Do Teacher Expectations Raise Students’ IQ?

Background: the most famous classroom experiment ever run

In 1968, Rosenthal and Jacobson told elementary-school teachers that a test had identified certain pupils as “intellectual bloomers” about to surge ahead. In reality the bloomers had been picked at random. At the end of the year, the researchers reported, those randomly blessed children had gained more IQ than their classmates. The study — Pygmalion in the Classroom — became one of the most cited results in psychology, and its one-line moral, that simply raising a teacher’s expectations makes children measurably smarter, entered textbooks, teacher training, and popular culture.

What entered the textbooks less often: replications began almost immediately, and most of them failed. By 1984, Raudenbush had assembled 18 experiments using the same broad design — deceive teachers about randomly chosen pupils, then measure IQ. This page carries 19 effect-size records because the Pellegrini & Hicks experiment contributes separate aware-tester and blind-tester records.

The evidence base

Across 19 effect-size records from 18 experiments, the pooled random-effects estimate of the expectancy effect is … standard deviations, 95% CI …, p = …. Between-study heterogeneity is moderate (I² = …, Q = …, p = …).

What the evidence actually says

Read as one undifferentiated pile, the literature is deflating: pooled across all effect-size records, the average expectancy effect is small and not statistically significant. If the Pygmalion effect were the robust, general phenomenon of popular telling, this is not the forest plot it would leave behind.

But the pile is not undifferentiated, and this is where the story turns. The experiments differ in one crucial, almost embarrassing way: how long the teachers had known their pupils before the researchers lied to them. You cannot easily plant a false expectation about a child the teacher has taught for months. The length of prior teacher–pupil contact is associated, under the fitted capped-linear model, with essentially all of the between-study differences — and once you look only at the studies where the deception had room to work, the effect reappears: in studies where teachers had known their pupils for at most one week, the expectancy induction produced a significant IQ gain.

A narrow descriptive summary is neither “Pygmalion is real” nor “Pygmalion failed to replicate.” In this historical, row-wise synthesis, larger estimates were associated with at most one week of prior contact under the authored subgroup and capped-linear analyses. This does not establish a causal window.

How solid is this?

Two robustness checks are built into this document. First, no single effect-size record changes whether the pooled estimate crosses p < 0.05 — this is a narrow leave-one-record-out threshold check, not a leave-one-experiment-out analysis or proof that the result is stable in magnitude. Second, the evidence base shows no signs of publication bias — though your agent may have opinions about how confidently that can be said. Every claim above carries a document-registered rule; none has run yet on your copy. A badge reports only whether that authored rule passed, failed, or was inconclusive — never scientific truth, validity, risk of bias, or evidence quality. That is deliberate. Don’t take this article’s word for its own claims — have them tested.

Reader’s Workbench

Audit ledger

A device-persistent SHA-256 hash chain of analyses, registered-rule outcomes and decisions. It survives reloads in this browser. The chain detects edits and reordering; it is not a trusted timestamp or author identity.

Reproducibility receipt

Create a signed receipt for the current scientific-state hash and audit-chain head. The signing key is generated for this page load and rotates on reload; each receipt is individually self-signed. A published receipt fixes only the audit prefix it covers, while later reader actions form a local suffix. Save the detached artifact receipt and pin its fingerprint in an external archive or repository release before using it as authorship evidence.

Tool console

Methods note

Pooling uses a random-effects model with REML estimation of τ² (DerSimonian–Laird and fixed-effect available through the tools); heterogeneity is reported as Cochran’s Q and I²; the moderator model is a mixed-effects meta-regression on prior contact truncated at three weeks, following Raudenbush (1984); publication-bias diagnostics use Egger’s regression test. The engine is dependency-free JavaScript running in this page; selected numerical outputs for this fixture reproduce R metafor reference values to the tested precision. That is a software reproduction check, not validation of the extracted data, model assumptions, or scientific conclusion. Statistical caveat: subgroup and moderator analyses here are observational comparisons across randomized experiments and can be confounded by other study features.

References