Ability grouping and tracking — four practices, four verdicts

Verdict

"Tracking" is four different practices with four different evidence bases, and the debate stays confused by pooling them (taxonomy anchor):

1. Within-class grouping (small groups inside one classroom): g≈0.19–0.30, essentially free.
2. Cross-grade subject grouping (Joplin plan — regroup across grades by subject readiness): g≈0.26, essentially free.
3. Between-class streaming with undifferentiated instruction: g≈0.04–0.06 — nothing. But the one large RCT (Duflo/Dupas/Kremer, Kenya) shows tracking with instructional adaptation raised scores across the entire distribution — bottom half +0.155 — via teach-to-level (median students gained equally in either section, rejecting peer-effects stories). Grouping works exactly insofar as instruction adapts. (Context caveats: grade 1, ~80-100-pupil baseline classes, contract teachers — don't export the magnitude.)
4. Separate high-achiever classrooms selected on demonstrated achievement: +0.5 SD for Black/Hispanic high achievers, persistent, no spillover harm, ~zero cost — while the same district's IQ-labelled gifted margin produces nothing, replicated in a second district by Bui, Craig & Imberman.

The essential scope limit (added by adversarial review): early, rigid, between-school tracking — sorting ten-year-olds into separate school types, European style — does harm low achievers on B-grade causal evidence. Two extra comprehensive years raise math and reading with the gain almost entirely in low achievers and nothing lost by high achievers; the cross-country diff-in-diff (weaker, C) points the same way; and three national detracking reforms agree — Sweden (attainment and earnings up for able children of unskilled fathers), Finland (father-son income elasticity down ~23%), and France (wages +4.7% at 40-45, concentrated in low-SES adults). Within-school flexible grouping and between-school child-sorting are different treatments; the harm evidence attaches to the latter, and where tracks are permeable, marginal track assignment has no long-run effect. Meanwhile the meta-analytic record shows no support for the claim that within-school grouping harms low-ability students — effects don't differ by ability level.

Detracking's flagship failed audit: San Francisco's celebrated reform rested on a repeat-rate statistic that was a policy-change artifact; AP-math participation fell 15% (−6pp) with ethnoracial gaps intact, and the district reversed course (source).

Hereditarian-lens assessment

Risk: low for the verdict (RCT/RD core). This is selection-confound central: raw tracked-vs-untracked comparisons mostly measure who sorts where (on substantially heritable ability), which is why only the experimental cells carry weight. The lens adds the design principle: since ability differences are real and persistent, the choice isn't whether children differ but whether instruction meets them where they are (teach-to-level, the one mechanism with RCT support) — while keeping placements permeable and re-measured, because locking children into tracks at 10 is where causal harm actually shows up.

Practical guidance

• Group by subject-specific measured readiness (a child can be ahead in math, behind in reading), with genuinely differentiated curriculum per group — otherwise don't bother.
• Reassess frequently (each term/year); permeability is what separates benign grouping from harmful tracking.
• Run achievement-selected advanced classrooms — selected on demonstrated performance, not IQ labels or teacher referral.
• Don't detrack on equity grounds — the causal record shows no within-school harm to remove, and the flagship detracking reform failed its audit.