CausalInferenceforEducation
Most education decisions get made on correlations that quietly fall apart under scrutiny. Attendance tracks with outcomes, so attendance must matter — except the students who attend more are also the ones with stable housing, engaged families and fewer disruptions. Funding the correlation buys you a program that looks effective and is not.
iThe problem
Interventions are expensive and hard to reverse. A program rolled out district-wide consumes budget and political capital that cannot be spent twice.
Randomization is usually unavailable — you cannot withhold a promised intervention from half a cohort of students, and no ethics board would let you.
The naive estimate is almost always biased upward, because selection into programs is rarely random and rarely observable.
iiThe approach
Difference-in-differences against a comparable, untreated cohort, with event-study plots to test whether trends were parallel before the intervention landed.
Propensity weighting as a cross-check, reporting the overlap region rather than silently extrapolating into territory with no comparable controls.
Sensitivity analysis on unobserved confounding: how strong would a hidden confounder have to be to erase the effect? If the answer is 'implausibly strong', the finding is durable.
Every specification is committed to version control with its data snapshot, so any number can be traced back to the exact run that produced it.
iiiWhat it produces
Estimates arrive with explicit uncertainty bounds and a written statement of what would falsify them.
Recommendations distinguish between 'we are confident this helps' and 'the data cannot answer this yet' — the second is a legitimate and frequently correct conclusion.
The specification registry means an estimate produced eighteen months ago can be regenerated byte-for-byte today.
Stack