Instrumental Variable Estimation under One-Sided Non-Compliance
When evaluating digital decision-support tools under full discretion, frontline staff frequently choose not to inspect algorithmic predictions, creating endogenous selection bias. Standard intent-to-treat estimates dilute actual software effects, while direct comparisons of users versus non-users suffer from self-selection confounding.
Picture this
Imagine testing whether wearing a new type of safety goggles reduces eye injuries in a factory. If workers randomly offered goggles choose to wear them only half the time, while unoffered workers can never get them, researchers can use the initial random offer as a lever to isolate the pure effect of actually wearing the goggles.
What the evidence says
IV estimates proved that downloading predictions had zero statistically significant impact on short-term caseworker compliance (estimates ranged from -0.05 to +0.03 across specifications, all statistically insignificant).
- Who
- N = 14,977 treatment jobseekers and N = 16,566 control jobseekers across 21 Swiss employment offices.
- How
- Two-Stage Least Squares (2SLS) Instrumental Variables regression using caseworker randomization status (Z) as an instrument for prediction download behavior (D) under one-sided non-compliance (Imbens & Angrist, 1994).
What to do
Instrument actual software usage with randomized assignment status to estimate local average treatment effects when software adoption is voluntary and non-compliance is one-sided.
From the source
"Since it was impossible for the members of the control group to download the predictions the monotonicity condition of IMBENS and ANGRIST (1994) is satisfied by definition, and this also means that the local average treatment effect (LATE) is the same as the average treatment effect on the treated (ATET) because the treated are the compliers."
Targeting_Labour_Market_Programmes_Results_from_a_Randomized_Experiment.pdf