Learning Science
The Testing Effect: What Retrieval Practice Actually Does
The testing effect is the finding that retrieving information from memory produces more durable learning than restudying the same information. It is one of the most heavily replicated findings in cognitive psychology — and also one of the most overstated in popular summaries, which tend to drop the comparison condition and the feedback requirement that determine its actual size.
The headline number, and what it's measured against
Yang, Luo, Vadillo, Yu & Shanks (2021, Psychological Bulletin, 222 studies, N = 48,478, conducted in real classrooms) report an overall testing effect of g = 0.499. But the size of the effect depends entirely on the comparison condition:
Against genuinely elaborative alternatives — not doing nothing, not rereading, but a competing study strategy that also engages with the material — the advantage nearly disappears. A tool that explains material well and elaborates well is not delivering materially inferior learning to one built purely around retrieval practice.
Feedback is the load-bearing variable, not retrieval itself
Rowland (2014, Psychological Bulletin, 159 effect sizes), reporting effect size as a function of initial retrieval success and whether feedback was given:
| Condition | g | 95% CI | k |
|---|---|---|---|
| No feedback, initial success <50% | 0.03 | [−0.21, 0.27] | 17 |
| No feedback, success 51–75% | 0.29 | [0.09, 0.49] | 31 |
| No feedback, success >75% | 0.56 | [0.42, 0.70] | 40 |
| Feedback provided | 0.73 | [0.61, 0.86] | 52 |
What this means
A failed retrieval attempt with no feedback is worth literally nothing — the confidence interval for that condition includes zero. The testing effect is substantially a re-exposure-plus-feedback effect, not a pure effect of the retrieval attempt itself.
Format matters: cued recall beats recognition
Within Rowland's (2014) same meta-analysis, cued recall on the practice test produced g = 0.61, against recognition's g = 0.29. Generation without answer options is the stronger practice format. Separately, Bertsch, Pesta, Wiscott & McDaniel (2007, Memory & Cognition, 445 effect sizes) found generating an answer produces d = 0.40 more learning than reading one.
Where retrieval practice does not transfer
Pan & Rickard (2018, Psychological Bulletin, 192 effect sizes, 122 experiments, N = 10,382) tested whether being quizzed on one piece of material improves learning of related but untested material. It largely does not:
Three moderators did survive both simultaneous fitting and the bias correction: test-format transfer (0.58), mediator cues (0.61), and application/inference transfer (0.32). The practical implication: knowledge does not reliably spread from what a learner was quizzed on to what they weren't. Coverage has to be enumerated item by item to be defensibly claimed at all.
What the effect is not caused by
Rejected mechanism
Yang et al. (2021) explicitly reject "retrieval effort" as the mechanism behind the testing effect, favoring additional exposure, transfer-appropriate processing, and motivation instead. Rowland's g = 0.03 no-feedback cell is difficult to explain under a pure effort account: if effortful retrieval itself were the active ingredient, a failed, effortful, unfed-back attempt should still help. It doesn't. "Desirable difficulty" has no effect size of its own in this literature and is a post hoc label, not a design principle.
Is the effect smaller for complex material? Genuinely contested
Live dispute
Van Gog & Sweller (2015) argued the testing effect shrinks as the complexity ("element interactivity") of the material rises. Karpicke & Aue (2015) rebutted that element interactivity was never quantitatively operationalized in that argument. Rowland's own data cuts against Van Gog & Sweller — prose material shows g = 0.58 versus single words at g = 0.39 — and Yang et al.'s classroom data shows the same pattern (concepts g = 0.644 > facts g = 0.524). The debate remains live; this is not a settled question.
Publication bias in the testing-effect literature itself
Rowland (2014) also reports a published-vs-unpublished gap of 0.58 vs. 0.25— a real signal that published studies overstate the effect — though the gap shrank to non-significance in his high-exposure subset. Separately, DellaVigna & Linos (2022, Econometrica, 126 RCTs, 23 million people) found behavioral-intervention effects average 8.7 percentage points in academic journals versus 1.4 percentage points in real government deployments — a gap the authors attribute fully to publication bias. Maier et al. (2022, PNAS 119(31)) argue nothing in adjacent nudge literatures survives bias correction at all; Mertens et al. dispute this. The general lesson: published effect sizes in this field are consistently larger than field-deployed ones.