Tutoring versus class size

Why does one-to-one tutoring transform results when cutting a class from thirty to twenty-five barely helps?

Everyone believes both halves of this: a private tutor works wonders, and smaller classes are better. The second belief is the expensive one — cutting classes by a fifth means a fifth more teachers and the rooms to put them in — and when it is done and measured, the gains are disappointing to the point of embarrassment. Yet tutoring, the same idea taken to its limit, plainly does something. If the ratio is what matters, why is the curve so flat until the very end of it?

Start with the claim that anchors the debate. In 1984 Benjamin Bloom reported that students taught one-to-one, with mastery learning, performed about two standard deviations above conventionally taught students — roughly moving an average student to the 98th percentile. That number, "2 sigma", is quoted constantly. It has not replicated at that magnitude. Later syntheses of human tutoring put the effect nearer 0.8 standard deviations, and the largest recent evidence base, a pooled analysis of dozens of randomised trials, lands near 0.4. Those are large effects by education standards, but they are not 2 sigma, and treating the original figure as a benchmark has distorted much thinking.

Now the other side. The Tennessee STAR experiment of the late 1980s randomly assigned young children to classes of roughly 13 to 17 or 22 to 25, and found gains of about 0.15 to 0.2 standard deviations, larger for disadvantaged pupils and concentrated in the earliest grades — a real effect, from a cut of about a third. When California cut class sizes statewide a decade later the measured benefit was slight, partly because staffing thousands of new classrooms at once meant hiring less experienced teachers.

Put the two together and a shape appears: something is very strong at one-to-one, weak at a third-sized cut in early grades, and undetectable at thirty-to-twenty-five. What kind of mechanism produces that?

Not a ratio. A ratio is a smooth quantity, and a smooth quantity should give you a fifth of the benefit for a fifth of the cut. So ask what a tutor has that a teacher of twenty-five lacks, and check whether the teacher of thirty acquires any of it by dropping to twenty-five.

Watch a tutor for ten minutes. The student attempts something; the tutor sees, in seconds, exactly where it went wrong — not that it was wrong, but which step, which misconception — and responds before the error is rehearsed. The next problem is chosen for this student's gap. Nothing is aimed at a median that does not exist. And silence is immediately interrogated, so nobody can sit quietly and be counted as following.

Every one of those is a property of the feedback loop, not of the headcount: how fast the teacher learns each student's state, how specific that information is, and how quickly instruction adjusts to it. In a class of thirty the information is coarse, delayed, and sampled from the few who answer — the loop closes at the pace of marked homework, days later, in aggregate. In a class of twenty-five it closes the same way. The mode of instruction is unchanged, so the loop is unchanged, so the outcome barely moves.

That also explains STAR. A class of fifteen five-year-olds is not merely 40 per cent smaller; it is small enough to change what the teacher does — more individual attention, less time on the behaviour management that dominates the youngest classrooms. Where the cut changes practice, results move. Where it only changes a number, they do not.

The claim is not that class size is irrelevant. It is that class size acts on outcomes only through what it lets a teacher do, and that relationship is lumpy rather than proportional — small cuts change nothing about the method, so they buy little; large cuts in early grades change the method, so they buy something real but modest.

Several honest qualifications. Effect sizes from different studies are not strictly comparable: they rest on different tests, ages and comparison conditions, and the small trials producing the largest numbers are exactly those most inflated by publication bias. Tutoring effects vary enormously with who does it — trained teachers outperform paraprofessionals, who outperform volunteers and parents — and with dosage, subject and grade. And tutoring at scale is not the tutoring in these trials: the programmes that worked ran in school hours with trained, supervised tutors, and staffing that nationally would meet the same dilution problem that blunted the Californian reform.

The practical inference is not "hire tutors instead of teachers", but that an intervention should be judged by whether it tightens the loop — frequent low-stakes checks and responsive re-teaching move information faster without a tutor for every child.

• Why formative assessment and mastery learning show sizeable effects without any change in staffing ratio.
• Whether computer tutoring systems close the gap to human tutors, and what they still cannot read in a student.

Every term the collection defines is gathered in the glossary.