Math curriculum choice — how much does the textbook matter?

Verdict

The "curriculum is the cheapest lever" thesis (Chingos & Whitehurst) survives only in a weakened, asymmetric form. What the causal record shows:

• The floor matters more than the ceiling. The only multi-curriculum RCT (Agodini) found three heterogeneous winners indistinguishable from each other and one stable loser (the minimal-guidance constructivist entrant) 0.22 SD behind everyone. Avoiding the worst product is the reliable win.
• The lever is era-dependent. Pre-CCSS, real differences existed: practice-heavy California Math beat three rivals by +0.05–0.08 SD schoolwide for ~4 years, and Saxon beat Scott Foresman-Addison Wesley in Indiana by +0.05–0.10 SD/yr, larger for low-income students (though that advantage faded). Post-CCSS, the same research team found textbook differences collapsed to ~0.00–0.03 SD (Blazar 6-state study) — plausibly because standards homogenized content and because teachers barely use the adopted text (7% exclusive use; 3 days of career textbook PD).
• Singapore imports are small wins, not gap-closers: +0.07 pooled in the EEF Mathematics Mastery trials, and +0.11-0.15 on the Stanford Achievement Test but NULL on the state accountability test in the same students in the US Math in Focus RCT — orders of magnitude below cross-national gaps, which are mostly population, not pedagogy.
• Don't trust WWC curriculum labels — the flagship reform text carries "potentially positive" from one weak study while an RCT co-winner is labeled "mixed" by procedure artifact (pathology exhibit).
• Early-childhood curricula show the biggest effects and the biggest fadeout: Building Blocks g=0.72 at end of pre-K on a developer-built measure, then >60% gone by the end of grade 1, ~72% of it tracking stable pre-existing differences — and the multi-site synthesis finds fadeout is faster where the control group later learns more, i.e. partly control-group catch-up.

Hereditarian-lens assessment

Risk: low (the verdict rests on the RCT and powered null studies). The fadeout mechanism decomposition is lens-relevant: ~72% of early-math fadeout tracks stable pre-existing differences — early curriculum effects are a head start on a genetically-anchored trajectory, not a trajectory change — but the Watts finding (faster fadeout where controls learn more later) means fadeout can't be read as purely genetic reversion either.

Practical guidance

• Choose once, carefully: structured, coherent, practice-heavy, explicitly-sequenced. Then invest in implementation (the thing schools demonstrably don't do — 3 days of career PD) rather than serial re-adoption.
• Hard exclusion list beats optimization: no minimal-guidance constructivist programs.
• For the knowledge-rich-curriculum angle in reading, see background knowledge — the strongest curriculum result in the database (Core Knowledge lottery) is a content effect, not a math-pedagogy one.