School groups
Compare a group-wide policy across schools like yours. A group pack layers your own policies and standards over your regulator's pack.
Policy lab · For school groups and education systems
Run a policy through hundreds of schools before it reaches a real one.
The policy lab runs the engine without the cast. Each school runs with and without the policy on the same random draws, and you compare how the results spread.
How a run works
Each school runs twice, and only the policy differs.
Shared random draws
Luck cancels out, so the policy shows.
When two branches draw their own luck, a storm or a bad flu week in one branch can swamp the effect you are looking for. Shared draws give both branches the same luck.
In the prototype, the effect of supporting 40 students varied a quarter to a third as much with shared draws as with separate ones. A run with identical decisions in both branches must show no difference at all.
Giving each decision its own random stream cut the runs needed by up to four orders of magnitude in one agent-based study (Klein et al., 2024).
Schematic · spread of the measured effect
Each branch draws its own luck
Branches share every draw
Stated assumptions
Every effect size has a source you can read.
A policy's assumed effect comes from published studies, and the lab shows results across the published range. For attendance support, that range runs from parent letters to intensive mentoring.
| Setting | Cut in daily absence | Where it comes from |
|---|---|---|
| Cautious | 0.5 points | Randomised trials of mailings to parents in the US, which cut absence by about 6% to 8% (Rogers and Feller, 2018; Robinson et al., 2018) |
| Central | 2.5 points | A randomised Chicago trial of mentoring, which cut absences in grades 5 to 7 by 4.22 days, or 22.9% (Guryan et al., 2020, working paper) |
| Optimistic | 5 points | New York City's Success Mentors, the largest estimate in our literature review, from a comparison that was not randomised (Balfanz and Byrnes, 2013) |
By design, the engine does not turn attendance gains into higher attainment automatically. The Chicago mentoring raised attendance with no significant effect on grades or test scores.
What the results mean
A way to compare options, labelled as illustrative.
Illustrative until validated
Results carry an illustrative label until validation supports stronger claims. A calibrated fit shows the engine can reproduce published figures, and forecasting is a separate test.
Two kinds of uncertainty
Intervals from shared draws measure simulation noise only. The sensitivity analysis shows how far each result depends on the engine's assumptions about behaviour.
Checked against the past
The planned retrodiction check builds synthetic schools from one district's published figures, replays an earlier year and compares the result with what that district went on to publish.
A tool for exploring options
Critics judge true prediction "infeasible for most social phenomena" (Edmonds and ní Aodha, 2019). The lab helps you see what could happen under stated assumptions.
Precedent
Education systems have been modelled this way before.
The closest precedent is a Washington State model of how teachers sort across schools (Goldhaber et al., 2022). Calibrated on 564,296 teacher-year records, it reproduced observed exposure to novice teachers within 90% intervals.
In a Chicago model of school choice, households that weighted test scores far above distance left new, better schools struggling to survive (Maroulis et al., 2010).
Who it is for
For the people who set policy for many schools.
Local authorities and districts
See how a change might play out in schools that differ as much as yours do, before choosing which to try for real.
Ministries and regulators
Explore options at system level on schools calibrated to your own published statistics, with every assumption written down.
Each run uses a pack for your system. See which school systems have packs in build or sources mapped.
Bring us a policy question.
Tell us the decision you face and the school system it sits in, and we will say plainly whether the lab can help.