language
How it works
For a binary instrument, divide its outcome difference by its treatment difference. This is the Wald ratio. With binary treatment, the usual causal reading is local to compliers, people whose treatment changes with the instrument.
Two-stage least squares first predicts treatment from the instrument and covariates. It then regresses the outcome on predicted treatment and the same covariates. This linear IV fit needs standard errors that account for both stages. Report the first-stage partial F statistic with its context; 10 is a warning rule, not a guarantee. Weak instruments can give biased estimates and misleading intervals.
For a binary outcome, it estimates a risk difference under a linear model; a logistic second stage does not give a causal odds ratio.
Worked example
A worked example
- Setting
- A trial randomly encourages some adults to start blood-pressure treatment.
- Question
- What is the treatment effect among people whose uptake changes with encouragement?
- Approach
- Uptake is 60 in 100 with encouragement and 20 in 100 without it. Mean blood pressure is 120 versus 124 mmHg.
- Result
- First-stage difference: 0.40. Outcome difference: -4 mmHg. Wald estimate: -10 mmHg per person whose treatment changes with encouragement.
- What it means
- Divide the outcome difference by the uptake difference. This is a complier effect under the stated assumptions, not an effect for every patient.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
A strong first stage does not prove an instrument is valid. Exclusion and independence need substantive arguments; the effect may apply only to people whose treatment the instrument changes.
- Calling any predictor of treatment an instrument.Instead: Explain independence and exclusion, not just prediction.
- Using ordinary second-stage regression standard errors.Instead: Use an IV estimator that accounts for both stages and any clustering.
- Treating a familiar first-stage cutoff as proof of validity.Instead: Weak-instrument bias and uncertainty depend on the design and estimator. Use appropriate diagnostics and weak-instrument-robust inference.
Inside Mallard
How Mallard uses this
Plans do not link to this method yet. You can find it in Learn and the Atlas.
instrumental-variables-2sls).analysis.instrumental_variable, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- An IV for the RCT: using instrumental variables to adjust for treatment contamination in randomised controlled trials (opens in a new tab) · BMJ · doi 10.1136/bmj.c2073
- Identification of causal effects using instrumental variables in randomized trials with stochastic compliance (opens in a new tab) · Biom J · doi 10.1002/bimj.201200104
- A review of instrumental variable estimators for Mendelian randomization (opens in a new tab) · Stat Methods Med Res · doi 10.1177/0962280215597579